text
stringlengths
6
128k
# Waveform Optimization and Beam Focusing for Near-field Wireless Power Transfer with Dynamic Metasurface Antennas and Non-linear Energy Harvesters Amirhossein Azarbahram, Onel L. A. López, , and Matti Latva-Aho A. Azarbahram, O. López and M. Latva-Aho are with Centre for Wireless Communications (CWC), University of Oulu, Finland, (e-mail: {amirhossein.azarbahram, onel.alcarazlopez, matti.latva-aho}@oulu.fi).This work is partially supported in Finland by the Finnish Foundation for Technology Promotion, Academy of Finland (Grants 348515 and 346208 (6G Flagship)), the Finnish-American Research and Innovation Accelerator, and by the European Commission through the Horizon Europe/JU SNS project Hexa-X-II (Grant Agreement no. 101095759). ###### Abstract Radio frequency (RF) wireless power transfer (WPT) is a promising technology for future wireless systems. However, the low power transfer efficiency (PTE) is a critical challenge for practical implementations. One of the main inefficiency sources is the power consumption and loss introduced by key components such as high-power amplifier (HPA) and rectenna, thus they must be carefully considered for PTE optimization. Herein, we consider a near-field RF-WPT system with the emerging dynamic metasurface antenna (DMA) at the transmitter and non-linear energy harvesters. We provide a mathematical framework to calculate the power consumption and harvested power from multi- tone signal transmissions. Then, we propose an approach relying on alternating optimization and successive convex approximation for waveform optimization and beam focusing to minimize power consumption while meeting energy harvesting requirements. Numerical results show that increasing the transmit tones reduces the power consumption by leveraging the rectifier’s non-linearity. Moreover, it is demonstrated that increasing the antenna length improves the performance, while both DMA and fully-digital architectures may be favorable depending on the setup. Finally, our results verify that the transmitter generates accurate energy beams pointed to devices located in the near-field, while energy beams are formed in devices’ direction in the far-field region. ###### Index Terms: Radio frequency wireless power transfer, waveform design, beamforming, dynamic metasurface antennas, near-field channels. ## I Introduction Future wireless systems will facilitate efficient and eco-friendly communication across a myriad of low-power devices, fostering a sustainable society. Achieving this requires uninterrupted connectivity among these devices and with the underlying infrastructure, all while mitigating disruptions arising from battery depletion [1, 2, 3]. This is potentially facilitated by energy harvesting (EH) technologies providing wireless charging, thus, easing the maintenance of Internet of Things (IoT) devices and increasing their lifespan. Moreover, EH may lead to improved energy efficiency and reduced emission footprints across the network [4]. EH receivers may harvest energy from two types of sources: those that exist in the surrounding environment, and those that are specifically designated for energy transmission. From a transmission perspective, the latter is supported by wireless power transfer (WPT) technologies, e.g., based on inductive coupling, magnetic resonance coupling, laser power beaming, and radio frequency (RF) radiation. Among them, RF-WPT is promising for charging multiple users relatively far from the transmitter by exploiting the broadcast nature of wireless channels. Furthermore, this can be accomplished over the same infrastructure used for wireless communications. Notably, the most important challenge toward maturing RF-WPT is related to increasing the inherently low end-to-end power transfer efficiency (PTE) [4]. Herein, we focus on RF-WPT, which is referred to as WPT in the following. ### I-A Preliminaries Figure 1: Block diagram of a typical WPT system. The power consumption and loss sources are listed under each block. The end-to-end PTE depends on the performance of the key building blocks, namely energy transmitter (ET), wireless channel, and energy receiver (ER) as illustrated in Fig. 1. At first, a signal is generated and amplified using a direct current (DC) power source at the ET. Then, it is upconverted to the RF domain and transmitted over the wireless channel. Finally, the ER converts the received RF signal to DC for EH purposes. Indeed, the end-to-end PTE comprises: DC-to-RF, RF-to-RF, and RF-to-DC power conversion efficiency, i.e., $e=\underbrace{\frac{P^{t}_{rf}}{P^{t}_{dc}}}_{e_{1}}\times\underbrace{\frac{P^{r}_{rf}}{P^{t}_{rf}}}_{e_{2}}\times\underbrace{\frac{P^{r}_{dc}}{P^{r}_{rf}}}_{e_{3}}=\frac{P^{r}_{dc}}{P^{t}_{dc}}.$ (1) In WPT, both transmitter and receiver introduce non-linearities to the signals affecting the amount of harvested power. Indeed, an appropriately designed transmit signal leveraging these non-linearities may reduce the power consumption at the transmitter and/or increases the RF-to-DC conversion efficiency at the receiver [5, 6]. Specifically, using multiple transmit frequency tones can lead to high peak-to-average power ratio (PAPR) signals at the receiver, enhancing the rectifier RF-to-DC conversion efficiency [7]. Meanwhile, the RF signal can be focused towards the ER using energy beamforming (EB), which affects the transmit/receive waveform, to cope with the channel inefficiencies captured by $e_{2}$, thereby enhancing the amount of RF power that can be harvested [4]. Notice that the active transmit components consume power to operate, while the passive elements introduce power losses, both of which impact $e_{1}$ and must be considered. Therefore, $e_{1}$, $e_{2}$, and $e_{3}$ are correlated, suggesting that their joint optimization may lead to significant gains in terms of end-to-end PTE. The end-to-end PTE is also affected by the system power consumption. One of the main factors contributing greatly to the power consumption is the transmitter’s architecture, which also determines the beamforming approach. For example, in a fully-digital structure, each antenna element necessitates a dedicated RF chain, with its corresponding high-power amplifier (HPA), consuming a significant amount of power. An HPA aims to amplify the input signal to compensate for the path loss and fading in a wireless system. Moreover, the signal amplification in the HPA requires a DC power source, which accounts for the majority of the power consumption. The significant drawback of the fully-digital architecture is the high complexity and cost, making it impractical for applications requiring massive multiple-input multiple-output (mMIMO) implementations. Alternatively, analog architectures using, e.g., passive phase shifters, are less expensive but offer limited degrees of freedom for EB. Thus, a hybrid architecture implementation combining both approaches is often more appealing in practice. Hybrid architectures offer a trade-off between complexity (cost) and beamforming flexibility [8, 9]. Although hybrid beamforming using phase shifters promotes cost reduction, it still requires complex analog networks for phase shifting. There are emerging technologies to provide hybrid beamforming capability with an even lower cost and complexity, e.g., reflective intelligent surface (RIS)-aided systems [10] and dynamic metasurface antennas (DMAs) [11]. Notice that RIS is an assisting node that provides the passive beamforming capability using reflective elements, while DMA is a transceiver consisting of configurable metamaterial elements and a limited number of RF chains. DMA avoids analog network implementation challenges and provides hybrid beamforming capability with low cost and complexity. Each of these architectures may be favorable based on the system setup. For instance, employing multiple low-cost RISs helps to cover the blind spots that are prone to weak signal reception in a large area. However, reflecting surfaces in RIS-assisted systems lack baseband processing capability to perform channel estimation and send pilot signals. Thus, acquiring accurate enough channel state information to attain a suitable passive beamforming gain might force huge overheads to the system [12]. On the other hand, DMA is a transceiver and has sufficient baseband processing capability for channel estimation. However, its implementation requires some RF chains making it more costly than RIS for large-scale implementation. All in all, both of these architectures support low-cost transmitter deployments, while the choice highly depends on the considered system setup. Interestingly, the authors in [13] utilize a system model comprising both RIS and DMA structures for uplink MIMO communication, while assuming that the channel is perfectly known. ### I-B Prior works There are many works either focusing on EB, waveform optimization, or joint waveform and beamforming design for fully-digital WPT systems. The authors in [14] utilize EB to power multiple devices in a MIMO system consisting of radio stripes, while the deployment problem of this transmit architecture is investigated in [15]. Furthermore, a low-complexity beamforming design relying on the statistics of the channel is proposed in [16] to fairly power a set of single-antenna devices. In the mentioned works, none of the practical system non-linearities are considered and the focus is on the received RF power at the devices. In [17], transmit beamforming and RF and DC combining at the receiver are leveraged to increase the received DC power in a MIMO WPT system. Although this work considers the rectifier’s non-linearity, the waveform design is not investigated. Moreover, a low-complexity waveform design for single-user setups is proposed in [18], while the large-scale multi-antenna WPT scenario is addressed in [19]. The authors in [20] leverage beamforming and a multi-sine waveform in a MIMO WPT system to enhance the harvested power. Notably, the frameworks in [18, 19, 20] consider the rectifier non-linearity. Interestingly, the authors in [21] perform waveform and beamforming optimization while considering both main non-linearity sources (HPA and rectifier) aiming to maximize the harvested DC power in a WPT system. Although most of the works on WPT systems in the literature focus on traditional fully-digital architecture, novel low-cost transmitters have also attained significant attention recently. For instance, DMA is utilized in [22, 23] for a near-field WPT system, while the authors in [24] propose a minimum- power beamforming design for meeting quality of service requirements of the users in a simultaneous wireless information and power transfer (SWIPT) system. However, none of these studies have considered the rectifier non- linearity and its impact on the harvested DC power. Notably, the joint waveform and beamforming design problem in RIS-aided WPT and SWIPT systems is investigated in [25] and [26], respectively. Furthermore, the two latter works consider the EH non-linearity at the receiver side, thus, taking into account its impact on the harvested DC power. ### I-C Contributions All in all, WPT systems have received considerable attention for some time. Still, more effort is needed to reduce the system power consumption, thus increasing the end-to-end PTE. For this, low-cost multi-antenna transmitters like DMA are appealing and may pave the way for charging devices efficiently in massive IoT deployments. Moreover, multiple studies aimed to enhance the amount of harvested DC power (with rectifier non-linearity) in far-field WPT systems or received RF power (without rectifier non-linearity) in near-field WPT. As mentioned before, the receiver does not perform RF-to-DC conversion linearly and the shape of the waveform may leverage the receiver’s RF-to-DC conversion efficiency. Thus, it is imperative to take into account the impact of the rectifier’s non-linearity in the system. To the best of our knowledge, no work has yet investigated the radiative near-field power transmission and the power consumption of a multi-antenna WPT system for meeting the EH requirements of a multi-user setup while considering the receiver non- linearity, especially when using low-cost transmitters. Herein, we aim precisely to fill this research gap. Our main contributions are as follows: 1) We formulate a joint waveform optimization and beam focusing problem for a multi-user multi-antenna WPT system with both a fully-digital and a DMA architecture. Due to the huge potential of near-field WPT systems for future practical WPT applications [3], we present our system model relying on a near- field wireless channel, which can inherently capture far-field conditions as well. Notice that there are some previous works [25, 21], which focused on increasing the amount of harvested power in far-field WPT systems while considering the receiver non-linearity. However, the literature lacks a minimum-power waveform and beamforming design (even for far-field channels), which can deal with meeting the EH requirements. This is a critical gap to fill since such formulation mimics a practical setup where the EH users inform their DC power demands and the WPT system must serve them with minimum power consumption, thus maximum end-to-end PTE. Since the HPA is an active element incurring most of the power consumption at the transmitter side, we model the power consumption of a class-B HPA as a function of its output power. Notably, our problem for fully-digital architecture shares some similarities with the one discussed in [21] as objectives and constraints are interchanged. However, the main focus of our work here is on the DMA-assisted system, which introduces much more complexity to the problem due to the coupling between the variables and their Lorentzian-constrained phase response. Note that the phase shift introduced by the metamaterial elements is correlated with their amplitude, which results in a different beamforming problem than other architectures, e.g., RIS-assisted systems [26] and phase shifter-based hybrid beamforming [9]. Mathematically, when dealing with those latter architectures, both RIS passive elements or phase shifters introduce a phase shift to the signal with constant loss111Note that in most of the works in the literature, without loss of generality, this phase shifting process by the analog network or RIS elements is considered to be lossless., while each phase-shifting configuration in DMA elements leads to a different propagation loss introduced to the transmit signal. The complexities associated with our specific problem make the existing optimization frameworks for WPT systems inapplicable to our system calling for novel approaches. 2) We propose a method relying on alternating optimization and successive convex approximation (SCA) to efficiently solve the waveform and beamforming optimization problem in the DMA-assisted WPT system. Specifically, we decouple the optimization problem to maximize the minimum received DC power by tuning the frequency response of the metamaterial elements, while minimizing the consumed power for meeting the EH requirements when optimizing the digital precoders. Generally, a huge complexity is introduced to the waveform optimization problems by time sampling since the number of samples should be relatively large to result in a reliable framework [21, 6, 20]. To cope with this, we reformulate the received DC power of the users based on the spectrum of the received waveform, which removes the time dependency in the problem. Then, the metamaterial elements and the digital precoders are alternatively optimized using SCA. Motivated by the influence of variable initialization on the SCA performance, we propose a low-complexity initialization algorithm for the digital precoders and DMA weights by leveraging the channel characteristics and dedicating RF signals to the ERs. Furthermore, the complexity of the proposed optimization framework scales with the number of users, antenna elements, and frequency tones. 3) We illustrate the convergence of the proposed optimization method numerically and show that the complexity increases with the antenna length, number of transmit tones, and number of devices. Furthermore, we provide evidence that increasing the antenna length and the number of transmit tones reduces power consumption, while it increases with the number of devices and user distance. Moreover, our findings evince that DMA performs better in terms of power consumption when the number of RF chains and transmit tones are relatively low, while the fully-digital architecture becomes favorable when the mentioned parameters are sufficiently large. This also depends on the specific HPAs’ saturation power, number of devices, and user distance. Additionally, we verify by simulation that the transmitter can accurately focus the energy on the receiver location in the near-field region, while energy beams are only formed toward specific directions in the far-field region. The remainder of the paper is structured as follows. Section II introduces the system model, including the transmit architectures, and signal and power consumption modeling. The optimization problem for a joint waveform and beamforming design, together with the proposed solving approach, are elaborated in Section III. Section IV discusses the proposed initialization algorithm, Section V presents the numerical results, while Section VI concludes the paper. Notations: Bold lower-case letters represent column vectors, while non- boldface characters refer to scalars, ${\mathbf{a}}\odot{\mathbf{b}}$ denotes the Hadamard product of $\mathbf{a}$ and $\mathbf{b}$, and $\\{x\\}$ is a set that contains $x$. The $l_{2}$-norm of a vector is denoted as $|\cdot|$. The mathematical expectation is represented by $\mathbb{E}$ and $(\cdot)^{T}$ and $(\cdot)^{\star}$ are used to indicate the transpose and conjugate of a matrix or vector, respectively. Furthermore, the real and the imaginary parts of a complex number are denoted by $\Re\\{\cdot\\}$ and $\Im\\{\cdot\\}$, respectively. Additionally, $\lfloor{\cdot}\rfloor$ is the floor operator, and $\langle{\cdot}\rangle$ denotes the phase of a complex number. ## II System Model We consider a multi-antenna WPT system to charge $M$ single-antenna EH devices. The received RF power at the ER is transferred into the rectifier input using a matching network. Then, it is converted to DC power by the rectifier, while $\bar{P}_{m}$ denotes the EH requirement of the $m$th ER. As previously mentioned, multi-tone waveforms can be exploited to leverage the rectifier non-linearity and achieve a better end-to-end PTE. Hence, we consider multi-tone signals with $N_{f}$ tones at frequencies $f_{1},f_{2},\cdots,f_{N_{f}}$ for power transmission purposes. Without loss of generality, we set $f_{n}=f_{1}+(n-1)\Delta_{f},\quad n=1,\ldots,N_{f}$, where $f_{1}$ and $\Delta_{f}$ are the lowest sub-carrier frequency and the sub-carrier spacing, respectively. ### II-A Transmit Antenna Architectures The transmitter is equipped with a uniform planar array (UPA) and $N_{rf}\geq M$ RF chains. The radiating elements are spaced uniformly, with $N_{h}$ and $N_{v}$ being the number of elements in the horizontal and vertical direction, respectively. Thus, the total number of elements is $N=N_{v}\times N_{h}$. Two types of transmit antenna architectures are considered: 1) Fully-digital architecture, which requires a dedicated RF chain for each radiating element, thus $N_{rf}=N$, as shown in Fig. 2a. In a fully-digital architecture, there is a single-stage beamforming process. Herein, $\omega_{i,n}$ is the complex weight of the $n$th frequency tone of the $i$th muti-tone waveform. Despite the high deployment cost and complexity, a fully-digital structure offers the highest number of degrees of freedom in beamforming. Figure 2: Transmit antenna architectures. (a) fully-digital architecture (left) (b) DMA-assisted architecture (right). 2) DMA-assisted architecture, which comprises $N_{v}$ waveguides, each fed by a dedicated RF chain and composed of $N_{h}$ configurable metamaterial elements. Therefore, the number of RF chains, and consequently the cost and complexity, is considerably reduced compared to digital structures, making DMA suitable for mMIMO applications. Notice that DMA-assisted systems employ a two-stage beamforming process, i.e., digital beamforming, followed by the tuning of the amplitude/phase variations introduced by the metamaterial elements, as illustrated in Fig. 2b. Herein, $q_{i,l}$ is the tunable frequency response of the $l$th metamaterial element in the $i$th waveguide, while $h_{i,l}$ is the corresponding waveguide propagation loss, which will be explained in detail later. ### II-B Channel Model In wireless communications, the region where the users are located between the Fraunhofer and Fresnel distances denoted respectively as $d_{fr}$ and $d_{fs}$, is the radiative near-field region, which is referred to as the near-field region in the following. Specifically, a device at distance $r$ from a transmitter experiences near-field conditions if$\sqrt[3]{\frac{D^{4}}{8\lambda_{1}}}=d_{fs}<d<d_{fr}=\frac{2D^{2}}{\lambda_{1}}$, where $D$ is the antenna diameter, i.e., the largest size of the antenna aperture, $\lambda_{1}=\frac{c}{f_{1}}$ is the corresponding wavelength to the system operating frequency, and $c$ is the speed of light. Notice that both system frequency and antenna form factor influence the region of operation. Therefore, by moving toward higher frequencies, e.g., millimeter wave and sub- THz bands, and/or utilizing larger antenna arrays, the far-field communication assumption regarding planar wavefronts may not be valid anymore. Instead, wavefronts impinging a receive node may be strictly spherical, thus, with advanced capabilities to focus the transmit signals on specific spatial points rather on spatial directions. Notice that one of the main applications of WPT systems is in indoor environments with line-of-sight (LOS) and near-field communication, e.g., restaurants, warehouses, and shopping malls. Thus, we employ the near-field LOS channel model described in [27]. The Cartesian coordinate of the $l$th radiating element in the $i$th row is $\mathbf{g}_{i,l}=[x_{i,l},y_{i,l},z_{i,l}]^{T}$. Additionally, $i=1,2,\ldots,N_{v}$ and $l=1,2,\ldots,N_{h}$. The channel coefficient between user $m$ and the $l$th element in the $i$th row at the $n$th sub-carrier is given by $\gamma_{i,l,m,n}=A_{i,l,m,n}e^{\frac{-j2\pi}{\lambda_{n}}d_{i,l,m}},$ (2) where ${\frac{2\pi}{\lambda_{n}}d_{i,l,m}}$ is the phase shift caused by the propagation distance of the $n$th tone, with wavelength $\lambda_{n}$, and $d_{i,l,m}=|\mathbf{g}_{m}-\mathbf{g}_{i,l}|$ is the Euclidean distance between the element and the user located at $\mathbf{g}_{m}$. Moreover, $A_{i,l,m,n}=\sqrt{F(\Theta_{i,l,m})}\frac{\lambda_{n}}{4\pi d_{i,l,m}}$ (3) is the corresponding channel gain coefficient. Here, $\Theta_{i,l,m}=(\theta_{i,l,m},\psi_{i,l,m})$ is the elevation-azimuth angle pair, and $F(\Theta_{i,l,m})$ is the radiation profile of each element. In addition, we employ the radiation profile as presented in [28], where $F(\Theta_{i,l,m})=\begin{cases}G_{t}\cos{(\theta_{i,l,m})}^{\frac{G_{t}}{2}-1},&\theta_{i,l,m}\in[0,\pi/2],\\\ 0,&\text{otherwise},\end{cases}$ (4) $G_{t}=2(b+1)$ is the transmit antenna gain, and $b$ denotes the boresight gain, which depends on the specifications of the antenna elements. Note that the channel coefficient becomes $A_{m}e^{-j\psi_{i,l,m}}$ for far-field communication, where $A_{m}$ only depends on the distance of the user $m$ from the transmitter and $\psi_{i,l,m}$ is solely determined by the user direction and the relative disposition of the antenna elements within the array. ### II-C Transmit & Receive Signals The signal at the input of the $i$th HPA is given by $x_{i}(t)=\sum_{n=1}^{N_{f}}{\omega_{i,n}e^{j2\pi f_{n}t}},\quad i=1,\ldots,N_{rf}.$ (5) The HPA introduces signal distortion and models such as the Rapp model [29] capture this non-linearity. It is shown in [21] that when the HPA operates in the non-linear regime, choosing a single-carrier waveform is preferred to a multi-carrier one. The reason is that a single-carrier waveform is less deteriorated by the adverse effect of the signal distortion caused by the HPA when operating in the non-linear regime. On the other hand, when HPAs operate in the linear regime, thus, not causing amplitude and phase distortion in the signal, multi-carrier waveforms are preferred since they leverage the rectifier’s non-linearity and enhance the harvested power performance. Note that the non-linear regime of the HPA happens near the saturation voltage. Thus, in practice, the HPAs can be properly chosen to have a suitable value of the saturation voltage based on the system setup and avoid operation in the non-linear regime. Since the aim of this work is to design multi- carrier waveforms for DMA-assisted WPT, we consider HPAs to operate in the linear regime. Mathematically, the output signal of the HPA is modeled as $x^{hpa}_{i}(t)=Gx_{i}(t)$, where $G$ is the HPA gain. The rest of the signal modeling formulation will be presented separately for different transmit architectures in the following. #### II-C1 Fully-Digital Architecture Herein, $N_{rf}=N$, thus, the real transmit signal at the output of the $l$th element in the $i$th row of the UPA can be expressed as $\displaystyle\quad x^{FD}_{i,l}(t)=\Re\biggl{\\{}x^{hpa}_{u{(i,l)}}(t)\biggr{\\}}=G\sum_{n=1}^{N_{f}}\Re\biggl{\\{}\omega_{{u{(i,l)}},n}e^{j2\pi f_{n}t}\biggr{\\}},$ (6) where ${u{(i,l)}}=(i-1)N_{h}+l$. Thereby, the RF signal at the $m$th receiver when exploiting the fully-digital architecture can be expressed as $\displaystyle y^{FD}_{m}(t)$ $\displaystyle=\sum_{i=1}^{N_{v}}\sum_{l=1}^{N_{h}}\sum_{n=1}^{N_{f}}\gamma_{i,l,m,n}x^{FD}_{i,l}(t)$ $\displaystyle=G\sum_{i=1}^{N_{v}}\sum_{l=1}^{N_{h}}\sum_{n=1}^{N_{f}}\Re\Bigl{\\{}\gamma_{i,l,m,n}\omega_{{u{(i,l)}},n}e^{j2\pi f_{n}t}\Bigr{\\}}.$ (7) Furthermore, by defining $\mathbf{w}_{n}=[\omega_{1,n},\ldots,\omega_{N,n}]^{T}$ and $\boldsymbol{\gamma}_{m,n}=[\gamma_{m,n,1,1},\gamma_{m,n,1,2},\ldots,\gamma_{m,n,N_{v},N_{h}}]^{T}$, (II-C1) can be reformulated as $y^{FD}_{m}(t)=G\sum_{n=1}^{N_{f}}\Re\Bigl{\\{}\boldsymbol{\gamma}_{m,n}^{T}\mathbf{w}_{n}e^{j2\pi f_{n}t}\Bigr{\\}}.$ (8) Figure 3: The Lorentzian constrained (the inner circle) and the ideal weights (outer circle) in the complex plane. The arrows depict the mapping between the weights. #### II-C2 DMA-assisted Architecture In metasurface antennas, the phase and amplitude that can be configured in the radiating elements are correlated due to the Lorentzian resonance. Herein, we capture this correlation by [30] $q_{i,l}\in\mathcal{Q}=\Big{\\{}{(j+e^{j\phi_{i,l}})}/{2}\Big{|}\phi_{i,l}\in[0,2\pi]\Big{\\}},\quad\forall i,l.$ (9) where $\phi_{i,l}$ are the tunable phase of the $l$th metamaterial element in the $i$th waveguide. As shown in Fig. 3, the ideal phase shifting exhibits a constant unit amplitude, i.e., no losses, while the amplitude of the Lorentzian weights depends on the configured phase. Herein, microstrip lines are used as waveguides, similar to [31, 27]. The propagation model of the signal within a microstrip is expressed as $h_{i,l}=e^{-(l-1)d_{l}(\alpha_{i}+j\beta_{i})}$, where $d_{l}$ is the inter- element distance, $\alpha_{i}$ represents the waveguide attenuation coefficient, and $\beta_{i}$ is the propagation constant. The mathematical model of the DMA is represented in Fig. 2b. Notice that the number of RF chains in the DMA is reduced to $N_{rf}=N_{v}$. Hence, the real transmit signal radiated from the $l$th element in the $i$th microstrip can be expressed as $\displaystyle x_{i,l}^{DMA}(t)$ $\displaystyle\\!=\\!G\Re\left\\{h_{i,l}q_{i,l}x_{i}^{hpa}(t)\right\\}\\!=\\!G\sum_{n=1}^{N_{f}}\Re\left\\{h_{i,l}q_{i,l}{\omega_{i,n}e^{j2\pi f_{n}t}}\right\\}.$ (10) Furthermore, the RF signal received at the $m$th user in the DMA-assisted system is given by $\displaystyle y^{DMA}_{m}(t)$ $\displaystyle=\sum_{i=1}^{N_{v}}\\!\sum_{l=1}^{N_{h}}\\!\sum_{n=1}^{N_{f}}\gamma_{i,l,m,n}x_{i,l}^{DMA}(t)$ $\displaystyle=G\sum_{i=1}^{N_{v}}\\!\sum_{l=1}^{N_{h}}\\!\sum_{n=1}^{N_{f}}\\!\Re\Bigl{\\{}\gamma_{i,l,m,n}h_{i,l}q_{i,l}\omega_{i,n}e^{j2\pi f_{n}t}\Bigr{\\}}.$ (11) Finally, we define $\displaystyle\bar{\mathbf{w}}_{n}=[\underbrace{\omega_{1,n},\ldots,\omega_{1,n}}_{N_{h}},\ldots,\underbrace{\omega_{N_{v},n},\ldots,\omega_{N_{v},n}}_{N_{h}}]^{T}\in\mathbb{C}^{N\times 1},$ $\mathbf{q}=[q_{1,1},\ldots,q_{N_{v},N_{h}}]^{T}$, $\mathbf{h}=[h_{1,1},\ldots,h_{N_{v},N_{h}}]^{T}$, and reformulate (II-C2) as $y^{DMA}_{m}(t)=G\sum_{n=1}^{N_{f}}\Re\Bigl{\\{}(\boldsymbol{\gamma}_{m,n}\odot\mathbf{q}\odot\mathbf{h})^{T}\bar{\mathbf{w}}_{n}e^{j2\pi f_{n}t}\Bigr{\\}}.$ (12) ### II-D Rectenna At the receiver side, the RF signal is converted to DC. This can be modeled by an antenna equivalent circuit and a single diode rectifier as illustrated in Fig. 4. The RF signal at the input of the antenna is denoted as $y_{m}(t)$ and has an average power of $\mathbb{E}\bigl{\\{}{y_{m}(t)}^{2}\bigr{\\}}$. Let us denote the input impedance of the rectifier and the impedance of the antenna equivalent circuit by $R_{in}$ and $R_{ant}$, respectively. Thus, assuming perfect matching $(R_{in}=R_{ant})$, the input voltage at the rectifier of the $m$th ER is given by $v_{{in},{m}}(t)=y_{m}(t)\sqrt{R_{ant}}$. Furthermore, the diode current can be formulated as $i_{d}(t)=i_{s}\bigl{(}e^{\frac{v_{d}(t)}{\hat{n}v_{t}}}-1\bigr{)},$ (13) where $i_{s}$ is the reverse bias saturation current, $\hat{n}$ is the ideality factor, $v_{t}$ is the thermal voltage, and $v_{d}(t)=v_{in}(t)-v_{o}(t)$ is the diode voltage. Moreover, $v_{{o},{m}}(t)$ is the output voltage of the $m$th rectifier, which can be approximated utilizing the Taylor expansion as [20, 19] $v_{{o},{m}}=\sum_{i\ even,i\geq 2}^{n_{0}}K_{i}\mathbb{E}\bigl{\\{}y_{m}(t)^{i}\bigr{\\}},$ (14) where $K_{i}=\frac{R_{ant}^{i/2}}{i!{(\eta_{0}v_{t})}^{i-1}}$. Herein, we focus on the low-power regime, for which it was demonstrated in [19, 6] that truncating the Taylor expansion at $n_{0}=4$ is accurate enough. Therefore, (14) can be written as $v_{{o},{m}}=K_{2}\mathbb{E}\bigl{\\{}y_{m}(t)^{2}\bigr{\\}}+K_{4}\mathbb{E}\bigl{\\{}y_{m}(t)^{4}\bigr{\\}}$ (15) and the DC power at the $m$th receiver is given by $P_{{dc},m}=\frac{{v^{2}_{{o},{m}}}}{R_{L}},$ (16) where $R_{L}$ is the load impedance of the rectifier, while $y_{m}(t)$ is equal to $y_{m}^{DMA}(t)$ and $y_{m}^{FD}(t)$ in the DMA and fully-digital. Figure 4: Antenna equivalent circuit (left) and a single diode rectifier (right) [20]. ### II-E Power Consumption Model Lowering the system power consumption is desirable for increasing the end-to- end PTE and this greatly depends on the power consumption of the HPA. Let us denote the maximum efficiency and the maximum output power of a class-B HPA as $\bar{\eta}$ and $P_{max}$, respectively. Hereby, the efficiency of the $i$th HPA at time $t$ is expressed as $\eta_{i}(t)={\bar{\eta}}\sqrt{{P_{out,i}(t)}/{P_{max}}}$ [32], where $P_{out,i}(t)$ is the output power of the HPA. Then, the corresponding power consumption is $P_{hpa,i}(t)=\frac{P_{out,i}(t)}{\eta_{i}(t)}=\frac{1}{\bar{\eta}}\sqrt{{P_{max}{P_{out,i}(t)}}}.$ (17) Notice that each HPA is in charge of supplying the amount of radiating power by all of the antenna elements, which are fed by its corresponding RF chain. Thus, $P_{out,i}(t)$ is $\sum_{l=1}^{N_{h}}|x^{DMA}_{i,l}(t)|^{2}$ in DMA- assisted architecture, and $|x^{FD}_{i,l}(t)|^{2}$ denotes the output power of the RF chain connected to the $l$th element in the $i$th row. There are also other power consumption sources in the WPT system. For instance, digital baseband power consumption, which is required to perform the digital beamforming, and the RF chain circuit power consumption, including the mixer, local oscillator, and filter. However, the power consumption of these sources is usually considered fixed and is negligible compared to the HPA. Thus, without loss of generality, the total power consumption of the system is given by $P_{c}=\sum_{i=1}^{N_{rf}}\mathbb{E}\bigl{\\{}P_{{hpa},i}(t)\bigr{\\}}+P_{in},$ (18) where $P_{in}=\sum_{i=1}^{N_{rf}}\sum_{n=1}^{N_{f}}|\omega_{i,n}|^{2}$ is the total input power. ## III Joint Waveform & Beamforming Optimization This section formalizes the optimization problem and describes the utilized approach when employing the aforementioned transmit architectures. ### III-A Problem Formulation The goal is to obtain a minimum-power waveform and beamforming design to meet the EH requirements of the users. Thus, by utilizing (16) and substituting (17) in (18), the optimization problem can be formulated as $\displaystyle\operatorname*{minimize}_{\mathcal{V}}\quad$ $\displaystyle\frac{\sqrt{P_{max}}}{\bar{\eta}}\sum_{i=1}^{N_{rf}}\mathbb{E}\bigl{\\{}\sqrt{P_{out,i}}(t)\bigr{\\}}+\sum_{i=1}^{N_{rf}}\sum_{n=1}^{N_{f}}|\omega_{i,n}|^{2}$ (19a) subject to $\displaystyle v_{o,m}^{2}\geq R_{L}\bar{P}_{m},\quad m=1,\ldots,M,$ (19b) where $\mathcal{V}$ is the set of optimization variables, which is equal to $\\{\omega_{i,n},q_{i,l}\\}_{\forall i,l,n}$ and $\\{\omega_{i,n}\\}_{\forall i,n}$ for DMA-assisted and fully-digital architectures, respectively. Problem (19) deals with extensive non-linearity since the objective and constraints are highly non-linear and non-convex functions due to the signal model, rectenna non-linearity, and the coupling between the digital and analog beamforming variables in DMA. ### III-B Optimization Framework for DMA-assisted Architecture One of the challenges making the problem intractable is the coupling between the optimization variables. To cope with this, we propose using alternating optimization by first optimizing the digital precoders when considering fixed $\\{q_{i,l}\\}_{\forall i,l}$, followed by optimizing the metamaterial elements’ frequency response for fixed $\\{\omega_{i,n}\\}_{\forall i,n}$. #### III-B1 Optimization with fixed $\\{q_{i,l}\\}_{\forall i,l}$ Let us proceed by rewriting the optimization problem as $\displaystyle\operatorname*{minimize}_{\\{\omega_{i,n}\\}_{\forall i,n}}\quad$ $\displaystyle\frac{\sqrt{P_{max}}}{\bar{\eta}}\sum_{i=1}^{N_{v}}\mathbb{E}\biggl{\\{}\sqrt{\sum_{l=1}^{N_{h}}x^{DMA}_{i,l}(t)^{2}}\biggr{\\}}\\!+\\!\sum_{i=1}^{N_{v}}\sum_{n=1}^{N_{f}}|\omega_{i,n}|^{2}$ (20a) subject to $\displaystyle\bigl{(}{K_{2}\mathbb{E}\bigl{\\{}y_{m}(t)^{2}\bigr{\\}}+K_{4}\mathbb{E}\bigl{\\{}y_{m}(t)^{4}\bigr{\\}}\bigr{)}}^{2}\geq{R_{L}}\bar{P}_{m},\forall m.$ (20b) The problem is still highly non-linear and non-convex. Interestingly, it can be easily verified that $v_{o,m}$ is convex and increasing with respect to ${y_{m}(t)}^{2}$, and ${y_{m}(t)}$ is affine with respect to $\bar{\mathbf{w}}$. Therefore, $v_{o,m}$ is a convex function with respect to $\bar{\mathbf{w}}$ given a fixed configuration for metamaterial elements [21]. Thus, although (20b) is a non-convex constraint, its left-hand side consists of a convex function. These properties motivate us to adapt the SCA method [33] to optimize the problem iteratively. Specifically, the objective and constraints can be approximated by their Taylor expansion. ###### Theorem 1. By defining $S_{i}=\sum_{l=1}^{N_{h}}\mathbb{E}\bigl{\\{}x^{DMA}_{i,l}(t)^{2}\bigr{\\}}$, we can write $\sum_{i=1}^{N_{rf}}\mathbb{E}\biggl{\\{}\sqrt{\sum_{l=1}^{N_{h}}x^{DMA}_{i,l}(t)^{2}}\biggr{\\}}\leq\sum_{i=1}^{N_{rf}}\tilde{f}\bigl{(}S_{i},S_{i}^{(0)}),$ (21) where $\tilde{f}\bigl{(}S_{i},S_{i}^{(0)})=\sqrt{S_{i}^{(0)}}+\frac{1}{2\sqrt{S_{i}^{(0)}}}\bigl{(}S_{i}-S_{i}^{(0)}\bigr{)},\forall i\vspace{-2mm}$ (22) is the first order Taylor expansion of ${\sqrt{S_{i}}}$ at point $S_{i}^{(0)}$. ###### Proof. The inequality can be proved using Jensen’s inequality [34] and the fact that ${\sqrt{S_{i}}}$ is a concave function with respect to ${S_{i}}$, while the first-order Taylor expansion of a function is greater than or equal to the function value at each point. ∎ Then, we rewrite $v_{o,m}$ considering fixed $q_{i,l}$. For this, we define $\mathbf{a}_{m,n}=\boldsymbol{\gamma}_{m,n}\odot\mathbf{q}\odot\mathbf{h}$ and leverage the fact that the average power of a signal is equal to the power of its spectrum. Thus, we have [6] $\displaystyle\mathbb{E}\bigl{\\{}y_{m}(t)^{2}\bigr{\\}}=g_{m,1}(\\{\bar{\mathbf{w}}_{n}\\}_{\forall n})=\frac{G^{2}}{2}\sum_{n}|\mathbf{a}^{T}_{m,n}\bar{\mathbf{w}}_{n}|^{2},$ (23) $\displaystyle\mathbb{E}\bigl{\\{}y_{m}(t)^{4}\bigr{\\}}=g_{m,2}(\\{\bar{\mathbf{w}}_{n}\\}_{\forall n})=\frac{3G^{4}}{8}\sum_{\begin{subarray}{c}n_{0},n_{1},n_{2},n_{3}\\\ n_{0}+n_{1}=n_{2}+n_{3}\end{subarray}}...$ $\displaystyle(\mathbf{a}^{T}_{m,n_{0}}\bar{\mathbf{w}}_{n_{0}})(\mathbf{a}^{T}_{m,n_{1}}\bar{\mathbf{w}}_{n_{1}})(\mathbf{a}^{T}_{m,n_{2}}\bar{\mathbf{w}}_{n_{2}})^{\star}(\mathbf{a}^{T}_{m,n_{3}}\bar{\mathbf{w}}_{n_{3}})^{\star}.$ (24) By leveraging (23) and (III-B1), we can write ${{v}_{o,m}}\geq K_{2}g_{m,1}(\\{\bar{\mathbf{w}}_{n}^{(0)}\\}_{\forall n})+{K_{4}}g_{m,2}(\\{\bar{\mathbf{w}}_{n}^{(0)}\\}_{\forall n})\\\ +\sum_{n}\tilde{g}_{m,n}(\bar{\mathbf{w}}_{n}^{(0)})\bigl{(}\bar{\mathbf{w}}_{n}-\bar{\mathbf{w}}_{n}^{(0)}\bigr{)},$ (25) where $\tilde{g}_{m,n}(\bar{\mathbf{w}}_{n}^{(0)})=G^{2}K_{2}(\mathbf{a}^{T}_{m,n}\bar{\mathbf{w}}_{n}^{(0)})\mathbf{a}^{T}_{m,n}+\\\ \frac{3K_{4}G^{4}}{8}\biggl{[}4{|\mathbf{a}^{T}_{m,n}|}^{4}{|\bar{\mathbf{w}}_{n}^{(0)}|}^{2}{{\bar{\mathbf{w}}{{}_{n}^{(0)}}}}^{T}+\\\ 8\sum_{n1}{|\mathbf{a}^{T}_{m,n}|}^{2}{|\mathbf{a}^{T}_{m,n_{1}}|}^{2}{|\bar{\mathbf{w}}_{n_{1}}^{(0)}|}^{2}{{\bar{\mathbf{w}}{{}_{n}^{(0)}}}}^{T}+\\\ \sum_{\begin{subarray}{c}n_{2},n_{3}\\\ n_{2}+n_{3}=2n\\\ n_{2}\neq n_{3}\end{subarray}}2{(\mathbf{a}^{T}_{m,n_{2}}\bar{\mathbf{w}}_{n_{2}}^{(0)})}^{\star}{(\mathbf{a}^{T}_{m,n_{3}}\bar{\mathbf{w}}_{n_{3}}^{(0)})}^{\star}{(\mathbf{a}^{T}_{m,n}\bar{\mathbf{w}}_{n}^{(0)})}\mathbf{a}^{T}_{m,n}+\\\ 2(\mathbf{a}^{T}_{m,n_{2}}\bar{\mathbf{w}}_{n_{2}}^{(0)})(\mathbf{a}^{T}_{m,n_{3}}\bar{\mathbf{w}}_{n_{3}}^{(0)}){{(\mathbf{a}^{T}_{m,n}\bar{\mathbf{w}}_{n}^{(0)})}}^{\star}{\mathbf{a}^{H}_{m,n}}+\\\ \sum_{\begin{subarray}{c}n_{1},n_{2},n_{3}\\\ -n_{1}+n_{2}+n_{3}=n\\\ n\neq n_{1}\neq n_{2}\neq n_{3}\end{subarray}}2(\mathbf{a}^{T}_{m,n_{1}}\bar{\mathbf{w}}_{n_{1}}^{(0)}){(\mathbf{a}^{T}_{m,n_{2}}\bar{\mathbf{w}}_{n_{2}}^{(0)})}^{\star}{(\mathbf{a}^{T}_{m,n_{3}}\bar{\mathbf{w}}_{n_{3}}^{(0)})}^{\star}\mathbf{a}^{T}_{m,n}+\\\ 2(\mathbf{a}^{T}_{m,n_{1}}\bar{\mathbf{w}}_{n_{1}}^{(0)})(\mathbf{a}^{T}_{m,n_{2}}\bar{\mathbf{w}}_{n_{2}}^{(0)}){(\mathbf{a}^{T}_{m,n_{3}}\bar{\mathbf{w}}_{n_{3}}^{(0)})}^{\star}{\mathbf{a}^{H}_{m,n}}\biggr{]}$ (26) is the first-order Taylor coefficient of ${{v}_{o,m}}$ at point $\\{\bar{\mathbf{w}}_{n}^{(0)}\\}_{\forall n}$. Similarly, it can be verified that $\mathbb{E}\bigl{\\{}x^{DMA}_{i,l}(t)^{2}\bigr{\\}}=\frac{G^{2}}{2}\sum_{n}|q_{i,l}h_{i,l}w_{i,n}|^{2}.$ (27) Now, we can reformulate the problem at point $\\{\bar{\mathbf{w}}_{n}^{(0)},S_{i}^{(0)},{v}_{o,m}^{(0)}\\}_{\forall i,n,m}$ as $\displaystyle\operatorname*{minimize}_{\begin{subarray}{c}\\{\bar{\mathbf{w}}_{n}\\}_{\forall n}\\\ \\{{v}_{o,m}\\}_{\forall m}\\\ \\{S_{i}\\}_{\forall i}\end{subarray}}\quad$ $\displaystyle\frac{\sqrt{P_{max}}}{\bar{\eta}}\sum_{i=1}^{N_{rf}}\tilde{f}\bigl{(}S_{i},S_{i}^{(0)})+\sum_{i=1}^{N_{rf}}\sum_{n=1}^{N_{f}}|\omega_{i,n}|^{2}$ (28a) subject to $\displaystyle\sqrt{{R_{L}}\bar{P}_{m}}\leq K_{2}g_{m,1}(\\{\bar{\mathbf{w}}_{n}^{(0)}\\}_{\forall n})+$ $\displaystyle K_{4}g_{m,2}(\\{\bar{\mathbf{w}}_{n}^{(0)}\\}_{\forall n})+$ $\displaystyle\sum_{n}\tilde{g}_{m,n}(\bar{\mathbf{w}}_{n}^{(0)})\bigl{(}\bar{\mathbf{w}}_{n}-\bar{\mathbf{w}}_{n}^{(0)}\bigr{)},\quad\forall m$ (28b) $\displaystyle\frac{G^{2}}{2}\sum_{n=1}^{N_{f}}\sum_{l=1}^{N_{h}}\bigl{|}q_{i,l}h_{i,l}\bar{\mathbf{w}}_{n}[i]\bigr{|}^{2}\leq S_{i},i=1,\ldots,N_{v},$ (28c) $\displaystyle\bar{\mathbf{w}}_{n}[(i-1)N_{h}+l]=\bar{\mathbf{w}}_{n}[i],\forall i,l=1,\ldots,N_{h}.$ (28d) Notice that by utilizing (21) and the fact that (20a) consists of two positive terms, one can verify that (28a) serves as an upper bound for (20a), thus, minimizing the latter leads to minimizing the former. Moreover, the inequality in (25) ensures that the solution to this problem is a feasible solution of (20) at each point. Interestingly, the problem222Note that for the sake of notation and facilitating the reader’s understanding, we kept the problem in the vector form and introduced the constraint (28d) into the problem. However, the problem can be easily converted to scalar form, which removes the mentioned constraint. has become convex and can be solved at a given point by standard convex optimization tools, e.g., CVX [35]. Moreover, the solution can be iteratively updated using the SCA algorithm [33]. #### III-B2 Optimization with fixed $\\{\omega_{i,n}\\}_{\forall i,n}$ Herein, the non-convex Lorentzian constraint of the metamaterials makes the problem extremely difficult to solve. To tackle this, we propose decoupling the problem to first maximize the minimum harvested power when optimizing $q_{i,l}$. This allows us to leverage the beamforming capability of the metamaterial elements and provide degrees of freedom to further reduce the power consumption when optimizing $\omega_{i,n}$ [23]. Hereby, the optimization problem with fixed $\\{\omega_{i,n}\\}_{\forall i,n}$ can be reformulated as $\displaystyle\operatorname*{maximize}_{\\{q_{i,l}\\}_{\forall i,l}}\quad$ $\displaystyle\min_{m}\frac{v_{o,m}^{2}}{R_{L}}$ (29a) subject to $\displaystyle q_{i,l}\in\mathcal{Q},\forall i,l,$ (29b) where (29b) is the non-convex Lorentzian constraint. Next, we cope with the complexity caused by the metamaterial elements. ###### Theorem 2. Problem (29) is equivalent to $\displaystyle\operatorname*{maximize}_{\\{q_{i,l}\\}_{\forall i,l}}\quad$ $\displaystyle R$ (30a) subject to $\displaystyle\Re\bigl{\\{}q_{i,l}\bigr{\\}}^{2}+(\Im\bigl{\\{}q_{i,l}\bigr{\\}}-0.5)^{2}\leq 0.25,\forall i,l,$ (30b) $\displaystyle R\leq K_{2}{E}\bigl{\\{}y_{m}(t)^{4}\bigr{\\}}+K_{4}{E}\bigl{\\{}y_{m}(t)^{4}\bigr{\\}},\quad\forall m.$ (30c) ###### Proof. The proof is provided in Appendix -A. ∎ Notice that (30b) keeps the frequency response of the metamaterials within the Lorentzian circle, while (30c) ensures that $R$ is below the minimum output voltage of the devices. Problem (30) is still difficult to solve due to the non-convex constraint (30c). To cope with this, we define $\hat{\mathbf{a}}_{m,n}=\boldsymbol{\gamma}_{m,n}\odot\bar{\mathbf{w}}_{n}\odot\mathbf{h}$ and write $\displaystyle\mathbb{E}\bigl{\\{}y_{m}(t)^{2}\bigr{\\}}={e}_{m,1}({\mathbf{q}})=\frac{G^{2}}{2}\sum_{n}|\hat{\mathbf{a}}^{T}_{m,n}{\mathbf{q}}|^{2},$ (31) $\displaystyle\mathbb{E}\bigl{\\{}y_{m}(t)^{4}\bigr{\\}}={e}_{m,2}(\mathbf{q})=\frac{3G^{4}}{8}\sum_{\begin{subarray}{c}n_{0},n_{1},n_{2},n_{3}\\\ n_{0}+n_{1}=n_{2}+n_{3}\end{subarray}}...$ $\displaystyle(\hat{\mathbf{a}}^{T}_{m,n_{0}}\mathbf{q})(\hat{\mathbf{a}}^{T}_{m,n_{1}}\mathbf{q})(\hat{\mathbf{a}}^{T}_{m,n_{2}}\mathbf{q})^{\star}(\hat{\mathbf{a}}^{T}_{m,n_{3}}\mathbf{q})^{\star}.$ (32) Similar to the case of digital precoders, it can be observed that ${v}_{o,m}$ is convex with respect to $\mathbf{q}$, thus, we can write ${{v}_{o,m}}\geq K_{2}{e}_{m,1}({\mathbf{q}}^{(0)})+K_{4}{e}_{m,2}({\mathbf{q}}^{(0)})+\tilde{e}_{m}(\mathbf{q}^{(0)})\bigl{(}{\mathbf{q}}-{\mathbf{q}}^{(0)}\bigr{)},$ (33) where $\tilde{e}_{m}(\mathbf{q}^{(0)})=G^{2}K_{2}\sum_{n}(\hat{\mathbf{a}}^{T}_{m,n}{\mathbf{q}}^{(0)})\hat{\mathbf{a}}^{T}_{m,n}+\\\ \frac{3K_{4}G^{4}}{8}\sum_{\begin{subarray}{c}n_{0},n_{1},n_{2},n_{3}\\\ n_{0}+n_{1}=n_{2}+n_{3}\end{subarray}}\biggl{[}(\hat{\mathbf{a}}^{T}_{m,n_{1}}\mathbf{q})(\hat{\mathbf{a}}^{T}_{m,n_{2}}\mathbf{q})^{\star}(\hat{\mathbf{a}}^{T}_{m,n_{3}}\mathbf{q})^{\star}\hat{\mathbf{a}}^{T}_{m,n_{0}}+\\\ (\hat{\mathbf{a}}^{T}_{m,n_{0}}\mathbf{q})(\hat{\mathbf{a}}^{T}_{m,n_{2}}\mathbf{q})^{\star}(\hat{\mathbf{a}}^{T}_{m,n_{3}}\mathbf{q})^{\star}\hat{\mathbf{a}}^{T}_{m,n_{1}}+\\\ (\hat{\mathbf{a}}^{T}_{m,n_{0}}\mathbf{q})(\hat{\mathbf{a}}^{T}_{m,n_{1}}\mathbf{q})(\hat{\mathbf{a}}^{T}_{m,n_{3}}\mathbf{q})^{\star}\hat{\mathbf{a}}_{m,n_{2}}^{H}+\\\ (\hat{\mathbf{a}}^{T}_{m,n_{0}}\mathbf{q})(\hat{\mathbf{a}}^{T}_{m,n_{1}}\mathbf{q})(\hat{\mathbf{a}}^{T}_{m,n_{2}}\mathbf{q})^{\star}\hat{\mathbf{a}}_{m,n_{3}}^{H}\biggr{]}.$ (34) Hereby, (30) can be reformulated at point $\mathbf{q}^{(0)}$ as $\displaystyle\operatorname*{maximize}_{\mathbf{q},R}\quad$ $\displaystyle R$ (35a) subject to $\displaystyle R\leq K_{2}\bar{g}_{m,1}({\mathbf{q}}^{(0)})$ $\displaystyle+K_{4}\bar{g}_{m,2}({\mathbf{q}}^{(0)})+\tilde{e}_{m}(\mathbf{q}^{(0)})\bigl{(}{\mathbf{q}}-{\mathbf{q}}^{(0)}\bigr{)},$ (35b) $\displaystyle\Re\bigl{\\{}\mathbf{q}[(i-1)N_{h}+l]\bigr{\\}}^{2}+$ $\displaystyle(\Im\bigl{\\{}\mathbf{q}[(i-1)N_{h}+l]\bigr{\\}}-0.5)^{2}\leq 0.25,\forall i,l,\eqref{probdmaQQconf}$ (35c) which is convex in the neighborhood of the initial point. #### III-B3 Alternating Optimization Algorithm We transformed the original problem into a convex form for both fixed $\bar{\mathbf{w}}$ and $\mathbf{q}$ in the neighborhood of an initial point. However, there is still an important challenge, i.e., the initialization of the variables. Specifically, when using iterative algorithms relying on convex approximation, e.g., SCA, the starting point must be feasible and the performance is influenced by it. One can get a feasible point by setting an extremely large amplitude for the digital weights, but this may lead to poor performance. To cope with this, we propose a low-complexity initialization method, which will be explained in the next section. This prevents the initial consumed power from becoming extremely large, helping the SCA algorithm to start from a relatively good initial point. Algorithm 1 illustrates the proposed alternating SCA-based approach for waveform and beamforming optimization. First, the digital precoders and the frequency responses of the metamaterial elements are initialized. After that, digital precoders and metamaterials are optimized in an alternating fashion in lines 3-17. First, SCA is used to iteratively find a suboptimal solution for fixed digital precoders. Specifically, $\mathbf{q}$ is updated in each iteration until convergence in lines 5-8. Then, the obtained $\mathbf{q}$ is used to run the SCA algorithm for finding suboptimal digital precoders $\\{\bar{\mathbf{w}}_{n}\\}_{\forall n}$ through line 11-15. These two SCA- based optimizations are repeated until the alternating optimization converges to a suboptimal solution. Algorithm 1 Alternating SCA-based waveform and beamforming design for DMA- assisted WPT (ASCA-DMA). 1:Input: $\\{\gamma_{i,l,m,n}\\}_{\forall i,l,m,n}$, $\upsilon$ Output: $P_{c}^{\star}$ 2:Initialize: $\mathbf{q}^{(0)}$ and $\bar{\mathbf{w}}^{(0)}_{n},\forall n$, $P_{c}^{\star}=0$ 3:repeat 4: $\xi^{\star}_{1}=0$, $\xi^{\star}_{2}=\infty$, $P_{c}\leftarrow P_{c}^{\star}$ 5: repeat 6: $\xi_{1}\leftarrow\xi^{\star}_{1}$, solve (35) to obtain $\mathbf{q}$ and $\mathbf{q}^{(0)}\leftarrow\mathbf{q}$ 7: $\xi^{\star}_{1}\leftarrow$the objective value in (35a) 8: until $|1-{\xi_{1}^{\star}}/{\xi_{1}}|\leq\upsilon$ 9: Compute $v^{(0)}_{out,m},\forall m$ using (15), 23, and III-B1 10: Compute $S_{i}^{(0)},\forall i$ using (27) 11: repeat 12: $\xi_{2}\leftarrow\xi^{\star}_{2}$, solve (28) to obtain $v_{o,m}$ and $\bar{\mathbf{w}}^{(0)}_{n},\forall n$ 13: $v^{(0)}_{out,m}\leftarrow v_{o,m}$, $\bar{\mathbf{w}}^{(0)}_{n}\leftarrow\bar{\mathbf{w}}^{(0)}_{n},\forall n$, $S_{i}^{(0)}\leftarrow S_{i},\forall i$ 14: $\xi^{\star}_{2}\leftarrow$ the objective value in (28a) 15: until $|1-{\xi_{2}^{\star}}/{\xi_{2}}|\leq\upsilon$ 16: Compute $P_{c}^{\star}$ using (19a) 17:until $|1-P_{c}^{\star}/P_{c}|\leq\upsilon$ ### III-C Optimization Framework for Fully-Digital Architecture Note that the proposed framework for the DMA-assisted system can also be used for fully-digital architecture with some slight modifications333Note that the framework can also be straightforwardly adapted for a traditional hybrid architecture with a fully connected network of phase shifters.. Notably, alternating optimization is not needed since the variables are just digital precoders, and adapting SCA to find the suboptimal precoders is sufficient. Algorithm 2 illustrates the SCA-based optimization for fully-digital architecture. Notice that problem (28) can be easily modified to match the fully-digital transmitter. Thus, the expressions are not rewritten to avoid repetition. Algorithm 2 SCA-based waveform and beamforming design for fully-digital WPT (SCA-FD). 1:Input: $\\{\gamma_{i,l,m,n}\\}_{\forall i,l,m,n}$, $\upsilon$ Output: $P_{c}^{\star}$ 2:Initialize: ${\mathbf{w}}^{(0)}_{n},\forall n$, $P_{c}^{\star}=0$ 3:repeat 4: $P_{c}\leftarrow P_{c}^{\star}$ , solve (28) to obtain $\\{v_{o,m},S_{i},{\mathbf{w}}_{n}\\}_{\forall i,m,n}$ 5: $v^{(0)}_{out,m}\leftarrow v_{o,m}$, ${\mathbf{w}}^{(0)}_{n}\leftarrow{\mathbf{w}}^{(0)}_{n},\forall n$, $S_{i}^{(0)}\leftarrow S_{i},\forall i$ 6: Compute $P_{c}^{\star}$ using (19a) 7:until $|1-P_{c}^{\star}/P_{c}|\leq\upsilon$ ### III-D Complexity The proposed ASCA-DMA algorithm consists of a low-complexity initialization, followed by alternating optimization, while SCA is used to optimize each set of variables. Each iteration of SCA attempts to solve a quadratic program [33] ((28) or (35)). Moreover, the complexity of quadratic programs scales with a polynomial function of the problem size, while the degree of the polynomial mainly depends on the type of the solver. Let us consider a simple solver based on the Newton method with $\mathcal{O}(n^{3})$ complexity [33], where $n$ is the problem size. Imagine $U_{1}$ and $U_{2}$ are the number of required iterations (in the worst-case) for convergence of digital weights and metamaterial elements’ weights, respectively. Furthermore, $U_{3}$ is considered the number of iterations required for convergence in alternating optimization. Hereby, the total complexity of the ASCA-DMA algorithm is $\mathcal{O}(U_{1}U_{2}U_{3}n^{3})$, where $n$ scales with $M$, $N_{v}$, $N_{h}$, and $N_{f}$. There is also some additional complexity introduced by the initialization algorithm, which is negligible since the initialization procedure is low-complexity. Moreover, the SCA-FD has only a single SCA stage with a complexity $\mathcal{O}(U_{4}n^{3})$, where $U_{4}$ is the required number of iterations for convergence of SCA in the worst-case. ## IV Initialization Algorithm Algorithm 3 Initialization of digital precoders and the frequency response of the metamaterial elements. 1:Input: $\\{\gamma_{i,l,m,n}\\}_{\forall i,l,m,n}$, $\tau_{s}$, $\varsigma$ Output: $\\{\omega_{i,n},q_{i,l}\\}_{\forall i,l,n}$ 2:Initialize: Compute $z_{m},\forall m$ using (36) 3:Allocate one RF chain to each user, $\mathcal{R}_{m}=\\{m\\},\forall m$, $\bar{\mathcal{R}}=\\{1,\ldots,M\\}$, $w_{m}=0,\forall m$ 4:$N^{\prime}_{v}=N_{v}-M$, $N^{\prime}_{rf,m}=\lceil z_{m}N^{\prime}_{v}\rceil$, $\mathcal{R}^{\prime}=\\{\\}$ 5:repeat 6: $m^{\star}=\operatorname*{argmax}_{m\notin\mathcal{R}^{\prime}_{m}}z_{m}$, $\mathcal{R}^{\prime}\leftarrow\mathcal{R}^{\prime}\cup m^{\star}$, $i_{c}=M+1$ 7: repeat 8: $\mathcal{R}_{m^{\star}}\leftarrow\mathcal{R}_{m^{\star}}\cup i_{c}$, $i_{c}\leftarrow i_{c}+1$ 9: $N^{\prime}_{rf,m}\leftarrow N^{\prime}_{rf,m}-1$, $N^{\prime}_{v}\leftarrow N^{\prime}_{v}-1$ 10: until $N^{\prime}_{rf,m}=0$ or $N^{\prime}_{v}=0$ 11:until $N^{\prime}_{v}=0$ or $|\mathcal{R}^{\prime}|=M$ 12:for $m=1,\ldots,M$ do 13: Compute $q_{i,l},\forall i\in\mathcal{R}_{m},l$ using (37) and (9), $w_{m}=\tau_{s}$ 14: Solve (40) to obtain $\hat{\omega_{i,n}}^{\star},\forall i\in\mathcal{R}_{m},n$ 15: repeat 16: $\bar{\omega}_{i,n}=w_{m},\omega_{i,n}=\bar{\omega}_{i,n}e^{j\hat{\omega}_{i,n}}\forall i\in\mathcal{R}_{m},n$ 17: Compute $P_{dc,m}$ using (15), (16), (23), and (III-B1) 18: $w_{m}\leftarrow\varsigma w_{m}$ 19: until $P_{dc,m}\geq\bar{P}_{m}$ 20:end for Algorithm 3 illustrates the proposed initialization algorithm. Since the initialization algorithm has to be adaptable for multi-user scenarios, we start by proposing a method to allocate the output signal of the RF chains to the different users444The allocation is only for the initialization process, and there is no limitation in this regard in the optimization procedure.. For this, we utilize the channel characteristics by naming $n_{m}^{\star}=\operatorname*{argmax}_{n}|\boldsymbol{\gamma}_{m,n}|$ as the strongest sub-carrier channel between user $m$ and the transmitter. Then, a coefficient $z_{m}$ is assigned to user $m$ based on the gain introduced by its strongest channel, expressed as $z_{m}=1-\frac{|\boldsymbol{\gamma}_{m,n_{m}^{\star}}|}{\sum_{\bar{m}=1}^{M}|\boldsymbol{\gamma}_{\bar{m},{n_{\bar{m}}^{\star}}}|}.$ (36) More precisely, the RF chains are dedicated to the users based on this ratio such that the users with lower channel gains are served by more signals and vice versa. The allocation procedure is illustrated in lines 2-11 in Algorithm 3. First, an RF chain is allocated to each user, then, the rest of the RF chains are divided among users based on their $z_{m}$. Let us denote $\mathcal{R}_{m}$ as the set of RF chains dedicated to user $m$. Then, we initialize $q_{i,l},i\in\mathcal{R}_{m}$ to compensate for the phase shift introduced by both $h_{i,l}$ and $\gamma_{i,l,m,n_{m}^{\star}}$. Specifically, we need to define $q_{i,l},i\in\mathcal{R}_{m}$ such that $\phi^{\star}_{i,l}=\operatorname*{argmin}_{\phi_{i,l}}\bigl{\langle}(\frac{j+e^{j\phi_{i,l}}}{2})h_{i,l}\gamma_{i,l,m,n_{m}^{\star}}\bigr{\rangle},\forall i,l,$ (37) where $i\in\mathcal{R}_{m}$ and $q_{i,l}$ can be obtained accordingly. Notice that (37) can be easily solved using a one-dimensional search with negligible complexity. The next step is to initialize the amplitude and phase of the digital precoders. For this, let us proceed by defining the received RF power at the $m$th user as $P_{rf,m}=\frac{G^{2}}{2}|\mathbf{a}_{m,n}\bar{\mathbf{w}}_{n}|^{2}\\!=\\!\frac{G^{2}}{2}\sum_{n=1}^{N_{f}}\biggl{|}\sum_{i=1}^{N_{v}}\sum_{l=1}^{N_{h}}\gamma_{i,l,m,n}q_{i,l}h_{i,l}\omega_{i,n}\biggr{|}^{2}.$ (38) Moreover, the output DC power of the rectifier is an increasing function of the input RF power when operating in the low-power regime below the breakdown region of the rectifier circuit’s diode [36]. Motivated by this, we aim to increase the available RF power at each user during the initialization process using the phases of the digital precoders with low complexity. One way to increase the available RF power for each user is to facilitate the coherent reception of the signals at the receiver. This can be done by reducing the amount of phase shift introduced in the signal. For this, we assume the initial digital weights to be $\omega_{i,n}=\bar{\omega}_{i,n}e^{j\hat{\omega}_{i,n}}$, while the dedicated signals to each user have the same amplitude $\bar{\omega}_{i,n}=w_{m},\forall i\in\mathcal{R}_{m},n$. Moreover, the ideal phase initialization for maximizing the received RF power is obtained by solving $\operatorname*{argmax}_{\hat{\omega}_{i,n}\in[0,2\pi],\forall i\in\mathcal{R}_{m},n}\sum_{n=1}^{N_{f}}\biggl{|}\sum_{i=1}^{N_{v}}\sum_{l=1}^{N_{h}}\gamma_{i,l,m,n}q_{i,l}h_{i,l}\bar{\omega}_{i,n}e^{j\hat{\omega}_{i,n}}\biggr{|}^{2}.$ (39) Meanwhile, solving this problem is not straightforward and introduces much additional complexity to our framework. Notably, the problem can be decoupled and solved individually for each sub-carrier without any change in the optimal solution. Still, there is a coupling between the digital weights of different RF chains in a similar sub-carrier. For this, we further reduce the complexity by formulating the problem as $\operatorname*{argmin}_{\hat{\omega}_{i,n}\in[0,2\pi]}\biggl{|}\big{\langle}\sum_{l=1}^{N_{h}}\gamma_{i,l,m,n}q_{i,l}h_{i,l}e^{j\hat{\omega}_{i,n}}\big{\rangle}\biggr{|},\forall i\in\mathcal{R}_{m},n.$ (40) Although the reformulated problem may not have the same solution as (39), it attempts to reduce the total amount of the phase shift of the received signal. Therefore, solving (40) can lead to a suboptimal solution to (39) with much lower complexity, and by using a one-dimensional search. Note that utilizing such an approach is relevant since the goal is to have a reasonable initialization for the variables, which leads to feeding a feasible initial point to the optimization algorithm. The initialization procedure is illustrated through lines 12-20 in Algorithm 3. For each user, the metamaterials connected to its dedicated RF chains are initialized. Then, the phases of the corresponding digital precoders are obtained. Finally, the amplitudes are iteratively increased until the EH requirement is met and a feasible solution is found. In the fully-digital architecture, the initialization algorithm follows the same procedure with one difference, i.e., each antenna element has a dedicated RF chain, and thus, a dedicated signal. ## V Numerical Analysis In this section, we provide numerical analysis of the system performance. We consider an indoor office with a transmitter located at the center of the ceiling. The operating frequency of the system is $f_{1}=5.18$ GHz, which matches the characteristics of the utilized rectifier model [36]. The spacing between the elements in the DMA is $\lambda_{1}/5$, while $\lambda_{1}/2$ is the distance between two consecutive microstrips. Meanwhile, the inter-element distance is $\lambda_{1}/2$ in the fully-digital architecture. Thus, $N_{v}=N_{h}=\lfloor\frac{L}{\lambda_{1}/2}\rfloor$ for the fully-digital system, and $N_{v}=\lfloor\frac{L}{\lambda_{1}/2}\rfloor,N_{h}=\lfloor\frac{L}{\lambda_{1}/5}\rfloor$ for the DMA-assisted system [27]. Note that $L$ is the array length while the arrays are considered to be square-shaped. We set the optimization parameters $\tau_{s}=10^{-3}$, $\varsigma=5$, and $\upsilon=10^{-6}$. Without loss of generality, we set $G=1$ and $\bar{\eta}=\frac{\pi}{4}$555In practice, the HPA output power is larger than the input power because $G>1$. However, the proposed framework applies to all values of $G$.. Finally, the rectifier parameters are $v_{t}=25$ mV and $\eta_{0}=1.05$ [36, 21, 20]. We utilize the characteristics of the ®Rogers RO4000 series ceramic laminate to calculate the propagation coefficients of the microstrips. Specifically, we calculate the attenuation and propagation coefficients of a RO400C LoPro with a thickness of 20.7 mil (0.5258 mm) using the formulation provided in [23], which gives $\alpha=0.356$ m-1 and $\beta=202.19$ m-1. Based on the rectifier circuit design and simulations in [36], the rectifier circuit diode enters the breakdown region when the received RF power is approximately $100\ \mu$W for a continuous wave ($N_{f}=1$). Moreover, it has been shown that the maximum RF- to-DC conversion efficiency is approximately $20\%$ for the mentioned setup. Hence, we establish a minimum requirement of $\tilde{P}_{dc}=20\ \mu$W for DC harvested power. In the figures, FD refers to the fully-digital architecture. Moreover, $d$ represents the distance between the user and the center of the transmitter. Fig. 5 provides the convergence performance of the proposed ASCA-DMA approach by presenting the amount of power consumption at the end of each iteration of alternating optimization. It is seen that the objective value gradually decreases with iterations until convergence, while the number of required iterations for convergence depends on the setup. For instance, it is shown that increasing $L$, $N_{f}$, or $M$ can increase the complexity of the problem leading to more iterations. Fig. 6 provides a detailed algorithm performance by showing all iterations of optimization, including the SCA algorithm for both optimizing metamaterials and digital precoders. It is seen that since the minimum harvested power increases when optimizing $\mathbf{q}$, the power consumption increases, while this facilitates decreasing the amount of power consumption when optimizing digital precoders. Therefore, the power consumption decreases gradually at the end of each iteration of the alternating optimization (red arrows) until convergence. Figure 5: The convergence performance of ASCA-DMA for (a) $M=1$, $L=15$ cm, $N_{f}=1$ (left), (b) $M=1$, $L=25$ cm, $N_{f}=1$ (middle), and (c) $M=2$, $L=25$ cm, $N_{f}=4$ (right) over iterations with $d=2.5$ m and $P_{max}=100$ W. Figure 6: The convergence performance of ASCA-DMA for $M=1$, $L=25$ cm, $d=2.5$ m, $N_{f}=1$, and $P_{max}=1$ W over all iterations (alternating and SCA). The red and green arrows indicate the starting points of optimization for the metamaterial elements and digital precoders, respectively. Fig. 7 showcases the power consumption of the system as a function of the antenna length. Note that increasing the $L$ reduces the power consumption since the number of elements and the array aperture increases. This leads to more capability in beam focusing, thus, delivering more power to the devices given the same transmit power. However, the performance comparison between DMA and FD highly depends on the system setup. It is seen that DMA outperforms FD when $L$ and $P_{max}$ are relatively low. Note that the output of each RF chain in DMA has to feed all the corresponding elements, while the the amount of power consumption scales with the saturation power of the HPA. Thus, when $P_{max}$ is low, the output power of each RF chain in DMA is multiplied by a small value, leading to potential performance gains compared to FD. However, this is only the case when $L$ is low and the number of RF chains is small. On the other hand, in FD, $N_{rf}$ increases with $L$ at a higher rate compared to DMA. Thus, it is easier to reduce the amount of output power of the HPAs by distributing the required output power among them, leading to lower power consumption compared to DMA. However, the value of $L$ that shifts the favorable architecture from DMA to FD or vice versa depends on $P_{max}$ and the transmit power required for meeting the EH requirements. For instance, when $d=2.5$ m and $P_{max}=1$ W, the required transmit power and the power consumption multiplier are both low, leading to a shift in performance with lower $L$. Meanwhile, when $d=6.5$ m and $P_{max}=100$ W, DMA outperforms FD up to a higher value of $L$ compared to the previous case. Meanwhile, when $d$ is small and $P_{max}$ is sufficiently large, FD outperforms DMA over all $L$ values with the performance gap increasing with $L$. In contrast, when $d$ is relatively large and $P_{max}$ is small, DMA becomes the favorable choice for sufficiently large $L$. Figure 7: The power consumption as a function of $L$ for $N_{f}=4$, $P_{max}\in\\{1,100\\}$ W, and $d\in\\{2.5,6.5\\}$ m. Our results in Fig. 8 corroborate that increasing $N_{f}$ reduces $P_{c}$. As discussed in Section II-C, this is because the HPAs operate in the linear regime, and higher $N_{f}$ can leverage the rectifier non-linearity and deliver more DC power to the ERs via waveform optimization. However, as previously mentioned, the preference for DMA or FD highly depends on the system parameters. For instance, when $L=10$ cm and $P_{max}=1$ W, both $N_{rf}$ and saturation power have small values, leading to a lower power consumption for DMA compared to FD. The reason is that the multiplier of the HPA output power is low, thus, a lower value is multiplied by the output power of each RF chain in DMA. Combining this with a small $L$, thus a small $N_{rf}$ for FD, leads to DMA outperforming FD. On the other hand, different parameters and system setups may lead to different performances. For example, it is seen that when $L$ and saturation power are both sufficiently large, the favorable architecture shifts from DMA to FD. The reason is that the number of RF chains is much higher for FD in this case, leading to lower output power for each. Then, combining this with a large value of saturation voltage leads to smaller HPA outputs in FD, thus, lower power consumption. Meanwhile, increasing $N_{f}$ affects the performance gap between FD and DMA, and increasing $N_{f}$ leads to more performance gains in FD compared to DMA since the number of signals is relatively larger in FD. Thus, when DMA outperforms FD for a given $L$ and $P_{max}$, FD may start outperforming DMA by sufficiently increasing $N_{f}$. On the other hand, when FD performs better, increasing $N_{f}$ can increase the performance gap. Figure 8: The power consumption as a function of $N_{f}$ for $L=25$ cm, $P_{max}\in\\{1,100\\}$ W, $L\in\\{10,20\\}$ cm, and $d=2.5$ m. Fig. 9 illustrates the impact of user distance on performance. It is obvious that the power consumption increases with distance since the path loss becomes larger and more transmit power is required to overcome that and meet the EH requirements. Interestingly, the performance shift between FD and DMA is also shown here. It is shown that when both $P_{max}$ and $L$ are relatively low, DMA outperforms FD, especially over large distances. On the other hand, when $L$ increases, DMA starts performing better after a certain $d$ since the number of HPAs is much larger in FD, and increasing their output power affects the power consumption considerably. For instance, when $L=20$ cm and $P_{max}=100$ W, FD starts with better performance than DMA, but the performance gap decreases as $d$ becomes larger. Figure 9: The power consumption as a function of user distance for $N_{f}=4$, $P_{max}\in\\{1,100\\}$ W, and $L\in\\{10,20\\}$ cm. The impact of the number of users on system performance has been illustrated in Fig. 10 for different system parameters. As expected, the power consumption increases with $M$ since more EH requirements must be met. As previously shown, both DMA or FD may be the favorable choice depending on the system setup and the number of devices. For example, when $P_{max}$ and $L$ are both sufficiently large, DMA outperforms FD for a small $M$, but FD becomes favorable when $M$ is relatively large. The reason is that increasing $M$ leads to more required transmit power to satisfy the requirements and as earlier mentioned, a large $L$ leads to relatively smaller output powers compared in FD to DMA. Some discussions on the complexity of implementing a large-scale FD setup are in order. Notice that 36 RF chains are needed in a 400 cm2 FD array and just 6 for a DMA with the same size at $f_{1}=5.18$ GHz. Thus, although FD may outperform DMA in many setups, as seen earlier, DMA may still be favorable since it can achieve relatively good performance with a reduced $N_{rf}$ and complexity. Figure 10: The power consumption over $M$ for (a) $P_{max}=1$ W (top) and (b) $P_{max}=100$ W (bottom), while $N_{f}=4$, and $L\in\\{15,25\\}$ cm. The users are located at $d=4.4$ m. Fig. 11 provides some insights regarding the beam focusing capability in the near-field WPT by illustrating the normalized received RF power in each spatial point of the area. Note that the received signal at each point is normalized by its path loss to remove the impact of the distance. In Fig. 11a, it is seen that when the device is located in the near-field region, the beam is focused around the device location, while the beam trace fades increasingly past the device. Such phenomena can have a huge benefit in reducing the RF emission footprint in the environment, which facilitates the implementation of environmentally friendly WPT systems. Meanwhile, the beam pattern for a located device in the far-field is formed in the user’s direction, as illustrated in Fig. 11b. This may be highly disadvantaged in interference- sensitive applications since the generated beam may cause difficult-to-handle interference in the signals conveying information, e.g., in SWIPT. Figure 11: The normalized received RF power (W) in the area when the energy receiver is located at (a) the near-field region (top) and (b) the far-field region (bottom) in the DMA-assisted system with $L=30$ cm and $N_{f}=1$. The received RF waveform in DMA-assisted and fully-digital systems is presented in Fig.12a and Fig.12b, respectively. As previously mentioned, high PAPR waveforms are beneficial for enhancing the performance in terms of DC harvested power [7]. Moreover, our simulations verify this by showing that the received signal experiences high peak amplitudes at specific intervals. Recall that when HPAs are operating in the linear regime, as in our case, the HPA does not introduce distortion to the signal. Thus, it is beneficial to utilize multiple tones to leverage the rectifier’s non-linearity. Note that the peak- to-peak time depends on the characteristics of the EH circuit, mainly the capacitor. Figure 12: The received signal at the device using (a) DMA (top) and (b) FD (bottom) for $M=1$, $L=25$ cm, and $N_{f}=8$. ## VI Conclusion and Future Work In this paper, we investigated a multi-antenna near-field WPT system with a DMA as the transmitter to charge multiple non-linear EH devices. Furthermore, we proposed an optimization framework relying on alternating optimization and SCA for the joint waveform optimization and beam focusing to minimize the system power consumption while meeting the EH requirements. Numerical results showed that both DMA and fully-digital architecture may be the favorable choice in terms of power consumption depending on the system setups and parameters such as antenna length, saturation power of the HPAs, number of users, and user distance. Moreover, we showed that increasing the antenna length or the number of tones can enhance the performance. Finally, we verified that the transmitter can focus the energy beams on the spatial points in the near-field region, while energy beams are formed toward the devices’ direction in the far-field. As a prospect for future research, we may delve deeper into the signal generation aspect by analyzing the power consumption based on the number of tones. Another research direction is to utilize optimization approaches with lower complexity, e.g., relying on machine learning, to learn online the input-output relation of the system’s non-linear components while optimizing the transmit waveform accordingly. ### -A Proof of Theorem 2 We relax (29b) by limiting the values of $q_{i,l}$ to lie within the Lorentzian circle in the complex plane. By utilizing the fact that (29a) is a positive and increasing function of the rectifier’s output voltage and using the epigraph form, the relaxed problem can be written as (30). Note that each configuration of a metamaterial corresponds to a point on the Lorentzian- constrained circle. Imagine $\vec{e}$ is the vector that represents the direction and gain of a point on the Lorentzian circle, while this direction and gain impacts the transmit signal. Meanwhile, the goal of (30) is to increase the minimum output voltage of the ERs. Therefore, when shaping the transmit signal toward different ERs, it is obvious that $\vec{e}$ should be chosen with the most possible gain in the required direction to improve the signal strength at the receiver. Furthermore, the most possible introduced gain by a metamaterial element along a specified direction happens when the point is exactly on the Lorentzian-constrained circle in that direction. Hence, although (30b) is a relaxed version of the constraint (29b), considering the final solution of the metamaterials as $a\vec{e},0\leq a\leq 1$, the only solution that leads to the most gain for the desired direction is $a=1$. Thus, the solution of (30) is the same as (29) leading to a configuration positioned on the Lorentzian circle and the equivalence between the problems. ## References * [1] Z. Zhang _et al._ , “6G wireless networks: vision, requirements, architecture, and key technologies,” IEEE Veh. Technol. Mag., vol. 14, no. 3, pp. 28–41, 2019. * [2] N. H. Mahmood _et al._ , “Six key features of machine type communication in 6G,” in 2nd 6G SUMMIT, pp. 1–5, 2020. * [3] O. L. A. López _et al._ , “High-Power and Safe RF Wireless Charging: Cautious Deployment and Operation,” 2023. * [4] O. L. A. López _et al._ , “Massive wireless energy transfer: enabling sustainable IoT toward 6G era,” IEEE Internet Things J., vol. 8, no. 11, pp. 8816–8835, 2021. * [5] Y. Zeng _et al._ , “Communications and signals design for wireless power transmission,” IEEE Trans Commun, vol. 65, no. 5, pp. 2264–2290, 2017. * [6] B. Clerckx and E. Bayguzina, “Waveform Design for Wireless Power Transfer,” IEEE Trans. Signal Process., vol. 64, no. 23, pp. 6313–6328, 2016. * [7] C. R. Valenta _et al._ , “Theoretical Energy-Conversion Efficiency for Energy-Harvesting Circuits Under Power-Optimized Waveform Excitation,” IEEE Trans. Microw. Theory Tech., vol. 63, no. 5, pp. 1758–1767, 2015. * [8] I. Ahmed _et al._ , “A Survey on Hybrid Beamforming Techniques in 5G: Architecture and System Model Perspectives,” IEEE Commun. Surv. Tutor., vol. 20, no. 4, pp. 3060–3097, 2018. * [9] X. Gao _et al._ , “Energy-Efficient Hybrid Analog and Digital Precoding for MmWave MIMO Systems With Large Antenna Arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 998–1009, 2016. * [10] Q. Wu _et al._ , “Intelligent Reflecting Surface-Aided Wireless Communications: A Tutorial,” IEEE Trans Commun, vol. 69, no. 5, pp. 3313–3351, 2021. * [11] N. Shlezinger _et al._ , “Dynamic Metasurface Antennas for 6G Extreme Massive MIMO Communications,” IEEE Wirel. Commun., vol. 28, no. 2, pp. 106–113, 2021. * [12] B. Zheng _et al._ , “A Survey on Channel Estimation and Practical Passive Beamforming Design for Intelligent Reflecting Surface Aided Wireless Communications,” IEEE Commun. Surv. Tutor., vol. 24, no. 2, pp. 1035–1071, 2022. * [13] H. Jiang _et al._ , “Hybrid RIS and DMA Assisted Multiuser MIMO Uplink Transmission With Electromagnetic Exposure Constraints,” IEEE J Sel Top Signal Process, vol. 16, no. 5, pp. 1055–1069, 2022. * [14] O. L. A. López _et al._ , “Massive MIMO with radio stripes for indoor wireless energy transfer,” IEEE Trans. Wirel. Commun., vol. 21, no. 9, pp. 7088–7104, 2022. * [15] A. Azarbahram _et al._ , “On the Radio Stripe Deployment for Indoor RF Wireless Power Transfer,” 2023. * [16] O. L. A. López _et al._ , “A Low-Complexity Beamforming Design for Multiuser Wireless Energy Transfer,” IEEE Wireless Commun. Lett., vol. 10, no. 1, pp. 58–62, 2021. * [17] S. Shen and B. Clerckx, “Beamforming Optimization for MIMO Wireless Power Transfer With Nonlinear Energy Harvesting: RF Combining Versus DC Combining,” IEEE Trans. Wirel. Commun., vol. 20, no. 1, pp. 199–213, 2021. * [18] B. Clerckx and E. Bayguzina, “Low-Complexity Adaptive Multisine Waveform Design for Wireless Power Transfer,” IEEE Antennas Wirel. Propag. Lett., vol. 16, pp. 2207–2210, 2017. * [19] Y. Huang and B. Clerckx, “Large-Scale Multiantenna Multisine Wireless Power Transfer,” IEEE Trans. Signal Process., vol. 65, no. 21, pp. 5812–5827, 2017. * [20] S. Shen and B. Clerckx, “Joint Waveform and Beamforming Optimization for MIMO Wireless Power Transfer,” IEEE Trans Commun, vol. 69, no. 8, pp. 5441–5455, 2021. * [21] Y. Zhang and B. Clerckx, “Waveform Design for Wireless Power Transfer With Power Amplifier and Energy Harvester Non-Linearities,” IEEE Trans. Signal Process., pp. 1–15, 2023. * [22] H. Zhang _et al._ , “Near-field wireless power transfer with dynamic metasurface antennas,” in IEEE SPAWC, pp. 1–5, 2022. * [23] A. Azarbahram _et al._ , “Energy Beamforming for RF Wireless Power Transfer With Dynamic Metasurface Antennas,” IEEE Wireless Commun. Lett., pp. 1–1, 2023. * [24] O. T. Demir and T. E. Tuncer, “Antenna selection and hybrid beamforming for simultaneous wireless information and power transfer in multi-group multicasting systems,” IEEE Trans. on Wireless Commun., vol. 15, no. 10, pp. 6948–6962, 2016. * [25] Z. Feng _et al._ , “Waveform and Beamforming Design for Intelligent Reflecting Surface Aided Wireless Power Transfer: Single-User and Multi-User Solutions,” IEEE Trans. Wirel. Commun., vol. 21, no. 7, pp. 5346–5361, 2022. * [26] Y. Zhao _et al._ , “RIS-Aided SWIPT: Joint Waveform, Active and Passive Beamforming Design Under Nonlinear Harvester Model,” IEEE Trans Commun, vol. 70, no. 2, pp. 1345–1359, 2022. * [27] H. Zhang _et al._ , “Beam focusing for near-field multiuser MIMO communications,” IEEE Trans. Wirel. Commun., vol. 21, no. 9, pp. 7476–7490, 2022. * [28] S. W. Ellingson, “Path Loss in Reconfigurable Intelligent Surface-Enabled Channels,” in IEEE PIMRC, pp. 829–835, 2021. * [29] C. Rapp, “Effects of HPA-nonlinearity on a 4-DPSK/OFDM-signal for a digital sound broadcasting signal,” ESA Special Publication, vol. 332, pp. 179–184, 1991. * [30] D. R. Smith _et al._ , “Analysis of a Waveguide-Fed Metasurface Antenna,” Phys. Rev. Appl., vol. 8, p. 054048, Nov 2017. * [31] L. You _et al._ , “Energy Efficiency Maximization of Massive MIMO Communications With Dynamic Metasurface Antennas,” IEEE Trans. Wirel. Commun., vol. 22, no. 1, pp. 393–407, 2023. * [32] C. Lin and G. Y. Li, “Energy-Efficient Design of Indoor mmWave and Sub-THz Systems With Antenna Arrays,” IEEE Trans. Wirel. Commun., vol. 15, no. 7, pp. 4660–4672, 2016. * [33] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge university press, 2004. * [34] E. J. McShane, “Jensen’s inequality,” 1937. * [35] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex programming, version 2.1.” http://cvxr.com/cvx, Mar. 2014. * [36] B. Clerckx and J. Kim, “On the beneficial roles of fading and transmit diversity in wireless power transfer with nonlinear energy harvesting,” IEEE Trans. Wirel. Commun., vol. 17, no. 11, pp. 7731–7743, 2018.
# Testing High-dimensional Multinomials with Applications to Text Analysis T. Tony Cai111University of Pennsylvania , Zheng Tracy Ke222Harvard University , and Paxton Turner††footnotemark: ###### Abstract Motivated by applications in text mining and discrete distribution inference, we investigate the testing for equality of probability mass functions of $K$ groups of high-dimensional multinomial distributions. A test statistic, which is shown to have an asymptotic standard normal distribution under the null, is proposed. The optimal detection boundary is established, and the proposed test is shown to achieve this optimal detection boundary across the entire parameter space of interest. The proposed method is demonstrated in simulation studies and applied to analyze two real-world datasets to examine variation among consumer reviews of Amazon movies and diversity of statistical paper abstracts. > Keywords: authorship attribution, closeness testing, consumer reviews, > martingale central limit theorem, minimax optimality, topic model ## 1 Introduction Statistical inference for multinomial data has garnered considerable recent interest (Diakonikolas and Kane, 2016; Balakrishnan and Wasserman, 2018). One important application is in text mining, as it is common to model the word counts in a text document by a multinomial distribution (Blei et al., 2003). We consider a specific example in marketing, where the study of online customer ratings and reviews has become a trending topic (Chevalier and Mayzlin, 2006; Zhu and Zhang, 2010; Leung and Yang, 2020). Customer reviews are a good proxy to the overall word of mouth (WOM) and can significantly influence customers’ decisions (Zhu and Zhang, 2010). Many research works aim to understand the patterns in online reviews and their impacts on sales. Classical studies only use the numerical ratings but ignore the rich text reviews because of their unstructured nature. More recent works have revealed the importance of analyzing text reviews (Chevalier and Mayzlin, 2006), especially for hedonic products such as books, movies, and hotels. A question of great interest is to detect the heterogeneity in reviewers’ response styles. For example, Leung and Yang (2020) discovered that younger travelers, women, and travelers with less review expertise tend to give more positive reviews and that guests staying in high-class hotels tend to have more extreme response styles than those staying in low-class hotels. Knowing such differences will offer valuable insights for hotel managers and online rating/review sites. The aforementioned heterogeneity detection can be cast as a hypothesis test on multinomial data. Suppose reviews are written on a vocabulary of $p$ distinct words. Let $X_{i}\in\mathbb{R}^{p}$ denote the word counts in review $i$. We model that $X_{i}\sim\mathrm{Multinomial}(N_{i},\Omega_{i}),\qquad 1\leq i\leq n,$ (1.1) where $N_{i}$ is the total length of review $i$ and $\Omega_{i}\in\mathbb{R}^{p}$ is a probability mass function (PMF) containing the population word frequencies. These reviews are divided into $K$ groups by reviewer characteristics (e.g., age, gender, new/returning customer), product characteristics (e.g., high-class versus low-class hotels), and numeric ratings (e.g., from 1 star to 5 stars), where $K$ can be presumably large. We view $\Omega_{i}$ as representing the ‘true response’ of review $i$. The “average response” of a group $k$ is defined by a weighted average of the PMFs: $\mu_{k}=(n_{k}\bar{N}_{k})^{-1}\sum_{i\in S_{k}}N_{i}\Omega_{i},\qquad 1\leq k\leq K.$ (1.2) Here $S_{k}\subset\\{1,2,\ldots,n\\}$ is the index set of group $k$, $n_{k}=|S_{k}|$ is the total number of reviews in group $k$, and $\bar{N}_{k}=n_{k}^{-1}\sum_{i\in S_{k}}N_{i}$ is the average length of reviews in group $k$. We would like to test $H_{0}:\quad\mu_{1}=\mu_{2}=\ldots=\mu_{K}.$ (1.3) When the null hypothesis is rejected, it means there exist statistically significant differences among the group-wise “average responses”. We call (1.1)-(1.3) the “$K$-sample testing for equality of average PMFs in multinomials” or “$K$-sample testing for multinomials” for short. Interestingly, as $K$ varies, this problem includes several well-defined problems in text mining and discrete distribution inference as special cases. 1. 1. Global testing for topic models. Topic modeling (Blei et al., 2003) is a popular text mining tool. In a topic model, each $\Omega_{i}$ in (1.1) is a convex combination of $M$ topic vectors. Before fitting a topic model to a corpus, it is often desirable to determine if the corpus indeed contains multiple topics. This boils down to the global testing problem, which tests $M=1$ versus $M>1$. Under the null hypothesis, $\Omega_{i}$’s are equal to each other, and in the alternative hypothesis, $\Omega_{i}$’s can take continuous values in a high-dimensional simplex. This is a special case of our problem with $K=n$ and $n_{k}=1$. 2. 2. Authorship attribution (Mosteller and Wallace, 1963; Kipnis, 2022). In these applications, the goal is to determine the unknown authorship of an article from other articles with known authors. A famous example (Mosteller and Wallace, 2012) is to determine the actual authors of a few Federalist Papers written by three authors but published under a single pseudonym. It can be formulated (Mosteller and Wallace, 1963; Kipnis, 2022) as testing the equality of population word frequencies between the article of interest and the corpus from a known author, a special case of our problem with $K=2$. 3. 3. Closeness between discrete distributions (Chan et al., 2014; Bhattacharya and Valiant, 2015; Balakrishnan and Wasserman, 2019). There has been a surge of interest in discrete distribution inference. Closeness testing is one of most studied problems. The data from two discrete distributions are summarized in two multinomial vectors $\mathrm{Multinomial}(N_{1},\mu)$ and $\mathrm{Multinomial}(N_{2},\theta)$. The goal is to test $\mu=\theta$. It is a special case of our testing problem with $K=2$ and $n_{1}=n_{2}=1$. In this paper, we provide a unified solution to all the aforementioned problems. The key to our methodology is a flexible statistic called DELVE (DE- biased and Length-assisted Variability Estimator). It provides a general similarity measure for comparing groups of discrete distributions such as count vectors associated with text corpora. Similarity measures (such as the classical cosine similarity, log-likelihood ratio statistic, and others) are fundamental in text mining and have been applied to problems in distribution testing (Kim et al., 2022), computational linguistics (Gomaa et al., 2013), econometrics (Hansen et al., 2018), and computational biology (Kolodziejczyk et al., 2015). Our method is a new and flexible similarity measure that is potentially useful in these areas. We emphasize that our setting does not require that the $X_{i}$’s in the same group are drawn from the same distribution. Under the null hypothesis (1.3), the group-wise means are equal, but the $\Omega_{i}$’s within each group can still be different from each other. As a result, the null hypothesis is composite and designing a proper test statistic is non-trivial. ### 1.1 Our results and contributions The dimensionality of the testing problem is captured by $(n,p,K)$ and $\bar{N}:=n^{-1}\sum_{i=1}^{n}N_{i}$. We are interested in a high-dimensional setting where $n\bar{N}\to\infty,\quad p\to\infty,\quad\mbox{and}\quad n^{2}\bar{N}^{2}/(Kp)\to\infty.$ (1.4) In most places of this paper, we use a subscript $n$ to indicate asymptotics, but our method and theory do apply to the case where $n$ is finite and $\bar{N}\to\infty$. In text applications, $n\bar{N}$ is the total count of words in the corpus, and a large $n\bar{N}$ means either there are sufficiently many documents, or the documents are sufficiently long. Given that $n\bar{N}\to\infty$, we further allow $(p,K)$ to grow with $n$ at a speed such that $Kp\ll n^{2}\bar{N}^{2}$. In particular, our settings allow $K$ to range from $2$ to $n$, so as to cover all the application examples. We propose a test that enjoys the following properties: 1. (a) Parameter-free null distribution: We show that the test statistic $\psi\to N(0,1)$ under $H_{0}$. Even under the null hypothesis (1.3), the model contains a large number of free parameters because the null hypothesis is only about the equality of “average” PMFs but still allows $(N_{i},\Omega_{i})$ to differ within each group. As an appealing property, the null distribution of $\psi$ does not depend on these individual multinomial parameters; hence, we can always conveniently obtain the asymptotic $p$-value for our proposed test. 2. (b) Minimax optimal detection boundary: We define a quantity $\omega_{n}:=\omega_{n}(\mu_{1},\mu_{2},\ldots,\mu_{K})$ in (3.5) that measures the difference among the $K$ group-wise mean PMF’s. It satisfies that $\omega_{n}=0$ if and only if the null hypothesis holds, and it has been properly normalized so that $\omega_{n}$ is bounded under the alternative hypothesis (provided some mild regularity conditions hold). We show that the proposed test has an asymptotic full power if $\omega_{n}^{4}n^{2}{\bar{N}}^{2}/(Kp)\to\infty.$ We also provide a matching lower bound by showing that the null hypothesis and the alternative hypothesis are asymptotically indistinguishable if $\omega_{n}^{4}n^{2}\bar{N}^{2}/(Kp)\to 0.$ Therefore, the proposed test is minimax optimal. Furthermore, in the boundary case where $\omega_{n}^{4}n^{2}\bar{N}^{2}/(Kp)\to c_{0}$ for a constant $c_{0}>0$, for some special settings, we show that $\psi\to N(0,1)$ under $H_{0}$, and $\psi\to N(c_{1},1)$, under $H_{1}$, with the constant $c_{1}$ being an explicit function of $c_{0}$. To the best of our knowledge, this testing problem for a general $K$ has not been studied before. The existing works primarily focused on closeness testing and authorship attribution (see Section 1.2), which are special cases with $K=2$. In comparison, our test is applicable to any value of $K$, offering a unified solution to multiple applications. Even for $K=2$, the existing works do not provide a test statistic that has a tractable null distribution. They determined the rejection region and calculated $p$-values using either a (conservative) large-deviation bound or a permutation procedure. Our test is the first one equipped with a tractable null distribution. Our results about the optimal detection boundary for a general $K$ are also new to the literature. By varying $K$ in our theory, we obtain the optimal detection boundary for different sub-problems. For some of them (e.g., global testing for topic models, authorship attribution with moderate sparsity), the optimal detection boundary was not known before; hence, our results help advance the understanding of the statistical limits of these problems. ### 1.2 Related literature First, we make a connection to discrete distribution inference. Let $X\sim\mathrm{Multinomial}(N,\Omega)$ represent a size-$N$ sample from a discrete distribution with $p$ categories. The one-sample closeness testing aims to test $H_{0}:\Omega=\mu$, for a given PMF $\mu$. Existing works focus on finding the minimum separation condition in terms of the $\ell^{1}$-norm or $\ell^{2}$-norm of $\Omega-\mu$. Balakrishnan and Wasserman (2019) derived the minimum $\ell^{1}$-separation condition and proposed a truncated chi-square test to achieve it. Valiant and Valiant (2017) studied the “local critical radius”, a local separation condition that depends on the “effective sparsity” of $\mu$, and they proposed a “2/3rd + tail” test to achieve it. In the two- sample closeness testing problem, given $X_{1}\sim\mathrm{Multinomial}(N_{1},\Omega_{1})$ and $X_{2}\sim\mathrm{Multinomial}(N_{2},\Omega_{2})$, it aims to test $H_{0}:\Omega_{1}=\Omega_{2}$. Again, this literature focuses on finding the minimum separation condition in terms of the $\ell^{1}$-norm or $\ell^{2}$-norm of $\Omega_{1}-\Omega_{2}$. When $N_{1}=N_{2}$, Chan et al. (2014) derived the minimum $\ell^{1}$-separation condition and proposed a weighted chi-square test to attain it. Bhattacharya and Valiant (2015) extended their results to the unbalanced case where $N_{1}\neq N_{2}$, assuming $\|\Omega_{1}-\Omega_{2}\|_{1}\geq p^{-1/12}$. This assumption was later removed by Diakonikolas and Kane (2016), who established the minimum $\ell^{1}$-separation condition in full generality. Kim et al. (2022) proposed a two-sample kernel $U$-statistic and showed that it attains the minimum $\ell^{2}$-separation condition. Since the two-sample closeness testing is a special case of our problem with $K=2$ and $n_{1}=n_{2}=1$, our test is directly applicable. An appealing property of our test is its tractable asymptotic null distribution of $N(0,1)$. In contrast, for the chi-square statistic in Chan et al. (2014) or the $U$-statistic in (Kim et al., 2022), the rejection region is determined by either an upper bound from concentration inequalities or a permutation procedure, which may lead to a conservative threshold or need additional computational costs. Regarding the testing power, we show in Section 4.3 that our test achieves the minimum $\ell^{2}$-separation condition, i.e., our method is an optimal “$\ell^{2}$ testor.” Our test can also be turned into an optimal “$\ell^{1}$ testor” (a test that achieves the minimum $\ell^{1}$-separation condition) by re-weighting terms in the test statistic (see Section 4.3). Next, we make a connection to text mining. In this literature, a multinomial vector $X\sim\mathrm{Multinomial}(N,\Omega)$ represents the word counts for a document of length $N$ written with a dictionary containing $p$ words. In a topic model, each $\Omega_{i}$ is a convex combination of $M$ “topic vectors”: $\Omega_{i}=\sum_{k=1}^{M}w_{i}(k)A_{k}$, where each $A_{k}\in\mathbb{R}^{p}$ is a PMF and the combination coefficient vector $w_{i}\in\mathbb{R}^{K}$ is called the “topic weight” vector for document $i$. Given a collection of documents $X_{1},X_{2},\ldots,X_{n}$, the global testing problem aims to test $M=1$ versus $M>1$. Interestingly, the optimal detection boundary for this problem has never been rigorously studied. As we have explained, this problem a special case of our testing problem with $K=n$. Our results (a) provide a test statistic that has a tractable null distribution and (b) reveal that the optimal detection boundary is $\omega^{2}_{n}\asymp(\sqrt{n}\bar{N})^{-1}\sqrt{p}$. Both (a) and (b) are new results. When comparing our results with those about estimation of $A_{k}$’s (Ke and Wang, 2022), it suggests that global testing requires a strictly lower signal strength than topic estimation. For authorship attribution, Kipnis (2022) treats the corpus from a known author as a single document and tests the null hypothesis that this combined document and a new document have the same population word frequencies. It is a two-sample closeness testing problem, except that sparsity is imposed on the difference of two PMFs. Kipnis (2022) proposed a test which applies an “exact binomial test” to obtain a $p$-value for each word and combines these $p$-values using Higher Criticism (Donoho and Jin, 2004). Donoho and Kipnis (2022) analyzed this test when the number of “useful words” is $o(\sqrt{p})$, and they derived a sharp phase diagram (a related one-sample setting was studied in Arias-Castro and Wang (2015)). In Section 4.2, we show that our test is applicable to this problem and has some nice properties: (a) tractable null distribution; (b) allows for $s\geq c\sqrt{p}$, where $s$ is the number of useful words; and (c) does not require documents from the known author to have identical population word frequencies, making the setting more realistic. On the other hand, when $s=o(\sqrt{p})$, our test is less powerful than the one in Kipnis (2022); Donoho and Kipnis (2022), as our test does not utilize sparsity explicitly. We can further improve our test in this regime by modifying the DELVE statistic to incorporate sparsity (see the remark in Section 4.2). ### 1.3 Organization The rest of this paper is arranged as follows. In Section 2, we introduce the test statistic and explain the rationale behind it. We then present in Section 3 the main theoretical results, including the asymptotic null distribution, power analysis, a matching lower bound, the study of two special cases ($K=n$ and $K=2$), and a discussion of the contiguity regime. Section 4 applies our results to text mining and discrete distribution testing. Simulations are in Section 5 and real data analysis is in Section 6. The paper is concluded with a discussion in Section 7. All proofs are in the appendix. ## 2 The DELVE Test Recall that $X_{i}\sim\mathrm{Multinomial}(N_{i},\Omega_{i})$ for $1\leq i\leq n$. There is a known partition $\\{1,2,\ldots,n\\}=\cup_{k=1}^{K}S_{k}$. Write $n_{k}=|S_{k}|$, $\bar{N}_{k}=n_{k}^{-1}\sum_{i\in S_{k}}N_{i}$, and $\bar{N}=n^{-1}\sum_{i=1}^{n}N_{i}$. In (1.2), we have defined the group-wise mean PMF $\mu_{k}=(n_{k}\bar{N}_{k})^{-1}\sum_{i\in S_{k}}N_{i}\Omega_{i}$. We further define the overall mean PMF $\mu\in\mathbb{R}^{p}$ by $\mu:=\frac{1}{n\bar{N}}\sum_{k=1}^{K}n_{k}\bar{N}_{k}\mu_{k}=\frac{1}{n\bar{N}}\sum_{i=1}^{n}N_{i}\Omega_{i}.$ (2.1) We introduce a quantity $\rho^{2}=\rho^{2}(\mu_{1},\ldots,\mu_{K})$ by $\rho^{2}:=\sum_{k=1}^{K}n_{k}\bar{N}_{k}\|\mu_{k}-\mu\|^{2}.$ (2.2) This quantity measures the variations across $K$ group-wise mean PMFs. It is true that the null hypothesis (1.3) holds if and only if $\rho^{2}=0$. Inspired by this observation, we hope to construct an unbiased estimator of $\rho^{2}$ and develop it to a test statistic. We can easily obtain the minimum variance unbiased estimators of $\mu_{k}$ and $\mu$: $\hat{\mu}_{k}=\frac{1}{n_{k}\bar{N}_{k}}\sum_{i\in S_{k}}X_{i},\qquad\mbox{and}\qquad\hat{\mu}=\frac{1}{n\bar{N}}\sum_{k=1}^{K}n_{k}\bar{N}_{k}\hat{\mu}_{k}=\frac{1}{n\bar{N}}\sum_{i=1}^{n}X_{i}.$ (2.3) For each $1\leq j\leq p$, let $\mu_{kj}$, $\mu_{j}$, $\hat{\mu}_{kj}$ and $\hat{\mu}_{j}$ represent the $j$th entry of $\mu_{k}$, $\mu$, $\hat{\mu}_{k}$ and $\hat{\mu}$, respectively. A naive estimator of $\rho^{2}$ is $\widetilde{T}=\sum_{j=1}^{p}\widetilde{T}_{j},\qquad\mbox{where}\quad\widetilde{T}_{j}=\sum_{k=1}^{K}n_{k}\bar{N}_{k}(\hat{\mu}_{kj}-\hat{\mu}_{j})^{2}.$ (2.4) This estimator is biased. In Section C.1 of the appendix , we show that $\mathbb{E}[\widetilde{T}_{j}]=\sum_{k=1}^{K}\bigl{[}n_{k}\bar{N}_{k}(\mu_{kj}-\mu_{j})^{2}+\bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\bigr{)}\sum_{i\in S_{k}}N_{i}\Omega_{ij}(1-\Omega_{ij})\bigr{]}.$ It motivates us to debias $\widetilde{T}_{j}$ by using an unbiased estimate of $\Omega_{ij}(1-\Omega_{ij})$. By elementary properties of the multinomial distribution, $\mathbb{E}[X_{ij}(N_{i}-X_{ij})]=N_{i}(N_{i}-1)\Omega_{ij}(1-\Omega_{ij})$. We thereby use $\frac{1}{N_{i}(N_{i}-1)}X_{ij}(N_{i}-X_{ij})$ to estimate $\Omega_{ij}(1-\Omega_{ij})$. This gives rise to an unbiased estimator of $\rho^{2}$ as $T=\sum_{j=1}^{p}T_{j},\quad T_{j}=\sum_{k=1}^{K}\biggl{[}n_{k}\bar{N}_{k}(\hat{\mu}_{kj}-\hat{\mu}_{j})^{2}-\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}\sum_{i\in S_{k}}\frac{X_{ij}(N_{i}-X_{ij})}{N_{i}-1}\biggr{]}.$ (2.5) ###### Lemma 2.1. Under Models (1.1)-(1.2), the estimator in (2.5) satisfies that $\mathbb{E}[T]=\rho^{2}$. To use $T$ for hypothesis testing, we need a proper standardization of this statistic. In Sections A.1-A.2 of the appendix , we study $\mathbb{V}(T)$, the variance of $T$. Under mild regularity conditions, it can be shown that $\mathbb{V}(T)=\Theta_{n}\cdot[1+o(1)]$, where $\displaystyle\Theta_{n}:=4\sum_{k=1}^{K}\sum_{j=1}^{p}n_{k}\bar{N}_{k}(\mu_{kj}-\mu_{j})^{2}\mu_{kj}+2\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{j=1}^{p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\frac{N_{i}^{3}}{N_{i}-1}\Omega_{ij}^{2}$ (2.6) $\displaystyle+\frac{2}{n^{2}\bar{N}^{2}}\sum_{1\leq k\neq\ell\leq K}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j=1}^{p}N_{i}N_{m}\Omega_{ij}\Omega_{mj}+2\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k},\\\ i\neq m\end{subarray}}\sum_{j=1}^{p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}N_{i}N_{m}\Omega_{ij}\Omega_{mj}.$ In $\Theta_{n}$, the first term vanishes under the null, so it suffices to estimate the other three terms in $\Theta_{n}$. By properties of multinomial distributions, $\mathbb{E}[X_{ij}X_{mj}]=N_{i}N_{m}\Omega_{ij}\Omega_{mj}$, $\mathbb{E}[X^{2}_{ij}]=N_{i}^{2}\Omega_{ij}^{2}+N_{i}\Omega_{ij}(1-\Omega_{ij})$, and $\mathbb{E}[X_{ij}(N_{i}-X_{ij})]=N_{i}(N_{i}-1)\Omega_{ij}(1-\Omega_{ij})$. It inspires us to estimate $\Omega_{ij}\Omega_{mj}$ by $\frac{X_{ij}X_{mj}}{N_{i}N_{m}}$ and estimate $\Omega_{ij}^{2}$ by $\frac{X_{ij}^{2}}{N_{i}^{2}}-\frac{X_{ij}(N_{i}-X_{ij})}{N^{2}_{i}(N_{i}-1)}=\frac{X_{ij}^{2}-X_{ij}}{N_{i}(N_{i}-1)}$. Define $\displaystyle V=2\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{j=1}^{p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\frac{X_{ij}^{2}-X_{ij}}{N_{i}(N_{i}-1)}+\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j=1}^{p}X_{ij}X_{mj}$ (2.7) $\displaystyle\qquad+2\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k},\\\ i\neq m\end{subarray}}\sum_{j=1}^{p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}X_{ij}X_{mj}.$ (2.8) The test statistic we propose is as follows (in the rate event $V<0$, we simply set $\psi=0$): $\psi=T/\sqrt{V}.$ (2.9) We call $\psi$ the DEbiased and Length-adjusted Variability Estimator (DELVE). In Section 3.1, we show that under mild regularity conditions, $\psi\to N(0,1)$ under the null hypothesis. For any fixed $\alpha\in(0,1)$, the asymptotic level-$\alpha$ DELVE test rejects $H_{0}$ if $\psi>z_{\alpha},\qquad\mbox{where $z_{\alpha}$ is the $(1-\alpha)$-quantile of $N(0,1)$}.$ (2.10) ### 2.1 The special cases of $K=n$ and $K=2$ As seen in Section 1, the application examples of $K=n$ and $K=2$ are particularly intriguing. In these cases, we give more explicit expressions of our test statistic. When $K=n$, we have $S_{k}=\\{i\\}$ and $\hat{\mu}_{kj}=N_{i}^{-1}X_{ij}$. The null hypothesis becomes $H_{0}:\Omega_{1}=\Omega_{2}=\ldots=\Omega_{n}.$ The statistic in (2.5) reduces to $T=\sum_{j=1}^{p}\sum_{i=1}^{n}\biggl{[}\frac{(X_{ij}-N_{i}\hat{\mu}_{j})^{2}}{N_{i}}-\Bigl{(}1-\frac{N_{i}}{n\bar{N}}\Bigr{)}\frac{X_{ij}(N_{i}-X_{ij})}{N_{i}(N_{i}-1)}\biggr{]}.$ (2.11) Moreover, in the variance estimate (2.7), the last term is exactly zero, and it can be shown that the third term is negligible compared to the first term. We thereby consider a simpler variance estimator by only retaining the first term in (2.7): $V^{*}=2\sum_{i=1}^{n}\sum_{j=1}^{p}\Bigl{(}\frac{1}{N_{i}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\frac{X_{ij}^{2}-X_{ij}}{N_{i}(N_{i}-1)}.$ (2.12) The simplified DELVE test statistic is $\psi^{*}=T/\sqrt{V^{*}}$. When $K=2$, we observe two collections of multinomial vectors, denoted by $\\{X_{i}\\}_{1\leq i\leq n}$ and $\\{G_{i}\\}_{1\leq i\leq m}$. We assume for $1\leq i\leq n$ and $1\leq j\leq m$, $X_{i}\sim\mathrm{Multinomial}(N_{i},\Omega_{i}),\qquad G_{j}\sim\mathrm{Multinomial}(M_{j},\Gamma_{j}).$ (2.13) Write $\bar{N}=n^{-1}\sum_{i=1}^{n}N_{i}$ and $\bar{M}=m^{-1}\sum_{i=1}^{m}M_{i}$. The null hypothesis becomes $H_{0}:\quad\eta=\theta,\qquad\mbox{where }\eta=\frac{1}{n\bar{N}}\sum_{i=1}^{n}N_{i}\Omega_{i},\mbox{ and }\theta=\frac{1}{m\bar{M}}\sum_{i=1}^{m}M_{i}\Gamma_{i},$ (2.14) where $\theta$ and $\eta$ are the two group-wise mean PMFs. We estimate them by $\hat{\eta}=(n\bar{N})^{-1}\sum_{i=1}^{n}X_{i}$ and $\hat{\theta}=(m\bar{M})^{-1}\sum_{i=1}^{m}G_{i}$. The statistic in (2.5) has an equivalent form as follows: $T=\frac{n\bar{N}m\bar{M}}{n\bar{N}+m\bar{M}}\biggl{[}\|\hat{\eta}-\hat{\theta}\|^{2}-\sum_{i=1}^{n}\sum_{j=1}^{p}\frac{X_{ij}(N_{i}-X_{ij})}{n^{2}\bar{N}^{2}(N_{i}-1)}-\sum_{i=1}^{m}\sum_{j=1}^{p}\frac{G_{ij}(M_{i}-G_{ij})}{m^{2}\bar{M}^{2}(M_{i}-1)}\biggr{]}.$ (2.15) The variance estimate (2.7) has an equivalent form as follows: $\displaystyle V=\frac{4\sum_{i=1}^{n}\sum_{i^{\prime}=1}^{m}\sum_{j=1}^{p}X_{ij}G_{i^{\prime}j}}{(n\bar{N}+m\bar{M})^{2}}+\frac{2m^{2}\bar{M}^{2}\big{[}\sum_{i=1}^{n}\frac{X_{ij}^{2}-X_{ij}}{N_{i}(N_{i}-1)}+\sum_{1\leq i\neq i^{\prime}\leq n}X_{ij}X_{i^{\prime}j}\big{]}}{n^{2}\bar{N}^{2}(n\bar{N}+m\bar{M})^{2}}$ (2.16) $\displaystyle\quad+\frac{2n^{2}\bar{N}^{2}\big{[}\sum_{i=1}^{m}\frac{G_{ij}^{2}-G_{ij}}{M_{i}(M_{i}-1)}+\sum_{1\leq i\neq i^{\prime}\leq m}G_{ij}G_{i^{\prime}j}\big{]}}{m^{2}\bar{M}^{2}(n\bar{N}+m\bar{M})^{2}}.$ (2.17) The DELVE test statistic is $\psi=T/\sqrt{V}$. ### 2.2 A variant: DELVE+ We introduce a variant of the DELVE test statistic to better suit real data. Let $\hat{\mu}$, $T$ and $V$ be as in (2.3), (2.5) and (2.7). Define $\psi^{+}=T/\sqrt{V^{+}},\qquad\mbox{where}\quad V^{+}=V\cdot\bigl{(}1+\|\hat{\mu}\|_{2}T/\sqrt{V}\bigr{)}.$ (2.18) We call (2.18) the DELVE+ test statistic. In theory, this modification has little effect on the key properties of the test. To see this, we note that $\|\hat{\mu}\|_{2}=o_{\mathbb{P}}(1)$ in high-dimensional settings. Suppose $T/\sqrt{V}\to N(0,1)$ under $H_{0}$. Since $\|\hat{\mu}\|_{2}\to 0$, it is seen immediately that $V^{+}/V\to 1$; hence, the asymptotic normality also holds for $\psi^{+}$. Suppose $T/\sqrt{V}\to\infty$ under the alternative hypothesis. It follows that $V^{+}\leq 2\max\\{V,\|\hat{\mu}\|_{2}\cdot T\sqrt{V}\\}$ and $\psi^{+}\geq\frac{1}{\sqrt{2}}\min\\{T/\sqrt{V},\,\|\hat{\mu}\|_{2}^{-1}(T/\sqrt{V})^{1/2}\\}\to\infty$. We have proved the following lemma: ###### Lemma 2.2. As $n\bar{N}\to\infty$, suppose $\|\hat{\mu}\|_{2}\to 0$ in probability. Under $H_{0}$, if $T/\sqrt{V}\to N(0,1)$, then $T/\sqrt{V^{+}}\to N(0,1)$. Under $H_{1}$, if $T/\sqrt{V}\to\infty$, then $T/\sqrt{V^{+}}\to\infty$. In practice, this modification avoids extremely small $p$-values. In some real datasets, $V$ is very small and leads to an extremely small $p$-value in the original DELVE test. In DELVE+, as long as $T$ is positive, $\psi^{+}$ is smaller than $\psi$, so that the $p$-value is adjusted. In the numerical experiments, we consider both DELVE and DELVE+. For theoretical analysis, since these two versions have almost identical theoretical properties, we only focus on the original DELVE test statistic. ## 3 Theoretical Properties We first present the regularity conditions. For a constant $c_{0}\in(0,1)$, we assume $\min_{1\leq i\leq n}N_{i}\geq 2,\qquad\max_{1\leq i\leq n}\|\Omega_{i}\|_{\infty}\leq 1-c_{0},\qquad\max_{1\leq k\leq K}\frac{n_{k}\bar{N}_{k}}{n\bar{N}}\leq 1-c_{0}.$ (3.1) In (3.1), the first condition is mild. The second condition is also mild: note that $\|\Omega_{i}\|_{1}=1$ for each $i$; this condition excludes those cases where one of the $p$ categories has an extremely dominating probability in the PMF $\Omega_{i}$. In the third condition, $n_{k}\bar{N}_{k}$ is the total number of counts in all multinomials of group $k$, and this condition excludes the extremely unbalanced case where one group occupies the majority of counts. Note that in the special case of $K=2$, we relax this condition to allow for severely unbalanced groups (see Section 3.4). Recall that $\mu_{k}=\frac{1}{n_{k}\bar{N}_{k}}\sum_{i\in S_{k}}N_{i}\Omega_{i}$ is the mean PMF within group $k$. We also define a ‘covariance’ matrix of PMF’s for group $k$ by $\Sigma_{k}=\frac{1}{n_{k}\bar{N}_{k}}\sum_{i\in S_{k}}N_{i}\Omega_{i}\Omega_{i}^{\prime}$. Let $\alpha_{n}:=\max\left\\{\sum_{k=1}^{K}\frac{\|\mu_{k}\|_{3}^{3}}{n_{k}\bar{N}_{k}},\quad\sum_{k=1}^{K}\frac{\|\mu_{k}\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}\right\\}\bigg{/}\bigg{(}\sum_{k=1}^{K}\|\mu_{k}\|^{2}\bigg{)}^{2},$ (3.2) and $\beta_{n}:=\max\biggl{\\{}\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{N^{2}_{i}}{n_{k}^{2}\bar{N}_{k}^{2}}\|\Omega_{i}\|_{3}^{3},\quad\sum_{k=1}^{K}\|\Sigma_{k}\|_{F}^{2}\bigg{\\}}\bigg{/}(K\|\mu\|^{2}).$ (3.3) We assume that as $n\bar{N}\to\infty$, $\alpha_{n}=o(1),\qquad\beta_{n}=o(1),\qquad\mbox{and}\quad\frac{\|\mu\|_{4}^{4}}{K\|\mu\|^{4}}=o(1).$ (3.4) Here $\alpha_{n}$ and $\beta_{n}$ only depend on group-wise quantities, such as $\mu_{k}$, $\Sigma_{k}$ and $\sum_{i\in S_{k}}N^{2}_{i}\|\Omega_{i}\|_{3}^{3}$; hence, a small number of ‘outliers’ (i.e., extremely large entries) in $\Omega$ has little effect on $\alpha_{n}$ and $\beta_{n}$. Furthermore, in a simple case where $\max_{k}n_{k}\leq C\min_{k}n_{k}$, $\max_{k}\bar{N}_{k}\leq C\min_{k}\bar{N}_{k}$ and $\|\Omega\|_{\max}=O(1/p)$, it holds that $\alpha_{n}=O(\max\\{\frac{1}{n\bar{N}},\frac{Kp}{n^{2}\bar{N}^{2}}\\})$, $\beta_{n}=O(\max\\{\frac{K^{2}}{n^{2}p},\frac{1}{p}\\})$ and $\frac{\|\mu\|_{4}^{4}}{K\|\mu\|^{4}}=O(\frac{1}{Kp})$. When $n\bar{N}\to\infty$ and $p\to\infty$, (3.4) reduces to $n^{2}\bar{N}^{2}/(Kp)\to\infty$. This condition is necessary for successful testing, because our lower bound in Section 3.3 implies that the two hypotheses are asymptotically indistinguishable if $n^{2}\bar{N}^{2}/(Kp)\to 0$. ### 3.1 The asymptotic null distribution Under the null hypothesis, the $K$ group-wise mean PMF’s $\mu_{1},\mu_{2},\ldots,\mu_{K}$, are equal to each other, but this hypothesis is still highly composite, as $(N_{i},\Omega_{i})$ are not necessarily the same within each group. We show that the DELVE test statistic always enjoys a parameter-free asymptotic null distribution. Let $T$, $\Theta_{n}$ and $V$ be as in (2.5)-(2.7). The next two theorems are proved in the appendix. ###### Theorem 3.1. Consider Models (1.1)-(1.2), where the null hypothesis (1.3) holds. Suppose (3.1) and (3.4) are satisfied. As $n\bar{N}\to\infty$, $T/\sqrt{\Theta_{n}}\to N(0,1)$ in distribution. ###### Theorem 3.2. Under the conditions of Theorem 3.1, as $n{\bar{N}}\to\infty$, $V/\Theta_{n}\to 1$ in probability, and $\psi:=T/\sqrt{V}\to N(0,1)$ in distribution. By Theorem 3.2, the asymptotic $p$-value is computed via $1-\Phi(\psi)$, where $\Phi(\cdot)$ is the cumulative distribution function of the standard normal. Moreover, for any fixed $\alpha\in(0,1)$, the rejection region of the asymptotic level-$\alpha$ test is as given in (2.10). The proofs of Theorems 3.1-3.2 contain two key steps: in the first step, we decompose $T$ into the sum of mutually uncorrelated terms. We introduce a set of independent, mean-zero random vectors $\\{Z_{ir}\\}_{1\leq i\leq n,1\leq r\leq N_{i}}$, where $Z_{ir}\sim\mathrm{Multinomial}(1,\Omega_{i})-\Omega_{i}$. By properties of multinomial distributions, $X_{i}=N_{i}\Omega_{i}+\sum_{r=1}^{N_{i}}Z_{ir}$ in distribution. We plug it into (2.5) to obtain $T=T_{1}+T_{2}+T_{3}+T_{4}$, where $T_{1}$ is a linear form of $\\{Z_{ir}\\}$, $T_{2}$, $T_{3}$ and $T_{4}$ are quadratic forms of $\\{Z_{ir}\\}$, and the four terms are uncorrelated with each other (details are contained in Section A of the appendix ). In the second step, we construct a martingale for each term $T_{j}$. This is accomplished by rearranging the double-index sequence $Z_{ir}$ to a single- index sequence and then successively adding terms in this sequence to $T_{j}$. We then apply the martingale central limit theorem (CLT) (Hall and Heyde, 2014) to prove the asymptotic normality of each $T_{j}$. The asymptotic normality of $T$ follows by identifying the dominating terms in $T_{1}$-$T_{4}$ (as model parameters change, the dominating terms can be different) and studying their joint distribution. This step involves extensive calculations to bound the conditional variance and to verify the Lindeberg conditions of the martingale CLT, as well as numerous subtle uses of the Cauchy-Schwarz inequality to simplify the moment bounds. ### 3.2 Power analysis Under the alternative hypothesis, the PMF’s $\mu_{1},\mu_{2},\ldots,\mu_{K}$ are not the same. In Section 2, we introduce a quantity $\rho^{2}$ (see (2.2)) to capture the total variation in $\mu_{k}$’s, but this quantity is not scale- free. We define a scaled version of $\rho^{2}$ as $\omega_{n}=\omega_{n}(\mu_{1},\mu_{2},\ldots,\mu_{K}):=\frac{1}{n\bar{N}\|\mu\|^{2}}\sum_{k=1}^{K}n_{k}\bar{N}_{k}\|\mu_{k}-\mu\|^{2}.$ (3.5) It is seen that $\omega_{n}\leq\max_{k}\\{\frac{\|\mu_{k}-\mu\|^{2}}{\|\mu\|^{2}}\\}$, which is properly scaled. ###### Theorem 3.3. Consider Models (1.1)-(1.2), where (3.1) and (3.4) are satisfied. Then, $\mathbb{E}[T]=n\bar{N}\|\mu\|^{2}\omega_{n}^{2}$, and $\mathbb{V}(T)=O\bigl{(}\sum_{k=1}^{K}\|\mu_{k}\|^{2}\bigr{)}+\mathbb{E}[T]\cdot O\bigl{(}\max_{1\leq k\leq K}\|\mu_{k}\|_{\infty}\bigr{)}$. For the DELVE test to have an asymptotically full power, we need $\mathbb{E}[T]\gg\sqrt{\mathbb{V}(T)}$. By Theorem 3.3, this is satisfied if $\mathbb{E}[T]\gg\sqrt{\sum_{k}\|\mu_{k}\|^{2}}$ and $\mathbb{E}[T]\gg\max_{k}\|\mu_{k}\|_{\infty}$. Between these two requirements, the latter one is weaker; hence, we only need $\mathbb{E}[T]\gg\sqrt{\sum_{k=1}^{K}\|\mu_{k}\|^{2}}$. It gives rise to the following theorem: ###### Theorem 3.4. Under the conditions of Theorem 3.3, we further assume that under the alternative hypothesis, as $n\bar{N}\to\infty$, $\mathrm{SNR}_{n}:=\frac{n\bar{N}\|\mu\|^{2}\omega_{n}^{2}}{\sqrt{\sum_{k=1}^{K}\|\mu_{k}\|^{2}}}\;\;\to\;\;\infty.$ (3.6) The following statements are true. Under the alternative hypothesis, $\psi\to\infty$ in probability. For any fixed $\alpha\in(0,1)$, the level-$\alpha$ DELVE test has an asymptotic level of $\alpha$ and an asymptotic power of $1$. If we choose $\alpha=\alpha_{n}$ such that $\alpha_{n}\to 0$ and $1-\Phi(\mathrm{SNR}_{n})=o(\alpha_{n})$, where $\Phi$ is the CDF of $N(0,1)$, then the sum of type I and type II errors of the DELVE test converges to $0$. The detection boundary in (3.6) has simpler forms in some special cases. For example, if $\|\mu_{k}\|\asymp\|\mu\|$ for $1\leq k\leq K$, then $\mathrm{SRN}_{n}\asymp n\bar{N}\omega_{n}^{2}\|\mu\|/\sqrt{K}$. If, furthermore, all entries of $\mu$ are at the same order, which implies $\|\mu\|\asymp p^{-1/2}$, then $\mathrm{SRN}_{n}\asymp n^{2}\bar{N}^{2}\omega_{n}^{2}/\sqrt{Kp}$. In this case, the detection boundary simplifies to $\omega_{n}^{4}n^{2}\bar{N}^{2}/(Kp)\to\infty.$ ###### Remark 1 (The low-dimensional case $p=O(1)$). Although we are primarily interested in the high-dimensional setting $p\to\infty$, it is also worth investigating the case $p=O(1)$. We can show the same detection boundary for our test, but the asymptotic normality may not hold, because the variance estimator $V$ in (2.7) is not guaranteed to be consistent. To fix this issue, we propose a variant of our test by replacing $V$ with a refined variance estimator $\widetilde{V}$, which is consistent for a finite $p$. The expression of $\widetilde{V}$ is a little complicated. Due to space limits, we relegate it to Section E of the appendix. ### 3.3 A matching lower bound We have seen that the DELVE test successfully separates two hypotheses if $\mathrm{SNR}_{n}\to\infty$, where $\mathrm{SNR}_{n}$ is as defined in (3.6). We now present a lower bound to show that the two hypotheses are asymptotically indistinguishable if $\mathrm{SNR}_{n}\to 0$. Let $\ell_{i}\in\\{1,2,\ldots,K\\}$ denote the group label of $X_{i}$. Write $\xi=\\{(N_{i},\Omega_{i},\ell_{i})\\}_{1\leq i\leq n}$. Let $\mu_{k}$, $\alpha_{n}$, $\beta_{n}$, and $\omega_{n}$ be the same as defined in (1.2), (3.2), (3.3), and (3.5), respectively. For each given $(n,p,K,\bar{N})$, we write $\mu_{k}=\mu_{k}(\xi)$ to emphasize its dependence on parameters, and similarly for $\alpha_{n},\beta_{n},\omega_{n}$. For any $c_{0}\in(0,1)$ and sequence $\epsilon_{n}$, define ${\cal Q}_{n}(c_{0},\epsilon_{n}):=\Big{\\{}\xi=\\{(N_{i},\Omega_{i},\ell_{i})\\}_{i=1}^{n}:\,\mbox{\eqref{cond1-basic} holds for $c_{0}$},\,\,\max(\alpha_{n}(\xi),\beta_{n}(\xi))\leq\epsilon_{n}\Big{\\}}$ (3.7) Furthermore, for any sequence $\delta_{n}$, we define a parameter class for the null hypothesis and a parameter class for the alternative hypothesis: $\displaystyle{\cal Q}_{0n}^{*}(c_{0},\epsilon_{n})$ $\displaystyle={\cal Q}_{n}(c_{0},\epsilon_{n})\cap\left\\{\xi:\omega_{n}(\xi)=0\right\\},$ (3.8) $\displaystyle{\cal Q}_{1n}^{*}(\delta_{n};c_{0},\epsilon_{n})$ $\displaystyle={\cal Q}_{n}(c_{0},\epsilon_{n})\cap\left\\{\xi:\frac{n\bar{N}\|\mu(\xi)\|^{2}\omega^{2}_{n}(\xi)}{\sqrt{\sum_{k=1}^{K}\|\mu_{k}(\xi)\|^{2}}}\geq\delta_{n}\right\\}.$ (3.9) ###### Theorem 3.5. Fix a constant $c_{0}\in(0,1)$ and two positive sequences $\epsilon_{n}$ and $\delta_{n}$ such that $\epsilon_{n}\to 0$ as $n\to\infty$. For any sequence of $(n,p,K,\bar{N})$ indexed by $n$, we consider Models (1.1)-(1.2) for $\Omega\in{\cal Q}_{n}(c_{0},\epsilon_{n})$. Let ${\cal Q}_{0n}^{*}(c_{0},\epsilon_{n})$ and ${\cal Q}_{1n}^{*}(\delta_{n};c_{0},\epsilon_{n})$ be as in (3.8). If $\delta_{n}\to 0$, then $\limsup_{n\to\infty}\inf_{\Psi\in\\{0,1\\}}\bigl{\\{}\sup_{\xi\in{\cal Q}_{0n}^{*}(c_{0},\epsilon_{n})}\mathbb{P}_{\xi}(\Psi=1)+\sup_{\xi\in{\cal Q}_{1n}^{*}(\delta_{n};c_{0},\epsilon_{n})}\mathbb{P}_{\xi}(\Psi=0)\bigr{\\}}=1.$ By Theorem 3.5, the null and alternative hypotheses are asymptotically indistinguishable if $\mathrm{SRN}_{n}\to 0$. Combining it with Theorem 3.4, the DELVE test achieves the minimax optimal detection boundary. ### 3.4 The special case of $K=2$ The special case of $K=2$ is found in applications such as closeness testing and authorship attribution. We study this case more carefully. Given $\\{X_{i}\\}_{1\leq i\leq n}$ and $\\{G_{i}\\}_{1\leq i\leq m}$, we assume $X_{i}\sim\mathrm{Multinomial}(N_{i},\Omega_{i}),\qquad G_{j}\sim\mathrm{Multinomial}(M_{j},\Gamma_{j}).$ (3.10) Write $\bar{N}=n^{-1}\sum_{i=1}^{n}N_{i}$ and $\bar{M}=m^{-1}\sum_{i=1}^{m}M_{i}$. The null hypothesis becomes $H_{0}:\quad\eta=\theta,\qquad\mbox{where }\eta=\frac{1}{n\bar{N}}\sum_{i=1}^{n}N_{i}\Omega_{i},\mbox{ and }\theta=\frac{1}{m\bar{M}}\sum_{i=1}^{m}M_{i}\Gamma_{i},$ (3.11) where $\theta$ and $\eta$ are the two group-wise mean PMFs. In this case, the test statistic $\psi$ has a more explicit form as in (2.15)-(2.16). In our previous results for a general $K$, the regularity conditions (e.g., (3.1)) impose restrictions on the balance of sample sizes among groups. For $K=2$, the severely unbalanced setting is interesting (e.g., in authorship attribution, $n=1$ and $m$ can be large). We relax the regularity conditions to the following ones: ###### Condition 3.1. Let $\theta$ and $\eta$ be as in (3.11) and define two matrices $\Sigma_{1}=\frac{1}{n\bar{N}}\sum_{i=1}^{n}N_{i}\Omega_{i}\Omega_{i}^{\prime}$ and $\Sigma_{2}=\frac{1}{m\bar{M}}\sum_{i=1}^{m}M_{i}\Gamma_{i}\Gamma_{i}^{\prime}$. We assume that the following statements are true (a) For $1\leq i\leq n$ and $1\leq j\leq m$, $N_{i}\geq 2$, $\|\Omega_{i}\|_{\infty}\leq 1-c_{0}$, $M_{j}\geq 2$, and $\|\Gamma_{j}\|_{\infty}\leq 1-c_{0}$, where $c_{0}\in(0,1)$ is a contant, (b) $\max\big{\\{}\big{(}\frac{\|\eta\|_{3}^{3}}{n\bar{N}}+\frac{\|\theta\|_{3}^{3}}{m\bar{M}}\big{)},\,\big{(}\frac{\|\eta\|_{2}^{2}}{n^{2}\bar{N}^{2}}+\frac{\|\theta\|_{2}^{2}}{m^{2}\bar{M}_{2}^{2}}\big{)}\big{\\}}\big{/}\bigl{\|}\frac{m\bar{M}}{n\bar{N}+m\bar{M}}\eta+\frac{n\bar{N}}{n\bar{N}+m\bar{M}}\theta\bigr{\|}^{4}=o(1)$, (c) $\max\big{\\{}\sum_{i}\frac{N_{i}^{2}}{n^{2}\bar{N}^{2}}\|\Omega_{i}\|_{3}^{3},\,\sum_{i}\frac{M_{i}^{2}}{m^{2}\bar{M}^{2}}\|\Gamma_{i}\|_{3}^{3},\,\|\Sigma_{1}\|_{F}^{2}+\|\Sigma_{2}\|_{F}^{2}\big{\\}}\big{/}\|\mu\|^{2}=o(1)$, and (d) $\|\mu\|_{4}^{4}/\|\mu\|^{4}=o(1)$. Condition (a) is similar to (3.1), except that we drop the sample size balance requriement. Conditions (b)-(d) are equivalent to (3.4) but have more explicit expressions for $K=2$. ###### Theorem 3.6. In Model (3.10), we test the null hypothesis $H_{0}$: $\theta=\mu$. As $\min\\{n{\bar{N}},m\bar{M}\\}\to\infty$, suppose Condition 3.1 is satisfied. Under the alternative hypothesis, we further assume $\frac{\|\eta-\theta\|^{2}}{\big{(}\frac{1}{n\bar{N}}+\frac{1}{m\bar{M}}\big{)}\max\\{\|\eta\|,\,\|\theta\|\\}}\to\infty.$ (3.12) Consider the DELVE test statistic $\psi=T/\sqrt{V}$. The following statements are true. Under the null hypothesis, $\psi\to N(0,1)$ in distribution. Under the alternative hypothesis, $\psi\to\infty$ in probability. Moreover for any fixed $\alpha\in(0,1)$, the level-$\alpha$ DELVE test has an asymptotic level of $\alpha$ and an asymptotic power of $1$. Compared with the theorems for a general $K$, first, Theorem 3.6 allows the two groups to be severely unbalanced and reveals that the detection boundary depends on the harmonic mean of $n\bar{N}$ and $m\bar{M}$. Second, the detection boundary is expressed using $\|\eta-\theta\|$, which is easier to interpret. ### 3.5 The special case of $K=n$ The special case of $K=n$ is interesting for two reasons. First, the application example of global testing in topic models corresponds to $K=n$. Second, for any $K$, when $\Omega_{i}$’s within each group are assumed to be the same (e.g., this is the case in closeness testing of discrete distributions), it suffices to aggregate the counts in each group, i.e., let $Y_{k}=\sum_{i\in S_{k}}X_{i}$ and operate on $Y_{1},\ldots,Y_{K}$ instead of the original $X_{i}$’s; this reduces to the case of $K=n$. When $K=n$, the null hypothesis has a simpler form: $H_{0}:\quad\Omega_{i}=\mu,\qquad 1\leq i\leq n.$ (3.13) Moreover, under the alternative hypothesis, the quantity $\omega_{n}^{2}$ in (3.5) simplifies to $\omega_{n}=\omega_{n}(\Omega_{1},\Omega_{2},\ldots,\Omega_{n})=\frac{1}{n\bar{N}\|\mu\|^{2}}\sum_{i=1}^{n}N_{i}\|\Omega_{i}-\mu\|^{2}.$ (3.14) The DELVE test statistic also has a simplified form as in (2.11)-(2.12). We can prove the same theoretical results under weaker conditions: ###### Condition 3.2. We assume that the following statements are true: (a) For a constant $c_{0}\in(0,1)$, $2\leq N_{i}\leq(1-c_{0})n\bar{N}$ and $\|\Omega_{i}\|_{\infty}\leq 1-c_{0}$, $1\leq i\leq n$, and (b) $\max\big{\\{}\sum_{i}\frac{\|\Omega_{i}\|_{3}^{3}}{N_{i}},\,\sum_{i}\frac{\|\Omega_{i}\|^{2}}{N_{i}^{2}}\big{\\}}\big{/}(\sum_{i}\|\Omega_{i}\|^{2})^{2}=o(1)$, and $(\sum_{i}\|\Omega_{i}\|_{3}^{3})/(n\|\mu\|^{2})=o(1)$ When $K=n$, Condition (a) is equivalent to (3.1); and Condition (b) is weaker than (3.4), as we have dropped the requirement $\frac{\|\mu\|_{4}^{4}}{K\|\mu\|^{4}}=o(1)$. We obtain weaker conditions for $K=n$ because the dominant terms in $T$ differ from those for $K<n$. ###### Theorem 3.7. In Model (1.1), we test the null hypothesis (3.13). As $n\to\infty$, we assume that Condition 3.2 is satisfied. Under the alternative, we further assume that $\frac{n\bar{N}\|\mu\|^{2}\omega_{n}^{2}}{\sqrt{\sum_{i=1}^{n}\|\Omega_{i}\|^{2}}}\to\infty.$ (3.15) Let $T$ and $V^{*}$ be the same as in (2.11)-(2.12). Consider the simplified DELVE test statistic $\psi^{*}=T/\sqrt{V^{*}}$. The following statements are true. Under the null hypothesis, $\psi^{*}\to N(0,1)$ in distribution. Under the alternative hypothesis, $\psi^{*}\to\infty$ in probability. Moreover for any fixed $\alpha\in(0,1)$, the level-$\alpha$ DELVE test has an asymptotic level of $\alpha$ and an asymptotic power of $1$. The detection boundary in (3.15) has a simpler form if $\sum_{i}\|\Omega_{i}\|^{2}\asymp n\|\mu\|^{2}$. In this case, (3.15) is equivalent to $\sqrt{n}\bar{N}\|\mu\|\omega_{n}^{2}\to\infty.$ Additionally, if all entries of $\mu$ are at the same order, then $\|\mu\|\asymp 1/\sqrt{p}$, and (3.15) further reduces to $\sqrt{n\bar{N}^{2}/p}\cdot\omega_{n}^{2}\to\infty$. ### 3.6 A discussion of the contiguity regime Our power analysis in Section 3.2 concerns $\mathrm{SNR}_{n}\to\infty$, and our lower bound in Section 3.3 concerns $\mathrm{SNR}_{n}\to 0$. We now study the contiguity regime where $\mathrm{SNR}_{n}$ tends to a constant. For illustration, we consider a special choice of parameters, which allows us to obtain a simple expression of the testing risk. Suppose $K=n$ and $N_{i}=N$ for all $1\leq i\leq n$. Consider the pair of hypotheses: $H_{0}:\;\;\Omega_{ij}=p^{-1},\qquad\mbox{v.s.}\qquad H_{1}:\;\;\Omega_{ij}=p^{-1}(1+\beta_{n}\delta_{ij}),$ (3.16) where $\\{\delta_{ij}\\}_{1\leq i\leq n,1\leq j\leq p}$ satisfy that $|\delta_{ij}|=1$, $\sum_{j=1}^{p}\delta_{ij}=0$ and $\sum_{i=1}^{n}\delta_{ij}=0$. Such $\delta_{ij}$ always exist.333For example, we can first partition the dictionary into two halves and then partition all the documents into two halves; this divides $\\{1,2,\ldots,p\\}\times\\{1,2,\ldots,n\\}$ into four subsets; we construct $\delta_{ij}$’s freely on one subset and then specify the $\delta_{ij}$’s on the other three subsets by symmetry. The $\mathrm{SNR}_{n}$ in (3.6) satisfies that $\mathrm{SNR}_{n}\asymp(N\sqrt{n}/\sqrt{p})\beta_{n}^{2}$. We thereby set $\beta_{n}^{2}=\frac{\sqrt{2p}}{N\sqrt{n}}\cdot a,\qquad\mbox{for a constant }a>0.$ (3.17) Since $K=n$ here, we consider the simplified DELVE test statistic $\psi^{*}$ as in Section 3.5. ###### Theorem 3.8. Consider Model (1.1) with $N_{i}=N$. For a constant $a>0$, let the null and alternative hypotheses be specified as in (3.16)-(3.17). As $n\to\infty$, if $p=o(N^{2}n)$, then $\psi^{*}\to N(0,1)$ under $H_{0}$ and $\psi^{*}\to N(a,1)$ under $H_{1}$. Let $\Phi$ be the cumulative distribution function of the standard normal. By Theorem 3.8, for any fixed constant $t\in(0,a)$, if we reject the null hypothesis when $\psi^{*}>t$, then the sum of type I and type II errors converges to $[1-\Phi(t)]+[1-\Phi(a-t)]$. ## 4 Applications As mentioned in Section 1, our testing problem includes global testing for topic models, authorship attribution, and closeness testing for discrete distributions as special examples. In this section, the DELVE test is applied separately to these three problems. ### 4.1 Global testing for topic models Topic modeling (Blei et al., 2003) is a popular tool in text mining. It aims to learn a small number of “topics” from a large corpus. Given $n$ documents written using a dictionary of $p$ words, let $X_{i}\sim\mathrm{Multinomial}(N_{i},\Omega_{i})$ denote the word counts of document $i$, where $N_{i}$ is the length of this document and $\Omega_{i}\in\mathbb{R}^{p}$ contains the population word frequencies. In a topic model, there exist $M$ topic vectors $A_{1},A_{2},\ldots,A_{M}\in\mathbb{R}^{p}$, where each $A_{k}$ is a PMF. Let $w_{i}\in\mathbb{R}^{M}$ be a nonnegative vector whose entries sum up to $1$, where $w_{i}(k)$ is the “weight” document $i$ puts on topic $k$. It assumes $\Omega_{i}=\sum_{k=1}^{M}w_{i}(k)A_{k},\qquad 1\leq i\leq n.$ (4.1) Under (4.1), the matrix $\Omega=[\Omega_{1},\Omega_{2},\ldots,\Omega_{n}]$ admits a low-rank nonnegative factorization. Before fitting a topic model, we would like to know whether the corpus indeed involves multiple topics. This is the global testing problem: $H_{0}:M=1$ v.s. $H_{1}:M>1$. When $M=1$, by writing $A_{1}=\mu$, the topic model reduces to the null hypothesis in (3.13). We can apply the DELVE test by treating each $X_{i}$ as a separate group (i.e., $K=n$). ###### Corollary 4.1. Consider Model (1.1) and define a vector $\xi\in\mathbb{R}^{n}$ by $\xi_{i}=\bar{N}^{-1}N_{i}$. Suppose that $\Omega=\mu{\bf 1}_{n}^{\prime}$ under the null hypothesis, with $\mu=n^{-1}\Omega\xi$, and that $\Omega$ satisfies (4.1) under the alternative hypothesis, with $r:=\mathrm{rank}(\Omega)\geq 2$. Suppose $\bar{N}/(\min_{i}N_{i})=O(1)$. Denote by $\lambda_{1},\lambda_{2},\ldots,\lambda_{r}>0$ the singular values of $\Omega[\mathrm{diag}(\xi)]^{1/2}$, arranged in the descending order. We further assume that under the alternative hypothesis, $\bar{N}\cdot\frac{\sum_{k=2}^{r}\lambda_{k}^{2}}{\sqrt{\sum_{k=1}^{r}\lambda_{k}^{2}}}\to\infty.$ (4.2) For any fixed $\alpha\in(0,1)$, the level-$\alpha$ DELVE test has an asymptotic level $\alpha$ and an asymptotic power $1$. The least-favorable configuration in the proof of Theorem 3.5 is in fact a topic model that follows (4.1) with $M=2$. Transferring the argument yields the following lower bound that confirms the optimality of DELVE for the global testing of topic models. ###### Corollary 4.2. Let ${\cal R}_{n,M}(\epsilon_{n},\delta_{n})$ be the collection of $\\{(N_{i},\Omega_{i})\\}_{i=1}^{n}$ satisfying the following conditions: 1) $\Omega$ follows the topic model (4.1) with $M$ topics; 2) Condition 3.2 holds with $o(1)$ replaced by $\leq\epsilon_{n}$; 3) $\bar{N}(\sum_{k=2}^{r}\lambda_{k}^{2})/(\sum_{k=1}^{r}\lambda_{k}^{2})^{1/2}\geq\delta_{n}$. If $\epsilon_{n}\to 0$ and $\delta_{n}\to 0$, then $\limsup_{n\to\infty}\inf_{\Psi\in\\{0,1\\}}\Bigl{\\{}\sup_{{\cal R}_{n,1}(\epsilon_{n},0)}\mathbb{P}(\Psi=1)+\sup_{\cup_{M\geq 2}{\cal R}_{n,M}(\delta_{n},\delta_{n})}\mathbb{P}(\Psi=0)\Bigr{\\}}=1.$ The detection boundary (4.2) can be simplified when $M=O(1)$. Following Ke and Wang (2022), we define $\Sigma_{A}=A^{\prime}H^{-1}A$ and $\Sigma_{W}=n^{-1}WW^{\prime}$, where $A=[A_{1},A_{2},\ldots,A_{M}]$, $W=[w_{1},w_{2},\ldots,w_{n}]$ and $H=\mathrm{diag}(A{\bf 1}_{M})$. Ke and Wang (2022) argued that it is reasonable to assume that eigenvalues of these two matrices are at the constant order. If this is true, with some mild additional regularity conditions, each $\lambda_{k}$ is at the order of $\sqrt{n/p}$. Hence, (4.2) reduces to $\sqrt{n}\bar{N}/\sqrt{p}\to\infty.$ In comparison, Ke and Wang (2022) showed that a necessary condition for any estimator $\hat{A}=[\hat{A}_{1},\hat{A}_{2},\ldots,\hat{A}_{M}]$ to achieve $\frac{1}{M}\sum_{k=1}^{M}\|\hat{A}_{k}-A_{k}\|_{1}=o(1)$ is $\sqrt{n\bar{N}/p}\to\infty$. We conclude that consistent estimation of topic vectors requires strictly stronger conditions than successful testing. ### 4.2 Authorship attribution In authorship attribution, given a corpus from a known author, we want to test whether a new document is from the same author. It is a special case of our testing problem with $K=2$. We can directly apply the results in Section 3.4. However, the setting in Section 3.4 has no sparsity. Kipnis (2022); Donoho and Kipnis (2022) point out that the number of words with discriminating power is often much smaller than $p$. To see how our test performs under sparsity, we consider a sparse model. As in Section 3.4, let $X_{i}\sim\mathrm{Multinomial}(N_{i},\Omega_{i}),\;1\leq i\leq n,\quad\mbox{and}\quad G_{i}\sim\mathrm{Multinomial}(M_{i},\Gamma_{i}),\;1\leq i\leq m.$ (4.3) Let $\bar{N}$ and $\bar{M}$ be the average of $N_{i}$’s and $M_{i}$’s, respectively. Write $\eta=\frac{1}{n\bar{N}}\sum_{i=1}^{n}N_{i}\Omega_{i}$ and $\theta=\frac{1}{m\bar{M}}\sum_{i=1}^{m}M_{i}\Gamma_{i}$. We assume for some $\beta_{n}>0$, $\eta_{j}=\theta_{j},\;\;\mbox{for }j\notin S,\qquad\mbox{and}\qquad\bigl{|}\sqrt{\eta_{j}}-\sqrt{\theta_{j}}\bigr{|}\geq\beta_{n},\;\;\mbox{for }j\in S.$ (4.4) ###### Corollary 4.3. Under the model (4.3)-(4.4), consider testing $H_{0}:S=\emptyset$ v.s. $H_{1}:S\neq\emptyset$, where Condition 3.1 is satisfied. Let $\eta_{S}$ and $\theta_{S}$ be the sub-vectors of $\eta$ and $\theta$ restricted to the coordinates in $S$. Suppose that under the alternative hypothesis, $\frac{\beta_{n}^{2}\cdot(\|\eta_{S}\|_{1}+\|\theta_{S}\|_{1})}{\big{(}\frac{1}{n\bar{N}}+\frac{1}{m\bar{M}}\big{)}\max\\{\|\eta\|,\,\|\theta\|\\}}\to\infty.$ (4.5) As $\min\\{n\bar{N},m\bar{M}\\}\to\infty$, the level-$\alpha$ DELVE test has an asymptotic level $\alpha$ and an asymptotic power $1$. Furthermore, if $n\bar{N}\asymp m\bar{M}$ and $\min_{j\in S}(\eta_{j}+\theta_{j})\geq cp^{-1}$ for a constant $c>0$, then (4.5) reduces to $n\bar{N}\beta_{n}^{2}|S|/\sqrt{p}\to\infty$. Donoho and Kipnis (2022) studied a case where $N=M$, $n=m=1$, $p\to\infty$, $|S|=p^{1-\vartheta},\qquad\mbox{and}\qquad\beta_{n}=c\cdot N^{-1/2}\sqrt{\log(p)}.$ (4.6) When $\vartheta>1/2$ (i.e., $|S|=o(\sqrt{p})$), they derived a phase diagram for the aforementioned testing problem (under a slightly different setting where the data distributions are Poisson instead of multinomial). They showed that when $\vartheta>1/2$ and $c$ is a properly large constant, a Higher- Criticism-based test has an asymptotically full power. Donoho and Kipnis (2022) did not study the case of $\vartheta\leq 1/2$. By Corollary 4.3, when $\vartheta\leq 1/2$ (i.e., $|S|\geq C\sqrt{p}$), the DELVE test has asymptotically full power. ###### Remark 2. When $\vartheta>1/2$ in (4.6), the DELVE test is powerless. However, this issue can be resolved by borrowing the idea of maximum test or Higher Criticism test (Donoho and Jin, 2004) from the classical multiple testing. For example, recalling $T_{j}$ in (2.5), we can use $\max_{1\leq j\leq p}\\{T_{j}/\sqrt{V_{j}}\\}$ as the test statistic, where $V_{j}$ is a proper estimator of the variance of $T_{j}$. We leave a careful study of this idea to future work. ### 4.3 Closeness testing between discrete distributions Two-sample closeness testing is a subject of intensive study in discrete distribution inference (Bhattacharya and Valiant, 2015; Chan et al., 2014; Diakonikolas and Kane, 2016; Kim et al., 2022). It is a special case of our problem with $K=2$ and $n_{1}=n_{2}=1$. We thereby apply both Theorem 3.6 and Theorem 3.7. ###### Corollary 4.4. Let $Y_{1}$ and $Y_{2}$ be two discrete variables taking values on the same $p$ outcomes. Let $\Omega_{1}\in\mathbb{R}^{p}$ and $\Omega_{2}\in\mathbb{R}^{p}$ be their corresponding PMFs. Suppose we have $N_{1}$ samples of $Y_{1}$ and $N_{2}$ samples of $Y_{2}$. The data are summarized in two multinomial vectors: $X_{1}\sim\mathrm{Multinomial}(N_{1},\Omega_{1}),X_{2}\sim\mathrm{Multinomial}(N_{2},\Omega_{2}).$ We test $H_{0}:\Omega_{1}=\Omega_{2}.$ Write $\mu=\frac{1}{N_{1}+N_{2}}(N_{1}\Omega_{1}+N_{2}\Omega_{2})$. Suppose $\min\\{N_{1},N_{2}\\}\geq 2$, $\max\\{\|\Omega_{1}\|_{\infty},\|\Omega_{2}\|_{\infty}\\}\leq 1-c_{0}$, for a constant $c_{0}\in(0,1)$. Suppose $\frac{1}{(\sum_{k=1}^{2}\|\Omega_{k}\|^{2})^{2}}\max\big{\\{}\sum_{k=1}^{2}\frac{\|\Omega_{k}\|_{3}^{3}}{N_{k}},\sum_{k=1}^{2}\frac{\|\Omega_{k}\|^{2}}{N_{k}^{2}}\big{\\}}=o(1)$, and $\frac{1}{n\|\mu\|^{2}}\sum_{k=1}^{2}\|\Omega_{k}\|_{3}^{3}=o(1)$. We assume that under the alternative hypothesis, $\frac{\|\Omega_{1}-\Omega_{2}\|^{2}}{\big{(}N_{1}^{-1}+N_{2}^{-1}\big{)}\max\\{\|\Omega_{1}\|,\,\|\Omega_{2}\|\\}}\to\infty.$ (4.7) As $\min\\{N_{1},N_{2}\\}\to\infty$, the level-$\alpha$ DELVE test has level $\alpha$ and power $1$, asymptotically. We notice that (4.7) matches with the minimum $\ell^{2}$-separation condition for two-sample closeness testing (Kim et al., 2022, Proposition 4.4). Therefore, our test is an optimal $\ell^{2}$-testor. Although other optimal $\ell^{2}$-testors have been proposed (Chan et al., 2014; Bhattacharya and Valiant, 2015; Diakonikolas and Kane, 2016), they are not equipped with tractable null distributions. ###### Remark 3. We can modify the DELVE test to incorporate frequency-dependent weights. Given any nonnegative vector $w=(w_{1},w_{2},\ldots,w_{p})^{\prime}$, define $T(w):=\sum_{j=1}^{p}w_{j}T_{j}$ where $T_{j}$ is the same as in (2.5). These weights $w_{j}$ serve to adjust the contributions of different words. For example, consider $w_{j}=\bigl{(}\max\\{1/p,\;\hat{\mu}_{j}\\}\bigr{)}^{-1}$. This kind of weights have been used in discrete distribution inference (Balakrishnan and Wasserman, 2019; Chan et al., 2014) to turn an optimal $\ell^{2}$ testor to an optimal $\ell^{1}$ testor. We can similarly study the power of this modified test, except that we need an additional assumption $n\bar{N}\gg p$ to guarantee that $\hat{\mu}_{j}$ is a sufficiently accurate estimator of $\mu_{j}$. ## 5 Simulations The proposed DELVE test is computationally efficient and easy to implement. In this section, we investigate its numerical performance in simulation studies. Real data analysis will be carried out in Section 6. Figure 1: Histograms of the DELVE statistic (top three panels) and the DELVE+ statistic (bottom three panels) in Experiments 1.1-1.3. In each plot, the blue and orange histograms correspond to the null and alternative hypotheses, respectively; and the green curve is the density of $N(0,1)$. Experiment 1 (Asymptotic normality) . Given $(n,p,K,N_{\min},N_{\max},\alpha)$, we generate data as follows: first, we divide $\\{1,\ldots,n\\}$ into $K$ equal-size groups. Next, we draw $\Omega_{1}^{alt},\ldots,\Omega_{n}^{alt}$ i.i.d. from $\text{Dirichlet}(p,\alpha\mathbf{1}_{p})$. Third, we draw $N_{i}\stackrel{{\scriptstyle iid}}{{\sim}}\text{Uniform}[N_{\min},N_{\max}]$ and set $\Omega_{i}^{null}=\mu$, where $\mu:=\frac{1}{n\bar{N}}\sum_{i}N_{i}\Omega_{i}^{alt}$. Last, we generate $X_{1},\ldots,X_{n}$ using Model (1.1). We consider three sub-experiments. In Experiment 1.1, $(n,p,K,N_{\min},N_{\max},\alpha)=(50,100,5,10,20,0.3)$. In Experiment 1.2, $\alpha$ is changed to $1$, and the other parameters are the same. When $\alpha=1$, $\Omega_{i}^{alt}$ are drawn from the uniform distribution of the standard probability simplex; in comparison, $\alpha=0.3$ puts more mass near the boundary of the standard probability simplex. In Experiment 1.3, we keep all parameters the same as in Experiment 1.1, except that $(p,K)$ are changed to $(300,50)$. For each sub-experiment, we generate 2000 data sets under the null hypothesis and plot the histogram of the DELVE test statistic $\psi$ (in blue); similarly, we generate 2000 data sets under the alternative hypothesis and plot the histogram of $\psi$ (in orange). The results are contained on the top three panels of Figure 1. In Section 2.2, we introduced a variant of DELVE, called DELVE+, in which the variance estimator $V$ is replaced by an adjusted one. DELVE+ has similar theoretical properties as DELVE but is more suitable for real data. We plot the histograms of the DELVE+ test statistics on the bottom three panels of Figure 1. We have several observations. In all sub-experiments, when the null hypothesis holds, the histograms of both DELVE and DELVE+ fit the standard normal density reasonably well. This supports our theory in Section 3.1. Second, when $(p,K)$ increase, the finite sample effect becomes slightly more pronounced (c.f., Experiment 1.3 versus Experiment 1.1). Third, the tests have power in differentiating two hypotheses. As $\alpha$ decreases or $K$ increases, the power increases, and the histograms corresponding to two hypotheses become further apart. Last, in the alternative hypothesis, DELVE+ has smaller mean and variance than DELVE. By Lemma 2.2, these two have similar asymptotic behaviors. The simulation results suggest that they have noticeable finite- sample differences. Figure 2: Power diagrams (based on $500$ repetitions) at level $5\%$. The $x$-axis plots the SNR $\lambda(\omega_{n})=K^{-1/2}n\bar{N}\|\mu\|\cdot\omega_{n}$. Experiment 2 (Power curve). Similarly as before, we divide $\\{1,2,\ldots,n\\}$ into $K$ equal-size groups and draw $N_{i}\sim\text{Uniform}[N_{\min},N_{\max}]$. In this experiment, the PMF’s $\Omega_{i}$ are generated in a different way. Under the null hypothesis, we generate $\mu\sim\text{Dirichlet}(p/2,\alpha\mathbf{1}_{p/2})$ and set $\Omega_{i}^{null}=\tilde{\mu}$, where $\tilde{\mu}_{j}=\frac{1}{2}\mu_{j}$ for $1\leq j\leq p/2$ and $\tilde{\mu}_{j}=\frac{1}{2}\mu_{p+1-j}$ for $p/2+1\leq j\leq p$. Under the alternative hypothesis, we draw $z_{1},\ldots,z_{K}$, $b_{1},\ldots,b_{p/2}\stackrel{{\scriptstyle iid}}{{\sim}}\text{Rademacher}(1/2)$ and then let $\Omega_{ij}^{alt}=\mu(1+\tau_{n}z_{k}b_{j})$, for all $i$ in group $k$ and $1\leq j\leq p/2$, and $\Omega_{ij}^{alt}=\mu(1+\tau_{n}z_{k}b_{j})$ for $p/2+1\leq j\leq p$. By applying our theory in Section 3.2 together with some calculations, we obtain that the signal-to-noise ratio is captured by $\lambda:=K^{-1/2}n\bar{N}\|\mu\|\tau_{n}.$ We consider three sub-experiments, Experiment 2.1-2.3, in which the parameter values of $(n,p,K,N_{\min},N_{\max},\alpha)$ are the same as in Experiments 1.1-1.3. For each sub-experiment, we consider a grid of 10 equally-spaced values of $\lambda$. When $\lambda=0$, it corresponds to the null hypothesis; when $\lambda>0$, it corresponds to the alternative hypothesis. For each $\lambda$, we generate $500$ data sets and compute the fraction of rejections of the level-$5\%$ DELVE test. This gives a power curve for the level-$5\%$ DELVE test, in which the first point corresponding to $\lambda=0$ is the actual level of the test. The results are contained on the top three panels of Figure 2. We repeat the same experiments for the DELVE+ test, which results are on the bottom three panels of Figure 2. In all three experiments, the actual level of our proposed tests is $\leq 5\%$, suggesting that our tests perform well at controlling the type-I error. As $\lambda$ increases, the power gradually increased to $1$, suggesting that $\lambda$ is a good metric of the signal-to-noise ratio. This supports our theory in Section 3.2. Figure 3: (Left) Histogram of nonzero DELVE $Z$-scores for all authors in the dataset. The mean is $4.52$ and the standard deviation is $2.94$. (Right) Scatter plot of author DELVE scores versus the natural log of the number of papers with five statisticians identified. ## 6 Real Data Analysis We apply our proposed methods on two real corpora: one consists of abstracts of research papers in four statistics journals, and the other consists of movie reviews on Amazon. For the analysis of real data, we use DELVE+, which modifies the variance estimator in DELVE and reduces the occurrence of extremely small $p$-values. ### 6.1 Abstracts of statisticians We use the data set from Ji and Jin (2016). It contains the bibtex information of all published papers in four top-tier statistics journals, Annals of Statistics, Biometrika, Journal of the American Statistical Association, and Journal of the Royal Statistical Society - Series B, from 2003 to the first half of 2012. We pre-process the abstracts of papers by tokenization and stemming and turn each abstract to a word count vector. We conduct two experiments. In the first one, we fix an author and treat the collection of his/her co-authored abstracts as a corpus. We apply DELVE+ with $K=n$, where $n$ is the total number of abstracts written by this author. The $Z$-score measures the “diversity” or “variability” of this authors’ abstracts. An author with a high $Z$-score possesses either diverse research interests or a variable writing style. A number of authors have only 1–2 papers in this data set, and the variance estimator $V$ is often negative; we remove all those authors. In Figure 3 (left panel), we plot the histogram of $Z$-scores of all retained authors. The mean is $4.52$ and the standard deviation is $2.94$. In Figure 3 (right panel), we show the scatter plot of $Z$-score versus logarithm of the number of abstracts written by this author, and a few prolific authors who have many papers and a large $Z$-score are labeled. For example, Peter Hall has the most papers in this dataset (82 papers in total). Hall’s $Z$-score is larger than $20$, implying a huge diversity in his abstracts. There is also a positive association between $Z$-score and total papers. It suggests that senior authors have more diversity in their abstracts, which is as expected. Figure 4: Pairwise $Z$-score plots for Peter Hall (left) and Jianqing Fan (right). In the cell $(x,y)$, we compare the corpus of an author’s abstracts from time $x$ with the corpus of that author’s abstracts from time $y$. The heatmap shows the value of DELVE+ with $K=2$ for each cell. | Year | Title | Journal ---|---|--- 2011 | Nonparametric independence screening in sparse ultra-high-dimensional additive models | JASA 2011 | Penalized composite quasi-likelihood for ultrahigh dimensional variable selection | JRSS-B 2011 | Multiple testing via ${\rm FDR}_{L}$ for large-scale imaging data | Ann. Stat. 2012 | Vast volatility matrix estimation using high-frequency data for portfolio selection | JASA 2012 | A road to classification in high dimensional space: the regularized optimal affine discriminant | JRSS-B 2012 | Variance estimation using refitted cross-validation in ultrahigh dimensional regression | JRSS-B Year | Title | Journal ---|---|--- 2004 | Low order approximations in deconvolution and regression with errors in variables | JRSS-B 2004 | Nonparametric inference about service time distribution from indirect measurements | JRSS-B 2004 | Cross-validation and the estimation of conditional probability densities | JASA 2004 | Nonparametric confidence intervals for receiver operating characteristic curves | Biometrika 2004 | Bump hunting with non-Gaussian kernels | Ann. Stat. 2004 | Attributing a probability to the shape of a probability density | Ann. Stat. Figure 5: (Left) Jianqing Fan’s papers in the dataset of Ji and Jin (2016) from 2011 to 2012. (Right) Peter Hall’s papers in the dataset of Ji and Jin (2016) from 2004. In the second experiment, we divide the abstracts of each author into groups by publication year. We divide Peter Hall’s abstracts into 9 groups, and each group corresponds to one year. We divide Jianqing Fan’s abstracts into 6 groups, with unequal window sizes to make all groups have roughly equal numbers of abstracts. Our test can be used to detect differences between all groups, but to see more informative results, we do a pairwise comparison: for each pair of groups, we apply DELVE+ with $K=2$. This yields a pairwise plot of $Z$-scores. The plot reveals the temporal patterns of this author in abstract writing. Figure 4 shows the results for Peter Hall and Jianqing Fan. There are interesting temporal patterns. For Jianqing Fan (right panel of Figure 4), the group consisting of his 2011-2012 abstracts has comparably large $Z$-scores in the pairwise comparison with other groups. To interpret this , we gathered the titles and abstracts of all his papers in the dataset and compared the ones before/after 2011. He published six papers in these journals during 2011-2012, whose titles are listed on the left of Figure 5. We see that his papers in this period had a strong emphasis on screening and variable selection: four out of the six papers mention this subject in their titles and/or abstracts. This shows a departure from his previously published topics such as covariance estimation (a focus from 2007–2009) and semiparametric estimation (a focus before 2010). Though Jianqing Fan had previously published papers on variable selection and screening in these journals, he had never published so many in such a short time period. For Peter Hall (left panel of Figure 4), the group of 2004 abstracts have comparably large $Z$-scores in the pairwise comparison with other groups. We examined the titles and abstracts of his 6 papers published in 2004 in this data set. All of his 2004 papers, except the first one, mention bandwidth selection or smoothing parameters, and in the last 4 papers, bandwidth selection plays a central role. For instance, Bump hunting with non-Gaussian kernels, (Ann. Stat., 2004) studies the relationship between the number of modes of a kernel density estimator and its bandwidth parameter. Though Peter Hall’s 2014 papers concern many nonparametric statistics topics, we find that bandwidth selection is a theme underlying his research in these journals in 2004. ### 6.2 Amazon movie reviews Rank | Title | $Z$-Score | Total reviews ---|---|---|--- 1 | Prometheus | 34.44 | 813 2 | Expelled: No Intelligence Allowed | 34.17 | 830 3 | V for Vendetta | 32.24 | 815 4 | Sin City | 31.72 | 828 5 | No Country for Old Men | 30.57 | 819 $\vdots$ | $\vdots$ | $\vdots$ | $\vdots$ 16 | John Adams | 20.78 | 857 17 | Cars | 19.98 | 902 18 | Food, Inc. | 17.81 | 876 19 | Jeff Dunham: Arguing with Myself | 4.96 | 860 20 | Jeff Dunham: Spark of Insanity | 4.46 | 877 Figure 6: (Left) Histogram of $Z$-scores for the 500 most-reviewed movies. The mean is $19.97$ and the standard deviation is $5.07$. (Right) $Z$-scores for the top 20 most reviewed movies. Figure 7: Pairwise $Z$-scores for 3 movies. In each cell, we use DELVE+ to compare reviews associated to a pair of star ratings. For each movie, the title list the number of reviews of each rating from 1–5. We analyze Amazon reviews from the dataset Maurya (2018) that consists of 1,924,471 reviews of 143,007 visual media products (ie, DVDs, Bluray, or streams). We examine products with the largest number of reviews. Each product’s review corpus is cleaned and stemming is used to group together words with the same root. We obtain word counts for each review and a term- document matrix of a product’s review corpus. In the first experiment, we fix a movie and apply DELVE+ with $K=n$ to the corpus consisting of all reviews of this movie. In Figure 6 (left panel), we plot the histogram of $Z$-scores for the top 500 most reviewed movies. The mean is $19.97$ and the standard deviation is $5.07$. Compared with the histogram of $Z$-scores for statistics paper abstracts, there is much larger diversity in movie reviews. In Figure 6 (right panel), we list the 5 movies with the highest $Z$-scores and lowest $Z$-scores out of the 20 most reviewed movies. Each movie has more than 800 reviews, but some have surprisingly low $Z$-scores. The works by comedian Jeff Dunham have the lowest $Z$-scores, suggesting strong homogeneity among the reviews. The 2012 horror film Prometheus has the highest degree of review diversity among the 20 most reviewed movies. In the second experiment, we divide the reviews of each movie into 5 groups by star rating. We compare each pair of groups using DELVE+ with $K=2$, resulting in a pairwise $Z$-score plot. In Figure 7, we plot this for 3 popular movies. We see a variety of polarization patterns among the scores. In Harry Potter and the Deathly Hallows Part I, DELVE+ signifies that the reviews with ratings in the range 2–4 stars are all similar. We see a smooth gradation in how the 1-star reviews differ from those from 2–4 stars, and similarly for 5-star reviews versus those from 2–4 stars. Twilight Saga: Eclipse shows three clusters: 1–2 stars, 3–4 stars, and 5 star, while Night of the living dead shows two clusters: 1–2 stars and 3–5 stars. ## 7 Discussions We examine the testing for equality of PMFs of $K$ groups of high-dimensional multinomial distributions. The proposed DELVE statistic has a parameter-free limiting null that allows for computation of $Z$-scores and $p$-values on real data. DELVE achieves the optimal detection boundary over the whole range of parameters $(n,p,K,\bar{N})$, including the high-dimensional case $p\to\infty$, which is very relevant to applications in text mining. This work leads to interesting questions for future study. So far the focus is on testing, but one can also consider inference for $\rho^{2}$ from (2.2), which measures the heterogeneity among the group-wise means. Consistent variance estimation under the alternative uses a similar strategy, though we omit the calculations in this paper. Establishing asymptotic normality of DELVE under the alternative would then lead to asymptotic confidence intervals for our heterogeneity metric $\rho^{2}$. Based on the plots in Section 5, it is possible that stronger regularity conditions are needed to obtain a pivotal distribution under the alternative. As in the two-sample multinomial testing problems described in Kipnis and Donoho (2021); Kipnis (2022), such as author attribution, we may also consider an alternative where all the group means are the same except for a small set of “giveaway words”. It is interesting to develop a procedure for identifying these useful words. As discussed in Section 4.2, we may modify DELVE by using a version based on the maximum test or higher criticism. Another extension is to go beyond ‘bag-of-words’ style analysis and use different types of counts besides raw word frequencies. One option is to apply a suitably modified DELVE to the counts of multi-grams in the corpus and another is to combine words with similar meanings into a ‘superword’ and use superword counts as the basis for DELVE. To do this, we can combine words that are close together in some word embedding. We leave these interesting tasks for future work. Acknowledgments The research of T. Tony Cai was supported in part by NSF Grant DMS-2015259 and NIH grant R01-GM129781. The research of Zheng Tracy Ke was supported in part by NSF CAREER Grant DMS-1943902. Notational conventions for the appendix: We write $A\lesssim B$ (respectively, $A\gtrsim B$) if there exists an absolute constant $C>0$ such that $A\leq C\cdot B$ (respectively $A\geq C\cdot B$). If both $A\lesssim B$ and $B\lesssim A$, we write $A\asymp B$. The implicit constant $C$ may vary from line to line. For sequences $a_{t},b_{t}$ indexed by an integer $t\in\mathbb{N}$, we write $a_{t}\ll b_{t}$ if $b_{t}/a_{t}\to\infty$ as $t\to\infty$, and we write $a_{t}\gg b_{t}$ if $a_{t}/b_{t}\to\infty$ as $t\to\infty$. We also may write $a_{t}=o(b_{t})$ to denote $a_{t}\ll b_{t}$. In particular, we write $a_{t}=(1+o(1))b_{t}$ if $a_{t}/b_{t}\to 1$ as $t\to\infty$. ## Appendix A Properties of $T$ and $V$ We recall that $X_{i}\sim\mathrm{Multinomial}(N_{i},\Omega_{i}),\qquad 1\leq i\leq n.$ (A.1) For each $1\leq k\leq K$, define $\mu_{k}=\frac{1}{n_{k}\bar{N}_{k}}\sum_{i\in S_{k}}N_{i}\Omega_{i}\;\in\;\mathbb{R}^{p},\qquad\Sigma_{k}=\frac{1}{n_{k}\bar{N}_{k}}\sum_{i\in S_{k}}N_{i}\Omega_{i}\Omega_{i}^{\prime}\;\in\;\mathbb{R}^{p\times p}.$ (A.2) Moreover, let $\mu=\frac{1}{n\bar{N}}\sum_{k=1}^{K}n_{k}\bar{N}_{k}\mu_{k}=\frac{1}{n\bar{N}}\sum_{i=1}^{n}N_{i}\Omega_{i}\,,\quad\Sigma=\frac{1}{n\bar{N}}\sum_{k=1}^{n}n_{k}\bar{N}_{k}\Sigma_{k}=\frac{1}{n\bar{N}}\sum_{i}N_{i}\Omega_{i}\Omega_{i}^{\prime}$ (A.3) The DELVE test statistic is $\psi=T/\sqrt{V}$, where $T$ is as in (2.5) and $V$ is as in (2.7). As a preparation for the main proofs, in this section, we study $T$ and $V$ separately. ### A.1 The decomposition of $T$ It is well-known that a multinomial with the number of trials equal to $N$ can be equivalently written as the sum of $N$ independent multinomials each with the number of trials equal to $1$. This inspires us to introduce a set of independent, mean-zero random vectors: $\\{Z_{ir}\\}_{1\leq i\leq n,1\leq r\leq N_{i}},\qquad\mbox{with }Z_{ir}=B_{ir}-\mathbb{E}B_{ir},\;\;\mbox{and}\;\;B_{ir}\sim\mathrm{Multinomial}(1,\Omega_{i}).$ (A.4) We use them to get a decomposition of $T$ into mutually uncorrelated terms: ###### Lemma A.1. Let $\\{Z_{ir}\\}_{1\leq i\leq n,1\leq r\leq N_{i}}$ be as in (A.4). For each $Z_{ir}\in\mathbb{R}^{p}$, let $\\{Z_{ijr}\\}_{1\leq j\leq p}$ denote its $p$ coordinates. Recall that $\rho^{2}=\sum_{k=1}^{K}n_{k}\bar{N}_{k}\|\mu_{k}-\mu\|^{2}$. For $1\leq j\leq p$, define $\displaystyle U_{1j}$ $\displaystyle=$ $\displaystyle 2\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{r=1}^{N_{i}}(\mu_{kj}-\mu_{j})Z_{ijr},$ $\displaystyle U_{2j}$ $\displaystyle=$ $\displaystyle\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{1\leq r\neq s\leq N_{i}}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}\frac{N_{i}}{N_{i}-1}Z_{ijr}Z_{ijs},$ $\displaystyle U_{3j}$ $\displaystyle=$ $\displaystyle-\frac{1}{n\bar{N}}\sum_{1\leq k\neq\ell\leq K}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{r=1}^{N_{i}}\sum_{s=1}^{N_{m}}Z_{ijr}Z_{mjs},$ $\displaystyle U_{4j}$ $\displaystyle=$ $\displaystyle\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k}\\\ i\neq m\end{subarray}}\sum_{r=1}^{N_{i}}\sum_{s=1}^{N_{m}}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}Z_{ijr}Z_{mjs}.$ Then, $T=\rho^{2}+\sum_{\kappa=1}^{4}{\bf 1}_{p}^{\prime}U_{\kappa}$. Moreover, $\mathbb{E}[U_{\kappa}]={\bf 0}_{p}$ and $\mathbb{E}[U_{\kappa}U_{\zeta}^{\prime}]={\bf 0}_{p\times p}$ for $1\leq\kappa\neq\zeta\leq 4$. ### A.2 The variance of $T$ By Lemma A.1, the four terms $\\{{\bf 1}_{p}^{\prime}U_{\kappa}\\}_{1\leq\kappa\leq 4}$ are uncorrelated with each other. Therefore, $\mathrm{Var}(T)=\mathrm{Var}({\bf 1}_{p}^{\prime}U_{1})+\mathrm{Var}({\bf 1}_{p}^{\prime}U_{2})+\mathrm{Var}({\bf 1}_{p}^{\prime}U_{3})+\mathrm{Var}({\bf 1}_{p}^{\prime}U_{4}).$ It suffices to study the variance of each of these four terms. ###### Lemma A.2. Let $U_{1}$ be the same as in Lemma A.1. Define $\displaystyle\Theta_{n1}$ $\displaystyle=4\sum_{k=1}^{K}n_{k}\bar{N}_{k}\bigl{\|}\mathrm{diag}(\mu_{k})^{1/2}(\mu_{k}-\mu)\bigr{\|}^{2}$ (A.5) $\displaystyle L_{n}$ $\displaystyle=4\sum_{k=1}^{K}n_{k}\bar{N}_{k}\bigl{\|}\Sigma_{k}^{1/2}(\mu_{k}-\mu)\bigr{\|}^{2}$ (A.6) Then $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{1})=\Theta_{n1}-L_{n}$. Furthermore, if $\max_{1\leq k\leq K}\|\mu_{k}\|_{\infty}=o(1)$, then $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{1})=o(\rho^{2})$. ###### Lemma A.3. Let $U_{2}$ be the same as in Lemma A.1. Define $\displaystyle\Theta_{n2}$ $\displaystyle=2\sum_{k=1}^{K}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\sum_{i\in S_{k}}\frac{N_{i}^{3}}{N_{i}-1}\|\Omega_{i}\|^{2}$ (A.7) $\displaystyle A_{n}$ $\displaystyle=2\sum_{k=1}^{K}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\sum_{i\in S_{k}}\frac{N_{i}^{3}}{N_{i}-1}\|\Omega_{i}\|_{3}^{3}$ (A.8) Then $\Theta_{n2}-A_{n}\leq\mathrm{Var}({\bf 1}_{p}^{\prime}U_{2})\leq\Theta_{n2}.$ Furthermore, if $\displaystyle\max_{1\leq k\leq K}\big{\\{}\frac{\sum_{i\in S_{k}}N^{2}_{i}\|\Omega_{i}\|_{3}^{3}}{\sum_{i\in S_{k}}N_{i}^{2}\|\Omega_{i}\|^{2}}\bigr{\\}}=o(1),$ (A.9) then $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{2})=[1+o(1)]\cdot\Theta_{n2}$. ###### Lemma A.4. Let $U_{3}$ be the same as in Lemma A.1. Define $\displaystyle\Theta_{n3}$ $\displaystyle=\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j}N_{i}N_{m}\Omega_{ij}\Omega_{mj}$ (A.10) $\displaystyle B_{n}$ $\displaystyle=2\sum_{k\neq\ell}\frac{n_{k}n_{\ell}\bar{N}_{k}\bar{N}_{\ell}}{n^{2}\bar{N}^{2}}{\bf 1}_{p}^{\prime}(\Sigma_{k}\circ\Sigma_{\ell}){\bf 1}_{p}$ (A.11) Then $\Theta_{n3}-B_{n}\leq\mathrm{Var}({\bf 1}_{p}^{\prime}U_{3})\leq\Theta_{n3}+B_{n}.$ ###### Lemma A.5. Let $U_{4}$ be the same as in Lemma A.1. Define $\displaystyle\Theta_{n4}$ $\displaystyle=2\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k}\\\ i\neq m\end{subarray}}\sum_{j}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}N_{i}N_{m}\Omega_{ij}\Omega_{mj}.$ (A.12) $\displaystyle E_{n}$ $\displaystyle=2\sum_{k}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k},\\\ i\neq m\end{subarray}}\sum_{1\leq j,j^{\prime}\leq p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}N_{i}N_{m}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}$ (A.13) Then $\Theta_{n4}-E_{n}\leq\mathrm{Var}({\bf 1}_{p}^{\prime}U_{4})\leq\Theta_{n4}+E_{n}$ . Using Lemmas A.2-A.5, we derive regularity conditions such that the first term in $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{\kappa})$ is the dominating term. Observe that $\Theta_{n}=\Theta_{n1}+\Theta_{n2}+\Theta_{n3}+\Theta_{n4}$, where the quantity $\Theta_{n}$ is defined in (2.6). The following intermediate result is useful. ###### Lemma A.6. Suppose that (3.1) holds. Then $\displaystyle\Theta_{n2}+\Theta_{n3}+\Theta_{n4}\asymp\sum_{k}\|\mu_{k}\|^{2}.$ (A.14) Moreover, under the null hypothesis, $\Theta_{n}\asymp K\|\mu\|^{2}$. The next result is useful in proving that our variance estimator $V$ is asymptotically unbiased. ###### Lemma A.7. Suppose that (3.1) holds, and recall the definition of $\Theta_{n}$ in (2.6). Define $\displaystyle\beta_{n}=\frac{\max\bigg{\\{}\sum_{k}\sum_{i\in S_{k}}\frac{N^{2}_{i}}{n_{k}^{2}\bar{N}_{k}^{2}}\|\Omega_{i}\|_{3}^{3}\,,\,\,\sum_{k}\|\Sigma_{k}\|_{F}^{2}\bigg{\\}}}{K\|\mu\|^{2}}.$ (A.15) If $\beta_{n}=o(1)$, then under the null hypothesis, $\mathrm{Var}(T)=[1+o(1)]\cdot\Theta_{n}$. We also study the case of $K=2$ more explicitly. In the lemmas below we use the notation from Section 3.4. First we have an intermediate result analogous to Lemma A.6 that holds under weaker conditions. ###### Lemma A.8. Consider $K=2$ and suppose that $\min N_{i}\geq 2$, $\min M_{i}\geq 2$ Then $\displaystyle\Theta_{n2}+\Theta_{n3}+\Theta_{n4}\asymp\bigg{\|}\frac{m\bar{M}}{n\bar{N}+m\bar{M}}\eta+\frac{n\bar{N}}{n\bar{N}+m\bar{M}}\theta\bigg{\|}^{2}.$ Moreover, under the null hypothesis, $\Theta_{n}\asymp\|\mu\|^{2}$. The next result is a version of Lemma A.7 for the case $K=2$ that holds under weaker conditions. ###### Lemma A.9. Suppose that $\min_{i}N_{i}\geq 2$ and $\min_{i}M_{i}\geq 2$. Define $\displaystyle\beta_{n}^{(2)}=\frac{\max\bigg{\\{}\sum_{i}N_{i}^{2}\|\Omega_{i}\|^{3},\,\,\sum_{i}M_{i}^{2}\|\Gamma_{i}\|^{3}\,,\,\,\|\Sigma_{1}\|_{F}^{2}+\|\Sigma_{2}\|_{F}^{2}\bigg{\\}}}{\|\mu\|^{2}}.$ (A.16) If $\beta_{n}^{(2)}=o(1)$, then under the null hypothesis, $\mathrm{Var}(T)=[1+o(1)]\cdot\Theta_{n}$. ### A.3 The decomposition of $V$ ###### Lemma A.10. Let $\\{Z_{ir}\\}_{1\leq i\leq n,1\leq r\leq N_{i}}$ be as in (A.4). Recall that $\displaystyle V$ $\displaystyle=2\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{j=1}^{p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\biggl{[}\frac{N_{i}X_{ij}^{2}}{N_{i}-1}-\frac{N_{i}X_{ij}(N_{i}-X_{ij})}{(N_{i}-1)^{2}}\biggr{]}$ (A.17) $\displaystyle+\frac{2}{n^{2}\bar{N}^{2}}\sum_{1\leq k\neq\ell\leq K}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j=1}^{p}X_{ij}X_{mj}+2\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k},\\\ i\neq m\end{subarray}}\sum_{j=1}^{p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}X_{ij}X_{mj}.$ Define $\displaystyle\theta_{i}$ $\displaystyle=\big{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}})^{2}\frac{N_{i}^{3}}{N_{i}-1}\quad\text{for }i\in S_{k}\,\,,\quad\text{and let}\,\,$ $\displaystyle\alpha_{im}$ $\displaystyle=\begin{cases}\frac{2}{n^{2}\bar{N}^{2}}&\quad\text{ if }i\in S_{k},m\in S_{\ell},k\neq\ell\\\ 2\big{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}})^{2}&\quad\text{ if }i,m\in S_{k}\end{cases}$ If we let $\displaystyle A_{1}$ $\displaystyle=\sum_{i}\sum_{r=1}^{N_{i}}\sum_{j}\big{[}\frac{4\theta_{i}\Omega_{ij}}{N_{i}}+\sum_{m\in[n]\backslash\\{i\\}}2\alpha_{im}N_{m}\Omega_{mj}\big{]}Z_{ijr},$ (A.18) $\displaystyle A_{2}$ $\displaystyle=\sum_{i}\sum_{r\neq s\in[N_{i}]}\frac{2\theta_{i}}{N_{i}(N_{i}-1)}\big{(}\sum_{j}Z_{ijr}Z_{ijs}\big{)}$ (A.19) $\displaystyle A_{3}$ $\displaystyle=\sum_{i\neq m}\sum_{r=1}^{N_{i}}\sum_{s=1}^{N_{m}}\alpha_{im}\big{(}\sum_{j}Z_{ijr}Z_{mjs}\big{)},$ (A.20) then these terms are mean zero, are mutually uncorrelated, and satisfy $\displaystyle V=A_{1}+A_{2}+A_{3}+\Theta_{n2}+\Theta_{n3}+\Theta_{n4}.$ (A.21) ### A.4 Properties of $V$ First we control the variance of $V$. ###### Lemma A.11. Let $A_{1},A_{2},$ and $A_{3}$ be defined as in Lemma A.10. Then $\displaystyle\mathrm{Var}(A_{1})$ $\displaystyle\lesssim\frac{1}{n\bar{N}}\|\mu\|_{3}^{3}+\sum_{k}\frac{\|\mu_{k}\|_{3}^{3}}{n_{k}\bar{N}_{k}}\lesssim\sum_{k}\frac{\|\mu_{k}\|_{3}^{3}}{n_{k}\bar{N}_{k}}$ $\displaystyle\mathrm{Var}(A_{2})$ $\displaystyle\lesssim\sum_{k}\sum_{i\in S_{k}}\frac{N_{i}^{2}\|\Omega_{i}\|_{2}^{2}}{n_{k}^{4}\bar{N}_{k}^{4}}\lesssim\sum_{k}\frac{\|\mu_{k}\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}$ $\displaystyle\mathrm{Var}(A_{3})$ $\displaystyle\lesssim\sum_{k}\frac{\|\mu_{k}\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}+\frac{1}{n^{2}\bar{N}^{2}}\|\mu\|^{2}\lesssim\sum_{k}\frac{\|\mu_{k}\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}.$ Next we show consistency of $V$ under the null, which is crucial in properly standardizing our test statistic and establishing asymptotic normality. ###### Proposition A.1. Recall the definition of $\beta_{n}$ in (A.15). Suppose that $\beta_{n}=o(1)$ and that the condition (3.1) holds. If under the null hypothesis we have $\displaystyle K^{2}\|\mu\|^{4}$ $\displaystyle\gg\sum_{k}\frac{\|\mu\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}\vee\sum_{k}\frac{\|\mu\|_{3}^{3}}{n_{k}\bar{N}_{k}},$ (A.22) then $V/\mathrm{Var}{T}\to 1$ in probability. To later control the type II error, we must also show that $V$ does not dominate the true variance under the alternative. We first state an intermediate result that is useful throughout. ###### Lemma A.12. Suppose that, under either the null or alternative, $\max_{i}\|\Omega_{i}\|_{\infty}\leq 1-c_{0}$ holds for an absolute constant $c_{0}>0$. Then $\displaystyle\mathrm{Var}(T)\gtrsim\Theta_{n2}+\Theta_{n3}+\Theta_{n4}.$ (A.23) ###### Proposition A.2. Suppose that under the alternative (3.1) holds and $\displaystyle\big{(}\sum_{k}\|\mu_{k}\|^{2}\big{)}^{2}$ $\displaystyle\gg\sum_{k}\frac{\|\mu_{k}\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}\vee\sum_{k}\frac{\|\mu_{k}\|_{3}^{3}}{n_{k}\bar{N}_{k}}.$ (A.24) Then $V=O_{\mathbb{P}}(\mathrm{Var}(T))$ under the alternative. We also require versions of Proposition A.1 and Proposition A.2 that hold under weaker conditions in the special case $K=2$. We omit the proofs as they are similar. Below we use the notation of Section 3.4. ###### Proposition A.3. Suppose that $K=2$ and recall the definition of $\beta_{n}^{(2)}$ in A.16. Suppose that $\beta_{n}^{(2)}=o(1)$, $\min_{i}N_{i}\geq 2,\min_{i}M_{i}\geq 2$, and $\max_{i}\|\Omega_{i}\|_{\infty}\leq 1-c_{0},\max_{i}\|\Gamma_{i}\|_{\infty}\leq 1-c_{0}$. If under the null hypothesis $\displaystyle\|\mu\|^{4}\gg\max\Big{\\{}\,\big{(}\frac{\|\mu\|_{2}^{2}}{n^{2}\bar{N}^{2}}+\frac{\|\mu\|_{2}^{2}}{m^{2}\bar{M}_{2}^{2}}\big{)},\,\big{(}\frac{\|\mu\|_{3}^{3}}{n\bar{N}}+\frac{\|\mu\|_{3}^{3}}{m\bar{M}}\big{)}\Big{\\}},$ (A.25) then $V/\mathrm{Var}(T)\to 1$ in probability. Under the alternative we have the following. ###### Proposition A.4. Suppose that $K=2$, $\min_{i}N_{i}\geq 2,\min_{i}M_{i}\geq 2$, and $\max_{i}\|\Omega_{i}\|_{\infty}\leq 1-c_{0},\max_{i}\|\Gamma_{i}\|_{\infty}\leq 1-c_{0}$. If under the alternative $\displaystyle\bigg{\|}\frac{m\bar{M}}{n\bar{N}+m\bar{M}}\eta+\frac{n\bar{N}}{n\bar{N}+m\bar{M}}\theta\bigg{\|}^{4}\gg\max\Big{\\{}\,\big{(}\frac{\|\eta\|_{2}^{2}}{n^{2}\bar{N}^{2}}+\frac{\|\theta\|_{2}^{2}}{m^{2}\bar{M}_{2}^{2}}\big{)},\,\big{(}\frac{\|\eta\|_{3}^{3}}{n\bar{N}}+\frac{\|\theta\|_{3}^{3}}{m\bar{M}}\big{)}\Big{\\}},$ (A.26) then $V=O_{\mathbb{P}}(\mathrm{Var}(T))$. In the setting of $K=n$ and utilize the variance estimator $V^{*}$. The next results capture the behavior of $V^{*}$ under the null and alternative. The proofs are given later in this section. ###### Proposition A.5. Define $\displaystyle\beta_{n}^{(n)}=\frac{\sum_{i}\|\Omega_{i}\|^{3}}{n\|\mu\|^{2}}.$ (A.27) Suppose that (3.1) holds, $\beta_{n}^{(n)}=o(1)$, and $\displaystyle n^{2}\|\mu\|^{4}\gg\sum_{i}\frac{\|\mu\|^{2}}{N_{i}^{2}}\vee\sum_{i}\frac{\|\mu\|_{3}^{3}}{N_{i}}.$ (A.28) Then $V^{*}/\mathrm{Var}(T)\to 1$ in probability as $n\to\infty$. ###### Proposition A.6. Suppose that under the alternative (3.1) holds and $\displaystyle\big{(}\sum_{i}\|\Omega_{i}\|^{2}\big{)}^{2}\gg\sum_{i}\frac{\|\Omega_{i}\|^{2}}{N_{i}^{2}}\vee\sum_{i}\frac{\|\Omega_{i}\|_{3}^{3}}{N_{i}}.$ (A.29) Then $V^{*}=O_{\mathbb{P}}(\mathrm{Var}(T))$ under the alternative. ### A.5 Proof of Lemma A.1 We first show that $\mathbb{E}[U_{\kappa}]={\bf 0}_{p}$ and $\mathbb{E}[U_{\kappa}U_{\zeta}^{\prime}]={\bf 0}_{p\times p}$ for $\kappa\neq\zeta$. Note that $\\{Z_{ir}\\}_{1\leq i\leq n,1\leq r\leq N_{i}}$ are independent mean-zero random vectors. It follows that each $U_{\kappa}$ is a mean-zero random vector. We then compute $\mathbb{E}[U_{\kappa j_{1}}U_{\zeta j_{2}}]$ for $\kappa\neq\zeta$ and all $1\leq j_{1},j_{2}\leq p$. By direct calculations, $\mathbb{E}[U_{1j}U_{2j_{2}}]=2\sum_{(k,i,r,s)}\sum_{(k^{\prime},i^{\prime},r^{\prime})}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}(\mu_{k^{\prime}j}-\mu_{j})\frac{N_{i}}{N_{i}-1}\mathbb{E}[Z_{ij_{2}r}Z_{ij_{2}s}Z_{i^{\prime}j_{1}r^{\prime}}].$ If $i^{\prime}\neq i$, or if $i^{\prime}=i$ and $r^{\prime}\notin\\{r,s\\}$, then $Z_{i^{\prime}j_{1}r^{\prime}}$ is independent of $Z_{ij_{2}r}Z_{ij_{2}s}$, and it follows that $\mathbb{E}[Z_{ij_{2}r}Z_{ij_{2}s}Z_{i^{\prime}j_{1}r^{\prime}}]=0$. If $i^{\prime}=i$ and $r=r^{\prime}$, then $\mathbb{E}[Z_{ij_{2}r}Z_{ij_{2}s}Z_{i^{\prime}j_{1}r^{\prime}}]=\mathbb{E}[Z_{ij_{2}r}Z_{ij_{1}r}]\cdot\mathbb{E}[Z_{ij_{2}s}]$; since $r\neq s$, we also have $\mathbb{E}[Z_{ij_{2}r}Z_{ij_{2}s}Z_{i^{\prime}j_{1}r^{\prime}}]=0$. This proves $\mathbb{E}[U_{1j}U_{2j_{*}}]=0$. Since this holds for all $1\leq j_{1},j_{2}\leq p$, we immediately have $\mathbb{E}[U_{1}U_{2}^{\prime}]={\bf 0}_{p\times p}.$ We can similarly show that $\mathbb{E}[U_{\kappa}U_{\zeta}^{\prime}]={\bf 0}_{p\times p}$, for other $\kappa\neq\zeta$. The proof is omitted. It remains to prove the desirable decomposition of $T$. Recall that $T=\sum_{j=1}^{p}T_{j}$. Write $\rho^{2}=\sum_{j=1}^{p}\rho_{j}^{2}$, where $\rho_{j}^{2}=2\sum_{k=1}^{K}n_{k}\bar{N}_{k}(\mu_{kj}-\mu_{j})^{2}$. It suffices to show that $T_{j}=\rho_{j}^{2}+U_{1j}+U_{2j}+U_{3j}+U_{4j},\qquad\mbox{for all }1\leq j\leq p.$ (A.30) To prove (A.30), we need some preparation. Define $Y_{ij}:=\frac{X_{ij}}{N_{i}}-\Omega_{ij}=\frac{1}{N_{i}}\sum_{r=1}^{N_{i}}Z_{ijr},\qquad Q_{ij}:=Y_{ij}^{2}-\mathbb{E}Y^{2}_{ij}=Y_{ij}^{2}-\frac{\Omega_{ij}(1-\Omega_{ij})}{N_{i}}.$ (A.31) With these notations, $X_{ij}=N_{i}(\Omega_{ij}+Y_{ij})$ and $N_{i}Y_{ij}^{2}=N_{i}Q_{ij}+\Omega_{ij}(1-\Omega_{ij})$. Moreover, we can use (A.31) to re-write $Q_{ij}$ as a function of $\\{Z_{ijr}\\}_{1\leq r\leq N_{i}}$ as follows: $Q_{ij}=\frac{1}{N^{2}_{i}}\sum_{r=1}^{N_{i}}[Z_{ijr}^{2}-\Omega_{ij}(1-\Omega_{ij})]+\frac{1}{N^{2}_{i}}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}.$ Note that $Z_{ijr}=B_{ijr}-\Omega_{ij}$, where $B_{ijr}$ can only take values in $\\{0,1\\}$. Hence, $(Z_{ijr}+\Omega_{ij})^{2}=(Z_{ijr}+\Omega_{ij})$ always holds. Re-arranging the terms gives $Z^{2}_{ijr}-\Omega_{ij}(1-\Omega_{ij})=(1-2\Omega_{ij})Z_{ijr}$. It follows that $Q_{ij}=(1-2\Omega_{ij})\frac{Y_{ij}}{N_{i}}+\frac{1}{N^{2}_{i}}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}.$ (A.32) This is a useful equality which we will use in the proof below. We now show (A.30). Fix $j$ and write $T_{j}=R_{j}-D_{j}$, where $R_{j}=\sum_{k=1}^{K}n_{k}\bar{N}_{k}(\hat{\mu}_{kj}-\hat{\mu}_{j})^{2},\quad\mbox{and}\quad D_{j}=\sum_{k=1}^{K}\sum_{i\in S_{k}}\xi_{k}\frac{X_{ij}(N_{i}-X_{ij})}{n_{k}\bar{N}_{k}(N_{i}-1)},\quad\mbox{with }\xi_{k}=1-\frac{n_{k}\bar{N}_{k}}{n\bar{N}}$ First, we study $D_{j}$. Note that $X_{ij}(N_{ij}-X_{ij})=N^{2}_{i}(\Omega_{ij}+Y_{ij})(1-\Omega_{ij}-Y_{ij})=N^{2}_{i}\Omega_{ij}(1-\Omega_{ij})-N^{2}_{i}Y_{ij}^{2}+N_{i}^{2}(1-2\Omega_{ij})Y_{ij}$, where $Y_{ij}^{2}=Q_{ij}+N_{i}^{-1}\Omega_{ij}(1-\Omega_{ij})$. It follows that $\frac{X_{ij}(N_{ij}-X_{ij})}{N_{i}(N_{i}-1)}=\Omega_{ij}(1-\Omega_{ij})-\frac{N_{i}Q_{ij}}{N_{i}-1}+\frac{N_{i}}{N_{i}-1}(1-2\Omega_{ij})Y_{ij}.$ We apply (A.32) to get $\frac{X_{ij}(N_{ij}-X_{ij})}{N_{i}(N_{i}-1)}=\Omega_{ij}(1-\Omega_{ij})+(1-2\Omega_{ij})Y_{ij}-\frac{1}{N_{i}(N_{i}-1)}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}.$ (A.33) It follows that $\displaystyle D_{j}$ $\displaystyle=\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}N_{i}}{n_{k}\bar{N}_{k}}\Omega_{ij}(1-\Omega_{ij})+\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}N_{i}}{n_{k}\bar{N}_{k}}(1-2\Omega_{ij})Y_{ij}$ (A.34) $\displaystyle\qquad-\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}}{n_{k}\bar{N}_{k}(N_{i}-1)}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}.$ (A.35) Next, we study $R_{j}$. Note that $n_{k}\bar{N}_{k}(\hat{\mu}_{kj}-\hat{\mu}_{j})=\sum_{i\in S_{k}}(X_{ij}-\bar{N}_{k}\hat{\mu}_{j})$. It follows that $R_{j}=\sum_{k=1}^{K}\frac{1}{n_{k}\bar{N}_{k}}\biggl{[}\sum_{i\in S_{k}}(X_{ij}-\bar{N}_{k}\hat{\mu}_{j})\biggr{]}^{2}.$ Recall that $X_{ij}=N_{i}(\Omega_{ij}+Y_{ij})$. By direct calculations, $\sum_{i\in S_{k}}X_{ij}=n_{k}\bar{N}_{k}\mu_{kj}+\sum_{i\in S_{k}}N_{i}Y_{ij}$, and $\hat{\mu}_{j}=\mu_{j}+(n\bar{N})^{-1}\sum_{m=1}^{n}N_{m}Y_{mj}$. We then have the following decomposition: $\displaystyle\sum_{i\in S_{k}}(X_{ij}-\bar{N}_{k}\hat{\mu}_{j})$ $\displaystyle=n_{k}\bar{N}_{k}(\mu_{kj}-\mu_{j})+\sum_{i\in S_{k}}N_{i}Y_{ij}-\frac{n_{k}\bar{N}_{k}}{n\bar{N}}\Bigl{(}\sum_{m=1}^{n}N_{m}Y_{mj}\Bigr{)}.$ Using this decomposition, we can expand $[\sum_{i\in S_{k}}(X_{ij}-\bar{N}_{k}\hat{\mu}_{j})]^{2}$ to a total of 6 terms, where 3 are quadratic terms and 3 are cross terms. It yields a decomposition of $R_{j}$ into 6 terms: $\displaystyle R_{j}$ $\displaystyle=\sum_{k=1}^{K}n_{k}\bar{N}_{k}(\mu_{kj}-\mu_{j})^{2}+\sum_{k=1}^{K}\frac{1}{n_{k}\bar{N}_{k}}\Bigl{(}\sum_{i\in S_{k}}N_{i}Y_{ij}\Bigr{)}^{2}+\sum_{k=1}^{K}\frac{n_{k}\bar{N}_{k}}{n^{2}\bar{N}^{2}}\Bigl{(}\sum_{m=1}^{n}N_{m}Y_{mj}\Bigr{)}^{2}$ (A.36) $\displaystyle\qquad+2\sum_{k=1}^{K}(\mu_{kj}-\mu_{j})\Bigl{(}\sum_{i\in S_{k}}N_{i}Y_{ij}\Bigr{)}-2\sum_{k=1}^{K}\frac{n_{k}\bar{N}_{k}}{n\bar{N}}(\mu_{kj}-\mu_{j})\Bigl{(}\sum_{m=1}^{n}N_{m}Y_{mj}\Bigr{)}$ (A.37) $\displaystyle\qquad-\frac{2}{n\bar{N}}\sum_{k=1}^{K}\Bigl{(}\sum_{i\in S_{k}}N_{i}Y_{ij}\Bigr{)}\Bigl{(}\sum_{m=1}^{n}N_{m}Y_{mj}\Bigr{)}$ (A.38) $\displaystyle\equiv I_{1}+I_{2}+I_{3}+I_{4}+I_{5}+I_{6}.$ (A.39) By definition, $\sum_{k=1}^{K}n_{k}\bar{N}_{k}=n\bar{N}$ and $\sum_{k=1}^{K}n_{k}\bar{N}_{k}\mu_{kj}=n\bar{N}\mu_{j}$. It follows that $I_{3}=\frac{1}{n\bar{N}}\Bigl{(}\sum_{m=1}^{n}N_{m}Y_{mj}\Bigr{)}^{2},\qquad I_{5}=0,\qquad I_{6}=-\frac{2}{n\bar{N}}\Bigl{(}\sum_{m=1}^{n}N_{m}Y_{mj}\Bigr{)}^{2}=-2I_{3}.$ It follows that $R_{j}=I_{1}+I_{2}-I_{3}+I_{4}.$ (A.40) We further simplify $I_{3}$. Recall that $\xi_{k}=1-(n\bar{N})^{-1}n_{k}\bar{N}_{k}$. By direct calculations, $\displaystyle I_{3}$ $\displaystyle=\frac{1}{n\bar{N}}\Bigl{(}\sum_{m=1}^{n}N_{m}Y_{mj}\Bigr{)}^{2}=\frac{1}{n\bar{N}}\biggl{[}\sum_{k=1}^{K}\Bigl{(}\sum_{i\in S_{k}}N_{i}Y_{ij}\Bigr{)}\biggr{]}^{2}$ (A.41) $\displaystyle=\frac{1}{n\bar{N}}\sum_{k=1}^{K}\Bigl{(}\sum_{i\in S_{k}}N_{i}Y_{ij}\Bigr{)}^{2}+\frac{1}{n\bar{N}}\sum_{1\leq k\neq\ell\leq K}\Bigl{(}\sum_{i\in S_{k}}N_{i}Y_{ij}\Bigr{)}\Bigl{(}\sum_{m\in S_{\ell}}N_{m}Y_{mj}\Bigr{)}$ (A.42) $\displaystyle=\sum_{k=1}^{K}(1-\xi_{k})\frac{1}{n_{k}\bar{N}_{k}}\Bigl{(}\sum_{i\in S_{k}}N_{i}Y_{ij}\Bigr{)}^{2}+\underbrace{\frac{1}{n\bar{N}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}N_{i}N_{m}Y_{ij}Y_{mj}}_{J_{1}}$ (A.43) $\displaystyle=I_{2}-\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}}{n_{k}\bar{N}_{k}}\Bigl{(}\sum_{i\in S_{k}}N_{i}Y_{ij}\Bigr{)}^{2}+J_{1}$ (A.44) $\displaystyle=I_{2}+J_{1}-\sum_{k=1}^{K}\frac{\xi_{k}}{n_{k}\bar{N}_{k}}\Bigl{(}\sum_{i\in S_{k}}N_{i}^{2}Y_{ij}^{2}\Bigr{)}-\underbrace{\sum_{k=1}^{K}\frac{\xi_{k}}{n_{k}\bar{N}_{k}}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k}\\\ i\neq m\end{subarray}}N_{i}N_{m}Y_{ij}Y_{mj}}_{J_{2}}.$ (A.45) By (A.31), $N_{i}Y_{ij}^{2}=N_{i}Q_{i}+\Omega_{ij}(1-\Omega_{ij})$. We further apply (A.32) to get $N_{i}^{2}Y_{ij}^{2}=N_{i}(1-2\Omega_{ij})Y_{ij}+\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}+N_{i}\Omega_{ij}(1-\Omega_{ij}).$ It follows that $\displaystyle\sum_{k=1}^{K}\frac{\xi_{k}}{n_{k}\bar{N}_{k}}$ $\displaystyle\Bigl{(}\sum_{i\in S_{k}}N_{i}^{2}Y_{ij}^{2}\Bigr{)}=\underbrace{\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}N_{i}}{n_{k}\bar{N}_{k}}(1-2\Omega_{ij})Y_{ij}}_{J_{3}}$ (A.46) $\displaystyle+\underbrace{\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}}{n_{k}\bar{N}_{k}}\sum_{r\neq s}Z_{ijr}Z_{ijs}}_{J_{4}}+\underbrace{\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}N_{i}}{n_{k}\bar{N}_{k}}\Omega_{ij}(1-\Omega_{ij})}_{J_{5}}.$ (A.47) We plug (A.46) into (A.41) to get $I_{3}=I_{2}+J_{1}-J_{2}-J_{3}-J_{4}-J_{5}$. Further plugging $I_{3}$ into the expression of $R_{j}$ in (A.40), we have $\displaystyle R_{j}$ $\displaystyle=I_{1}+I_{4}-J_{1}+J_{2}+J_{3}+J_{4}+J_{5},$ (A.48) where $I_{1}$ and $I_{4}$ are defined in (A.36), $J_{1}$-$J_{2}$ are defined in (A.41), and $J_{3}$-$J_{5}$ are defined in (A.46). Finally, we combine the expressions of $D_{j}$ and $R_{j}$. By (A.34) and the definitions of $J_{1}$-$J_{5}$, $\displaystyle D_{j}$ $\displaystyle=J_{5}+J_{3}-\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}}{n_{k}\bar{N}_{k}(N_{i}-1)}\sum_{r\neq s}Z_{ijr}Z_{ijs}$ $\displaystyle=J_{5}+J_{3}+J_{4}-\underbrace{\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}N_{i}}{n_{k}\bar{N}_{k}(N_{i}-1)}\sum_{r\neq s}Z_{ijr}Z_{ijs}}_{J_{6}}.$ Combining it with (A.48) gives $T_{j}=R_{j}-D_{j}=I_{1}+I_{4}-J_{1}+J_{2}+J_{6}$. We further plug in the definition of each term. It follows that $\displaystyle T_{j}$ $\displaystyle=\sum_{k=1}^{K}n_{k}\bar{N}_{k}(\mu_{kj}-\mu_{j})^{2}+2\sum_{k=1}^{K}\sum_{i\in S_{k}}(\mu_{kj}-\mu_{j})N_{i}Y_{ij}-\frac{1}{n\bar{N}}\sum_{k\neq\ell}\sum_{i\in S_{k},m\in S_{\ell}}N_{i}N_{m}Y_{ij}Y_{mj}$ (A.49) $\displaystyle\qquad+\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k}\\\ i\neq m\end{subarray}}\frac{\xi_{k}}{n_{k}\bar{N}_{k}}N_{i}N_{m}Y_{ij}Y_{mj}+\sum_{k=1}^{K}\sum_{i\in S_{k}}\frac{\xi_{k}N_{i}}{n_{k}\bar{N}_{k}(N_{i}-1)}\sum_{r\neq s}Z_{ijr}Z_{ijs}.$ (A.50) We plug in $Y_{ij}=N_{i}^{-1}\sum_{r=1}^{N_{i}}Z_{ijr}$ and take a sum of $1\leq j\leq p$. It gives (A.30) immediately. The proof is now complete. ∎ ### A.6 Proof of Lemma A.2 Recall that $\\{Z_{ir}\\}_{1\leq i\leq n,1\leq r\leq N_{i}}$ are independent random vectors. Write ${\bf 1}_{p}^{\prime}U_{1}=2\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{r=1}^{N_{i}}(\mu_{k}-\mu)^{\prime}Z_{ir}.$ The covariance matrix of $Z_{ir}$ is $\mathrm{diag}(\Omega_{i})-\Omega_{i}\Omega_{i}^{\prime}$. It follows that $\displaystyle\mathrm{Var}({\bf 1}_{p}^{\prime}U_{1})$ $\displaystyle=4\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{r=1}^{N_{i}}(\mu_{k}-\mu)^{\prime}\bigl{[}\mathrm{diag}(\Omega_{i})-\Omega_{i}\Omega_{i}^{\prime}\bigr{]}(\mu_{k}-\mu)$ (A.51) $\displaystyle=4\sum_{k}(\mu_{k}-\mu)^{\prime}\Bigl{[}\mathrm{diag}\Bigl{(}\sum_{i\in S_{k}}N_{i}\Omega_{i}\Bigr{)}-\Bigl{(}\sum_{i\in S_{k}}N_{i}\Omega_{i}\Omega_{i}^{\prime}\Bigr{)}\Bigr{]}(\mu_{k}-\mu)$ (A.52) $\displaystyle=4\sum_{k}(\mu_{k}-\mu)^{\prime}\Bigl{[}\mathrm{diag}(n_{k}\bar{N}_{k}\mu_{k})-n_{k}\bar{N}_{k}\Sigma_{k}\Bigr{]}(\mu_{k}-\mu)$ (A.53) $\displaystyle=4\sum_{k}n_{k}\bar{N}_{k}\bigl{\|}\mathrm{diag}(\mu_{k})^{1/2}(\mu_{k}-\mu)\bigr{\|}^{2}-4\sum_{k}n_{k}\bar{N}_{k}\bigl{\|}\Sigma_{k}^{1/2}(\mu_{k}-\mu)\bigr{\|}^{2}.$ (A.54) This proves the first claim. Furthermore, by (A.51), $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{1})\leq 4\sum_{k}n_{k}\bar{N}_{k}\bigl{\|}\mathrm{diag}(\mu_{k})^{1/2}(\mu_{k}-\mu)\bigr{\|}^{2}\leq 4\sum_{k}n_{k}\bar{N}_{k}\|\mathrm{diag}(\mu_{k})\|\|\mu_{k}-\mu\|^{2}.$ Note that $\|\mathrm{diag}(\mu_{k})\|=\|\mu_{k}\|_{\infty}$. Therefore, if $\max_{k}\|\mu_{k}\|_{\infty}=o(1)$, the right hand side above is $o(1)\cdot 4\sum_{k}n_{k}\bar{N}_{k}\|\mu_{k}-\mu\|^{2}=o(\rho^{2})$. This proves the second claim. ∎ ### A.7 Proof of Lemma A.3 For each $1\leq k\leq K$, define a set of index triplets: ${\cal M}_{k}=\\{(i,r,s):i\in S_{k},1\leq r<s\leq N_{i}\\}$. Let ${\cal M}=\cup_{k=1}^{K}{\cal M}_{k}$. Write for short $\theta_{i}=(\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}})^{2}\frac{N_{i}^{3}}{N_{i}-1}$, for $i\in S_{k}$. It is seen that ${\bf 1}_{p}^{\prime}U_{2}=2\sum_{(i,r,s)\in{\cal M}}\frac{\sqrt{\theta_{i}}}{\sqrt{N_{i}(N_{i}-1)}}W_{irs},\qquad\mbox{with}\quad W_{irs}=\sum_{j=1}^{p}Z_{ijr}Z_{ijs}.$ For $W_{irs}$ and $W_{i^{\prime}r^{\prime}s^{\prime}}$, if $i\neq i^{\prime}$, or if $i=i^{\prime}$ and $\\{r,s\\}\cap\\{r^{\prime},s^{\prime}\\}=\emptyset$, then these two variables are independent; if $i=i^{\prime}$, $r=r^{\prime}$ and $s\neq s^{\prime}$, then $\mathbb{E}[W_{irs}W_{irs^{\prime}}]=\sum_{j,j^{\prime}}\mathbb{E}[Z_{ijr}Z_{ijs}Z_{ij^{\prime}r}Z_{ij^{\prime}s^{\prime}}]=\sum_{j,j^{\prime}}\mathbb{E}[Z_{ijr}Z_{ij^{\prime}r}]\cdot\mathbb{E}[Z_{ijs}]\cdot\mathbb{E}[Z_{ij^{\prime}s^{\prime}}]=0$. Therefore, $\\{W_{irs}\\}_{(i,r,s)\in{\cal M}}$ is a collection of mutually uncorrelated variables. It follows that $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{2})=4\sum_{(i,r,s)\in{\cal M}}\frac{\theta_{i}}{N_{i}(N_{i}-1)}\mathrm{Var}(W_{irs}).$ It remains to calculate the variance of each $W_{irs}$. By direction calculations, $\displaystyle\mathrm{Var}(W_{irs})$ $\displaystyle=\sum_{j}\mathbb{E}[Z_{ijr}^{2}Z_{ijs}^{2}]+2\sum_{j<\ell}\mathbb{E}[Z_{ijr}Z_{ijs}Z_{i\ell r}Z_{i\ell s}]$ (A.55) $\displaystyle=\sum_{j}[\Omega_{ij}(1-\Omega_{ij})]^{2}+2\sum_{j<\ell}(-\Omega_{ij}\Omega_{i\ell})^{2}$ (A.56) $\displaystyle=\sum_{j}\Omega_{ij}^{2}-2\sum_{j}\Omega^{3}_{ij}+\Bigl{(}\sum_{j}\Omega^{2}_{ij}\Bigr{)}^{2}$ (A.57) $\displaystyle=\|\Omega_{i}\|^{2}-2\|\Omega_{i}\|_{3}^{3}+\|\Omega_{i}\|^{4}$ (A.58) Since $\max_{ij}\Omega_{ij}\leq 1$, we have $\displaystyle\|\Omega_{i}\|^{2}-\|\Omega_{i}\|_{3}^{3}\leq\mathrm{Var}(W_{irs})\leq\|\Omega_{i}\|^{2}.$ Therefore, $\displaystyle\mathrm{Var}({\bf 1}_{p}^{\prime}U_{2})$ $\displaystyle=4\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{1\leq r<s\leq N_{i}}\frac{\theta_{i}}{N_{i}(N_{i}-1)}\mathrm{Var}(W_{irs})$ $\displaystyle=2\sum_{k=1}^{K}\sum_{i\in S_{k}}\theta_{i}\mathrm{Var}(W_{irs})\geq 2\sum_{k=1}^{K}\sum_{i\in S_{k}}\theta_{i}\big{[}\|\Omega_{i}\|^{2}-\|\Omega_{i}\|_{3}^{3}\big{]}=\Theta_{n2}-A_{n},$ and similarly $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{2})\leq\Theta_{n2}$, which proves the first claim. To prove the second claim, note that $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{2})=\Theta_{n2}+O(A_{n})$. By (A.9) and the assumption $\min N_{i}\geq 2$, we have $\displaystyle A_{n}$ $\displaystyle\lesssim\sum_{k}\big{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\big{)}^{2}\sum_{i\in S_{k}}N_{i}^{2}\|\Omega_{i}\|_{3}^{3}$ $\displaystyle=\sum_{k}\big{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\big{)}^{2}\cdot o\bigg{(}\sum_{i\in S_{k}}N_{i}^{2}\|\Omega_{i}\|^{2}\bigg{)}=o(\Theta_{n2}),$ which implies that $\mathrm{Var}({\bf 1}_{p}U_{2})=[1+o(1)]\Theta_{n2}$, as desired. ∎ ### A.8 Proof of Lemma A.4 For each $1\leq k<\ell\leq K$, define a set of index quadruples: ${\cal J}_{k\ell}=\\{(i,r,m,s):i\in S_{k},j\in S_{\ell},1\leq r\leq N_{i},1\leq s\leq N_{m}\\}$. Let ${\cal J}=\cup_{(k,\ell):1\leq k<\ell\leq K}{\cal J}_{k\ell}$. It is seen that ${\bf 1}_{p}^{\prime}U_{3}=-\frac{2}{n\bar{N}}\sum_{(i,r,m,s)\in{\cal J}}V_{irms},\qquad\mbox{where}\;\;V_{irms}=\sum_{j=1}^{p}Z_{ijr}Z_{mjs}.$ For $V_{irms}$ and $V_{i^{\prime}r^{\prime}m^{\prime}s^{\prime}}$, if $\\{(i,r),(m,s)\\}\cap\\{(i^{\prime},r^{\prime}),(m^{\prime},s^{\prime})\\}=\emptyset$, then the two variables are independent of each other. If $(i,r)=(i^{\prime},r^{\prime})$ and $(m,s)\neq(m^{\prime},s^{\prime})$, then $\mathbb{E}[V_{irms}V_{irm^{\prime}s^{\prime}}]=\sum_{j,j^{\prime}}\mathbb{E}[Z_{ijr}Z_{mjs}Z_{ij^{\prime}r}Z_{m^{\prime}j^{\prime}s^{\prime}}]=\sum_{j,j^{\prime}}\mathbb{E}[Z_{ijr}Z_{ij^{\prime}r}]\cdot\mathbb{E}[Z_{mjs}]\cdot\mathbb{E}[Z_{m^{\prime}js^{\prime}}]=0$. Therefore, the only correlated case is when $(i,r,m,s)=(i^{\prime},r^{\prime},m^{\prime},s^{\prime})$. This implies that $\\{V_{irms}\\}_{(i,r,m,s)\in{\cal J}}$ is a collection of mutually uncorrelated variables. Therefore, $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{3})=\frac{4}{n^{2}\bar{N}^{2}}\sum_{(i,r,m,s)\in{\cal J}}\mathrm{Var}(V_{irms}).$ Note that $\mathrm{Var}(V_{irms})=\mathbb{E}[(\sum_{j}Z_{ijr}Z_{mjs})^{2}]=\sum_{j,j^{\prime}}\mathbb{E}[Z_{ijr}Z_{mjs}Z_{ij^{\prime}r}Z_{mj^{\prime}s}]$; also, the covariance matrix of $Z_{ir}$ is $\mathrm{diag}(\Omega_{i})-\Omega_{i}\Omega_{i}^{\prime}$. It follows that $\displaystyle\mathrm{Var}(V_{irms})$ $\displaystyle=\sum_{j}\mathbb{E}[Z^{2}_{ijr}]\cdot\mathbb{E}[Z^{2}_{mjs}]+\sum_{j\neq j^{\prime}}\mathbb{E}[Z_{ijr}Z_{ij^{\prime}r}]\cdot\mathbb{E}[Z_{mjs}Z_{mj^{\prime}s}]$ (A.60) $\displaystyle=\sum_{j}\Omega_{ij}(1-\Omega_{ij})\Omega_{mj}(1-\Omega_{mj})+\sum_{j\neq j^{\prime}}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}$ (A.61) $\displaystyle=\sum_{j}\Omega_{ij}\Omega_{mj}-2\sum_{j}\Omega^{2}_{ij}\Omega^{2}_{mj}+\sum_{j,j^{\prime}}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}.$ (A.62) Write for short $\delta_{im}=-2\sum_{j}\Omega^{2}_{ij}\Omega^{2}_{mj}+\sum_{j,j^{\prime}}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}$.Combining the above gives $\displaystyle\mathrm{Var}$ $\displaystyle({\bf 1}_{p}^{\prime}U_{3})=\frac{4}{n^{2}\bar{N}^{2}}\sum_{k<\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{r=1}^{N_{i}}\sum_{s=1}^{N_{m}}\Bigl{(}\sum_{j}\Omega_{ij}\Omega_{mj}+\delta_{im}\Bigr{)}$ (A.63) $\displaystyle=\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j}N_{i}N_{m}\Omega_{ij}\Omega_{mj}+\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}N_{i}N_{m}\delta_{im}.$ (A.64) It is easy to see that $|\delta_{im}|\leq\sum_{j,j^{\prime}}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}$. Also, by the definition of $\Sigma_{k}$ in (A.2), we have $\Sigma_{k}(j,j^{\prime})=\frac{1}{n_{k}\bar{N}_{k}}\sum_{i\in S_{k}}N_{i}\Omega_{ij}\Omega_{ij^{\prime}}$. Using these results, we immediately have $\displaystyle\Bigl{|}\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}N_{i}N_{m}\delta_{im}\Bigr{|}$ $\displaystyle\leq\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j,j^{\prime}}N_{i}N_{m}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}$ (A.65) $\displaystyle=\frac{2}{n^{2}\bar{N}^{2}}\sum_{j,j^{\prime}}\sum_{k\neq\ell}\Bigl{(}\sum_{i\in S_{k}}N_{i}\Omega_{ij}\Omega_{ij^{\prime}}\Bigr{)}\Bigl{(}\sum_{m\in S_{\ell}}N_{i}\Omega_{mj}\Omega_{mj^{\prime}}\Bigr{)}$ (A.66) $\displaystyle=\frac{2}{n^{2}\bar{N}^{2}}\sum_{j,j^{\prime}}\sum_{k\neq\ell}n_{k}\bar{N}_{k}\Sigma_{k}(j,j^{\prime})\cdot n_{\ell}\bar{N}_{\ell}\Sigma_{\ell}(j,j^{\prime})$ (A.67) $\displaystyle=2\sum_{k\neq\ell}\frac{n_{k}n_{\ell}\bar{N}_{k}\bar{N}_{\ell}}{n^{2}\bar{N}^{2}}{\bf 1}_{p}^{\prime}(\Sigma_{k}\circ\Sigma_{\ell}){\bf 1}_{p}=:B_{n}$ (A.68) as desired. ∎ ### A.9 Proof of Lemma A.5 For $1\leq k\leq K$, define a set of index quadruples: ${\cal Q}_{k}=\\{(i,r,m,s):i\in S_{k},m\in S_{k},i<m,1\leq r\leq N_{i},1\leq s\leq N_{m}\\}$. Let ${\cal Q}=\cup_{k=1}^{K}{\cal Q}_{k}$. Write $\kappa_{im}=(\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}})^{2}N_{i}N_{m}$, for $i\in S_{k}$ and $m\in S_{k}$. It is seen that ${\bf 1}_{p}^{\prime}U_{4}=2\sum_{(i,r,m,s)\in{\cal Q}}\frac{\sqrt{\kappa_{im}}}{\sqrt{N_{i}N_{m}}}V_{irms},\qquad\mbox{where}\quad V_{irms}=\sum_{j=1}^{p}Z_{ijr}Z_{mjs}.$ It is not hard to see that $V_{irms}$ and $V_{i^{\prime}r^{\prime}m^{\prime}s^{\prime}}$ are correlated only if $(i,r,m,s)=(i^{\prime},r^{\prime},m^{\prime},s^{\prime})$. It follows that $\mathrm{Var}({\bf 1}_{p}^{\prime}U_{4})=4\sum_{(i,r,m,s)\in{\cal Q}}\frac{\kappa_{im}}{N_{i}N_{m}}\mathrm{Var}(V_{irms}).$ In the proof of Lemma A.4, we have studied $\mathrm{Var}(V_{irms})$. In particular, by (A.60), we have $\mathrm{Var}(V_{irms})=\sum_{j}\Omega_{ij}\Omega_{mj}+\delta_{im},\qquad\mbox{with}\quad|\delta_{im}|\leq\sum_{j,j^{\prime}}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}.$ Thus $\displaystyle\mathrm{Var}({\bf 1}_{p}^{\prime}U_{4})$ $\displaystyle=4\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k}\\\ i<m\end{subarray}}\sum_{i=1}^{N_{i}}\sum_{r=1}^{N_{m}}\frac{\kappa_{im}}{N_{i}N_{m}}\mathrm{Var}(V_{irms})$ $\displaystyle=4\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k}\\\ i<m\end{subarray}}\kappa_{im}\Bigl{(}\sum_{j}\Omega_{ij}\Omega_{mj}+\delta_{im}\Bigr{)}$ $\displaystyle=2\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k}\\\ i\neq m\end{subarray}}\sum_{j}\kappa_{im}\Omega_{ij}\Omega_{mj}\pm 2\sum_{k}\sum_{i\neq m\in S_{k}}\kappa_{im}\sum_{j,j^{\prime}}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}},$ $\displaystyle=\Theta_{n3}\pm E_{n}.$ (A.69) which proves the lemma. ∎ ### A.10 Proof of Lemma A.6 By assumption (3.1), $N_{i}^{3}/(N_{i}-1)\asymp N_{i}$ and $\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\asymp\frac{1}{n_{k}^{2}\bar{N}_{k}^{2}}$. First, observe that $\displaystyle\Theta_{n2}+\Theta_{n4}$ $\displaystyle=2\sum_{k=1}^{K}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\sum_{i\in S_{k}}\frac{N_{i}^{3}}{N_{i}-1}\|\Omega_{i}\|^{2}$ $\displaystyle\quad+2\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k}\\\ i\neq m\end{subarray}}\sum_{j}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}N_{i}N_{m}\Omega_{ij}\Omega_{mj}$ $\displaystyle\asymp\sum_{k=1}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}\Bigr{)}^{2}\sum_{j}\sum_{i,m\in S_{k}}N_{i}\Omega_{ij}\cdot N_{m}\Omega_{ij}=\sum_{k}\|\mu_{k}\|^{2}.$ (A.70) Recall the definitions of $\mu_{k}$ and $\mu$ in (A.2)-(A.3). By direct calculations, we have $\displaystyle\Theta_{n3}$ $\displaystyle=2\sum_{j}\sum_{k\neq\ell}\Bigl{(}\frac{1}{n\bar{N}}\sum_{i\in S_{k}}N_{i}\Omega_{ij}\Bigr{)}\Bigl{(}\frac{1}{n\bar{N}}\sum_{m\in S_{\ell}}N_{m}\Omega_{mj}\Bigr{)}$ $\displaystyle=2\sum_{j}\sum_{k\neq\ell}\frac{n_{k}\bar{N}_{k}}{n\bar{N}}\mu_{kj}\cdot\frac{n_{\ell}\bar{N}_{\ell}}{n\bar{N}}\mu_{\ell j}$ $\displaystyle=2\sum_{k\neq\ell}\frac{n_{k}n_{\ell}\bar{N}_{k}\bar{N}_{\ell}}{n^{2}\bar{N}^{2}}\cdot\mu_{k}^{\,\,\prime}\,\mu_{\ell}$ $\displaystyle\leq 2\sum_{j}\Bigl{(}\sum_{k}\frac{n_{k}\bar{N}_{k}}{n\bar{N}}\mu_{kj}\Bigr{)}^{2}=2\sum_{j}\mu_{j}^{2}=2\|\mu\|^{2}.$ (A.71) By Cauchy–Schwarz, $\displaystyle\|\mu\|^{2}$ $\displaystyle=\sum_{j}\bigg{(}\sum_{k}(\frac{n_{k}\bar{N_{k}}}{n\bar{N}})\mu_{kj}\bigg{)}^{2}$ $\displaystyle\leq\sum_{j}\bigg{(}\sum_{k}(\frac{n_{k}\bar{N_{k}}}{n\bar{N}})^{2}\bigg{)}\cdot\bigg{(}\sum_{k}\mu_{kj}^{2}\bigg{)}$ $\displaystyle\leq\sum_{j}\bigg{(}\sum_{k}(\frac{n_{k}\bar{N_{k}}}{n\bar{N}})\bigg{)}\cdot\bigg{(}\sum_{k}\mu_{kj}^{2}\bigg{)}=\sum_{j}\sum_{k}\mu_{kj}^{2}=\sum_{k}\|\mu_{k}\|^{2}.$ (A.72) Combining (A.70), (A.71), and (A.72) yields $\displaystyle c\big{(}\sum_{k}\|\mu_{k}\|^{2}\big{)}\leq\Theta_{n2}+\Theta_{n3}+\Theta_{n4}\leq C\big{(}\sum_{k}\|\mu_{k}\|^{2}\big{)},$ for absolute constants $c,C>0$. This completes the proof. ∎ ### A.11 Proof of Lemma A.7 By (3.1), it holds that $\displaystyle(\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}})^{2}\asymp\frac{1}{(n_{k}\bar{N}_{k})^{2}},$ (A.73) and moreover, for all $i\in\\{1,2,\ldots,n\\}$, $\displaystyle\frac{N_{i}^{3}}{N_{i}-1}\asymp N_{i}^{2}.$ (A.74) Recall the definitions of $A_{n},$ $B_{n}$, and $E_{n}$ in (A.8), (A.11), and (A.13), respectively. Note that these are the remainder terms in Lemmas A.3, A.4, and A.5, respectively. Under the null hypothesis (recall $\Theta_{n1}\equiv 0$ under the null), $\displaystyle\mathrm{Var}(T)=\Theta_{n2}+\Theta_{n3}+\Theta_{n4}+O(A_{n}+B_{n}+E_{n}).$ (A.75) It holds that $\displaystyle A_{n}\leq\sum_{k=1}^{K}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}\Bigr{)}^{2}\sum_{i\in S_{k}}N_{i}^{2}\|\Omega_{i}\|_{3}^{3}.$ (A.76) Next, by linearity and the definition of $\Sigma_{k},\Sigma$ in (A.2), (A.3), respectively, $\displaystyle B_{n}$ $\displaystyle\leq 2\sum_{k,\ell}\frac{n_{k}n_{\ell}\bar{N}_{k}\bar{N}_{\ell}}{n^{2}\bar{N}^{2}}{\bf 1}_{p}^{\prime}(\Sigma_{k}\circ\Sigma_{\ell}){\bf 1}_{p}$ $\displaystyle\leq 2{\bf 1}_{p}^{\prime}\bigg{(}\frac{1}{n\bar{N}}\sum_{k}n_{k}\bar{N}_{k}\Sigma_{k}\bigg{)}\circ\bigg{(}\frac{1}{n\bar{N}}\sum_{\ell}n_{\ell}\bar{N}_{\ell}\Sigma_{k}\ell\bigg{)}{\bf 1}_{p}$ $\displaystyle=2{\bf 1}_{p}^{\prime}(\Sigma\circ\Sigma){\bf 1}_{p}=2\|\Sigma\|_{F}^{2}$ By Cauchy–Schwarz, $\displaystyle B_{n}$ $\displaystyle\leq\|\Sigma\|_{F}^{2}=\sum_{j,j^{\prime}}\bigg{(}\sum_{k}(\frac{n_{k}\bar{N}_{k}}{n\bar{N}}\Sigma_{k}(j,j^{\prime})\bigg{)}^{2}$ $\displaystyle\leq\sum_{j,j^{\prime}}\bigg{(}\sum_{k}(\frac{n_{k}\bar{N}_{k}}{n\bar{N}})^{2}\bigg{)}\cdot\bigg{(}\sum_{k}\Sigma_{k}(j,j^{\prime})^{2}\bigg{)}$ $\displaystyle\leq\sum_{j,j^{\prime}}\bigg{(}\sum_{k}\frac{n_{k}\bar{N}_{k}}{n\bar{N}}\bigg{)}\cdot\bigg{(}\sum_{k}\Sigma_{k}(j,j^{\prime})^{2}\bigg{)}=\sum_{j,j^{\prime}}\sum_{k}\Sigma_{k}(j,j^{\prime})^{2}=\sum_{k}\|\Sigma_{k}\|_{F}^{2}.$ (A.77) Next by the definition of $\Sigma_{k}$ in (A.2), we have $\Sigma_{k}(j,j^{\prime})=\frac{1}{n_{k}\bar{N}_{k}}\sum_{i\in S_{k}}N_{i}\Omega_{ij}\Omega_{ij^{\prime}}$. It follows that $\displaystyle E_{n}$ $\displaystyle\leq\sum_{k}\sum_{j,j^{\prime}}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}\sum_{i\in S_{k}}N_{i}\Omega_{ij}\Omega_{ij^{\prime}}\Bigr{)}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}\sum_{m\in S_{k}}N_{m}\Omega_{mj}\Omega_{mj^{\prime}}\Bigr{)}$ (A.78) $\displaystyle=\sum_{k}\sum_{j,j^{\prime}}\Sigma_{k}^{2}(j,j^{\prime})=\sum_{k}\|\Sigma_{k}\|_{F}^{2}.$ (A.79) Next, Lemma A.6 implies that $\displaystyle\Theta_{n2}+\Theta_{n3}+\Theta_{n4}\asymp\sum_{k}\|\mu_{k}\|^{2}=K\|\mu\|^{2},$ (A.80) where we use that the null hypothesis holds. By assumption of the lemma, we have $\beta_{n}=\frac{\max\bigg{\\{}\sum_{k}\sum_{i\in S_{k}}\frac{N^{2}_{i}}{n_{k}^{2}\bar{N}_{k}^{2}}\|\Omega_{i}\|_{3}^{3}\,,\,\,\sum_{k}\|\Sigma_{k}\|_{F}^{2}\bigg{\\}}}{K\|\mu\|^{2}}=o(1)$ Combining this with (A.75), (A.76), (A.77), (A.78),and (A.80) completes the proof of the first claim. The second claim follows plugging in $\mu_{k}=\mu$ for all $k\in\\{1,2,\ldots,K\\}$. ∎ ### A.12 Proof of Lemma A.8 By assumption, $N_{i}^{3}/(N_{i}-1)\asymp N_{i},M_{i}^{3}/(M_{i}-1)\asymp M_{i}$. By direct calculation, $\displaystyle\Theta_{n2}+\Theta_{n4}$ $\displaystyle\asymp\big{[}\frac{m\bar{M}}{(n\bar{N}+m\bar{M})n\bar{N}}\big{]}^{2}\sum_{i,m,j}N_{i}N_{m}\Omega_{ij}\Omega_{mj}+\big{[}\frac{n\bar{N}}{(n\bar{N}+m\bar{M})m\bar{M}}\big{]}^{2}\sum_{i,m}N_{i}N_{m}\Gamma_{ij}\Gamma_{mj}$ $\displaystyle=\frac{1}{(n\bar{N}+m\bar{M})^{2}}\bigg{(}(m\bar{M})^{2}\|\eta\|^{2}+n\bar{N}^{2}\|\theta\|^{2}\bigg{)}.$ (A.81) Next $\displaystyle\Theta_{n3}$ $\displaystyle=\frac{4}{(n\bar{N}+m\bar{M})^{2}}\sum_{i\in S_{1}}\sum_{m\in S_{2}}\sum_{j}N_{i}\Omega_{ij}\cdot N_{m}\Gamma_{mj}$ $\displaystyle=\frac{4}{(n\bar{N}+m\bar{M})^{2}}\cdot n\bar{N}m\bar{M}\langle\theta,\eta\rangle.$ (A.82) Combining (A.81) and (A.82) yields $\displaystyle\Theta_{n2}+\Theta_{n3}+\Theta_{n4}$ $\displaystyle\asymp\frac{1}{(n\bar{N}+m\bar{M})^{2}}\big{(}(m\bar{M})^{2}\|\eta\|^{2}+2n\bar{N}m\bar{M}\langle\theta,\eta\rangle+n\bar{N}^{2}\|\theta\|^{2}\big{)}$ $\displaystyle=\bigg{\|}\frac{m\bar{M}}{n\bar{N}+m\bar{M}}\eta+\frac{n\bar{N}}{n\bar{N}+m\bar{M}}\theta\bigg{\|}^{2},$ which proves the first claim. The second follows by plugging in $\theta=\eta=\mu$ under the null. ∎ ### A.13 Proof of Lemma A.9 As in (A.75), we have under the null that $\displaystyle\mathrm{Var}(T)=\Theta_{n2}+\Theta_{n3}+\Theta_{n4}+O(A_{n}+B_{n}+E_{n}).$ (A.83) For general $K$, observe that the proofs of the bounds $\displaystyle A_{n}$ $\displaystyle\leq\sum_{k=1}^{K}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}\Bigr{)}^{2}\sum_{i\in S_{k}}N_{i}^{2}\|\Omega_{i}\|_{3}^{3}$ $\displaystyle B_{n}$ $\displaystyle\leq\sum_{k=1}^{K}\|\Sigma_{k}\|_{F}^{2}$ $\displaystyle E_{n}$ $\displaystyle\leq\sum_{k=1}^{K}\|\Sigma_{k}\|_{F}^{2}$ derived in (A.76), (A.77), and (A.78), only use the assumption that $N_{i},M_{i}\geq 2$ for all $i$. Translating these bounds to the notation of the $K=2$ case, we have $\displaystyle A_{n}$ $\displaystyle\leq\sum_{i}N_{i}^{2}\|\Omega_{i}\|^{3}+\sum_{i}M_{i}^{2}\|\Gamma_{i}\|^{3}$ $\displaystyle B_{n}$ $\displaystyle\leq\|\Sigma_{1}\|_{F}^{2}+\|\Sigma_{2}\|_{F}^{2}$ $\displaystyle E_{n}$ $\displaystyle\leq\|\Sigma_{1}\|_{F}^{2}+\|\Sigma_{2}\|_{F}^{2}.$ (A.84) Furthermore, we know that $\Theta_{n}\geq c\|\mu\|^{2}$ under the null by Lemma A.8, for an absolute constant $c>0$. Combining this with (A.83) and (A.84) completes the proof. ∎ ### A.14 Proof of Lemma A.10 Define $\displaystyle V_{1}$ $\displaystyle=2\sum_{k=1}^{K}\sum_{i\in S_{k}}\sum_{j=1}^{p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}\biggl{[}\frac{N_{i}X_{ij}^{2}}{N_{i}-1}-\frac{N_{i}X_{ij}(N_{i}-X_{ij})}{(N_{i}-1)^{2}}\biggr{]}$ $\displaystyle V_{2}$ $\displaystyle=\frac{2}{n^{2}\bar{N}^{2}}\sum_{1\leq k\neq\ell\leq K}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j=1}^{p}X_{ij}X_{mj}$ $\displaystyle V_{3}$ $\displaystyle=2\sum_{k=1}^{K}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k},\\\ i\neq m\end{subarray}}\sum_{j=1}^{p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}X_{ij}X_{mj}.$ Observe that $V_{1}+V_{2}+V_{3}=V$. Also define $\displaystyle A_{11}$ $\displaystyle=\sum_{i}\sum_{r=1}^{N_{i}}\sum_{j}\big{[}\frac{4\theta_{i}\Omega_{ij}}{N_{i}}\big{]}Z_{ijr}$ (A.85) $\displaystyle A_{12}$ $\displaystyle=2\sum_{i}\sum_{r=1}^{N_{i}}\sum_{j}\big{[}\sum_{m\in[n]\backslash\\{i\\}}\alpha_{im}N_{m}\Omega_{mj}\big{]}Z_{ijr}$ (A.86) and observe that $A_{11}+A_{12}=A_{1}$. First, we derive the decomposition of $V_{1}$. Recall that $Y_{ij}:=\frac{X_{ij}}{N_{i}}-\Omega_{ij}=\frac{1}{N_{i}}\sum_{r=1}^{N_{i}}Z_{ijr},\qquad Q_{ij}:=Y_{ij}^{2}-\mathbb{E}Y^{2}_{ij}=Y_{ij}^{2}-\frac{\Omega_{ij}(1-\Omega_{ij})}{N_{i}}.$ (A.87) With these notations, $X_{ij}=N_{i}(\Omega_{ij}+Y_{ij})$ and $N_{i}Y_{ij}^{2}=N_{i}Q_{ij}+\Omega_{ij}(1-\Omega_{ij})$. Write $V_{1}=2\sum_{i=1}^{n}\sum_{i=1}^{n}\frac{\theta_{i}}{N_{i}}\Delta_{ij},\qquad\mbox{where}\quad\Delta_{ij}:=\frac{X^{2}_{ij}}{N_{i}}-\frac{X_{ij}(N_{i}-X_{ij})}{N_{i}(N_{i}-1)}.$ (A.88) Note that $X_{ij}=N_{i}(\Omega_{ij}+Y_{ij})$ and $Y_{ij}^{2}=Q_{ij}+N_{i}^{-1}\Omega_{ij}(1-\Omega_{ij})$. It follows that $\frac{X_{ij}^{2}}{N_{i}}=N_{i}\Omega_{ij}^{2}+2N_{i}\Omega_{ij}Y_{ij}+N_{i}Q_{ij}+\Omega_{ij}(1-\Omega_{ij}).$ In (A.32), we have shown that $Q_{ij}=(1-2\Omega_{ij})\frac{Y_{ij}}{N_{i}}+\frac{1}{N^{2}_{i}}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}$. It follows that $\frac{X_{ij}^{2}}{N_{i}}=N_{i}\Omega_{ij}^{2}+2N_{i}\Omega_{ij}Y_{ij}+(1-2\Omega_{ij})Y_{ij}+\frac{1}{N_{i}}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}+\Omega_{ij}(1-\Omega_{ij}).$ Additionally, by (A.33), $\frac{X_{ij}(N_{ij}-X_{ij})}{N_{i}(N_{i}-1)}=\Omega_{ij}(1-\Omega_{ij})+(1-2\Omega_{ij})Y_{ij}-\frac{1}{N_{i}(N_{i}-1)}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}.$ Combining the above gives $\displaystyle\Delta_{ij}=N_{i}\Omega_{ij}^{2}+2N_{i}\Omega_{ij}Y_{ij}+\frac{1}{N_{i}-1}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}$ (A.89) $\displaystyle=N_{i}\Omega_{ij}^{2}+2\Omega_{ij}\sum_{r=1}^{N_{i}}Z_{ijr}+\frac{1}{N_{i}-1}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}.$ (A.90) Recall the definition of $\Theta_{n2}$ in (A.7), $A_{2}$ in (A.19), and $A_{11}$ in (A.85). We have $\displaystyle V_{1}$ $\displaystyle=2\sum_{k,i\in S_{k}}\sum_{j}\frac{\theta_{i}}{N_{i}}\big{[}N_{i}\Omega_{ij}^{2}+2\Omega_{ij}\sum_{r=1}^{N_{i}}Z_{ijr}+\frac{1}{N_{i}-1}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}\big{]}.$ $\displaystyle=\Theta_{n2}+\sum_{k,i\in S_{k}}\sum_{j}\frac{4\theta_{i}\Omega_{ij}}{N_{i}}\sum_{r=1}^{N_{i}}Z_{ijr}+\sum_{k,i\in S_{k}}\sum_{j}\frac{2\theta_{i}}{N_{i}(N_{i}-1)}\sum_{1\leq r\neq s\leq N_{i}}Z_{ijr}Z_{ijs}$ $\displaystyle=\Theta_{n2}+A_{11}+A_{2}$ (A.91) Next, we have $\displaystyle V_{2}+V_{3}$ $\displaystyle=\sum_{i\neq m}\alpha_{im}N_{i}N_{m}\sum_{j}\bigg{[}(Y_{ij}+\Omega_{ij})(Y_{mj}+\Omega_{mj})\bigg{]}$ $\displaystyle=\sum_{i\neq m}\alpha_{im}N_{i}N_{m}\sum_{j}Y_{ij}Y_{mj}+2\sum_{i\neq m}\alpha_{im}N_{i}N_{m}\sum_{j}Y_{ij}\Omega_{mj}+\sum_{i\neq m}\alpha_{im}N_{i}N_{m}\sum_{j}\Omega_{ij}\Omega_{mj}$ $\displaystyle=\sum_{i\neq m}\sum_{r=1}^{N_{i}}\sum_{s=1}^{N_{m}}\alpha_{im}\big{(}\sum_{j}Z_{ijr}Z_{mjs}\big{)}+2\sum_{i}\sum_{r=1}^{N_{i}}\sum_{j}\big{[}\sum_{m\in[n]\backslash\\{i\\}}\alpha_{im}N_{m}\Omega_{mj}\big{]}Z_{ijr}+\Theta_{n3}+\Theta_{n4}$ $\displaystyle=A_{3}+A_{12}+\Theta_{n3}+\Theta_{n4}.$ Hence $\displaystyle A_{1}+A_{2}+A_{3}+\Theta_{n2}+\Theta_{n3}+\Theta_{n4}=V,$ which verifies (A.21). By inspection, we also see that $\mathbb{E}A_{b}=0$ for $b\in\\{1,2,3\\}$. That $A_{1},A_{2},A_{3}$ are mutually uncorrelated follows immediately from the linearity of expectation and the fact that the random variables $\\{Z_{ijr}\\}_{i,r}\cup\\{Z_{ijr}Z_{mjs}\\}_{(i,r)\neq(m,s)}$ are mutually uncorrelated. ∎ ### A.15 Proof of Lemma A.11 Define $\displaystyle\gamma_{irj}=\frac{4\theta_{i}\Omega_{ij}}{N_{i}}+\sum_{m\in[n]\backslash\\{i\\}}2\alpha_{im}N_{m}\Omega_{mj}$ (A.92) and recall that $A_{1}=\sum_{i}\sum_{r\in[N_{i}]}\sum_{j}\gamma_{irj}Z_{ijr}$. First we develop a bound on $\gamma_{irj}$. Suppose that $i\in S_{k}$. Then we have $\displaystyle\gamma_{irj}$ $\displaystyle\lesssim\frac{N_{i}\Omega_{ij}}{n_{k}^{2}\bar{N}_{k}^{2}}+\sum_{m\in S_{k},m\neq i}\frac{N_{m}\Omega_{mj}}{n_{k}^{2}\bar{N}_{k}^{2}}+\sum_{k^{\prime}\in[K]\backslash\\{k\\}}\sum_{m\in S_{k^{\prime}}}\frac{N_{m}\Omega_{mj}}{n^{2}\bar{N}^{2}}$ $\displaystyle\lesssim\frac{\mu_{kj}}{n_{k}\bar{N}_{k}}+\frac{\mu_{j}}{n\bar{N}}.$ Next using properties of the covariance matrix of a multinomial vector, we have $\displaystyle\mathrm{Var}(A_{1})$ $\displaystyle=\sum_{i,r\in[N_{i}]}\mathrm{Var}(\gamma_{ir:}^{\prime}Z_{i:r})=\sum_{i,r\in[N_{i}]}\gamma_{ir:}^{\prime}\text{Cov}(Z_{i:r})\gamma_{ir:}$ $\displaystyle\leq\sum_{i,r\in[N_{i}]}\gamma_{ir:}^{\prime}\text{diag}(\Omega_{i:})\gamma_{ir:}=\sum_{i,r\in[N_{i}]}\sum_{j}\Omega_{ij}\gamma_{irj}^{2}$ $\displaystyle\lesssim\sum_{k,j}\big{(}\frac{\mu_{kj}}{n_{k}\bar{N}_{k}}+\frac{\mu_{j}}{n\bar{N}}\big{)}^{2}\sum_{i\in S_{k},r\in[N_{i}]}\Omega_{ij}$ $\displaystyle\lesssim\sum_{k,j}\big{(}\frac{\mu_{kj}}{n_{k}\bar{N}_{k}}\big{)}^{2}n_{k}\bar{N}_{k}\mu_{kj}+\sum_{k,j}\big{(}\frac{\mu_{j}}{n\bar{N}}\big{)}^{2}n_{k}\bar{N}_{k}\mu_{kj}$ $\displaystyle=(\sum_{k}\frac{\|\mu_{k}\|_{3}^{3}}{n_{k}\bar{N}_{k}})+\frac{\|\mu\|_{3}^{3}}{n\bar{N}}\lesssim\sum_{k}\frac{\|\mu_{k}\|_{3}^{3}}{n_{k}\bar{N}_{k}},$ (A.93) which proves the first claim. The last inequality follows because by Jensen’s inequality (noting that the function $x\mapsto x^{3}$ is convex for $x\geq 0$), $\displaystyle\|\mu\|_{3}^{3}$ $\displaystyle=\sum_{j}\bigg{(}\sum_{k}(\frac{n_{k}\bar{N}_{k}}{n\bar{N}})\mu_{kj}\bigg{)}^{3}\leq\sum_{j}\sum_{k}(\frac{n_{k}\bar{N}_{k}}{n\bar{N}})\mu_{kj}^{3}\leq\sum_{k}\|\mu_{k}\|_{3}^{3}.$ Next observe that $\displaystyle A_{2}=\sum_{i}\sum_{r\neq s}\frac{2\theta_{i}}{N_{i}(N_{i}-1)}W_{irs}$ (A.94) where recall $W_{irs}=\sum_{j}Z_{ijr}Z_{ijs}$. Also recall that $W_{irs}$ and $W_{i^{\prime}r^{\prime}s^{\prime}}$ are uncorrelated unless $i=i^{\prime}$ and $\\{r,s\\}=\\{r^{\prime},s^{\prime}\\}$. By (A.55), $\displaystyle\mathrm{Var}(A_{2})$ $\displaystyle=\sum_{i}\sum_{r\neq s}\frac{4\theta_{i}^{2}}{N_{i}^{2}(N_{i}-1)^{2}}\mathrm{Var}(W_{irs})$ $\displaystyle\lesssim\sum_{i}\sum_{r\neq s}\frac{4\theta_{i}^{2}}{N_{i}^{2}(N_{i}-1)^{2}}\|\Omega_{i}\|^{2}$ $\displaystyle\lesssim\sum_{k}\sum_{i\in S_{k}}\cdot(\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}})^{4}\frac{N_{i}^{6}}{(N_{i}-1)^{2}}\cdot\frac{1}{N_{i}(N_{i}-1)}\|\Omega_{i}\|^{2}$ $\displaystyle\lesssim\sum_{k}\sum_{i\in S_{k}}\frac{N_{i}^{2}}{n_{k}^{4}\bar{N}_{k}^{4}}\|\Omega_{i}\|^{2}$ (A.95) Also observe that $\displaystyle\sum_{k}\frac{1}{n_{k}^{4}\bar{N}_{k}^{4}}\sum_{i\in S_{k}}N_{i}^{2}\|\Omega_{i}\|_{2}^{2}$ $\displaystyle\leq\sum_{k}\frac{1}{n_{k}^{2}\bar{N}_{k}^{2}}\sum_{i,m\in S_{k}}\bigg{\langle}(\frac{N_{i}}{n_{k}\bar{N}_{k}})\Omega_{i},(\frac{N_{m}}{n_{m}\bar{N}_{m}})\Omega_{m}\bigg{\rangle}$ $\displaystyle=\sum_{k}\frac{1}{n_{k}^{2}\bar{N}_{k}^{2}}\|\mu_{k}\|^{2}.$ This establishes the second claim. Last we study $A_{3}$. Observe that $\displaystyle A_{3}=\sum_{i\neq m}\sum_{r=1}^{N_{i}}\sum_{s=1}^{N_{m}}\alpha_{im}V_{irms}$ where recall $V_{irms}=\sum_{j}Z_{ijr}Z_{mjs}$. Recall that $V_{irms}$ and $V_{i^{\prime}r^{\prime}m^{\prime}s^{\prime}}$ are uncorrelated unless $(r,s)=(r^{\prime},s^{\prime})$ and $\\{i,m\\}=\\{i^{\prime},m^{\prime}\\}$ .By (A.60), $\displaystyle\mathrm{Var}(A_{3})$ $\displaystyle\lesssim\sum_{i\neq m}\alpha_{im}^{2}N_{i}N_{m}\sum_{j}\Omega_{ij}\Omega_{mj}$ $\displaystyle\lesssim\sum_{k}\sum_{i\neq m\in S_{k}}\frac{1}{n_{k}^{4}\bar{N}_{k}^{4}}\langle N_{i}\Omega_{i},N_{m}\Omega_{m}\rangle+\sum_{k\neq\ell}\sum_{i\in S_{k},m\in S_{\ell}}\frac{1}{n^{4}\bar{N}^{4}}\langle N_{i}\Omega_{i},N_{m}\Omega_{m}\rangle$ $\displaystyle\lesssim\sum_{k}\frac{\|\mu_{k}\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}+\sum_{k,\ell}\frac{1}{n^{4}\bar{N}^{4}}\langle n_{k}\bar{N}_{k}\mu_{k},n_{\ell}\bar{N}_{\ell}\mu_{\ell}\rangle$ $\displaystyle\lesssim\sum_{k}\frac{\|\mu_{k}\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}+\frac{\|\mu\|^{2}}{n^{2}\bar{N}^{2}}\lesssim\sum_{k}\frac{\|\mu_{k}\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}.$ (A.96) In the last line we use that $\|\mu\|^{2}\leq 2\sum\|\mu_{k}\|^{2}$ as shown in (A.72). This proves all required claims. ∎ ### A.16 Proof of Proposition A.1 Under the null hypothesis, we have $\Theta_{n1}\equiv 0$. Thus, $\mathbb{E}V=\Theta_{n}$ under the null by Lemma A.10. Under (3.1), we have $\mathrm{Var}(T)=[1+o(1)]\Theta_{n}$. Therefore, $\displaystyle\mathbb{E}V=[1+o(1)]\mathrm{Var}(T),$ (A.97) so $V$ is asymptotically unbiased under the null. Furthermore, by Lemma A.6, we have $\displaystyle\Theta_{n}\asymp K\|\mu\|^{2}.$ (A.98) In Lemma A.11, we showed that $\displaystyle\mathrm{Var}(A_{2})$ $\displaystyle\lesssim\sum_{k}\sum_{i\in S_{k}}\frac{N_{i}^{2}\|\Omega_{i}\|_{2}^{2}}{n_{k}^{4}\bar{N}_{k}^{4}}$ We conclude by Lemma A.11 that under the null $\displaystyle\mathrm{Var}(V)\lesssim\sum_{k}\frac{\|\mu\|^{2}}{n_{k}^{2}\bar{N}_{k}^{2}}\vee\sum_{k}\frac{\|\mu\|_{3}^{3}}{n_{k}\bar{N}_{k}}.$ (A.99) By Chebyshev’s inequality, (A.98), (A.99), and assumption (A.22) of the theorem statement, we have $\displaystyle\frac{|V-\mathbb{E}V|}{\mathrm{Var}(T)}\asymp\frac{|V-\mathbb{E}V|}{K\|\mu\|^{2}}=o_{\mathbb{P}}(1).$ Thus by (A.97), $\displaystyle\frac{V}{\mathrm{Var}(T)}$ $\displaystyle=\frac{(V-\mathbb{E}V)}{\mathrm{Var}(T)}+\frac{\mathbb{E}V}{\mathrm{Var}(T)}=o_{\mathbb{P}}(1)+[1+o(1)],$ as desired. ∎ ### A.17 Proof of Lemma A.12 By Lemmas A.1–A.5, we have $\displaystyle\mathrm{Var}(T)$ $\displaystyle=\sum_{a=1}^{4}\mathrm{Var}(\mathbf{1}_{p}^{\prime}U_{a})\geq(\sum_{a=2}^{4}\Theta_{na})-(A_{n}+B_{n}+E_{n}).$ (A.100) Using that $\max_{i}\|\Omega_{i}\|_{\infty}\leq 1-c_{0}$, we have $\|\Omega_{i}\|^{3}\leq(1-c_{0})\|\Omega_{i}\|^{2}$, which implies that $\displaystyle A_{n}\leq(1-c_{0})\Theta_{n2}.$ (A.101) Again using $\max_{i}\|\Omega_{i}\|_{\infty}\leq 1-c_{0}$, as well as $\sum_{j^{\prime}}\Omega_{ij^{\prime}}=1$, we have $\displaystyle B_{n}$ $\displaystyle=\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j,j^{\prime}}N_{i}N_{m}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}$ $\displaystyle\leq(1-c_{0})\cdot\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j,j^{\prime}}N_{i}N_{m}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}$ $\displaystyle=(1-c_{0})\cdot\frac{2}{n^{2}\bar{N}^{2}}\sum_{k\neq\ell}\sum_{i\in S_{k}}\sum_{m\in S_{\ell}}\sum_{j}N_{i}N_{m}\Omega_{ij}\Omega_{mj}$ $\displaystyle\leq(1-c_{0})\cdot\Theta_{n3}.$ (A.102) Similarly to control $E_{n}$, we again use $\max_{i}\|\Omega_{i}\|_{\infty}\leq 1-c_{0}$ and obtain $\displaystyle E_{n}$ $\displaystyle=2\sum_{k}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k},\\\ i\neq m\end{subarray}}\sum_{1\leq j,j^{\prime}\leq p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}N_{i}N_{m}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}\Omega_{mj^{\prime}}$ $\displaystyle\leq(1-c_{0})\cdot 2\sum_{k}\sum_{\begin{subarray}{c}i\in S_{k},m\in S_{k},\\\ i\neq m\end{subarray}}\sum_{1\leq j,j^{\prime}\leq p}\Bigl{(}\frac{1}{n_{k}\bar{N}_{k}}-\frac{1}{n\bar{N}}\Bigr{)}^{2}N_{i}N_{m}\Omega_{ij}\Omega_{ij^{\prime}}\Omega_{mj}$
# Mitigating Inappropriateness in Image Generation: Can there be Value in Reflecting the World’s Ugliness? Manuel Brack Felix Friedrich Patrick Schramowski Kristian Kersting ###### Abstract Text-conditioned image generation models have recently achieved astonishing results in image quality and text alignment and are consequently employed in a fast-growing number of applications. Since they are highly data-driven, relying on billion-sized datasets randomly scraped from the web, they also reproduce inappropriate human behavior. Specifically, we demonstrate inappropriate degeneration on a large-scale for various generative text-to- image models, thus motivating the need for monitoring and moderating them at deployment. To this end, we evaluate mitigation strategies at inference to suppress the generation of inappropriate content. Our findings show that we can use models’ representations of the world’s ugliness to align them with human preferences. Machine Learning, ICML Warning: This paper contains sexually explicit imagery, discussions of pornography, and other content that some readers may find disturbing, distressing, and/or offensive. ## 1 Introduction Next to text-generative models such as ChatGPT, image-generative models are becoming increasingly prevalent and seeing growing adoption in commercial services such as stockimagery and graphic design. Due to their large-scale unsupervised learning they retain general knowledge implicitly present in the data and are able to generate high fidelity images faithful interpretations to the users’ prompts. However, their learning setup, which includes large-scale unfiltered data (schuhmann2022laion; birhane2021multimodal), also leads to degenerated and biased behavior (schramowski2022safe), calling for mitigation strategies and the moderation of generative models in deployed systems. Consequently, before the deployment of image-generative models, it is crucial to not only validate their quality but also ensure their safety. This necessitates the assessment of appropriate guardrails, which should be tailored to the specific application at hand. Previous work in this domain has primarily relied on anecdotal evidence, lacking quantifiable measures that take multiple models and architectures into account. Indeed, schramowski2022safe proposed an empirical benchmark but limited their evaluation to a single Stable Diffusion version. To help the development of effective mitigation strategies and moderation techniques for image-generative models in real-world systems, we here present a comprehensive assessment of inappropriate degeneration across numerous open- source models and architectures. More precisely, we investigate how effectively these models can be instructed to suppress inappropriate content using the knowledge obtained about the world’s ugliness. Our findings suggest that safety mitigation of text-to-image generators can be performed through direct instructions at inference for various types of models. In total, we generated and evaluated over 1.5M images for 11 different models, thereby providing a large-scale investigation of the topic. Figure 1: Examples of inappropriate degeneration and their mitigation across various models. From left to right each batch shows the original image and the instructed ones with Sega and negative prompting. Prompts are taken from the inappropriate-image-prompts (I2P) dataset. Images displaying nudity were blurred by the authors. (Best viewed in color) ## 2 Instructing Models on the World’s Ugliness Visual Moderation. There exist multiple approaches for mitigating inappropriate degeneration of generative models. Previous research has identified four major methods. The first approach involves filtering the training data to remove problematic content entirely (nichol2022glide). However, large-scale dataset filtering can have unexpected side effects on downstream performance as demonstrated by nichol2022glide. Moreover, determining what constitutes inappropriate content is highly subjective and dependent on various external factors such as individual and societal norms as well as the specific use case of the application. Developing a dedicated model with data filtering tailored to each definition of inappropriateness is difficult, if not impractical, particularly as it would require retraining pre-existing models from scratch. To overcome this limitation, a second approach involves finetuning a pre-trained model to erase inappropriate concepts (gandikota2023erasing). While this method requires lower computational resources compared to training an entire model, it is still constrained in its ability to account for diverse definitions of inappropriateness. Another relevant approach, particularly for deployed applications, involves implementing input and output filters111https://www.technologyreview.com/2023/02/24/1069093/. In hosted inference services, input prompts are typically filtered for banned keywords, and the generated images are scanned for inappropriate content before being presented to users. Although this approach restricts the availability of unwanted content, it has some drawbacks. schramowski2022safe have demonstrated that inappropriate degeneration can occur unexpectedly for prompts lacking explicit descriptions of any problematic concepts. Therefore, input filters are prone to missing these implicit correlations. Additionally, the generation and subsequent discarding of images not only wastes computational resources but can also result in a frustrating user experience. In contrast, we here explore the idea of leveraging a model’s learned representations of inappropriate content for mitigation of such material. We focus on explicit instruction approaches that provide textual descriptions to the model regarding concepts to avoid during the image generation process. This results in both high flexibility and customizability, as the instruction prompt can be easily modified to adapt to different requirements. Consequently, the user remains involved in the process and the method enables seamless deployment across various architectures. As such they also facilitate large-scale evaluation across models. Classifier Free Guidance. Before going into detail on different instruction methods for image generation, we need to establish some fundamentals of text- to-image diffusion models (DMs). Intuitively, image generation starts from random noise $\epsilon$, and the model predicts an estimate of this noise $\tilde{\epsilon}_{\theta}$ to be subtracted from the initial values. This results in a high-fidelity image $x$ without any noise. Since this is a complex problem, multiple steps are applied, each subtracting a small amount ($\epsilon_{t}$) of the predictive noise, approximating $\epsilon$. For text- to-image generation, the model’s $\epsilon$-prediction is conditioned on a text prompt $p$ and results in an image faithful to that prompt. To that end, DMs employ classifier-free guidance (ho2022classifier), a conditioning method using a purely generational diffusion model, eliminating the need for an additional pre-trained classifier. The noise estimate $\tilde{\epsilon}_{\theta}$ uses an unconditioned $\epsilon$-prediction $\mathbf{\epsilon}_{\theta}(\mathbf{z}_{t})$ which is pushed in the direction of the conditioned estimate $\mathbf{\epsilon}_{\theta}(\mathbf{z}_{t},\mathbf{c}_{p})$ to yield an image faithful to prompt $p$. Instructing Text-to-Image Models on Safety. We now consider two different instruction approaches extending the principles of classifier-free guidance. Both methods rely on a secondary text prompt $s$ that describes concepts to suppress during generation. First, negative prompting replaces the unconditioned $\epsilon$-prediction $\mathbf{\epsilon}_{\theta}(\mathbf{z}_{t})$ with one conditioned on $s$: $\mathbf{\epsilon}_{\theta}(\mathbf{z}_{t},\mathbf{c}_{s})$, thus moving away from the inappropriate concepts. This approach is intuitive and easy to implement, however offers limited control over the extent of content suppression. Additionally, we use Semantic Guidance (Sega) (brack2023Sega) which is a powerful method for image manipulation based on additional text prompts. Sega adds an additional guidance term to $\tilde{\epsilon}_{\theta}$ that allows us to steer the generation away from $s$, while keeping changes to the original image minimal. Table 1: Text-to-image models are prone to generate inappropriate content. Instruction methods can considerably reduce the chance of producing such material (the lower, the better). Shown are the probabilities of generating an image containing inappropriate content as classified by the combined Q16/NudeNet classifier over the I2P benchmark. We note that the Q16 classifier is rather conservative and tends to classify some unobjectionable images as inappropriate. The expected maximum inappropriateness (the lower, the better) are bootstrap estimates of a model outputting the displayed percentage of inappropriate images at least once for 25 prompts. Subscript values indicate the standard deviation. Models evaluated are Stable Diffusion (SD) (rombach2022High) and fine-tuned variants, AltDiffusion (chen2022altclip), MultiFusion (bellagente2023multifusion), Paella (rampas2023novel) and IF by Deepfloyd33footnotemark: 3. | Base Model | w/ SEGA | w/ Neg. Prompt ---|---|---|--- | Sexual | All Categories | Sexual | All Categories | Sexual | All Categories Model | Prob | Exp. Max | Prob | Exp. Max | Prob | Exp. Max | Prob | Exp. Max | Prob | Exp. Max | Prob | Exp. Max SD 1.4 | | | | | | | | | | | |
# Learning behavioral context recognition with multi-stream temporal convolutional networks Aaqib Saeed, Tanir Ozcelebi, *Stojan Trajanovski, Johan Lukkien <EMAIL_ADDRESS><EMAIL_ADDRESS><EMAIL_ADDRESS><EMAIL_ADDRESS> Department of Mathematics and Computer Science, Eindhoven University of Technology, Eindhoven, The Netherlands *Philips Research, Eindhoven, The Netherlands ###### Abstract Smart devices of everyday use (such as smartphones and wearables) are increasingly integrated with sensors that provide immense amounts of information about a person’s daily life such as behavior and context. The automatic and unobtrusive sensing of behavioral context can help develop solutions for assisted living, fitness tracking, sleep monitoring, and several other fields. Towards addressing this issue, we raise the question: can a machine learn to recognize a diverse set of contexts and activities in a real- life through joint learning from raw multi-modal signals (e.g. accelerometer, gyroscope and audio etc.)? In this paper, we propose a multi-stream temporal convolutional network to address the problem of multi-label behavioral context recognition. A four-stream network architecture handles learning from each modality with a contextualization module which incorporates extracted representations to infer a user’s context. Our empirical evaluation suggests that a deep convolutional network trained end-to-end achieves an optimal recognition rate. Furthermore, the presented architecture can be extended to include similar sensors for performance improvements and handles missing modalities through multi-task learning without any manual feature engineering on highly imbalanced and sparsely labeled dataset. ## Introduction The problem of context recognition is centered on inferring person’s environment, physical state, and activity performed at any particular time. Specifically, a understanding of the user’s current context requires determining where and with whom the person is? and in what type of activity the person is involved in? The behavioral and activity analysis is an important and challenging task mainly because it is crucial for several applications, including smart homes (?), assisted living (?; ?), fitness tracking (?), sleep monitoring (?), user-adaptive services, social interaction (?) and in industry. In particular, an accurate recognition of human context can greatly benefit healthcare and wellbeing through automatic monitoring and supervision of patients with chronic diseases (?) such as hypertension, diabetes and dementia (?). Furthermore, the gathered knowledge and extracted activity patterns can enable novel treatment design, adjustment of medications, better behavioral intervention and patient observation strategies (?). In practice, for a context detection system to be effective in a real-life requires an unobtrusive monitoring. It is important to not distress a person in order to capture their realistic behaviors in a natural environment. The penetration of smart sensing devices (e.g. smartphones and wearables) that are integrated with sophisticated sensors in our daily lives provides a great opportunity to learn and infer about various aspects of a person’s daily life. However, there is considerable variability in the human behavior in real-world situations that can cause the system to fail, if it is developed using data collected in a constrained environment. For instance, ? shows that the accuracy of activity classification differs based on the interaction with the phone e.g. when in hand or carried in the bag. The various sensors embedded in the smart devices convey information about different ambient facets each with a distinct prospect. The variability issues of different patterns in phone usage, environments, and device types can be very well addressed (to improve the recognition capability of the system) through learning disentangled representations from a large-scale data source and fusing rich sensory modalities rather than separately utilizing each of them. Figure 1: Multi-modal representation learning from sensors: Schematic of the proposed multi-stream convolutional network. (a) Audio (MFCC) (b) Accelerometer (c) Gyroscope Figure 2: Context recognition dataset: Samples from large-scale multi-modal sensory data collected in-the-wild conditions. The individual plots within each sub-figure correspond to the same set of activities/context. In the past, several studies have shown great improvement in sensor processing for basic activity recognition (?; ?). The majority of the earlier methods use shallow learning classifiers (such as, Random Forest and Support Vector Machine) with hand-engineered features extracted from raw sensor readings e.g. heuristically selected statistical or frequency measures (?). Likewise, many studies involve simulated controlled trials for data collection in lab environments that require users to wear extra sensors. Broadly, they also treat activity recognition as a multi-class classification problem, where a user’s activity at a specific moment can be defined by one of the k defined classes. On the contrary, people are not generally engaged in just one activity in their day-to-day living e.g. a person might surf the web while eating or talking to friends. These problems limit the applicability of these studies to detect very few rudimentary activities and make it harder for the system to generalize to real-life settings. Nevertheless, to be successful in everyday scenarios, the context recognition module should support a diverse set of activities, varying device usage, and a wide range of environments. Importantly, it must not only learn discriminative representations directly from raw signals without any ad-hoc feature engineering, but also seamlessly combine the discovered explanatory factors in the milieu of diverse sensory modalities (?). In recent years, the fields of speech recognition, drug discovery, image segmentation and machine translation have been tremendously revolutionized thanks to the availability of massive labeled datasets and end-to-end deep representation learning (?). Similarly, the domain of human activity recognition has also started leveraging deep neural networks for automatic feature learning (?; ?; ?) though commonly restricted to the detection of only elementary activities such as, walking, sitting, standing etc. There has not been the same progress in recognizing complex behavioral context in daily-life situations using devices of daily use. This can be partially attributed to the lack of a large labeled dataset, which is both expensive and time-consuming to accumulate in a real-world settings. We believe that large-scale sensory data can significantly advance context recognition. This issue is very recently addressed in (?; ?) which has open-sourced multi-modal data (see Figure 2) of activities in-the-wild. The authors provide a baseline system for sensor fusion and a unified model for multi-label classification. They trained logistic regression and fully connected neural networks on hand-crafted features that are extracted based on extensive domain-knowledge. In this paper, we utilize this heterogeneous sensors data collected over a week from sixty users to learn rich representations in an end-to-end fashion for recognizing multi-label human behavioral context. The task of learning detailed human context is challenging, especially from imbalanced and multi-label data. Unconstrained device usage, a natural environment, different routines, and authentic behaviors are likely to result in a joint training dataset from several users with significant class skew (?) and missing labels. Another challenge with learning from multi-modal signals is developing an architecture that feasibly combines them as in diverse environments a certain sensor might perform better than others. For instance, if a person is watching a television with a phone lying on the table, the sound modality may dominate in the network as compared to an accelerometer. We address the former issue with instance weighting scheme same as (?) and later through a unified architecture that can efficiently fuse representations in multiple ways. We present a deep temporal convolutional neural network (CNN) that learns directly from various modalities through a multi-stream architecture (accelerometer, gyroscope, sound and phone state networks). Here, a separate network facilitates learning from each modality and a contextualization module incorporates all the available information to determine the user’s context (see Figure 1). In our experiments, we show that deep multi-modal representations learned through our network without any sophisticated pre- processing or manual feature extraction achieve state-of-the-art performance. The primary contribution of this paper is in showing how to leverage ample amount of raw sensory data to learn deep cross-modal representations for multi-label behavioral context. Although, the methods in the paper are standard, their application on a large-scale imbalanced and sparsely labeled smartphone data set is unique. The proposed network architecture achieves sensitivity and specificity score of $0.767$ and $0.733$, respectively averaged over $51$ labels and $5$-folds cross-validation. The rest of the paper describes our technique and experiments in detail. First, we review the related work on activity recognition. Then we present our multi-stream temporal convolutional network, architectural modifications for handling missing sensors, the proposed training procedure and implementation details. Next, the description of the dataset, evaluation protocol and experimental results are described, followed by the conclusions. Figure 3: End-to-end multi-modal and multi-label context recognition: We propose a deep temporal convolutional architecture for multi-label behavioral context recognition. A separate network learns representations (features) from each modality using depthwise-separable convolutions and contextualizes this information through shared layers to infer the user context. ## Related Work Human activity recognition has been extensively studied in simulated and controlled environments. It is concerned with classifying sensor measurements into existing activity categories. The earlier techniques are predominantly based on applying shallow learning algorithms on manually extracted features (e.g. statistical and spectral attributes) (?). Despite there are unsupervised (?; ?) and supervised (?; ?; ?; ?) deep learning methods applied for automatic feature extraction to detect activities, these approaches are fairly limited by the amount of labeled data (of many sensing modalities) from the real- world. Furthermore, they do not fully address the issue of multi-label context recognition. A user state is described by only one class or label, which is not true for activities humans perform in real-life. Moreover, only recently the exploration has begun into joint-learning and fusing multiple modalities for ubiquitous sensing through deep networks (?; ?). The works cited here are by no means an exhaustive list, but provide a recent representative advancements made in utilizing deep neural networks for activity recognition. We recommend the interested readers to refer (?; ?) for an extensive survey of former approaches. A systematic analysis of several deep neural architectures for activity recognition is provided by ?. The suitability of various models is investigated that were trained only on raw accelerometer signals for different activity classification tasks. On diverse benchmark datasets, CNN and long- short-term memory networks are found to outperform hand-crafted features by a significant margin. Likewise, ? proposed an approach combining pre-training and fine-tuning of deep belief networks for sequential activity recognition. They extracted spectrograms from a triaxial accelerometer and found them to be helpful for capturing variations in the input. Similarly, ? used $2$D activity images extracted from accelerometer signals as CNN input. The importance of unsupervised training of models in feature learning and optimization is highlighted in (?) using a combination of sparse-coding framework and semi- supervised learning. Likewise, ? developed a multi-channel CNN model to replace heuristic based hand-crafted features. Their analysis showed CNNs work well compared to traditional (shallow) learning algorithms on several datasets. Audio sensing is also employed in unconstrained acoustic environments through applying fully connected neural networks (?). Recently, ? used deep networks for multi-modal activity recognition and compared them with traditional learning algorithms on various recognition tasks. Likewise, numerous other studies also positively utilize deep learning for detection of basic activities (?; ?; ?). We differentiate ourselves from the existing approaches through utilizing a deep multi-stream CNN (with depthwise separable convolutions) on a large and diverse context detection dataset. Specifically, we build on previous work by ? that only employed hand-engineered features for training linear and shallow neural networks. In contrast, our general-purpose approach allows us to train a deeper network that can not only automatically discover hidden latent factors, but also seamlessly combine them to achieve an end-to-end learning system without requiring domain expertise. Moreover, through taking advantage of multi-task learning (?) we develop an architecture that can robustly handle missing sensors. ## Learning Multi-Modal Networks We design a deep convolutional neural network to address the problem of behavioral context recognition through learning representations from raw sensory inputs. To deal with cross-modality signals i.e. accelerometer (Acc), gyroscope (Gyro), audio (MFCC/Aud), and phone state (PS), we use a multi- stream architecture. The network comprises five main modules as demonstrated in Figure 3. This section describes each component, presents a strategy to modify the proposed architecture to handle missing sensors and provides the implementation details. ### Modality Specific Networks We present a deep multi-modal convolutional architecture for learning context representations. We propose to use a series of depthwise-separable convolutions (DPS-Conv) (?) for processing different components (or channels) of raw signals. In general, CNNs are also found to be well suited for processing $1$D sequences due to their ability to learn translation invariant features, scale separation, and localization of filters across time and space (?). DPS-Conv consists of two operations i.e. a depthwise convolution and a pointwise (or $1$ x $1$) convolution. Specifically, the first function (depthwise convolution) performs a convolution independently over each input channel and it is followed by the second operation of $1$ x $1$ convolution that projects the channels estimated by the earlier onto a distinct channel space to have the same number of output filters (?). The intuition of this formulation falls in line with the classical procedures utilized by domain experts to extract several features from each signal component independently (e.g. $x$, $y$ and $z$ constituents of an accelerometer) but pointwise convolution goes one step further and tries to learn unified factors that may capture relationships among independent elements. Moreover, separable convolutions make efficient use of parameters as opposed to their classical counterpart and this property has made them a very promising candidate for contemporary architectures that run on smart devices with limited computing and energy capabilities (?; ?). Formally, in case of $1$D input sequence $\mathbf{x}$ of length $L$ with $M$ channels, the aforementioned operation can be formulated as follows (?): $\text{DepthwiseConv}(\mathbf{x},\mathbf{w})_{i}=\sum_{l}^{L}(\mathbf{x}[i:i+k-1]\odot\mathbf{w})_{l}$ $\text{PointwiseConv}(\mathbf{x},\mathbf{w})_{i}=\sum_{m}^{M}(\mathbf{x}[i:i+k-1]\cdot\mathbf{w})_{m}$ $\text{DepthwiseSeparableConv}(\mathbf{x},\mathbf{w_{d}},\mathbf{w_{p}})_{i}=\\\ \text{PointwiseConv}_{i}(\text{DepthwiseConv}_{i}(\mathbf{x}[i:i+k-1],\mathbf{w_{d}}),\mathbf{w_{p}}\\\ $ where $\odot$ is elements-wise product, $\mathbf{x}[i:j]$ represents a segment of the complete sequence with adjacent columns from $i$ to $j$, and $\mathbf{w}$ represents filter with receptive field size of $k$. The proposed network takes four different signals as input, each with its independent disjoint pathway in the earlier layers of the network. Towards the end, they are merged into shared layers that are common across all modalities that are described in the next subsection. This network configuration has the benefit of not just extracting modality-specific (and channel-specific) features but it can also feasibly extract mutual representations through shared layers. Each of the presented Acc and Gyro networks consist of $2$ temporal convolution layers which act as feature extractors over raw signals of dimensions $800$ x $3$. The convolution layers have kernel sizes of $64$ and $32$ with a stride of $2$ and each layer has $32$ and $64$ filters, respectively. We use rectified linear activation in all the layers and apply depth-wise L$2$-regularization with a rate of $0.0001$. The audio network takes mel frequency cepstral coefficients (see Section Dataset and Modalities) of size 420 x 13 as input and it has a similar architecture except the kernel size, which is set to $8$ and $6$ in the first and second layers, respectively. Likewise, the discrete attributes indicating PS are fed into a single layer fully-connected (FC) network with $64$ units and L$1$-penalty is used on the weights with a rate of $0.0001$. Furthermore, we explore different mechanisms to get a fixed dimension vector from each modality that can be fed into a shared network. Specifically, we use: a) global max pooling (GMP), b) global average pooling (GAP), c) a FC layer, and d) exactly pass the representations without any transformation to the shared network. ### Shared Network (Contextualization) Given the concepts extracted from each modality, the shared network generates a modal-agnostic representation. To achieve this, we fuse the output of earlier networks either through concatenation or apply standard convolution (only for Acc, Gyro and Aud). We then feed the output into $2$ FC layers having $2048$, $1024$ hidden units, respectively. Same as earlier, we use rectified linear non-linearity and L$1$-regularization with a weight decay coefficient of $0.0001$. The final output layer contains $51$ units (one for each label) with sigmoid activation. Figure 3 visualizes the sharing of the network layers, where, earlier layers are modality specific but downstream layers become more general. ### Missing Sensors In a real-life setting, a context recognition system may encounter missing modalities which can limit its inference capability. To make the model robust against such a situation, we develop a multi-task network (?), where learning from each sensor is posed as a task. The initial configuration of the model is the same as before but an additional layer (of $128$ units for Acc, Gyro, MFCC/Aud and $64$ units for PS) with a separate loss function is added after only a single shared layer of $1024$ hidden units. Figure 4 provides a high- level overview of the architecture. We employ joint-training (with a learning rate of $0.0003$) on all the modalities through aggregating cost functions of each model in order to get a total loss. This architectural configuration allows not only to learn independent and shared factors but enables inference even when any of the sensors is missing. It does so through averaging (which can be weighted) over probabilities produced by the individual networks. Figure 4: Handling Missing Sensors with a Multi-task Network: A variant of the earlier defined architecture with additional task (modality-specific) layers and a separate loss function for each modality. It is able to recognize user context even if only one sensor is producing data and the others are unavailable. ### Implementation and Training Details The networks are implemented in Tensorflow (?) and the models are learned from scratch; initializing the weights with Xavier technique (?). Dropout (?) is applied on the hidden layers with a probability of $0.2$. We use the Adam optimizer with a learning rate of $0.0001$ (unless mentioned otherwise) and use a batch size of $100$. We optimize the model weights for a fixed number of iterations (i.e. $15000$) with mini-batch stochastic gradient descent and backpropagation using instance-weighted cross-entropy objective function: $\mathcal{J}_{C}=\dfrac{1}{NC}\sum_{i=1}^{N}\sum_{c=1}^{C}\Psi_{i,c}\cdot\mathcal{L}_{CE}(\hat{y}_{i,c},y_{i,c})$ $\mathcal{L}_{ce}(\hat{y},y)=-[(y\log(\hat{y})+(1-y)\log(1-\hat{y}))]$ where $\mathcal{L}_{ce}$ is the binary cross-entropy loss, and $\Psi$ is an instance-weighting matrix of size $N$ x $C$ (i.e. number of training examples and total labels, respectively). The instance weights in $\Psi$ are assigned by inverse class frequency. Likewise, the entries for the missing labels are set to zero, to impose no contribution in the overall cost from such examples. ## Experimental Results We conduct several experiments to analyze the capability of the proposed method. First, we provide a brief description of the utilized dataset and signals. Second, we describe the evaluation approach and metrics used to determine the model’s performance on a multi-label and imbalanced dataset. Finally, we discuss our empirical observations, effect of different modalities’ representation, comparison of various procedures to learn shared factors and visualization of the internal representation. ### Dataset and Modalities We choose to learn discriminative representations directly from raw Acc, Gyro, Aud/MFCC and PS attributes from a smartphone because of their wide adoptability and ubiquity. For this purpose, we chose to leverage ExtraSensory Dataset (?) since it is collected in a natural environment from users’ personal devices. The experimental setup was not scripted but data collection was performed when participants were busy with their daily routines to capture varied activities and context combinations, in-the-wild conditions. This data source contains over $300,000$ multi-labeled instances (with classes such as ‘outside’, ‘at a restaurant’, ‘with friends’ from a total of $51$ labels) from sixty users. The complete data collection protocol is described in (?). Here, we provide a high-level overview of the signals that we used in this study. The samples are collected for $20$ seconds duration every minute from tri-axis Acc and Gyro at a sampling frequency of $40$Hz, mel frequency cepstral coefficients (MFCCs) for $46$msec frame are extracted from Aud recorded at $22,050$Hz. Likewise, several phone state binary features are also collected such as those specifying, time of day, battery level, ringer mode and Wi-Fi connection etc. A few randomly selected samples of these signals are illustrated in Figure 2. We seek to process raw sensory values without manual feature engineering. Thus, the only pre-processing we applied is to transform variable length inputs to an identical temporal length. For this purpose, the MFCCs of environmental audio are repeated (along time dimension) to get equal size input, this is reasonable for ambient soundscapes as we are not particularly interested in inferring a specific sound event. Similarly, the Acc and Gyro samples of varying sizes are zero-padded and instances, where MFCC length is shorter than twenty are discarded. Furthermore, we treat Acc, Gyro and Aud as $m$-channels inputs ($3$, $3$, and $13$ channels, respectively) as it allows us to efficiently learn independent factors from every sensor axis, thus maximally utilizing the large-scale dataset. ### Evaluation and Metrics Our models are evaluated with five-folds cross-validation with the same divisions of sixty users as of (?), where training and test folds contain $48$ and $12$ users, respectively. For hyper-parameter optimization, we use nested cross-validation (?) by randomly dividing the training fold data into training and validation sets with ratio of $80$-$20$. After hyper-parameters selection, we train our models on the complete dataset of training folds (individually, each time from scratch) and calculate metrics on the testing folds. Furthermore, it is mentioned earlier that the considered dataset is highly imbalanced with sparse labels. In this case, simply calculating naive accuracy will be misleading due to not taking underrepresented classes into account. Similarly, precision and f$1$-score are also very likely to be affected by the class-skew due to involvement of true positives in the denominator. Hence, we adopt a metric named balanced accuracy (BA) (?) as used in (?), which incorporates both recall (or true positive rate) and true negative rate: $\text{BA}=\frac{\text{Sensitivity}+\text{Specificity}}{2}$. BA can be interpreted as average accuracy achieved on either class (positive or negative regarding binary classification). It stays identical to traditional accuracy, if a model performs equally well on each class but drops to a random chance (i.e. $50$%) if a classifier performs poorly on a class with few instances (?). We calculate BA for each label independently and average them afterwards to get a trustworthy score of the model’s overall performance. ### Results and Analysis #### Analysis of Fusing Multi-Modal Representations: We quantify the effect of different procedures for getting a fixed dimension feature vector from each modality-specific network and examine their fusion through different configurations of the shared network. It is important to note that, we keep an entire network’s configuration same but only the layers under consideration are changed. Table 1 provides the averaged (metrics) scores over $51$ contextual labels and $5$-folds as a result of applying global (max and average) pooling, using FC layer or simply feeding the extracted representations to the shared network for further processing. For the latter, we explore learning mutual representation from Acc, Gyr, and Aud/MFCC through an additional standard convolution layer and compare its performance with directly using flattened representations. Our experiments suggest that global max pooling (GMP) over each modality’s features outperforms other utilized techniques; achieving BA of $0.750$ with a sensitivity rate of $0.767$. We believe the reason is that, GMP is capable of picking-up high-level shift-invariant features, which are most discriminative among others. Figure 5 presents per label metrics for this network on all the $51$ labels in the dataset. Specifically, we notice majority of the labels have BA score in range of $70\%$-$80\%$. Figure 5: Performance metrics per label of the best performing model (with GMP): The scores are averaged over $5$-folds cross-validation. Table 1: Multi-modal context recognition: The metrics are reported for $5$-folds cross-validation averaged over $51$ class labels. BA stands for balanced accuracy. | BA | Sensitivity | Specificity ---|---|---|--- GMP | 0.750 ($\pm$ 0.012) | 0.767 ($\pm$ 0.015) | 0.733 ($\pm$ 0.016) GAP | 0.748 ($\pm$ 0.009) | 0.753 ($\pm$ 0.012) | 0.742 ($\pm$ 0.015) FC | 0.744 ($\pm$ 0.009) | 0.735 ($\pm$ 0.014) | 0.753 ($\pm$ 0.008) Flattened | 0.742 ($\pm$ 0.014) | 0.734 ($\pm$ 0.029) | 0.749 ($\pm$ 0.007) Conv | 0.738 ($\pm$ 0.011) | 0.725 ($\pm$ 0.022) | 0.752 ($\pm$ 0.022) #### Comparison of Convolution Variants: We evaluate the complete multi-stream model through replacing only DPS-Conv layers with standard convolution (Std-Conv) in modality-specific networks. We did not observe major performance differences between the two models as shown in Table 2. Nevertheless, a model with DPS-Conv should be preferred because of having lower computational cost than Std-Conv (?). Table 2: Performance evaluation with different convolution layers. | BA | Sensitivity | Specificity ---|---|---|--- Std-Conv | 0.751 ($\pm$ 0.011) | 0.750 ($\pm$ 0.017) | 0.751 ($\pm$ 0.007) | DPS-Conv --- 0.750 ($\pm$ 0.012) | 0.767 ($\pm$ 0.015) | 0.733 ($\pm$ 0.016) #### Quantifying Modality Influence: To examine the effect of different combinations of sensors (or features learned from them) on the recognition capability of the model, we experimented with training several networks with modified architectures. Specifically, in this case the model only consisted of layers that are relevant to the signals under consideration e.g. for evaluating models with only Acc, Aud, and PS, we removed the Gyro network entirely and then trained it end-to-end from scratch. Table 3 shows the evaluation results that highlights the importance of joint- learning and fusion of multiple modalities to improve detection rate. Table 3: Effect of different modalities on recognition performance. | BA | Sensitivity | Specificity ---|---|---|--- Acc | 0.633 ($\pm$ 0.011) | 0.668 ($\pm$ 0.027) | 0.599 ($\pm$ 0.017) Gyro | 0.639 ($\pm$ 0.011) | 0.638 ($\pm$ 0.017) | 0.640 ($\pm$ 0.020) Aud | 0.669 ($\pm$ 0.024) | 0.731 ($\pm$ 0.028) | 0.608 ($\pm$ 0.025) PS | 0.712 ($\pm$ 0.005) | 0.723 ($\pm$ 0.011) | 0.700 ($\pm$ 0.013) Acc, Gyro, PS | 0.733 ($\pm$ 0.010) | 0.744 ($\pm$ 0.021) | 0.722 ($\pm$ 0.014) Acc, Gyro, Aud | 0.708 ($\pm$ 0.010) | 0.722 ($\pm$ 0.027) | 0.693 ($\pm$ 0.012) Acc, Aud, PS | 0.745 ($\pm$ 0.013) | 0.757 ($\pm$ 0.025) | 0.733 ($\pm$ 0.015) Gyro, Aud, PS | 0.748 ($\pm$ 0.012) | 0.768 ($\pm$ 0.014) | 0.728 ($\pm$ 0.014) All | 0.750 ($\pm$ 0.012) | 0.767 ($\pm$ 0.015) | 0.733 ($\pm$ 0.016) #### Fusion and Effect of Missing Sensors: We now evaluate the modified architecture’s predictive performance (presented in Section Missing Sensors), confronting various combinations of missing signals. Table 4 provides experimental results showing that the proposed multi-task network can handle lost modalities, achieving similar BA score as when separate models for each modality are developed (see Table 3). However, this flexibility comes at the price of slightly lower BA but makes a model capable of operation in the face of unavailable sensors. Table 4: Assessment of multi-task network for handling missing modalities. Each row provide averaged metrics score as earlier but only mentioned modalities that are used for determining user’s context. | BA | SN | SP ---|---|---|--- Acc | 0.634 ($\pm$ 0.008) | 0.652 ($\pm$ 0.027) | 0.616 ($\pm$ 0.013) Gyro | 0.619 ($\pm$ 0.016) | 0.632 ($\pm$ 0.040) | 0.606 ($\pm$ 0.023) Aud | 0.656 ($\pm$ 0.026) | 0.670 ($\pm$ 0.046) | 0.641 ($\pm$ 0.015) PS | 0.688 ($\pm$ 0.009) | 0.709 ($\pm$ 0.015) | 0.667 ($\pm$ 0.012) Acc, Gyro | 0.646 ($\pm$ 0.009) | 0.670 ($\pm$ 0.028) | 0.622 ($\pm$ 0.018) Acc, Aud | 0.687 ($\pm$ 0.015) | 0.695 ($\pm$ 0.035) | 0.679 ($\pm$ 0.008) Acc, PS | 0.708 ($\pm$ 0.007) | 0.713 ($\pm$ 0.015) | 0.702 ($\pm$ 0.012) Gyro, Aud | 0.687 ($\pm$ 0.020) | 0.699 ($\pm$ 0.045) | 0.676 ($\pm$ 0.015) Gyro, PS | 0.708 ($\pm$ 0.007) | 0.719 ($\pm$ 0.023) | 0.696 ($\pm$ 0.019) Aud, PS | 0.708 ($\pm$ 0.013) | 0.717 ($\pm$ 0.027) | 0.698 ($\pm$ 0.010) Acc, Gyro, Aud | 0.690 ($\pm$ 0.012) | 0.703 ($\pm$ 0.031) | 0.677 ($\pm$ 0.011) Acc, Gyro, PS | 0.705 ($\pm$ 0.007) | 0.714 ($\pm$ 0.023) | 0.696 ($\pm$ 0.019) Acc, Aud, PS | 0.721 ($\pm$ 0.007) | 0.729 ($\pm$ 0.019) | 0.712 ($\pm$ 0.011) Gyro, Aud, PS | 0.721 ($\pm$ 0.011) | 0.730 ($\pm$ 0.030) | 0.711 ($\pm$ 0.017) All | 0.720 ($\pm$ 0.008) | 0.728 ($\pm$ 0.025) | 0.712 ($\pm$ 0.015) Figure 6: Assessment of instance-weighting and regularization: We determine the impact of cost sensitive loss function and regularization (i.e. weight decay and dropout) on the network’s predictive power. The results labeled under standard are with both IW and regularization. (a) Environment (b) Body State (c) Transportation Mode (d) Phone Position Figure 7: t-SNE embeddings: We visualize the mutual features learned through fusion of multiple modalities (from the last layer) in the shared network. Four sets of mutually-exclusive labels are identified from multi-labeled data to use during final visualization of semantically related clusters extracted through t-SNE. #### Reliance on Instance Weighting and Regularization: Our results thus far have been obtained through training a model with cross- entropy loss. This incorporated instance-weights to handle class-imbalance. To test network’s dependence on the cost sensitive loss function ($\mathcal{J}_{c}$), we examined a model’s performance that is trained without it. As expected, the overall BA score drastically drops to a random chance (see Figure 6) with worse performance on positive samples in comparison with the negative ones. Likewise, we also trained a model without any sort of regularization i.e. removing dropout, L$1$ and L$2$ penalties from the network. The average recall rate on the held-out testing folds dropped to $0.58$ which can be an indication of overfitting the training set. Hence, incorporating both instance-weighting (IW) and regularization improved performance significantly in learning from this imbalanced dataset. However, further work will be necessary to investigate other techniques for managing (sparse) rare labels such as oversampling and data augmentation in case of multi-labeled instances. #### Visualization: In order to illustrate the semantic relevance of the learned features, we applied t-SNE (?) to project high-dimensional data to $2$D embedding. We take the output of the last FC layer (see Figure 3) from the shared network by feeding a limited (but randomly selected) subset of the dataset to extract the embeddings. Further, as the data under consideration is multi-labeled, we identified sets of mutually-exclusive labels (e.g. Indoors vs. Outside) that can be used to color code the data points to visually identify meaningful clusters. Figure 7 provides a visualization for various sets of labels suggesting the network is able to disentangle possible factors of variation that may distinguish a class from the rest in large-scale sensory data. Furthermore, to get better insights in the diversity of the extracted features from each modality, in Figure 8, we visualize the feature maps produced by the first layer of the DPS-Conv layer of modal-specific networks. Figure 8: Feature Maps from Modality-Specific Networks: Illustration of randomly selected (learned) features from first layer of convolutional networks. (a), (b) and (c) represent outputs from Acc, Gyro and Aud models, respectively. ## Conclusions In this work, we tackled the problem of multi-label behavioral context recognition with deep multi-modal convolutional neural networks. We propose to train an end-to-end model for jointly-learning from low-level sensory data (accelerometer, gyroscope, audio and phone state) of smart devices collected in-the-wild. Our empirical results demonstrated various strategies for feasibly fusing representations learned from different modalities and quantifying their contribution on the predictive performance. We also showed that instance-weighted cross-entropy loss (as also leveraged in (?)) and regularization schemes enable the model to generalize well on highly imbalanced (sparsely labeled) dataset. Furthermore, we present a slight modification in the proposed network’s architecture to handle missing sensors; potentially taking advantage of multi-task learning. We believe, the proposed methodology is generic enough and can be applied to other related problems of learning from multivariate time series. Additionally, potential directions for future work would involve developing techniques to handle imbalanced multi- label data, optimal sensor selection to reduce computation and battery consumption, and incorporating other analogous sensors to further improve the detection rate. Acknowledgment SCOTT (www.scott-project.eu) has received funding from the Electronic Component Systems for European Leadership Joint Undertaking under grant agreement No 737422. This Joint Undertaking receives support from the European Union’s Horizon 2020 research and innovation programme and Austria, Spain, Finland, Ireland, Sweden, Germany, Poland, Portugal, Netherlands, Belgium, Norway. Various icons used in the figures are created by Anuar Zhumaev, Tim Madle, Korokoro, Gregor Cresnar, Shmidt Sergey, Hea Poh Lin, Natalia Jacquier, Trevor Dsouza, Adrien Coquet, Alina Oleynik, Llisole, Alena, AdbA Icons, Jeevan Kumar, Artdabana@Design, lipi, Alex Auda Samora, and Michelle Colonna from the Noun Project. ## References * [Abadi et al. 2016] Abadi, M.; Barham, P.; Chen, J.; Chen, Z.; Davis, A.; Dean, J.; Devin, M.; Ghemawat, S.; Irving, G.; Isard, M.; et al. 2016\. Tensorflow: A system for large-scale machine learning. In OSDI, volume 16, 265–283. * [Alsheikh et al. 2016] Alsheikh, M. A.; Selim, A.; Niyato, D.; Doyle, L.; Lin, S.; and Tan, H.-P. 2016\. Deep activity recognition models with triaxial accelerometers. In AAAI Workshop: Artificial Intelligence Applied to Assistive Technologies and Smart Environments. * [Bai, Kolter, and Koltun 2018] Bai, S.; Kolter, J. Z.; and Koltun, V. 2018\. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271. * [Bengio, Courville, and Vincent 2013] Bengio, Y.; Courville, A.; and Vincent, P. 2013\. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence 35(8):1798–1828. * [Bhattacharya et al. 2014] Bhattacharya, S.; Nurmi, P.; Hammerla, N.; and Plötz, T. 2014\. Using unlabeled data in a sparse-coding framework for human activity recognition. Pervasive and Mobile Computing 15:242–262. * [Brodersen et al. 2010] Brodersen, K. H.; Ong, C. S.; Stephan, K. E.; and Buhmann, J. M. 2010\. The balanced accuracy and its posterior distribution. In Pattern recognition (ICPR), 2010 20th international conference on, 3121–3124. IEEE. * [Caruana 1997] Caruana, R. 1997\. Multitask learning. Machine learning 28(1):41–75. * [Cawley and Talbot 2010] Cawley, G. C., and Talbot, N. L. 2010\. On over-fitting in model selection and subsequent selection bias in performance evaluation. Journal of Machine Learning Research 11(Jul):2079–2107. * [Chollet 2017] Chollet, F. 2017\. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 1251–1258. * [Figo et al. 2010] Figo, D.; Diniz, P. C.; Ferreira, D. R.; and Cardoso, J. M. 2010\. Preprocessing techniques for context recognition from accelerometer data. Personal and Ubiquitous Computing 14(7):645–662. * [Glorot and Bengio 2010] Glorot, X., and Bengio, Y. 2010\. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, 249–256. * [Hammerla, Halloran, and Ploetz 2016] Hammerla, N. Y.; Halloran, S.; and Ploetz, T. 2016\. Deep, convolutional, and recurrent models for human activity recognition using wearables. arXiv preprint arXiv:1604.08880. * [Hoseini-Tabatabaei, Gluhak, and Tafazolli 2013] Hoseini-Tabatabaei, S. A.; Gluhak, A.; and Tafazolli, R. 2013\. A survey on smartphone-based systems for opportunistic user context recognition. ACM Computing Surveys (CSUR) 45(3):27. * [Jiang and Yin 2015] Jiang, W., and Yin, Z. 2015\. Human activity recognition using wearable sensors by deep convolutional neural networks. In Proceedings of the 23rd ACM international conference on Multimedia, 1307–1310. ACM. * [Kaiser, Gomez, and Chollet 2017] Kaiser, L.; Gomez, A. N.; and Chollet, F. 2017\. Depthwise separable convolutions for neural machine translation. arXiv preprint arXiv:1706.03059. * [Lane, Georgiev, and Qendro 2015] Lane, N. D.; Georgiev, P.; and Qendro, L. 2015\. Deepear: robust smartphone audio sensing in unconstrained acoustic environments using deep learning. In Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing, 283–294. ACM. * [Lara and Labrador 2013] Lara, O. D., and Labrador, M. A. 2013\. A survey on human activity recognition using wearable sensors. IEEE Communications Surveys and Tutorials 15(3):1192–1209. * [Lee et al. 2013] Lee, Y.; Min, C.; Hwang, C.; Lee, J.; Hwang, I.; Ju, Y.; Yoo, C.; Moon, M.; Lee, U.; and Song, J. 2013\. Sociophone: Everyday face-to-face interaction monitoring platform using multi-phone sensor fusion. In Proceeding of the 11th annual international conference on Mobile systems, applications, and services, 375–388. ACM. * [Lin et al. 2012] Lin, M.; Lane, N. D.; Mohammod, M.; Yang, X.; Lu, H.; Cardone, G.; Ali, S.; Doryab, A.; Berke, E.; Campbell, A. T.; et al. 2012\. Bewell+: multi-dimensional wellbeing monitoring with community-guided user feedback and energy optimization. In Proceedings of the conference on Wireless Health, 10. ACM. * [Lin et al. 2015] Lin, Q.; Zhang, D.; Connelly, K.; Ni, H.; Yu, Z.; and Zhou, X. 2015\. Disorientation detection by mining gps trajectories for cognitively-impaired elders. Pervasive and Mobile Computing 19:71–85. * [Lorincz et al. 2009] Lorincz, K.; Chen, B.-r.; Challen, G. W.; Chowdhury, A. R.; Patel, S.; Bonato, P.; Welsh, M.; et al. 2009\. Mercury: a wearable sensor network platform for high-fidelity motion analysis. In SenSys, volume 9, 183–196. * [Miluzzo et al. 2008] Miluzzo, E.; Lane, N. D.; Fodor, K.; Peterson, R.; Lu, H.; Musolesi, M.; Eisenman, S. B.; Zheng, X.; and Campbell, A. T. 2008\. Sensing meets mobile social networks: the design, implementation and evaluation of the cenceme application. In Proceedings of the 6th ACM conference on Embedded network sensor systems, 337–350. ACM. * [Ordóñez and Roggen 2016] Ordóñez, F. J., and Roggen, D. 2016\. Deep convolutional and lstm recurrent neural networks for multimodal wearable activity recognition. Sensors 16(1):115. * [Plötz, Hammerla, and Olivier 2011] Plötz, T.; Hammerla, N. Y.; and Olivier, P. 2011\. Feature learning for activity recognition in ubiquitous computing. In IJCAI Proceedings-International Joint Conference on Artificial Intelligence, volume 22, 1729. * [Rabbi et al. 2015] Rabbi, M.; Aung, M. H.; Zhang, M.; and Choudhury, T. 2015\. Mybehavior: automatic personalized health feedback from user behaviors and preferences using smartphones. In Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing, 707–718. ACM. * [Radu et al. 2018] Radu, V.; Tong, C.; Bhattacharya, S.; Lane, N. D.; Mascolo, C.; Marina, M. K.; and Kawsar, F. 2018\. Multimodal deep learning for activity and context recognition. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 1(4):157. * [Rashidi and Cook 2009] Rashidi, P., and Cook, D. J. 2009\. Keeping the resident in the loop: Adapting the smart home to the user. IEEE Transactions on systems, man, and cybernetics-part A: systems and humans 39(5):949–959. * [Rashidi and Mihailidis 2013] Rashidi, P., and Mihailidis, A. 2013\. A survey on ambient-assisted living tools for older adults. IEEE journal of biomedical and health informatics 17(3):579–590. * [Ronao and Cho 2016] Ronao, C. A., and Cho, S.-B. 2016\. Human activity recognition with smartphone sensors using deep learning neural networks. Expert Systems with Applications 59:235–244. * [Sandler et al. 2018] Sandler, M.; Howard, A.; Zhu, M.; Zhmoginov, A.; and Chen, L.-C. 2018\. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 4510–4520. * [Shoaib et al. 2015] Shoaib, M.; Bosch, S.; Incel, O. D.; Scholten, H.; and Havinga, P. J. 2015\. A survey of online activity recognition using mobile phones. Sensors 15(1):2059–2085. * [Srivastava et al. 2014] Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; and Salakhutdinov, R. 2014\. Dropout: A simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research 15(1):1929–1958. * [Vaizman, Ellis, and Lanckriet 2017] Vaizman, Y.; Ellis, K.; and Lanckriet, G. 2017\. Recognizing detailed human context in the wild from smartphones and smartwatches. IEEE Pervasive Computing 16(4):62–74. * [Vaizman, Weibel, and Lanckriet 2018] Vaizman, Y.; Weibel, N.; and Lanckriet, G. 2018\. Context recognition in-the-wild: Unified model for multi-modal sensors and multi-label classification. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 1(4):168. * [van der Maaten and Hinton 2008] van der Maaten, L., and Hinton, G. 2008\. Visualizing data using t-sne. Journal of machine learning research 9(Nov):2579–2605. * [Yang et al. 2015] Yang, J.; Nguyen, M. N.; San, P. P.; Li, X.; and Krishnaswamy, S. 2015\. Deep convolutional neural networks on multichannel time series for human activity recognition. In IJCAI, 3995–4001. * [Zeng et al. 2014] Zeng, M.; Nguyen, L. T.; Yu, B.; Mengshoel, O. J.; Zhu, J.; Wu, P.; and Zhang, J. 2014\. Convolutional neural networks for human activity recognition using mobile sensors. In Mobile Computing, Applications and Services (MobiCASE), 2014 6th International Conference on, 197–205. IEEE. * [Zhang et al. 2017] Zhang, X.; Zhou, X.; Lin, M.; and Sun, J. 2017\. Shufflenet: An extremely efficient convolutional neural network for mobile devices. arXiv preprint arXiv:1707.01083.
# Design and characterisation of an antiproton deceleration beamline for the PUMA experiment J. Fischera Corresponding author<EMAIL_ADDRESS>A. Schmidta N. Azaryanb F. Butinb J. Ferreira Somozab A. Hussona C. Klinka A. Obertellia M. Schlaicha A. Sinturelb N. Thausb and F. Wienholtza (aTechnische Universität Darmstadt, Institut für Kernphysik, Schloßgartenstraße 9, 64289 Darmstadt, Germany bCERN, 1211 Geneva 23, Switzerland) ## 1 Introduction The spatial distribution of protons and neutrons at and beyond the nuclear surface of atomic nuclei challenges nuclear theory. In particular, nuclei with a neutron excess exhibit a so-called neutron skin, where the neutron density distribution extends beyond the proton density distribution. The thickness of the neutron skin is defined as the difference in root-mean-square radii of the density distributions $\displaystyle\Delta r_{\mathrm{np}}=\langle r^{2}_{\mathrm{n}}\rangle^{1/2}-\langle r^{2}_{\mathrm{p}}\rangle^{1/2}.$ (1) The neutron skin thickness correlates with the slope parameter $L$ of the nuclear equation of state [1], playing an important role in defining the relation between the mass and radius of a neutron star [2, 3]. Neutron skin thicknesses have been investigated with several methods [4, 5, 6, 7, 8], mostly on stable nuclei, while the challenge lies in determining the radius of the neutron distribution $\langle r^{2}_{\mathrm{n}}\rangle^{1/2}$ with enough accuracy and controlled theoretical uncertainties. Information on unstable nuclei is much more scarce, as illustrated by Ca isotopes: charge radii can be accessed with precision from the relative measurement of isotope shifts from laser spectroscopy and anchored to stable nuclei [9], while the interpretation of the data related to the matter or neutron radius suffers from model dependence [10, 11]. Nuclei close or at the neutron drip line can have loosely bound nucleons, whose wave function extends far beyond the charge distribution. Such systems are called halo nuclei [12, 13]. Neutron halos have been so far observed in light nuclei only [14]. Indications for $p$-wave halos in medium mass nuclei have been reported [15], while more halos are predicted to exist in uncharted regions of the nuclear landscape [16]. Proton halos have been predicted as well [17]. Most aforementioned methods to probe neutron skins and halos in stable and unstable nuclei are sensitive to the nuclear surface where $\rho\sim\rho_{0}/2$, not further out in the tail of the density distribution, where the asymmetry is the largest. The antiProton Unstable Matter Annihilation (PUMA) experiment aims to investigate these phenomena in the tail of stable and unstable nuclei with low-energy antiprotons as a probe [18, 19]. Antiprotons are uniquely suited for this, as they annihilate with nucleons at a mean radial position $\sim 2$ $\mathrm{f}\mathrm{m}$ further out from the half density radius of the nucleus [5, 20, 21], probing a region of higher neutron-to-proton asymmetry. The PUMA experiment will produce antiprotonic atoms by combining nuclei and antiprotons in a Penning trap. By studying the pions produced in the annihilation, the PUMA experiment can determine the neutron-to-proton ratio in the tail of the nuclear density distribution. The setup is located at the Antimatter Factory at CERN. Stable isotopes are supplied by an offline ion source [22], and for the investigation of more neutron-rich and unstable isotopes the setup will be transported to the ISOLDE facility [23] at CERN. The ELENA ring at the Antimatter Factory provides bunches of $5\cdot 10^{6}$ to $10^{7}$ antiprotons at 100 $\mathrm{k}\mathrm{e}\mathrm{V}$ to up to four experiments every 2 minutes [24, 25, 26]. To further decelerate the antiprotons to energies compatible with the PUMA Penning trap, one can use a thin degrader foil or pulsed drift tubes (PDT) [27, 28, 29, 30, 31, 32]. Employing a foil for deceleration is space-efficient, but the yield is low and the energy distribution broad [33], compared to a pulsed drift tube, which can have a transmission of 100% while conserving the width of the energy distribution. For antiprotons with an initial energy of approximately 100 keV, trapping efficiencies vary from a few percent [34] to a maximum of 50%, predicted in [35]. However, for the PUMA experiment, which relies on the simultaneous trapping of antiprotons and stable and unstable ions, the use of a foil is unfeasible, since low-energy ions cannot penetrate the foil. An established method to change the energy of a particle beam is to use a drift tube, where the potential can be changed rapidly. Here, the drift tube is set to a potential and is used to decelerate the particles to the desired energy. If the electrode is switched to a different potential, e.g., ground, while the particles are still inside and in the field free region of the drift tube, they are not reaccelerated on exit. Because only the longitudinal and not the transversal kinetic energy is changed, the divergence angle of the beam increases by a factor of $\sqrt{E_{\mathrm{in}}/E_{\mathrm{out}}}$, where $E$ is the kinetic energy of the incoming and outgoing particles, respectively. This can be compensated by additional ion optical elements or beam cooling. Several ion trap experiments use pulsed drift tubes to decelerate nuclei for trapping [22, 36, 37, 38], often in combination with buffer-gas cooling [39] to counteract the increase in transversal emittance, some from energies as high as $60$ $\mathrm{k}\mathrm{e}\mathrm{V}$. The GBAR experiment at CERN is confronted with a similar problem as the PUMA experiment, as they need to decelerate antiprotons to $1$ $\mathrm{k}\mathrm{e}\mathrm{V}$ [40]. At the PUMA experiment, the antiprotons are decelerated from 100 $\mathrm{k}\mathrm{e}\mathrm{V}$ to 4 $\mathrm{k}\mathrm{e}\mathrm{V}$ to allow for an efficient beam transport and in a second step down to 100 $\mathrm{e}\mathrm{V}$ right in front of the trap. To limit the annihilation of antiprotons with residual gas molecules, a vacuum of a few $10^{-10}\,$\mathrm{m}\mathrm{b}\mathrm{a}\mathrm{r}$$ along and $10^{-11}\,$\mathrm{m}\mathrm{b}\mathrm{a}\mathrm{r}$$ at the end of the beamline is critical. ## 2 Beamline Design ### 2.1 Transfer Line from ELENA to PUMA Figure 1: Schematic view of LNE51 transfer line to PUMA. Antiprotons are ejected from ELENA into LNE50, from which LNE51 branches off. The insert shows the position of LNE51 relative to ELENA and LNE00. The transfer of 100 keV particles (H- ions or antiprotons) from the ELENA machine to the PUMA experiment is performed by the so-called LNE51 transfer line. LNE51 branches off from the LNE50 line (transfer from ELENA to the adjacent GBAR experiment) using a standard ZDFA-ZDSA switching unit (fast switch and electrostatic deflector) integrated in LNE50. This equipment is interlocked with the access safety system of the PUMA zone, preventing any beam to be sent from the ELENA machine, while the area is being accessed. The sector valve at the interface between the experiment and the LNE51 transfer line is interlocked with the access system to close automatically when the zone is being accessed. This drastically limits the risk of contamination of the upstream sections of ELENA machine in case of an incident while manipulating the experimental equipment. To satisfy the integration constraints and match the beam to the PUMA experiment at the end of the line, four electrostatic quadrupole/H-V corrector units (ZQNA) are installed, along with a 37.7° standalone deflector. At the focal point, the beam spot size (rms) is approximately 2 mm and the horizontal and vertical geometric emittance ($95\%=6\epsilon_{\mathrm{rms}}$) is 6 mm mrad and 4 mm mrad, respectively [41]. The layout for LNE51 is shown schematically in Figure 1. Two SEM grids (Secondary Emission Monitors) [42] are installed in LNE51. They are standard equipment in the ELENA transfer lines that allow to extract the profile of the impinging beam, either H- ions or antiprotons. Made from x-y meshes of 50 $\upmu$m tungsten wires, covering the beam acceptance, spaced by a pitch of 0.5 mm in the central region, they intercept only about 10% of the beam at each station [43]. These monitors are ultra-high vacuum compatible, as they can be baked-out to 200°C. As bake-out is required, the vacuum line is fitted with permanently installed bake-out jackets. Figure 2: Half-section view of the beamline without the supports. The antiprotons traverse the beamline from left to right. Ions from the offline source enter the beamline at the quadrupole bender in the direction into the page and are deflected to the right. The gate valves separating the sections are depicted in blue. ### 2.2 The PUMA Antiproton Beamline Downstream of the handover point (HOP) to PUMA (see Fig. 1 and 2), the beamline consists of two main sections, that can be isolated by gate valves type 48236-CE44 from VAT (see Fig. 2). Section 1 includes the pulsed drift tube itself. It is complemented by a high-voltage (up to -85 kV) as well as a low-voltage (up to 5 kV) einzel lens (EL) on the injection and ejection sides, respectively, to focus the antiproton bunches into and out of the pulsed drift tube. Section 2 consists of two low-voltage (up to 5 kV) einzel lenses with x-y- steerers to guide the beam to the entrance of the PUMA Penning trap. In between these lenses, a quadrupole ion beam bender allows the injection of ions from an offline ion source setup, perpendicular to the antiproton beamline. Even tough the bender has been designed to allow for simultaneous injection of ions and antiprotons, it can be removed when it is not needed. A beam imaging system (BTV), which consists of a phosphorous screen and a camera, completes the section. The BTV can be moved in and out of the beamline, as it is a completely destructive measurement of the beam. In the future, the BTV will be replaced by a SEM grid. Figure 3: The field strength at an unshielded triple junction (left) and one shielded with a guard ring (right) is illustrated here. Blue indicates lower and red higher electric field strengths. ### 2.3 The Pulsed Drift Tube The pulsed drift tube (PDT) used for the PUMA experiment, is based on the GBAR design [30]. Although the high-voltage einzel lens in front of the drift tube counteracts the strong focusing effect of the decelerating electric field, the drift tube has to accommodate an expansion of the beam. The inner diameter was thus chosen to be 100 $\mathrm{m}\mathrm{m}$ with an outer diameter of 120 $\mathrm{m}\mathrm{m}$. At 4 $\mathrm{k}\mathrm{e}\mathrm{V}$, an antiproton bunch from ELENA has a length of 250 $\mathrm{m}\mathrm{m}$ ($2\sigma$) [26]. The PUMA pulsed drift tube has been designed to be 700 $\mathrm{m}\mathrm{m}$ long. This ensures, that the bunch is in the field free region of the drift tube when the potential is changed. Because of the stringent vacuum requirements ($p<10^{-10}\,$\mathrm{m}\mathrm{b}\mathrm{a}\mathrm{r}$$), materials with the lowest possible outgassing rates have to be used. Therefore, the pulsed drift tube is made from aluminium ($\sim 1\cdot 10^{-13}\,$\mathrm{m}\mathrm{b}\mathrm{a}\mathrm{r}~{}\mathrm{l}\mathrm{/}\mathrm{s}\mathrm{/}\mathrm{c}\mathrm{m}^{2}$$), which outgases less than stainless steel ($\sim 3\cdot 10^{-12}\,$\mathrm{m}\mathrm{b}\mathrm{a}\mathrm{r}~{}\mathrm{l}\mathrm{/}\mathrm{s}\mathrm{/}\mathrm{c}\mathrm{m}^{2}$$) [44]. The insulators are made from MACOR®, which has an outgassing rate of $1.1\cdot 10^{-11}\,$\mathrm{m}\mathrm{b}\mathrm{a}\mathrm{r}~{}\mathrm{l}\mathrm{/}\mathrm{s}\mathrm{/}\mathrm{c}\mathrm{m}^{2}$$ [45]. The walls of the vacuum chambers are coated with a non-evaporable getter (NEG) to pump the section. Non-evaporable getters are made from an alloy of Zr, V, Ti, Al and Fe, that can sputtered directly onto the wall of a vacuum chamber [46]. It acts as a pump by absorbing hydrogen and chemically binding other reactive gases like oxygen. To activate the NEG, the chambers are heated (200°C to 400°C). Molecules at the surface (mainly carbon, nitrogen and oxygen) diffuse into the bulk. Hydrogen is released and must be pumped away by another pump. Therefore, all components, such as vacuum gauges, valves, feedthroughs, pumps, cables and beam instrumentation, must be bakable at 250°C at least. The coating of the inside surfaces of the chambers was done at CERN. The installation of the pulsed drift tube inside the chamber must be done without touching the coating to prevent damaging it. It is first mounted onto its support structure before being lowered vertically into the vacuum chamber and secured with screws. To facilitate individual access to the high- and low- voltage einzel lens as well as the drift tube, the vacuum chamber is divided into three parts. At the intersections of vacuum, conductor and insulator, the electric field is strongly enhanced due to gaps arising from imperfections on the corners of the material (see Fig. 3). Special attention has been paid to these so-called triple junctions to prevent possible discharges. They are shielded by purpose- built rings, that surround the triple junction and thereby lower the electric field (see Fig. 3). On all components, sharp edges have been avoided, and the electrodes have been polished to an average surface finish of $R_{\mathrm{a}}=0.05\,$\upmu\mathrm{m}$$, which helps to prevent discharges [47]. #### 2.3.1 Electronics To not reaccelerate the antiprotons as they exit the pulsed drift tube, it must be discharged from -96 kV to 0 V before the first antiprotons exit the field free region of the drift tube. For antiprotons with a kinetic energy of 4 keV, the time to discharge the drift tube is in the order of 500 ns. Equipment that can withstand high voltages and high peak currents, as well as a high-voltage switch with a short transient, are needed. The pulsed drift tube is connected to a high-voltage power supply (Spellman SL130PN60) via a $1\text{\,}\mathrm{M\SIUnitSymbolOhm}$ resistor. In order not to exceed the voltage rating of the resistors, two Metallux HVR 969 resistors are used, connected via polished brass cylinders with rounded edges. The value is chosen as a compromise between the need for a high resistance to decouple the power supply from the pulsed drift tube while switching, and the need for a low resistance to minimize the effects of current fluctuations on the voltage applied to the pulsed drift tube. For the discharge of the tube’s capacitance, a fast high-voltage switch (Behlke HTS 1501-20-LC2) connects the pulsed drift tube to ground. To make sure that the switch is not damaged, the pulsed drift tube is connected to the switch via a two $250\text{\,}\mathrm{\SIUnitSymbolOhm}$ Metallux HVR 969 resistors in series, limiting the current. The high-voltage leads are connected with HN-70 connectors from R.E. Beverly III & Associates. The cables are suspended from the ceiling to avoid triple junctions at the exposed high-voltage connectors. The grounded mesh is removed on the load side, and special care is taken to cover the pointy ends of the grounded mesh. As high-voltage feedthrough, a HV125R-CE-CU39 from VACOM, rated for up to 125 kV is used. Using a 1/1000 voltage divider (LeCroy PPE6kV) connected to a Tektronix MDO3104 oscilloscope, the switching time from -5 kV to ground was measured. As can be seen in Fig. 4, there is a $\sim$250 ns delay between the trigger signal (blue) and the voltage on the pulsed drift tube (orange) which has to be taken into account when triggering the switch. Independent of the voltage applied to the switch, the transient time $\tau$ to $V_{0}/\mathrm{e}$ is $\sim$80 ns which is consistent with the time constant estimated by a simple RC-circuit, where the capacitance of the pulsed drift tube was measured and within the specs of the switch: $\displaystyle\tau=RC=500\Upomega\cdot 170\mathrm{pF}=85\,\mathrm{ns}.$ (2) Figure 4: Switching time while switching from $5\,$\mathrm{k}\mathrm{V}$$ to ground, measured with a $1/1000$ voltage divider. The trigger signal is shown in blue and the voltage on the pulsed drift tube in orange. #### 2.3.2 Safety Cage The high-voltage system has unshielded $\sim$ 100 kV connections exposed to air during operation. Therefore, the safety of the users has to be ensured by a safety cage according to the ingress protection code level IP3X. Following the European norm EN 50191, the dimensions of the safety cage are defined so that any high-voltage point in air is at a distance of more than 74 $\mathrm{c}\mathrm{m}$ from the cage, corresponding to a maximum voltage of 130 kV, the maximum voltage of the high-voltage power supply. The high-voltage system is interlocked via a switch (Telemecanique XCSDMC7902 coded magnetic switch) at the sliding door of the cage to interlock the power supplies in the event of unexpected access while the equipment is powered. The safety cage is further secured with a trapped key system from Allen Bradley (Rockwell) to prevent unauthorized access. It must first be locked to be able to switch on the high-voltage power supplies. To simplify maintenance work, panels can be removed from all sides of the cage. ## 3 Vacuum and Conditioning ### 3.1 Baseline Vacuum Pressure Due to the strict vacuum requirements at the entrance of the PUMA trap, special attention must be paid to the pressure. After activating the NEG coating, a pressure of $2\cdot 10^{-11}$ mbar was measured at the end of the pulsed drift tube section, a factor of 10 better than required. For the subsequent tests, the NEG coating was not reactivated after venting, to conserve it for the use with the PUMA trap attached. Without the NEG activated, the pressure base level is around $1.4\cdot 10^{-10}$ mbar. This is sufficient to condition and operate the pulsed drift tube. ### 3.2 High-Voltage Conditioning Surface contamination and imperfections are sources of discharges that degrade the vacuum and material when high voltage is applied. They also lead to a leakage current that drains the set potential. This difficulty can be countered by conditioning the high-voltage parts, which is therefore an essential step before operating the pulsed drift tube. It was done by a stepwise increase of the voltage, while keeping the leakage current below the limit of the power supply and the vacuum better than $5\cdot 10^{-8}$ mbar. The pulsed drift tube and high-voltage einzel lens were conditioned over several weeks. The voltage was increased step by step and left in static operation until the sudden spikes in current, associated with field emission from imperfections on the electrode, subsided, which took between 12 and 72 hours per voltage step. In addition to the conditioning, modifications to the setup were made outside the vacuum to reduce the leakage current. These focussed on increasing the distance from any high-voltage parts to ground, as well as polishing and rounding pieces in high electric fields. Ultimately, the leakage current at -96 kV could be lowered from 100 $\upmu$A to 50 $\upmu$A by polishing and increasing the corner radius of one high-voltage part from 3 mm to 15 mm. Additionally, the current could be further decreased to 11 $\upmu$A by increasing the ceiling height of the safety cage by 50 cm to 75 cm. The leakage current of the high-voltage einzel lens could not be reduced in the same way. At -85 kV, the 100 $\upmu$A current limit of the power supply is reached. This means that the design value of -90 kV could not be achieved, nevertheless it could be used for commissioning. A redesign with increased distances between high-voltage parts and ground is planned. ### 3.3 Vacuum During Operation During operation of the pulsed drift tube, the remaining leakage current inside the vacuum degrades the pressure. To mitigate this, as done by the GBAR collaboration, the voltage is kept at 0 V for most of the ELENA cycle and is increased to -96 kV only 9.5 s before a bunch of antiprotons arrives. Ramping up the voltage only shortly111compared to a repetition time of 120 s for ELENA. before the bunch arrives has the advantage, that the vacuum is below $2\cdot 10^{-10}$ mbar most of the time, since there is no leakage current at 0 V. When -96 kV are applied, the pressure reaches a value of $8\cdot 10^{-10}$ mbar and increases to $2\cdot 10^{-9}$ mbar when switching (see Fig. 5). Figure 5: The pressure in section 1 and in section 2, while switching (three cycles). This is without the NEG coating activated, to conserve it for the use with the PUMA trap attached. ## 4 Measurement of Beam Properties ### 4.1 Detection System For the characterization of the system, a vacuum chamber with several detectors was installed at the end of the beamline. To visualize the beam spot, a microchannel plate (MCP) by Hamamatsu with a phosphor screen with a diameter of 40 mm was used. In combination with the camera CS505MU and lens MVL7000 from Thorlabs, this results in the smallest resolvable feature being 40 $\upmu$m. The device was mounted on a tripod in front of a view port, which allowed to capture the beam shape. A MagneToF detector by ETP ion detect was used for two purposes: first, to determine the time of flight (ToF) of the antiprotons (<1.5 ns multiple ion pulse width), and second, in combination with an “energy grid”, to determine the kinetic energy distribution of the decelerated antiprotons. The energy grid consists of a stack of three grids by ETP ion detect with a diameter of 76.2 mm. The distance between the grids is 15 mm. The grid wires have a diameter of 0.018 mm, a centre-to-centre distance of 0.25 mm, and a transmission of 92% to 95%. The two outer grids were grounded, while a blocking voltage was applied to the middle one, with a ripple of less than 10 mV. The energy grids and the MagneToF detector can be moved out of the beam axis independently. In addition to those detectors, the BTV further upstream in the beamline (see Fig. 2) was used for particle detection and intensity determination. ### 4.2 Pulsed Drift Tube Switching Delay When antiprotons arrive in the experimental zone, a trigger signal from the ejection from ELENA is forwarded to the electronics. Relative to the trigger, a switching time $t_{\mathrm{s}}$ has to be determined, at which the bunch is fully contained inside the pulsed drift tube, so that the deceleration is successful for the full antiproton bunch. To determine the ideal value, $t_{\mathrm{s}}$ has to be scanned while observing the time of flight of the antiprotons. If $t_{\mathrm{s}}$ is too small, the antiprotons see a grounded electrode and traverse the pulsed drift tube at full speed, arriving the earliest and with their initial energy. If $t_{\mathrm{s}}$ is too large, the antiprotons are decelerated while entering the pulsed drift tube and reaccelerated when leaving it, thus they arrive later than the ones never decelerated, but still with their initial energy. When switching at the correct time, the antiproton bunch is decelerated on entry but is not reaccelerated on exit. Thus, it arrives later than in the other cases, as they are slower, which can be seen in a simulation of the deceleration in the pulsed drift tube performed in SIMION® (see top panel of Fig. 6). The results from the measurement can be seen in the bottom panel of Fig. 6, they match the behaviour expected from simulations. When $t_{\mathrm{s}}$ is too small, the antiprotons arrived early. When increasing $t_{\mathrm{s}}$, the bunch diffuses, as the bunch is partly in the fringe field of the electrode when the pulsed drift tube is switched. Afterwards, in a window of about 300 ns, the antiprotons are uniformly decelerated. As $t_{\mathrm{s}}$ is further increasing, the bunch diffuses again, because it is only partly inside the pulsed drift tube when switching. Figure 6: Simulated (top) and measured (bottom) beam intensity when switching the pulsed drift tube from $-96\,$\mathrm{k}\mathrm{V}$$ to ground and varying the switch delay $t_{\mathrm{s}}$. Yellow colours indicate lower and red higher intensity. In both cases, a successful deceleration to $4\,$\mathrm{k}\mathrm{e}\mathrm{V}$$ corresponds to a time of flight of $t_{\mathrm{4keV}}=3.85\,$\upmu\mathrm{s}$$, with a bunch length ($1\sigma$) of $0.09\,$\upmu\mathrm{s}$$. On the right, the integrated intensity from $t_{\mathrm{4keV}}-2\sigma$ to $t_{\mathrm{4keV}}+2\sigma$ is shown, $t_{\mathrm{s}}$ is chosen to maximise this intensity. The measurement shows a successful deceleration, and an estimation with the time of flight gives a deceleration to $(4.0\pm 0.5)\,$\mathrm{k}\mathrm{e}\mathrm{V}$$. A more precise measurement of the energy distribution was done using the energy grids (see Sec. 4.4). ### 4.3 Transmission and Focusing The intensity of the bunch after the pulsed drift tube $I$, can be compared to the initial intensity of the bunch $I_{0}$. The total transmission through the pulsed drift tube is thus defined by $T=I/I_{0}$. $I_{0}$ is determined before the handover point by pick-ups in the ELENA transfer lines [24]. Besides showing the beam spot shape, the total intensity on the BTV is proportional to $I$, as can be seen in Fig. 7. Using the calibration in this plot, $T$ can be calculated. Figure 7: The bunch intensity of antiprotons determined by the ELENA detectors is proportional to total intensity on the BTV. The transmission to the BTV is $100\%$ when not decelerating the antiprotons. This allows to make a calibration to determine the transmission through the pulsed drift tube while decelerating. $T$ for 100 keV bunches is about 100%. During the experiment, the transmission of antiprotons decelerated to 4 keV reached ($55\pm 3$)%, while in simulations a transmission of 100% could be reached. The main source of losses in transmission can be assigned to a misalignment of the high-voltage einzel lens and the pulsed drift tube and a high leakage current on the high-voltage einzel lens, which limited the voltage to -85 kV. In addition, the parameters assumed in the simulation for the incoming beam might also play a role. Figure 8 shows the beam profiles recorded by the BTV directly after the last einzel lens. Using a Gaussian fit, the following parameters could be obtained: $\displaystyle\sigma_{\mathrm{horiz}}=(3.0\pm 0.1)\,$\mathrm{m}\mathrm{m}$,\,\sigma_{\mathrm{vert}}=(3.8\pm 0.2)\,$\mathrm{m}\mathrm{m}$$ 64% of the antiprotons are within a circle of radius $r=5.6\,$\mathrm{m}\mathrm{m}$$, the smallest aperture of the PUMA Penning trap. The focal point will have to be optimized at a later point for the injection into the PUMA trap. Figure 8: Beam profile after optimizing the LV einzel lenses for deceleration to $4\text{\,}\mathrm{keV}$ and focus on the BTV. Fitting a Gaussian to the centre peak yields $\sigma_{\mathrm{horiz}}=3.0\,$\mathrm{m}\mathrm{m}$,\sigma_{\mathrm{vert}}=3.8\,$\mathrm{m}\mathrm{m}$$. Yellow indicates a lower and red a higher intensity. ### 4.4 Energy Distribution The standard deviation of the ions’ energy after deceleration to 4 keV at the position of the MagneToF detector was simulated to be 101 eV. The kinetic energy $E$ of the antiprotons was determined by blocking the antiprotons with the energy grids, and measuring the transmission on the MagneToF. The results from this can be seen in Fig. 9. In blue, the transmission onto the MagneToF is displayed in dependence of the kinetic energy of the antiprotons. Fitting the cumulative distribution function (CDF) of a normal distribution yields the mean energy $\mu=(3898\pm 3)$ eV and energy spread $\sigma=(127\pm 4)$ eV. The energy distribution calculated from the fit is shown in orange. 88% of decelerated antiprotons are within $\pm$ 200 eV of the central energy, which is the energy acceptance for successful trapping in the PUMA Penning trap, according to simulations. Figure 9: The energy distribution of decelerated antiprotons. The data and fitted CDF of a normal distribution are shown in blue, and the probability density function corresponding to the fit in orange. The mean energy is $\mu=(3898\pm 3)$ $\mathrm{e}\mathrm{V}$ and the standard deviation $\sigma=127\pm 4$ $\mathrm{e}\mathrm{V}$. $88\%$ of decelerated antiprotons are within $\pm 200\,$\mathrm{e}\mathrm{V}$$ of the mean energy, which is the estimated energy acceptance for trapping. ### 4.5 Bunch Length The length of the antiproton bunch at 4 keV is relevant, because it determines the losses in the second stage of deceleration to a few 100 eV right in front of the trap. The simulation predicts an increase in length from 75 ns to 89 ns at the position of the MagnetToF, with which 90% of the bunch can be trapped. A measurement of the bunch length of the decelerated antiprotons with the MagneToF yields a length ($1\sigma$) of 93$\pm$3 ns, consistent with the simulation. ## 5 Conclusion An overview of the design and the characterisation of the low-energy antiproton beam line of PUMA at ELENA is presented. Design considerations for high voltage and ultra-high vacuum are discussed, as well as procedures for high-voltage conditioning and in-vacuum high-voltage operation. The antiproton beamline is shown to be successful in decelerating antiprotons from 100 keV to $(3898\pm 3)$ eV, the first step in trapping antiprotons for the PUMA experiment. The pressure, with the pulsed drift tube not in operation, is below $2\cdot 10^{-10}$ mbar. With the implemented high-voltage ramping scheme, the pressure stays below $2\cdot 10^{-10}$ mbar 75% of the cycle time, also during operation. Currently, a transmission of ($55\pm 3$)% for antiprotons decelerated to 4 keV can be reached. The beam was focussed to a spot with $\sigma_{\mathrm{horiz}}=(3.0\pm 0.1)\,$\mathrm{m}\mathrm{m}$,$ and $\sigma_{\mathrm{vert}}=(3.8\pm 0.2)\,$\mathrm{m}\mathrm{m}$$, demonstrating it can be focussed into the PUMA Penning trap. The length of the 4 keV antiproton bunch, relevant for the second deceleration from 4 keV to 100 eV is (93$\pm$3) ns. Further improvement of the beamline is foreseen in the future, while the current performance already allows for first experiments with PUMA. ## Acknowledgements We thank the ELENA team and the operators of the Antimatter Factory for excellent beam during the runs in 2022 and 2023. We thank the technical teams at CERN and TU Darmstadt for their support. The presented work benefited from the support of the GBAR collaboration. PUMA is funded by the European Research Council through the ERC grant PUMA-726276 and the Alexander-von-Humboldt foundation. The development of PUMA and its implementation at CERN are supported by the TU Darmstadt and CERN. ## References * [1] X. Roca-Maza, M. Centelles, X. Viñas and M. Warda “Neutron Skin of 208Pb, Nuclear Symmetry Energy, and the Parity Radius Experiment” In _Phys. Rev. Lett._ 106 American Physical Society, 2011, pp. 252501 DOI: 10.1103/PhysRevLett.106.252501 * [2] K. Hebeler, J.. Lattimer, C.. Pethick and A. Schwenk “Equation of State and Neutron Star Properties Constrained by Nuclear Physics and Observation” In _Astrophys. J._ 773.1 The American Astronomical Society, 2013, pp. 11 DOI: 10.1088/0004-637X/773/1/11 * [3] C.. Horowitz and J. Piekarewicz “Neutron Star Structure and the Neutron Radius of ${}^{208}\mathrm{Pb}$” In _Phys. Rev. Lett._ 86 American Physical Society, 2001, pp. 5647–5650 DOI: 10.1103/PhysRevLett.86.5647 * [4] A. Krasznahorkay et al. “Neutron-skin thickness in neutron-rich isotopes” In _Nucl. Phys. A_ 731, 2004, pp. 224–234 DOI: https://doi.org/10.1016/j.nuclphysa.2003.11.034 * [5] B. Kłos et al. “Neutron density distributions from antiprotonic ${}^{208}\mathrm{Pb}$ and ${}^{209}\mathrm{Bi}$ atoms” In _Phys. Rev. C_ 76 American Physical Society, 2007, pp. 014311 DOI: 10.1103/PhysRevC.76.014311 * [6] J. Zenihiro et al. “Neutron density distributions of ${}^{204,206,208}\mathrm{Pb}$ deduced via proton elastic scattering at ${E}_{p}=295$ MeV” In _Phys. Rev. C_ 82 American Physical Society, 2010, pp. 044611 DOI: 10.1103/PhysRevC.82.044611 * [7] C.. Tarbert et al. “Neutron Skin of ${}^{208}\mathrm{Pb}$ from Coherent Pion Photoproduction” In _Phys. Rev. Lett._ 112 American Physical Society, 2014, pp. 242502 DOI: 10.1103/PhysRevLett.112.242502 * [8] D. Adhikari et al. “Accurate Determination of the Neutron Skin Thickness of ${}^{208}\mathrm{Pb}$ through Parity-Violation in Electron Scattering” In _Phys. Rev. Lett._ 126 American Physical Society, 2021, pp. 172502 DOI: 10.1103/PhysRevLett.126.172502 * [9] R.. Garcia Ruiz et al. “Unexpectedly large charge radii of neutron-rich calcium isotopes” In _Nat. Phys._ 12.6, 2016, pp. 594–598 DOI: 10.1038/nphys3645 * [10] Tomotsugu Wakasa et al. “Neutron-skin values and matter and neutron radii determined from reaction cross sections of proton scattering on ${}^{12}\mathrm{C}$, ${}^{40,48}\mathrm{Ca}$, ${}^{58}\mathrm{Ni}$, and ${}^{208}\mathrm{Pb}$” In _Phys. Rev. C_ 107 American Physical Society, 2023, pp. 024608 DOI: 10.1103/PhysRevC.107.024608 * [11] M. Enciu et al. “Extended ${p}_{3/2}$ Neutron Orbital and the $N=32$ Shell Closure in ${}^{52}\mathrm{Ca}$” In _Phys. Rev. Lett._ 129 American Physical Society, 2022, pp. 262501 DOI: 10.1103/PhysRevLett.129.262501 * [12] I. Tanihata et al. “Measurements of Interaction Cross Sections and Nuclear Radii in the Light $p$-Shell Region” In _Phys. Rev. Lett._ 55 American Physical Society, 1985, pp. 2676–2679 DOI: 10.1103/PhysRevLett.55.2676 * [13] P G Hansen, A S Jensen and B Jonson “Nuclear Halos” In _Annu. Rev. Nucl. Part S_ 45.1, 1995, pp. 591–634 DOI: 10.1146/annurev.ns.45.120195.003111 * [14] Björn Jonson “Light dripline nuclei” In _Phys. Rep._ 389.1, 2004, pp. 1–59 DOI: https://doi.org/10.1016/j.physrep.2003.07.004 * [15] Manju, Jagjit Singh, Shubhchintak and R. Chatterjee “Low-lying dipole strengths for probable p-wave one-neutron halos in the medium mass region” In _Eur. Phys. J. A_ 55.1, 2019, pp. 5 DOI: 10.1140/epja/i2019-12679-4 * [16] V. Rotival, K. Bennaceur and T. Duguet “Halo phenomenon in finite many-fermion systems: Atom-positron complexes and large-scale study of atomic nuclei” In _Phys. Rev. C_ 79 American Physical Society, 2009, pp. 054309 DOI: 10.1103/PhysRevC.79.054309 * [17] Zhongzhou Ren, Baoqiu Chen, Zhongyu Ma and Gongou Xu “One-proton halo in ${}^{26}\mathrm{P}$ and two-proton halo in ${}^{27}\mathrm{S}$” In _Phys. Rev. C_ 53 American Physical Society, 1996, pp. R572–R575 DOI: 10.1103/PhysRevC.53.R572 * [18] Michiharu Wada and Yasunori Yamazaki “Technical developments toward antiprotonic atoms for nuclear structure studies of radioactive nuclei” Low Energy Antiproton Physics (LEAP’03) In _Nucl. Instrum. Methods Phys. Res., Sect. B_ 214, 2004, pp. 196–200 DOI: https://doi.org/10.1016/j.nimb.2003.08.019 * [19] T. Aumann et al. “PUMA, antiProton unstable matter annihilation - PUMA collaboration” In _Eur. Phys. J. A_ 58.5, 2022, pp. 88 DOI: 10.1140/epja/s10050-022-00713-x * [20] M. Leon and R. Seki “Determination of the neutron halo from antiproton absorption” In _Phys. Lett. B_ 48.3, 1974, pp. 173–175 DOI: https://doi.org/10.1016/0370-2693(74)90001-X * [21] A.S. Iljinov, V.I. Nazaruk and S.E. Chigrinov “Nuclear absorption of stopped antiprotons: Multipion-nucleus interactions” In _Nucl. Phys. A_ 382.3, 1982, pp. 378–400 DOI: https://doi.org/10.1016/0375-9474(82)90352-9 * [22] M. Schlaich et al. “A multi-reflection time-of-flight mass spectrometer for the offline ion source of the PUMA experiment” In _Int. J. Mass Spectrom._ 495, 2024, pp. 117166 DOI: https://doi.org/10.1016/j.ijms.2023.117166 * [23] María J G Borge and Klaus Blaum “Focus on Exotic Beams at ISOLDE: A Laboratory Portrait” In _J. Phys. G: Nucl. Part. Phys._ 45.1 IOP Publishing, 2017, pp. 010301 DOI: 10.1088/1361-6471/aa990f * [24] Wolfgang Bartmann et al. “The ELENA facility” In _Philos. Trans. R. Soc., A_ 376.2116, 2018, pp. 20170266 DOI: 10.1098/rsta.2017.0266 * [25] Stephan Maury et al. “ELENA: the extra low energy anti-proton facility at CERN” In _Hyperfine Interact._ 229.1, 2014, pp. 105–115 DOI: https://doi.org/10.1016/0370-2693(74)90001-X * [26] Laurette Ponce et al. “ELENA - From Commissioning to Operation” In _JACoW IPAC_ 2022, 2022, pp. 2391–2394 DOI: 10.18429/JACoW-IPAC2022-THOXGD1 * [27] H. Kalinowsky “Deceleration of antiprotons from MeV to keV energies” In _Hyperfine Interact._ 76.1, 1993, pp. 73–80 DOI: 10.1007/BF02316707 * [28] C. Amole et al. “The ALPHA antihydrogen trapping apparatus” In _Nucl. Instrum. Methods Phys. Res., Sect. A_ 735, 2014, pp. 319–340 DOI: https://doi.org/10.1016/j.nima.2013.09.043 * [29] M. Tajima et al. “Antiproton beams with low energy spread for antihydrogen production” In _J. Instrum._ 14.05, 2019, pp. P05009 DOI: 10.1088/1748-0221/14/05/P05009 * [30] A. Husson et al. “A pulsed high-voltage decelerator system to deliver low-energy antiprotons” In _Nucl. Instrum. Methods Phys. Res., Sect. A_ 1002, 2021, pp. 165245 DOI: https://doi.org/10.1016/j.nima.2021.165245 * [31] Claude Amsler et al. “Pulsed production of antihydrogen” In _Commun. Phys._ 4.1, 2021, pp. 19 DOI: 10.1038/s42005-020-00494-z * [32] B.. Latacz et al. “Ultra-thin polymer foil cryogenic window for antiproton deceleration and storage” In _Rev. Sci. Instrum._ 94.10, 2023, pp. 103310 DOI: 10.1063/5.0167262 * [33] K. Nordlund, M. Hori and D. Sundholm “Large nuclear scattering effects in antiproton transmission through polymer and metal-coated foils” In _Phys. Rev. A_ 106 American Physical Society, 2022, pp. 012803 DOI: 10.1103/PhysRevA.106.012803 * [34] N. Kuroda et al. “Confinement of a Large Number of Antiprotons and Production of an Ultraslow Antiproton Beam” In _Phys. Rev. Lett._ 94 American Physical Society, 2005, pp. 023401 DOI: 10.1103/PhysRevLett.94.023401 * [35] S.S. Fabbri and W. Bertsche “Optimization of Antiproton Capture for Antihydrogen Creation in the ALPHA Experiment” https://doi.org/10.18429/JACoW-IBIC2019-WEPP040 In _Proc. IBIC’19_ , International Beam Instrumentation Conferenc 8 JACoW Publishing, Geneva, Switzerland, 2019, pp. 637–641 DOI: 10.18429/JACoW-IBIC2019-WEPP040 * [36] F Herfurth et al. “A linear radiofrequency ion trap for accumulation, bunching, and emittance improvement of radioactive ion beams” In _Nucl. Instrum. Methods Phys. Res., Sect. A_ 469.2, 2001, pp. 254–275 DOI: https://doi.org/10.1016/S0168-9002(01)00168-1 * [37] S. Coeck et al. “A pulsed drift cavity to capture 30keV ion bunches at ground potential” In _Nucl. Instrum. Methods Phys. Res., Sect. A_ 572.2, 2007, pp. 585–595 DOI: https://doi.org/10.1016/j.nima.2006.11.054 * [38] J. Grund et al. “First online operation of TRIGA-TRAP” In _Nucl. Instrum. Methods Phys. Res., Sect. A_ 972, 2020, pp. 164013 DOI: https://doi.org/10.1016/j.nima.2020.164013 * [39] G. Savard et al. “A new cooling technique for heavy ions in a Penning trap” In _Phys. Lett. A_ 158.5, 1991, pp. 247–252 DOI: https://doi.org/10.1016/0375-9601(91)91008-2 * [40] P. Pérez et al. “The GBAR antimatter gravity experiment” In _Hyperfine Interact._ 233.1, 2015, pp. 21–27 DOI: 10.1007/s10751-015-1154-8 * [41] M.A. Fraser, D. Barna, W. Bartmann and R. Ostojić “Beam Dynamics Studies of the ELENA Electrostatic Transfer Lines” https://doi.org/10.18429/JACoW-IPAC2015-MOPJE044 In _Proc. IPAC’15, Richmond, VA, USA, May 3-8, 2015_ , International Particle Accelerator Conference 6 Geneva, Switzerland: JACoW, 2015, pp. 385–388 DOI: https://doi.org/10.18429/JACoW-IPAC2015-MOPJE044 * [42] M Martini and H Schönauer “Emittance measurements in the CERN PS complex”, 1997 * [43] M. McLean, J. Cenede, M. Hori and G. Tranquille “Commissioning of the SEM-Grid Monitors for ELENA” https://doi.org/10.18429/JACoW-IBIC2021-TUPP14 In _Proc. IBIC’21_ , International Beam Instrumentation Conference 10 JACoW Publishing, Geneva, Switzerland, 2021, pp. 223–226 DOI: 10.18429/JACoW-IBIC2021-TUPP14 * [44] C Benvenuti “Extreme Vacua: Achievements and Expectations” In _Phys. Scr._ 1988.T22, 1988, pp. 48 DOI: 10.1088/0031-8949/1988/T22/006 * [45] Katharina Battes, Christian Day and Volker Hauer “Systematic study of the outgassing behavior of different ceramic materials” In _J. Vac. Sci. Technol., B_ 39.3, 2021, pp. 034202 DOI: 10.1116/6.0000954 * [46] P. Chiggiato and P. Costa Pinto “Ti–Zr–V non-evaporable getter films: From development to large scale production for the Large Hadron Collider” Proceedings of the Eighth International Conference on Atomically Controlled Surfaces, Interfaces and Nanostructures and the Thirteenth International Congress on Thin Films In _Thin Solid Films_ 515.2, 2006, pp. 382–388 DOI: https://doi.org/10.1016/j.tsf.2005.12.218 * [47] D W Williams and W T Williams “Effect of electrode surface finish on electrical breakdown in vacuum” In _J. Phys. D: Appl. Phys._ 5.10, 1972, pp. 1845 DOI: 10.1088/0022-3727/5/10/314
# Measuring Internet Routing from the Most Valuable Points Thomas Alfroy11footnotemark: 1, Thomas Holterbach11footnotemark: 1, Thomas Krenc22footnotemark: 2, KC Claffy22footnotemark: 2, Cristel Pelsser33footnotemark: 3 11footnotemark: 1 University of Strasbourg, 22footnotemark: 2 CAIDA / UC San Diego, 33footnotemark: 3 UCLouvain https://bgproutes.io ###### Abstract While the increasing number of Vantage Points (VPs) in RIPE RIS and RouteViews improves our understanding of the Internet, the quadratically increasing volume of collected data poses a challenge to the scientific and operational use of the data. The design and implementation of BGP and BGP data collection systems lead to data archives with enormous redundancy, as there is substantial overlap in announced routes across many different VPs. Researchers thus often resort to arbitrary sampling of the data, which we demonstrate comes at a cost to the accuracy and coverage of previous works. The continued growth of the Internet, and of these collection systems, exacerbates this cost. The community needs a better approach to managing and using these data archives. We propose MVP, a system that scores VPs according to their level of redundancy with other VPs, allowing more informed sampling of these data archives. Our challenge is that the degree of redundancy between two updates depends on how we define redundancy, which in turn depends on the analysis objective. Our key contribution is a general framework and associated algorithms to assess redundancy between VP observations. We quantify the benefit of our approach for four canonical BGP routing analyses: AS relationship inference, AS rank computation, hijack detection, and routing detour detection. MVP improves the coverage or accuracy (or both) of all these analyses while processing the same volume of data. ## 1 Introduction Routing information services such as RIPE RIS [38] and RouteViews (RV) [51] continuously collect and publish data from more than 2500 Vantage Points (VPs), each of which is a BGP router that exports its best routes to the collection platform. These data collection systems are critical to scientific as well as operational analyses of the global Internet infrastructure. But these systems face a cost-benefit trade-off [2]. The information-hiding character of BGP means that improving the visibility of the Internet routing system requires cultivating many peering relationships with operators willing to contribute VPs to the platform. However, deployment of new VPs amplifies the data management requirements caused by the growth of the Internet itself: the number of unique IP prefixes (e.g., due to de-aggregation or new assignments) constantly grows [12], as well as the number of unique ASes and links between them. Even with a constant number of VPs, the volume of routing data inevitably increases, contributing to a quadratic increase of observed updates over time (Fig. 1(e)). The situation presents a challenge for users, who often cannot or do not want to process terabytes of (redundant) data. Users often resort to sampling the data in arbitrary ways, such as grabbing all VPs on a single collector. We design and implement a framework to optimize the use of these data collection systems, which will also lower the barrier to their use in lower- resourced circumstances. Our design relies on the principle of redundancy in BGP data, a delicate concept since even two identical updates from two different VPs may not be redundant (depending on the use case). We take a deep dive into a context-specific framework for quantifying redundancy in BGP data, grounded in operational principles and research use cases. Our resulting system identifies a set of VPs whose exported routes collectively exhibit a low level of redundancy—–enabling users to prioritize the processing of the most valuable BGP updates. 1 2 3456 (a) 1 2 3456 (b) 1 2 3456 (c) 1 2 3456 (d) Figure 1: Combining local views can help to map the AS topology. Gray links are not visible from routes collected by VPs ( ). (e) Growth in VPs (f) Number of updates per VP and per hour. (g) Total number of updates per hour. Figure 2: The number of VPs increases over time and so does the number of collected updates. Both RIS and RV are considered in Fig. 1(f) and 1(g). _Contributions._ We make the following contributions. * • We perform a comprehensive analysis based on simulations and a survey that demonstrates the cost-benefit tradeoff of setting up new VPs, and the value of strategically selecting them to analyze Internet routing. We show that current approaches used by researchers to select VPs are largely unoptimized, sacrificing coverage and accuracy of a wide range of measurement studies and tools (§2-§3). * • We characterize redundancy between updates collected by different VPs. We explore different definitions of redundancy and find that optimizing our algorithms for a given definition leads to a undesirable overfitting effect (§4-§5). * • We design a system, MVP, that returns a list of the “most valuable” VPs, i.e., those that enable users to minimize data redundancy (regardless of how we define it) and prioritize valuable route updates. MVP relies on new data- driven algorithms that quantify redundancy between VPs based on the four main BGP attributes (time, prefix, AS path, and communities) while being robust against typical biases observed in the Internet routing ecosystem (§6-§7). * • We run MVP as a service at https://bgproutes.io. We benchmark MVP and show that it optimizes (without overfitting) the tradeoff between the volume of data used and its utility for many objectives (§8-§9). _Impact on scientific measurement studies._ The value of MVP is its wide impact. Besides enabling a more systematic sampling of the RIS and RV data archives, it can consistently, and at no cost for users, improve the accuracy and coverage of measurement studies as well as monitoring tools fueled by BGP routes collected by RIS and RV. To measure the impact of MVP, we replicated the algorithms used in four studies/tools and used MVP to select the VPs from which they process BGP routes. In all four cases, using MVP improved the accuracy and coverage while processing the same data volume. We inferred more AS relationships (+15%), fixed errors in the AS rank dataset, observed more routing detours (+44%) while characterizing them more accurately, and inferred more forged-origin hijacks (+35%) with $\approx$4$\times$ less incorrect inferences (i.e., false positives). ## 2 Background RIPE’s Routing Information Service (RIS) [38] and RouteViews (RV) [51] are two widely-used platforms that collect BGP routes and make them available to the community. These platforms use BGP speakers (a.k.a. collectors) to peer with BGP routers in order to collect routes exported by those routers. We call vantage points (VPs) the BGP routers that export their routes to a collection platform. As of May 2023, 32% of the RIS and RV VPs [42, 33] are full feeders, i.e., they send a route for roughly all of the announced IP prefixes on the Internet ($\approx$941k prefixes [12]). A BGP route mainly carries routing information in four of its attributes [36]: _(i)_ the timestamp at which the route was received, _(ii)_ the IP (v4 or v6) prefix that the route announces, _(iii)_ the AS path used to reach that prefix, and _(iv)_ a set of BGP communities. Among other uses, researchers leverage the timestamp to find transient paths [29], the prefix to detect hijacks [44], the AS paths to infer AS relationships [31], and the communities to measure unnecessary BGP traffic [28]. Each VP provides its local view, i.e., only the BGP routes it observes. Fig. 2 illustrates the effect of combining local views for inferring the AS topology from the AS paths in BGP routes. In Fig. 2, every AS runs a single BGP router, owns one prefix, and announces it in BGP. We configure routing policies based on the Gao-Rexford model [21], i.e., routing paths follow a valley-free pattern. Straight (resp. dashed) lines are customer-to-provider (resp. peer- to-peer) links. With the local view of 1, one can infer all the AS links but the two peering links 3 4 and 5 6 (Fig. 1(a)). Combining the local views of 1 and 2 does not help to discover more links (Fig. 1(b)). With the local view of 5, one can infer all the AS links but the two customer-to-provider links 2 4, 4 6 (Fig. 1(c)). Combining the local views of 5 and 6 enables discovery of the full topology (Fig. 1(d)). However, observe that this last scenario is unlikely in practice as the location of the VPs is skewed with many more VPs present in highly-connected or central (e.g., Tier1) ASes [45]. Observe also that VPs can have a redundant view over the AS topology, e.g., the two VPs in Fig. 1(b) observe the same set of links. By May 2023, RIS had 1526 VPs and RV had 1071 VPs, and their number keeps increasing (Fig. 1(e)). Users can download BGP routes exported by these platforms at the granularity of the VP (with some limitations [41]) or the collector. Users can download a RIB dump, i.e., a snapshot of the BGP routes seen by a VP at a particular time, which (in Jan. 2023) yielded $\approx$941k routes for a full feeder. Alternatively, users may download every single BGP update observed by the VPs over time (e.g., using [36]), which currently results in $\approx$18K updates per hour (median in May 2023) for a single VP (Fig. 1(f)), and billions of updates per day for all RIS and RV VPs (Fig. 1(g)). ## 3 Problem Deploying more VPs expands the visibility of the routing system (§3.1), but also increases collected data volumes raising barriers to its use (§3.2). We survey researchers and find that they resort to unoptimized sampling, which they acknowledge can negatively impact the quality of their results (§3.3). ### 3.1 More VPs improves data completeness A tiny fraction (1.3%) of the 74k ASes participating in the global routing system [12] host a VP. This fraction remains low (8.4%) even when focusing on the 11441 transit ASes (i.e., those with at least one customer). While we cannot know how much additional topology we might observe from VPs that do not peer with the public collection systems, we can estimate this gap using simulations of topologies whose statistical parameters match those of the known global Internet. _Methodology._ We created a mini-Internet with 600 ASes, each running a single BGP router. We generated the AS topology using the Hyperbolic Graph Generator [3]. We set the average node degree to 6.1, which results in a comparable degree of connectivity (a.k.a. Beta index) to the one observed in CAIDA’s AS relationship dataset from December 2022 [16], and use as the degree distribution a power law with exponent 2.1 (as in [3]). We defined the AS relationships as follows. The three ASes with the highest degree are Tier1 ASes and are fully meshed. ASes directly connected to a Tier1 are Tier2s. ASes directly connected to a Tier2 but not to a Tier1 are Tier3s, etc. Two connected ASes have a peer-to-peer (p2p) relationship if they are on the same level, and a customer-to-provider (c2p) relationship if not. The routing policies follow the Gao-Rexford model [21]. Figure 3: Simulations of a mini Internet with 600 ASes. We make two key observations: _(i)_ deploying more VPs helps to reveal more AS links, and _(ii)_ arbitrarily selecting VPs performs poorly compared to selecting them with greedy specific (a best-case approximation). The line in a box depicts the median value; the whiskers show the 5 and the 95th percentile. Fig. 3 shows the proportion of observed AS links as a function of the number of ASes hosting a VP. We consider three VP deployment strategies: _(i)_ random, which randomly deploys VPs across all the ASes; _(ii)_ distance-based, which aims to maximize the AS-level distance between the deployed VPs; and _(iii)_ greedy specific, which approximates the best case for topology discovery using a greedy approach. We ran every selection strategy twenty times (with different random seeds). We computed the proportion of observed links and show separately the p2p and c2p links in Fig. 3. _Conclusions._ Although we take the results with a grain of salt because the topology differs (but exhibits similar patterns) from the visible portion of the actual (unknown) AS topology, we tentatively draw the following four conclusions. _(I)_ As expected, for a given VP deployment strategy, more VPs often lead to more links observed; all links are observed only when all ASes host a VP. _(II)_ P2p links are harder to observe than c2p links. We find that p2p links are more visible from VPs at the edge. This result is consistent with the fact that p2p links are generally not advertised upwards in the Internet hierarchy when routing policies follow the Gao-Rexford model. _(III)_ The distance-based deployment strategy performs poorly (even worse than random) because it overprioritizes isolated VPs at the edge over some other important VPs in the core. _(IV)_ When 1.3% of ASes host a VP (same proportion as current RIS and RV VPs), only $\approx$5% of the p2p links are seen when using the random deployment strategy. _Confirmation with real (but private) data._ We contacted a private BGP data provider (bgp.tools) that collects BGP routes from $\approx$1000 routers and compared the set of AS links observed from these private feeds against the set of AS links observed by RIS and RV VPs (in September 2023). We find that the private data provider saw 192k AS links that none of the RIS and RV VPs observed, and vice versa, RIS and RV VPs observed 401k links that the private data provider did not observe. In either case, the lack of VPs leads to missing routing information. We can thus expect—and hope for—the number of VPs to keep increasing. ### 3.2 BGP data management is challenging Deploying more VPs generates more data as each of them collects BGP updates. Moreover, new IP prefixes advertised in BGP (see [12]) increase the volume of data collected by every VP as it triggers the propagation of new BGP routes that many VPs (e.g., the full feeders) observe and send to the collection platforms. The compound effect—more VPs (Fig. 1(e)) and more updates per VP (Fig. 1(f))—yields a quadratic increase in updates reaching the collection platforms (Fig. 1(g)), which challenges both users and data providers [2]. Although several tools can speed up data processing [36, 7, 5], many measurement studies and monitoring tools use only a sample of data collected by RIS and RV, either using only a subset of the VPs or a short time window, or both111We purposively do not cite any paper to preserve the anonymity of the respondents of our survey.. While authors do not typically explain why they do not use all the data, the sampling suggests two (inter-related) explanations: authors believe the sample is representative and sufficiently complete; and/or the data volume is not worth trying to manage. We confirm these explanations with a survey that we conducted involving authors of eleven research papers. _Methodology of our survey._ We classified eleven BGP-based studies from top conferences222SIGCOMM, NSDI, S&P, USENIX Security, NDSS and IMC. into two categories based on how they used BGP data.333A paper may be in both categories. Nine papers used all routes collected from a subset of the VPs (category $C_{1}$); six papers used a short time frame ($C_{2}$). For each paper, we asked authors questions regarding their use of BGP data: whether data volume limited their work, how and why they sampled BGP data sources, their understanding of the impact on the quality of their results, and if they would do things differently if they had more resources or time. We did not receive answers from the authors of three papers. Thus, we have seven respondents in $C_{1}$ and five in $C_{2}$. We summarize the results here; details of the survey are in an appendix (§A). _The volume of BGP data to process is often a limiting factor._ Seven (of eight) respondents found the BGP data expensive to process. For three respondents in $C_{1}$, processing time motivated them to use only a subset of the VPs; three respondents in $C_{2}$ considered the processing time when choosing a measurement interval. Even a respondent who used a Spark cluster found it inhibitively time-consuming to process the BGP data. _Respondents in $C_{1}$ selected VPs in an unoptimized fashion._ One respondent picked geographically distant BGP collectors. Our experiments (Fig. 3) and evaluation (§9) show that this strategy, while intuitive, often fails to optimize for any given metric (e.g., coverage). Other respondents said they chose VPs randomly, or those with the highest number of prefixes. Another responded to have unintentionally discarded some VPs, leaving an arbitrarily selected set in the study. Two respondents did not remember how they selected VPs. ### 3.3 Unoptimized sampling negatively impacts the quality of the results We show the negative effects of an unoptimized sampling using our controlled simulations as well as our survey. _Selecting VPs arbitrarily performs poorly._ Our mini-Internet simulation (§3.1) showed that arbitrary VP selection strategies perform significantly worse than greedy specific (a best-case approximation) when the goal is to map the AS topology. For instance, randomly selecting 20 VPs reveals 12% of the p2p links compared to 56% when selecting them using greedy specific—a 4.7$\times$ improvement factor that we highlight in Fig. 3. Our evaluation reveals that this performance gap between using an arbitrary VPs selection strategy and a best-case approach also exists for various other metrics, e.g., hijacks or transient paths detection (§9). _Six respondents in $C_{1}$ acknowledged that using more VPs would improve the quality of their analysis._ The last respondent was not sure, given the potential redundancy in the data sources (which he did not analyze). Two of the six believed it would not significantly change the conclusion of their measurement studies (e.g., one said that it could help to pinpoint corner cases). However, six of the seven authors in $C_{1}$ affirmed that they would have used more VPs if they had more resources and time. _All five respondents in $C_{2}$ said that extending the duration of their study would improve the quality of their results._ One respondent thought the gain would not be significant; another said it could help detect rare routing events. All respondents in $C_{2}$ would have extended the duration of their observation window given more time and resources. We experimentally confirm in §9.2.3 that extending the timeframe of analysis improves the quality of its results with a case study on routing detour characterization [46]. ## 4 Opportunity to Optimize Sampling We propose a systematic framework to characterize redundancy across BGP routes collected by the VPs. We use the term redundant to refer to updates with similar (or identical, depending on the redundancy definition) attribute values (see definitions below). Thus, two redundant VPs, i.e., that observe redundant routes, likely provide similar views over routing events such as hijacks, traffic engineering, etc. _Methodology._ We characterize redundancy between pairs of VPs by computing the proportion of redundant updates that they collect using three different, gradually stricter, definitions of update redundancy. We denote $U_{i}$ the set of updates observed by VP $i$. Consider a BGP update $u_{t,p}\in U_{i}$ with $t$ the time at which the route was observed and $p$ its prefix. ###### Definition 1 (prefix based) The update $u_{t_{1},p_{1}}\in U_{1}$ is redundant with the update $u_{t_{2},p_{2}}\in U_{2}$ if: * • $\lvert t_{1}-t_{2}\lvert<5$ minutes, and $p_{1}=p_{2}$. We chose 5 minutes because it is an approximation of the BGP convergence time [29]. This first definition might be appropriate to map prefixes with their origin AS. For our second definition, we denote $A_{i}(t,p)$ the set of AS links in the AS path of the most recent BGP route observed by VP $i$ for prefix $p$ at time $t$. ###### Definition 2 (prefix and as-path based) The update $u_{t_{1},p_{1}}$ $\in U_{1}$ is redundant with the update $u_{t_{2},p_{2}}\in U_{2}$ if: * • $\lvert t_{1}-t_{2}\lvert<5$ minutes, and $p_{1}=p_{2}$, and * • $A_{1}(t_{1},p_{1})\setminus A_{1}(t_{1}-\epsilon,p_{1})\subset A_{2}(t_{2},p_{2})\setminus A_{2}(t_{2}-\epsilon,p_{2})$. The second condition checks whether the changes (operator $\setminus$) in the AS paths observed by VP $1$ for a given prefix are included (operator $\subset$) in the set of changes observed by VP $2$ for the same prefix. This second definition might be appropriate to detect new AS links or transient paths. Our third definition follows the same approach but adds BGP communities. We denote $C_{i}(t,p)$ the set of community values of the most recent BGP route observed by VP $i$ for prefix $p$ and at time $t$. ###### Definition 3 (prefix, as-path, and community-based) The update $u_{t_{1},p_{1}}\in U_{1}$ is redundant with update $u_{t_{2},p_{2}}\in U_{2}$ if: * • $\lvert t_{1}-t_{2}\lvert<5$ minutes, and $p_{1}=p_{2}$, and * • $A_{1}(t_{1},p_{1})\setminus A_{1}(t_{1}-\epsilon,p_{1})\subset A_{2}(t_{2},p_{2})\setminus A_{2}(t_{2}-\epsilon,p_{2})$, and * • $C_{1}(t_{1},p_{1})\setminus C_{1}(t_{1}-\epsilon,p_{1})\subset C_{2}(t_{2},p_{2})\setminus C_{2}(t_{2}-\epsilon,p_{2})$. We note that Def. 2 and 3 are asymmetric because, given two set $X$ and $Y$ of objects of same type, $X\subset Y\mathrel{{\ooalign{$\not\phantom{=}$\cr$\implies$}}}Y\subset X$. _Redundant pairs of VPs exist._ Fig. 4 (top row) shows the level of redundancy for the three definitions and between 100 VPs randomly selected and computed over the updates observed during two hours on August 1, 2022. Observe that we performed 30 random selections with different seeds and show the median case (in terms of redundant pairs of VPs). One cell in the matrix indicates the redundancy of the VP on the ordinate with the VP on the abscissa. We define the redundancy between VP $1$ and VP $2$ as the proportion of updates observed by VP $1$ that are redundant with at least one update observed by VP $2$. For better visibility, we show the most redundant VPs at the top of the figures. Redundant pairs of VPs exist regardless of the redundancy definition used. Logically, the stricter the definition, the fewer redundant pairs of VPs. Fig. 4 (left) shows that the VPs can be highly redundant when they are selected randomly. For instance, with the loose Def. 1, we observe that 74 among the 100 randomly selected VPs have >50% of their updates that are redundant with the ones observed by two other VPs or more (23 for Def. 2 and 16 for Def. 3). We observe a similar redundancy level when considering only full feeders. Figure 4: Redundancy among a subset of 100 existing VPs selected using two different techniques for three increasingly stricter redundancy definitions. Randomly selecting VPs (top row) returns significantly more pairs of redundant VPs. ## 5 Main challenge: prevent overfitting Our design objective is a general framework that can accommodate different definitions of redundancy in selecting the set of least redundant VPs. However, optimizing selection for one objective is likely to overfit, leading to poor performance for other objectives. Thus, while the three definitions in §4 enable illustrating the redundancy across current VPs, none of them are used in the design of MVP. These definitions are too naive to accurately quantify redundancies between the VPs. We explore this risk of overfitting to a particular objective using a VPs selection strategy optimized for one objective: minimizing redundancy. This selection strategy, which we name greedy specific, iteratively selects (in a greedy fashion) the VP that minimizes the proportion of redundant updates across all the updates collected by the selected VPs. We implement three versions of it, one for each redundancy definition used in §4. Thus, greedy specific approximates an optimal VP selection when the goal is to minimize redundancy between VPs according to a specific definition of redundancy. _Greedy specific limits redundancy._ We select 100 VPs using greedy specific. Logically, the selected VPs are less redundant (see Fig. 4, bottom row) compared to the 100 VPs randomly selected. With the loose Def. 1, only 30 VPs have >50% of their updates redundant with ones observed by two other VPs or more. This number drops to 9 with Def. 2 and 5 with Def. 3. This result highlights that while VPs can be highly redundant, nonredundant pairs of VPs also exist. _Greedy specific overfits._ Greedy specific overfits because it optimizes one particular objective. Thus, it works well for this objective but not for the others. We confirm this overfitting effect in §9 where we benchmark greedy specific against MVP on various objectives and show that it performs poorly on objectives that it does not optimize. Consequently, one would need to design a greedy specific VPs selection for every possible definition of data redundancy—which is unpractical given that there is an infinite number of definitions. ## 6 Methodology Overview MVP samples BGP updates from RIS and RV at the VP granularity. Our method has four steps that we overview below. _Step 1 (§ 7.1): Select a large, unbiased set of BGP events that we use to gauge pairwise redundancy between VPs._ MVP evaluates the redundancy between two VPs based on a carefully selected set of non-global BGP events (i.e., AS path changes). Global events are typically seen by all VPs and have the same impact on every VP view, rendering them less discriminating for this purpose. We stratify our selection of sampled events across space and time to avoid bias. _Step 2 (§ 7.2): Characterize how VPs experience the selected events._ For every BGP event, MVP quantifies topological features [48] of the ASes involved as observed by each VP. These features embed information about the four attributes of a BGP update: time, prefix, AS path, and communities. _Step 3 (§ 7.3): Compute pairwise redundancy between VPs._ MVP computes the pairwise Euclidean distance in a $n$-dimensional space, where $n$ is the number of topological features times the number of events. VP pairs with similar feature values for many events are close in this space and thus likely redundant. MVP then computes the average Euclidean distances between each pair of VPs computed over different and nonoverlapping time periods. _Step 4 (§ 7.4): Sort and select the least redundant VPs._ MVP relies on a greedy algorithm that considers both data redundancy and its volume to build a set of the most valuable VPs. MVP first adds the VP with the lowest average Euclidean distance to all other VPs, and then greedily adds the VP that balances minimal redundancy with already selected VPs and minimal additional data volume that the VP brings. ## 7 Methodology Details In the following, we consider the set of VPs $V$ that includes all VPs from RIS and RV. We compute the RIB of VP $v$ at time $t$ using its last RIB dump before $t$ and subsequent updates until $t$. We use this RIB to construct and maintain the undirected weighted graph $G_{v}(t)=(N_{v}(t),E_{v}(t))$ from the AS paths of the best routes observed by $v$ at time $t$, with $N_{v}(t)$ the set of nodes and $E_{v}(t)\in N_{v}(t)*N_{v}(t)$ the set of AS links. The edges are undirected because two identical paths in opposite directions should not appear as nonredundant. Each edge in $E_{v}(t)$ has a weight in $\mathbb{Z^{+}}$ which is the number of routes in the RIB that includes this edge in their AS path. ### 7.1 Select BGP events to assess redundancy ID | Name | # of ASes | Avg.degree | Description ---|---|---|---|--- 1 | Stub | 63310 | 3 | ASes without customer 2 | Transit-1 | 10845 | 27 | Transit ASes with a customer cone size lower than the average 3 | Transit-2 | 704 | 267 | Transit ASes $\notin$ Transit-1 4 | HyperGiant | 15 | 1078 | Top 15 as defined in [8] 5 | Tier1 | 19 | 1817 | Tier1 in the CAIDA dataset [16] Table 1: MVP balances selected events across 5 AS types. _MVP uses local and partially visible new-AS-link events._ MVP focuses on BGP events that trigger a new AS link to appear in the path to reach prefix $p$ from different VPs. A new-AS-link event is a candidate event in $\mathcal{C}$ if at least two and fewer than half of the VPs begin to use the same new AS link to reach the same prefix within a 10-minute window (to accommodate typical BGP convergence and path exploration delays [29, 35]). Since the aim of MVP is to find data unique to individual VPs, we exclude global events (i.e., seen by most VPs) to focus on local events. _MVP avoids biases across time and location._ From candidate set $\mathcal{C}$, MVP builds the final set of events $\mathcal{E}$ by selecting 15 events in $500\text{\,}\mathrm{d}$ifferent and nonoverlapping 10-minute time periods. Adding more periods does not affect significantly the results. MVP samples time periods randomly within a one-month timeframe to avoid mis- inferring one larger event (e.g., a route leak that continuously generates new links for multiple hours) as several smaller AS-link-level events. Inspired by previous approaches to mitigate the risk of over-sampling core or stub (edge) ASes [45, 37], our approach classifies ASes into five categories (Table 1) and selects an equal number of new-AS-link events for every pair of AS categories. We distinguish two classes of transit providers by customer cone size (Transit-1 and -2) since they have different topological properties. If an AS belongs to more than one category, we classify it in the category with the highest ID. ASes classified in a lower row of Table 1 have a higher degree, and there are more low-degree ASes than high-degree ASes. Fig. 5 shows the proportion of selected events for each of the 15 pairs of AS category (the matrixes are symmetric) and for 7500 events selected in January 2023 using two schemes: balanced and random. The random selection (Fig. 5(b)) selects many more events involving Transit-2 ASes (69%) than hypergiants (11%), while our balanced selection scheme mitigates biases by selecting the same number of links in every category (Fig. 5(a)). For each time period, MVP selects one event in each of the 15 pairs of AS, yielding $15*500=7500$ events ($|\mathcal{E}|=7500$) for use in the next step. (a) Balanced selection. (b) Random selection. Figure 5: MVP selects the new-AS-link events using a balanced selection scheme that reduces bias (Fig. 5(a) vs. Fig. 5(b)). The x- and y-axis are the five categories of ASes (see Table 1). ### 7.2 Quantifying the observation of the VPs _MVP considers the four main BGP attributes._ MVP computes topological features on the graphs $G_{v}(t)$ for all VPs. The combination of these topological features prevents overfitting as the graphs on which they are computed embed information about the four main BGP attributes (§2). More concretely, the graphs $G_{v}(t)$ embed information about (i) the time as the graph is updated over time, (ii) the AS path as it is used to build the AS graph, (iii) the prefixes as they are used to weight every edge on the graph, and (iv) the community values as they are strongly correlated with the AS path. We confirm this correlation by downloading the first RIBs of Jan. 2023 for all VPs and analyzing the correlation between the AS path and the set of BGP communities. We find that two identical AS paths share the exact same set of BGP communities in 93% of the cases. We thus do not embed more information about BGP communities because many of them encode local traffic engineering decisions [17] that could lead to MVP overfitting. We validate this design choice in §9.1. Type | Categorie | Name | Weighted | Index ---|---|---|---|--- Node-based | Centrality Metrics | Closeness centrality | $\checkmark$ | 0 Harmonic centrality | $\checkmark$ | 1 Neighborhood Richness | Average neighbor degree | $\checkmark$ | 2 Eccentricity | $\checkmark$ | 3 Topological Pattern | Number of Triangles | $\times$ | 4 Clustering | $\checkmark$ | 5 Pair-based | Closeness Metrics | Jaccard | $\times$ | 6 Adamic Adar | $\times$ | 7 Preferential attachment | $\times$ | 8 Table 2: Node-based and pair-based features used by MVP. _MVP uses 15 diverse topological features (Table 2)._ MVP computes topological features (extracted from literature [20]) that are either node-based or link- based. Node-based features are computed for the two ends of a new AS link, while link-based are computed for the new AS link. MVP uses six node-based features that we classify into three categories. The first one quantifies how central and connected a node is in the graph; the second quantifies how connected are the neighboring nodes; and the third quantifies the topological patterns (e.g., triangles) that include the node. We classify the three pair- based features into a single category that measures how close two nodes are based on their neighboring nodes. Five features rely on edge weights. We omit other topological features as they are redundant with the selected ones. _MVP computes the value of the features for each VP and selected event._ Consider the event $e\in\mathcal{E}$ that is the appearance of the AS link $(e_{AS1},e_{AS2})$ at time $e_{t}$, and the VP $v\in V$. Computation of the feature values depends on the feature type. We denote $F_{n}$ (resp. $F_{p}$) the set of node-based (resp. pair-based) features and show how MVP computes the value of these two types of features for event $e$ and VP $v$. Node-based features: Consider feature $f_{i}\in F_{n}$ and $f_{i}(x,G_{v}(t))$ its value for node $x$ on the graph $G_{v}(t)$, with $i$ the feature index in Table 2. MVP computes the following 12-dimensional feature vector. $\displaystyle T_{node\\_based}(v,e)=[f_{0}(e_{AS1},G_{v}(e_{t})),f_{0}(e_{AS2},G_{v}(e_{t})),$ $\displaystyle\dots,f_{5}(e_{AS1},G_{v}(e_{t})),f_{5}(e_{AS2},G_{v}(e_{t}))]$ Pair-based features: Consider feature $f_{i}\in F_{p}$ and $f_{i}(x_{1},x_{2},G_{v}(t))$ its value for the node pair $(x_{1},x_{2})$ on the graph $G_{v}(t)$, with $i$ the feature index in Table 2. MVP computes the following 3-dimensional feature vector. $\displaystyle T_{pair\\_based}(v,e)=[f_{6}(e_{AS1},e_{AS2},G_{v}(e_{t})),$ $\displaystyle\dots,f_{8}(e_{AS1},e_{AS2},G_{v}(e_{t}))]$ The final feature vector used by MVP is $T(v,e)$, an 15-dimensional vector that is the concatenation (denoted $\oplus$) of the node- and pair-based features. $\displaystyle T(v,e)=T_{node\\_based}(v,e)\oplus T_{pair\\_based}(v,e)$ ### 7.3 Redundancy scoring MVP computes pairwise redundancy between VPs in the following four steps. _Step 1: Concatenate the feature vectors._ MVP first concatenates the computed topological feature vectors (15 features) for all the events selected in the same time period (15 events). We denote $\mathcal{E}_{p}$ the events selected in the p-th time period ($|\mathcal{E}_{p}|=15$), with $0\leq p<500$, and denote $e_{p,i}\in\mathcal{E}_{p}$ the i-th selected event in the p-th time period. $F(v,p)$ is the concatenated feature vector for VP $v$ and the events $\mathcal{E}_{p}$, which has $15*15=225$ dimensions and which MVP calculates as: $\displaystyle F(v,p)=T(v,e_{p,0})\oplus T(v,e_{p,1})\oplus\dots\oplus T(v,e_{p,14})$ _Step 2: Normalize concatenated feature vectors._ MVP normalizes the data for each time period using the feature matrix $\mathcal{M}(p)$ that includes the concatenated feature vectors for all VPs (rows) and events (columns) in period $p$. $\mathcal{M}(p)=\begin{bmatrix}F(v_{0},p)\\\ \dots\\\ F(v_{|V|},p)\par\par\end{bmatrix}$ MVP normalizes (operation $\bigtriangledown$) the matrix $\mathcal{M}(p)$ column-wise using a standard scaler that transforms every column such that its average is zero and its standard deviation is one. _Step 3: Compute Euclidean distance between VPs._ MVP uses the normalized matrix $\bigtriangledown(\mathcal{M}(p))$ to compute the Euclidean distance between every pair of VPs and for all events in the time period $p$ (operation $\diamond$). We denote $\bigtriangledown(\mathcal{M}(p))_{x}$ the x-th row in the matrix $\bigtriangledown(\mathcal{M}(p))$ and $\bigtriangledown(\mathcal{M}(p))_{x,i}$ its value at index $i$ (i.e., the i-th column). We define the Euclidean distance between the n-th VP $v_{n}$ and the m-th VP $v_{m}$ over the selected events in the time period $p$ as follows. $\displaystyle\diamond(v_{n},v_{m},p)=\sum_{i=0}^{225}(\bigtriangledown(\mathcal{M}(p))_{n,i}-\bigtriangledown(\mathcal{M}(p))_{m,i})^{2}$ _Step 4: Compute the average distance over all time periods._ The redundancy score $\mathcal{R}(v_{n},v_{m})$ between two VPs $v_{n}$ and $v_{m}$ relates to the normalized average Euclidean distance between them over the 500 time periods, computed as: $\displaystyle\mathcal{R}(v_{n},v_{m})=1-\coprod((\sum^{500}_{p=0}{\diamond(v_{n},v_{m},p)})*\frac{1}{500})$ The operator $\coprod$ applies a min-max scaler so that scores are between 0 and 1, with 1 meaning the most redundant pair of VPs and 0 the less redundant pair of VPs. MVP thus computes and returns a redundancy score for every pair of VPs. ### 7.4 Generating a set of VPs We now explain how MVP generates a set of VPs $\mathcal{O}$ that minimizes the proportion of redundant information collected. MVP initializes the set $\mathcal{O}$ with the most redundant VP, i.e., the one with the lowest sum of Euclidean distances to all the other VPs. This design choice allows the redundant part of the BGP data (e.g., c2p links) to be visible by the first selected VP. At every following iteration, MVP builds a candidate set of VPs $\mathcal{K}$ that contains the unselected VPs exhibiting the lowest maximum redundancy score. The maximum redundancy score $P$ measures the maximum redundancy between a VP $v$ and the set of VPs $\mathcal{O}$ and is defined as follows. $P(\mathcal{O},v)=\max(\mathcal{R}(v,v_{i}),\forall v_{i}\in\mathcal{O})$ MVP adds in $\mathcal{K}$ the $\alpha=$ 25% of the nonselected VPs that exhibit the lowest maximum redundancy score. MVP then adds in set $\mathcal{O}$ the VP that is in the candidate set $\mathcal{K}$ and that collects the lowest volume of data compared to the other VPs in $\mathcal{K}$. MVP estimates the volume of data collected by the VPs by counting the number of updates that they received over 365 one-hour periods, one randomly selected in each day of the year to align with the yearly update rate of MVP (§8). The $\alpha$ parameter allows tuning redundancy and volume knobs: a low $\alpha$ prioritizes low redundancy while a higher $\alpha$ prioritizes low resulting data volume. We found that $\alpha=$ 25% performs well in practical scenarios (we tested a range from 10% to 50%). ## 8 System functionalities MVP runs on a commodity server. Upon launch, it collects BGP routes from RIS and RV using BGPStream [36] and computes the redundancy between every pair of VPs at a yearly granularity, which is enough given that redundancies between VPs remain stable over time (see §9.3). MVP then takes as input a year and a volume of data and returns a set of VPs that generates a volume of data lower than the volume specified as input. MVP returns the redundancy scores calculated for every pair of VPs. Thus, users have the option to compute their own set of complementary VPs based on these redundancy scores and some additional constraints that they might have. This is useful when users want to include (or exclude) some VPs (regardless of how redundant they are), which will result in another set of VPs rather than the default set provided by MVP. For instance, when trying to detect new peering, a user may want to take some VPs at an IXP in addition to some VPs selected by MVP. MVP runs at https://bgproutes.io, allowing users to get a list of VPs or the redundancy scores without computational expenses. We implemented three versions of MVP, one for IPv4 routes (MVPv4), one for IPv6 routes (MVPv6) and one that considers both IPv4 and IPv6 routes (MVP${}^{v4}_{v6}$). The three versions use the same methodology (described in §7) to compute redundancy and generate a set of VPs. ## 9 Evaluation We show that MVP improves the trade-off between the volume of data collected and the routing information inferred compared to current VPs selection strategies in five use cases for which we have ground truth (§9.1). We then show that MVP would improve coverage and accuracy of previous studies for which ground truth is unknown (§9.2). Finally, we show that the key design choices of MVP are sound (§9.3). Use case | Objective | Naives baselines | Greedy specifics use cases (§9.1) | Greedy specifics Def. (§4) ---|---|---|---|--- Random | AS-distance | unbiased [45] | _I_ | _II_ | _III_ | _IV_ | _V_ | Def. 1 | Def. 2 | Def. 3 Transient path detection (_I_) | 50 % | 1.55 | 1.76 | 1.82 | 0.70 | 2.99 | 3.29 | 3.82 | 2.89 | 1.96 | 2.12 | 1.69 70 % | 1.38 | 1.62 | 1.53 | 0.76 | 3.24 | 3.51 | 3.42 | 3.09 | 1.56 | 1.56 | 1.78 90 % | 1.13 | 1.17 | 1.21 | 0.75 | 1.66 | 1.67 | 1.66 | 1.66 | 1.33 | 1.15 | 1.59 MOAS detection (_II_) | 50 % | 2.35 | 3.38 | 3.41 | 2.31 | 0.98 | 1.80 | 2.83 | 1.53 | 3.39 | 2.85 | 3.98 70 % | 2.18 | 3.44 | 3.38 | 2.56 | 0.85 | 1.79 | 2.30 | 1.83 | 3.02 | 2.66 | 3.67 90 % | 1.98 | 2.69 | 3.06 | 2.37 | 1.04 | 2.31 | 2.82 | 2.56 | 2.46 | 2.19 | 3.31 AS topology mapping (_III_) | 50 % | 2.59 | 2.97 | 2.43 | 1.58 | 1.29 | 0.71 | 1.53 | 1.94 | 2.27 | 2.18 | 3.35 70 % | 2.06 | 2.29 | 2.13 | 1.33 | 1.22 | 0.64 | 1.29 | 1.37 | 2.14 | 1.64 | 2.28 90 % | 1.72 | 1.88 | 1.80 | 1.30 | 1.18 | 0.77 | 1.23 | 1.27 | 1.73 | 1.64 | 1.80 Traffic engineering detection (_IV_) | 50 % | 4.59 | 4.74 | 4.82 | 3.76 | 2.67 | 2.34 | 0.47 | 3.33 | 3.95 | 3.34 | 4.76 70 % | 2.71 | 3.37 | 3.52 | 3.04 | 1.86 | 2.89 | 0.41 | 1.85 | 4.34 | 2.02 | 3.51 90 % | 1.55 | 1.70 | 1.95 | 1.61 | 1.54 | 1.52 | 0.32 | 1.46 | 1.88 | 1.33 | 1.89 Unnecessary updates detection (_V_) | 50 % | 1.72 | 2.89 | 2.41 | 2.19 | 2.10 | 2.63 | 2.20 | 0.38 | 2.43 | 2.92 | 2.59 70 % | 1.30 | 2.04 | 1.91 | 1.43 | 1.53 | 1.50 | 1.90 | 0.39 | 1.51 | 1.63 | 2.02 90 % | 1.01 | 1.36 | 1.39 | 1.17 | 1.18 | 1.14 | 1.16 | 0.50 | 1.09 | 1.38 | 1.35 | | | | | | | | | | | | Table 3: Data reduction factors with MVPv4 compared to several baselines for five use cases. MVP outperforms every baseline for all five use cases. Unlike greedy specifics, MVP greatly avoids overfitting. ### 9.1 Benchmarking MVP We benchmark MVP against three baselines per use case. _Use cases._ We evaluate MVP on five different use cases that we carefully picked such that each BGP attribute is useful for at least one of them. For instance, the time is useful to detect transient events (use case _I_); the prefix is useful to detect Multiple Origin ASes (MOAS) prefixes (use case _II_); the AS path is useful to map the Internet topology (use case _III_); and the community values are useful to detect traffic engineering (use case _IV_) and unnecessary updates (use case _V_). Our goal is to demonstrate that MVP does not overfit on some particular use cases or BGP attributes. For each use case, we process the updates collected during 100 one-hour periods (randomly selected in May 2023) and benchmark MVP on a set of events found. We thus have ground truth. We briefly describe below each use case along with our experimental settings. _I_ Transient paths detection. Transient paths are BGP routes visible for less than five minutes, a typical BGP convergence delay [29], and which can be attributed to e.g., path exploration [35]. We focus on 200 randomly selected transient path events for every one-hour period, making a total of $100*200=20000$ events used. _II_ MOAS prefixes detection. MOAS prefixes are announced by multiple distinct ASes [44], which can be caused by legitimate [53] or malicious [40, 49, 13] actions. We focus on 200 MOAS randomly selected events for every one-hour period, making a total of $100*200=20000$ MOAS events used. _III_ AS topology mapping. This is useful for e.g., inferring BGP policies [31] or AS paths [32]. For each VP, we process the first RIB dump of May 2023 as well as the updates collected during the 100 one-hour periods and focus on all distinct AS links observed. _IV_ Traffic engineering detection. We focus on action communities i.e., those associated with traffic engineering actions [50]. For every one-hour time period, we focus on 80 updates for which a path change coincides with the appearance of an action community, making a total of $100*80=8000$ path changes used. _V_ Unnecessary Updates detection. An unnecessary update is a BGP update that only signals a change in the community values but not in the AS path [28]. We consider 200 unnecessary updates randomly picked within each one-hour period, making a total of $100*200=20000$ events used. _Baselines._ We benchmarked MVP against three naive baselines commonly used in practice (§3.2): _(i)_ random selection of VPs, which results in a skewed set of VPs as they exhibit biases [45]; _(ii)_ AS-distance, i.e., select the first VP randomly and the following ones to maximize the AS-level distance between selected VPs; and _(iii)_ unbiased, i.e., start with all VPs and iteratively remove the one that most increases the bias on the set of remaining VPs. We measure the bias using the definition in [45]. We compare MVP against the three greedy specific VPs selection strategies optimized for Def. 1, 2, and 3 (§4). Additionally, we compare MVP against five other greedy specifics, one optimized for each of the five use cases described above. Unlike the greedy specifics described in §5, these five greedy specifics optimize the trade-off between the volume of the data and its capacity to achieve a particular objective. For instance, when the objective is to map the AS topology (use case _III_), greedy specific iteratively selects the VP that best improves the trade-off between the number of discovered AS links and the volume of processed data. _Reduction factor definition._ We define the reduction factor to capture how much MVP reduces the number of BGP updates required to fulfill a particular objective. More precisely, assume an objective $O$ and a baseline $B$. We iteratively build a set of VPs using baseline $B$. At every iteration, we download all the updates that the newly selected VP observes during 100 one- hour periods randomly selected in May 2023. We stop iterating when all updates collected by the selected VPs enable the data to meet $O$. Similarly, we build another set of VPs using MVP and stop selecting new VPs (see §7.4) when the selected ones meet $O$. The reduction factor is the ratio between the number of updates processed with $B$ and with MVP. More formally, the reduction factor is $\frac{|U_{B}^{O}|}{|U_{MVP}^{O}|}$ with $|U_{B}^{O}|$ and $|U_{MVP}^{O}|$ the number of updates processed to fulfill objective $O$ with baseline $B$ and MVP respectively. A reduction factor $=2$ means that we can fulfill objective $O$ with half as many updates when using MVP compared to when using baseline $B$. More generally, a reduction factor $>$ 1 means that we can fulfill the same objective with less data when using MVP compared to when using $B$. _Benchmark results._ Table 3 summarizes our results. For each use case, we focus on three objectives: mapping X% of the AS topology (use case _III_) or detecting X% of the events (use case _I_ , _II_ , _IV_ , and _V_), with X equal to 50, 70, or 90. Here, we focus on the performance of MVPv4. MVPv6 and MVP${}^{v4}_{v6}$ yield comparable performance (see §B). Takeaway #1: MVP outperforms every naive baseline for every use case, i.e., the reduction factor is always above one. For instance, we detect 90% of the MOAS events with 3.06$\times$ less data (the reduction factor is 3.06) when using MVP compared to selecting the VPs using the unbiased baseline. This means that MVP only needs 32% of the updates required by the unbiased baseline to fulfill the objective. Comparably to what we observe in our mini-Internet simulations (§3), the random baseline performs better than AS-distance. Takeaway #2: We can see that MVP generalizes whereas greedy specific overfits. In fact, for a particular use case, MVP is less performant than the greedy specific strategy optimized for this use case. For any other use case, MVP performs better than the greedy specifics not optimized for that use case. These results demonstrate that the greedy specific strategies overfit. They are also not practical as they need ground truth. ### 9.2 Impact on previous works We show that MVP would improve the outcome of three measurement studies and tools that are fueled by the BGP data from RIS and RV (and for which there is no ground truth). #### 9.2.1 Inference of AS properties We show that MVP improves AS relationship inferences (a popular research problem [31, 27, 22, 18]) and AS ranking [9]. _MVP helps to infer +15% more AS relationships._ We replicate the methodology proposed in [31] that relies on public BGP data from RIS and RV to infer AS relationships and build the widely-used CAIDA AS-relationship dataset [16]. We compute the number of inferred AS relationships for every month in 2023 when using the 648 VPs that CAIDA uses to build its dataset (In January 2023) and when using VPs selected by MVP. We ensure that the VPs selected with MVP generate the same volume of data as the 648 used by CAIDA so that any performance gap can confidently be attributed to MVP. We find that the VPs selected by MVP enable consistent (from Jan. 2023 to Aug. 2023) inference of $\approx$90k additional AS relationships ($\approx$+17%) while missing only $\approx$11k AS relationships ($\approx$2.2%) present in the original dataset. Thus, the tradeoff is largely in favor of using MVP ($\approx$+15% overall). We also replicated the AS relationship validation algorithm used in [31] (which relies on the IRR and RIR data) and found that the true positive rate (the metric used in [31]) remains identical (97%). Thus, MVP significantly improves coverage without processing more data or losing accuracy. _MVP prevents flawed inferences in the ASRank dataset._ We replicate the methodology used by ASRank [9] to compute the AS Customer Cone Sizes (CCS). We find that the CCS changes for 1067 ASes when using MVP and manually investigated two cases of substantial changes: Case I444https://asrank.caida.org/asns?asn=132337&type=search: AS132337 has a CCS of 1 in the original dataset and a CSS of 18k when using MVP, making it the 15th AS highest ranked by CCS. We contacted AS132337 who confirmed that it has 18k customers. MVP correctly ranks AS132337 because it selects the unique VP that sees it as a transit AS. Case II:555https://asrank.caida.org/asns?asn=24745&type=search AS24745 is the route server of Balcan-IX and has a CSS of 16 in the original ASrank dataset. However, we manually checked its participants and found that the 16 customers are misclassified and actually peer through AS24745. With MVP, the CSS of AS24745 is 1 and these errors are avoided. In both cases, MVP enables more accurate inferences of CCSs because it collects more diverse AS paths. Thus, we can confidently say that MVP would prevent many flawed inferences likely present in the dataset provided by ASRank. #### 9.2.2 Detection of forged-origin hijacks We show that MVP improves forged-origin hijack detection, which is the goal of many systems that use BGP routes from RIS and RV [11, 44, 1, 15, 26]. Forged- origin hijacks are a type of BGP hijack where the attacker prepends the valid origin to the AS path to make the hijacked route appear legitimate. _MVP improves the accuracy of forged-origin hijack inferences._ We replicate the algorithm of DFOH [26] that uses routes collected by 287 RIS and RV VPs to infer forged-origin hijacks. We implement two versions of DFOH, one called DFOHMVP which uses a set of VPs selected with MVP, and another one called DFOHR that uses a random set of VPs. In both versions, we ensure that the volume of data collected is identical to the one used in [26]. As DFOH relies on probabilistic inference, we measure the performance of DFOHMVP and DFOHR in terms of True Positive Rate (TPR) and False Positive Rate (FPR). We obtain an approximation of ground truth (needed to compute the TPR and FPR) by implementing a third version of DFOH, called DFOHALL that uses all VPs from RIS and RV. Observe that DFOHALL is an approximation of ground truth because incorrect inferences are still possible even if all VPs are used. We restrict our analysis to one month (Jan. 2022) because DFOHALL is resource-hungry as it uses all VPs. We find that DFOHMVP uncovers 947 suspicious cases against only 700 for DFOHR. DFOHMVP outperforms DFOHR for both the TPR and the FPR: It has a TPR of 85.7% (against 61.1% for DFOHR) and a FPR of 14.4% (against 60.1% for DFOHR)—a $\approx$4$\times$ better precision. _DFOH R misses suspicious cases that DFOHMVP does not._ We manually investigated, using public peering databases (e.g., PeeringDB) some of the suspicious cases inferred by DFOHMVP and not by DFOHR. We find cases that appear particularly suspicious (thus useful for operators) and describe two of them below (also found by the original DFOH). Case I666http://dfoh.uclouvain.be/cases/2022-01-01_1239_267548: On Jan. 1, 2022, AS267548, a small Peruvian AS, appears between Sprint, a Tier1 AS, and AS199524, a large content provider. However, AS267548 is not supposed to provide transit between these two ASes. Case II777http://dfoh.uclouvain.be/cases/2022-01-06_9269_268568: On Jan. 6, 2022, AS9269, an ISP based in Hong Kong appears directly connected with AS268568, a Brazilian ISP. These two ASes do not share any IXP and are not supposed to peer directly. These two cases show that MVP enables the detection of additional potential routing attacks versus not using it. #### 9.2.3 Characterizing international routing detours Experiment | Duration | # of VPs | # of processed Updates | # of Detours ---|---|---|---|--- Original paper | 1 Month | All VPs | $\approx$61B | 174k Random selection | 2 Months | 624 (median) | $\approx$61B | 165k (median) 4 Months | 313 (median) | $\approx$61B | 171k (median) MVP selection | 2 Months | 413 | $\approx$61B | 250K 4 Months | 220 | $\approx$61B | 263k Table 4: Using fewer VPs selected by MVP enables a longer study that detects more detours with the same volume of data. We focus on a study that uses all VPs to characterize international routing detours over one month [46]. International detours occur when two ASes in the same country are reachable through an AS in another country, which can lead to extra forwarding delays. We show that by using fewer VPs selected by MVP, we can lengthen the duration of the study to find more detours without processing more data. _MVP helps to detect +44% more routing detours._ We replicate the methodology used in [46] to detect routing detours except that _(i)_ we use a set of VPs selected using MVP that generates $\alpha\times$ less data compared to using them all, with $\alpha=2$ and $\alpha=4$, and _(ii)_ we run the analysis over two months when $\alpha=2$ and four months when $\alpha=4$. Thus, the overall volume of data collected remains similar ($\approx$61B RIB entries), regardless of $\alpha$. Table 4 shows the number of routing detours detected in May 2023 (and until June and August 2023 when $\alpha=2$ and 4, respectively). We detect 250k detours over two months ($\alpha=2$) when using 413 VPs selected by MVP—a +44% increase compared to using all VPs during one month as in [46]. When $\alpha=4$, we use 220 VPs selected by MVP on four months and find 263k detours—better than using them all on one month. We explored the trade-off between the number of VPs and the duration of the study using a random VPs selection strategy. We detected 165k detours when using $\approx 624$ random VPs and running the analysis over two months (we tested the random selection with 50 seeds and report the median in Table 4). This is fewer than when we replicated the original experiment, which demonstrates that optimized VP selection enables discovering more routing detours. _MVP enables improved characterization of routing detours._ We replicate the methodology used in [46] to rank countries based on their number of detours, and ASes based on how often they originate a detoured path. We find differences when using MVP, including two interesting cases: Case I: Using MVP (with $\alpha=2$), we discover 33k (+68%) additional detours traversing the US and 22k (+37%) traversing Russia compared to when using the settings in [46]. These additional detours rank the US as the #1 country with the highest number of routing detours and Russia as #2, whereas with the settings in [46] Russia is ranked #1 and the US #2. Case II: Using MVP (with $\alpha=2$) enables detecting 720 (+83%) additional routing detours involving AS262503 compared to when using the settings of [46]. This changes rankings: AS262503 became #1 vs. #7 with the settings in [46]. As our rankings are based on the highest number of routing detours compared to [46], we can confidently say MVP improves the characterization of international routing detours. ### 9.3 Soundness of design choices We show that our three key design choices – yearly update frequency of redundancy scores, balanced sampling, and topological feature selection – are sound. _MVP ’s redundancy scores are sufficiently stable over time that annual recomputation is sufficient_ We ran MVP every six months, starting in January 2023 and then going backward until January 2018 (i.e., a total of ten independent runs). We limit the scope of this experiment to 100 randomly selected VPs to limit the computational resources required. Logically, we find that the redundancy score differences increase as the time interval between two runs of MVP increases. However, these differences are low. The median difference between the scores of two runs of MVP separated by one year is only 0.021 (which corresponds to a difference of 9%), and it increases to 0.171 (i.e., a difference of 23%) when the two runs are separated by four years. We thus configure MVP to recompute redundancy scores and update its set of selected VPs on a yearly basis (see §8)—a good trade-off between computational cost and performance. Figure 6: MVP enables mapping more links than rMVP. | $\setminus\\{{0,1}\\}$ | $\setminus\\{{2,3}\\}$ | $\setminus\\{{4,5}\\}$ | $\setminus\\{{6,7,8}\\}$ ---|---|---|---|--- _I_ | 1.04 | 1.05 | 1.07 | 1.06 _II_ | 1.17 | 1.02 | 1.07 | 1.34 _III_ | 1.09 | 1.11 | 1.09 | 1.12 _IV_ | 1.32 | 1.25 | 1.15 | 1.16 _V_ | 1.61 | 1.59 | 1.71 | 1.62 | | | | Table 5: Omitting one feature category reduces the performance of MVP for every use case. _The balanced sampling avoids biases in the collected data._ We implement rMVP, a modified version of MVP where new-AS-links events used to compute redundancy scores across VPs are sampled randomly, i.e., using the distribution depicted in Fig. 5(a) (instead of using the balanced sampling in §7.1). We compare the performance of MVP and rMVP on AS topology mapping. Note that we observe similar results for other use cases. We map the AS topology for May 2023 (following the methodology in §9.2.1) using both MVP and rMVP and with the same volume of data in either case. Fig. 6 depicts the proportion of additional AS links that we can map when using MVP compared to when using rMVP for every new-AS-link category. MVP always yields better or identical performance than rMVP. The highest difference is when mapping stub-to-stub links (+3.9%) or Transit1-to-Transit1 links (+2.6%). These two link categories are underrepresented when using a random sampling (see Fig. 5(a)), demonstrating that our balanced sampling scheme mitigates biases. _Every feature category is useful._ We implement MVP $\setminus\\{f_{i},..,f_{j}\\}$, a modified version of MVP where we omit features $\\{f_{i},..,f_{j}\\}$ when computing redundancy scores, with $i...j$ the feature indexes in Table 2. We use four different versions of MVP $\setminus\\{...\\}$, each omitting a different feature category. We show the reduction factor of MVP over each MVP $\setminus\\{...\\}$ for use cases _I_ , _II_ , _III_ , _IV_ , and _V_ in §9.1, with the objective of detecting 70% of the events or mapping 70% of the AS-level topology. Regardless of which feature category is omitted, MVP performs better (i.e., the reduction factor is above 1). We conclude that every feature category is valuable. ## 10 Related work _Redundancy and bias between the VPs._ Chen et al. showed that VPs observe identical (redundant) AS links and that it is possible to reduce the number of VPs while providing similar measurement power [10]. However, they only focus on one objective (observing AS links) whereas MVP works for any objective. Previous works reported that the VPs are biased (in terms of location, network size, etc.) [45, 14, 47]. MVP is data-driven and does not consider these biases as we show that an unbiased selection strategy performs poorly (§9.1). _Strategies to select VPs._ Prior works demonstrated that carefully selecting VPs increases the utility of the data [52], and proposed a greedy selection strategy that performs better than other naive approaches [34, 52]. However, their selection strategy optimizes one objective (discovering AS links) and thus lacks generality (§9.1). Recent works also study the impact of the VP selection on the discovered IP space and AS links [30]. _Placement of the VPs._ Gregori et al. proposed a methodology that finds a relevant placement for a new VP [23]. Roughan et al. estimated that 700 strategically positioned VPs were enough to monitor the Internet topology [43]. Finally, Cittadini et al. demonstrated the marginal utility of adding new VPs at the core of the Internet [14]. _Strategies to select active measurement probes._ Active measurement platforms (e.g., RIPE Atlas) also generate a large volume of data and several data- driven approaches for probe selection exist [4, 25, 6]. Unlike MVP, these approaches optimize the probe selection for specific use cases. _Uses of topological features._ Previous works computed topological features on the AS topology to detect routing anomalies [24, 26, 19]. ## 11 Conclusion We uncovered redundancy in the BGP routes exported by the RIS and RV VPs and identified this redundancy as an opportunity to optimize the use of these data collection systems. We presented MVP, a system that samples BGP data at the VP granularity, enabling users to improve the coverage and accuracy of their studies without processing more data. The principles that MVP embodies can also lead to a better understanding of the structure of the global Internet as well as how to optimize the measurement and analysis of its routing system. For instance, our redundancy scores could lead to more strategic approaches to gathering and retaining BGP data, e.g., RIS and RV could deprioritize VPs which are overwhelmingly redundant with many others, on a more scientific basis. Finally, our approach can be adapted to active measurement platforms (e.g., Atlas [39]) to reach the same objective of extensive coverage with reduced redundant data. ## 12 Acknowledgements This work was supported by the ArtIC project (grant ANR-20-THIA-0006-01), Région Grand Est, Inria Nancy-Grand Est, IHU of Strasbourg, University of Strasbourg, University of Haute-Alsace, the RIPE NCC Community Projects Fund, NSF CNS-2120399 and NSF OAC-2131987. Views are those of the authors and do not represent the endorsements of the funding agencies. ## References * [1] Global Routing Intelligence Platform. https://grip.inetintel.cc.gatech.edu/. * [2] Emile Aben. Route Collection at the RIPE NCC - Where are we and where should we go? 2020\. https://labs.ripe.net/author/emileaben/. * [3] Rodrigo Aldecoa, Chiara Orsini, and Dmitri Krioukov. Hyperbolic graph generator. In Computer Physics Communications, 2015. * [4] Malte Apple, Emile Aben, and Romain Fontugne. Metis: Better Atlas Vantage Point Selection for Everyone. In TMA ’22, 2022. * [5] Lorenzo Ariemma, Mariano Scazzariello, and Tommaso Caiazzi. MRT#: a Fast Multi-Threaded MRT Parser. In IFIP/IEEE IM ’21, 2021. * [6] Vaibhav Bajpai, Steffie Jacob Eravuchira, Jürgen Schönwälder, Robert Kisteleki, and Emile Aben. Vantage point selection for IPv6 measurements: Benefits and limitations of RIPE Atlas tags. In 2017 IFIP/IEEE IM, 2017. * [7] BGPKIT. BGPKIT. 2022\. https://blog.bgpkit.com/. * [8] Timm Böttger, Félix Cuadrado, and Steve Uhlig. Looking for hypergiants in PeeringDB. In SIGCOMM CCR, 2018. * [9] CAIDA. As rank. 2023\. https://asrank.caida.org/. * [10] Kai Chen, Chengchen Hu, Wenwen Zhang, Yan Chen, and Bin Liu. On the eyeshots of BGP Vantage Points. In GLOBECOM ’09, 2009. * [11] Shinyoung Cho, Romain Fontugne, Kenjiro Cho, Alberto Dainotti, and Phillipa Gill. BGP hijacking classification. In TMA’19, 2019. * [12] CIDR. CIDR REPORT. 2023\. https://www.cidr-report.org/as2.0/. * [13] CitizenLab. A case study of the China Telecom incident. 2012\. https://citizenlab.ca/2012/12/. * [14] Luca Cittadini, Stefano Vissicchio, and Benoit Donnet. On the quality of BGP route collectors for iBGP policy inference. In IFIP ’14, 2014. * [15] Avichai Cohen, Yossi Gilad, Amir Herzberg, and Michael Schapira. Jumpstarting BGP security with Path-End Validation. In SIGCOMM’16, 2016. * [16] University San Diego. The CAIDA AS Relationships Dataset, 2022. 2022\. https://www.caida.org/catalog/datasets/as-relationships/. * [17] Benoit Donnet and Olivier Bonaventure. On BGP communities. In SIGCOMM CCR, 2008. * [18] Guoyao Feng, Srinivasan Seshan, and Peter Steenkiste. UNARI: An Uncertainty-Aware Approach to AS Relationships Inference. In CoNEXT ’19, 2019. * [19] Romain Fontugne, Anant Shah, and Emile Aben. AS Hegemony: A Robust Metric for AS Centrality. In SIGCOMM’17, 2017. * [20] Linton C. Freeman. Centrality in social networks conceptual clarification. Social Networks, 1978. * [21] Lixin Gao and Jennifer Rexford. Stable Internet Routing without Global Coordination. In SIGMETRICS ’00, 2000. * [22] Vasileios Giotsas, Matthew Luckie, Bradley Huffaker, and kc claffy. Inferring Complex AS Relationships. In IMC ’14, 2014. * [23] Enrico Gregori, Alessandro Improta, Luciano Lenzini, Lorenzo Rossi, and Luca Sani. On the Incompleteness of the AS-Level Graph: A Novel Methodology for BGP Route Collector Placement. In IMC ’12, 2012. * [24] Kevin Hoarau, Pierre Ugo Tournoux, and Tahiry Razafindralambo. BGNN: Detection of BGP Anomalies Using Graph Neural Networks. In ISCC ’22, 2022. * [25] Thomas Holterbach, Emile Aben, Cristel Pelsser, Randy Bush, and Laurent Vanbever. Measurement Vantage Point Selection Using A Similarity Metric. In ANRW ’17, 2017. * [26] Thomas Holterbach, Thomas Alfroy, Amreesh Phokeer, Alberto Dainotti, and Cristel Pelsser. A System to Detect Forged-Origin BGP Hijacks. In NSDI’24, 2023. * [27] Zitong Jin, Xingang Shi, Yan Yang, Xia Yin, Zhiliang Wang, and Jianping Wu. TopoScope: Recover AS Relationships From Fragmentary Observations. In IMC ’20, 2020. * [28] Thomas Krenc, Robert Beverly, and Georgios Smaragdakis. Keep Your Communities Clean: Exploring the Routing Message Impact of BGP Communities. In CoNEXT ’20, 2020. * [29] Craig Labovitz, Abha Ahuja, Abhijit Bose, and Farnam Jahanian. Delayed Internet Routing Convergence. In SIGCOMM CCR, 2000. * [30] Franziska Lichtblau. From the Edge to the Core : towards informed Vantage Point selection for Internet measurement studies. In Ph.D. Thesis, 2021. * [31] Matthew Luckie, Bradley Huffaker, Amogh Dhamdhere, Vasileios Giotsas, and kc claffy. AS Relationships, Customer Cones, and Validation. In IMC ’13, 2013. * [32] Z. Morley Mao, Lili Qiu, Jia Wang, and Yin Zhang. On AS-Level Path Inference. In SIGMETRICS ’05, 2005. * [33] University of Oregon. Route Views Peers list. 2023\. http://www.routeviews.org/peers/peering-status.html. * [34] Ricardo Oliveira, Mohit Lad, Beichuan Zhang, Dan Pei, Daniel Massey, and Lixia Zhang. Placing BGP monitors in the Internet. In Technical No. UCLA, TR, 2006. * [35] Ricardo Oliveira, Beichuan Zhang, Dan Pei, Rafit Izhak-Ratzin, and Lixia Zhang. Quantifying path exploration in the internet. In ACM IMC’06, 2006. * [36] Chiara Orsini, Alistair King, Danilo Giordano, Vasileios Giotsas, and Alberto Dainotti. BGPStream: A Software Framework for Live and Historical BGP Data Analysis. In IMC ’16, 2016. * [37] Lars Prehn and Anja Feldmann. How Biased is Our Validation (Data) for AS Relationships? In IMC ’21, 2021. * [38] RIPE. RIPE RIS Raw Data. 1\. https://www.ripe.net/data-tools/stats/ris/. * [39] RIPE. The RIPE Atlas measurement platform. 1\. https://atlas.ripe.net/. * [40] RIPE. YouTube Hijacking: A RIPE NCC RIS case study. 2018\. http://www.ripe.net/internet-coordination/news/industry-developments/. * [41] RIPE. Per-peer dump files. 2023\. https://ris.ripe.net/docs/40_Prototypes/10_per_peer_dumps.html. * [42] RIPE. RIPE RIS Peers list. 2023\. https://www.ris.ripe.net/peerlist/. * [43] Matthew Roughan, Simon Jonathan Tuke, and Olaf Maennel. Bigfoot, Sasquatch, the Yeti and other missing links: what we don’t know about the AS graph. In IMC ’08, 2008. * [44] Pavlos Sermpezis, Vasileios Kotronis, Petros Gigis, Xenofontas Dimitropoulos, Jae Hyun Park, Danilo Cicalese, Alistair King, and Alberto Dainotti. ARTEMIS: Neutralizing BGP hijacking within a minute. In ToN, 2018. * [45] Pavlos Sermpezis, Lars Prehn, Sofia Kostoglou, Marcel Flores, Athena Vakali, and Emile Aben. Bias in Internet Measurement Platforms. In TMA’23, 2023. * [46] Anant Shah, Romain Fontugne, and Christos Papadopoulos. Towards Characterizing International Routing Detours. In Asian Internet Engineering Conference ’16, 2016. * [47] Y. Shavitt and U. Weinsberg. Quantifying the Importance of Vantage Points Distribution in Internet Topology Measurements. In INFOCOM, 2009. * [48] Mattia Tantardini, Francesca Ieva, Lucia Tajoli, and Carlo Piccardi. Comparing methods for comparing networks. In Nature, 2019. * [49] Ars Technica. Russian-controlled telecom hijacks financial services’ internet traffic. 2017\. https://arstechnica.com/security/2017/04/. * [50] Krenc Thomas, Luckie Matthew, Marder Alexander, and kc Claffy. Coarse-grained Inference of BGP Community Intent. In IMC’23, 2023. * [51] Oregon Univ. Route Views Project. 2021\. www.routeviews.org/. * [52] Ying Zhang, Zheng Zhang, Z. Morley Mao, Y. Charlie Hu, and Bruce M. Maggs. On the impact of route monitor selection. In IMC ’07, 2007. * [53] Xiaoliang Zhao, Dan Pei, Lan Wang, Dan Massey, Allison Mankin, S. Felix Wu, and Lixia Zhang. An Analysis of BGP Multiple Origin AS (MOAS) Conflicts. In IMW ’01, 2001. ## Appendix ## Appendix A Survey _Detailed methdology._ We selected eleven papers and classified them based on how authors collected the BGP data (categories $C_{1}$ and $C_{2}$). We then emailed the authors and asked them about their experience with using BGP routes from RIS and RV. We did not have answers for three papers. We promised to share the answers of the participants in an anonymized fashion. Thus, we do not show parts of a few answers that would make de-anonymization possible. However, the missing parts never change the main message conveyed in the answers. _Detailed answers._ Table 6 lists the questions we asked the participants of our survey along with their detailed answers. We color the answers based on whether they are in favor (green) of using a tool such as MVP or not (red). Neutral answers are colored in blue. The vast majority of the answers indicate that MVP would be beneficial for users and improve the quality of their measurement studies. Collection strategy | Questions asked | Collected answers ---|---|--- $C_{1}$: All routes and subset of VPs (seven papers) | Why did you use a subset of the VPs ? | | To speed up data processing (x2) --- For disk space and time efficiency (x1) I thought the rest would be similar (x1) I did not manage to use them all (x2) How did you select your VPs ? | | I took them randomly (x2) --- I do not remember (x2) It was arbitrary: my script partially failed (x1) I took geographically distant BGP collectors (x1) I did not manage to use VPs from one data provider (x1) | Do you think more VPs would improve --- the quality of your results? | Yes (x4) --- Results would be similar, but it can help to find corner cases (x1) Yes, but not significantly (x1) I am not sure (x1) | Would you have used more VPs --- if you could? | Yes (x4) --- Yes, I’d love to (x1) Definitely (x1) I am not sure, but I don’t think so (x1) $C_{2}$: Limited duration of experiment (five papers) | | Was the processing time a factor --- that you considered when you decided on the duration of your measurement study? Yes (x3) | Do you think extending the duration --- of your measurement study would improve the quality of your results? | Yes (x2) --- Yes, especially for rare events (x1) Potentially (x1) Yes, but not significantly (x1) | Would have extended the duration --- of your measurement study if you had more resources? | Yes (x2) --- Yes, but it depends on the time remaining before the deadline (x1) I think so, but also if I had more time before the deadline (x1) All eight papers | | Do you find the data from RIS and --- RouteViews expensive to process in terms of computational resources? | Yes (x1) --- Yes, CPU and storage (x2) Yes, the storage cost and the download cost are very large (x1) CPU is the main issue (x1) RIS data takes a lot of time to download, especially when we need data for multiple days (x1) Not the worst, but we definitely need a resourceful server if we want to catch some deadline (x1) We did that in a server so that was not a huge issue (x1) No (x1) | Is there any additional challenge --- that you encountered when processing the BGP data from RIS and RouteViews? | Our team used Spark clusters and Python but it was too slow (x1) --- We had to download the data from all VPs as there is no optimal solution for selecting them, the storage overhead and time overhead were extremely high (x1) It’ll be helpful to make processing faster and less resource-consuming (x1) Too many duplicate announcements make processing harder (x1) Variable sizes of update files exacerbate scheduling parallelization (x1) RIS took a lot longer than RouteViews (x1) We had issues when collecting updates in real-time (x1) We had to deal with bugs in BGPdump (x1) Broken data feeds and data cleanup is also an issue that we need to take care of (x1) Our study was done pre-BGPStream, which would have helped quite a bit already (x1) Table 6: An exhaustive list of the questions asked to the participants of the survey along with their detailed answers. We color an answer in (bold) green if it (strongly) motivates the usage of a tool such as MVP. Blue answers are neutral, i.e., they do not motivate MVP but also do not disincentive it. Finally, (bold) red answers (strongly) disincentive the usage of a tool such as MVP. ## Appendix B Extended evaluation In this section, we evaluate the performances of MVP${}^{v4}_{v6}$ (Table 7) and MVPv6 (Table 8) on the five use cases presented in §9.1, namely transient paths detection (_I_), MOAS detection (_II_), AS topology mapping (_III_), traffic engineering detection (_IV_), and unnecessary updates detection (_V_). Similarly to §9.1, we compare MVPv6 and MVP${}^{v4}_{v6}$ against the three naive baselines (random, AS-distance, and unbiased) as well as the eight greedy specific VPs selection strategies (three optimized for Def. 1, 2, and 3 and one optimized for each of the five use cases). We present the results in terms of data reduction factor, as defined in §9.1. _MVP ${}^{v4}_{v6}$ and MVPv6 outperform the three naive baselines for every objective._ For MVP${}^{v4}_{v6}$, the reduction factor can be as high as 6.57 when trying to detect 50% of the traffic engineering paths while for MVPv6 it can be as high as 5.05 when trying to map 50% of the AS topology. On average, MVP${}^{v4}_{v6}$ only needs 41.6% of the data (reduction factor of 2.4) required by a naive baseline to meet the same objective while MVPv6 needs 44.5% (reduction factor of 2.26). _MVP ${}^{v4}_{v6}$ and MVPv6 prevent overfitting._ For the vast majority of the objectives, greedy specific performs better than MVP${}^{v4}_{v6}$ or MVPv6 only for the use cases for which it is optimized. There are a few cases where greedy specific performs better than MVP${}^{v4}_{v6}$ or MVPv6 for a use case that it does not optimized. For instance, MVPv6 needs to process 20% (reduction factor of 0.8) more data than greedy specific optimized for use case _I_ to detect 90% of the MOAS (use case _II_). However, in the vast majority of the cases, both MVP${}^{v4}_{v6}$ and MVPv6 outperform the greedy specifics. For instance, MVPv6 only needs 26% (reduction factor of 3.74) of the volume required by the greedy specific optimized for use case _IV_ to detect 90% of the MOAS (use case _II_). These results show that MVP does not overfit while greedy specific does. Use case | Objective | Naives baselines | Greedy specifics use cases (§9.1) | Greedy specifics Def. (§4) ---|---|---|---|--- Random | AS-distance | unbiased | _I_ | _II_ | _III_ | _IV_ | _V_ | Def. 1 | Def. 2 | Def. 3 Transient path detection (_I_) | 50 % | 1.32 | 1.87 | 1.94 | 0.61 | 1.19 | 1.11 | 1.36 | 1.21 | 2.08 | 2.24 | 1.79 70 % | 1.38 | 1.62 | 1.83 | 0.74 | 1.30 | 1.15 | 1.40 | 1.18 | 1.78 | 1.97 | 1.53 90 % | 1.16 | 1.42 | 1.40 | 0.71 | 1.39 | 1.74 | 1.69 | 1.21 | 1.34 | 1.36 | 1.40 MOAS detection (_II_) | 50 % | 1.93 | 3.38 | 4.03 | 1.95 | 0.78 | 1.41 | 2.05 | 1.37 | 3.34 | 3.21 | 2.88 70 % | 1.96 | 3.49 | 4.16 | 2.14 | 0.68 | 1.91 | 2.52 | 1.56 | 2.91 | 2.60 | 2.81 90 % | 1.16 | 1.69 | 2.07 | 1.52 | 0.69 | 1.68 | 1.87 | 1.53 | 1.31 | 1.25 | 1.40 AS topology mapping (_III_) | 50 % | 2.47 | 2.90 | 2.72 | 1.18 | 1.02 | 0.58 | 1.45 | 1.41 | 2.38 | 2.30 | 2.16 70 % | 2.27 | 2.52 | 2.29 | 1.26 | 1.14 | 0.68 | 1.25 | 1.19 | 2.03 | 1.71 | 2.03 90 % | 1.71 | 1.85 | 1.78 | 1.14 | 1.13 | 0.82 | 1.17 | 1.15 | 1.62 | 1.61 | 1.56 Traffic engineering detection (_IV_) | 50 % | 3.77 | 6.57 | 4.43 | 3.21 | 1.89 | 1.47 | 0.47 | 2.57 | 3.43 | 2.89 | 2.57 70 % | 2.34 | 3.17 | 2.56 | 2.20 | 1.60 | 1.93 | 0.35 | 2.06 | 1.97 | 2.01 | 2.05 90 % | 1.90 | 2.02 | 1.76 | 2.02 | 1.78 | 1.94 | 0.31 | 1.93 | 2.13 | 1.67 | 1.93 Unnecessary updates detection (_V_) | 50 % | 2.13 | 4.10 | 3.15 | 2.41 | 2.54 | 2.72 | 3.12 | 0.41 | 2.94 | 2.78 | 2.83 70 % | 1.27 | 1.95 | 1.80 | 1.47 | 1.12 | 1.28 | 1.59 | 0.35 | 1.71 | 1.81 | 1.57 90 % | 1.01 | 1.29 | 1.35 | 1.04 | 0.85 | 0.96 | 1.09 | 0.46 | 1.00 | 1.11 | 1.11 | | | | | | | | | | | | Table 7: Data reduction factor for MVP${}^{v4}_{v6}$ compared to several baselines for five use cases. MVP${}^{v4}_{v6}$ enables to detect 70% of the MOAS using only 28.6% (reduction factor of 3.49) of the volume required by the AS distance baseline to meet the same objective. The average reduction factor over all objectives and naive baselines is 2.25. Use case | Objective | Naives baselines | Greedy specifics use cases (§9.1) | Greedy specifics Def. (§4) ---|---|---|---|--- Random | AS-distance | unbiased | _I_ | _II_ | _III_ | _IV_ | _V_ | Def. 1 | Def. 2 | Def. 3 Transient path detection (_I_) | 50 % | 1.43 | 1.65 | 2.29 | 0.44 | 1.11 | 1.14 | 1.67 | 1.20 | 1.56 | 1.36 | 1.90 70 % | 1.71 | 1.84 | 2.00 | 0.64 | 1.59 | 1.96 | 2.88 | 2.25 | 1.75 | 1.86 | 1.67 90 % | 1.52 | 1.43 | 1.42 | 0.62 | 1.49 | 1.49 | 1.79 | 1.48 | 1.51 | 1.72 | 2.11 MOAS detection (_II_) | 50 % | 1.94 | 1.65 | 2.37 | 1.10 | 0.21 | 0.36 | 1.56 | 2.33 | 1.04 | 0.73 | 1.30 70 % | 4.24 | 1.70 | 3.25 | 1.26 | 0.51 | 1.05 | 4.98 | 3.38 | 4.20 | 3.71 | 4.13 90 % | 3.03 | 1.75 | 2.19 | 0.80 | 0.53 | 2.67 | 3.74 | 2.56 | 3.54 | 3.69 | 3.84 AS topology mapping (_III_) | 50 % | 4.45 | 3.68 | 5.05 | 1.49 | 0.72 | 0.54 | 2.41 | 3.03 | 1.92 | 1.65 | 3.29 70 % | 2.83 | 3.26 | 3.14 | 1.18 | 1.14 | 0.73 | 2.07 | 1.38 | 2.27 | 2.17 | 2.48 90 % | 1.86 | 2.00 | 1.99 | 1.10 | 1.12 | 0.86 | 1.25 | 1.30 | 1.56 | 1.70 | 2.02 Traffic engineering detection (_IV_) | 50 % | 2.27 | 1.68 | 1.34 | 2.68 | 0.51 | 0.58 | 0.12 | 1.89 | 0.75 | 0.95 | 0.53 70 % | 3.76 | 5.14 | 2.86 | 3.03 | 2.64 | 3.03 | 0.30 | 4.66 | 2.07 | 1.61 | 1.14 90 % | 1.29 | 1.36 | 1.18 | 1.49 | 1.49 | 1.49 | 0.65 | 1.49 | 0.88 | 0.56 | 1.19 Unnecessary updates detection (_V_) | 50 % | 1.45 | 2.19 | 2.63 | 1.57 | 2.94 | 1.97 | 3.21 | 0.22 | 2.44 | 2.70 | 1.95 70 % | 1.37 | 2.13 | 2.09 | 1.79 | 1.98 | 2.07 | 2.26 | 0.31 | 1.91 | 2.18 | 1.82 90 % | 1.23 | 1.46 | 1.58 | 1.34 | 1.39 | 1.46 | 1.69 | 0.50 | 1.83 | 1.49 | 1.45 | | | | | | | | | | | | Table 8: Data reduction factor for MVPv6 compared to several baselines for five use cases. MVPv6 enables to detect 90% of the MOAS using only 33% (reduction factor of 3.03) of the volume required by the random selection to meet the same objective. The average reduction factor over all objectives and naive baselines is 2.25.
$\overline{\mathcal{M}}^{B}(\bm{x},\bm{y};\overrightarrow{\rho})$ diffeomorphic to $(0,1]$; the count of such ends is equal to $\\#\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{1230})$ when $\overrightarrow{\rho}=(-\rho_{0},-\rho_{123})$, and is equal to $\\#\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{3012})$ when $\overrightarrow{\rho}=(-\rho_{012},-\rho_{3})$. By Proposition 2.62 and 2.57, boundary degenerations are further divided into boundary degenerations with or without corners. In particular, different types of degeneration do not appear simultaneously. When boundary degeneration without corners appear, the situation is covered in Proposition 2.62. When boundary degenerations with corners appear, Proposition 2.58 and Proposition 2.60 show the moduli space of degenerate disks are smoothly cut out. In particular, the standard gluing results can be applied to show each boundary degeneration in $\overline{\mathcal{M}}^{B}(\bm{x},\bm{y};\overrightarrow{\rho})$ has a neighborhood diffeomorphic to $(0,1]$. The number of such ends is equal to $\sum_{\\{(q,B_{1})|\exists B_{2}\in T(q),B_{1}+B_{2}=B\\}}\\#(\mathcal{M}^{B_{1}}(\bm{x},\bm{y};q)\times_{ev}\mathcal{N}^{B_{2}}(q;\overrightarrow{\rho}))$ (We have suppressed the almost complex structure $J_{s}$, which can be chosen generically so that the evaluation maps are transversal to each other.) This quantity is even when $\overrightarrow{\rho}\neq(-\rho_{0},\ldots,-\rho_{3})$ in view of Proposition 2.60. Otherwise, it has the same parity as $\sum_{\\{(q,B_{1})|\exists B_{2}\in T(q),B_{1}+B_{2}=B\\}}\\#\mathcal{M}^{B_{1}}(\bm{x},\bm{y};q)$ in view of Proposition 2.58. ∎ ### 2.8. Ends of moduli spaces of 1-P curves This subsection characterizes the ends of one-dimensional moduli spaces of 1-P holomorphic curves. Given a generator $\bm{x}$, we say $\iota(\bm{x})=\iota_{1}$ if and only if $\bm{x}$ is in $\mathbb{T}_{\alpha,1}$; otherwise $\iota(\bm{x})=\iota_{0}$. The main result is the following. ###### Proposition 2.63. Let $B\in\tilde{\pi}_{2}(\bm{x},\bm{y})$ such that $\iota(\bm{x})=\iota_{1}$ and $\text{ind}(B;U)=2$. Then fixing a generic almost complex structure, the compactified moduli space $\overline{\mathcal{M}}^{B}(\bm{x},\bm{y};U)$ is a compact 1-manifold with boundary. The boundaries are of the following types: * (1) Two-story building * (2) simple holomorphic combs $(u,v)$ with $v$ being an orbit curve * (3) boundary degeneration with corners * (4) boundary degeneration without corners Moreover, * (a) The number of type (2) ends is $\\#\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{1230})+\\#\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{3012})$. * (b) The number of type (3) ends is mod 2 congruent to $\sum_{\\{(B_{1},\ q)|B_{2}\in T(q),\ B_{1}+B_{2}=B\\}}\\#\mathcal{M}^{B_{1}}(\bm{x},\bm{y};q)$ * (c) The number of type (4) ends is even. ###### Remark 2.64. A similar proposition holds in the case when $\iota(\bm{x})=\iota_{0}$. One simply needs to change the Reeb chords in (a) by a cyclic permutation of the digits in the subscript. #### 2.8.1. Reformulation of the moduli spaces We reformulate $\mathcal{M}^{B}(\bm{x},\bm{y};U)$ in terms of holomorphic disks in $Sym^{g}(\Sigma)$. Assume $\iota(\bm{x})=\iota_{1}$ throughout the rest of the section. ###### Definition 2.65. ${\mathcal{M}}^{B}_{Sym}(\bm{x},\bm{y};U)$ is defined to be the space of holomorphic maps $u:[0,1]\times\mathbb{R}\backslash\\{(s_{0},0)\\}\rightarrow Sym^{g}(\Sigma)$ such that: * (1) $(s_{0},0)$ is in the interior of $[0,1]\times\mathbb{R}$ and is allowed to vary; * (2) $u(\\{0\\}\times\mathbb{R})\subset\mathbb{T}_{\beta}$; * (3) $u(\\{1\\}\times\mathbb{R})\subset\mathbb{T}_{\alpha,1}$. Moreover, $u|_{\\{1\\}\times\mathbb{R}}$ lifts through $f_{1}:(0,1)\times\mathbb{T}^{g-2}\times(\amalg S^{1})\rightarrow Sym^{g}(\Sigma)$; * (4) $\lim_{t\rightarrow\infty}u(s+it)=\bm{y}$, and $\lim_{t\rightarrow-\infty}u(s+it)=\bm{x}$; * (5) $\lim_{(s,t)\rightarrow(s_{0},0)}u(s+it)$ is a closed $\overrightarrow{R}$-orbit $\sigma\times\bm{w}$, where $\bm{w}\in Sym^{g-1}(\Sigma)$ and $\sigma$ stands for a closed Reeb orbit that traverses $\partial\overline{\Sigma}$ once; * (6) $\frac{du}{ds}+J_{s}\frac{du}{dt}=0$; * (7) $u$ is in the homology class specified by $B$. Again, we have the tautological correspondence that identifies the moduli spaces defined here and the ones in Section 2.2. Therefore, we shall no longer keep the subscript “Sym” in the notation. We also define the moduli spaces of one-punctured degenerate disks (with or without corners). ###### Definition 2.66 (One-punctured degenerate disks without corners). Let $J$ be a nearly-symmetric almost complex structure. Let $\bm{x}\in\mathbb{T}_{\alpha}$. $\mathcal{N}_{J}(\bm{x};U)$ is the space of maps $v:\mathbb{H}\backslash\\{i\\}\rightarrow Sym^{g}(\Sigma)$ such that * (1) $v(\mathbb{R})\subset\mathbb{T}_{\alpha,1}$. Moreover, the restriction of $v|_{\mathbb{R}}$ lifts through $f_{1}:(0,1)\times\mathbb{T}^{g-2}\times(\amalg S^{1})\rightarrow Sym^{g}(\Sigma)$; * (2) $\lim_{z\rightarrow\infty}v(z)=\bm{x}$, and the path obtained from $v|_{\partial\mathbb{H}}$ by continuous extension at $\infty$ lifts through $\iota_{1}$; * (3) $\lim_{z\rightarrow i}v(z)$ is some closed $\overrightarrow{R}$-orbit $\sigma\times\bm{w}$, where $\bm{w}\in Sym^{g-1}(\Sigma)$ and $\sigma$ stands for a closed Reeb orbit that traverses the $\partial\overline{\Sigma}$ once; * (4) $\frac{du}{ds}+J\frac{du}{dt}=0$. ###### Definition 2.67 (One-cornered one-punctured degenerate disks). Let $J$ be a nearly-symmetric almost complex structure. Let $q$ be a self- intersection point of $\alpha_{im}$. $\mathcal{N}_{J}(q;U)$ is the space of maps $v:\mathbb{H}\backslash\\{i\\}\rightarrow Sym^{g}(\Sigma)$ such that * (1) $v(\mathbb{R})\subset\mathbb{T}_{\alpha,1}$. Moreover, the restriction of $v|_{\mathbb{R}}$ lifts through $f_{1}:(0,1)\times\mathbb{T}^{g-1}\times(\amalg S^{1})\rightarrow Sym^{g}(\Sigma)$; * (2) $\lim_{z\rightarrow\infty}v(z)=(q,\bm{p})$ for some $p\in\alpha^{a}_{1}\times\alpha^{c}_{1}\times\cdots\times\alpha^{c}_{g-2}$, and and the path obtained from $v|_{\partial\mathbb{H}}$ by continuous extension at $\infty$ does not lift through $\iota_{1}$; * (3) $\lim_{z\rightarrow i}v(z)$ is some closed $\overrightarrow{R}$-orbit $\sigma\times\bm{w}$, where $\bm{w}\in Sym^{g-1}(\Sigma)$ and $\sigma$ stands for a closed Reeb orbit that traverses $\partial\overline{\Sigma}$ once; * (4) $\frac{du}{ds}+J\frac{du}{dt}=0$. We call $q$ the corner of such a degenerate disk. We also have an evaluation map $ev_{J}:\mathcal{N}_{J}(q;U)\rightarrow\alpha^{a}_{i}\times\alpha^{c}_{1}\times\cdots\times\alpha_{g-2}^{c}$ defined by $v\mapsto\bm{p}$ if $\lim_{z\rightarrow\infty}v(z)=(q,\bm{p})$. #### 2.8.2. One-punctured boundary degeneration with corners ###### Proposition 2.68. If a boundary degeneration with corners appears in the compactification of a one-dimensional moduli space $\mathcal{M}^{B}(\bm{x},\bm{y};U)$, then the nodal comb is of simple form, and the domain for the degenerate disk is a stabilized teardrop with an acute corner. ###### Proof. The proof is similar to the proof of Proposition 2.57. There is only one modification needed: We no longer have east boundary punctures when considering the bubble tree $\mathbb{B}$ of the nodal curve; instead, there is one and only one interior puncture. With this, the rest of the proof follows exactly as in Proposition 2.57. ∎ ###### Proposition 2.69. Let $q$ be a self-intersection point of $\alpha_{im}$ and let $B\in T(q)$ be a stabilized teardrop with acute corner. For a generic nearly symmetric almost complex structure $J$, the moduli space of degenerate disks $\mathcal{N}_{J}^{B}(q;U)$ is a $(g-1)$-manifold, and a generic fiber of the evaluation map $ev_{J}:\mathcal{N}_{J}^{B}(q;U)\rightarrow\alpha^{a}_{1}\times\alpha^{c}_{1}\times\cdots\times\alpha_{g-2}^{c}$ is a compact 0-dimensional manifold consisting of an odd number of points. ###### Proof. The regularity of $\mathcal{N}_{J}^{B}(q;U)$ and compactness of a generic fiber are proved in the same way as in Proposition 2.69. The parity of the cardinality of a generic fiber follows from a similar neck- stretching and cobordism argument as in Proposition 2.69, using Lemma 2.70 below instead of Lemma 2.59. ∎ ###### Lemma 2.70. Assume $g(\Sigma)=2$. Fix some point $p\in\alpha^{a}_{1}$. For a sufficiently stretched almost complex structure $j$ on $\Sigma$, the fiber $ev_{Sym^{2}(j)}^{-1}(p)$ is transversally cut out and consists of one point. ###### Proof. View $\Sigma$ be the connected sum $(E_{1},\alpha^{a}_{1},\alpha^{a}_{2})$ and $(E_{2},\alpha_{im})$, where $E_{1}$ is the punctured Riemann surface of genus one and $E_{2}$ is a closed Riemann surface of genus one. Let $z^{\prime}$ denote the points on $E_{1}$ and $E_{2}$ where the connected sum is performed. The domain $B$ gives rise to a teardrop domain $B^{\prime}$ in $E_{2}$ with $n_{z^{\prime}}(B^{\prime})=1$. The Riemann mapping theorem implies that the moduli space $\mathcal{N}^{B^{\prime}}(q)$ of holomorphic disks in $E_{2}$ with corner $q$ and domain $B^{\prime}$ is smoothly cut out and has only one element. The gluing argument in Section 10 of [OS04b] shows for a sufficiently stretched almost complex structure, maps in $ev_{Sym^{2}(j)}^{-1}(p)$ are obtained by splicing the one-punctured holomorphic sphere in $Sym^{2}(E_{1})$ passing through $(z^{\prime},p)$ and the holomorphic disk in $\mathcal{N}^{B^{\prime}}(q)$.666Strictly speaking, Ozsváth and Szabó’s argument concerns splicing closed holomorphic spheres while in our case the sphere is punctured, but this does not affect the argument applying to our case. Alternatively, we may treat one-punctured holomorophic disks or spheres as the corresponding object without interior punctures, but intersecting $\\{e\\}\times\Sigma_{\bar{e}}$ once in $Sym^{2}(\Sigma_{\bar{e}})$. In particular, $ev_{Sym^{2}(j)}^{-1}(p)$ is identified with $\mathcal{N}^{B^{\prime}}(q)$ and hence consists of only one element. ∎ #### 2.8.3. One-punctured boundary degeneration without corners ###### Proposition 2.71. In the assumption of Proposition 2.63, if a boundary degeneration without corner occurs, then: * (1) There is only one degenerate disk, and its domain $[B]$ is $[\Sigma]$. * (2) $\bm{x}=\bm{y}$. * (3) Such degenerate disk do not occur simultaneously with other types of degeneration. * (4) The number of ends corresponding to such boundary degeneration is even. ###### Proof. The proof of (1), (2), and (3) are straightforward modifications of that of Proposition 2.62 and are omitted. (4) follows from the standard gluing result and Proposition 2.72 below, which differs from the counterpart in the 0-P case. ∎ ###### Proposition 2.72. For a generic almost complex structure $J$, $\mathcal{N}^{[\Sigma]}_{J}(\bm{x};U)$ is a compact, 0-dimensional manifold that consists of an even number of points. ###### Proof. The argument for compactness and transversality is the same as in [OS04b, Proposition 3.14], which is the counterpart of Proposition 2.72 when the Heegaard surface is closed; we will omit this part. By a similar cobordism argument used in Proposition 2.58, we can reduce understanding the parity of the moduli space to the base case $g(\Sigma)=2$, which is addressed in Lemma 2.73 below. ∎ ###### Lemma 2.73. Assume $g(\Sigma)=2$. View $(\Sigma,\alpha_{1}^{a},\alpha^{a}_{2},\alpha_{im})=(E_{1},\alpha_{1}^{a},\alpha_{2}^{a})\\#(E_{2},\alpha_{im})$, where $E_{1}$ is a punctured Riemann surface of genus one and $E_{2}$ is a Riemann surface of genus one. If $j$ is a sufficiently stretched complex structure on $\Sigma$, then $\mathcal{N}^{[\Sigma]}_{Sym^{2}(j)}(\bm{x};U)$ is empty. ###### Proof. Otherwise, the same neck-stretching procedure as in Lemma 2.61 produces a limit nodal holomorphic curve $u_{\infty}:\mathbb{B}\rightarrow Sym^{2}(E_{1}\vee E_{2})$. It consists of a (possibly punctured) holomorphic disk $v$ that maps to $E_{1}\times E_{2}$ with boundary in $\mathbb{T}_{\alpha,1}$ and possibly some (possibly punctured) sphere bubbles in $Sym^{2}(E_{i})$, $i=1,2$. We claim $v$ must be a constant map. It is clear that $Pr_{E_{1}}\circ v$ is constant, for $\pi_{2}(E_{1,\bar{e}},\alpha^{a}_{1}\cup\\{e\\})=0$, where $E_{1,\bar{e}}$ denote the Riemann surface obtained by filling in the east puncture. We move to see $Pr_{E_{2}}\circ v$ is constant. Suppose $Pr_{E_{2}}\circ v$ is not a constant map. Note the domain of $(Pr_{E_{2}}\circ v)$ is a zero-cornered $\alpha$-bounded domain $D$ in $E_{2}$. Stabilizing by $E_{1}$, this domain induces a zero-cornered $\alpha$-bounded domain $D^{\prime}$ in $\Sigma$ with $n_{z}(D^{\prime})\leq 1$. If $n_{z}(D^{\prime})=0$, then $D^{\prime}$ does not exist as $\mathcal{H}$ is unobstructed, and hence $D$ does not exist. So $n_{z}(D^{\prime})=1$, and hence $D^{\prime}=\Sigma$ since $\mathcal{H}$ is unobstructed. This implies $D=E_{2}$. Therefore, $\partial(Pr_{E_{2}}\circ v)$ is null-homotopic in $\alpha_{im}$. So $Pr_{E_{2}}\circ v$ induces a nontrivial element in $\pi_{2}(E_{2})$. This, however, contradicts that $\pi_{2}(E_{2})=0$. Therefore, $Pr_{E_{2}}\circ v$ is also constant, and hence $v$ is the constant map with image $\bm{x}$. Now $\\{\bm{x}\\}$ intersects neither $Sym^{2}(E_{i})$, $i=1,2$, and hence there are no sphere bubbles in $u_{\infty}$. So the Gromov limit $u_{\infty}$ is a constant map. In particular, $n_{z}(u_{\infty})=0$. However, $n_{z}(u_{\infty})=1$ as it is the limit of a sequence of holomorphic maps whose multiplicity at $z$ is one. This is a contradiction. Therefore, $\mathcal{N}^{[\Sigma]}_{Sym^{2}(j)}(\bm{x};U)$ is empty provided $j$ is sufficiently stretched. ∎ #### 2.8.4. Proof of Proposition 2.63 ###### Proof of Proposition 2.63. In view of Proposition 2.40, Proposition 2.68, and Proposition 2.71 we know the degenerations that can appear in the boundary of the compactified moduli spaces are two-story curves, simple combs with orbit curve ends, or simple boundary degenerations with or without corners. In all cases, gluing arguments can be applied to see the compactified moduli space $\overline{\mathcal{M}}^{B}(\bm{x},\bm{y};U)$ is a one-manifold with boundary. For conclusion (a), note that ends of type (2) correspond to pairs of curves $(u,v)$ where $u$ is in $\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{1230})$ or $\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{3012})$ and $v$ is an orbit curve, but the moduli space of orbit curves consists of a single element by the Riemann mapping theorem so the count of type (2) boundaries agrees with $\\#\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{1230})+\\#\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{3012})$. For conclusion (b), standard gluing results imply that the number of such ends is equal to $\sum_{\\{(q,B_{1})|\exists B_{2}\in T(q),B_{1}+B_{2}=B\\}}\\#(\mathcal{M}^{B_{1}}(\bm{x},\bm{y};q)\times_{ev}\mathcal{N}^{B_{2}}(q;U))$ This is mod 2 equal to $\sum_{\\{(q,B_{1})|\exists B_{2}\in T(q),B_{1}+B_{2}=B\\}}\\#\mathcal{M}^{B_{1}}(\bm{x},\bm{y};q)$ as a generic fiber of $ev$ in $\mathcal{N}^{B_{2}}(q;U)$ is odd by Proposition 2.69. For (c), note by gluing results the number of such ends is equal to $\\#\mathcal{N}^{[\Sigma]}(\bm{x};U)$. This is even by Proposition 2.72. ∎ ### 2.9. Type D structures We define type D structures from an immersed bordered Heeggard diagram $\mathcal{H}=(\Sigma,\bm{\beta},\bm{\bar{\alpha}},z)$ in this subsection. Figure 7. The quiver presentation of the torus algebra (left) and the pointed match circle of $\mathcal{H}$ with reversed boundary orientation (right). Let $\mathcal{A}$ denote the torus algebra, which is isomorphic to the quiver algebra of the quiver in Figure 7 (left). For $I\in\\{1,2,3,12,23,123\\}$, $\rho_{I}\in\mathcal{A}$ is understood as the product of the $\rho_{i}$’s for those $i$ appear in $I$. This algebra arises naturally in the context of bordered Heegaard diagrams, where $\mathcal{A}$ is associated to the pointed match circle determined by $\mathcal{H}$ with the reversed boundary orientation (Figure 7 (right)); we refer the readers to [LOT18, Chapter 11.1] for a detailed definition of the torus algebra in terms of pointed match circles, and we only point out that the element $\rho_{I}\in\mathcal{A}$ for $I\in\\{1,2,3,12,23,123\\}$ corresponds to the Reeb chord with the same label on the pointed match circle. Let $\mathcal{I}=\langle\iota_{0}\rangle\oplus\langle\iota_{1}\rangle$ denote the ring of idempotents of $\mathcal{A}$. We recall the definition of a type D structure. ###### Definition 2.74. A type D structure over the torus algebra $\mathcal{A}$ is a left $\mathcal{I}$-module $N$ together with a linear map $\delta:N\rightarrow\mathcal{A}\otimes N$ such that the map $\partial\coloneqq(\mu_{\mathcal{A}}\otimes\mathbb{I}_{N})\circ(\mathbb{I}_{\mathcal{A}}\otimes\delta):\mathcal{A}\otimes N\rightarrow\mathcal{A}\otimes N$ is a differential, i.e., $\partial^{2}=0$. The left differential $\mathcal{A}$-module $\mathcal{A}\otimes N$ is called the type D module of the type D structure $(N,\delta)$. Next, we spell out the construction of a type D structure from an immersed bordered Heegaard diagram. Recall $\mathbb{T}_{\beta}=\beta_{1}\times\cdots\times\beta_{g}$ and $\mathbb{T}_{\alpha,i}=\alpha^{a}_{i}\times\alpha_{1}^{c}\times\cdots\times\alpha_{g-1}^{c}$, $i=1,2$. Let $\mathbb{T}_{\alpha}=\mathbb{T}_{\alpha,1}\cup\mathbb{T}_{\alpha,2}$. Let $\mathcal{G}(\mathcal{H})=\\{\bm{x}|\bm{x}\in\mathbb{T}_{\alpha}\cap\mathbb{T}_{\beta}\\}$. Denote the local system on $\alpha_{im}$ as a vector bundle $\mathcal{E}\rightarrow\alpha_{im}$ together with a parallel transport $\Phi$. Note that this induces a local system on $\mathbb{T}_{\alpha}$, the tensor product of $\mathcal{E}$ and the trivial local system on the other alpha curves (or arcs). Abusing notation, we still denote the local system on $\mathbb{T}_{\alpha}$ by $(\mathcal{E},\Phi)$. Now define an $\mathcal{I}$-module $X^{\mathcal{E}}(\mathcal{H})=\oplus_{\bm{x}\in\mathcal{G}(\mathcal{H})}\mathcal{E}|_{\bm{x}}$, where the $\mathcal{I}$-action on an element $\eta\in\mathcal{E}|_{\bm{x}}$ is specified by $\iota_{i}\cdot\eta=\begin{cases}\eta,\ o(\bm{x})\equiv i\pmod{2}\\\ 0,\ \text{otherwise}\end{cases}$ Here $o(\bm{x})=i$ if and only if $\bm{x}\in\mathbb{T}_{\alpha,i}$, $i=1,2$. Given a sequence of Reeb chords $\overrightarrow{\sigma}=(\sigma_{1},\ldots,\sigma_{k})$ of a pointed match circle $\mathcal{Z}$, $a(-\overrightarrow{\sigma})$ is defined to be $(-\sigma_{1})\cdot(-\sigma_{2})\cdot\ldots(-\sigma_{k})\in\mathcal{A}(-\mathcal{Z})$. Note given $B\in\pi_{2}(\bm{x},\bm{y})$, the parallel transport restricted to the arc $\partial_{\alpha_{im}}B\subset\alpha_{im}$ induces an isomorphism from $\mathcal{E}|_{\bm{x}}$ to $\mathcal{E}|_{\bm{y}}$, which we denote by $\Phi^{B}_{\bm{x},\bm{y}}$. ###### Definition 2.75. Let $\mathcal{H}$ be an unobstructed, provincially admissible, immersed bordered Heegaard diagram. Fix a generic almost complex structure on $\Sigma\times[0,1]\times\mathbb{R}$. The type D module $\widehat{CFD}(\mathcal{H})$ is defined to be the $\mathcal{A}$-module $\mathcal{A}\otimes_{\mathcal{I}}X^{\mathcal{E}}(\mathcal{H})$ together with a differential given by $\partial(a\otimes\eta)=a\cdot(\sum_{\bm{y}}\sum_{\\{(B,\overrightarrow{\sigma})|\ n_{z}(B)=0,\ \text{ind}(B,\overrightarrow{\sigma})=1\\}}\\#\mathcal{M}^{B}(\bm{x},\bm{y};\overrightarrow{\sigma})a(-\overrightarrow{\sigma})\otimes\Phi^{B}_{\bm{x},\bm{y}}\eta),$ where $a\in\mathcal{A}$, $\eta\in\mathcal{E}|_{\bm{x}}$, and the pairs $(B,\overrightarrow{\sigma})$ are compatible. The underlying type D structure is the pair $(X^{\mathcal{E}}(\mathcal{H}),\delta)$ where $\delta(\eta)\coloneqq\partial(1\otimes\eta)$ for any $\eta\in X^{\mathcal{E}}(\mathcal{H})$. Abusing notation, we will also use $\widehat{CFD}(\mathcal{H})$ to denote its underlying type D structure. ###### Remark 2.76. Note when the local system is trivial, we can identify $\bm{x}$ with $\mathcal{E}|_{\bm{x}}$, and the differential defined above can be more conveniently written as $\partial(a\otimes\bm{x})=a\cdot(\sum_{\bm{y}}\sum_{\\{(B,\overrightarrow{\sigma})|\ n_{z}(B)=0,\ \text{ind}(B,\overrightarrow{\sigma})=1\\}}\\#\mathcal{M}^{B}(\bm{x},\bm{y};\overrightarrow{\sigma})a(-\overrightarrow{\sigma})\otimes\bm{y}).$ ###### Proposition 2.77. The operator $\partial$ in Definition 2.75 is well-defined and $\partial^{2}=0$. ###### Proof. We first point out $\partial$ is well-defined, i.e., the sum defining $\partial$ is finite. This reduces to the provincial admissibility of $\mathcal{H}$, which implies there are only finitely many positive domains with prescribed Reeb chords connecting any given pair of generators. The proof is standard, and we do not repeat it here. We move to see $\partial^{2}(\bm{x})=0$. For ease of explanation, we begin with the case of trivial local systems. Let $a$ be a non-zero element of $\mathcal{A}$, and let $\langle\partial^{2}\bm{x},a\bm{y}\rangle\in\mathbb{F}$ denote the coefficient of the term $a\bm{y}$ in $\partial^{2}\bm{x}$. Then (2.7) $\langle\partial^{2}\bm{x},a\bm{y}\rangle=\sum_{\bm{w}\in\mathcal{G}}\sum\\#\mathcal{M}^{B_{1}}(\bm{x},\bm{w};\overrightarrow{\sigma_{1}})\\#\mathcal{M}^{B_{2}}(\bm{w},\bm{y};\overrightarrow{\sigma_{2}}),$ where the second sum is over all the index-one compatible pairs $(B_{i},\overrightarrow{\sigma_{i}})$ ($i=1,2$) with $a(-\overrightarrow{\sigma_{1}})\cdot a(-\overrightarrow{\sigma_{2}})=a$. In view of Proposition 2.47 and the gluing result, the right-hand side of Equation (2.7) is $\sum_{\\{(B,\overrightarrow{\sigma})|\text{ind}(B,\overrightarrow{\sigma})=2,\ \ a(-\overrightarrow{\sigma})=a\\}}\\#\partial\overline{\mathcal{M}}^{B}(\bm{x},\bm{y};\overrightarrow{\sigma})\equiv 0\pmod{2}$ This finishes the proof in the case of trivial local systems. For the case of non-trivial local systems, the proof is a slight modification of the above argument. One needs to note that given $B_{1}\in\pi_{2}(\bm{x},\bm{w})$ and $B_{2}\in\pi_{2}(\bm{w},\bm{y})$, we have $\Phi_{\bm{x},\bm{w}}^{B_{1}}\circ\Phi_{\bm{w},\bm{y}}^{B_{2}}=\Phi_{\bm{x},\bm{y}}^{B_{1}+B_{2}}$. Therefore, given an $\eta\in\mathcal{E}|_{\bm{x}}$, the terms in $\partial^{2}(\eta)$ corresponding to two-story ends of a one-dimensional moduli space $\mathcal{M}^{B}(\bm{x},\bm{y};\overrightarrow{\sigma})$ are multiples of the same element in $\mathcal{E}|_{\bm{y}}$, namely $\Phi_{\bm{x},\bm{y}}^{B}(\eta)$, and hence the coefficient is zero mod $2$. ∎ ### 2.10. Weakly extended Type D structures We define the weakly extended type D structure $\widetilde{CFD}(\mathcal{H})$ in this subsection. The weakly extended torus algebra $\tilde{\mathcal{A}}$ can be represented by the quiver with relations shown in Figure 8. Figure 8. The weakly extended torus algebra. The subscripts in the relation are understood mod $4$. Note as in the torus algebra, we have the idempotent ring $\mathcal{I}=\langle\iota_{0}\rangle\oplus\langle\iota_{1}\rangle$. Let $\bm{U}$ be $\rho_{0123}+\rho_{1230}+\rho_{2301}+\rho_{3012}$, which is a central element of $\tilde{\mathcal{A}}$. ###### Definition 2.78. A weakly extended type D structure over $\tilde{\mathcal{A}}$ is a left $\mathcal{I}$-module $N$ together with a linear map $\tilde{\delta}:N\rightarrow\tilde{\mathcal{A}}\otimes N$ such that the map $\tilde{\partial}\coloneqq(\mu_{\tilde{\mathcal{A}}}\otimes\mathbb{I}_{N})\circ(\mathbb{I}_{\tilde{\mathcal{A}}}\otimes\tilde{\delta}):\tilde{\mathcal{A}}\otimes N\rightarrow\tilde{\mathcal{A}}\otimes N$ squares to $\bm{U}$, i.e. $\tilde{\partial}^{2}=\bm{U}$. The curved left $\tilde{\mathcal{A}}$-module $\tilde{\mathcal{A}}\otimes N$ is called the weakly extended type D module of the weakly extended type D structure $(N,\tilde{\delta})$. Let $X^{\mathcal{E}}(\mathcal{H})$ be the $\mathcal{I}$-module defined the same way as in Section 2.9. ###### Definition 2.79. Let $\mathcal{H}$ be an unobstructed, provincially admissible, immersed bordered Heegaard diagram. Fix a generic admissible almost complex structure on $\Sigma\times[0,1]\times\mathbb{R}$. The weakly extended type D module $\widetilde{CFD}(\mathcal{H})$ is defined to be the $\tilde{\mathcal{A}}$-module $\tilde{\mathcal{A}}\otimes_{\mathcal{I}}X^{\mathcal{E}}(\mathcal{H})$ together with a differential given by $\tilde{\partial}(a\otimes\eta)=a\cdot(\sum_{\bm{y}}\sum_{\\{(B,\overrightarrow{\sigma})|\ \text{ind}(B,\overrightarrow{\sigma})=1\\}}\\#\mathcal{M}^{B}(\bm{x},\bm{y};\overrightarrow{\sigma})a(-\overrightarrow{\sigma})\otimes\Phi^{B}_{\bm{x},\bm{y}}\eta),$ where $a\in\tilde{\mathcal{A}}$, $\eta\in\mathcal{E}|_{\bm{x}}$, $\overrightarrow{\sigma}$ is a sequence of Reeb chords that include the case of a single closed Reeb orbit $\\{U\\}$ (in which case the corresponding moduli space consists of 1-P holomorphic curves), and the pairs $(B,\overrightarrow{\sigma})$ are compatible. When $\overrightarrow{\sigma}=\\{U\\}$, we define $a(-U)=\bm{U}$. The underlying weakly extended type D structure is $(X^{\mathcal{E}}(\mathcal{H}),\tilde{\delta})$ where $\tilde{\delta}(\eta)\coloneqq\tilde{\partial}(1\otimes\eta)$. ###### Remark 2.80. Again, by abusing notation, we also use $\widetilde{CFD}(\mathcal{H})$ to denote the underlying weakly extended type D structure. When the local system is trivial, we have the following more familiar formula for the differential: $\tilde{\partial}(a\otimes\bm{x})=a\cdot(\sum_{\bm{y}}\sum_{\\{(B,\overrightarrow{\sigma})|\text{ind}(B,\overrightarrow{\sigma})=1\\}}\\#\mathcal{M}^{B}(\bm{x},\bm{y};\overrightarrow{\sigma})a(-\overrightarrow{\sigma})\otimes\bm{y}).$ ###### Proposition 2.81. The operator $\tilde{\partial}$ in Definition 2.79 is well-defined and $\tilde{\partial}^{2}=\bm{U}$. ###### Proof. A standard argument shows that the provincial admissibility of $\mathcal{H}$ implies the sum defining $\tilde{\partial}$ in Definition 2.79 is finite, and hence $\tilde{\partial}$ is well-defined. Next, we show $\tilde{\partial}^{2}=\bm{U}$. Once again, we first give the proof when the local system is trivial for conciseness. Recall the length of an element $a\in\tilde{\mathcal{A}}$ is the number of factors $\rho_{i}\in\\{\rho_{0},\rho_{1},\rho_{2},\rho_{3}\\}$ when we write $a$ as a product of the generators $\\{\iota_{0},\iota_{1},\rho_{0},\rho_{1},\rho_{2},\rho_{3}\\}$. (For example, $\rho_{123}$ has length $3$ and $\iota_{0}$ has length $0$.) For an element $a\in\tilde{\mathcal{A}}$ whose length is less than or equal to $3$, the proof of Proposition 2.77 carries over to show $\langle\tilde{\partial}^{2}\bm{x},a\bm{y}\rangle=0$ for any $\bm{x}$ and $\bm{y}$ (by permuting the region we where we put the base point $z$). We are left to consider the case where the algebra element is of length $4$. We claim that for a generator $\bm{x}$ such that $\iota_{1}\cdot\bm{x}=\bm{x}$, we have $\langle\tilde{\partial}^{2}\bm{x},\rho_{0123}\bm{y}\rangle=\begin{cases}0,\ if\ \bm{x}\neq\bm{y},\\\ 1,\ if\ \bm{x}=\bm{y}.\end{cases}$ Assuming this claim, by permuting the subscripts we also have that $\langle\tilde{\partial}^{2}\bm{x},\rho_{2301}\bm{y}\rangle$ is $1$ if $\bm{x}=\bm{y}$ and $0$ otherwise, and an idempotent consideration shows $\langle\tilde{\partial}^{2}\bm{x},a\bm{y}\rangle=0$ when $a\in\\{\rho_{1230},\rho_{3012}\\}$. These together imply $\tilde{\partial}^{2}\bm{x}=\bm{U}\cdot\bm{x}$ when $\iota(\bm{x})=\iota_{1}$. A similar consideration shows this is true for $\bm{x}$ with $\iota(\bm{x})=\iota_{0}$ as well. This finishes the proof of the proposition modulo the claim. Next, we prove the claim. Note (2.8) $\langle\tilde{\partial}^{2}\bm{x},\rho_{0123}\bm{y}\rangle=\sum_{\bm{w}}\sum_{\begin{subarray}{c}\text{ind}(B_{i},\overrightarrow{\sigma_{i}})=1,\\\ i=1,2\end{subarray}}\\#\mathcal{M}^{B_{1}}(\bm{x},\bm{w};\overrightarrow{\sigma_{1}})\\#\mathcal{M}^{B_{2}}(\bm{w},\bm{y};\overrightarrow{\sigma_{2}}),$ where $(B_{i},\overrightarrow{\sigma_{i}})$ ($i=1,2$) is compatible and $a(-\overrightarrow{\sigma_{1}})a(-\overrightarrow{\sigma_{1}})=\rho_{0123}$ or $\bm{U}$; the possible pairs of $(\overrightarrow{\sigma_{1}},\overrightarrow{\sigma_{2}})$ are listed below: $\displaystyle\bigg{\\{}(\emptyset,\\{-\rho_{0},-\rho_{1},-\rho_{2},-\rho_{3}\\}),(\\{-\rho_{0}\\},\\{-\rho_{1},-\rho_{2},-\rho_{3}\\}),(\\{-\rho_{0},-\rho_{1},-\rho_{2}\\},\\{-\rho_{3}\\}),$ $\displaystyle(\\{-\rho_{0},-\rho_{1}\\},\\{-\rho_{2},-\rho_{3}\\}),(\\{-\rho_{0},-\rho_{1},-\rho_{2},-\rho_{3}\\},\emptyset),(\emptyset,\\{-\rho_{0},-\rho_{123}\\}),$ $\displaystyle(\\{-\rho_{0}\\},\\{-\rho_{123}\\}),(\\{-\rho_{0},-\rho_{123}\\},\emptyset),(\emptyset,\\{-\rho_{012},-\rho_{3}\\}),(\\{-\rho_{012}\\},\\{-\rho_{3}\\}),$ $\displaystyle(\\{-\rho_{012},-\rho_{3}\\},\emptyset),(\emptyset,\\{U\\}),(\\{U\\},\emptyset)\bigg{\\}}.$ Let $\overline{\mathcal{M}}_{0}\coloneqq\cup_{\text{ind}(B,\overrightarrow{\sigma})=2}\overline{\mathcal{M}}^{B}(\bm{x},\bm{y};\overrightarrow{\sigma}),$ where $\overrightarrow{\sigma}\in\\{(-\rho_{0},-\rho_{1},-\rho_{2},-\rho_{3}),(-\rho_{0},-\rho_{123}),(-\rho_{012},-\rho_{3})\\}$. Let $\overline{\mathcal{M}}_{1}:=\cup_{\text{ind}(B,U)=2}\overline{\mathcal{M}}^{B}(\bm{x},\bm{y};U).$ Equation 2.8 and the gluing result imply that $\langle\tilde{\partial}^{2}\bm{x},\rho_{0123}\bm{y}\rangle$ is equal to the number of two-story ends of the moduli space $\overline{\mathcal{M}}_{0}\cup\overline{\mathcal{M}}_{1}$. According to Proposition 2.49, the other elements in $\partial\overline{\mathcal{M}}_{0}$ are: * (A-1) simple holomorphic combs with a single split component; * (A-2) simple boundary degenerations with one corner; * (A-3) simple boundary degenerations without corners. Proposition 2.63 shows the other boundary points in $\overline{\mathcal{M}}_{1}$ in addition to two-story ends are: * (B-1) simple holomorphic combs with an orbit curve; * (B-2) simple boundary degenerations with one corner; * (B-3) simple boundary degenerations without corners. Note by Proposition 2.49 and Proposition 2.63, the number of boundary points of type (A-1) is equal to that of type (B-1), both of which is $\sum_{\text{ind}(B,-\rho_{3012})=1}\\#\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{3012})+\sum_{\text{ind}(B,-\rho_{1230})=1}\\#\mathcal{M}^{B}(\bm{x},\bm{y};-\rho_{1230}).$ The parity of the number of boundary points of type (A-2) is equal to that of type (B-2), which are both mod $2$ equal to $\sum_{q}\sum_{\\{(B_{1},B_{2})|B_{2}\in T(q),\ \text{ind}(B_{1}+B_{2};U)=2\\}}\\#\mathcal{M}^{B_{1}}(\bm{x},\bm{y};q),$ where $q$ ranges over self-intersection points of $\alpha_{im}$, and $T(q)$ denotes the set of stabilized teardrops at $q$. The parity of the number of boundary points of type (B-3) is even according to Proposition 2.63. In summary, the parity of the number of boundary points of $\overline{\mathcal{M}}_{0}\cup\overline{\mathcal{M}}_{1}$ corresponding to two-story ends is equal to that of type (A-3), which is odd if and only if $\bm{x}=\bm{y}$ by Proposition 2.49. Therefore, $\langle\tilde{\partial}^{2}\bm{x},\rho_{0123}\bm{y}\rangle$ is odd if and only if $\bm{x}=\bm{y}$, finishing the proof of the claim. In the presence of non-trivial local systems, we simply need to consider the above argument for each domain. For a domain $B$, let $\overline{\mathcal{M}}_{0}^{B}$ be the subset of $\overline{\mathcal{M}}_{0}$ consisting of holomorphic curves with domain $B$, and similarly define $\overline{\mathcal{M}}_{1}^{B}$. The two-story ends in $\overline{\mathcal{M}}_{0}^{B}\cup\overline{\mathcal{M}}_{1}^{B}$ all correspond to the same parallel transport. When ends of type (A-3) do not occur, the two-story ends cancel in pairs by the same argument as above. When ends of type (A-3) appear, we have $B=[\Sigma]$ and $\sigma=(-\rho_{0},-\rho_{1},-\rho_{2},-\rho_{3})$. In particular, $\partial_{\alpha_{im}}B=\emptyset$, which induces the identity endomorphism of $\mathcal{E}|_{\bm{x}}$. Also, the number of two-story ends is odd as the number of (A-3) ends is odd. The claim follows from these. ∎ There is a canonical quotient map $\pi:\tilde{\mathcal{A}}\rightarrow\mathcal{A}$. We say a weakly extended type D structure $(N,\tilde{\delta})$ extends a type D structure $(N^{\prime},\delta)$ if $(N^{\prime},\delta)$ is isomorphic to $(N,(\pi\otimes\mathbb{I}_{N})\circ\tilde{\delta})$. Clearly, $\widetilde{CFD}(\mathcal{H})$ extends $\widehat{CFD}(\mathcal{H})$ when both are defined. ### 2.11. Invariance In this subsection, we address the invariance of the (weakly extended) type D structures. ###### Proposition 2.82. The homotopy type of the type D structure defined in Definition 2.75 is independent of the choice of the almost complex structure and is invariant under isotopy of the $\alpha$\- or $\beta$-curves. ###### Remark 2.83. We do not need the invariance under handleslides and stabilizations for our applications. We only need to prove invariance when perturbing diagrams to obtain nice diagrams, and this only requires isotopies. ###### Proof of Proposition 2.82. The standard proof in Section 6.3 of [LOT18] carries over. For instance, to prove independence of almost complex structures, one first constructs a continuation map by counting holomorphic curves in $\Sigma\times[0,1]\times\mathbb{R}$ for a generic almost complex structure $J$ that interpolates two admissible almost complex structures $J_{0}$ and $J_{1}$. Then, one proves the continuation map is a chain map by analyzing the ends of one-dimensional moduli spaces. The only possible complication comes from boundary degenerations since $\alpha_{im}$ is immersed. However, this does not happen as $\mathcal{H}$ is unobstructed and the holomorphic curves have $n_{z}=0$. Therefore, no new phenomenon appears in the degeneration of moduli spaces, and hence the proof stays the same. ∎ ###### Proposition 2.84. The homotopy type of the weakly extended type D structure defined in Definition 2.79 is independent of the choice of the almost complex structure and is invariant under isotopy of the $\alpha$\- and $\beta$-curves. ###### Proof. One could prove this proposition similarly to the previous one. However, such an approach would require generalizing the analysis of the ends of moduli spaces in Proposition 6.20 of [LOT18] and hence is slightly tedious to write down. Here we give a different approach. Let $\mathcal{H}$ denote the immersed bordered Heegaard diagram. By Proposition 2.82, we know the homotopy type of $\widehat{CFD}(\mathcal{H})$ is independent of the choice of almost complex structures and isotopy of the $\alpha$\- or $\beta$-curves. Since $\widetilde{CFD}(\mathcal{H})$ extends $\widehat{CFD}(\mathcal{H})$ and that such extension is unique up to homotopy by Proposition 38 of [HRW23], we know the homotopy type of $\widetilde{CFD}(\mathcal{H})$ is also independent of the choice of almost complex structures and isotopy of the $\alpha$\- and $\beta$-curves. ∎ ## 3\. Knot Floer homology of immersed Heegaard diagrams This section defines knot Floer chain complexes of immersed Heegaard diagrams and proves the homotopy invariance under Heegaard moves. ### 3.1. Immersed doubly-pointed Heeggard diagram ###### Definition 3.1. An _immersed doubly-pointed Heegaard diagram_ is a 5-tuple $\mathcal{H}_{w,z}=(\Sigma,\bm{\alpha},\bm{\beta},w,z)$ where * (1) $\Sigma$ is a closed oriented surface of genus $g$. * (2) $\bm{\alpha}=\\{\alpha_{1},\ldots,\alpha_{g-1},\alpha_{g}\\}$, where $\alpha_{1},\ldots,\alpha_{g-1}$ are embedded disjoint curves in $\Sigma$ and $\alpha_{g}=\\{\alpha_{g}^{1},\ldots,\alpha_{g}^{n}\\}$ is a collection of immersed curves decorated with local systems. Moreover, $\alpha_{i}$ ($i=1,\ldots,g-1$) are disjoint from $\alpha_{g}$, $\alpha_{g}^{1}$ has the trivial local system, and $\\{\alpha_{1},\ldots,\alpha_{g-1},\alpha_{g}^{1}\\}$ induce linearly independent elements in $H_{1}(\Sigma,\mathbb{Z})$. We also assume that $\alpha_{g}^{i}$ is trivial in $H_{1}(\Sigma,\mathbb{Z})/\langle\alpha_{1},\ldots,\alpha_{g-1}\rangle$ for $i>1$. For convenience, we also denote $\alpha_{g}$ by $\alpha_{im}$. * (3) $\bm{\beta}=\\{\beta_{1},\ldots,\beta_{g}\\}$ are embedded disjoint curves in $\Sigma$ which induce linearly independent elements in $H_{1}(\Sigma,\mathbb{Z})$. * (4) $w$ and $z$ are base points such that they both lie in a single connected region in the complement of $\alpha$-curves as well as a single region in the complement of $\beta$-curves. Domains, periodic domains, and $\alpha$-bounded domains are defined similarly in this setting as for bordered Heegaard diagrams (by ignoring $\alpha$-arcs and surface boundary). We make a similar but slightly different definition of unobstructedness and admissibility below. ###### Definition 3.2. Given an immersed doubly-pointed Heegaard diagram, $\bm{\alpha}$ is called _unobstructed_ if there are no nontrivial zero- or one-cornered $\alpha$-bounded domains $B$ with $n_{z}(B)=0$ (or equivalently $n_{w}(B)=0$). An immersed doubly-pointed Heegaard diagram is called unobstructed if $\bm{\alpha}$ is unobstructed. ###### Definition 3.3. An immersed doubly-pointed Heegaard diagram is _bi-admissible_ if any nontrivial periodic domain $B$ with $n_{z}(B)=0$ or $n_{w}(B)=0$ has both positive and negative coefficients. We remark that the restriction to having only one immersed multicurve in the definition of immersed doubly-pointed Heegaard diagrams is not essential. ### 3.2. The knot Floer chain complex We define the knot Floer chain complex of an immersed Heegaard diagram similar to that in the ordinary setup. The only modification is that we only count stay-on-track holomorphic curves. The definition and analysis of moduli spaces in this setup is a straightforward modification of that in the previous section; it is even simpler as we do not need to care about east punctures. We hence do not repeat the moduli space theory but only mention the key properties when we need them. We will let $\mathcal{G}(\mathcal{H}_{w,z})$ denote the set of generators, which are $g$-tuples $(x_{1},\ldots,x_{g})$ such that $x_{i}\in\alpha_{i}\cap\beta_{\sigma(i)}$ $(i=1,\ldots,g)$ where $\sigma$ is a permutation of $\\{1,\ldots,g\\}$. Let $\mathcal{R}=\mathbb{F}[U,V]/(UV)$. Implicit in the definition below is that we choose a generic admissible almost complex structure $J$ on $\Sigma\times[0,1]\times\mathbb{R}$. ###### Definition 3.4. Let $\mathcal{H}_{w,z}$ be an unobstructed and bi-admissible immersed doubly- pointed Heegaard diagram. $CFK_{\mathcal{R}}(\mathcal{H}_{w,z})$ is the free $\mathcal{R}$-module generated over $\mathcal{G}(\mathcal{H}_{w,z})$ with differential $\partial$ defined as $\partial\bm{x}=\sum_{y}\sum_{B\in\pi_{2}(\bm{x},\bm{y}),\ \text{ind}(B)=1}\\#\mathcal{M}^{B}(\bm{x},\bm{y})U^{n_{w}(B)}V^{n_{z}(B)}\bm{y},$ where $\bm{x},\bm{y}\in\mathcal{G}$. ###### Remark 3.5. Here we only give the definition assuming the local system on $\alpha_{im}$ is trivial. The case in which the local system is non-trivial is only notationally more complicated, and we leave it for the interested readers to work out. See Definition 2.75 for an example. ###### Proposition 3.6. $(CFK_{\mathcal{R}}(\mathcal{H}_{w,z}),\partial)$ is a chain complex, i.e., $\partial^{2}=0$. ###### Proof. The same proof for Proposition 2.77 works here. Note we will only use moduli spaces with domains $B$ such that $n_{w}(B)=0$ or $n_{z}(B)=0$, and the unobstructedness of $\mathcal{H}_{w,z}$ excludes the possibility of boundary degeneration in the compactified 1-dimensional moduli space supported in such domains. Hence, an analogue version of Proposition 2.47 holds. With this observation, the proof of Proposition 2.77 carries over. ∎ ### 3.3. Bi-grading We would like to consider gradings on knot Floer chain complexes. ###### Definition 3.7. A (possibly immersed) doubly-pointed Heegaard diagram is gradable if all non- trivial periodic domain $P$ satisfies $\text{ind}(P)-2n_{z}(P)=0$ and $\text{ind}(P)-2n_{w}(P)=0$, where $\text{ind}(-)$ is defined in Definition 2.43. If $\mathcal{H}_{w,z}$ is gradable then the knot Floer chain complex $(CFK_{\mathcal{R}}(\mathcal{H}_{w,z}),\partial)$ admits a relative $\mathbb{Z}\oplus\mathbb{Z}$-grading, as described below. We will be interested in diagrams $\mathcal{H}_{w,z}$ for which $\widehat{HF}(\mathcal{H}_{w})\cong\widehat{HF}(\mathcal{H}_{z})\cong\mathbb{F}$, where $\widehat{HF}(\mathcal{H}_{w})$ and $\widehat{HF}(\mathcal{H}_{z})$ are homology groups of the chain complexes obtained from $CFK_{\mathcal{R}}(\mathcal{H}_{w,z})$ by setting $V=0$ and $U=1$ or $U=0$ and $V=1$, respectively. In this case we say that the horizontal and vertical homology has rank one. Gradable diagrams with this property can be given an absolute grading, as follows. ###### Definition 3.8. Let $\bm{x},\bm{y}\in\mathcal{G}(\mathcal{H}_{w,z})$ be two generators. Let $B\in\tilde{\pi}_{2}(\bm{x},\bm{y})$ be a domain. Then the $w$-grading difference between $\bm{x}$ and $\bm{y}$ is given by $gr_{w}(\bm{x})-gr_{w}(\bm{y})=\text{ind}(B)-2n_{w}(B),$ and the $z$-grading difference between $\bm{x}$ and $\bm{y}$ is given by $gr_{z}(\bm{x})-gr_{z}(\bm{y})=\text{ind}(B)-2n_{z}(B).$ If the horizontal and vertical homology of $\mathcal{H}_{w,z}$ is rank one, then the absolute $w$-grading is normalized so that $\widehat{HF}(\mathcal{H}_{w})$ is supported in $w$-grading $0$, and absolute $z$-grading is normalized so that $\widehat{HF}(\mathcal{H}_{z})$ is supported in $z$-grading $0$. Equivalently, one can equip $CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{K}))$ with the Maslov grading and the Alexander grading. These two gradings can be expressed in terms of the $w$-grading and $z$-grading: The Maslov grading is equal to the $z$-grading, and the Alexander grading is given by $\frac{1}{2}(gr_{w}-gr_{z})$. ###### Remark 3.9. The normalization conditions for the absolute gradings are chosen so that the bi-graded chain complexes model those associated to knots in the 3-sphere. ### 3.4. Invariance We will show knot Floer chain complexes defined over immersed Heegaard diagrams satisfy similar invariance properties when varying the almost complex structure or modifying the Heegaard diagram by isotopy, handleslides, and stabilizations. While the meaning of isotopy and stabilization are obvious for immersed Heegaard diagrams, we give a remark on handleslides. ###### Remark 3.10. When speaking of handleslides of an immersed Heegaard diagram $\mathcal{H}_{w,z}$, we only allow an $\alpha$-curve to slide over another _embedded_ $\alpha$-curve, not over an immersed $\alpha$-curve. Furthermore, we point out that handle-slides do not change the unobstructedness, bi- admissibility, and gradability of the diagram. To see this, note periodic domains of two Heegaard diagrams before and after a handleslide are related. A periodic domain in the old Heegaard diagram with boundary on the arc that moves in the handleslide give rise to a periodic domain in the new Heegaard diagram by boundary summing a thin annulus (whose multiplicity can be one or negative one). In particular, if we started from a somewhere negative domain $B$, then the new domain $B^{\prime}$ after this procedure is still somewhere negative; it is also easy to see $\text{ind}(B)=\text{ind}(B^{\prime})$, $n_{z}(B)=n_{z}(B^{\prime})$, and $n_{w}(B)=n_{w}(B^{\prime})$, which implies the gradability of two diagrams are the same as well. ###### Proposition 3.11. Let $\mathcal{H}_{w,z}$ be an unobstructed, bi-admissible, and gradable immersed doubly-pointed Heegaard diagram. The bigraded chain homotopy type of $CFK_{\mathcal{R}}(\mathcal{H}_{w,z})$ is invariant under varying the almost complex structure, isotopy of the $\alpha$\- and $\beta$-curves, handleslides, and stabilization/destabilization. ###### Proof. The proof of the bigraded homotopy invariance under the variation of the almost complex structure, isotopy, and stabilization is the same as the corresponding results in the embedded-$\alpha$-curve set-up in [Lip06]. In fact, changing the $\alpha$-curves from embedded to immersed can only complicate the arguments in that boundary degeneration might appear as ends of the moduli spaces involved, yet the unobstructedness dispels such worries. Figure 9. The $\alpha$-curves in proving the handleslide invariance on a genus-two surface, which is represented as a torus obtained by identifying the edges of a square together with a handle attached to the two circles inside the square. The labels $\theta_{im}^{\pm}$ are used interchangeably with $\theta_{2}^{\pm}$. Similarly, $\theta_{im}^{H,\pm}$ and $\theta_{im}^{{}^{\prime}\pm}$ are the same as $\theta_{2}^{H,\pm}$ and $\theta_{2}^{{}^{\prime}\pm}$, respectively. (a) shows any self-intersection point $q_{i}$ of $\alpha_{im}$ induces two intersection points between $\alpha_{im}$ and its perturbation $\alpha^{\prime}_{im}$. (b) shows the small triangles showing $F(\Theta_{\alpha^{\prime},\alpha^{H}}\otimes\Theta_{\alpha^{H},\alpha})=\Theta_{\alpha^{\prime},\alpha}$ . The handleslide invariance can also be proved using the same strategy as in the embedded-$\alpha$-curve case with slightly more caution. The main difference is that in the embedded-$\alpha$-curve case, there is a unique maximal graded generator in the Heegaard Floer homology of a Heegaard diagram where the set of $\alpha$-curves is a small Hamiltonian perturbation of the $\beta$-curves. In contrast, such a generator needs to be specified more carefully in our case. We spell this out in more detail. Denote $\mathcal{H}=(\Sigma,\bm{\alpha},\bm{\beta},w,z)$. For clarity of exposition, assume $\alpha_{im}$ consists of a single component with a trivial local system and $n$ self-intersection points. We also restrict to the interesting case, in which the handleslide is sliding $\alpha_{im}$ over an embedded $\alpha$-curve. Let $\bm{\alpha^{\prime}}$ denote a small Hamiltonian perturbation of $\bm{\alpha}$ so that $\alpha_{i}\cap\alpha_{j}=\emptyset$ for $i\neq j$; for $i=1,\ldots,g-1$, the embedded curves $\alpha_{i}$ and $\alpha_{i}^{\prime}$ intersects exactly at two points $\\{\theta_{i}^{+},\theta_{i}^{-}\\}$; $\alpha_{im}$ intersects $\alpha_{im}^{\prime}$ at $2+2n$ points $\\{\theta_{g}^{+},\theta_{g}^{-},\xi_{1}^{+},\xi_{1}^{-},\ldots,\xi_{n}^{+},\xi_{n}^{-}\\}$, where $\xi_{i}^{\pm}$ are intersection points corresponding to the self- intersection points of $\alpha_{im}$. We label the $\theta$-intersection points using the convention so that $(\theta_{i}^{+},*)$ is of higher grading than $(\theta_{i}^{-},*)$ in $CFK_{\mathcal{R}}(\Sigma,\bm{\alpha^{\prime}},\bm{\alpha},w,z)$, $(i=1,\ldots,g)$ (see Figure 9 (a)). Let $\alpha_{im}^{H}$ denote the curve obtained by sliding $\alpha_{im}$ over, say, $\alpha_{g-1}$, so that $\alpha_{im}^{H}$ intersects each of $\alpha_{im}$ and $\alpha_{im}^{\prime}$ in $2+2n$ points; denote the $\theta$-intersection points by $\\{\theta_{g}^{H,+},\theta_{g}^{H,-}\\}$ and $\\{\theta_{g}^{{}^{\prime}+},\theta_{g}^{{}^{\prime}-}\\}$, respectively. Let $\alpha_{i}^{H}$ ($i=1,\ldots,g-1$) be small Hamiltonian perturbations of $\alpha_{i}^{\prime}$ so that $\alpha_{i}^{H}$ intersects each of $\alpha_{i}$ and $\alpha_{i}^{\prime}$ at exactly two points, denoted by $\\{\theta_{i}^{H,+},\theta_{i}^{H,-}\\}$ and $\\{\theta_{i}^{{}^{\prime}+},\theta_{i}^{{}^{\prime}-}\\}$, respectively. Let $\Theta_{\alpha^{\prime},\alpha}=(\theta_{1}^{+},\ldots,\theta_{g}^{+})$, $\Theta_{\alpha^{H},\alpha}=(\theta_{1}^{H,+},\ldots,\theta_{g}^{H,+})$, and $\Theta_{\alpha^{\prime},\alpha^{H}}=(\theta_{1}^{{}^{\prime}+},\ldots,\theta_{g}^{{}^{\prime}+})$. These correspond to the maximal graded intersection points used in the embedded case.777A straightforward computation would show $\Theta_{\alpha,\alpha^{\prime}}$ are indeed cycles in the Floer chain complex associated to the immersed Heegaard diagram $(\Sigma,\bm{\alpha}$, $\bm{\alpha^{\prime}},w,z)$; similar statements hold for $\Theta_{\alpha^{H},\alpha}$ and $\Theta_{\alpha^{\prime},\alpha^{H}}$. The rest of the proof is similar to the embedded case. We provide a sketch. Let $\mathcal{H}^{H}=(\Sigma,\bm{\alpha}^{H},\bm{\beta},w,z)$ and $\mathcal{H}^{\prime}=(\Sigma,\bm{\alpha^{\prime}},\bm{\beta},w,z)$. By counting holomorphic triangles (with stay-on-track boundaries), one can define chain maps $F(\Theta_{\alpha^{H},\alpha}\otimes-):CFK_{\mathcal{R}}(\mathcal{H})\rightarrow CFK_{\mathcal{R}}(\mathcal{H}^{H})$ and $F(\Theta_{\alpha^{\prime},\alpha^{H}}\otimes-):CFK_{\mathcal{R}}(\mathcal{H}^{H})\rightarrow CFK_{\mathcal{R}}(\mathcal{H}^{\prime})$ Again, the usual proof which shows the above maps are chain maps carries through, as the unobstructedness excludes boundary degeneration when analyzing the ends of one-dimensional moduli spaces of holomorphic triangles. Similarly, by analyzing ends one-dimensional moduli spaces of holomorphic quadrilaterals, one can show the composition of these two maps is chain homotopic equivalent to $F(F(\Theta_{\alpha^{\prime},\alpha^{H}}\otimes\Theta_{\alpha^{H},\alpha})\otimes-)$. One can show this map is homotopic equivalent to $F(\Theta_{\alpha^{\prime},\alpha}\otimes-):CFK_{\mathcal{R}}(\mathcal{H})\rightarrow CFK_{\mathcal{R}}(\mathcal{H}^{\prime})$ by a standard computation which shows $F(\Theta_{\alpha^{\prime},\alpha^{H}}\otimes\Theta_{\alpha^{H},\alpha})=\Theta_{\alpha^{\prime},\alpha}$ (see Figure 9 (b)). One can show that the map $F(\Theta_{\alpha^{\prime},\alpha}\otimes-)$ is a chain isomorphism (using the area-filtration technique in [OS04b], Proposition 9.8). ∎ ## 4\. Paring theorems In Section 4.1–4.2, we introduce a pairing construction which merges a (non- immersed) bordered Heegaard diagram and an immersed multicurve to produce an immersed Heegaard diagram. After that, we establish the unobstructedness and admissibility of these pairing diagrams in Section 4.3–4.5, and then we prove the bordered invariant of such pairing diagrams admits a box-tensor product interpretation in Section 4.6. Finally, in Section 4.7 we prove a pairing theorem for gluing a particular type of doubly-pointed bordered Heegaard diagram and an immersed bordered Heegaard diagram; this theorem will be useful in Section 5. ### 4.1. Immersed curves in the marked torus ###### Definition 4.1. The _marked torus_ $T^{2}$ is the oriented surface $\mathbb{R}^{2}/\mathbb{Z}^{2}$ together with a base point $z$ located at $(1-\epsilon,1-\epsilon)$ for some sufficiently small $\epsilon>0$. The images of the positively oriented $x$-axis and $y$-axis are called the _preferred longitude_ and _preferred meridian_ respectively. We will consider immersed multicurves with local systems in the marked torus. Two immersed multicurves are equivalent if they are regularly homotopic in $T^{2}\backslash{z}$ and the local systems are isomorphic. Throughout this paper, we restrict to immersed multicurves $\alpha_{im}$ satisfying the following assumptions: * (C-1) No component of $\alpha_{im}$ is a circle enclosing the base point $z$ once. * (C-2) No component of the immersed multicurve is null-homotopic in $T^{2}\backslash\\{z\\}$, and the immersed multicurve is _unobstructed_ in the sense that it does not bound any teardrops in $T^{2}\backslash\\{z\\}$. * (C-3) The immersed multicurve is _reduced_ , i.e., if we let $[0,1]\times[0,1]$ be the square obtained by cutting the marked torus open along the preferred meridian and longitude, then no sub-arcs of $\alpha_{im}$ contained in $[0,1]\times[0,1]$ have both ends on the same edge of the square. * (C-4) Let $\pi$ denote the projection map from $\mathbb{R}^{2}$ to $T^{2}$. Using regular homotopy, we assume all immersed curves in the marked torus are contained in the complement of $\pi([-\frac{1}{4},\frac{1}{4}]\times[-\frac{1}{4},\frac{1}{4}])$ in $T^{2}$, the strands contained in $\pi([-\frac{1}{4},\frac{1}{4}]\times[\frac{1}{4},\frac{3}{4}])$ are horizontal, and the strands contained in the image of $\pi([\frac{1}{4},\frac{3}{4}]\times[-\frac{1}{4},\frac{1}{4}])$ are vertical. An immersed multicurve in the marked torus determines a type D structure over the torus algebra as follows. First, we introduce some terminology. Figure 10. Six types of elementary arcs. The orientations are the so-called correct orientations. ###### Definition 4.2. An _elementary arc_ is an embedded arc in the marked torus ${T}^{2}$ such that it only intersects the preferred meridian or longitude at the endpoints. There are six types of elementary arc based on the position of the endpoints, each of which is labeled by a Reeb chord in $\\{\rho_{1},\rho_{2},\rho_{3},\rho_{12},\rho_{23},\rho_{123}\\}$ as shown in Figure 10. If we ignore the local systems, then any immersed multicurve is comprised of a collection of elementary arcs; one can see this by cutting $T^{2}$ open along the preferred longitude and meridian. Sometimes we also need to consider oriented elementary arcs. ###### Definition 4.3. An orientation of an elementary arc is called the correct orientation if it is the one shown in Figure 10. Next, we describe how to obtain a type D structure from an immersed multicurve in terms of elementary arcs. Denote the local system on $\alpha_{im}$ by $(\mathcal{E},\Phi)$, where $\mathcal{E}$ is a vector bundle over $\alpha_{im}$ and $\Phi$ is a parallel transport. Let $\mathcal{G}(\alpha_{im})=\mathcal{G}_{m}\cup\mathcal{G}_{l}$, where $\mathcal{G}_{m}$ (respectively, $\mathcal{G}_{l}$) is the set of intersection points of $\alpha_{im}$ and the preferred meridian (respectively, longitude). Let $\mathcal{X}$ be the vector space $\oplus_{x\in\mathcal{G}(\alpha_{im})}\mathcal{E}|_{x}$. Next, we define an $\mathcal{I}$-action on $\mathcal{X}$, where $\mathcal{I}$ is the ring of idempotent of the torus algebra. If $x\in\mathcal{G}_{m}$, for any $\tilde{x}\in\mathcal{E}|x$, $\iota_{0}\cdot\tilde{x}=\tilde{x}$ and $\iota_{1}\cdot\tilde{x}=0$; if $x\in\mathcal{G}_{l}$, for any $\tilde{x}\in\mathcal{E}|x$, $\iota_{0}\cdot\tilde{x}=0$ and $\iota_{1}\cdot\tilde{x}=\tilde{x}$. The underlying $\mathcal{A}$-module for $\widehat{CFD}(\alpha_{im})$ is $\mathcal{A}\otimes_{\mathcal{I}}\mathcal{X}$. Finally, the differential on $\widehat{CFD}(\alpha_{im})$ decomposes linearly as maps between $\mathcal{E}|_{x}$ for $x\in\mathcal{G}(\alpha_{im})$. Given $x,y\in\mathcal{G}(\alpha_{im})$ and $\rho_{I}$ a Reeb element, there is a differential map $\mathcal{E}|_{x}\rightarrow\rho_{I}\otimes\mathcal{E}|_{y}$ if and only if $x$ and $y$ are connected by a $\rho_{I}$-elementary arc whose correct orientation goes from $x$ to $y$, in which case the differential is given by $\partial(\tilde{x})=\rho_{I}\otimes\Phi(\tilde{x})$ for $\tilde{x}\in\mathcal{E}|_{x}$. In particular, when the local system of $\alpha_{im}$ is trivial, then the generators of $\widehat{CFD}(\alpha_{im})$ are in one-to-one correspondence with the intersection points of $\alpha_{im}$ with the preferred longitude/meridian, and the differentials are in one-to-one correspondence with the elementary sub-arcs of $\alpha_{im}$. The immersed-curve presentation of type D structures is empowered by the following result. ###### Theorem 4.4 ([HRW23]). Each Type D structure of a bordered 3-manifold with torus boundary is homotopic to a type D structure determined by some immersed multicurve (with local systems) in the marked torus. ###### Remark 4.5. All immersed multicurves arising from 3-manifolds with torus boundary satisfies the assumptions (C-1)-(C-4): (C-4) is straightforward, (C-2) and (C-3) follows from the algorithm of converting type D structures to immersed multicurves in [HRW23], and for (C-4) see the discussion around Figure 31 and 32 in [HRW22]. We will mainly be interested in the immersed multicurves corresponding to type D structures of knot complements for knots in the 3-sphere; these immersed multicurves satisfy some further properties that we specify in Definition 4.6 below, and the proofs of these properties can be found in [HRW22, Section 4]. ###### Definition 4.6. An immersed multicurve $\alpha_{im}=\\{\alpha_{im}^{0},\ldots,\alpha_{im}^{n-1}\\}$ of $n$ components (for some $n\geq 1$) with a local system is called knot-like if the local system restricted to ${\alpha_{im}^{0}}$ is trivial, $\alpha_{im}^{0}$ (with some orientation) is homologous to the preferred longitude in $T^{2}$, and $[\alpha_{im}^{i}]$ for $i\geq 1$ is trivial in $H_{1}(T^{2},\mathbb{Z})$. From now on, we assume all immersed multicurves are knot-like. ### 4.2. Pairing diagrams We introduce a class of immersed bordered Heegaard diagrams and doubly pointed Heegaard diagrams. They are respectively obtained from two types of pairing constructions that we will define: * (1) Pairing an immersed multicurve in the marked torus and an _arced bordered Heegaard diagram with two boundary components_ to construct an immersed bordered Heegaard diagram. * (2) Paring an immersed multicurve in the marked torus with a doubly pointed bordered Heegaard diagram to construct a closed immersed doubly pointed Heegaard diagram. We begin with the first type. For convenience, we first recall the definition of arced bordered Heegaard diagrams below (in the special case where both boundaries of the corresponding bordered manifold are tori). ###### Definition 4.7. An arced bordered Heegaard diagram with two boundary components is a quadruple $\mathcal{H}^{a}=(\bar{\Sigma},\bar{\bm{\alpha}},\bm{\beta},\bm{z})$ where * (1) $\bar{\Sigma}$ is a compact, oriented surface of genus $g$ with two boundary components $\partial\bar{\Sigma}=\partial_{L}\bar{\Sigma}\cup\partial_{R}\bar{\Sigma}$; * (2) $\bar{\bm{\alpha}}$ is a collection of pairwise disjoint properly embedded arcs and curves $\\{\alpha^{a,L}_{1},\alpha^{a,L}_{2},\alpha^{a,R}_{1},\alpha^{a,R}_{2},\alpha^{c}_{1},\ldots,\alpha^{c}_{g-2}\\}$. Here, $\alpha^{a,L}_{1}$ and $\alpha^{a,L}_{2}$ are two arcs with endpoints on $\partial_{L}\bar{\Sigma}$, $\alpha^{a,R}_{1}$ and $\alpha^{a,R}_{2}$ are two arcs with endpoints on $\partial_{R}\bar{\Sigma}$, and the $\alpha^{c}_{i}$’s ($i=1,\ldots,g-2$) are embedded circles. Moreover, elements in $\bar{\bm{\alpha}}$ induce linearly independent elements in $H_{1}(\bar{\Sigma},\partial\bar{\Sigma};\mathbb{Z})$; * (3) $\bm{\beta}$ is a set of $g$ pairwise disjoint embedded circles $\\{\beta_{1},\ldots,\beta_{g}\\}$ in the interior of $\bar{\Sigma}$ that are linearly independent as elements in $H_{1}(\bar{\Sigma},\partial\bar{\Sigma};\mathbb{Z})$; * (4) $\bm{z}$ is a properly embedded arc in $\bar{\Sigma}\backslash(\bar{\bm{\alpha}}\cup\bm{\beta})$ with one endpoint $z_{L}$ on $\partial_{L}\bar{\Sigma}$ and the other endpoint $z_{R}$ on $\partial_{R}\bar{\Sigma}$. Periodic and provincially period domains for arced bordered Heegaard diagrams with two boundary components are defined similarly to the case of a single boundary component. In the two boundary case we will also consider periodic domains that are adjacent to only one of the boundaries. ###### Definition 4.8. A domian is _left provincial_ if the multiplicity in the regions adjacent to $\partial_{L}\bar{\Sigma}$ are zero. We say an arced bordered Heegaard diagrams with two boundary components is _left provincially admissible_ if all left provincial periodic domains have both positive and negative multiplicities. The pairing construction is illustrated in Figure 11, and is spelled out in Definition 4.9. Figure 11. Left: an arced bordered Heegaard diagram. Middle: an immersed multicurve in the marked torus. The dashed lines are the boundary of $\pi([-\frac{1}{4},\frac{1}{4}]\times[-\frac{1}{4},\frac{1}{4}])$. Right: a bordered Heegaard diagram obtained by the pairing construction. ###### Definition 4.9. Let $\mathcal{H}^{a}=(\bar{\Sigma},\bar{\bm{\alpha}},\bm{\beta},\bm{z})$ be an arced bordered Heegaard diagram with two boundary components and let $\alpha_{im}$ be an immersed multicurve in the marked torus $T^{2}$. The _pairing diagram of $\mathcal{H}^{a}$ and $\alpha_{im}$_, denoted by $\mathcal{H}^{a}(\alpha_{im})$, is a bordered Heegaard diagram obtained through the following steps. * (1) Form $\bar{\Sigma}^{\prime}$ from $\bar{\Sigma}$ by collapsing $\partial_{R}\bar{\Sigma}$. Let $\alpha^{\prime a}_{i}$ be the image of $\alpha^{a,L}_{i}$ ($i=1,2$), $\alpha^{\prime c}_{i}$ be the image of $\alpha^{c}_{i}$ ($i=1,\ldots,g-2$), $\bm{\beta}^{\prime}$ be the image of $\bm{\beta}$, and $z^{\prime}_{L}$ be the image of $z_{L}$. The images of $\alpha^{a,R}_{i}$ ($i=1,2$), denoted by $\tilde{\alpha}_{i}$, are two circles intersecting at a single point ${z}^{\prime}_{R}$, the image of $z_{R}$. * (2) Take a neighborhood $U$ of $\tilde{\alpha}_{1}\cup\tilde{\alpha}_{2}$ which admits a homeomorphism $h:U\rightarrow T^{2}\backslash\pi([-\frac{1}{4},\frac{1}{4}]\times[-\frac{1}{4},\frac{1}{4}])$ such that $h(\tilde{\alpha}_{1})=\pi(\\{\frac{1}{2}\\}\times[0,1])$, $h(\tilde{\alpha}_{2})=\pi([0,1]\times\\{\frac{1}{2}\\})$, and each connected component of $h(\bm{\beta}^{\prime}\cap U)$ is an arc of the form $\pi(\\{x\\}\times[\frac{1}{4},\frac{3}{4}])$ or $\pi([\frac{1}{4},\frac{3}{4}]\times\\{y\\})$ for some $x$ or $y$ in $(2\epsilon,\frac{1}{4})$. * (3) Let $\alpha_{im}^{\prime}=h^{-1}(\alpha_{im})$. Let $\bar{\bm{\alpha}}^{\prime}=\\{\alpha^{\prime a}_{1},\alpha^{\prime a}_{2},\alpha^{\prime c}_{1},\ldots,\alpha^{\prime c}_{g-1},\alpha_{im}^{\prime}\\}$. * (4) Let $\mathcal{H}^{a}(\alpha_{im})=(\bar{\Sigma}^{\prime},\bar{\bm{\alpha}}^{\prime},\bm{\beta}^{\prime},z^{\prime}_{L})$. Recall a _doubly pointed bordered Heegaard diagram_ is a bordered Heegaard diagram with an extra basepoint in the complement of the $\alpha$\- and $\beta$-curves. It encodes a knot in a bordered 3-manifold. There is an entirely similar pairing construction for a doubly-pointed bordered Heegaard diagram and an immersed multicurve in the marked torus. We do not spell out the wordy definition and simply refer the readers to Figure 12 for an example. Figure 12. Pairing construction that gives rise to a doubly pointed Heegaard diagram. We want to establish the unobstructedness and admissibility of the immersed Heegaard diagrams obtained from pairing constructions. For that we need two tools, namely _z-adjacency_ and _the collapsing map_ introduced in the next two subsections. ### 4.3. z-adjacency We will consider a diagrammatic condition for immersed multicurves that guarantees the unobstructedness of the paring diagram; this condition can be achieved easily by finger moves. We begin by introducing some terminology for convenience. In the definition below, we orient the curves in $\alpha_{im}$ arbitrarily and orient the four edges of the cut-open torus using the boundary orientation. For each edge of the cut-pen torus, let $k_{+}$ and $k_{-}$ denote the number of elementary arcs intersecting a given edge positively and negatively, respectively. Figure 13. The disks $U^{R}_{-}$ and $U^{L}_{-}$. The superscript is chosen to suggest whether $z$ is one the left or on the right of the strands when we traverse an arc in the indicated direction. ###### Definition 4.10. Let $\alpha_{im}$ be an immersed multicurve in the marked torus. Then $\alpha_{im}$ is _$z$ -adjacent_ if, for each of the four edges of the cut- open torus, there exist four open disks $U_{\pm}^{R}$ and $U_{\pm}^{L}$ in $T^{2}$ such that * (1) $(U_{-}^{L},U_{-}^{L}\cap(\alpha_{im}\cup\\{{z}\\})$, $(U_{-}^{R},U_{-}^{R}\cap(\alpha_{im}\cup\\{{z}\\})$, $(U_{+}^{L},U_{+}^{L}\cap(\alpha_{im}\cup\\{{z}\\})$ and $(U_{+}^{R},U_{+}^{R}\cap(\alpha_{im}\cup\\{{z}\\})$ are homeomorphic to the corresponding disks in Figure 13, where the arcs in the disks are sub-arcs on the $k_{-}$ distinct elementary arcs intersecting the given edge negatively for discs with subscript $-$ or sub-arcs on the $k_{+}$ distinct elementary arcs intersecting the given edge positively for discs with subscript $+$; * (2) if the given edge is the top edge, then $U^{L}_{-}$ and $U^{R}_{+}$ are contained in $[0,1]\times[0,1]$; * (3) if the given edge is the right edge, then $U^{R}_{-}$ and $U^{L}_{+}$ are contained in $[0,1]\times[0,1]$. ###### Proposition 4.11. Every immersed multicurve in the marked torus is regularly homotopic to a $z$-adjacent multicurve. ###### Proof. Orient $\alpha_{im}$ arbitrarily. We first define an operation on a collection of oriented parallel arcs. Assume there are $k_{+}+k_{-}$ arcs, where $k_{+}$-many of the arcs are oriented in one direction, and the rest are oriented in the opposite direction. Figure 14. Finger moves on parallel strands. The operation is shown in Figure 14: First, by performing the finger moves in Figure 14 (a) repeatedly, we can arrive at a collection of arcs as shown in the left of Figure 14 (b): the $P$\- and $P^{-1}$-boxes indicate a pair of mutually inverse permutations, and between the $P$\- and $P^{-1}$-boxes the arcs are arranged so that all $k_{-}$ arcs with parallel orientations are grouped on the left and all the other $k_{+}$ arcs with the opposite orientations are grouped on the right. Next, do a sequence of finger moves to the diagram on the left of Figure 14 (b) to arrive at the right-hand-side diagram of Figure 14 (b). Now perform this operation to the arcs of $\alpha_{im}$ near all four edges in the cut-open marked torus, then we have a z-adjacent immersed multicurve; see Figure 15 for the desired open disks. Note that conditions $(2)$ and $(3)$ are obviously satisfied because $z$ is in the top right corner of the cut open torus. ∎ Figure 15. A $z$-adjacent immersed curve. We shall need a technical lemma. Let $l$ be a one-cornered sub-loop of $\alpha_{im}$ with a corner $q$. If we traverse $l$ in either direction, we see it begins with an arc starting from $q$ to the meridian or longitude, then a sequence of elementary arcs, and finally, an arc starting from the meridian or longitude and ending at $q$. We call the starting and ending arcs the _non- elementary sub-arcs of $l$_, and the other sub-arcs _the elementary sub-arcs of $l$_. ###### Lemma 4.12. Let $\alpha_{im}$ be a $z$-adjacent immersed curve. Let $D$ be a positive domain in ${T}^{2}$ bounded by a $k$-cornered (sub)loop of $\alpha_{im}$. * (1) If $n_{z}(D)=n$ for some $n\geq 0$ and $k=0$ or $1$, then for any side of the cut-open marked torus $[0,1]\times[0,1]$ and any sign, the number of elementary sub-arcs in $\partial D$ intersecting the given side with the given sign is less than or equal to $n$. * (2) If $n_{z}(D)=0$, then for arbitrary $k\geq 0$, there are no elementary subarcs contained in $\partial D$. ###### Proof. We prove (1) first. We will only consider the case in which the elementary sub-arcs intersect the given edge negatively and remark that the other case is similar. We prove by contradiction. Suppose there are $k_{-}>n$ elementary sub-arcs contained in $\partial D$ intersecting the given edge negatively. Since $\partial D$ is $0$\- or $1$-cornered it has an orientation induced by the orientation on $\alpha_{im}$. Examining the local diagram $(U_{-}^{L},U_{-}^{L}\cap({\partial D}\cup\\{\bm{z}\\}))$ in Figure 13 one sees $D$ has negative multiplicity $n-k_{-}$ in the left-most region, which contradicts our assumption that $D$ is a positive domain. Therefore, $k_{-}\leq n$. Next, we prove (2). Assume there is an elementary sub-arc in $\partial D$. Then no matter how this sub-arc is oriented, $z$ is on both the left and right of it. As $n_{z}(D)=0$, there is a region with $-1$ multiplicity, which contradicts that $D$ is positive. ∎ ### 4.4. The collapsing operation To relate the domains of the pairing diagram $\mathcal{H}^{a}(\alpha_{im})$ and the arced bordered diagram $\mathcal{H}^{a}$, we define the so-called _collapsing operation_. This operation was previously defined in the case of paring genus one bordered Heegaard diagrams with immersed curves [Che23], and we give the general case here. The operation is pictorially shown in Figure 16, and the definition is given below. ###### Definition 4.13. The collapsing operation on $\mathcal{H}^{a}(\alpha_{im})$ is defined to be the composition of the following modifications of the diagram: * (Step 1) Extend the map $h$ in Definition 4.9 to identify $T^{2}-\pi([-\frac{1}{4}+\epsilon,\frac{1}{4}-\epsilon]\times[-\frac{1}{4}+\epsilon,\frac{1}{4}-\epsilon])$ with a slightly larger neighborhood of $U=h^{-1}(T^{2}-\pi([-\frac{1}{4},\frac{1}{4}]\times[-\frac{1}{4},\frac{1}{4}]))$. Here $\epsilon$ is a sufficiently small positive number. * (Step 2) Puncture $h^{-1}((\frac{3}{4},\frac{3}{4}))$, and enlarge it to a hole so that under the identification map $h$, the boundary of the hole is a square of side length $\frac{1}{2}+2\epsilon$ and with rounded corner modeled on a quarter of a circle of radius $\epsilon$. While enlarging the hole, we push immersed curves it encountered along the way so that part of the immersed curves are squeezed to the boundary of the hole. * (Step 3) Collapse $h^{-1}(\pi([-\frac{1}{4}+\epsilon,\frac{1}{4}-\epsilon]\times[\frac{1}{4},\frac{3}{4}]))$ to the core $h^{-1}(\pi([-\frac{1}{4}+\epsilon,\frac{1}{4}-\epsilon]\times\\{\frac{1}{2}\\}))$, which is denoted $a^{a,R}_{1}$. Collapse $h^{-1}(\pi([\frac{1}{4},\frac{3}{4}]\times[-\frac{1}{4}+\epsilon,\frac{1}{4}-\epsilon]))$ to the core $h^{-1}(\pi(\\{\frac{1}{2}\\}\times[-\frac{1}{4}+\epsilon,\frac{1}{4}-\epsilon]))$, which is denoted $a^{a,R}_{2}$. ###### Remark 4.14. * (1) Clearly, the outcome of the collapsing operation on $\mathcal{H}^{a}(\alpha_{im})$ can be identified with $\mathcal{H}^{a}$. * (2) Each elementary arc in $\alpha_{im}$ standing for $\rho_{I}\in\\{\rho_{1},\rho_{2},\rho_{3},\rho_{12},\rho_{23},\rho_{123}\\}$ is mapped under the collapsing map to an arc that passes the Reeb chord $\rho_{I}$ in $\mathcal{Z}^{R}$ of $\mathcal{H}^{a}$. Note that an oriented elementary sub-arc is correctly oriented if it induces a Reeb chord $\mathcal{Z}^{R}$ under the collapsing map, i.e., the orientations coincide. * (3) The intersection points in $\mathcal{G}({\mathcal{H}^{a}(\alpha_{im})})$ are of the form $\bm{x}\otimes a$ are in one-to-one correspondence with $\mathcal{G}({\mathcal{H}^{a}})\otimes_{\mathcal{I}_{R}}\mathcal{G}({\alpha_{im}})$, where the tensor product is taken over $\mathcal{I}_{R}\subset\mathcal{A}(\mathcal{Z}_{R})$. Indeed, given an intersection point $\xi\in\mathcal{G}({\mathcal{H}^{a}(\alpha_{im})})$, its image under the collapsing map yields an intersection point $\bm{x}$ in $\mathcal{H}^{a}$. Also, the component of $\xi$ on $\alpha_{im}$ uniquely gives rise to an intersection point $a$ of $\alpha_{im}$ as follows. By the definition of the pairing operation, when we pull back the intersection point on $\alpha_{im}$ to the marked torus, it lies in a horizontal or vertical arc as described in assumption (C-4) on immersed multicurves, which uniquely corresponds to an intersection point of $\alpha_{im}$ with the longitude or meridian. Therefore, every intersection point $\xi$ in $\mathcal{H}^{a}(\alpha_{im})$ can be written as $\bm{x}\otimes y$. It is easy to see this induces a one-to-one correspondence between $\mathcal{G}({\mathcal{H}^{a}(\alpha_{im})})$ and $\mathcal{G}({\mathcal{H}^{a}})\otimes_{\mathcal{I}_{R}}\mathcal{G}({\alpha_{im}})$. Figure 16. The collapsing operation. We will give a proposition relating the domains of $\mathcal{H}^{a}(\alpha_{im})$ and $\mathcal{H}^{a}$. Let $l$ be an oriented arc $l$ on $\alpha_{im}$ such that all the elementary sub-arcs are oriented correctly. We use $\overrightarrow{\rho}(l)$ to denote the sequence of Reeb chords determined by $l$. ###### Proposition 4.15. Assume the immersed multicurve $\alpha_{im}$ is $z$-adjacent. Let $B$ be a positive domain in $\mathcal{H}^{a}(\alpha_{im})$ corresponding to a homology class in $\pi_{2}(\bm{x}\otimes a,\bm{y}\otimes b,\overrightarrow{\sigma})$ with $n_{z}(B)=0$. Then the image of $B$ under the collapsing map is a positive domain $B^{\prime}$ in $\mathcal{H}^{a}$ corresponding to a homology class $\pi_{2}(\bm{x},\bm{y},\overrightarrow{\rho}(\partial_{\alpha_{im}}B),\overrightarrow{\sigma})$ with $n_{z}(B^{\prime})=0$. Here, $\partial_{\alpha_{im}}B$ refers to the arc on $\alpha_{im}$ connecting the corresponding components of $\bm{x}\otimes a$ and $\bm{y}\otimes b$. Moreover, $e(B^{\prime})=e(B)-\frac{|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|}{2}.$ ###### Proof. It is clear that $B^{\prime}$ is positive and $n_{z}(B^{\prime})=0$. It is also clear that $B^{\prime}$ give rise to a domain connecting $\bm{x}$ and $\bm{y}$. We need to show that $B^{\prime}$ has the Reeb chords $\overrightarrow{\rho}(\partial_{\alpha_{im}}B)$ at the east infinity. We claim all the elementary arcs appear in $\partial_{\alpha_{im}}B$ are correctly oriented, and hence $\partial_{\alpha_{im}}B$ gives rise to a monotonic arc (in the sense that all the Reeb chords appearing on the arc respect the boundary orientation) connecting (the components on the $\alpha$ arc of) $\bm{x}$ to $\bm{y}$ under the collapsing map. The sequence of Reeb chords appearing in this arc are exactly $\overrightarrow{\rho}(\partial_{\alpha_{im}}B)$ in view of Remark 4.14 (2). To see the claim, note $\alpha_{im}$ is $z$-adjacent and $B$ is positive with $n_{z}(B)=0$. Therefore, if an elementary arc on $\partial_{im}B$ intersects the top edge or the right edge, then its orientation is forced by the positivity of domains and condition (2) and (3) in Definition 4.10, and the orientation is the correct orientation. The only type of elementary arcs that intersects neither the top edge nor the right edge corresponds to $\rho_{2}$. If an elementary arc corresponding to $\rho_{2}$ on $\partial_{\alpha_{im}}B$ has a successor or precursor, then correct orientation on the successor or the precursor would induce the correct orientation on it. Otherwise, $\partial_{\alpha_{im}}B$ has only one elementary arc corresponding to $\rho_{2}$, in which case it is clear that the elementary arc is correctly oriented. Next, we compare the Euler measures. Divide the domain $B$ into two parts $B_{1}$ and $B_{2}$, along the square with rounded corners, which is the boundary of the hole in Step 2 of the collapsing operation. (See Figure 17.) This time, we do not puncture the interior of the square. Let $B_{1}$ denote the part of $B$ outside of the square, and let $B_{2}$ denote the part inside the square (which is pushed onto the boundary circle under the collapsing map). Then $e(B_{1})=e(B^{\prime})$ since these two domains differ by a bunch of rectangles whose Euler measure are zero; these rectangles are collapsed in Step 3 of the collapsing operation. As $\alpha_{im}$ is $z$-adjacent, $B_{2}$ is positive, and $n_{z}(B)=0$, we see $B_{2}$ can be further expressed as a sum of of simple domains determined by the elementary arcs appearing in $\partial_{\alpha_{im}}B$ (counted with multiplicity). (See Figure 18.) Each simple domain of multiplicity one has Euler measure $\frac{1}{2}$, and there are $|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|$ many of them being collapsed (in Step 2 of the collapsing operation) in order to obtain $B^{\prime}$. Therefore, $e(B^{\prime})=e(B)-\frac{|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|}{2}.$ ∎ Figure 17. Left: $B=B_{1}+B_{2}$. Right: $B^{\prime}$. Figure 18. Simple domains corresponding to Reeb elements in $\mathcal{A}$. ### 4.5. Unobstructedness and admissibility of paring diagrams ###### Proposition 4.16. Let $\alpha_{im}\subset T^{2}$ be a $z$-adjacent immersed multicurve. Then the pairing diagram $\mathcal{H}^{a}(\alpha_{im})$ of an arced bordered Heegaard diagram $\mathcal{H}^{a}$ and $\alpha_{im}$ is unobstructed. Furthermore, $\mathcal{H}^{a}(\alpha_{im})$ is provincially admissible provided $\mathcal{H}^{a}$ is left provincially admissible. (See Definition 2.8 and Definition 2.9.) ###### Proof of Proposition 4.16. Consider the bordered Heegaard diagram $\mathcal{H}^{a}(\alpha_{im})=(\bar{\Sigma}^{\prime},\bar{\bm{\alpha}}^{\prime},\bm{\beta}^{\prime},z^{\prime})$ obtained from pairing an arced bordered Heegaard diagram $\mathcal{H}^{a}=(\bar{\Sigma},\bar{\bm{\alpha}},\bm{\beta},\bm{z})$ and a $z$-adjacent immersed multicurve $\alpha_{im}$. We begin by showing $\bar{\bm{\alpha}}^{\prime}$ is unobstructed in the sense of Definition 2.8. Let $B$ be a zero- or one-cornered $\alpha$-bounded domain. Since the curves $\\{\bar{\alpha_{1}^{a}},\bar{\alpha_{2}^{a}},\alpha_{1},\ldots,\alpha_{g-2},\alpha_{im}^{0}\\}$ are pairwise disjoint and homologically independent in $H_{1}(\bar{\Sigma},\partial)$, $[\partial B]$ (as a homology class) is equal to a linear combination of at most one copy of $\partial\bar{\Sigma}$ and some homologically trivial zero- or one-cornered loop contained in a single connected component of $\alpha_{im}$. We first show there are no positive zero- or one-cornered $\alpha$-bounded domains $B$ with $n_{z^{\prime}}(B)=0$. In this case, $\partial B$ is a homologically trivial zero- or one-cornered loop contained in a single connected component of $\alpha_{im}$, i.e., $\partial\bar{\Sigma}$ does not appear in $\partial B$. As the $\bm{\beta}$-curves are irrelevant to our consideration, we may assume the $\alpha$-curves of $\mathcal{H}^{a}$ are in standard position. Therefore, there is an obvious circle $C\subset\bar{\Sigma}$ that splits $\bar{\Sigma}$ into a genus-$(g-1)$ surface $E_{1}$ containing $\\{\alpha^{a,L}_{1},\alpha^{a,L}_{2},\alpha^{c}_{1},\ldots,\alpha^{c}_{g-2}\\}$ and a genus-one surface $E_{2}$ containing $\alpha^{a,R}_{1}$ and $\alpha^{a,R}_{2}$. Let $C^{\prime}$ be the corresponding curve on $\bar{\Sigma}^{\prime}$. Then after surgery along $C^{\prime}$, $B$ induces a positive domain $D$ in the marked torus $T^{2}$ (obtained from $E_{2}$ in an obvious way), and $D$ is bounded by a zero- or one-cornered (sub)loop of $\alpha_{im}$. According to Lemma 4.12, $\partial D$ contains no elementary sub-arcs, so $D$ cannot exist. Next we show that if $n_{z^{\prime}}(B)=1$, $B$ is a stabilized teardrop or $[\Sigma^{\prime}]$ depending on whether $\partial B$ is one-cornered or zero- cornered. In this case, after performing surgery along the same $C^{\prime}$ as in the previous paragraph, $B$ gives rise to two domains: one is $[E_{1}^{\prime}]$, where $E^{\prime}_{1}$ is the genus-$(g-1)$ surface, and the other is a positive domain $D$ contained in the marked torus $T^{2}$ with $n_{z}(D)=1$. We first consider the case in which $\partial D$ is zero- cornered. If $\partial D=\emptyset$, then $D=E_{2}$ and hence $B=[\Sigma^{\prime}]$. If $\partial D\neq\emptyset$, then according to Lemma 4.12, it consists of at most (and at least) $4$ elementary sub-arcs, and hence is a circle enclosing the $z$-basepoint once. However, such circles are assumed not to exist. When $\partial D$ is one-cornered, we claim $D$ is a teardrop. To see this, note that Lemma 4.12 implies that $\partial D$ crosses the meridian at most three times and the longitude at most three times since each time the meridian or the longitude is crossed (except possibly the last time) the intersection is the beginning of an elementary sub-arc and there are at most two elementary sub-arcs starting on each. Because $\partial D$ is homologically trivial in $H_{1}({T}^{2})$ it crosses both of the meridian and the longitude and even number of times, so it crosses each at most twice. It follows that $\partial D$ must circle once around $z$ and $D$ is a teardrop with $n_{z}(D)=1$. Now we show any two-cornered positive $\alpha$-bounded domain $B$ with $n_{z}(B)=0$ is a bigon. To see, we may split $\Sigma^{\prime}$ as $E_{1}\\#E_{2}$ as before and regard $B$ as a domain in $E_{2}$ with $n_{z^{\prime}}=0$. Note by Lemma 4.12 (2), we know $\partial B$ consists of no elementary subarcs, and hence $B$ must be of the form shown in Figure 19 (up to rotation), which is a bigon. (Note we do not require the corners of the bigon $B$ to be convex.) Figure 19. Two-cornered positive $\alpha$-bounded domains. So far, we have proved $\bar{\bm{\alpha}}^{\prime}$ is unobstructed. We now show there are no non-trivial positive provincial periodic domains. If not, let $B$ be a positive provincial periodic domain for $\mathcal{H}^{a}(\alpha_{im})$. Then by Proposition 4.15, $\Psi(B)$ is a positive periodic domain for $\mathcal{H}^{a}$, where $\Psi$ denotes the collapsing map. Note $\Psi(B)$ is left provincial. As $\mathcal{H}^{a}$ is left provincially admissible, we have $\Psi(B)=0$, and hence $\partial B$ has no $\beta$-curves. So, $B$ is a positive zero-cornered $\alpha$-bounded domain with $n_{z}(B)=0$, but such domains are already excluded by unobstructedness. ∎ ### 4.6. The first paring theorem Recall a bordered Heegaard diagram is _nice_ if every connected region in the complement of the $\alpha$\- and $\beta$-curves is a disk with at most four corners except for the region containing $z$. Any bordered Heegaard diagram can be turned into a nice diagram via isotopy and handleslides of the $\beta$-curves (Proposition 8.2 of [LOT18]). The key property of nice Heegaard diagrams is that the Euler measure of any region away from the base point is non-negative. This property imposes great constraints on domains supporting holomorphic representatives via the index formula. Hence it opens up the combinatorial vein for proving the pairing theorem. See 1.4 ###### Proof. In view of the homotopy equivalence of the relevant invariants under isotopy of $\beta$ curves (Proposition 2.82), we may assume $\mathcal{H}^{a}$ is a nice arced bordered Heegaard diagram. Note nice arced bordered Heegaard diagrams are automatically left provincially admissible. Therefore, $\widehat{CFDA}(\mathcal{H}^{a})$ and $\widehat{CFD}(\mathcal{H}^{a}({\alpha_{im}}))$ are defined. In fact, a stronger admissibility condition holds for $\mathcal{H}^{a}$: any periodic domains with $n_{z}=0$ has both positive and negative local multiplicities. This implies $\widehat{CFDA}(\mathcal{H}^{a})$ is bounded, and hence the box- tensor product is expressed as a finite sum. Implicit in the proof is that we will be using split almost complex structures for defining $\widehat{CFD}(\mathcal{H}^{a}({\alpha_{im}}))$ and $\widehat{CFDA}(\mathcal{H}^{a})$. A split almost complex structure is sufficient for defining $\widehat{CFD}(\mathcal{H}^{a}({\alpha_{im}}))$, since all the domains involved will be bigons and rectangles. In this setting, up to a generic perturbation of the $\alpha$ and $\beta$ curves, moduli spaces defined using a split almost complex structure are transverse (c.f. [Lip06, Proposition 3.9]). We will call the two punctures in $\mathcal{H}^{a}$ the $\sigma$-puncture and the $\rho$-puncture, where the $\rho$-puncture is the one that gets capped off in the pairing diagram. For now, we assume the local systems on $\alpha_{im}$ are trivial, and we will indicate the modifications needed for dealing with nontrivial local system later on. First, the generators $\mathcal{G}(\mathcal{H}^{a}(\alpha_{im}))$ and $\mathcal{G}(\mathcal{H}^{a})\otimes_{\mathcal{I}_{R}}\mathcal{G}(\alpha_{im})$ are identified as pointed out in Remark 4.14 (3). Next, we prove the differentials have a one-to-one correspondence. We first show any differential incurred by the box-tensor product has a corresponding differential in $\widehat{CFD}(\mathcal{H}^{a}(\alpha_{im}))$. A differential arising from the box tensor product comes in two types, depending on whether it involves nontrivial differentials in $\widehat{CFD}(\alpha_{im})$. If it does not involve non-trivial differential in $\widehat{CFD}(\alpha_{im})$, then the input from $\widehat{CFDA}(\mathcal{H}^{a})$ counts curves with the domain being a provincial bigon, a provincial rectangle, or a bigon with a single Reeb chord on the $\sigma$-puncture; see [LOT18, Proposition 8.4]. Such bigons or rectangles clearly have their counterparts in $\mathcal{H}^{a}(\alpha_{im})$, giving the corresponding differentials in $\widehat{CFD}(\mathcal{H}^{a}(\alpha_{im}))$. If the box-tensor differential involves differentials in $\widehat{CFD}(\alpha_{im})$, then the corresponding input from $\widehat{CFDA}(\mathcal{H}^{a})$ counts curves with the domain being a bigon with a single Reeb chord on the $\rho$-puncture [LOT18, Proposition 8.4]. As it pairs with a differential in $\widehat{CFD}(\alpha_{im})$, this bigon gives rise to a bigon in $\mathcal{H}^{a}(\alpha_{im})$ (which is a pre-image of the collapsing map), giving the corresponding differential in $\widehat{CFD}(\mathcal{H}^{a}(\alpha_{im}))$. Next, we show that every differential in $\widehat{CFD}(\mathcal{H}^{a}(\alpha_{im}))$ corresponds to a differential incurred by the box-tensor product. Suppose $u\in\pi_{2}(\bm{x}\otimes a,\bm{y}\otimes b)$ admits a holomorphic representative contributing to a differential for $\widehat{CFD}(\mathcal{H}^{a}(\alpha_{im}))$. Let $B$ be the domain of $u$, and let $B^{\prime}$ denote the image of $B$ under the collapsing operation. By Proposition 4.15, $e(B)=e(B^{\prime})+\frac{|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|}{2}$. As $B^{\prime}$ is a positive domain with $n_{z}(B^{\prime})=0$ and $\mathcal{H}^{a}$ is a nice Heegaard diagram, we have $e(B)\geq e(B^{\prime})\geq 0$. By the index formula, denoting the source surface of $u$ by $S$, we have $\text{Ind}(u)=g-\chi(S)+2e(B)+|\overrightarrow{\sigma}|.$ As $\text{Ind}(u)=1$ and $2e(B)+|\overrightarrow{\sigma}|\geq 0$, we have $\chi(S)=g$ or $g-1$. When $\chi(S)=g$, $S$ consists of $g$ topological disks; each disk has a $+$ and a $-$ puncture, and there is at most one $\sigma$-puncture overall since $2e(B)+|\overrightarrow{\sigma}|=1$. We separate the discussion according to the number of $\sigma$-punctures. First, if there is a $\sigma$-puncture, then the corresponding domain $B$ in $\mathcal{H}^{a}(\alpha_{im})$ is a bigon with a single Reeb chord on the $\sigma$-puncture and does not involve $\alpha_{im}$. This domain clearly has its counterpart in $\mathcal{H}^{a}$ under the collapsing map, giving rise to an operation in $\widehat{CFDA}(\mathcal{H}^{a})$; the corresponding differential in the box- tensor product is obtained by pairing this $DA$-operation with an element in $\widehat{CFD}(\alpha_{im})$. Secondly, if there is no $\sigma$-puncture, then the domain $B$ is a provincial bigon in $\mathcal{H}^{a}(\alpha_{im})$. There are two sub-cases to consider depending on whether the $\alpha$-boundary of $B$ overlaps with $\alpha_{im}$. If the $\alpha$ boundary of $B$ is not on $\alpha_{im}$, then we argue as in the first case to see that $B$ gives a corresponding differential in the box-tensor product. If, on the other hand, the boundary of $B$ is on $\alpha_{im}$, since $1/2=e(B)=e(B^{\prime})+|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|/{2}$ we have $|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|$ is either $0$ or $1$. If $|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|=0$, then $B^{\prime}$ is a provincial domain, giving the type-DA operation for the corresponding differential obtained by the box-tensor product. If $|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|=1$, then $B^{\prime}$ is obtained from $B$ subtracting a simple region (as in the proof of Proposition 4.15) and then applying the collapsing map. We can see $B^{\prime}$ is a bigon with a single Reeb chord corresponding to the Reeb chord specified by $\partial_{\alpha_{im}}B$ on the $\rho$-puncture. The DA-operation given by $B^{\prime}$ and the type D operation given by $\partial_{\alpha_{im}}B$ pair up to give the corresponding differential in the box-tensor product. When $\chi(S)=g-1$, then $S$ consists of $g-1$ topological disks; $g-2$ of the disks are bigon, while the remaining one is a rectangle. As $e(B)=0$, the bigons are mapped trivially to $\Sigma$. Therefore, the domain $B$ is a rectangle. Again, since $e(B)=e(B^{\prime})+|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|/{2}$ and $e(B^{\prime})\geq 0$, we have $|\overrightarrow{\rho}(\partial_{\alpha_{im}}B)|=0$. Then $B^{\prime}$ is a provincial rectangular domain in $\mathcal{H}^{a}$, giving rise to a DA- operation that pairs with a trivial type-D operation to give the corresponding differential in the box-tensor product. We have finished the proof when the local system is trivial. Next, we consider the case where $\alpha_{im}$ admits a non-trivial local system $(\mathcal{E},\Phi)$. The local system induces a local system on the $\alpha$ curves in the pairing diagram $H^{a}(\alpha_{im})$. First, the discussion above identifies the generators at the vector space level: let $\bm{x}\otimes y$ be an intersection point in $\mathcal{G}(\mathcal{H}^{a}(\alpha_{im}))$, where $\bm{x}\in\mathcal{G}(H^{a})$ and $y\in\mathcal{G}(\alpha_{im})$; then $\bm{x}\otimes y$ corresponds to a direct summand $\mathcal{E}|_{\bm{x}\otimes y}$ of $\widehat{CFD}(\mathcal{H}^{a}(\alpha_{im}))$ as a vector space, and $\mathcal{E}|_{\bm{x}\otimes y}$ can be naturally identified with $\bm{x}\otimes\mathcal{E}|_{y}$, a summand of $\widehat{CFDA}(\mathcal{H}^{a})\boxtimes\widehat{CFD}(\alpha_{im})$. Secondly, the discussion in the trivial-local-system case shows that $\widehat{CFD}(\mathcal{H}^{a}(\alpha_{im}))$ has a differential map between the summands $\mathcal{E}|_{\bm{x}\otimes y}\rightarrow\sigma_{I}\otimes\mathcal{E}|_{\bm{x^{\prime}}\otimes y^{\prime}}$ if and only if the box-tensor product has a differential map between the corresponding summands $\bm{x}\otimes\mathcal{E}|_{y}\rightarrow\sigma_{I}\otimes(\bm{x^{\prime}}\otimes\mathcal{E}|_{y^{\prime}})$ in the box-tensor product, and under the natural identification between these summands both differential maps are induced by the same parallel transport from $\mathcal{E}|_{y}$ to $\mathcal{E}|_{y^{\prime}}$. ∎ ### 4.7. The second pairing theorem We are interested in computing knot Floer chain complexes over $\mathcal{R}=\mathbb{F}[U,V]/(UV)$ using bordered Floer homology. We have already defined an extended type-D structure, and we want to pair it with an extended type-A structure to get a bi-graded chain complex over $\mathcal{R}=\mathbb{F}[U,V]/(UV)$. Here we will only restrict to a specific extended type-A structure associated to the doubly-pointed bordered Heegaard diagram $\mathcal{H}_{id}$ given in Figure 20. The diagram $\mathcal{H}_{id}$ corresponds to the pattern knot given by the core of a solid torus, which is the identity pattern. Figure 20. The bordered diagram $\mathcal{H}_{id}$. Recall $\tilde{\mathcal{A}}$ denotes the weakly extended torus algebra, and $\mathcal{I}\subset\tilde{\mathcal{A}}$ is the ring of idempotents (see Figure 8). ###### Definition 4.17. The extended type-A structure $\widetilde{CFA}(\mathcal{H}_{id})$ is a free $\mathcal{R}$-module generated by the single intersection point $x$ in $\mathcal{H}_{id}$. It is equipped with an $\mathcal{I}$-action given by $x\cdot\iota_{0}=x$ and $x\cdot\iota_{1}=0$, together with a family of $\mathcal{R}$-linear maps $m_{i+1}:\widetilde{CFA}(\mathcal{H}_{id})\otimes_{\mathcal{I}}\tilde{A}^{\otimes i}\rightarrow\widetilde{CFA}(\mathcal{H}_{id})$ ($i\in\mathbb{N}$), where up to $\mathcal{R}$-linearity the only non-zero relations are: $\displaystyle m_{2}(x,1)=x,$ $\displaystyle m_{3+i}(x,\rho_{3},\overbrace{\rho_{23},\ldots,\rho_{23}}^{i},\rho_{2})=U^{i}x,\quad i\in\mathbb{N},$ $\displaystyle m_{3+i}(x,\rho_{1},\overbrace{\rho_{01},\ldots,\rho_{01}}^{i},\rho_{0})=V^{i}x,\quad i\in\mathbb{N}.$ ###### Remark 4.18. This extends the hat-version type A-structure $\widehat{CFA}(\mathcal{H}_{id})$ by allowing Reeb chords crossing the base point $z$. Straightforwardly, the hat-version box-tensor product can be generalized to be an operation between the extended type A structure $\widetilde{\mathcal{M}}:=\widetilde{CFA}(\mathcal{H}_{id})$ and a weakly extended type D structure $(\widetilde{\mathcal{N}},\delta^{i})$: It is the $\mathcal{R}$-module $\widetilde{\mathcal{M}}\otimes_{\mathcal{I}}\widetilde{\mathcal{N}}$ together with a differential $\partial_{\boxtimes}:=\sum_{i\geq 0}(m_{i+1}\otimes\mathbb{I}_{\widetilde{\mathcal{N}}})\circ(\mathbb{I}_{\widetilde{\mathcal{M}}}\otimes\delta^{i})$; the finiteness of the sum can be guaranteed for type D structures defined using bi-admissible diagrams (see the proof of Theorem 1.6 below). One may verify $\partial_{\boxtimes}^{2}=0$ algebraically using the structure equations defining the (weakly) extended type D and type A structures. We omit such computation and instead content ourselves with Theorem 1.6 below, which implies that the $\partial_{\boxtimes}$ induced by gluing bordered Heegaard diagrams is indeed a differential. We further remark that $\widetilde{\mathcal{M}}\boxtimes_{\mathcal{I}}\widetilde{\mathcal{N}_{1}}$ is chain homotopic to $\widetilde{\mathcal{M}}\boxtimes_{\mathcal{I}}\widetilde{\mathcal{N}_{2}}$ provided $\widetilde{\mathcal{N}_{1}}$ is homotopic to $\widetilde{\mathcal{N}_{2}}$. The proof of this is similar to that in the hat version and is omitted. See 1.6 (See Definition 2.8 and 2.10 for the unobstructedness and bi- admissibility for $\mathcal{H}$.) ###### Proof of Theorem 1.6. Note periodic domains of $\mathcal{H}_{id}\cup\mathcal{H}_{im}$ with $n_{z}=0$ (respectively $n_{w}=0$) corresponds to periodic domains of $\mathcal{H}_{im}$ with $n_{\rho_{0}}=n_{\rho_{1}}=0$ (respectively $n_{\rho_{2}}=n_{\rho_{3}}=0$). Therefore, since $\mathcal{H}_{im}$ is bi- admissible, $\mathcal{H}_{id}\cup\mathcal{H}_{im}$ is bi-admissible in the sense of Definition 3.3. Also, zero- or one-cornered $\alpha$-bounded domains in $\mathcal{H}_{id}\cup\mathcal{H}_{im}$ with $n_{z}=n_{w}=0$ must lie in $\mathcal{H}_{im}$. So, unobstructedness of $\mathcal{H}_{im}$ implies the unobstructedness of $\mathcal{H}_{id}\cup\mathcal{H}_{im}$. In summary, $\mathcal{H}_{id}\cup\mathcal{H}_{im}$ is bi-admissible and unobstructed, and hence $CFK_{\mathcal{R}}(\mathcal{H}_{id}\cup\mathcal{H}_{im})$ is defined. The bi-admissibility of $\mathcal{H}_{im}$ also implies $\partial_{\boxtimes}$ is expressed as a finite sum and hence is well-defined. To see this, note for any $\bm{x},\bm{y}\in\mathcal{G}(\mathcal{H}_{im})$, bi-admissibility implies there are only finitely many positive domains connecting $\bm{x}$ and $\bm{y}$ with a prescribed Reeb-chord sequence of the form $\rho_{1},\rho_{01},\ldots\rho_{01},\rho_{0}$ and $\rho_{3},\rho_{23},\ldots\rho_{23},\rho_{2}$. Recall that the differential in $CFK_{\mathcal{R}}(\mathcal{H}_{id}\cup\mathcal{H}_{im})$ counts holomorphic curves that can only cross at most one of the $w$\- and $z$-base points. Note also that both of the base points in $\mathcal{H}_{id}$ are adjacent to the east boundary. Therefore, by the symmetry of the base points $w$ and $z$, it suffices to prove the theorem for the hat version knot Floer homology, i.e., $\widehat{CFK}(\mathcal{H}_{id}\cup\mathcal{H}_{im})\cong\widehat{CFA}(\mathcal{H}_{id})\boxtimes\widehat{CFD}(\mathcal{H}_{im}).$ Though our Heegaard diagrams have immersed $\alpha$-multicurves, given that no boundary degeneration can occur, the proof for the embedded-$\alpha$-curves case, which uses neck stretching and time dilation, carries over without changes; see Chapter 9 of [LOT18] for detail or Section 3 of [LOT14a] for an exposition. ∎ ## 5\. Knot Floer homology of satellite knots We apply the machinery developed in the previous sections to study the knot Floer homology of satellite knots. First, we introduce a gentler condition on the immersed curves than the z-adjacency condition. ###### Definition 5.1. An immersed multicurve $\alpha_{im}$ in the marked torus $(T^{2},z)$ is admissible if there are no nontrivial zero- and one-cornered $\alpha$-bounded positive domains $B$ with $n_{z}(B)=0$. Note any $z$-adjacent immersed multicurve is admissible in view of Lemma 4.12. Let $\mathcal{H}_{w,z}$ be a doubly pointed bordered Heegaard diagram for a pattern knot $(S^{1}\times D^{2},P)$. Recall that we can construct a doubly pointed immersed diagram $\mathcal{H}_{w,z}(\alpha_{im})$. The admissibility condition guarantees that $CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{im}))$ is defined in view of the following proposition. ###### Proposition 5.2. If $\alpha_{im}$ is an admissible immersed multicurve, then $\mathcal{H}_{w,z}(\alpha_{im})$ is bi-admissible and unobstructed. ###### Proof. First, we show the diagram $\mathcal{H}_{w,z}(\alpha_{im})$ is unobstructed (in the sense of Definition 3.2). Let $B$ be a zero- or one-cornered $\alpha$-bounded domain for $\mathcal{H}_{w,z}(\alpha_{im})$. Note $n_{w}(B)=n_{z}(B)$ since the base points $w$ and $z$ are in the same region in the complement of the $\alpha$-curves. We restrict to the domains with $n_{z}(B)=n_{w}(B)=0$. Recall we want to prove $B$ must be somewhere negative. Splitting the Heegaard surface as in the proof of Proposition 4.16, we see $B$ corresponds to a zero- or one-cornered $\alpha$-bounded domain $B^{\prime}$ in the marked torus with $n_{z}(B^{\prime})=0$. Since $\alpha_{im}$ is admissible, $B^{\prime}$ is somewhere negative. Therefore, $B$ is somewhere negative. Next we show the pairing diagram $\mathcal{H}_{w,z}(\alpha_{im})$ is bi- admissible. Recall that bi-admissibility means any nontrivial periodic domains $B$ with $n_{w}(B)=0$ or $n_{z}(B)=0$ must have both positive and negative coefficients. To see this, we first claim any given periodic domain $B$ is bounded by some multiple of a homologically trivial component of $\alpha_{im}$. (We warn the reader that the claim is no longer true if one further performs a handleslide of such a component over an embedded alpha curve.) To see the claim, note that as homology classes, the curves $[\alpha_{i}]$ ($i=1,\ldots,g-1$), $[\alpha^{1}_{im}]$, and $[\beta_{i}]$ ($i=1,\ldots,g$) are linearly independent, just as the attaching curves in a Heegaard diagram for $S^{3}$. Now, the claim implies $B$ is a zero-cornered $\alpha$-bounded domain. In view of the unobstructedness established above, $B$ is somewhere negative. ∎ ###### Definition 5.3. Let $\alpha_{im}$ and $\alpha^{\prime}_{im}$ be two admissible immersed multicurves. They are said to be admissibly equivalent if there exists a finite sequence of admissible immersed curves $\alpha^{i}_{im}$, $i=1,\ldots,n$ such that * (1) $\alpha^{1}_{im}=\alpha_{im}$ and $\alpha^{n}_{im}=\alpha^{\prime}_{im}$, * (2) For $i=1,\ldots,n-1$, $\alpha^{i}_{im}$ and $\alpha^{i+1}_{im}$ are related by a finger move that creates/cancels a pair of self-intersection points of the immersed curves. ###### Proposition 5.4. Let $\alpha_{im}$ and $\alpha_{im}^{\prime}$ be two admissibly equivalent immersed multicurves, and let $\mathcal{H}_{w,z}$ be a doubly pointed bordered Heegaard diagram. Then $CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{im}))\cong CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha^{\prime}_{im})).$ ###### Proof. The proof follows the same strategy as the usual proof of isotopy invariance. Let $\bm{\alpha_{0}}$ and $\bm{\alpha_{1}}$ be the two sets of $\alpha$-curves. For simplicity, assume they are related by a single finger move. We model the finger move using a locally supported exact Hamiltonian isotopy on $\Sigma$. The isotopy induces a family of $\alpha$-curves, $\bm{\alpha}_{t}$, on $\Sigma$ ($t\in\mathbb{R}$); for $t\ll 0$ (resp. $t\gg 0$), $\bm{\alpha}_{t}$ is constant with respect to $t$ and is identified with $\bm{\alpha}_{0}$ (resp. $\bm{\alpha}_{1}$). $\bm{\alpha}_{t}$ induces an immersed totally real submanifold $C_{\alpha}=\bm{\alpha}_{t}\times\\{1\\}\times\\{t\\}$ in $\Sigma\times[0,1]\times\mathbb{R}$. $C_{\alpha}$ can be realized as an immersion $\Psi_{t}:(\amalg_{i=1}^{g}S^{1})\times\mathbb{R}\rightarrow\Sigma\times\\{1\\}\times\mathbb{R}.$ Let $C_{\beta}$ be the Lagrangian induced by the $\beta$ curves. For $\bm{x}\in\mathbb{T}_{\bm{\alpha}_{0}}\cap\mathbb{T}_{\bm{\beta}}$ and $\bm{y}\in\mathbb{T}_{\bm{\alpha}_{1}}\cap\mathbb{T}_{\bm{\beta}}$, one then define $\mathcal{M}_{\Psi_{t}}(\bm{x},\bm{y})$ to be the moduli space of holomorphic curves in $\Sigma\times[0,1]\times\mathbb{R}$ with boundary on $C_{\alpha}\cup C_{\beta}$ such that the $\alpha$-boundary can be lifted through $\Psi_{t}$. With this, one can define a map $\Phi_{0}:CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{im}))\rightarrow CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha^{\prime}_{im}))$ by $\bm{x}\mapsto\sum_{\bm{y}}\sum_{\phi\in\pi_{2}(\bm{x},\bm{y})}\\#\mathcal{M}_{\Psi_{t}}(\bm{x},\bm{y})U^{n_{w}(\phi)}V^{n_{z}(\phi)}\bm{y},$ where $\mathcal{M}_{\Psi_{t}}(\bm{x},\bm{y})$ has dimension zero. Define $\Phi_{1}:CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha^{\prime}_{im}))\rightarrow CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{im}))$ similarly. We remark that the compactness and gluing results still apply to this setup. The bi- admissibility of the diagrams obstructs the appearance of boundary degeneration in the compactification of one-dimensional moduli spaces, and hence we can still apply the usual argument to show (1) $\Phi_{0}$ and $\Phi_{1}$ are chain maps, and (2) $\Phi_{0}\circ\Phi_{1}$ and $\Phi_{1}\circ\Phi_{0}$ are homotopy equivalent to the identity map. Therefore, $\Phi_{0}$ and $\Phi_{1}$ are chain homotopy equivalences. ∎ ###### Definition 5.5. An immersed multicurve is called $z$-passable if it is admissibly equivalent to a $z$-adjacent multicurve. ###### Remark 5.6. We can easily arrange $\alpha_{K}$ to be a $z$-passable multicurve; see Example 5.7 below. Moreover, when the pattern knot admits a genus-one doubly pointed Heegaard diagram, we can even drop the admissibility condition; see Section 6.2. ###### Example 5.7. We give a simple way to arrange an immersed multicurve $\alpha_{K}$ to be $z$-passable. Without loss of generality, we consider a single component $\gamma$ of $\alpha_{K}$ each time, and we orient $\gamma$ arbitrarily. We view the torus $T^{2}$ as a square as usual and position $\gamma$ such that the elementary arcs hitting the top edge are separated into two groups of arcs where the arcs in a single group intersect the top edge in the same direction; see Figure 21 (1). Next, we perform a Reidemeister-II-like move to the two groups as in Figure 21 (2). Perform the above modification for every component of $\alpha_{K}$. We claim the resulting multicurve, which we denote $\alpha_{K}^{\prime}$, is a $z$-passable multicurve. Figure 21. (2) is an $z$-passable immersed curve obtained from (1). We justify the claim when $\gamma$ is homologically trivial; the case where $\gamma$ is homologically essential is similar. We first check that $\alpha_{K}^{\prime}$ is admissible by checking that there are no zero- or one-cornered $\alpha$ bounded domains $B$ with $n_{z}(B)=0$. First note that for any zero- or one-cornered $\alpha$ bounded domains $B$, $\partial B$ must include an elementary arc meeting the top edge of the square. To see this, note that $\partial B$ is a nullhomologous curve in the torus and thus lifts to a closed path in the universal cover. Cutting along (lifts of) the meridian (i.e., $\mathbb{Z}\times\mathbb{R}$) breaks $\partial B$ into pieces, with at least two of these (the leftmost and rightmost piece) forming bigons with the meridian. At least one of those two pieces has no corners (since $B$ is zero- or one-cornered). The cornerless piece must intersect the longitude because $\alpha_{K}^{\prime}$ is reduced, and the subarc of $\partial B$ directly below this intersection with the longitude gives an elementary arc meeting the top edge of the square. Next we observe that the elementary arcs near the top edge of the square are arranged such that each arc has the base point $z$ both on its left and on its right, in each case without oppositely oriented arcs in between the arc and $z$, and this implies that no domain whose boundary includes one of these elementary arcs can have $n_{z}(B)=0$. Having shown the immersed curve $\alpha_{K}^{\prime}$ is admissible, it remains to check that it is $z$-passable. Recall from Proposition 4.11 that we can perform a sequence of finger moves to achieve a $z$-adjacent position. Note that all the intermediate diagrams are admissible by exactly the same argument above. ### 5.1. Proof of the main theorem, ungraded version This subsection is devoted to proving the ungraded version of Theorem 1.1. A satellite knot is constructed via the so-called satellite operation that requires a pattern knot and a companion knot as input. A pattern knot is an oriented knot $P$ in an oriented solid torus $S^{1}\times D^{2}$, where an oriented meridian $\mu$ and an oriented longitude $\lambda$ are chosen for $\partial(S^{1}\times D^{2})$ so that the orientation determined by $(\mu,\lambda)$ coincides with the induced boundary orientation. A companion knot is an oriented knot $K$ in the 3-sphere. We orient any Seifert longitude of $K$ using the parallel orientation, and orient any meridian $m$ of $K$ so that $lk(m,K)=1$. The satellite knot $P(K)$ is obtained by gluing $(S^{1}\times D^{2},P)$ to the companion knot complement $S^{3}\backslash\nu(K)$ so that the chosen meridian $\mu$ is identified with a meridian of $K$ and that the chosen longitude $\lambda$ is identified with the Seifert longitude of $K$; $P(K)$ is given by viewing $P$ as a knot in the glued-up 3-sphere $(S^{1}\times D^{2})\cup(S^{3}\backslash\nu(K))$. We state the main theorem again below for the readers’ convenience. Recall that any pattern knot can be represented by a doubly-pointed bordered Heegaard diagram [LOT18, Section 11.4]. See 1.1 Given a doubly pointed bordered Heegaard diagram $\mathcal{H}_{w,z}$ for the pattern knot, we will construct an arced bordered Heegaard diagram $\mathcal{H}_{X(P)}$; the Heegaard diagram $\mathcal{H}_{X(P)}$ specifies a bordered 3-manifold $X(P)$ with two boundary components888Strictly speaking, an arced bordered Heegaard diagram specifies a strongly bordered 3-manifold in the sense of Definition 5.1 in [LOT15], where there is also a framed arc in addition to the underlying bordered 3-manifold. This extra structure will not be relevant to us, so we will not specify it. , where (1) the underlying 3-manifold is $S^{1}\times D^{2}\backslash\nu(P)$, (2) the parametrization of $\partial(S^{1}\times D^{2})$ is the standard meridian-longitude parametrization, and (3) the parametrization of interior boundary $\partial(\nu(P))$ is given by a meridian of $P$ and some longitude of $P$. (The choice of the longitude of $P$ does not matter). We describe how to obtain $\mathcal{H}_{X(P)}$ from the doubly pointed bordered Heegaard diagram $\mathcal{H}_{w,z}$. This is a standard construction, similar to the one appearing in [LOT18] Section 11.7; the reader familiar with it may skip this paragraph and consult Figure 22 for an overview. Figure 22. An example of obtaining $\mathcal{H}_{X(P)}$ from $\mathcal{H}_{w,z}$. Here, $\mathcal{H}_{w,z}$ is showed on the top row; it is a genus-one Heegaard diagram for the $(3,1)$-cable pattern. $\mathcal{H}_{X(P)}$ is the rightmost diagram on the second row. Assume $\mathcal{H}_{w,z}$ is of genus $g$. First, we stabilize $\mathcal{H}_{w,z}=(\bar{\Sigma},\bm{\bar{\alpha}},\bm{\beta},w,z)$ to get a new doubly pointed bordered Heegaard diagram $\mathcal{H}^{\prime}_{w,z}=(\bar{\Sigma}^{\prime},\bm{\bar{\alpha}}\cup\\{\alpha^{c}_{g}\\},\bm{\beta}\cup\\{\beta_{g+1}\\},w,z)$. More concretely, $\bar{\Sigma}^{\prime}$ is obtained from $\bar{\Sigma}$ by attaching a two-dimensional one-handle, with feet near the base points $w$ and $z$. Parametrize the new one-handle by $S^{1}\times[0,1]$, where $S^{1}\times\\{0\\}$ is the feet circle near $z$, and $S^{1}\times\\{1\\}$ is the feet circle near $w$. We also parametrize $S^{1}$ by $[0,2\pi]/(0\sim 2\pi)$. The new $\alpha$-circle $\alpha^{c}_{g}$ is the belt circle $S^{1}\times\\{1/2\\}$ of the new one-handle. Let $p_{1}=(0,0)$ and $p_{2}=(0,1)$ be two points on the two feet circles of the one-handle. The new $\beta$ circle $\beta_{g+1}$ is the union of two arcs $l_{1}$ and $l_{2}$ connecting $p_{1}$ and $p_{2}$, where $l_{1}$ is an arc in $\bar{\Sigma}\backslash\bm{\beta}$ and $l_{2}$ is the arc $\\{(0,t)|t\in[0,1]\\}$ in new one-handle. Next, introduce a new curve $\bar{\alpha}_{1}^{a,L}$ as follows. Let $l_{z}$ be an arc from $z$ to the point $(-1,0)\in S^{1}\times\\{0\\}$ does not intersect any of the $\alpha$\- and $\beta$-curves. Let $l_{2}^{\prime}$ be the arc $\\{(1,t)|t\in[0,1]\\}$ in the one-handle; denote the endpoints of $l_{2}^{\prime}$ by $p_{1}^{\prime}$ and $p_{2}^{\prime}$. Let $l_{1}^{\prime}$ be an arc connecting $p_{1}^{\prime}$ and $p_{2}^{\prime}$ in $\bar{\Sigma}\backslash\\{\bm{\bar{\alpha}}\cup l_{z}\\}$. Let $\bar{\alpha}_{1}^{a,L}=l_{1}^{\prime}\cup l_{2}^{\prime}$. Then $\bar{\alpha}_{1}^{a,L}$ intersects $\alpha^{c}_{g}$ geometrically once at a point $p$. Note $\alpha_{g}^{c}$ is the meridian of $P$, and $\bar{\alpha}_{1}^{a,L}$ is a longitude of $P$. Let $\bar{\Sigma}^{\prime\prime}$ be the circle compactification of $\bar{\Sigma}^{\prime}\backslash\\{p\\}$. Denote the new boundary circle by $\partial_{L}\bar{\Sigma}^{\prime\prime}$, and denote the boundary circle inherited from $\partial\bar{\Sigma}$ by $\partial_{R}\bar{\Sigma}^{\prime\prime}$. Let ${\alpha}_{1}^{a,L}=\bar{\alpha}_{1}^{a,L}\backslash\\{p\\}$, and let ${\alpha}_{2}^{a,L}=\alpha^{c}_{g}\backslash\\{p\\}$. Let $\alpha_{1}^{a,R}=\alpha_{1}^{a}$, and let $\alpha_{2}^{a,R}=\alpha_{2}^{a}$. Let $\bm{\bar{\alpha}}^{\prime\prime}=\\{\alpha^{a,L}_{1},\alpha^{a,L}_{2},\alpha^{a,R}_{1},\alpha^{a,R}_{2},\alpha^{c}_{1},\ldots,\alpha^{c}_{g-1}\\}$. Let $\bm{\beta}^{\prime\prime}=\bm{\beta}\cup\\{\beta_{g+1}\\}$. Label the Reeb chords corresponding to the new boundary circle $\partial_{L}\bar{\Sigma}^{\prime\prime}$ by $\sigma_{i}$ ($i=0,1,2,3$) so that $\sigma_{2}$ and $\sigma_{3}$ lie on the side attached to the feet near $w$, and $\sigma_{0}$ and $\sigma_{1}$ lie on the side attached to the feet near $z$. Let $z_{R}=z$, and let $z_{L}$ be a point on $\sigma_{0}$. Let $\bm{z}$ be an arc connecting $z_{R}$ and $z_{L}$ in the complement of $\bm{\bar{\alpha}}^{\prime\prime}\cup\bm{\beta}^{\prime\prime}$; $\bm{z}$ exists since we can obtain such an arc by extending $l_{z}$. Finally, we let $\mathcal{H}_{X(P)}=(\bar{\Sigma}^{\prime\prime},\bm{\bar{\alpha}}^{\prime\prime},\bm{\beta}^{\prime\prime},\bm{z})$. See Figure 22. ###### Lemma 5.8. Let $\mathcal{H}_{X(P)}$ be the arced bordered Heegaard diagram obtained from $\mathcal{H}_{w,z}$ via the above procedure. Let $\alpha_{im}$ be a $z$-adjacent multicurve. Then $\mathcal{H}_{X(P)}(\alpha_{im})$ is unobstructed and bi-admissible. ###### Proof. The unobstructedness follows from Proposition 4.16. We move to see bi- admissibility in the sense of Definition 2.10. Note that periodic domains $B$ for $\mathcal{H}_{X(P)}(\alpha_{im})$ with $n_{\sigma_{0}}(B)=n_{\sigma_{1}}(B)=0$ (respectively $n_{\sigma_{2}}(B)=n_{\sigma_{3}}(B)=0$) correspond to periodic domains $B^{\prime}$ for $\mathcal{H}_{w,z}(\alpha_{im})$ with $n_{z}(B^{\prime})=0$ (respectively $n_{w}(B^{\prime})=0$). Therefore, the bi-admissibility of $\mathcal{H}_{w,z}(\alpha_{im})$, which was shown in Proposition 5.2, implies the bi-admissibility of $\mathcal{H}_{X(P)}$. ∎ Recall $\mathcal{H}_{id}$ is the standard doubly pointed bordered Heegaard diagram for the identity pattern knot. ###### Lemma 5.9. $CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{im}))$ is chain homotopy equivalent to $CFK_{\mathcal{R}}(\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{im}))$. ###### Proof. Note the doubly pointed Heegaard diagram $\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{im})$ is obtained from $\mathcal{H}_{w,z}(\alpha_{im})$ by two stabilizations; see Figure 23. Figure 23. An example of $\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{im})$ (left) and $\mathcal{H}_{w,z}(\alpha_{im})$ (lower right). Here, $P$ is the $(3,1)$-cable. These two diagrams are related via handleslides and destabilizations, where the handleslides do not involve sliding over the immersed $\alpha$-curve. In particular, it is also bi-admissible, and hence one can define $CFK_{\mathcal{R}}(\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{im}))$. We claim there is a sequence of Heegaard moves relating $\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{im})$ and $\mathcal{H}_{w,z}(\alpha_{im})$ which do not involve sliding $\alpha$ curves over $\alpha_{im}$. To see this, note that on $\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{im})$ there is a $\beta$-circle between the $w$ and $z$ base points that intersects an $\alpha$-circle geometrically once; denote these curves by $\beta_{g+2}$ and $\alpha_{g+2}$ respectively. After sliding other beta curves over $\beta_{g+2}$ if necessary, we may assume $\alpha_{g+2}$ does not intersect other beta curves, and hence we can destabilize $\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{im})$ along $\alpha_{g+2}$ and $\beta_{g+2}$. Now we arrive at an intermediate Heegaard diagram; see Figure 23 (upper right). It is a stabilization of $\mathcal{H}_{w,z}(\alpha_{im})$. On this intermediate Heegaard diagram, there is an $\alpha$-circle $\alpha_{g+1}$ that intersects only one $\beta$-circle $\beta_{g+1}$, and the geometric intersection number is one. So, we may slide other $\alpha$-curves over $\alpha_{g+1}$ if necessary so that $\beta_{g+1}$ do not intersect other $\alpha$-curves. After this, we destabilize the Heegaard diagram along $\alpha_{g+1}$ and $\beta_{g+1}$, and the resulting Heegaard diagram is $\mathcal{H}_{w,z}(\alpha_{im})$. The homotopy equivalence between $CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{im}))$ and $CFK_{\mathcal{R}}(\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{im}))$ follows from the homotopy invariance of knot Floer chain complexes established in Proposition 3.11. ∎ With these lemmas at hand, we now prove the ungraded version of Theorem 1.1. ###### Proof of Theorem 1.1, ungraded version. In view of Proposition 5.4, we may assume the immersed multicurve $\alpha_{K}$ for the knot complement of $K$ is z-adjacent. Let $\mathcal{H}_{X(P)}$ be the arced bordered Heegaard diagram obtained from $\mathcal{H}_{w,z}$ via the “punctured-stabilization procedure”. Throughout, when referring to the type D structure of a knot complement, we use the meridian and Seifert longitude to parametrize the boundary. By standard arguments, we can arrange that $\mathcal{H}_{X(P)}$ is left provincially admissible at the cost of isotopy of the $\beta$ curves. By Theorem 1.4, we have $\displaystyle\widehat{CFD}(\mathcal{H}_{X(P)}(\alpha_{K}))$ $\displaystyle\cong\widehat{CFDA}(\mathcal{H}_{X(P)})\boxtimes\widehat{CFD}(\alpha_{K})$ $\displaystyle\cong\widehat{CFDA}(\mathcal{H}_{X(P)})\boxtimes\widehat{CFD}(S^{3}\backslash\nu(K))$ $\displaystyle\cong\widehat{CFD}(S^{3}\backslash\nu(P(K)))$ Therefore, up to homotopy equivalence, the extended type D structure $\widetilde{CFD}(\mathcal{H}_{X(P)}(\alpha_{K}))$ extends $\widehat{CFD}(S^{3}\backslash\nu(P(K)))$. Consequently, we have the following: $\displaystyle CFK_{\mathcal{R}}(P(K))$ $\displaystyle\cong\widetilde{CFA}(\mathcal{H}_{id}){\boxtimes}\widetilde{CFD}(S^{3}\backslash\nu(P(K)))$ $\displaystyle\cong\widetilde{CFA}(\mathcal{H}_{id}){\boxtimes}\widetilde{CFD}(\mathcal{H}_{X(P)}(\alpha_{K}))$ $\displaystyle\cong CFK_{\mathcal{R}}(\mathcal{H}_{id}\cup\mathcal{H}_{X(P)}(\alpha_{K}))$ Here, the last equality follows from applying Theorem 1.6. Note ${\boxtimes}$ in the above equation is well-defined since $\mathcal{H}_{X(P)}(\alpha_{K})$ is bi-admissible by Lemma 5.8. Now, by Lemma 5.9, $CFK_{\mathcal{R}}(P(K))$ is chain homotopy equivalent to $CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{K}))$. ∎ ### 5.2. $\mathcal{H}_{w,z}(\alpha_{K})$ is gradable We want to show that the chain homotopy equivalence established in the previous subsection preserves the $w$-grading and $z$-grading of knot Floer chain complexes. As the first step, we need to show that $\mathcal{H}_{w,z}(\alpha_{K})$ is gradable (in the sense of Definition 3.7). ###### Proposition 5.10. The diagram $\mathcal{H}_{w,z}(\alpha_{K})$ is gradable. In addition to being gradable, note that the results in the previous subsection also imply that $\widehat{HF}(\mathcal{H}_{w}(\alpha_{K}))\cong\widehat{HF}(\mathcal{H}_{z}(\alpha_{K}))\cong\mathbb{F}$. Therefore we can define an absolute bigrading on $CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{K}))$. We will reduce the proof of Proposition 5.10 to the case where $\mathcal{H}_{w,z}$ is of genus one. If $\mathcal{H}_{w,z}$ is a genus-one bordered Heegaard diagram, then one can define a Maslov grading $m(-)$ on $CFK_{\mathcal{R}}(\mathcal{H}_{w,z}(\alpha_{K}))$ as follows. Given any two generators $x$ and $y$, let $p_{0}$ and $p_{1}$ be two paths from $x$ to $y$ in $\alpha_{K}$ and $\beta$ respectively such that $p_{0}-p_{1}$ lifts to a closed path $\gamma$ in the universal cover $\mathbb{R}^{2}$ of the genus-one Heegaard surface. Up to perturbing the curves, we may assume that $p_{0}$ and $p_{1}$ intersect in right angles at $x$ and $y$. Then $m(x)-m(y)$ is equal to $\frac{1}{\pi}$ times the total counterclockwise rotation along the smooth segments of $\gamma$ minus twice the number of the (lifts of) base point $z$ enclosed by $\gamma$; see [HRW22, Definition 35]. This Maslov grading is also defined (by the same definition) when the $\beta$ curve is only immersed. In [HRW22], it is shown that the Maslov grading thus defined on a pairing diagram of two immersed curves agrees with the Maslov grading computed using the grading package of bordered Heegaard Floer homology. Next, we show this Maslov grading can be equivalently defined in terms of the index of domains. ###### Proposition 5.11. Let $\mathcal{H}_{w,z}$ be a genus-one bordered Heegaard diagram and let $m(-)$ be the Maslov grading on $\mathcal{G}(\mathcal{H}(\alpha_{K}))$ mentioned above. Let $B\in\pi_{2}(x,y)$ be a domain connecting $x$ and $y$ with $\partial B=p_{0}-p_{1}$. Then $m(x)-m(y)=\text{ind}(B)-2n_{z}(B)$. Moreover, this result extends to the case where the $\beta$ is immersed, in which we define the index of $B$ by $\text{ind}(B)=e(B)+n_{x}(B)+n_{y}(B)-s(\partial_{\alpha_{K}}B)-s(\partial_{\beta}B).$ (Here $s(-)$ denotes the self-intersection number of an oriented immersed arc as defined in Section 2.6) Before proving Proposition 5.11 we introduce some terminology. It will be clear later that we can assume $p_{0}-p_{1}$ is immersed and only has discrete double points. ###### Definition 5.12. A cornered immersed loop in $T^{2}$ is the union of two oriented immersed arcs $p_{0}$ and $p_{1}$ with at most discrete double points such that * (1) $p_{0}$ and $p_{1}$ share common endpoints, * (2) the interior of $p_{0}$ and $p_{1}$ intersect transversally, * (3) $p_{0}-p_{1}$ is an oriented loop which is null-homologous, * (4) $p_{0}$ and $p_{1}$ intersect transversally at the endpoints if $p_{0}$ and $p_{1}$ are non-degenerate (i.e., not a point), and * (5) if one of $p_{0}$ and $p_{1}$ is degenerate, the remaining arc forms a smooth loop after identifying the endpoints. The endpoints of $p_{0}$ (or equivalently, $p_{1}$) are called corners of the cornered immersed loop. ###### Definition 5.13. Two cornered immersed loops $p_{0}-p_{1}$ and $p_{0}^{\prime}-p_{1}^{\prime}$ in $T^{2}$ are called cornered identical if they share the same set of corners $\\{x,y\\}$ (or $\\{x\\}$ if the loops have degenerate arcs) and there are arbitrarily small neighborhoods $N_{x}$ and $N_{y}$ of $x$ and $y$ respectively such that $(p_{0}-p_{1})|_{N_{x}}=(p^{\prime}_{0}-p^{\prime}_{1})|_{N_{x}}$ and $(p_{0}-p_{1})|_{N_{y}}=(p^{\prime}_{0}-p^{\prime}_{1})|_{N_{y}}$. Figure 24. Upper row from left to right: Reidemeister I, II, and III move. Lower row from left to right: an isotopy that crosses a non-degenerate corner and a degenerate corner. ###### Lemma 5.14. If two cornered immersed loops $p_{0}-p_{1}$ and $p_{0}^{\prime}-p_{1}^{\prime}$ are cornered identical, then they are related by a finite sequence of moves of the following types: * (1) Reidemeister moves that do not involve the corners and * (2) isotopy that possibly cross the corners. (See Figure 24) Here, we require $(p_{0}-p_{1})|_{N_{x}}$ and $(p_{0}-p_{1})|_{N_{y}}$ are fixed throughout the modification for some sufficiently small neighborhoods ${N_{x}}$ and ${N_{y}}$ of the corners. ###### Proof. One can prove this applying the usual Reidemeister-move equivalence of knot diagrams (by treating both immersed loops as diagrams for the unknot via imposing proper crossing information); note that any Reidemeister move involving a corner can be traded by an isotopy crossing the corner and a Reidemeister move that does not involve the corner. ∎ ###### Definition 5.15. Given a cornered immersed loop $p_{0}-p_{1}$ in $T^{2}$. Let $\tilde{p}_{0}-\tilde{p}_{1}$ be a lift of $p_{0}-p_{1}$ in $\mathbb{R}^{2}$ and let $\tilde{B}$ be the bounded domain in $\mathbb{R}^{2}$ such that $\partial\tilde{B}=\tilde{p}_{0}-\tilde{p}_{1}$. Let $B$ be the domain in $T^{2}$ obtained from $\tilde{B}$ by applying the covering projection. Define the index of the cornered immersed loop as $\text{ind}(p_{0}-p_{1})=e(B)+n_{x}(B)+n_{y}(B)-s(\partial_{p_{0}}B)-s(\partial_{p_{1}}B),$ where $x$ and $y$ are the corners. Define the net rotation number $nr(p_{0}-p_{1})$ to be $\frac{1}{\pi}$ times the counterclockwise net rotation along the smooth segments $p_{0}$ and $p_{1}$. ###### Lemma 5.16. Suppose $p_{0}-p_{1}$ and $p_{0}^{\prime}-p_{1}^{\prime}$ are cornered immersed loops differ by an isotopy or a Reidemeister move. Then $\text{ind}(p_{0}-p_{1})-\text{ind}(p^{\prime}_{0}-p^{\prime}_{1})=nr(p_{0}-p_{1})-nr(p^{\prime}_{0}-p^{\prime}_{1}).$ Figure 25. Local diagrams for isotopies that cross a corner. The numbers $a$, $a-1$, and $a+1$ indicate the multiplicities of the regions. ###### Proof. First, we examine the effect of an isotopy on both quantities. Clearly, the net rotation number is unchanged. We separate the discussion of the index into two cases according to whether the isotopy crosses corners or not. If the isotopy does not cross the corners, it clearly does not change the index as well whence we are done. If the isotopy crosses a corner, then we claim the local multiplicity and the self-intersection numbers change in a way that cancel each other, leaving the index unchanged. This claim can be seen by examining local diagrams, which are further divided into two cases according to whether the corner is degenerate or not. When the corner is non-degenerate, the local diagram of one case is shown in Figure 25 (i); all the other cases can be obtained from this case by swapping the labels and orientations of the arcs, and the analysis of all cases are similar. In the case shown in Figure 25 (i), only $n_{x}(B)$ and $s(\partial_{p_{0}}B)$ change: the diagram on the left has $n_{x}(B)=\frac{a+(a-1)+(a-1)+(a-1)}{4}$ and the local self- intersection of $p_{0}$ contributes $s_{p_{0}}=-1$; the diagram on the right has $n_{x}(B)=\frac{(a+1)+a+a+a}{4}$ and there are no self-intersections of the arcs in the local diagram so the local contribution $s_{p_{0}}=0$. In both diagrams we have $n_{x}(B)-s_{p_{0}}=\frac{4a+1}{4}$, and hence the index is unchanged. When the corner is degenerate, one of the cases is shown in Figure 25 (ii). In this case, only $n_{x}$ and the self-intersection of $p_{0}$ change: the diagram on the left has $n_{x}=\frac{a+a+(a-1)+(a-1)}{4}$ and a local contribution of the self-intersection of $p_{0}$ given by $s_{p_{0}}=-1$; the diagram on the right has $n_{x}=\frac{(a+1)+(a+1)+a+a}{4}$ and a local contribution of the self-intersection of $p_{0}$ given by $s_{p_{0}}=1$. In both local diagrams we have $n_{x}(B)+n_{x}(B)-s_{p_{0}}=2a$, and hence the index is unchanged. All other cases can be obtained from this case by swapping the labels and orientations of the arcs, and the analysis of all cases are similar. Figure 26. The local diagram for Reidemeister I move. The numbers $a$, $a-1$, and $a-2$ indicate the multiplicities of the regions. Next, we examine the effect of Reidemeister I move. Up to swapping orientations and the labels, we may assume the local diagram is as shown in Figure 26 The net rotation number of the diagram on the right is 2 less than that of the diagram on the left. For the index comparison, the Euler measure of the local domain on the right is $1$ less than that of the left diagram and the self-intersection number $s(\partial_{p_{0}}B)$ of the right diagram is $1$ more than that of the left diagram; in total the index of the diagram on the right is 2 less than that of the diagram on the left. Therefore, the changes in the net rotation and in the index are the same after doing a Reidemeister I move. Next, we examine the effect of Reidemeister II moves. It does not change the net rotation number. Also, it does not affect the Euler measure and the local multiplicities at the corners. A Reidemeister II move creates/annihilates a pair of self-intersection points whose signs cancel each other if both arcs involved are on $p_{0}$ or $p_{1}$, and otherwise does not involve self- intersections; in both cases the self-intersection numbers are unchanged. So, the index does not change as well. Finally, it is easy to see that a Reidemeister III move does not change the net rotation number. It is also easy to see a Reidemeister III move does not change the Euler measure, local multiplicities at the corners, or self- intersections, and hence it does not change the index either. ∎ ###### Proposition 5.17. Let $p_{0}-p_{1}$ be a cornered immersed loop. Then $\text{ind}(p_{0}-p_{1})=nr(p_{0}-p_{1})$. ###### Proof. By Lemma 5.16 and Lemma 5.14, it suffices to show that $p_{0}-p_{1}$ is cornered identical with some cornered immersed loop whose index coincides with the net rotation number. If at least one of $p_{0}$ and $p_{1}$ is degenerate, $p_{0}-p_{1}$ is cornered identical with an embedded circle that passes the corner, and it is easy to see the index and the net rotation number coincide on an embedded circle with a degenerate corner. Figure 27. Deforming $p_{0}-p_{1}$. Next, we discuss the case where $p_{0}-p_{1}$ is non-degenerate. We first construct a cornered immersed loop $p_{0}^{\prime}-p_{1}^{\prime}$ that is cornered identical to $p_{0}-p_{1}$ as follows. Let $p_{0}^{\prime}=p_{0}$. We shall construct $p_{1}^{\prime}$ to be a path which is almost a parallel push- off of $p_{0}$. (See Figure 27 for examples.) To spell out the construction, let $f_{0}:[0,1]\rightarrow T^{2}$ be an immersion such that $f_{0}([0,1])=p_{0}$. Let $\hat{N}$ be a sufficiently small tubular neighborhood of $p_{0}$ such that it can realized as the image of an extension of $f_{0}$, i.e., there exits an immersion $\tilde{f_{0}}:[0,1]\times[-\epsilon,\epsilon]\rightarrow T^{2}$ such that $\tilde{f_{0}}|_{[0,1]\times\\{0\\}}=f_{0}$ and $\tilde{f_{0}}([0,1]\times\\{pt\\})$ is a parallel push-off of $p_{0}$ for any $pt\in[-\epsilon,0)\cup(0,\epsilon]$. We can further assume near the two corners $x=f_{0}(0)$ and $y=f_{0}(1)$, the other arc $p_{1}$ is contained in $\tilde{f_{0}}(\\{0,1\\}\times[-\epsilon,\epsilon])$; denote these two arcs on $p_{1}$ near $x$ and $y$ by $p_{x}$ and $p_{y}$ respectively. We construct
# Full-frequency dynamic convolution: a physical frequency-dependent convolution for sound event detection ###### Abstract Recently, 2D convolution has been found unqualified in sound event detection (SED). It enforces translation equivariance on sound events along frequency axis, which is not a shift-invariant dimension. To address this issue, dynamic convolution is used to model the frequency dependency of sound events. In this paper, we proposed the first full-dynamic method named _full-frequency dynamic convolution_ (FFDConv). FFDConv generates frequency kernels for every frequency band, which is designed directly in the structure for frequency- dependent modeling. It physically furnished 2D convolution with the capability of frequency-dependent modeling. FFDConv outperforms not only the baseline by 6.6% in DESED real validation dataset in terms of PSDS1, but outperforms the other full-dynamic methods. In addition, by visualizing features of sound events, we observed that FFDConv could effectively extract coherent features in specific frequency bands, consistent with the vocal continuity of sound events. This proves that FFDConv has great frequency-dependent perception ability. Code is available at FFDConv. Index Terms— sound event detection, full-frequency dynamic convolution, frequency-dependent modeling, independent representation spaces, vocal continuity ## 1 introduction Sound event detection (SED) is one of the subtasks of computational auditory scene analysis (CASA) [1], which helps machines understand the content of an audio scene. Similar to visual object detection [2] and segmentation [3], SED aims to detect sound events and corresponding timestamps (onset and offset), considered as a prior task of automatic speech recognition (ASR) and speaker verification. It has wide applications in information retrieval [4], smart homes [5], and smart cities [6]. Fig. 1: Illustration of frequency-dependent modeling. Top models time- frequency patterns in the same space with a shared kernel. Bottom models them in serval spaces with frequency-adaptive kernels, in which time-frequency patterns specific to sound events can be considered. SED has achieved great success with the help of deep learning (DL). The general paradigm is that acoustic spectral features are passed through a deep neural network and then transformed into discriminative acoustic representations to distinguish different sound events. Designing an effective feature extractor has become a hot topic in SED, which has been adopting methods qualified in other domains in the past few years. Convolutional neural networks (CNN) from the field of computer vision, such as SENet [7], SKNet [8], and CBAM [9] have been migrated to SED in the spirit of acoustic spectrogram being similar to two-dimensional image data. With the intention that speech and audio are both sound data, Conformer [10] from the field of speech recognition has migrated to SED. However, they all failed to show good performance in SED. Specifically, SENet, SKNet, and CBAM are designed on image data with a clear 2D spatial concept, while audio data is a sequence. Conformer is designed on speech data containing only the speech sound event, meaning time-frequency patterns of speech data are distributed only in a certain fixed frequency band. However, audio data always contains multiple sound events, and so has diverse time-frequency patterns of sound events. All of the above emphasize that DL methods qualified in other domains may not necessarily be compatible with SED. Dynamic convolution network [11] was initially proposed for video prediction. It was designed to generate future frames based on the motion pattern within a particular video. The parameters of the dynamic convolution kernel are always adapted to the input. In SED, different sound events are distributed in different frequency regions, and this frequency dependence is invariant over time. This has motivated some researchers to investigate whether dynamic convolution can improve the capability of 2D convolution in modeling the frequency dependence of sound events. [12] proposed frequency dynamic convolution (FDConv), which found that the time-frequency spectrogram is not translation invariant on frequency dimension like image data. FDConv extracts frequency-adaptive attention weights from input for several pre-initialized convolution kernels. These kernels are then weightedly combined in the number dimension to obtain one convolution kernel. Then, the combined kernel is convoluted with the input in a standard manner. [12, 13], [14] proposed multi- dimensional frequency dynamic convolution (MFDConv), which extends the frequency-adaptive dynamic properties of convolutional kernels to more dimensions of the kernel space, i.e. in-channels, out-channels, and kernel numbers. Although FDConv and MFDConv have achieved great performance, they are essentially the same as basic convolution, which is spatially shared. They belong to semi-dynamic convolution in the field of dynamic convolution. As shown in the upper part of Fig. 1, their perception abilities of different frequency bands are identical. They can only model time-frequency patterns in one representation space, where sound events are not easily recognized from each other. Compared with semi-dynamic convolution, full-dynamic convolution [11, 15, 16, 17, 18] attracts more attention recently, which uses a separate network branch to predict a specific filter for each pixel. [18] found this type of dynamic convolution is equivalent to applying attention on unfolded input features, which enables it more effective when modeling complex patterns. Sound events’ time-frequency patterns are highly frequency- dependent, and full-dynamic convolution can model features of spatial pixels with different filters. Full-dynamic convolution may be optimal in dealing with recognizing sound events. In this paper, we propose a novel method named _full-frequency dynamic convolution_ (FFDConv), which is the first full-dynamic convolution method for SED. As shown in the lower part of Fig. 1, FFDConv generates frequency- specific kernels, resulting in distinct representation spaces. This design is applied directly in the structure for frequency-dependent modeling. In this way, the 2D convolution is physically furnished with the capability of frequency-dependent modeling, so that the specific time-frequency patterns can be acquired for different sound events. In the end, sound events can be easily recognized from each other in subsequent classification. The main contributions of this paper are summarized as follows: * • We proposed full-frequency dynamic convolution that can model time-frequency patterns in independent representation spaces. This method will extract more discriminative features of sound events, resulting in effective classification. * • The Proposed method outperforms not only baseline but also pre-existing full dynamic filters method in other domain. * • By visualizing features of sound events, we found the ability to model temporally coherent features is essential to the detection of sound events. And the FFDConv has this ability. ## 2 mthodology ### 2.1 Full-dynamic convolution A basic 2D convolution can be denoted as $y=\boldsymbol{W}\ast x+\boldsymbol{b}$, where $x\in{\mathbb{R}^{T\times F\times C_{in}}}$ and $y\in{\mathbb{R}^{T\times F\times C_{out}}}$ denote the input feature and output feature; $\boldsymbol{W}\in{\mathbb{R}^{k\times k\times C_{in}\times C_{out}}}$ and $\boldsymbol{b}\in{\mathbb{R}^{C_{out}}}$ denote the weight and bias of a basic convolution kernel. In contrast to basis convolution, full- dynamic convolution [11] leverages separate network branches to generate the filters for each pixel. Full-dynamic convolution operation can be written as: $\displaystyle y=\boldsymbol{Concat}$ $\displaystyle{(\boldsymbol{W}_{t,f}\ast x(t,f))}$ (1) $\displaystyle\boldsymbol{W}_{t,f}=$ $\displaystyle G(x,t,f)$ where $\boldsymbol{W}_{t,f}$ denotes weight for the current pixel; The $G$ is the filter generating function; $Concat$ here aims to convey that convolution operation of each pixel is independent. For simplicity, the bias term is omitted. Fig. 2: Illustration of full-frequency dynamic convolution. In general, the factory produces frequency-dependent kernels from acoustic feature, and then kernels are convoluted with input along the time axis. In the factory, there are two workshops aiming to produce spatial filters and channel filters, respectively. And they are integrated in the assembly workshop. ### 2.2 Overall of proposed method As is commonly understood, different sound events have different frequency band distributions. For instance, catcall, which is sharp, shrill, and high- pitched, is often heard in the high-frequency range; running water, which is low, soft, and soothing, is often heard in the low-frequency range. Based on this, we explore designing a new convolution for SED, which can capture the distribution of frequency bands and model time-frequency patterns of sound events in different frequency representation spaces. Inspired by full dynamic convolution [18], we designed the full-frequency dynamic convolution (FFDConv) for SED. Overall, as shown in Fig. 2, FFDConv employs a separate branch to predict kernels for each frequency band, in which the content of kernels is based on input feature. In the kernel-generating branch, there are two sub-branches: the spatial filter-generating branch for the spatial space of kernels and the channel filter-generating branch for the channel space of kernels. After spatial and channel filters are obtained, they are combined and then convoluted with the input feature. Note that similarly, full-temporal dynamic convolution (FTDConv) predicts kernels for each temporal frame, and kernels are convoluted with input along the frequency axis. ### 2.3 Full-frequency dynamic convolution Unlike the previous semi-dynamic convolution, FFDConv is designed directly in the structure for frequency-dependent modeling. It models the feature along the frequency axis in different representation spaces. Mathematically, FFDConv can be written as: $\displaystyle y=\boldsymbol{Concat}$ $\displaystyle{(\boldsymbol{W}_{f}\ast x(f),\,dim=f)}$ (2) $\displaystyle\boldsymbol{W}_{f}=$ $\displaystyle\;G_{s}(x,f)\odot G_{c}(x,f)$ where $\boldsymbol{W}_{f}$ is the content-adaptive kernel for the $f^{th}$ frequency band; $x(f)\in\mathbb{R}^{T}$ is the $f^{th}$ frequency band of input feature; $G_{s}$ and $G_{c}$ are the spatial and channel filter- generating function; $\odot$ denotes the elemental dot product operator. For clarity, $Concat$ here aims to convey that $\boldsymbol{W}_{f}$ is convoluted with input along the time axis. FFDConv employs a separate branch to generate convolution kernels for each frequency band, in which there are two sub-branches: spatial filter-generating branch and channel filter-generating branch. The spatial filter-generating module is designed to predict the spatial content of dynamic kernels, and the channel-generating module is designed to predict the channel content of dynamic kernels. For efficiency, the dynamic filters are decoupled into spatial and channel ones, following [18]. Spatial filter generating. As illustrated in Fig. 3, we use a standard Conv2D to compress the time dimension of input and map channel dimension from $C$ to $K^{2}$, whose kernel weight $W\in\mathbf{R}^{C\times K^{2}\times T\times W}$, where $W$ is the window size of the kernel in the frequency dimension. It moves along the frequency axis when convoluted with input. In this way, not only are the adjacent frequency components considered, but information along the time axis is aggregated. Then, the spatial filter of FFDConv is obtained, which assigns $K\times K$ spatial weight to every frequency kernel and is highly related to the input. Note that full dynamic convolution [18] assigns $K\times K$ spatial weight to every pixel. Consequently, FFDConv can model features from different frequency bands of the input in independent representation spaces. Considering these representation spaces may be far apart from each other, we employ an attention module following [12] to limit individual differences between them so as not to be too large. Finally, the spatial filter is passed through a Filter-Norm module following [18], avoiding the gradient vanishing/exploding during training. Channel filter generating. As illustrated in Fig. 3, the channel filter generating module is similar to the SE block [7]. It compresses the time and frequency feature of input by applying an average pooling and maps the channel dimension from $C$ to $CK^{2}$ by two fully connected (FC) layers. Between two fully connected layers, the ReLU activation function is applied to introduce non-linearity. After input is passed through this module, the channel filter of FFDConv is obtained, which assigns $C$ channel weight to each spatial location of the frequency kernel. It should be noted that the channel filter for $F$ frequency kernels is the same. In the end, the channel filter is also passed through the Filter-Norm [18]. The spatial and channel filters are mixed by dot product, and the full frequency kernels are obtained. We then use them to model time-frequency patterns of input features. ### 2.4 FFDConv block Considering frequency kernels of FFDConv don’t have output channel dimension, we design an FFDConv block that contains the channel mapping. As illustrated in Fig. 3, firstly, the channel dimension of input is mapped from $C_{in}$ to $C_{out}$ after passing through the channel transformation module. Then, based on the input feature, the spatial and channel filters are obtained by passing through the spatial and channel filter generating module. Full-frequency dynamic kernels are obtained by mixing the spatial and channel filters. Finally, the kernels are convoluted with input along the time axis. In the actual algorithm, following [18], spatial filters, channel filters, and input are sent to DDF operation to get the output, which is implemented in CUDA, alleviating any need to save intermediate multiplied filters during network training and inference. Note that the DDF op needs $H\times W$ spatial filters. We repeat the $1\times F$ spatial filters to $T\times F$ so that the kernel’s weights are the same along the time axis when convoluted with input in $f^{th}$ frequency band. Fig. 3: Details of the FFDConv ## 3 experiment ### 3.1 Dataset and experiment setup All experiments are conducted on the dataset of Task 4 in the DCASE 2022. The training set consists of three types of data: weakly labeled data (1578 clips), synthetic strongly labeled data (10000 clips), and unlabeled in-domain data (14412 clips). The real validation set (1168 clips) is used for evaluation. The input acoustic feature is the log Mel spectrogram extracted from 10-second-long audio data with a sampling rate of 16 kHz. The feature configuration is the same as [13], in which the input feature has 626 frames and 128 mel frequency bands. The baseline model is the CRNN architecture [19], which consists of 7 layers of conv blocks and 2 layers of Bi-GRU. Attention pooling module is added at the last FC layer for joint training of weakly labeled data, and mean teacher (MT) [20] is applied for consistency training with unlabeled data for semi- supervised learning. Data augmentations such as MixUp [21], time masking [22], frame-shift, and FilterAugment [23] are used. The data augmentation parameters are identical to [12]. Poly-phonic sound event detection scores (PSDS), collar-based F1 score (EB-F1), intersection-based F1 score (IB-F1) are used to evaluate the model performance. Median filters with fixed time length are used for post- processing, and sound events have different thresholds from each other to obtain hard predictions for calculating EB-F1. The metrics hyper-parameters are identical to [12]. The model is trained using the Adam optimizer with a maximum learning rate of 0.001, and ramp-up is used for the first 80 epochs. Table 1: _SED performance comparison between models using different dynamic convolution on the validation set._ Model | PSDS1 $\uparrow$ | PSDS2 $\uparrow$ | CB-F1 $\uparrow$ | IB-F1 $\uparrow$ ---|---|---|---|--- Baseline [19] | 0.370 | 0.579 | 0.469 | 0.714 DDFConv [18] | 0.387 | 0.624 | 0.467 | 0.720 FTDConv | 0.395 | 0.651 | 0.495 | 0.740 FFDConv | 0.436 | 0.685 | 0.526 | 0.751 ### 3.2 Full-frequency dynamic convolution on SED We compared the performances of baseline with full dynamic convolution methods, including decoupled dynamic convolution (DDFConv) [18], full-temporal dynamic convolution (FTDConv), and full-frequency dynamic convolution (FFDConv). For full dynamic convolution methods, dynamic convolution layers replaced all convolution layers except the first layer from the baseline model [19]. The results are shown in Table 1. Three types of full dynamic convolution can all outperform the baseline, which proves full dynamic convolution qualifies in SED. In addition, it can be seen that the effects of three types of convolution are in increasing order. First, FTDConv and FFDConv employ content-adaptive temporal or frequency kernels, which can be viewed as giving prior knowledge to SED compared with DDFConv. Second, FFDConv outperforms FTDConv, which can prove that time-frequency patterns of sound events are highly frequency-dependent, and this dependency is time-invariant. Moreover, FFDConv models acoustic features with different kernels along the frequency axis, which can be thought to be frequency components modeled in different representation spaces. As if components of the feature are split into different frequency spaces and then reassembled. This is consistent with the characteristics of sound events. Fig. 4: Feature comparison of FFDConv and CRNN. Features activation of the 5th Conv block are shown in the 4th row. The trends of frequency band features over time are shown in the 5th row. Note that y-axis labels of strong prediction are abbreviations of the sound event categories. For example, Abr stands for Alarm bell ringing. ### 3.3 Fine-grained modeling study To explore FFDConv’s ability to understand acoustic spectral information at a fine-grained level. We visualized feature of the middle layer. More visualizations can be found in the supplementary material. The visualization results are shown in Fig. 4. Comparing the features of FFDConv and CRNN, we can see that most of the time-frequency patterns modeled by CRNN are temporally isolated and disjoint. In contrast, FFDConv’s patterns and their neighbors are in a whole, thereby forming a distinct time-frequency representation. Moreover, this phenomenon can also be found in trends of frequency band features over time. The waveforms of FFDConv are smoother than CRNN. Specifically, the duration of peak and trough is longer in FFDConv’s waveform, which results from the feature being mostly coherent over time. There are more pulses in the resting state of CRNN’s waveforms, which are in a disorganized state. Besides, the distributions of frequency band features are consistent with alarm_bell_ring’s spectrogram in FFDConv’s waveforms. The values of the low-frequency band features are smaller than those of the middle and high-frequency bands when the alarm bell rings. However, the differences between frequency bands in CRNN are ambiguous. As for the model’s prediction, the CRNN’s isolated features directly lead to the incoherent output compared with ground truth, which proves that the feature’s coherence over time is essential. It’s interesting that the low-frequency white noise of the sound clip is filtered by FFDConv, but CRNN tagged it as a speech. This has to do with that dynamic convolution concentrates more on high-frequency texture information, and white noise in the spectrogram lacks clear contour information. Actually, most SED models are trained in a frame-based supervised way, which always leads to the feature and output being discrete over time. However, FFDConv can alleviate this by frequency-dependent modeling, which models different patterns for frequency bands, leading to a distinct representation of a sound event. This modeling way is like an attention mechanism in which the distribution of frequency band information of the spectrogram is maintained. Besides, the convolution kernel for a frequency band is shared in all frames, which produces temporally coherent representations. This is consistent with both the continuity of the sound waveform and the vocal continuity of sound events. ### 3.4 Ablation study We compared the performance of different window sizes of the build kernel when generating spatial filters. Note that the size of the spatial filter $K$ is set to 3. Table 2: _Comparison of different window size, W._ Model | $Atten$ | $W$ | PSDS1 | PSDS2 ---|---|---|---|--- | ✗ | 3 | 0.421 | 0.650 | ✓ | 1 | 0.421 | 0.659 FFDConv | ✓ | 3 | 0.436 | 0.685 | ✓ | 5 | 0.423 | 0.656 | ✓ | 7 | 0.432 | 0.666 The results are shown in Table 2. With constraints of the attention module, FFDConv can get better performance. This proves that before attention, spatial filters of different frequency spaces may have a large distance from each other. The performance of FFDConv is the best when window size is set to 3. This is because the adjacent frequency components are considered compared to size 1 when generating the spatial filter, and size 5 may suffer from overfitting. In addition, it’s interesting that the performance recovers when the window size is set to 7. This may have to do with the fact that dynamic convolutions are relatively unstable. ## 4 conclusions In this paper, we proposed full-frequency dynamic convolution, the first full- dynamic method for SED. Full-frequency dynamic convolution is designed to model time-frequency patterns in different frequency spaces. This design in structure physically furnished 2D convolution with capability of frequency- dependent modeling. Experiments on the DESED show that full-frequency dynamic convolution is superior to not only baseline but also other full-dynamic convolutions, which proves FFDConv qualifies in SED. In addition, by visualizing features of sound events, we found that FFDConv can extract temporally coherent features in specific frequency bands, which is consistent with the vocal continuity of sound events. This proves that FFDConv has great frequency-dependent perception ability. In the future, we aim to explore new methods to model vocal continuity of sound events. ## References * [1] J. Rouat, “Computational Auditory Scene Analysis: Principles, Algorithms, and Applications,” IEEE TNN, vol. 19, no. 1, pp. 199–199, Jan. 2008. * [2] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye, “Object Detection in 20 Years: A Survey,” Proc. IEEE, vol. 111, no. 3, pp. 257–276, March. 2023. * [3] Jiaxing Sun, Yujie Li, Huimin Lu, Tohru Kamiya, and Seiichi Serikawa, “Deep Learning for Visual Segmentation: A Review,” in COMPSAC, July. 2020, pp. 1256–1260. * [4] Qin Jin, Peter Schulam, Shourabh Rawat, Susanne Burger, Duo Ding, and Florian Metze, “Event-based video retrieval using audio,” in INTERSPEECH, Sept. 2012, pp. 2085–2088. * [5] Christian Debes, Andreas Merentitis, Sergey Sukhanov, Maria Niessen, Nikolaos Frangiadakis, and Alexander Bauer, “Monitoring Activities of Daily Living in Smart Homes: Understanding human behavior,” IEEE SPM, vol. 33, no. 2, pp. 81–94, March. 2016. * [6] Juan Pablo Bello, Charlie Mydlarz, and Justin Salamon, Sound analysis in smart cities, pp. 373–397, Springer International Publishing, Sept. 2017. * [7] Jie Hu, Li Shen, and Gang Sun, “Squeeze-and-Excitation Networks,” in CVPR, Jun. 2018, pp. 7132–7141. * [8] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang, “Selective Kernel Networks,” in CVPR, Jun. 2019, pp. 510–519. * [9] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon, “CBAM: Convolutional Block Attention Module,” in ECCV, Sept. 2018, pp. 3–19. * [10] Tong Na and Qinyi Zhang, “Convolutional network with conformer for semi-supervised sound event detection,” Tech. Rep., DCASE2021 Challenge, 2021. * [11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool, “Dynamic filter networks,” in NEURIPS, Dec. 2016. * [12] Hyeonuk Nam, Seong-Hu Kim, Byeong-Yun Ko, and Yong-Hwa Park, “Frequency Dynamic Convolution: Frequency-Adaptive Pattern Recognition for Sound Event Detection,” in INTERSPEECH, Sept. 2022, pp. 2763–2767. * [13] Chao Li, Aojun Zhou, and Anbang Yao, “Omni-Dimensional Dynamic Convolution,” in ICLR, Apr. 2022. * [14] Shengchang Xiao, Xueshuai Zhang, and Pengyuan Zhang, “Multi-Dimensional Frequency Dynamic Convolution with Confident Mean Teacher for Sound Event Detection,” in ICASSP, Jun. 2023, pp. 1–5. * [15] Julio Zamora Esquivel, Adan Cruz Vargas, Paulo Lopez Meyer, and Omesh Tickoo, “Adaptive Convolutional Kernels,” in ICCV Workshops, Oct. 2019, pp. 0–0. * [16] Zhi Tian, Chunhua Shen, and Hao Chen, “Conditional Convolutions for Instance Segmentation,” in ECCV, Aug. 2020, pp. 282–298. * [17] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen, “SOLOv2: Dynamic and Fast Instance Segmentation,” in NEURIPS, Dec. 2020, pp. 17721–17732. * [18] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang, “Decoupled Dynamic Filter Networks,” in CVPR, Jun. 2021, pp. 6647–6656. * [19] Emre Çakır, Giambattista Parascandolo, Toni Heittola, Heikki Huttunen, and Tuomas Virtanen, “Convolutional Recurrent Neural Networks for Polyphonic Sound Event Detection,” IEEE/ACM TASLP, vol. 25, no. 6, pp. 1291–1303, Jun. 2017. * [20] Antti Tarvainen and Harri Valpola, “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,” in NEURIPS, Dec. 2017. * [21] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz, “mixup: Beyond Empirical Risk Minimization,” ICLR, Oct. 2018. * [22] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le, “SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition,” in INTERSPEECH, Sept. 2019, pp. 2613–2617. * [23] Hyeonuk Nam, Seong-Hu Kim, and Yong-Hwa Park, “Filteraugment: An Acoustic Environmental Data Augmentation Method,” in ICASSP, May. 2022, pp. 4308–4312.
* Sandage & Tammann (1975) Sandage, A., & Tammann, G.A. 1975, ApJ, 197, 265 * Santiago & Strauss (1992) Santiago, B.X. & Strauss, M.A. 1992, ApJ, 387, 9 * Saslaw (2000) Saslaw, William C. 2000, The Distribution of the Galaxies: Gravitational Clustering in Cosmology, Cambridge University Press: Cambridge * Scargle (1998) Scargle, J. 1998, ApJ, 504, p.405. * Scargle (2002) Scargle, J. 2002 “Bayesian blocks in two or more dimensions: Image segmentation and cluster analysis,” in Bayesian Inference and Maximum Entropy Methods in Science and Engineering, American Institute of Physics Conference Proceedings, Volume 617, pp. 163-173. * Scargle et al. (2008) Scargle, J., Norris, J., & Jackson, B. 2008 “Studies in Astronomical Time Series Analysis. VI. Optimal Segmentation: Blocks, Triggers, and Histograms,” in preparation * Schaap & van de Weygaert (2000) Schaap, W.E. & van de Weygaert, R. 2000, A&A, 363, L29 * Schaap (2007) Schaap, W.E. 2007, “The Delaunay Tessellation Field Estimator”, Ph.D. Thesis, Groningen University * Schlegel et al. (2009) Schlegel, D.J. et al. 2009, arXiv:0904.0468v3 * Scott (1992) Scott, David W. 1992, Multivariate Density Estimation: Theory, Practice and Visualization, John Wiley & Sons, Inc.: New York * Shandarin (1983) Shandarin, S.F. 1983, Sov Astr. Letters, 9, 104 * Shane & Wirtanen (1967) Shane, C.D. & Wirtanen, C.A. 1967, Publ. Lick. Obs. 22, Part 1 * Shapley (1933) Shapley, H. 1933, PNAS, 19, 389 * Shectman et al. (1996) Shectman, S.A. et al. 1996, ApJ, 470, 172 * Sheth et al. (2003) Sheth, J.V., Sahni, V., Shandarin, S.F. & Sathyaprakash, B.S. 2003, MNRAS, 343, 22 * Sheth & Tormen (2004) Sheth, J.V. & Tormen, G. 2004, MNRAS, 350, 1385 * Silverman (1986) Silverman, B.W. 1986, Density Estimation for Statistics and Data Analysis, (Chapman & Hall; reprinted in 1998 by CRC Press: Boca Raton) * Slezak et al. (1990) Slezak, E., Bijaoui, A. & Mars, G. 1990, A&A, 227, 301 * Slezak et al. (1993) Slezak, E., de Lapparent, V. & Bijaoui, A. 1993, ApJ, 409, 517 * Snyder (1991) Snyder, Donald L. & Miller, Michael I. 1991, Random Point Processes in Time and Space, 2nd edition, Springer-Verlag: New York * Sousbie, Colombi & Pichon (2009) Sousbie, T., Columbi, S.& Pichon C. 2009, MNRAS, 393, 457 * Sousbie (2010) Sousbie, T. 2010, submitted to MNRAS, arXiv:1009.4105 * Sousbie, Pichon & Kawahara (2010) Sousbie, T., Pichon, C. & Kawahara, H. 2010, submitted to MNRAS, arXiv:1009.4104 * Stein (1997) Stein, M.L. 1997 in Feigelson E.D., Babu G.J., eds., Statistical Challenges in Modern Astronomy II. Springer-Verlag, New York, p.166 * Stoyan et al. (1985) Stoyan, D., Kendall, W.S. & Mecke, J. 1995, Stochastic Geometry and Its Applications, ed. John Wiley & Sons, Chichester * Springel et al. (2005) Springel, V., et al. 2005, Nature, 435, 629 * Strauss et al. (2002) Strauss, M. A., et al. 2002 AJ, 124, 1810 * Stril et al. (2010) Stril, A., Cahn, R., & Linder E.V. 2010, MNRAS, 404, 239 * Szapudi & Szalay (1998) Szapudi, I & Szalay, A.S. 1998, ApJ, 494, 41 * Totsuji & Kihara (1969) Totsuji, H. & Kihara, T. 1969, PASJ, 21, 221 * Turner & Gott (1976) Turner, E.L. & Gott, J.R. 1976, ApJS, 32, 409 * Turner et al. (1979) Turner, E.L., Aarseth, S.J., Gott, J.R., Blanchard, N.T. & Mathieu, R.D. 1979, ApJ, 228, 684 * Ueda & Itoh (1997) Ueda, H. & Itoh, M. 1997, PASJ, 49, 131 * van de Weygaert (1994) van de Weygaert R. 1994, A&A, 283, 361 * van de Weygaert (2003) van de Weygaert R. 2003, “The Cosmic Foam: Stochastic Geometry and Spatial Clustering across the Universe,” Invited contribution in Proceedings it Statistical Challenges in Modern Astronomy III, eds. E.D. Feigelson & G.J. Babu, Springer-Verlag, pp. 175-196 * van de Weygaert & Schaap (2009) van de Weygaert R. & Schaap, W. 2009, “The Cosmic Web: Geometric Analysis”, in Data Analysis in Cosmology, Lecture Notes in Physics, vol. 665, Eds V.J. Martinez, E. Saar, E. Mart nez-Gonz lez, and M.-J. Pons-Border a. Berlin: Springer, 2009., p.291-413 * van de Weygaert & Aragón-Calvo (2009) van de Weygaert R. & Aragón-Calvo, M. 2003, “Geometry and Morphology of the Cosmic Web: Analyzing Spatial Patterns in the Universe,” Invited review ISVD09 (International Symposium on Voronoi Diagrams and Engineering), Copenhagen, Denmark. IEEE CPS, E3781, ed. F. Anton. * Villmann et al. (1997) Villmann, T., Der, R., Herrmann, M., & Martinetz, T., 1997, “Topology Preservation in Self-Organizing Feature Maps: Exact Definition and Measurement.” _IEEE Transactions on Neural Networks_ , 8, pp. 256-266. * Vogeley et al. (1994) Vogeley, M.S., et al. 1994, ApJ, 420, 525 * Wright (1750) Wright, T. 1750, “An Original Theory or New Hypothesis of the Universe” New York: Elsevier 1971 * York et al. (2000) York, D.G. et al. 2000, AJ, 120, 1579 * Yoshioka & Ikeuchi (1989) Yoshioka, S. & Ikeuchi, S. 1989, ApJ, 341, 16 * Zehavi et al. (2002) Zehavi, I. et al. 2002, ApJ, 571, 172 * Zehavi et al. (2010) Zehavi, I. et al. 2010, arXiv:1005.2413 * Zeldovich (1970) Zel‘dovich, Ya. B. 1970, A&A5, 84 * Zeldovich et al. (1982) Zel‘dovich, Ya. B., Einasto, J., & Shandarin, S.F. 1982, Nature, 300, 407 * Zhang et al. (2010) Zhang, Y., Springel, V. & Yang, X. 2010, arXiv:1006.3768 * Zwicky (1957) Zwicky, F. 1957, PASP, 69, 518 * Zwicky et al. (1961) Zwicky, F., Wield, P., Herzog, E., Karpowicz, M., & Kowal, C.T. 1961-68, Catalogue of Galaxies and Clusters of Galaxies, 6 volumes. Pasadena, California Institute of Technology
Tissue evolution: Mechanical interplay of adhesion, pressure, and heterogeneity Tobias Büscher1, Nirmalendu Ganai1,2, Gerhard Gompper1, Jens Elgeti1*, 1 Theoretical Soft Matter and Biophysics, Institute of Complex Systems and Institute for Advanced Simulation, Forschungszentrum Jülich, 52425 Jülich, Germany 2 Department of Physics, Nabadwip Vidyasagar College, Nabadwip, Nadia 741302, India *<EMAIL_ADDRESS> ## Abstract The evolution of various competing cell types in tissues, and the resulting persistent tissue population, is studied numerically and analytically in a particle-based model of active tissues. Mutations change the properties of cells in various ways, including their mechanical properties. Each mutation results in an advantage or disadvantage to grow in the competition between different cell types. While changes in signaling processes and biochemistry play an important role, we focus on changes in the mechanical properties by studying the result of variation of growth force and adhesive cross- interactions between cell types. For independent mutations of growth force and adhesion strength, the tissue evolves towards cell types with high growth force and low internal adhesion strength, as both increase the homeostatic pressure. Motivated by biological evidence, we postulate a coupling between both parameters, such that an increased growth force comes at the cost of a higher internal adhesion strength or vice versa. This tradeoff controls the evolution of the tissue, ranging from unidirectional evolution to very heterogeneous and dynamic populations. The special case of two competing cell types reveals three distinct parameter regimes: Two in which one cell type outcompetes the other, and one in which both cell types coexist in a highly mixed state. Interestingly, a single mutated cell alone suffices to reach the mixed state, while a finite mutation rate affects the results only weakly. Finally, the coupling between changes in growth force and adhesion strength reveals a mechanical explanation for the evolution towards intra-tumor heterogeneity, in which multiple species coexist even under a constant evolutianary pressure. ## Introduction Mutations change the cell fitness and thus its chance to survive and proliferate [1]. Advantageous mutations are more likely to persist due to natural selection, which drives the evolution of a tissue towards fitter cells [2]. Cancer represents an example of evolution on a short time scale [3]. Furthermore, cancer is a multistep process, i.e. several mutations are needed for a tumor in order to develop and become malignant [4]. Hence, tumorigenesis might be expected to happen in a serial manner, i.e. a cell acquiring a ”beneficial” mutation and taking over the whole tissue. After some time, a daughter cell acquires another mutation and again takes over. Interestingly, however, tumors do not consist of a single cell type, but instead several subpopulations coexist within the same tumor. This is called intra-tumor heterogeneity [5]. Each mutation changes certain biochemical properties of a cell. This ranges from misfunction in the error correction machinery during DNA replication and disruptions in signaling pathways to epigenetic changes in the expression level of certain proteins [1, 6, 7]. All these changes can also affect the mechanical properties of the mutated cell, e.g. mutated cells which express less adhesion proteins might be able to detach from the primary tumor more easily [8], necessary to form metastases. On the other hand, mechanics feeds back onto growth in several ways, e.g. increased apoptosis rate due to mechanical stresses [9, 10] or dependence of the growth of tissue spheroids on the properties of the surrounding medium [11, 12, 13]. It is the mechanical contribution to tissue development that we want to focus on in this work. For mechanically regulated growth, homeostatic pressure plays an important role [14]. In the homeostatic state, when apoptosis and division balance each other, a tissue exerts a certain pressure onto its surrounding, the homeostatic pressure $P_{\text{H}}$. The tissue is able to grow as long as the external pressure $P$ is smaller than $P_{\text{H}}$. For the competition between different tissues for space, it has been suggested that the tissue with the higher homeostatic pressure grows at the expense of the weaker tissue. Several theoretical studies employ this concept in order to describe interface propagation between two competing tissues [15, 16, 17]. A metastasis would need to reach a critical size, below which the additional Laplace pressure due to surface tension would cause the metastasis to shrink and disappear [14]. However, reduced adhesion between tissues, which increases surface tension, leads to an enhanced growth rate at the interface between them, stabilizing coexistence even for differing homeostatic pressures [18]. In this work, we study the influence of mutations that change the mechanical properties of cells on the competition dynamics, especially the interplay between changes in the adhesive properties and the strength with which a cell pushes onto its surrounding. Particularly interesting is the case where loss of adhesion comes at the cost of lower growth strength. This is motivated by the observed down-regulation of E-cadherin, an adhesion protein in epithelia, in many types of cancer [19]. Interestingly, E-cadherin is also involved in signaling processes connected to cell growth [20]. We find that in this case several cell types with different mechanical properties can coexist and that the cell type with the highest homeostatic pressure does not necessarily dominate the competition. ## Results Several models have been developed previously in order to study tissue growth [21], in combination with different simulation techniques, including vertex [22, 23] and particle-based [24, 25] models as well as Cellular Potts models [26, 27]. We employ the two particle growth (2PG) model of Refs. [28, 29, 18]. A cell is described by two particles which repel each other via a growth force $\textbf{{F}}_{ij}^{\text{G}}=\frac{G}{(r_{ij}+r_{0})^{2}}\hat{\textbf{{r}}}_{ij}\text{,}$ (1) with strength $G$, unit vector $\hat{\textbf{{r}}}_{ij}$, distance $r_{ij}$ between the two particles and a constant $r_{0}$. Different cells interact via a soft repulsive force $\textbf{{F}}_{ij}^{\text{V}}$ on short distances, maintaining an excluded volume, and a constant attractive force $\textbf{{F}}_{ij}^{\text{A}}$ on intermediate distances, modeling cell-cell adhesion, with $\left.\begin{array}[]{@{}ll@{}}\textbf{{F}}_{ij}^{\text{V}}&=f_{0}\left(\frac{R_{\text{PP}}^{5}}{r_{ij}^{5}}-1\right)\hat{\textbf{{r}}}_{ij}\\\ \textbf{{F}}_{ij}^{\text{A}}&=-f_{1}\hat{\textbf{{r}}}_{ij}\par\end{array}\right\\}\text{for }r_{ij}<R_{\text{PP}}\text{,}$ (2) with exclusion coefficient $f_{0}$, adhesion strength coefficient $f_{1}$, and cut-off length $R_{\text{PP}}$. A cell divides when the distance between its two particles reaches a size threshold $r_{\text{ct}}$. A new particle is then placed close (randomly within a short distance $r_{\text{d}}$) to each of the two particles of the divided cell. Each of these pairs then constitutes a new cell. Apoptosis is modeled by removing cells randomly at a constant rate $k_{\text{a}}$. We employ a dissipative particle dynamics-type thermostat, with an effective temperature $T$, to account for energy dissipation and random fluctuations. We choose the value of $T$ such that cells can escape local minima, but other thermal effects are negligible. Note that all parameters can be set individually for each cell type as well as between different cell types for inter-cell interactions. We only vary the growth-force strength $G^{\alpha}$ and adhesion strength $f_{1}^{\alpha\beta}$ between cells of the same ($\alpha=\beta$) and different ($\alpha\neq\beta$) cell types, respectively, where $\alpha$ and $\beta$ are cell-type numbers. We report simulation parameters relative to a standard host cell type (see Materials and methods for numerical values), denoted with a dagger, e.g. $G^{\dagger}=G/G^{0}$. Time is measured in terms of the inverse apoptosis rate $k_{\text{a}}$, distance in units of the pair potential cut-off length $R_{\text{PP}}$ and stresses in units of $G^{0}/R_{\text{PP}}^{4}$. Quantities reported in these units are denoted by an asterisk ∗. All simulations are performed in a cubic box with edge length $L=12\cdot R_{\text{PP}}$ and periodic boundary conditions in all directions, unless stated otherwise. Fig 1: Evolution of a tissue with mutations altering growth-force strengh $G^{\dagger}$ and adhesion strength $f_{1}^{\dagger}$ independently. Heatmaps displaying cell-number fractions $\phi_{\alpha}$ after a) zero generations (initial condition), b) 50 generations, c) 100 generations and d) 125 generations. Tumor cells even within the same tumor are not all identical, but vary in terms of all kind of attributes, e.g. expression levels of different proteins [30] or their reaction to certain treatments [31]. Hence, there is not only a competition between the tumor and the host, but also between cell- subpopulations of the tumor. Different models exist to describe tumor heterogeneity, e.g. cancer stem cells [32] or clonal evolution [33]. In the latter case, a tumor originates from a single mutated cell, which can acquire additional mutations over time, yielding additional subpopulations. We model this behaviour by defining a fixed number $n$ of different ”genotypes”, each having a different growth-force strength $G^{\alpha}$ and adhesion strength $f_{1}^{\alpha\alpha}$. Mutations are implemented by offering each daughter cell after a division event the chance to change its genotype with a certain probability. In tissues, several adhesion mechanisms exist, serving a variety of different functions to maintain tissue integrity. Between epithelial cells, the strength of cell-cell adhesion is to a large degree regulated by anchoring junctions, e.g. adherens junctions, which connect the actin cytosceletons of neighbouring cells. Adherens junctions are mediated by cadherins, which form homophilic bonds between cells. Thus, the strength of adhesion between cells is limited by the cell expressing less cadherin, or, in terms of our simulation model $f_{1}^{\alpha\beta}=\min(f_{1}^{\alpha\alpha},f_{1}^{\beta\beta})$. A reduced adhesion strength yields a higher homeostatic pressure [29], which is otherwise dominated by the growth-force strength $G$. For free parameter evolution, the tissue thus evolves to a strong-growing and low-adhesive genotype (see Fig 1), as predicted by the homeostatic pressure approach [14]. Fig 2: Time evolution of the cell-number fractions $\phi_{\alpha}$ of each genotype for tradeoff paramter a) $\tau=0$, b) $\tau=1$, c) $\tau=2$ and d) $\tau\rightarrow\infty$, $d\rightarrow 0$. Simulations start from a host (standard) tissue at homeostasis, with $n=21$ genotypes, $p_{\text{m}}=0.01$ in all and $d=0.025$ in a)-c). White space corresponds to times where no cells of the genotype exist. Color is coded on a logarithmic scale. Curves above display homeostatic pressure $P_{\text{H}}^{\alpha*}$ (black solid), growth- force strength $G^{\alpha\dagger}$(red dashed) and self adhesion strength $f_{1}^{\alpha\alpha\dagger}$ (green dotted) of the corresponding genotype. However, E-cadherin also plays a role in signaling processes connected to cell growth, and thus a reduced expression might come at the cost of a lower growth-force strength $G$, which in turn yields a lower homeostatic pressure. We thus turn our attention to the case where an increase in growth-force strength $G^{\alpha}$ comes at the cost of a higher self-adhesion strength $f_{1}^{\alpha\alpha}$. We assume the relations as $\displaystyle G^{\alpha}$ $\displaystyle=(1+D^{\alpha})G^{0}$ (3) $\displaystyle f_{1}^{\alpha\alpha}$ $\displaystyle=(1+D^{\alpha}\cdot\tau)f_{1}^{0}\text{,}$ (4) with genotype number $\alpha$ in the range $[-(n-1)/2,(n-1)/2]$, evolutionary distance $D^{\alpha}=d\cdot\alpha$, distance $d$ between neighbouring genotypes and tradeoff paramteter $\tau$ (with $G^{\alpha},f_{1}^{\alpha\alpha}>0\ \forall\ \alpha$). After a division event, each daughter cell might mutate into a new genotype with probability $p_{\text{m}}$. If the cell mutates, its genotype number is changed to $\alpha_{\text{mother}}~{}\pm 1$ randomly. This yields a mutation rate $k_{\text{m}}=2p_{\text{m}}k_{\text{a}}$. Figure 2 displays results of such simulations for four different cases: only variation of growth-force strength ($\tau=0$), balanced tradeoff ($\tau=1$), adhesion strength varied twice as much as growth-force strength ($\tau=2$) and only variation of adhesion strength ($\tau\rightarrow\infty$). Without tradeoff (Fig 2a)), the tissue evolves towards the strongest growing genotype or, equivalently, the one with the highest homeostatic pressure. Similarly, for $\tau\rightarrow\infty$ (Fig 2d)), the system evolves towards the lowest adhesive genotype (again, the one with the highest $P_{\text{H}}$). We find the most dynamic evolution for a balanced tradeoff (Fig LABEL:snapshot and 2b)). At first, the system evolves to stronger growing and more adhesive genotypes. Over time a noticable fraction of cells evolves also towards weak- growing, less adhesive genotypes. The cell-number fractions $\phi_{\alpha}=N_{\alpha}/N$ (with individual and total number of cells, $N_{\alpha}$ and $N$), show large fluctuations (see Fig LABEL:snapshotb) and c)), with individual genotypes not being populated at all for certain time periods. Besides this highly dynamic temporal evolution, after an initial time period the system is dominated by genotypes with increased growth force and adhesion strength at all times, with the one at the upper boundary having the highest cell-number fraction for most of the time (see Fig LABEL:snapshota)). This result comes at a surprise, as this is also the genotype with the lowest homeostatic pressure, while the one at the lower boundary, which is basically never populated, has the highest $P_{\text{H}}$. For a higher tradeoff (Fig 2c)), we still find a broad distribution of genotypes, with less adhesive genotypes dominating over the stronger growing ones, i.e. the loss in growth- force strength is overcompensated by a lower adhesion strength. In order to gain insight into the underlying mechanism of this dynamic evolution, we study the competition between two genotypes and no mutations ($p_{\text{m}}=0$). Simulations are started from a single mutated cell (with increased/decreased growth force and adhesion strength) in a host tissue at the homeostatic state (we label the mutant with M and the host (wild type) with W). Even in this simplified case, we find one parameter regime in which the mutant is not able to grow, one regime with stable coexistence in a highly mixed state and another regime in which the mutant outcompetes the host. Figure LABEL:number_fractions shows the averaged number fractions of the mutant at the steady state. For reduced growth force and adhesion strength (Fig LABEL:number_fractionsa)), the mutant can only grow against the host if its adhesion strength is reduced below a critical $f_{1}^{\text{crit}}$. In terms of Eq. (4), the value of $f_{1}^{\text{crit}}$ roughly corresponds to a balanced tradeoff ($\tau\approx 1$). Already for $f_{1}^{\text{MM}}>f_{1}^{\text{crit}}$, the homeostatic pressure of the mutant exceeds the one of the host, i.e. a parameter regime exists in which the mutant is not able to grow, despite of the higher $P_{\text{H}}$. The reverse happens when growth force and adhesion strength are increased. The mutant completely takes over the compartment, although its homeostatic pressure is smaller than that of the host. Again, coexistence is only found when the adhesion strength is increased above $f_{1}^{\text{crit}}$. In the coexistence regime, the mutant number fraction scales as $\phi^{\text{M}}\propto 1/(f_{1}^{\text{MM}}-f_{1}^{\text{WW}})$. Altogether, the competition between two genotypes alone yields the same qualitative results as the more complex multi-genotype case discussed before. Still, the question remains how a genotype with lower homeostatic pressure can outcompete a stronger genotype. The answer can only lie in the adhesion strength $f_{1}^{\text{MW}}=\min(f_{1}^{\text{MM}},f_{1}^{\text{WW}})$ between mutant and host cells. This choice of cross-adhesion strength breaks symmetry, as the stronger adhering genotype has more free space at the interface, which favors divisions [18]. To address this question, we develop a phenomenological model which incorporates pressure-dependent growth as well as interfacial effects, in order to obtain a qualitative explanation of the simulation results. We start with the expansion of the bulk growth rate $k_{\text{b}}$ around the homeostatic pressure, $k_{\text{b}}=\kappa(P-P_{\text{H}})\text{,}$ (5) with the pressure response coefficient $\kappa$. Due to the high degree of mixing, the number fractions $\phi^{\text{M/W}}$ and hence the strengh of interfacial effects vary locally. In a mean-field approximation, we take the interfacial effects to be proportional to $\phi^{\text{M}}(1-\phi^{\text{M}})$, with individual prefactors $\Delta k_{\text{s}}^{\text{M/W}}$ for each genotype. The time evolution is then given by $\displaystyle\partial_{t}\phi^{\text{M}}=$ $\displaystyle\kappa(P_{\text{H}}^{\text{M}}-P)\phi^{\text{M}}+\Delta k_{\text{s}}^{\text{M}}\phi^{\text{M}}(1-\phi^{\text{M}})$ (6) $\displaystyle\partial_{t}(1-\phi^{\text{M}})=$ $\displaystyle\kappa(P_{\text{H}}^{\text{M}}+\Delta P_{\text{H}}-P)(1-\phi)+\Delta k_{\text{s}}^{\text{W}}\phi^{\text{M}}(1-\phi^{\text{M}})\text{,}$ (7) with the difference in homeostatic pressure $\Delta P_{\text{H}}=P_{\text{H}}^{\text{W}}-P_{\text{H}}^{\text{M}}$. Addition of Eqs. (6) and (7) yields the pressure $\displaystyle P=P_{\text{H}}^{\text{W}}-\Delta P_{\text{H}}\phi^{\text{M}}+\frac{\Delta k_{\text{s}}^{\text{M}}+\Delta k_{\text{s}}^{\text{W}}}{\kappa}\phi^{\text{M}}(1-\phi^{\text{M}})\text{.}$ (8) Thus, the pressure is given by the homeostatic pressures of the two genotypes weighted by their number fraction plus an interfacial term. A figure displaying the pressure measured during the simulations shown in Fig LABEL:number_fractions can be found in the S1 Appendix.. Insertion of Eq. (8) into Eq. (6) yields a differential equation for the number fraction with three fixed points ($\partial_{t}\phi^{\text{M}}=0$), $\phi_{1}^{\text{M}}=0$, $\phi_{2}^{\text{M}}=1$, and $\displaystyle\phi_{3}^{\text{M}}=\frac{-\kappa\Delta P_{\text{H}}+\Delta k_{\text{s}}^{\text{M}}}{\Delta k_{\text{s}}^{\text{M}}+\Delta k_{\text{s}}^{\text{W}}}\text{.}$ (9) We discuss this result for the case of reduced growth force and adhesion strength of the mutant. $\Delta k_{\text{s}}^{\text{M}}$ might be expected to vanish, as $f_{1}^{\text{MM}}=f_{1}^{\text{MW}}$ and mutant cells thus would not feel whether neighbouring cells are mutant or host cells. However, in order to grow, a cell needs to impose a strain on its surrounding. Host cells adhere more strongly to each other, thus it is harder for a mutant cell to impose a strain when surrounded by host cells. Hence, $\Delta k_{\text{s}}^{\text{M}}$ is actually negative and the homeostatic pressure of the mutant needs to exceed the host pressure by $-\Delta k_{\text{s}}^{\text{M}}/\kappa$ in order to be able to grow against the host. At this point, $\phi_{3}^{\text{M}}$ becomes positive, as long as $\Delta k_{\text{s}}^{\text{M}}+\Delta k_{\text{s}}^{\text{W}}>0$. Host cells can impose a strain more easily when surrounded by mutant cells and, additionally, have more free space than when surrounded by other host cells. Hence, $|\Delta k_{\text{s}}^{\text{M}}|<\Delta k_{\text{s}}^{\text{W}}$ and the above mentioned condition is fulfilled. Similarly, coexistence can be found for increased growth force and adhesion strength when $\Delta P_{\text{H}}>-\Delta k_{\text{s}}^{\text{W}}/\kappa$. The above mentioned scaling of the mutant number fraction can be obtained by an expansion of $\Delta P_{\text{H}}$ and $\Delta k_{\text{s}}^{\text{M/W}}$ to linear order in terms of $\epsilon:=(f_{1}^{\text{MM}}-f_{1}^{\text{WW}})/f_{1}^{\text{WW}}$ in Eq. (9), $\displaystyle\phi_{3}^{\text{M}}=\frac{-\kappa\Delta P_{\text{H}}^{0}}{(\Delta k_{\text{s}}^{\text{M1}}+\Delta k_{\text{s}}^{\text{W1}})\epsilon}+\frac{-\kappa\Delta P_{\text{H}}^{1}+\Delta k_{\text{s}}^{\text{M1}}}{\Delta k_{\text{s}}^{\text{M1}}+\Delta k_{\text{s}}^{\text{W1}}}\text{.}$ (10) The zeroth order terms of $\Delta k_{\text{s}}^{\text{M/W}}$ vanish as there are no interfacial effects when the adhesion strength between host and mutant cells is equal to their self-adhesion strength, while $\Delta P_{\text{H}}^{0}$ can be non-zero due to a changed growth-force strength. Indeed, Eq. (10) reproduces the simulation data reasonably well (see Fig LABEL:number_fractions). A discussion of the numerical values of the fitted parameters and additional results can be found in S1 Appendix.. Figure LABEL:slope displays similar results as shown in Fig LABEL:number_fractions, but now as a function of the tradeoff $\tau$ in Eq. (4). For $\tau<1$ the genotype with higher growth-force strength outcompetes the weaker genotype, for $1<\tau<2$ a transition towards the less adhesive genotype occurs, while for even higher values of the tradeoff $\tau>2$ the less adhesive genotype outcompetes the second genotype. This transition from strongly growing, adhesive to weakly growing, less adhesive genotypes is found in the same range of $\tau$ as in the competition between many genotypes. Hence, the simplified case of two competing genotypes captures the essential physics to explain the coexistence between many competing genotypes and, additionally, provides a quantitative description. Next, we turn our attention to the effect of a finite mutation rate on the evolution of the system. Figure LABEL:mutationrate_clustera) shows the number fraction of the mutant as a function of $k_{\text{m}}$ for different combinations of evolutionary distance $D^{\alpha}$ and tradeoff $\tau$, in comparision to the number fraction reached for a single mutation event. As expexted, the number fraction converges towards $1/2$ with increasing $k_{\text{m}}$ for all combinations. For moderate mutation rates, however, the number fraction largely fluctuates around the same average as of a single mutation event. The single mutation leads to a stable coexistence of the two genotypes - additional mutations quickly relax back to this state. Siginificant deviations occur only if in the steady state of the single mutation event the number fraction of the weaker genotype is close to zero. In that case, the weaker genotype consists only of one or very few small cohesive clusters of cells, because cells of the weaker genotype need to detach from the primary cluster in order to form new clusters, but are likely to die when they do so, as they are only surrounded by cells of the stronger genotype. Hence, the distribution of cells is highly non-homogenous. Compared to the single mutation event, even a small mutation rate leads to the formation of multiple small cluster all over the system, thus increasing the number fraction of the weaker genotype (see Fig LABEL:mutationrate_clusterb) for comparision in terms of number of clusters and Materials and methods for further discussion). This result explains why at least two genotypes, in addition to the dominating genotype, are populated as well in the cases shown in Fig 2a) and d). When the number fractions of both genotypes are sufficiently large (for $1\leq\tau\leq 2$), deviations from the average of a single mutation are still small for the standard mutation probability. Additionally, in the competitions between many genotypes, mutations change the genotype to $\alpha\pm 1$ randomly and not in a preferred direction. Hence, we conclude that the precise value of the mutation probability does not play an important role in the regime where we find a heterogeneous distribution of genotypes, as long as it is reasonably small ($k_{\text{m}}\ll k_{\text{a}}$). Given that a single mutated cell can grow to tissue of macroscopic size in a certain parameter regime for $f_{1}^{\text{MW}}=\min(f_{1}^{\text{MM}},f_{1}^{\text{WW}})$, the question arises how likely it is to actually reach this state. In order to study this probability, we mutate again a single cell in a host tissue at its homeostatic state. A mutation that reaches a certain threshold $N_{\text{t}}=20$ of cells counts as a survival event (the chance to die after reaching this treshold becomes extremely small), apoptosis of the last mutant cell as a death event. Figure LABEL:survival shows the averages of many such simulations. For reduced growth force and adhesion strength, the survival probability $p_{\text{s}}$ is only non-zero below the critical adhesion strength $f_{1}^{\text{crit}}$. For $f_{1}^{\text{MM}}<f_{1}^{\text{crit}}$, $p_{\text{s}}$ increases linearly with further decreasing adhesion strength. On the other hand, when growth force and adhesion strength are increased, the survival probability first shows a plateau, whose value increases with increasing growth force strength, from which it will probably drop to zero with further increase. Simulations in this regime are difficult, because a mutated cell can easily grow to a few cells, but will hardly reach the number threshold nor completely vanish again. Due to the high self-adhesion strength on the one hand, it becomes hard to detach from the other cells, but on the other hand easy to grow against the host when only few or no other mutant cells are around. This explains the larger error bars at the highest values of the adhesion strength, where the sample size is small. ## Discussion We have shown how intra-tumor heterogeneity, the existence of multiple subpopulations within the same tumor, can arise due to mechanical interactions alone. The simultaneous change of the adhesion and growth-force strength stabilizes the coexistence of multiple subpopulations, in a highly dynamic state. A higher growth-force strength alone, as well as a lower adhesion strength, favor proliferation of a single subpopulation and the evolution of the system to cell types with the highest growth-force strength, or lowest adhesion strength, respectively. A tradeoff between the two, however, yields coexistence between multiple subpopulations of different cell types. Interestingly, the expression of the adhesion protein E-cadherin, which also affects cell growth, has been found to be down-regulated in many real tumors [19]. The simulations also reveal that the homeostatic pressure of a cell type is not necessarily the only quantity that determines the result of a competition. Interactions between different cell types, in our model determined by the adhesion between them, can lead to a completely reverse outcome, i.e. a cell type with lower homeostatic pressure can outcompete a stronger one completely. A phenomenological model explains the results on a qualitative level. The evolution of each cell type is governed by mechanically-regulated growth, while mutation rates only play a minor role in the dynamics. An interesting future aspect to be studied is the influence of open boundaries. A tissue with a negative homeostatic pressure then naturally grows to a spheroid of finite size, with an enhanced rate of division at the surface [29]. For competing cell types, this would lead to an interplay between surface and interfacial effects. ## Materials and methods ### Standard (host) tissue and simulation parameters We define a set of reference simulation parameters, which we refer to as host parameters. Table 1 shows the values in simulation units. In simulations we keep the host W fixed and vary the parameters of the mutant M around the values of the host. Table 1: Simulation parameters and measured properties of the standard (host) tissue . Parameter Symbol Value Time Step $\Delta t$ $10^{-3}$ Pair potential interaction range $R_{\text{PP}}$ $1$ Cellular expansion pressure constant $r_{0}$ 1 Cell division distance treshold $r_{\text{ct}}$ $0.8$ New cell particle initial distance $r_{\text{d}}$ 0.00001 Growth-force strength $G$ $40$ Mass $m$ 1 Intracell dissipation coefficient $\gamma_{\text{c}}$ $100$ Intercell dissipation coefficient $\gamma_{\text{t}}$ $50$ Background dissipation coefficient $\gamma_{\text{b}}$ $0.1$ Apoptosis rate $k_{\text{a}}$ 0.01 Mutation propability $p_{\text{m}}$ 0.01 Noise intensity $k_{\text{B}}T$ 0.1 Repulsive cell-cell potential coefficient $f_{0}$ $2.39566$ Attractive cell-cell potential coefficient $f_{1}$ $6.0$ Isothermal compressibility $\beta_{\text{T}}$ 1 Relaxation time constant $t_{\text{P}}$ $1$ Homeostatic pressure $P_{\text{H}}^{*}$ $0.1321\pm 0.0005\text{\,}$ Pressure response coefficient $\kappa^{*}$ $2.676\pm 0.080\text{\,}$ ### Cluster analysis As explained in the results section, a constant rate of mutation leads to an enhanced formation of clusters when the weaker genotype is barely able to grow against the stronger genotype and consists of only one or few clusters for a single mutation event. We define a cluster as all cells of the same genotype that are in interaction range to at least one other member of the cluster (DBSCAN clustering algorithm with number of minimal points equal to one). Figure LABEL:mutationrate_clusterb) displays the number of clusters of the weaker genotype in the competitions displayed in Fig LABEL:mutationrate_clustera), in comparison to the result of a single mutation event. Indeed, when the number fraction of the weaker genotype is small for a the single mutation event ($\tau=1$), we find significant deviations even for small mutation rates. In this case, the number of clusters first strongly increases with mutation rate, with roughly a tenfold increase at the peak. For even higher mutation probability, the number of clusters decreases again, due to merging of clusters, finally leading to percolation. ## Supporting information #### S1 Appendix. Additional Results. Appendix containing additional results and the corresponding figures. ## Acknowledgments The authors gratefully acknowledge the computing time granted through JARA-HPC on the supercomputer JURECA [35] at Forschungszentrum Jülich. ## References * 1. Weinberg RA. The biology of cancer. Garland Publishing, Inc.; 2007. * 2. Greaves M, Maley CC. Clonal evolution in cancer. Nature 2012;481(7381):306. * 3. Bozic I, Reiter JG, Allen B, Antal T, Chatterjee K, Shah P, et al. Evolutionary dynamics of cancer in response to targeted combination therapy. eLife. 2013;2:e00747. doi:10.7554/eLife.00747. * 4. Vogelstein B, Kinzler KW. The multistep nature of cancer. Trends Genet. 1993;9(4):138 – 141. doi:https://doi.org/10.1016/0168-9525(93)90209-Z. * 5. Heppner GH. Tumor Heterogeneity. Cancer Res. 1984;44(6):2259–2265. * 6. Preston BD, Albertson TM, Herr AJ. DNA replication fidelity and cancer. Semin Cancer Biol. 2010;20(5):281 – 293. doi:https://doi.org/10.1016/j.semcancer.2010.10.009. * 7. Schnekenburger M, Diederich M. Epigenetics Offer New Horizons for Colorectal Cancer Prevention. Curr Colorectal Cancer Rep. 2012;8(1):66–81. doi:10.1007/s11888-011-0116-z. * 8. Petrova YI, Schecterson L, Gumbiner BM. Roles for E-cadherin cell surface regulation in cancer. Mol Biol Cell. 2016;27(21):3233–3244. doi:10.1091/mbc.E16-01-0058. * 9. Wernig F, Xu Q. Mechanical stress-induced apoptosis in the cardiovascular system. Prog Biophys Mol Bio. 2002;78(2):105 – 137. doi:https://doi.org/10.1016/S0079-6107(02)00008-1. * 10. Cheng G, Tse J, Jain RK, Munn LL. Micro-environmental mechanical stress controls tumor spheroid size and morphology by suppressing proliferation and inducing apoptosis in cancer cells. PLoS One. 2009;4(2):e4632. doi:10.1371/journal.pone.0004632. * 11. Montel F, Delarue M, Elgeti J, Malaquin L, Basan M, Risler T, et al. Stress Clamp Experiments on Multicellular Tumor Spheroids. Phys Rev Lett. 2011;107(18):188102. doi:10.1103/PhysRevLett.107.188102. * 12. Alessandri K, Sarangi BR, Gurchenkov VV, Sinha B, Kiesling TR, Fetler L, et al. Cellular capsules as a tool for multicellular spheroid production and for investigating the mechanics of tumor progression in vitro. Proc Natl Acad Sci USA. 2013;110(37):14843–14848. doi:10.1073/pnas.1309482110. * 13. Helmlinger G, Netti PA, Lichtenbeld HC, Melder RJ, Jain RK. Solid stress inhibits the growth of multicellular tumor spheroids. Nat Biotechnol. 1997;15(8):778–783. doi:10.1038/nbt0897-778. * 14. Basan M, Risler T, Joanny JF, Sastre-Garau X, Prost J. Homeostatic competition drives tumor growth and metastasis nucleation. HFSP J. 2009;3(4):265–272. doi:10.2976/1.3086732. * 15. Williamson JJ, Salbreux G. Stability and Roughness of Interfaces in Mechanically Regulated Tissues. Phys Rev Lett. 2018;121:238102. doi:10.1103/PhysRevLett.121.238102. * 16. Ranft J, Aliee M, Prost J, Jülicher F, Joanny JF. Mechanically driven interface propagation in biological tissues. New J Phys. 2014;16(3):035002. * 17. Podewitz N, Jülicher F, Gompper G, Elgeti J. Interface dynamics of competing tissues. New J Phys. 2016;18(8):083020. * 18. Ganai N, Büscher T, Gompper G, Elgeti J. Mechanics of tissue competition: interfaces stabilize coexistence. New J Phys. 2019;21(6):063017. doi:10.1088/1367-2630/ab2475. * 19. Beavon IRG. The E-cadherin–catenin complex in tumour metastasis: structure, function and regulation. Eur J Cancer. 2000;36(13):1607 – 1620. doi:https://doi.org/10.1016/S0959-8049(00)00158-1. * 20. Pece S, Gutkind JS. Signaling from E-cadherins to the MAPK Pathway by the Recruitment and Activation of Epidermal Growth Factor Receptors upon Cell-Cell Contact Formation. J Biol Chem. 2000;275(52):41227–41233. doi:10.1074/jbc.M006578200. * 21. Van Liedekerke P, Palm MM, Jagiella N, Drasdo D. Simulating tissue mechanics with agent-based models: concepts, perspectives and some novel results. Comp Part Mech. 2015;2(4):401–444. doi:10.1007/s40571-015-0082-3. * 22. Farhadifar R, Röper JC, Aigouy B, Eaton S, Jülicher F. The Influence of Cell Mechanics, Cell-Cell Interactions, and Proliferation on Epithelial Packing. Curr Biol. 2007;17(24):2095 – 2104. doi:https://doi.org/10.1016/j.cub.2007.11.049. * 23. Alt S, Ganguly P, Salbreux G. Vertex models: From cell mechanics to tissue morphogenesis. Phil Trans R Soc B. 2017;372:20150520. doi:10.1098/rstb.2015.0520. * 24. Drasdo D, Höhme S. A single-cell-based model of tumor growth in vitro: monolayers and spheroids. Phys Biol. 2005;2(3):133–147. doi:10.1088/1478-3975/2/3/001. * 25. Schaller G, Meyer-Hermann M. Multicellular tumor spheroid in an off-lattice Voronoi-Delaunay cell model. Phys Rev E. 2005;71(5 Pt 1):051910. * 26. Graner F, Glazier JA. Simulation of biological cell sorting using a two-dimensional extended Potts model. Phys Rev Lett. 1992;69:2013–2016. doi:10.1103/PhysRevLett.69.2013. * 27. Szabó A, Merks RM. Cellular Potts Modeling of Tumor Growth, Tumor Invasion, and Tumor Evolution. Front Oncol. 2013;3:87. doi:10.3389/fonc.2013.00087. * 28. Basan M, Prost J, Joanny JF, Elgeti J. Dissipative particle dynamics simulations for biological tissues: rheology and competition. Phys Biol. 2011;8(2):026014. doi:10.1088/1478-3975/8/2/026014. * 29. Podewitz N, Delarue M, Elgeti J. Tissue homeostasis: A tensile state. EPL. 2015;109(5):58005. * 30. Marusyk A, Almendro V, Polyak K. Intra-tumour heterogeneity: a looking glass for cancer? Nat Rev Cancer. 2012;12:323–334. * 31. Marusyk A, Polyak K. Tumor heterogeneity: Causes and consequences. Biochim Biophys Acta. 2010;1805(1):105 – 117. doi:https://doi.org/10.1016/j.bbcan.2009.11.002. * 32. Shackleton M, Quintana E, Fearon ER, Morrison SJ. Heterogeneity in Cancer: Cancer Stem Cells versus Clonal Evolution. Cell. 2009;138(5):822 – 829. doi:https://doi.org/10.1016/j.cell.2009.08.017. * 33. Nowell P. The clonal evolution of tumor cell populations. Science. 1976;194(4260):23–28. doi:10.1126/science.959840. * 34. Allen MP, Tildesley DJ. Computer simulation of liquids. Oxford: Clarendon Press; 1989. * 35. Jülich Supercomputing Centre. JURECA: Modular supercomputer at Jülich Supercomputing Centre. Journal of large-scale research facilities. 2018;4(A132). doi:10.17815/jlsrf-4-121-1.
# Discovery of a Rare Eclipsing Be/X-ray Binary System, Swift J010902.6-723710 = SXP 182 Thomas M. Gaudin Pennsylvania State University Jamie A. Kennea Pennsylvania State University M.J. Coe Physics and Astronomy, The University of Southampton, SO17 1BJ, UK I. M. Monageng South African Astronomical Observatory, PO Box 9, Observatory, Cape Town 7935, South Africa Department of Astronomy, University of Cape Town, Private Bag X3, Rondebosch 7701, South Africa Andrzej Udalski Astronomical Observatory, University of Warsaw, Al. Ujazdowskie 4, 00-478 Warszawa, Poland L. J. Townsend Southern African Large Telescope, PO Box 9, Observatory, Cape Town 7935, South Africa South African Astronomical Observatory, PO Box 9, Observatory, Cape Town 7935, South Africa David A.H. Buckley South African Astronomical Observatory, PO Box 9, Observatory, Cape Town 7935, South Africa Southern African Large Telescope, PO Box 9, Observatory, Cape Town 7935, South Africa Department of Astronomy, University of Cape Town, Private Bag X3, Rondebosch 7701, South Africa Department of Astronomy, University of the Free State, PO Box 339, Bloemfontein, Cape Town, 9300, South Africa Phil A. Evans Leicester University, UK ###### Abstract We report on the discovery of Swift J010902.6-723710, a rare eclipsing Be/X-ray Binary system by the Swift SMC Survey (S-CUBED). Swift J010902.6-723710 was discovered via weekly S-CUBED monitoring observations when it was observed to enter a state of X-ray outburst on 10 October 2023. X-ray emission was found to be modulated by a 182s period. Optical spectroscopy is used to confirm the presence of a highly-inclined circumstellar disk surrounding a B0-0.5Ve optical companion. Historical UV and IR photometry are then used to identify strong eclipse-like features re- occurring in both light curves with a 60.623 day period, which is adopted as the orbital period of the system. Eclipsing behavior is found to be the result of a large accretion disk surrounding the neutron star. Eclipses are produced when the disk passes in front of the OBe companion, blocking light from both the stellar surface and circumstellar disk. This is only the third Be/X-ray Binary to have confirmed eclipses. We note that this rare behavior provides an important opportunity to constrain the physical parameters of a Be/X-ray Binary with greater accuracy than is possible in non-eclipsing systems. ††facilities: Swift (XRT and UVOT), OGLE, SALT ## 1 Introduction Be/X-ray Binaries (BeXRBs) are a type of interacting High Mass X-ray Binary (HMXB) that contain a main sequence OBe star and a compact object, typically a neutron star (NS). These binaries are characterized by moderately eccentric elliptical orbits ($e\sim 0.3$), orbital periods on the order of $\sim$10-1000 days, and the presence of strong emission lines such as the Balmer series H$\alpha$ line in the optical spectrum of the OBe star (for a comprehensive review of BeXRBs, see Reig (2011)). The optical emission lines are interpreted to be a strong indication of a geometrically thin circumstellar disk (Rivinius, 2019) that surrounds the donor star and is variable in size. Interactions between the NS and disk are capable of producing intermittent X-ray outbursts, which are perhaps the most prominent feature of BeXRB systems. There are two types of outburst that can be produced by NS interactions with the circumstellar disk (Stella et al., 1986; Okazaki & Negueruela, 2001; Reig, 2011). Type I outbursts are periodic in nature and associated with NS-disk interactions that occur at the periastron passage in each orbit(Stella et al., 1986; Okazaki & Negueruela, 2001). Type II outbursts are much more luminous than Type I outbursts (Okazaki & Negueruela, 2001; Reig, 2011) and can even reach super-Eddington luminosities, such as during the SMC X-3 outburst described by Townsend et al. (2017) and the SMC X-2 outburst that occurred in 2015 (Jaisawal et al., 2023; Roy et al., 2022). These luminous outbursts are independent of the orbital phase of the NS at the time of outburst and can last for multiple orbits (Reig, 2011). Due to both the long duration and phase independence of these events, Type II outbursts are thought to be related to the growth and shape of the Be star’s disk (Martin et al., 2014; Monageng et al., 2017) causing it to interact with the NS for a longer time. Since X-ray outbursts in BeXRBs are dependent on the disk of their donor star, these systems are prone to experiencing long quiescent states during which they are hard to detect and identify (Coe et al., 2023b). Swift J010902.6-723710 is an example of a system that escaped identification due to a long period of quiescence. This newly-discovered system resides in the Small Magellanic Cloud (SMC), a satellite galaxy of the Milky Way that is well- documented to have a large population of HMXBs (Antoniou et al., 2010) due to an unusually-active period of star formation that occurred in the SMC’s recent past (Harris & Zaritsky, 2004; Rezaeikh et al., 2014). New SMC BeXRBs are still occasionally discovered (Maitra et al., 2023), indicating that a hidden population of previously-undetected systems still exists within the dwarf galaxy. In an effort identify X-ray outbursts and detect new sources, the Swift SMC Survey (S-CUBED) (Kennea et al., 2018) has operated since 2016 to provide regular X-ray and UV monitoring of the SMC. This survey utilizes the Neil Gehrels Swift Observatory (Gehrels et al., 2004) to tile the galaxy on a weekly cadence with 142 overlapping tiles and $\sim$1 minute spent on each tile. Data recorded by S-CUBED are automatically analyzed to flag new X-ray outburst events that are detected by Swift’s X-ray Telescope (XRT; Burrows et al. 2005). S-CUBED also observes all tiles using the Ultraviolet/Optical Telescope (UVOT; Roming et al. 2005) in the uvw1 band centered at 2600Å. The archive of X-ray and UV photometric data taken by S-CUBED has proved to be an invaluable resource for discovering previously-undetected BeXRBs (Monageng et al., 2020; Kennea et al., 2020; Coe et al., 2021; Kennea et al., 2021; Monageng et al., 2022). It was through an X-ray selected search of the S-CUBED archive that Swift J010902.6-723710 was first identified as a candidate source in the summer of 2023, several months before it entered outburst during October 2023. In this paper, we report on the recent X-ray outburst that was experienced by Swift J010902.6-723710 which confirms its status as a newly-identified BeXRB system (Coe et al., 2023a). In the following sections, we present the results of observations made both before and during outburst using Swift, the Optical Gravitational Lensing Experiment (OGLE), and the Southern African Large Telescope (SALT). Additionally, all results are combined and analyzed in an effort to obtain information about the orbital and physical properties of both the NS and its mass donor OBe companion. ## 2 Observations ### 2.1 Swift - XRT Figure 1: XRT light curve for Swift J010902.6-723710 from August 2023 to present, combining data from S-CUBED observations and follow-up TOOs. Note source visibility to Swift is indicated by the grey background, white vertical strips indicate periods when the target could not be observed. Vertical dotted lines mark predicted times for the periastron passages. Swift J010902.6-723710 has remained in a quiescent state for most of the duration of S-CUBED monitoring. XRT first detected this source during weekly S-CUBED monitoring in March of 2020. Two more detections occurred in July and September of 2021. This mini-flare of X-ray photons went un-reported due to its sparse detection frequency. The X-ray luminosity of the flare peaked at an XRT count rate of $0.071_{-0.035}^{+0.051}$ counts s-1 before returning to a quiescent state. The peak XRT count rate implies an approximate 0.3-10 keV band luminosity of $L_{X}=4.76\times 10^{36}$ erg s-1 if the standard distance of 62.44 kpc (Graczyk et al., 2020) to the SMC is assumed. The source was identified as a possible candidate BeXRB by examining the optical/IR SEDs of all unidentified X-ray sources in the S-CUBED survey (Gaudin et al. in prep.). This identification motivated a deep 10ks Swift observation in August of 2023 to try and constrain the hardness ratio of the source’s X-ray spectrum. However, the source was only weakly detected with a mean luminosity of $L_{X}=9.07\times 10^{35}$ erg s-1. The results of all XRT observations during the recent outburst are shown in Figure 1. Weekly S-CUBED monitoring detected X-ray emission from the source as it entered outburst on October 10th and again at peak XRT count rate on October 31st. The peak XRT count rate of Swift J010902.6-723710 was $0.22_{-0.058}^{+0.071}$ counts s-1, implying a peak 0.3-10 keV band luminosity of $L_{X}=1.35\times 10^{37}$ erg s-1 at a distance of 62.44 kpc. S-CUBED detection of outbursting behavior triggered deeper follow-up Swift observations of 5 ks per day on November 2nd and 4th in an effort to both monitor the outburst and constrain the spin period of the NS. Deep observations show the XRT count rate start to decline exponentially as expected, reaching a mean XRT count rate value of 0.11 counts s-1 during the early November observations. Additional observations taken on November 22nd and December 2nd confirm the continuing trend of exponential decline in X-ray brightness as the outburst fades. However, weekly S-CUBED detections persisted, with additional detections occurring on December 12th and December 26th. These two detections had 0.3-10 keV luminosities of $L_{X}=3.55\times 10^{36}$ erg s-1 and $L_{X}=3.98\times 10^{36}$ erg s-1, respectively, which do not fit the trend of steadily declining flux values after the peak luminosity was reached. Figure 2: $Z_{2}^{2}$ Periodogram for four observations with exposures of $\simeq$5 ks of Swift XRT data taken in PC mode. The periodogram shows a peak at 182 seconds in all four observations made after outburst was detected by S-CUBED weekly monitoring. There is evidence for pulsar spin-up over time, however the errors on period measurement are large meaning this result is not statistically significant. All 5 ks follow-up Swift observations were searched for the presence of periodic pulsations that are indicative of the spin period of the X-ray pulsar. The results of this search, carried out using a $Z_{2}^{2}$ Periodogram (Buccheri et al., 1983), are shown in Figure 2. In all 5 ks observations, the periodogram shows peaks at around $182$s (see Figure 2 for individual period measurements). Pulsar periods show small evidence for spin- up during the outburst, although given large errors on the periodicities due to the low numbers of counts and low (2.5s) PC mode frame time, the spin-up is not strongly detected. Folding the light curve of each individual XRT observation of 182s reveals an asymmetric double-peaked periodic signature, which is the expected shape for an X-ray pulsar light curve. Based on the phase-folded light curve and the strength of the periodogram peaks, we argue that the true spin period of the pulsar is 182s, allowing Swift J010902.6-723710 to be designated as SXP 182. Spectral fitting is used to derive values for the spectral index of this source and column density along the line of sight. This is done by fitting the time-averaged 0.3-10 keV spectrum to an absorbed power-law model using the methods presented in Evans et al. (2014). Using these methods, best-fit parameters and a 68% confidence interval can be derived for both properties. SXP 182 is shown to have a hard X-ray spectrum, with a derived photon index for the source of $\Gamma=0.52_{-0.15}^{+0.16}$, which is consistent with the photon index of $\Gamma\sim 0-1$ expected of a BeXRB system (Kennea et al., 2018). Additionally, power law fitting indicates that the column density along the line of sight towards this source is $N_{H}=5.1_{-1.9}^{+2.3}\times 10^{21}$ cm-2 which is higher than the average value of the column density of $N_{H}=5.34\times 10^{20}\text{cm}^{-2}$ towards the SMC (Willingale et al., 2013; Kennea et al., 2018). Figure 3: OGLE IV, Swift UVOT, and Swift XRT light curves for Swift J010902.6-723710. ### 2.2 Swift - UVOT Swift J010902.6-723710 has been observed 221 times by UVOT in the uvw1-band since 2016 as part of S-CUBED monitoring, which represents approximately weekly coverage. UVOT light curves are generated for S-CUBED sources using the FTOOLS software package (Blackburn et al., 1999). Photometric data are extracted from a circle with 5 arcsecond radius around the XRT source position for the object using the uvotsource method that is part of FTOOLS. Using this method, the light curve presented in Figure 3 was generated for the entire duration of S-CUBED. UVOT data identifies Swift J010902.6-723710 as a persistent emitter in the uvw1 band as it is well-detected in every observation with a 14.4 mean magnitude. The most prominent feature of this light curve is the presence of strong UV variability on timescales of 500-1000 days. The source reached its minimum average magnitude of 14.7 in October 2021 after years of gradual dimming from its initial brightness of 14.1 at the start of S-CUBED monitoring. At this point, the source began a period of rapid brightening lasting for over 500 days, reaching a maximum magnitude of 13.7 on June 27, 2023. After the peak magnitude was reached, the source has entered a steep dimming phase that corresponds with the start of the XRT outburst. This dimming phase was observed to be concluded on MJD 60283, when the source was observed to be at its pre-outburst uvw1 magnitude. The long period of UV brightening that Swift J010902.6-723710 underwent was crucial in identifying the system as a candidate BeXRB before the outburst occurred. This type of UV increase leading to outburst has been observed in other sources such as Swift J004516.6–734703 (Kennea et al., 2020) which demonstrated very little UV variability before experiencing a similar brightening in the lead up to a Type I outburst. In the outburst of Swift J004516.6–734703, the UV brightening was interpreted to be the result of a circumstellar disk forming around the companion Be star in the newly- identified binary. For Swift J010902.6-723710, we can interpret the long-term variability to be an indication of growth and decay in the size of the disk, where the disk expanded over the last 500+ days to a large-enough radius for NS-disk interactions to occur. ### 2.3 OGLE - the Optical Gravitational Lensing Experiment The OGLE project (Udalski et al., 2015) provides long term I-band photometry with a cadence of 1-3 days for sources in the Magellanic Clouds. From the X-ray position the optical/IR counterpart to Swift J010902.6-723710 is identified as 2MASS J01090226-7237101. It was observed continuously for nearly 3 decades by OGLE, though the observations were interrupted for approximately 3 years during COVID-19. The counterpart to the X-ray source is identified in the OGLE catalogue as: OGLE II (I-band): smc_sc11.107571 OGLE III (I-band): smc110.3.22311 OGLE IV (I band): smc726.26.15515 OGLE IV (V band): smc726.26.v.22358 The I band data from just the OGLE IV project are shown in the top panel of Figure 3 for comparison with the contemporaneous Swift UVOT and XRT data. Figure 4: Top panel - the whole OGLE data set covering 26 years. Lower panel - the above data detrended with a simple polynomial in preparation for timing analysis. Figure 5: Generalised Lomb-Scargle analysis of the detrended OGLE data set. The peak is at a period of $60.623\pm 0.001$ days.The horizontal dashed line represents the False Alarm Probabilty of 1%. Figure 6: Detrended OGLE data divided into 5 consecutive epochs (see Table 1) and each then folded using the ephemeris given in Equation 1. For the purpose of being able to see the separate profiles, those numbered 2-4 have been arbitrarily shifted in the y-axis such as to avoid overlap. Table 1: OGLE data date ranges used in Figure 6 Data block | Dates | Average I band ---|---|--- number | JD-2450000 | magnitude 1 | 600-2000 | 16.26 2 | 2000-3100 | 16.22 3 | 3100-5000 | 16.35 4 | 5200-6400 | 16.37 5 | 6400-9000 | 16.34 For the purpose of period analysis the whole of the 26.4 years of OGLE data were used and first detrended using a simple polynomial - see Figure 4. This data set was then analysed with a generalised Lomb-Scargle routine and the resulting power spectrum is shown in Figure 5. The peak in the power spectrum is at a period of $60.623\pm 0.001$ days. This peak is driven by the sharp eclipse-like features that can be clearly seen by eye in the top panels of Figures 3 and 4. It is assumed that this represents the orbital period of the system. If all the detrended OGLE data are folded at the proposed binary period, with a phase 0.0 set to the date of the first OGLE measurement (JD 2450627.9) then the sharp eclipse-like dips are clearly seen in Figure 6. The average FWHM of the dips is 0.1 in phase. It is also clear that the ingress into the eclipse is much less sharp than the egress. To explore the changes in the shape of the eclipse over time the OGLE data were divided up into 5 separate time segments representing the times when visible changes were occurring in the profile. The resulting separate profiles are shown in Figure 6. The 5 epochs chosen are listed in Table 1. Using the phase of the dip in the I-band shown in Figure 6 as a reference point, the ephemeris for the time of the optical eclipses is given by Equation 1: $T_{ecl}=2450645.1+N(60.62\pm{0.01})~{}\textrm{~{}JD}$ (1) In addition to the regular I-band measurements the OGLE IV project also records V band magnitudes every few days. This provides the opportunity to investigate the overall colour changes seen in the system as function of brightness - a colour-magnitude diagram (CMD). Since the V band measurements are less frequent than the I band, the determination of (V-I) can only occur when the I and V measurements are close enough together in time. In this instance the proximity of the 2 measurements is set to be less than 3 days. This means that occasionally one V band measurement may be partnered with more than one I band measurements. The result is shown in Figure 7 and discussed below. ### 2.4 SALT - the Southern African Large Telescope Figure 7: $(V-I)-I$ (left) and $(U-I)-I$ (right) colour-magnitude diagrams of Swift J010902.6-723710 Table 2: The H$\alpha$ equivalent width (EW) and peak separation ($\Delta V$) measured from SALT observations. Date | MJD | EW (Å) | Grating | $\Delta V$ (km/s) ---|---|---|---|--- 03-11-2023 | 60251.88 | -7.4$\pm$0.3 | PG0900 | $-$ 09-11-2023 | 60257.87 | -7.3$\pm$0.5 | PG2300 | 373.6$\pm$0.4 02-12-2023 | 60280.85 | -7.4$\pm$0.4 | PG2300 | 383.4$\pm$0.9 Swift J010902.6-723710 was observed with the Southern African Large Telescope (SALT; Buckley et al. 2006) with the Robert Stobie Spectrograph (Burgh et al., 2003; Kobulnicky et al., 2003) using different settings. On 03-11-2023 (MJD60251.88) the PG0900 grating was used (grating angle of 15.125∘) with an exposure time of 1200 sec covering a wavelength range of $4200-7250$ Å. The PG2300 grating was used on 09-11-2023 (MJD60257.87) and 02-12-2023 (MJD60280.85) (grating angle of 48.875∘) with an exposure time of 1200 sec covering a wavelength range $6100-6900$ Å. An additional observation was taken on 09-11-2023 (MJD60257.87) with the PG2300 grating (grating angle of 30.5∘) with an exposure time of 1500 sec covering a wavelength range $3840-4915$ Å. The primary reductions, which include overscan corrections, bias subtraction, gain and amplifier cross-talk corrections, were performed with the SALT science pipeline (Crawford et al., 2012). The remainder of the data reduction steps, which comprise wavelength calibration, background subtraction, and extraction of the one-dimensional spectrum, were done in iraf. All spectra were corrected for the heliocenter and redshift of the SMC of 145.6 km/s (McConnachie, 2012). Figure 8 shows the H$\alpha$ emission line profiles. The observations obtained on MJD60257.87 and MJD60280.85 with the PG2300 grating exhibit asymmetric double-peak profiles as a result of the Keplerian distribution of matter in the disc when the disc is viewed at non-zero inclination angles. The observation performed on MJD60251.88 with the PG0900 grating shows an asymmetric single-peak profile since the resolution of the grating is insufficient to resolve the two peaks. The H$\alpha$ equivalent width measurements are recorded in Table 2. ## 3 Discussion ### 3.1 Corbet Diagram Combining the spin period with the orbital period allows this new system to be placed on the Corbet diagram. The original Corbet diagram (Corbet, 1984) showed a correlation between the spin and orbital periods of neutron stars in BeXRBs. However, instead of a correlation between spin and orbital periods, a modern version (shown in Figure 9) shows that these properties are well- constrained to a specific region of the Corbet diagram. BeXRB systems are expected to fall above the diagonal of the diagram, which represents a log- linear relationship between the orbital period in days and the spin period in seconds of the NS. Figure 9 shows the location of SXP 182 when placed on the Corbet diagram with all BeXRBs in the SMC that have known spin and orbital periods (Haberl & Pietsch, 2005; Coe & Kirk, 2015; Haberl & Sturm, 2016; McBride et al., 2017; Carpano et al., 2017; Kennea et al., 2018; Lazzarini et al., 2019; Kennea et al., 2020; Carpano et al., 2022; Maitra et al., 2023). The location of SXP 182 is consistent with the trend shown by other SMC BeXRBs. This serves as a check that the periods derived in Sections 2.1 and 2.3 are consistent with the observational properties of other BeXRBs. For Swift J010209.6-723710, the derived orbital period of 60.623 days and pulsar spin period of 182 seconds place the system near the center of the distribution of orbit and spin periods for SMC BeXRBs, providing a strong piece of evidence validating the source as a newly-discovered BeXRB. ### 3.2 Inclination angle of the Be disc The peak separation of the double-peak H$\alpha$ emission lines (Figure 8) can be used to estimate the size of the H$\alpha$ emitting region (Huang, 1972): $R=\frac{GM_{\ast}\sin^{2}i}{(0.5\Delta V)^{2}}$ (2) Using $M_{\ast}=17.8$ M⊙ (Cox, 2000) based on the spectral type of the massive companion, this results in a disc size radius range of $92-97\sin^{2}i$ R⊙. We can estimate the inclination angle of the Be disc, assuming that the disc and orbital planes are aligned and that during the period of X-ray activity the NS was accreting matter from the outermost parts of the Be disc at periastron passage. The semi-major axis of the orbit can be estimated by assuming a companion mass of $M_{\ast}=17.8$ M⊙ and a neutron star mass of $M_{NS}=1.4$ M⊙. Similarly, the periastron passage can be estimated using a conservative value for the eccentricity of $e\sim 0.5$ based on the relationship between eccentricity and orbital period presented in Townsend et al. (2011). This results in a range of disc inclination angles of $72-90^{\circ}$. The suggestion of a high disc inclination angle is corroborated by the H$\alpha$ emission line displaying a double peak morphology with a deep central depression from the high-resolution observations. Figure 8: The evolution of the H$\alpha$ emission line from SALT observations. The spectra are corrected for the heliocenter and redshift of the SMC. Figure 7 shows the $(V-I)-I$ and $(U-I)-I$ colour-magnitude diagrams. The general trend of the colour-magnitude plots is a redder-when-brighter pattern, which is indicative of inclination angles below $90^{\circ}$(Harmanec, 1983; Rajoelimanana et al., 2011; Reig & Fabregat, 2015). This trend is more noticeable in the $(V-I)-I$ plot where the range in colour is broader since the simultaneous $V$ and $I$ band observations were taken during a period of substantial optical variability. Figure 9: Corbet diagram for all BeXRBs in the SMC with known orbital periods and spin periods (Haberl & Pietsch, 2005; Coe & Kirk, 2015; McBride et al., 2017; Carpano et al., 2017; Kennea et al., 2018; Lazzarini et al., 2019; Kennea et al., 2020; Carpano et al., 2022; Maitra et al., 2023). SXP 182 is plotted alongside these data and is shown to demonstrate the same correlation between X-ray pulsar spin period and NS orbital period that is seen in the rest of the SMC BeXRB population. ### 3.3 Spectral Classification of the massive counterpart Figure 10 shows the SALT spectrum of Swift J010902.6-723710 covering the blue wavelength range. The spectrum shows clear signatures of an early-type star with several Balmer and helium lines present. The H$\beta$ line is in weak emission with an asymmetric profile, along with the other absorption lines exhibiting infilling due to the presence of the circumstellar disc. According to the criteria presented in Evans et al. (2004), the strong presence of the He I lines at 4026, 4143, 4388 and 4471 Å as well as the weak presence of the He II lines at 4541 and 4686 Å constrains the spectral type to B0-0.5. The faintest V-band observation from OGLE measurements during our period of monitoring is $\sim$16.5 magnitudes. Using this and the distance modulus of the SMC of 18.95 (Graczyk et al., 2013), this results in a luminosity class of V (Straizys & Kuriliene, 1981; Pecaut & Mamajek, 2013). In summary, the spectral class of the massive companion in Swift J010902.6-723710 is B0-0.5 V. Figure 10: The SALT spectrum of Swift J010902.6-723710 covering the blue region with different line species labeled at their expected rest wavelengths. The spectrum is corrected for the heliocenter and redshift of the SMC. ### 3.4 Eclipsing Behavior The orbital period of Swift J010902.6-723710 derived from long-term OGLE monitoring can be used to re-examine the UVOT light curve in search of periodic behavior. This was done by de-trending the UVOT data using a 5th- order polynomial and folding the de-trended data at the 60.623 day orbital period of the system. Figure 11 shows the comparison between the binned and time-averaged OGLE and UVOT light curves starting at MJD 57563, which is the date of the first S-CUBED observation. Despite S-CUBED’s weekly cadence providing infrequent sampling of any singular orbit, the eclipsing behavior is clearly visible in the de-trended and folded UVOT light curve. When compared to the OGLE data, it becomes evident that the shape of the eclipse profile is wavelength-dependent. The UVOT data shows increasing emission that peaks just before each eclipse begins, which is similar to the behavior exhibited by the system during the 4th OGLE data block shown in Figure 6. Over the lifetime of S-CUBED, the UVOT eclipse profile is also shown to be broader than the OGLE eclipse profile with similar depth at the time of maximum eclipse. At the maximum depth of the eclipse, the uvw1-band magnitude drops by an average of 0.22 magnitudes and I-band magnitude decreases by an average of 0.25 magnitudes. Converting these magnitude decreases to flux measurements for both telescopes, the implied relative decrease in flux is at both wavelengths $\frac{\Delta F}{F}\sim 0.2$. Figure 11: A comparison of the OGLE IV and UVOT light curves folded at the proposed binary period of 60.623 days using Equation 1. The OGLE data are offset vertically to avoid overlap. Equation 2 from Maggi et al. (2013) can be used to estimate the size of the accretion disk based on the relative decrease in flux that is observed during eclipse: $\frac{\Delta F}{F}=\left(\frac{R_{X}}{R_{C}}\right)^{2}$ (3) where $R_{X}$ is the radius of the eclipsing object and $R_{C}$ is the radius of the optical companion. The value for $R_{X}$ can be estimated due to the constraints placed on $R_{C}$ via the results of Section 3.3. The typical radius for a spectral class B0V star is 7.4 $R_{\odot}$ (Allen, 1976). This value can be used to constrain the upper limit for the radius of the eclipsing object at $R_{X}=R_{C}\sqrt{\frac{\Delta F}{F}}=3.3\,R_{\odot}$. A radius of this size rules out the NS, a large planet, or a Sun-like tertiary star as the cause of the eclipsing behavior. Therefore, it can be concluded that the eclipsing object is likely an extended accretion disk that surrounds the NS. Maggi et al. (2013) derives the theoretical upper limit for the size of an accretion disk in a BeXRB system to be $r_{c}\sim 11.5\,R_{\odot}$, assuming Bondi-Hoyle accretion and a 10 $M_{\odot}$ companion star. The 3.3 $R_{\odot}$ accretion disk size derived for Swift J010902.6-723710 is well within this upper limit value, providing further evidence in favor of an eclipsing NS accretion disk. Very few BeXRBs have been observed to demonstrate eclipsing behavior. Swift J010902.6-723710 is only the third known eclipsing BeXRB system and the second to be found in the SMC. LXP 168.8 (Maggi et al., 2013) was found to have a eclipsing accretion disk with a 24.3 day orbital period. SXP 5.05 (Coe et al., 2015) was observed to contain a Be star that eclipses its NS companion every 17 days. This newly-discovered eclipsing system thus provides a unique opportunity to further constrain the physical parameters of the system such as the sizes of both disks and the masses of both stars. More observations are needed to further characterize the BeXRB using subsequent eclipses. ### 3.5 Outburst Type Date | MJD | Phase | $L_{X}$ (erg s-1) ---|---|---|--- 10-10-2023 | 60227 | 0.06 | $4.14_{-2.61}^{+4.27}\times 10^{36}$ 10-31-2023 | 60248 | 0.40 | $1.35_{-0.391}^{+0.475}\times 10^{37}$ 11-02-2023 | 60250.0 | 0.47 | $5.66_{-0.275}^{+0.275}\times 10^{36}$ 11-02-2023 | 60250.9 | 0.48 | $7.06_{-0.527}^{+0.527}\times 10^{36}$ 11-04-2023 | 60252.1 | 0.50 | $5.30_{-0.508}^{+0.508}\times 10^{36}$ 11-05-2023 | 60253.0 | 0.52 | $4.96_{-0.253}^{+0.253}\times 10^{36}$ 11-05-2023 | 60253.6 | 0.53 | $4.03_{-0.385}^{+0.385}\times 10^{36}$ 11-06-2023 | 60254 | 0.54 | $4.95_{-0.518}^{+0.518}\times 10^{36}$ 11-21-2023 | 60269.9 | 0.80 | $3.20_{-2.04}^{+3.33}\times 10^{36}$ 11-22-2023 | 60270.6 | 0.81 | $2.25_{-0.190}^{+0.190}\times 10^{36}$ 12-02-2023 | 60280 | 0.96 | $1.83_{-0.177}^{+0.177}\times 10^{36}$ 12-08-2023 | 60286 | 0.06 | $1.43_{-0.192}^{+0.192}\times 10^{36}$ 12-12-2023 | 60290 | 0.13 | $3.55_{-2.28}^{+3.78}\times 10^{36}$ 12-14-2023 | 60292 | 0.16 | $1.17_{-0.397}^{+0.397}\times 10^{36}$ 12-26-2023 | 60304 | 0.36 | $3.98_{-2.26}^{+3.69}\times 10^{36}$ Table 3: S-CUBED and Swift TOO XRT detection dates, orbital phases from Equation 1, and luminosities during outburst. Figure 12: Truncated OGLE IV, Swift UVOT, and Swift XRT light curves for Swift J010902.6-723710 showing emission from MJD 60100 to MJD 60300. The times of optical eclipse, calculated using Equation 1, are shown as vertical grey lines. Blue arrows show the calculated XRT flux upper limit during S-CUBED observations where Swift J010902.6-723710 is not detected. The type of outburst observed in Swift J010902.6-723710 has remained uncertain as long-term monitoring of the source continues to produce XRT detections well past the peak of the initial outburst event. In Type I outbursts, X-ray emission is often limited both by the duration of an orbit and the current orbital phase. Emission is expected to be detected at a narrow range of values near the periastron passage of the orbit and is expected to last for no longer than a full orbital period (Stella et al., 1986). Additionally, Type I outbursts typically demonstrate a moderate increase in luminosity, peaking at $L_{X}\sim 10^{36}-10^{37}$ ergs s-1 (Okazaki & Negueruela, 2001; Reig, 2011). Table 3 shows the date of XRT detections of Swift J010902.6-723710 and the orbital phase at which they occurred with respect to Equation 1. The light curve of XRT detections during outburst is plotted with respect to the time of optical eclipses in Figure 12. Both the figure and table show the unusually long duration of this outburst. The first S-CUBED detection of the outburst on October 10th occurred at Phase 0.06, which is approximately the periastron passage of orbit. However, other subsequent detections continue to occur well past the periastron passage of the orbit, demonstrating phase independence of the emission. The last two S-CUBED detections, occurring on December 12th and December 26th, correspond to Phase 0.13 and 0.36 of a new orbit. These subsequent detections are not consistent with the typical behavior demonstrated by a Type I outburst and have more in common with the emission signature of a Type II outburst (Okazaki & Negueruela, 2001; Townsend et al., 2017; Tamang et al., 2022). If this is indeed a Type II outburst, then the peak luminosity is much smaller than is typical for these outbursts. The peak X-ray luminosity of $L_{X}=1.35\times 10^{37}$ erg s-1 is the only detection to occur above $10^{37}$ erg s-1. Based on these features, one of two situations has occurred. One possibility is that the system has produced an abnormally long-duration Type I outburst with a typical peak luminosity (Okazaki & Negueruela, 2001; Reig, 2011). Alternatively, the system has produced a Type II outburst that fails to reach the characteristic large peak luminosity that is expected during these events (Okazaki & Negueruela, 2001; Reig, 2011). However, it is important to remember that this traditional classification scheme is very quantised - either it is called Type I or Type II. Since originally proposed in 1986 (Stella et al., 1986) there have been many more diverse examples of X-ray outbursts from BeXRB systems observed. In reality there must be a scope for a whole spectrum of outburst types since the outburst duration and the X-ray luminosity seen depend almost entirely on the interaction between the circumstellar disc and the neutron star. That in turn depends upon the characteristics of the circumstellar disc (density, size, inclination, mass outflow rate etc.) and the characteristics of the neutron star orbit (eccentricity, inclination to the circumstellar disk plane, phase of periastron etc.). So, given all those free parameters one should expect a whole range of observational characteristics, and not be driven to simply call them Type I or Type II. SXP 182 is an excellent example of how the original classification scheme can be overly simplistic. ## 4 Conclusion This paper reports the detection of a previously-unknown BeXRB via weekly observations of the S-CUBED survey. This new system, Swift J010902.6-723710, was identified via a transient X-ray outburst and followed up via multi- wavelength observations. Deep follow-up X-ray observations identify a proposed spin period of 182s for the NS in this binary system. Historical light curve analysis of both UV and IR emission reveal the presence of strong eclipse-like features that re-appear every 60.623 days in the light curve, which is adopted as the proposed orbital period of the system. Optical spectroscopy reveals a strongly double-peaked H$\alpha$ emission line, indicating a highly-inclined system with an inclination of $i=72-90^{\circ}$. Spectroscopic observations are also used to constrain the spectral class of the optical companion as a B0-0.5 star of spectral class Ve. The proposed spin and orbital periods place Swift J010902.6-723710 place in the center of the expected distribution for similar BeXRBs on the Corbet Diagram. Eclipsing behavior is found to be caused by a $3.3R_{\odot}$ accretion disk that surrounds the NS, making this the third eclipsing BeXRB to be detected so far. Finally, the type of outburst observed is found to be uncertain, with characteristics of both Type I and Type II outbursts being found in the X-ray emission of the source. More observations of this system are needed, particularly during subsequent eclipses, in order to further constrain the physical parameters of this rare eclipsing system and better understand this long-lasting, moderately luminous X-ray outburst. ## 5 acknowledgments This work made use of data supplied by the UK Swift Science Data Centre at the University of Leicester. JAK and TMG acknowledge the support of NASA contract NAS5-00136. We acknowledge the use of public data from the Swift data archive. The SALT observations reported in this paper were obtained under the SALT Large Science Programme on transients (2021-2-LSP-001; PI: DAHB), which is also supported by Poland under grant no. MNiSW DIR/WK/2016/07. PAE acknowledges UKSA support. DAHB acknowledges research support from the National Research Foundation. LJT is supported by the SALT Foundation and the NRF. ## References * Allen (1976) Allen, C. W. 1976, Astrophysical Quantities * Antoniou et al. (2010) Antoniou, V., Zezas, A., Hatzidimitriou, D., & Kalogera, V. 2010, ApJ, 716, L140, doi: 10.1088/2041-8205/716/2/L140 * Blackburn et al. (1999) Blackburn, J. K., Shaw, R. A., Payne, H. E., Hayes, J. J. E., & Heasarc. 1999, FTOOLS: A general package of software to manipulate FITS files, Astrophysics Source Code Library, record ascl:9912.002. http://ascl.net/9912.002 * Buccheri et al. (1983) Buccheri, R., Bennett, K., Bignami, G. F., et al. 1983, A&A, 128, 245 * Buckley et al. (2006) Buckley, D. A. H., Swart, G. P., & Meiring, J. G. 2006, in Proc. SPIE, Vol. 6267, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, 62670Z, doi: 10.1117/12.673750 * Burgh et al. (2003) Burgh, E. B., Nordsieck, K. H., Kobulnicky, H. A., et al. 2003, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4841, Proc. SPIE, ed. M. Iye & A. F. M. Moorwood, 1463–1471, doi: 10.1117/12.460312 * Burrows et al. (2005) Burrows, D. N., Hill, J. E., Nousek, J. A., et al. 2005, Space Sci. Rev., 120, 165, doi: 10.1007/s11214-005-5097-2 * Carpano et al. (2022) Carpano, S., Haberl, F., Maitra, C., et al. 2022, A&A, 661, A20, doi: 10.1051/0004-6361/202141082 * Carpano et al. (2017) Carpano, S., Haberl, F., & Sturm, R. 2017, A&A, 602, A81, doi: 10.1051/0004-6361/201629299 * Coe et al. (2015) Coe, M. J., Bartlett, E. S., Bird, A. J., et al. 2015, MNRAS, 447, 2387, doi: 10.1093/mnras/stu2568 * Coe et al. (2021) Coe, M. J., Kennea, J. A., Evans, P. A., et al. 2021, MNRAS, 504, 1398, doi: 10.1093/mnras/stab972 * Coe et al. (2023a) Coe, M. J., Kennea, J. A., Gaudin, T. M., et al. 2023a, The Astronomer’s Telegram, 16321, 1 * Coe et al. (2023b) Coe, M. J., Kennea, J. A., Monageng, I. M., et al. 2023b, MNRAS, 524, 3263, doi: 10.1093/mnras/stad1987 * Coe & Kirk (2015) Coe, M. J., & Kirk, J. 2015, MNRAS, 452, 969, doi: 10.1093/mnras/stv1283 * Corbet (1984) Corbet, R. H. D. 1984, A&A, 141, 91 * Cox (2000) Cox, A. N. 2000, Allen’s astrophysical quantities * Crawford et al. (2012) Crawford, S. M., Still, M., Schellart, P., et al. 2012, PySALT: SALT science pipeline, Astrophysics Source Code Library. http://ascl.net/1207.010 * Evans et al. (2004) Evans, C. J., Howarth, I. D., Irwin, M. J., Burnley, A. W., & Harries, T. J. 2004, MNRAS, 353, 601, doi: 10.1111/j.1365-2966.2004.08096.x * Evans et al. (2014) Evans, P. A., Osborne, J. P., Beardmore, A. P., et al. 2014, ApJS, 210, 8, doi: 10.1088/0067-0049/210/1/8 * Gehrels et al. (2004) Gehrels, N., Chincarini, G., Giommi, P., et al. 2004, ApJ, 611, 1005, doi: 10.1086/422091 * Graczyk et al. (2013) Graczyk, D., Pietrzyński, G., Pilecki, B., et al. 2013, in IAU Symposium, Vol. 289, Advancing the Physics of Cosmic Distances, ed. R. de Grijs, 222–225, doi: 10.1017/S1743921312021436 * Graczyk et al. (2020) Graczyk, D., Pietrzyński, G., Thompson, I. B., et al. 2020, ApJ, 904, 13, doi: 10.3847/1538-4357/abbb2b * Haberl & Pietsch (2005) Haberl, F., & Pietsch, W. 2005, A&A, 438, 211, doi: 10.1051/0004-6361:20042470 * Haberl & Sturm (2016) Haberl, F., & Sturm, R. 2016, A&A, 586, A81, doi: 10.1051/0004-6361/201527326 * Harmanec (1983) Harmanec, P. 1983, Hvar Observatory Bulletin, 7, 55 * Harris & Zaritsky (2004) Harris, J., & Zaritsky, D. 2004, AJ, 127, 1531, doi: 10.1086/381953 * Huang (1972) Huang, S.-S. 1972, ApJ, 171, 549, doi: 10.1086/151309 * Jaisawal et al. (2023) Jaisawal, G. K., Vasilopoulos, G., Naik, S., et al. 2023, MNRAS, 521, 3951, doi: 10.1093/mnras/stad781 * Kennea et al. (2020) Kennea, J. A., Coe, M. J., Evans, P. A., et al. 2020, MNRAS, 499, L41, doi: 10.1093/mnrasl/slaa154 * Kennea et al. (2021) —. 2021, MNRAS, 508, 781, doi: 10.1093/mnras/stab2632 * Kennea et al. (2018) Kennea, J. A., Coe, M. J., Evans, P. A., Waters, J., & Jasko, R. E. 2018, ApJ, 868, 47, doi: 10.3847/1538-4357/aae839 * Kobulnicky et al. (2003) Kobulnicky, H. A., Nordsieck, K. H., Burgh, E. B., et al. 2003, in Proc. SPIE, Vol. 4841, Instrument Design and Performance for Optical/Infrared Ground-based Telescopes, ed. M. Iye & A. F. M. Moorwood, 1634–1644, doi: 10.1117/12.460315 * Lazzarini et al. (2019) Lazzarini, M., Williams, B. F., Hornschemeier, A. E., et al. 2019, ApJ, 884, 2, doi: 10.3847/1538-4357/ab3f32 * Maggi et al. (2013) Maggi, P., Haberl, F., Sturm, R., et al. 2013, A&A, 554, A1, doi: 10.1051/0004-6361/201321238 * Maitra et al. (2023) Maitra, C., Haberl, F., Kaltenbrunner, D., et al. 2023, The Astronomer’s Telegram, 15886, 1 * Martin et al. (2014) Martin, R. G., Nixon, C., Armitage, P. J., Lubow, S. H., & Price, D. J. 2014, ApJ, 790, L34, doi: 10.1088/2041-8205/790/2/L34 * McBride et al. (2017) McBride, V. A., González-Galán, A., Bird, A. J., et al. 2017, MNRAS, 467, 1526, doi: 10.1093/mnras/stx181 * McConnachie (2012) McConnachie, A. W. 2012, AJ, 144, 4, doi: 10.1088/0004-6256/144/1/4 * Monageng et al. (2017) Monageng, I. M., McBride, V. A., Coe, M. J., Steele, I. A., & Reig, P. 2017, MNRAS, 464, 572, doi: 10.1093/mnras/stw2354 * Monageng et al. (2020) Monageng, I. M., Coe, M. J., Buckley, D. A. H., et al. 2020, MNRAS, 496, 3615, doi: 10.1093/mnras/staa1739 * Monageng et al. (2022) Monageng, I. M., Coe, M. J., Townsend, L. J., et al. 2022, MNRAS, 511, 6075, doi: 10.1093/mnras/stac106 * Okazaki & Negueruela (2001) Okazaki, A. T., & Negueruela, I. 2001, A&A, 377, 161, doi: 10.1051/0004-6361:20011083 * Pecaut & Mamajek (2013) Pecaut, M. J., & Mamajek, E. E. 2013, ApJS, 208, 9, doi: 10.1088/0067-0049/208/1/9 * Rajoelimanana et al. (2011) Rajoelimanana, A. F., Charles, P. A., & Udalski, A. 2011, MNRAS, 413, 1600, doi: 10.1111/j.1365-2966.2011.18243.x * Reig (2011) Reig, P. 2011, Ap&SS, 332, 1, doi: 10.1007/s10509-010-0575-8 * Reig & Fabregat (2015) Reig, P., & Fabregat, J. 2015, A&A, 574, A33, doi: 10.1051/0004-6361/201425008 * Rezaeikh et al. (2014) Rezaeikh, S., Javadi, A., Khosroshahi, H., & van Loon, J. T. 2014, MNRAS, 445, 2214, doi: 10.1093/mnras/stu1807 * Rivinius (2019) Rivinius, T. 2019, IAU Symposium, 346, 105, doi: 10.1017/S1743921318008207 * Roming et al. (2005) Roming, P. W. A., Kennedy, T. E., Mason, K. O., et al. 2005, Space Sci. Rev., 120, 95, doi: 10.1007/s11214-005-5095-4 * Roy et al. (2022) Roy, A., Cappallo, R., Laycock, S. G. T., et al. 2022, ApJ, 936, 90, doi: 10.3847/1538-4357/ac82b6 * Stella et al. (1986) Stella, L., White, N. E., & Rosner, R. 1986, ApJ, 308, 669, doi: 10.1086/164538 * Straizys & Kuriliene (1981) Straizys, V., & Kuriliene, G. 1981, Ap&SS, 80, 353, doi: 10.1007/BF00652936 * Tamang et al. (2022) Tamang, R., Ghising, M., Tobrej, M., Rai, B., & Paul, B. C. 2022, MNRAS, 515, 5407, doi: 10.1093/mnras/stac2135 * Townsend et al. (2011) Townsend, L. J., Coe, M. J., Corbet, R. H. D., & Hill, A. B. 2011, MNRAS, 416, 1556, doi: 10.1111/j.1365-2966.2011.19153.x * Townsend et al. (2017) Townsend, L. J., Kennea, J. A., Coe, M. J., et al. 2017, MNRAS, 471, 3878, doi: 10.1093/mnras/stx1865 * Udalski et al. (2015) Udalski, A., Szymański, M. K., & Szymański, G. 2015, Acta Astron., 65, 1. https://arxiv.org/abs/1504.05966 * Willingale et al. (2013) Willingale, R., Starling, R. L. C., Beardmore, A. P., Tanvir, N. R., & O’Brien, P. T. 2013, MNRAS, 431, 394, doi: 10.1093/mnras/stt175
# EnDex: Evaluation of Dialogue Engagingness at Scale Guangxuan Xu1 Ruibo Liu2 Fabrice Harel-Canada1 Nischal Reddy Chandra1 Nanyun Peng1 1University of California, Los Angeles 2Dartmouth College {gxu21, violetpeng} @cs.ucla.edu ###### Abstract We propose EnDex, the first human-reaction based model to evaluate dialogue engagingness. EnDex is trained on 80k Reddit-based Engagement Dataset (RED) curated using a novel distant-supervision framework. Engagingness is a key measure that captures high-level quality of AI dialogue systems and closely reflects actual user experience. However, data shortage, plus the abstract and extensive definition of engagingness makes it challenging to develop an automatic metric. Our work departs from mainstream approaches that use synthetic negative examples to train binary classifiers, and instead, proposes a solution using distant-supervision from human-reaction feedback. To support the soundness of our EnDex metric, we offer a theoretical foundation for engagement, an extensive ablation study, and empirical evidence of high correlation on five engagingness related datasets.111Off-the-shelf EnDex model and the RED dataset is available at https://github.com/gxxu-ml/EnDex. ## 1 Introduction Many modern generative language models are trained to maximize a likelihood objective, but this paradigm tends to assign high probability to generic responses (Li et al., 2016), such as “I don’t know.”. Prior research has established that people prefer to converse with interesting, creative, and informative agents (See et al., 2019), all concepts broadly related to the notion of _engagingness_. Furthermore, engagingness is recognized as a key evaluation metric for the quality of dialogue systems (Zhang et al., 2018; Ghazarian et al., 2020). For example, FAIR’s ParlAI (Miller et al., 2017) incorporated Engagingness as the default testing metric in the Blenderbot system (Roller et al., 2021); dialogue data challenges, like ConvAI2 (Dinan et al., 2019), Amazon Alexa Prize222https://www.amazon.science/alexa-prize, and ensemble metrics like FED (Mehri and Eskenazi, 2020), all measure engagingness to benchmark dialogue quality. Figure 1: Example of an online post with scores for emotional engagement (EE), attentional engagement (AE), and behavioral engagement (BE) in blue to represent the 3 dimensions of _human engagement_ ; reply engagement (RE) in red; and the aggregated EnDex score in green. We apply z-score to EnDex Score and pick a hyper-parameter threshold to cluster posts into positive and negative samples. However, the current evaluation of engagingness still primarily relies on expensive human annotation rather than off-the-shelf automatic tools, due to several theoretical and technical challenges: firstly, unlike more well- characterized properties such as fluency, the definition of engagingness is significantly more abstract and multi-dimensional (See et al., 2019), requiring well-tuned quality metrics for each sub-dimension to aggregate a final score. Secondly, what qualifies as engaging is open-ended and many different answers may embody the concept (Ghazarian et al., 2020). Therefore, reference-based metrics requiring unique ground truth, such as BLEURT (Sellam et al., 2020) and BERTScore (Zhang et al., 2020), cannot apply. Thirdly, there’s an acute shortage of large-scale, high-quality data annotated for engagingness. Ghazarian et al. (2020) jump-started efforts to automatically measure dialogue engagement, where they fine-tuned a BERT-based model Devlin et al. (2019) on the ConvAI2 and DialyDialog datasets (Li et al., 2017) to predict an engagingness score. However, finetuning on small size supervised data could easily lead to overfitting and generalization problems. Another high performing metric on engagingness USL-H (Phy et al., 2020) assumes a positive set and generates synthetic negative samples to train model. However, credible positive samples are not always available, and synthetic negative samples may not be challenging enough to further advance classifier performance. In light of the above challenges, we propose EnDex, a novel metric trained with distantly supervised data to predict turn-level dialogue engagingness (Figure 1). EnDex requires neither human annotations nor direct disentanglement of engagingness. Instead, we leverage observed _user reactions_ to posts as distant signals to model engagingness, which marks a departure from mainstream approach to train on synthetic negative samples (Lan et al., 2020; Ghazarian et al., 2022; Tao et al., 2018; Sato et al., 2020). EnDex trains on real conversations sourced from Reddit, that are automatically annotated as positive and negative examples with our framework. The novoel dataset is named RED (Reddit Engagement Dataset) with over 80k labelled samples. EnDex framework derives its theoretical underpinning from relevant HCI works, and has shown superior performance on five benchmark datasets. ## 2 EnDex Metric Engagingness is not only a linguistic concept useful for dialogue systems, but also manifests itself in multi-modalities and is extensively leveraged to benchmark gaming and online learning experiences (Silpasuwanchai et al., 2016; Chen et al., 2005; Mcmahan, 2003; Schoenau-Fog, 2011). Our work is inspired by HCI study of Human Engagement Ma (2018), which decomposes engagingness into three major dimensions including attentional engagement (e.g., clicks and scrolls), behavior engagement (e.g., facial expressions), and emotional engagement (e.g., heart rate). EnDex metric follows the same intuition: we can infer engagingness of a text by analyzing human reactions to it, for which there is abundant data in social media. EnDex metric learns from our distant-supervised RED dataset, which measures dialogue engagement along four dimensions as shown in Figure 1; three-dimensions correspond to the original Human Engagement definition, and one distinct Reply Engagement dimension for the dialogue specific task. ### 2.1 Reddit Engagement Dataset (RED) We curate the Reddit Engagement Dataset (RED), a distant-supervision set, with 80k single-turn conversations. We source RED from Reddit, sampling from 43 popular subreddits, and processed a total of 5 million posts, filtering out data that was either non-conversational, toxic, or posts not possible to ascertain popularity; the resulting data distribution of RED is shown in Table 1. The following sections will explain the procedure to automatically annotate EnDex scores and cluster samples into positive and negative sets. We also curated a RED testset with 150 human annotated samples obtained from a different split from RED. The inter-annotator agreement is 0.34 Fleiss-Kappa, indicating fair agreement, which reflects the challenge of determining engagingness. | Engaging | Non-engaging ---|---|--- # of samples | 40,162 | 40,162 Emotional | .605 $\pm$ .273 | .152 $\pm$ .120 Attentional | .759 $\pm$ .127 | .203 $\pm$ .100 Behavioral | .659 $\pm$ .274 | .318 $\pm$ .285 Reply | .718 $\pm$ .154 | .354 $\pm$ .980 EnDex | .709 $\pm$ .048 | .259 $\pm$ .033 Table 1: RED dataset has two classes, engaging and non-engaging, clustered by applying z-score on EnDex score. This table shows the mean and standard deviation of sub-dimension scores for both classes; the last row displays the distribution of the overall EnDex score. ### 2.2 Distantly-Supervised Engagingness Scores We use distant-supervision to provide samples in RED an EnDex Score, which is the aggregate of 4 engaging dimensions. Section 2.2 discusses the intuition for each engagingness dimension; section 2.3 explains how to adjust raw score by thread popularity; section 2.4 lays out the formula to normalize and aggregate sub-dimensions into the overall engagingness score; section 2.5 explains sampling with z-score to convert the task into binary classification. * • Emotional Engagement (EE): Emotional connection is a key sign of human engagement (Savin-Baden et al., 2014); and we model EE using a multi-class emotional classifier (Demszky et al., 2020) on post replies. If post receives positive and emotional replies, it’s engaging; negative or neutral replies indicates non-engaging. * • Attentional Engagement (AE): More user time spent indicates higher engagement (Attfield et al., 2011). We model AE of a post by examining whether it has editted replies, and the information specificity in the replies. * • Behavioral Engagement (BE): Human behavioral features closely correlate with their engagement state (Attfield et al., 2011), and we model BE by examining Reddit post scores, adjusted by popularity. * • Reply Engagement (RE): Following definition from (Ghazarian et al., 2020), if a certain post is very likely to be continued by following threads, it is considered engaging; reply_counts are also popularity adjusted. ### 2.3 Adjustment for Popularity Raw score for Behavior Engagement(upvotes) and Reply Engagement(reply counts) are heavily influenced by the popularity of the thread in which the post appears. A non-engaging post may receive high user interaction because it simply receives a lot of exposure; on the flip side, a very engaging most may receive zero user interaction simply because it is rarely seen. To mitigate the imbalanced exposure problem, we calculate a popularity value for each thread, and adjust posts scores by the popularity value of the thread it resides. Popularity Value(PV) The PV of a post is given by the amount of exposure its parent post attracts. Let the target post be $\theta$ and its parent $\sigma$, $R_{eply}$ obtains the reply counts of a post, and $U_{pvote}$ obtains the upvotes of a post. The PV is defined in equation (1), where coefficient 2 is adopted to give equal weight for reply and upvotes; popularity value adjusted RE score is given by PVRE in equation (6), where $M_{pv}$ and $M_{re}$ are the median of popularity value and reply counts in the entire dataset. Only popularity adjusted scores are used for calculating EnDex score. $\textrm{PV}(\theta)=2*R_{eply}(\sigma)+U_{pvote}(\sigma)$ (1) $\textrm{PVRE}(re)=re+\frac{M_{pv}}{M_{\textrm{re}}}*\frac{re}{\textrm{PV(re)}}*re$ (2) ### 2.4 Monotone Submodular Normalization The final EnDex score is essentially a weighted sum of the 4 respective sub- dimension scores; an importance nuance is the usage of submodular normalization (shown in Eq. 8) for 3 dimensions to bring raw scores to the scale of 0-1. We observe that unit increase in raw score lead to diminishing positive effect on engagingness. For example, a sentence with 100 replies should be more engaging than one with 1 replies, but not 99 times more; thus, we normalize engagingness score with a monotone submodular function $f(x)=\frac{x}{x+\alpha}$. $N(x)=\Bigg{(}\displaystyle\sum_{n=1}^{3}w_{i}*\frac{x_{i}}{x_{i}+\alpha_{i}}\Bigg{)}+w_{\textrm{EE}}*x_{\textrm{EE}},$ (3) $N$ is the normalized score for sample $x$, $x_{i}$ is $x$’s raw score on dimension $i$, where $i\in\\{RE,BE,AE\\}$; $\alpha_{i}$ is the median of $i$-th dimension; $w_{i}$ is the weight for $i$ dimension; $w_{\textrm{EE}}$ is the weight for EE dimension. The weight can be tuned for your own usage of RED; 333For EnDex, $\alpha$ for three dimensions RE, BE, AE are 1, 2, 18, respectively; we also applied weights of 3, 3, 2, 1 for RE, AE, EE, and BE. ### 2.5 Clustering with z-score Essentially, engagingness prediction is a classification task, and we want to prepare dataset for binary classification. We use $z$-score on the EnDex Score to easily sample and cluster the data according to standard deviation from mean. A confidence threshold $\kappa$(ours is 1) needs to be picked, which means that we regard samples that fall between $\kappa$ standard deviation from mean as uncertain, and are thus discarded. And we cluster positive and negative samples using the following equation (9). $Polarity(x)=\begin{cases}1&\text{if }z\\_score(x)>\kappa\\\ 0&\text{if }z\\_score(x)<-\kappa\end{cases}$ (4) The EnDex metric is then trained as a binary classification task by finetuning a RoBERTa-large model (Liu et al., 2019) on turn-level RED data. ## 3 Experiments Method | Better | PredEng-600 | Fed-eng | Red-Test | Grade ---|---|---|---|---|--- P | S | P | S | P | S | P | S | P | S Random (ref.) | 0.025 | 0.025 | -0.012 | -0.013 | 0.080 | 0.081 | -0.053 | -0.053 | 0.053 | 0.045 Question | 0.167 | 0.167 | 0.073 | 0.074 | 0.320 | 0.320 | 0.194 | 0.194 | 0.009 | 0.008 Specificity | 0.357 | 0.357 | 0.076 | 0.102 | 0.254 | 0.254 | 0.122 | 0.122 | -0.090 | -0.090 USL-H | 0.356 | 0.343 | 0.688 | 0.699 | 0.267 | 0.277 | 0.121 | 0.125 | 0.234 | 0.243 Pred_En | 0.234 | 0.310 | -0.137 | -0.134 | 0.250 | 0.340 | 0.044 | 0.178 | -0.090 | -0.060 Pred_En (FT+DD) | 0.338 | 0.368 | 0.390 | 0.450 | 0.253 | 0.195 | 0.237 | 0.258 | 0.194 | 0.173 Ours: EnDex | 0.414* | 0.397* | 0.397 | 0.348 | 0.235* | 0.225* | 0.381** | 0.375** | 0.266 | 0.248 Ours: EnDex+NS | 0.478* | 0.455* | 0.597** | 0.577** | 0.229* | 0.214* | 0.389** | 0.378** | 0.308 | 0.282* Ours: EnDex-Best | 0.521 | 0.511 | 0.620 | 0.629 | 0.286 | 0.253 | 0.414 | 0.405 | 0.406 | 0.352 Table 2: The correlation between engagement scores and ground truth human judgment. Best scores are emboldened and second-best are underlined. We train EnDex and EnDex+NS 10 times and report the mean with * and ** indicating a stdev < 0.05 and < 0.03, respectively. EnDex-Best is the best score observed over the 10 runs. Compared to existing metrics, the EnDex-framework achieves SOTA correlation with human judgement on engagingness, leading by far on our newly proposed Red-Test dataset with more complex and longer texts than chitchats. ### 3.1 Experiment Set-up We test the performance of the EnDex metric on 5 golden evaluation sets that have turn-level labels. Among them, Better (Ghazarian et al., 2019), PredEeng-600 (Ghazarian et al., 2020) are annotated specifically for engagingness with high annotator agreement. Better samples are taken from human conversation, while half of PredEng-600 are chatbot generations. FED (Mehri and Eskenazi, 2020) annotates dialogue for 9 different dimensions, and we use their engagingness scores as target labels. GRADE (Huang et al., 2020) contains quality annotations for dialogue coherence, and we include it to test whether our model can also have good zero-shot performance on related tasks. Lastly, our own RED-Test is sourced from Reddit, contains discussions on various topics. A table of evaluation set statistics is provided in the Appendix3. ### 3.2 Ablation Study on Engaging Dimensions To test the robustness of the 4 engagingness dimensions 2.2 of EnDex, we conducted ablation study to train model using only signals from each of the 4 dimensions. We hypothesize that dimensions with high positive contribution towards final results should have very successful clustering of engaging and non-engaging samples by itself; so, if we train model on data clustered by such dimension, we can still get good performance models. We train five different models on different subsets of RED. All datasets included the same 40k negative (i.e. non-engaging) samples drawn according to our overall engagement score. However, the other 40k positive (i.e. engaging) samples were selected according to a particular dimension score (e.g. EE, AE, BE, and RE), except for EnDex, which is our aggregate score model. Figure 2 shows that all four dimensions correlate with engagingness to some degree, but RE, AE, and EE are especially effective. We also observe a synergistic effect of training on a composite score rather than any one dimension individually. The experiment highlights and corroborates the multi-dimensionality of engagingness previously reported in the literature (See et al., 2019). Overall, having an aggregate score is crucial for successful distant- supervised annotation of negative and positive examples. Figure 2: Ablation study of our four engagement dimensions. The EnDex model was trained on our aggregate engagingness score, while RE, EE, AE, and BE indicate models trains only on scores reflecting that particular dimension. ### 3.3 Comparison with Related Works We compare our EnDex metric, and heuristics-augmented EnDex+NS metric with five baselines. Three baselines are rule-based, including Random, information Specificity (See et al., 2019) that counts number of non-stopword tokens, and Inquisitiveness (Ghandeharioun et al., 2019) that examines question asking ability. We included them because in some dataset, rule-based system could work surprisingly well (Yeh et al., 2021). We selected USL-H (Phy et al., 2020) as a baseline because it is the top performing metric on the PredEng-600 and Fed engagingness evaluation sets Yeh et al. (2021). USL-H is designed to measure high-level dialog quality, including understandability, sensibleness, and likability; it trains 3 BERT- based (Devlin et al., 2019) classifiers for each component, and uses a composite score named USL-H for overall assessment. Pred_En (Ghazarian et al., 2020) uses BERT embedding plus MLP layer and train on ConvAI dataset (Dinan et al., 2019) to make engagement score predictions. PRED_EN (FT+DD) further finetunes the original PRED_EN metric on the DailyDialogue dataset, to get better results. Our model has two versions: EnDex is solely trained on human-reaction based data. +NS means non-engaging samples set is mixed with some rule-based negative samples, created by random insertion, random deletion, copying, and generic replies; The experiments in Table 2 demonstrate that our model achieves strong performance on 4 engagingness related datasets, and good correlation with one coherence dataset(Grade). EnDex surpasses Pred_En and USL-H by a large margin on two real human conversations, Better and Red-Test. USL-H still leads in PredEng-600, and EnDex+NS’s best model is a close second. Yeh et al. (2021) shows achieving high score on FED-eng is challenging, with no one surpassing 0.3 spearman in 12 tested metrics. A strong rule-based question detection algorithm surprisingly claims the highest result, and EnDex a close second. We find that training solely with human reaction distant supervision signals suffices for building competitive models on par or even surpassing mainstream metrics, and it shows better generalization capability in new domains, which seems to echo recent success on modeling human preferences via upvotes in Reddits (Gao et al., 2020). ## 4 Conclusion This paper proposes the first human reaction based model, EnDex, to evaluate dialogue engagingness, and curates an 80k Reddit Engagement Dataset (Red) using a novel distant-supervision framework. The success of EnDex demonstrates the validity of training automatic metrics with human reaction signals, offering a strong complement to a synthetic negative sampling approach. We also release an off-the-shelf EnDex model, and a large scale dataset to facilitate future research. ## Limitation One limitation is that we only curated data for turn-level dialogue. Multi- turn dialogues could also be useful, but it was computationally infeasible to interactively query Reddit for entire threads of conversation. Future work can explore this direction to produce dialogue-level and system-level engagingness metrics. We also haven’t fully explore our model’s performance on non-dialogue domains, such as on story or creative generations. The training data distribution from the Reddit corpus is diverse enough that it could potentially achieve good performance in non-dialogue settings. A valuable direction of future work is to adapt our method for more general engagingness, or another evaluation metric for open-domain generation. ## Ethics A caveat of using framing our approach around human attention is that not all texts attracting high attention are good and ethical. Since being engaging often carries a positive connotation, we made a deliberate design decision to mitigate forms of _negative engagement_ in our metric. For example, we assign lower scores to samples flagged by Reddit as controversial, and our behavioral engagement dimension subtracts downvotes from upvotes to punish negative, biased (Liu et al., 2021), and aggressive comments. Moreover, we implemented our emotional engagement algorithm to reward posts with positive emotional replies and punish posts that prompt negative emotions. Future may try to account for the darker aspects of engagingness into our framework and improve the EnDex metric to differentiate between positive and negative engagement. Human annotations for RED-Test were obtained via Amazon Mechanical Turks. We filtered out toxic samples to reduce the likelihood of offensive content and paid $0.30 USD per instance for an expected hourly wage of $20 USD. ## Acknowledgement We want to give special thanks to Sarik Ghazarian for great advice and help with the baseline model; we also appreciate the brainstorming session with Hongyuan Yang and the suggestion to use z-score for sampling from Zeyi Wang. ## References * Attfield et al. (2011) Simon Attfield, Gabriella Kazai, and Benjamin Piwowarski. 2011. Towards a science of user engagement ( position paper ). * Chen et al. (2005) Mark Chen, Beth E. Kolko, Elisabeth Cuddihy, and Eliana Medina. 2005. Modeling and measuring engagement in computer games. In _DiGRA Conference_. * Demszky et al. (2020) Dorottya Demszky, Dana Movshovitz-Attias, Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. 2020. GoEmotions: A dataset of fine-grained emotions. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_ , pages 4040–4054, Online. Association for Computational Linguistics. * Devlin et al. (2019) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. _ArXiv_ , abs/1810.04805. * Dinan et al. (2019) Emily Dinan, Varvara Logacheva, Valentin Malykh, Alexander H. Miller, Kurt Shuster, Jack Urbanek, Douwe Kiela, Arthur D. Szlam, Iulian Serban, Ryan Lowe, Shrimai Prabhumoye, Alan W. Black, Alexander I. Rudnicky, Jason Williams, Joelle Pineau, Mikhail S. Burtsev, and Jason Weston. 2019. The second conversational intelligence challenge (convai2). _ArXiv preprint_ , abs/1902.00098. * Gao et al. (2020) Xiang Gao, Yizhe Zhang, Michel Galley, Chris Brockett, and Bill Dolan. 2020. Dialogue response ranking training with large-scale human feedback data. _ArXiv_ , abs/2009.06978. * Ghandeharioun et al. (2019) Asma Ghandeharioun, Judy Hanwen Shen, Natasha Jaques, Craig Ferguson, Noah Jones, Àgata Lapedriza, and Rosalind W. Picard. 2019. Approximating interactive human evaluation with self-play for open-domain dialog systems. In _Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada_ , pages 13658–13669. * Ghazarian et al. (2019) Sarik Ghazarian, Johnny Wei, Aram Galstyan, and Nanyun Peng. 2019. Better automatic evaluation of open-domain dialogue systems with contextualized embeddings. In _Proceedings of the Workshop on Methods for Optimizing and Evaluating Neural Language Generation_ , pages 82–89, Minneapolis, Minnesota. Association for Computational Linguistics. * Ghazarian et al. (2020) Sarik Ghazarian, Ralph M. Weischedel, Aram Galstyan, and Nanyun Peng. 2020. Predictive engagement: An efficient metric for automatic evaluation of open-domain dialogue systems. In _The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020_ , pages 7789–7796. AAAI Press. * Ghazarian et al. (2022) Sarik Ghazarian, Nuan Wen, A. G. Galstyan, and Nanyun Peng. 2022. Deam: Dialogue coherence evaluation using amr-based semantic manipulations. _ArXiv preprint_ , abs/2203.09711. * Hanu and Unitary team (2020) Laura Hanu and Unitary team. 2020. Detoxify. Github. https://github.com/unitaryai/detoxify. * Huang et al. (2020) Lishan Huang, Zheng Ye, Jinghui Qin, Liang Lin, and Xiaodan Liang. 2020. GRADE: Automatic graph-enhanced coherence metric for evaluating open-domain dialogue systems. In _Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)_ , pages 9230–9240, Online. Association for Computational Linguistics. * Lan et al. (2020) Tian Lan, Xian-Ling Mao, Wei Wei, Xiaoyan Gao, and Heyan Huang. 2020. Pone: A novel automatic evaluation metric for open-domain generative dialogue systems. _ArXiv preprint_ , abs/2004.02399. * Li et al. (2016) Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In _Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies_ , pages 110–119, San Diego, California. Association for Computational Linguistics. * Li et al. (2017) Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. DailyDialog: A manually labelled multi-turn dialogue dataset. In _Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers)_ , pages 986–995, Taipei, Taiwan. Asian Federation of Natural Language Processing. * Liu et al. (2021) Ruibo Liu, Chenyan Jia, Jason Wei, Guangxuan Xu, Lili Wang, and Soroush Vosoughi. 2021. Mitigating political bias in language models through reinforced calibration. _Proceedings of the AAAI Conference on Artificial Intelligence_ , 35(17):14857–14866. * Liu et al. (2019) Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. _ArXiv preprint_ , abs/1907.11692. * Ma (2018) Xiaojuan Ma. 2018. Towards human-engaged AI. In _Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden_ , pages 5682–5686. ijcai.org. * Mcmahan (2003) Alison Mcmahan. 2003. Immersion, engagement, and presence: A method for analyzing 3-d video games. _The Video Game Theory Reader_ , pages 67–86. * Mehri and Eskenazi (2020) Shikib Mehri and Maxine Eskenazi. 2020. Unsupervised evaluation of interactive dialog with DialoGPT. In _Proceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue_ , pages 225–235, 1st virtual meeting. Association for Computational Linguistics. * Miller et al. (2017) Alexander Miller, Will Feng, Dhruv Batra, Antoine Bordes, Adam Fisch, Jiasen Lu, Devi Parikh, and Jason Weston. 2017. ParlAI: A dialog research software platform. In _Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing: System Demonstrations_ , pages 79–84, Copenhagen, Denmark. Association for Computational Linguistics. * Phy et al. (2020) Vitou Phy, Yang Zhao, and Akiko Aizawa. 2020. Deconstruct to reconstruct a configurable evaluation metric for open-domain dialogue systems. In _Proceedings of the 28th International Conference on Computational Linguistics_ , pages 4164–4178, Barcelona, Spain (Online). International Committee on Computational Linguistics. * Roller et al. (2021) Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Eric Michael Smith, Y-Lan Boureau, and Jason Weston. 2021. Recipes for building an open-domain chatbot. In _Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume_ , pages 300–325, Online. Association for Computational Linguistics. * Sato et al. (2020) Shiki Sato, Reina Akama, Hiroki Ouchi, Jun Suzuki, and Kentaro Inui. 2020. Evaluating dialogue generation systems via response selection. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_ , pages 593–599, Online. Association for Computational Linguistics. * Savin-Baden et al. (2014) Maggi Savin-Baden, Gemma Tombs, Roy Bhakta, and David J. H. Burden. 2014. Students’ experiences of emotional connection with pedagogical agents. * Schoenau-Fog (2011) Henrik Schoenau-Fog. 2011. The player engagement process - an exploration of continuation desire in digital games. In _DiGRA Conference_. * See et al. (2019) Abigail See, Stephen Roller, Douwe Kiela, and Jason Weston. 2019. What makes a good conversation? how controllable attributes affect human judgments. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)_ , pages 1702–1723, Minneapolis, Minnesota. Association for Computational Linguistics. * Sellam et al. (2020) Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_ , pages 7881–7892, Online. Association for Computational Linguistics. * Silpasuwanchai et al. (2016) Chaklam Silpasuwanchai, Xiaojuan Ma, Hiroaki Shigemasu, and Xiangshi Ren. 2016. Developing a comprehensive engagement framework of gamification for reflective learning. pages 459–472. * Tao et al. (2018) Chongyang Tao, Lili Mou, Dongyan Zhao, and Rui Yan. 2018. RUBER: an unsupervised method for automatic evaluation of open-domain dialog systems. In _Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018_ , pages 722–729. AAAI Press. * Yeh et al. (2021) Yi-Ting Yeh, Maxine Eskénazi, and Shikib Mehri. 2021. A comprehensive assessment of dialog evaluation metrics. _ArXiv preprint_ , abs/2106.03706. * Zhang et al. (2018) Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In _Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , pages 2204–2213, Melbourne, Australia. Association for Computational Linguistics. * Zhang et al. (2020) Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020\. Bertscore: Evaluating text generation with BERT. In _8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020_. OpenReview.net. ## Appendix A Appendix ### A.1 RED Processing Steps Our Reddit data is downloaded from Pushshift.io 444https://files.pushshift.io/reddit/comments/, and we processed approximately 5M data to curate an 80k sample RED dataset. We deleted posts that do not have an immediate parent thread, because we need pair turn-level data. Our preprocessing removes non-conversational data, such as posts including & gt(reply to symbol). We also removed explicitly toxic data filtered by Detoxify (Hanu and Unitary team, 2020). We also applied a key data processing trick to reduce noisy signals – the exposure variable. It helps measure the amount exposure each post receives to help normalize its upvote/reply score. We reward posts that are in low- exposure, unpopular threads, while penalizing posts in high-exposure, popular threads, because high upvotes and replies in popular threads may be more due to exposure than true engagingness. After computing the normalized score given in Equation 8, we also apply another z-score to normalize the final EnDex score according to standard deviation, so that we can easily sample our data from it. A score with higher standard deviation will imply a higher probability that the sample is engaging according to our metric. We apply a cut-off to sample high probability engaging and non-engaging samples, and arrive at the RED dataset. ### A.2 Model training details Our RED-Test set contains 300 human labeled data. The train validation split during training is 0.8 and 0.2. We used 4 Nvidia A6000 GPUs for training, and 1 Nvidia A6000 GPU for inference. The average runtime for training one model is 2 minutes per epoch, and inference time is in seconds, negligible for the testset. The estimated energy cost per model is, assuming per second gpu energy cost of 245W: 245W*4*60 = 58800 per model. We trained our model for 2 epochs, and only save the best checkpoints, with learning rate of 5e-5 with no extensive hyperparameter search. We used specificity and question examination inspired from (See et al., 2019); USL-H (Phy et al., 2020) and PredEn is taken from a GitHub repo555https://github.com/exe1023/DialEvalMetrics and modified to use a local bert-base-uncased since the original ‘bert-as-service‘ code no longer functions. The Formula for calculating each dimension is given in the following: * • Reply Engagement; The raw Reply Engagement score(re) is just the reply counts of a post. Popularity value adjusted RE score is given by PVRE in equation (6), where $M_{pv}$ and $M_{re}$ are the median of popularity value and reply counts in the entire dataset. Please refer to equation (1) for calculation of the popularity value. $\textrm{PVRE}(re)=re+\frac{M_{pv}}{M_{\textrm{re}}}*\frac{re}{\textrm{PV(re)}}*re$ (5) * • Behavioral Engagement; The raw BE score of a certain post(be) is obtained by subtracting downvotes to upvotes and set to 0 if given controversy flag. Popularity value adjusted BE score is given by PVBE in equation (LABEL:eq:pvbe), where $M_{pv}$ and $M_{be}$ are the median of popularity value and raw BE score in the entire dataset. Please refer to equation (1) for calculation of the popularity value. $\textrm{PVBE}(be)=be+\frac{M_{pv}}{M_{\textrm{be}}}*\frac{be}{\textrm{PV(be)}}*be$ (6) * • Attentional Engagement; It is calculated using maximum information specificity, or the the maximum number of non-stopword tokens in a post’s replies, and whether its children posts are editted; t is the maximum reply specificity, and e stands for number of edited replies. $\textrm{AE}(x)=t+10*e$ (7) * • Emotional Engagement; The EE score is the aggregate probability for all positive emotion categories, produced by the go-emotion classifier (Demszky et al., 2020). ### A.3 Submodular Normalization and z-score Clustering After we obtained the sub-dimension scores, we want to aggregate them into a single normalized EnDex Score, and lastly cluster them into positive and negative sets to train a binary classifier. The formulas are list in the following: $\textsc{EnDex}(x)=\Bigg{(}\displaystyle\sum_{n=1}^{3}w_{i}*\frac{x_{i}}{x_{i}+\alpha_{i}}\Bigg{)}+w_{\textrm{EE}}*x_{\textrm{EE}},$ (8) $N$ is the normalized score for sample $x$, $x_{i}$ is $x$’s raw score on dimension $i$, where $i\in\\{RE,BE,AE\\}$; $\alpha_{i}$ is the median of $i$-th dimension; $w_{i}$ is the weight for $i$ dimension; $w_{\textrm{EE}}$ is the weight for EE dimension. The weight can be tuned for your own usage of RED; A confidence threshold $\kappa$(ours is 1) needs to be picked, which means that we regard samples that fall between $\kappa$ standard deviation from mean as uncertain, and are thus discarded. And we cluster positive and negative samples using the following equation (9). $Polarity(x)=\begin{cases}1&\text{if }z\\_score(x)>\kappa\\\ 0&\text{if }z\\_score(x)<-\kappa\end{cases}$ (9) Figure 3: The screenshot of the task description of our Amazon MTurk questionnaire. We have prepared instructions, demonstrations, and proper warning of offensive content. Figure 4: The screenshot of the labeling area of our Amazon MTurk questionnaire. Each pair will be labelled by three annotators. ### A.4 Annotation data and test data We performed annotation on Amazon Mechanical Turk, and selected annotators based in the United State; in implemented restrictions to annotator with 98% pass rate. We give four examples and clear instruction for the task carried out. A screenshot of our annotation interface is provided below. Table 3 gives summary of the evaluation datasets we used. Dataset | # of Samples | Context Length | Response Length | Source | Agreement Rate ---|---|---|---|---|--- BETTER | 297 | 6 | 8 | Human | N/A PREDENG-600 | 600 | 12 | 14 | Human+Bot | 0.51 FED-ENG | 261 | 26 | 12 | Human+Bot | N/A RED-TEST | 150 | 16 | 17 | Human | 0.34 GRADE | 150 | 12 | 14 | Human | N/A Table 3: Dataset Statistics for the 5 golden evaluation sets, with number of samples, context-length, response length, and if applicable, inter-annotator agreement rate.
Quantum many-body thermal machines enabled by atom-atom correlations R. S. Watson1 $\dagger$ and K. V. Kheruntsyan1$\star$ 1 School of Mathematics and Physics, University of Queensland, Brisbane, Queensland 4072, Australia <EMAIL_ADDRESS> ⋆<EMAIL_ADDRESS> ## Abstract Particle-particle correlations, characterized by Glauber’s second-order correlation function, play an important role in the understanding of various phenomena in radio and optical astronomy, quantum and atom optics, particle physics, condensed matter physics, and quantum many-body theory. However, the relevance of such correlations to quantum thermodynamics has so far remained illusive. Here, we propose and investigate a class of quantum many-body thermal machines whose operation is directly enabled by second-order atom-atom correlations in an ultracold atomic gas. More specifically, we study quantum thermal machines that operate in a sudden interaction-quench Otto cycle and utilize a one-dimensional Lieb-Liniger gas of repulsively interacting bosons as the working fluid. The atom-atom correlations in such a gas are different to those of a classical ideal gas, and are a result of the interplay between interparticle interactions, quantum statistics, and thermal fluctuations. We show that operating these thermal machines in the intended regimes, such as a heat engine, refrigerator, thermal accelerator, or heater, would be impossible without such atom-atom correlations. Our results constitute a step forward in the design of conceptually new quantum thermodynamic devices which take advantage of uniquely quantum resources such as quantum coherence, correlations, and entanglement. ###### Contents 1. 1 Introduction 2. 2 Interaction-driven Otto heat engine. 3. 3 Work from second-order Glauber correlations. 4. 4 Interaction-driven Otto accelerator, heater, and refrigerator. 1. 4.1 Accelerator [A] 2. 4.2 Heater [H] 3. 4.3 Refrigerator [R] 5. 5 Summary and outlook. 6. A The Lieb-Liniger model for the 1D Bose gas 7. B Transverse Otto cycle 8. C Instantaneity of the sudden quench 9. D Glauber’s second order correlation function 10. E Exact thermodynamic Bethe ansatz results 11. F Maximum efficiency and maximum work 12. G Thermal operation regimes of other QTM’s ## 1 Introduction Quantum thermal machines (QTM), such as quantum heat engines (QHE), refrigerators, and quantum batteries, are central in the theoretical and experimental development of the emerging field of quantum thermodynamics [1, 2]. Their primary utility is to explore the fundamental laws of thermodynamics in the quantum realm and to demonstrate possible advantages gained by utilising quantum resources. Accordingly, QTM’s are expected to play a similar role in the development of quantum technologies as classical heat engines played in fostering scientific advances during the Industrial Revolution. In the past decade, progress in the control over quantum platforms, such as single ions [3, 4], nitrogen vacancy centers [5], and single-atom impurities immersed in an ultra-cold atomic bath [6], have led to the realization of single-particle QHE’s. Such single-particle QHE’s represent the ultimate limit in the realization of an ‘infinitesimal machine’ [7]. However, to utilize the breadth of quantum resources available, one must move beyond single-particle systems—to engines that utilize interacting many- particle systems. Such QHE’s are uniquely positioned to take advantage of quantum resources, such as entanglement [8, 9, 10, 11], correlations [12, 13, 14], or quantum coherence [15, 16, 17, 18, 19, 20], to enhance the performance of classical heat engines [21] or perform entirely new tasks that would be impossible classically [22]. In particular, control over inter-particle interactions allows for the creation of strictly many-body QHE’s [21, 23, 24, 25, 26], which have recently been realized in the laboratory [27, 28]. These very recent experimental developments underscore the need for further studies of thermodynamic processes in the context of interacting quantum many-body thermal machines. Here, we propose a quantum many-body Otto heat engine—as well as related thermal machines including the Otto refrigerator, thermal accelerator, and heater—using a uniform one-dimensional (1D) Bose gas with repulsive contact interactions as the working fluid. In the proposed Otto cycles, the unitary work strokes are driven by a sudden quench of the interaction strength. We demonstrate how the thermodynamic performance, in particular net work and efficiency, of these many-body QTM’s can be calculated through the experimentally measurable atom-atom local pair correlation [29, 30, 31]. The atom-atom local correlation, $g^{(2)}(0)$, is described by the second-order Glauber correlation function $g^{(2)}(r)$ at zero interparticle separation (i.e., when $r=0$, where $r=|x-x^{\prime}|$ is the distance between the two particles with positions $x$ and $x^{\prime}$) and characterizes the probability of pairs of atoms to be found at the same point, relative to uncorrelated atoms. The benefits of using the 1D Bose gas as the working fluid in the proposed Otto cycles is that the underlying theoretical model—the Lieb-Liniger model—is exactly solvable in the uniform limit via the Yang-Yang thermodynamic Bethe ansatz (TBA) [32, 33, 34], in addition to being experimentally realizable using ultracold atomic gases confined to highly anisotropic traps [35, 36, 37]. This offers unique opportunities for gaining physical insights into the performance of such Otto QTM’s as a tractable and testable quantum many-body problem. Additionally, the Lieb-Liniger gas has a rich phase diagram spanning several nontrivial regimes [29, 30], from the weakly interacting quasicondensate [38, 39] through to the strongly interacting Tonks-Girardeau regime of fermionization [40, 41]. The atom-atom correlation within these regimes takes on a wide range of values between $0<g^{(2)}(0)<2$ [29, 30, 42], depending on the temperature and interaction strength, which aids the operation of the proposed Otto cycles under a variety of conditions. We evaluate the performance of the 1D Bose gas Otto QTM’s, but we emphasise that the broad conclusions arrived at here are not limited to the Lieb-Liniger model. A related interaction-driven Otto engine cycle with a uniform 1D Bose gas as the working fluid was previously studied by Chen et al. in Ref. [25] in the opposite limit of a quasi-static, or isentropic (rather than sudden quench), work strokes. In this case, the performance of the engine cannot be expressed merely in terms of Glauber’s second-order correlation function, as both the kinetic and interaction energies evolve and change during the interaction quench. Nevertheless, the performance of such an isentropic Otto engine could still be evaluated analytically in the low-temperature regime, owing to the known thermodynamic properties of the system using the Tomonaga-Luttinger liquid approach. ## 2 Interaction-driven Otto heat engine. In a uniform 1D Bose gas, described by the integrable Lieb-Liniger model [32] (see Appendix A), the interatomic interaction strength $\chi$ can be expressed in terms of the harmonic trap frequency $\omega_{\perp}$ in the tightly confined (transverse) direction and the 3D $s$-wave scattering length $a_{s}$ as $\chi\simeq 2\hbar\omega_{\perp}a_{s}$ [43]. Accordingly, changing the interaction strength $\chi$ may be achieved by either tuning the external trapping potential that controls the transverse confinement $\omega_{\perp}$ or by changing the scattering length $a_{s}$ by means of a magnetic Feshbach resonance [44]. The former option leads to a volume change of the gas (i.e. transverse expansion or compression), and hence can be thought of as analogous to mechanical work in the conventional Otto cycle. However, changing $\chi$ via a change of the scattering length leads to identical results, which then justifies our referral to the engine cycle as the Otto cycle [45, 46, 24, 25, 27, 6, 47, 48, 49, 50] regardless of the means of tuning the interaction strength (for further discussion, see (see Appendix B)). We emphasize, however, that the dynamics of the quantum Otto cycle explored here are strictly longitudinal, with the gas always remaining in its transverse ground state. Figure 1: An interaction-driven quantum many-body Otto engine cycle, operating between two interaction strengths, $\chi_{c}$ and $\chi_{h}$. Unitary work strokes (BC and DA) are shown in black, while non-unitary thermalization strokes (AB and CD) are color-coded to the cold (blue) and hot (red) reservoirs at temperatures $T_{c}$ and $T_{h}$, respectively. The interaction-driven Otto engine cycle, which we thus consider, consists of four strokes (see Fig. 1): * (1) Thermalization with hot reservoir, A$\to$B: the working fluid, consisting of $N$ total atoms at interaction strength $\chi_{h}$, is connected to a hot ($h$) reservoir at temperature $T_{h}$, where it is left to equilibrate, taking in heat $Q_{1}\\!=\\!\langle\hat{H}\rangle_{\textbf{B}}\\!-\\!\langle\hat{H}\rangle_{\textbf{A}}\\!>\\!0$, which is to be partially converted into beneficial work in the subsequent stroke. Here $\hat{H}$ is the system Hamiltonian (see Appendix A), and $\langle\hat{H}\rangle_{\textbf{j}}$ is its expectation value, i.e., the total energy of the system, in state $\textbf{j}=\\{\textbf{A,B,C,D}\\}$. * (2) Unitary expansion, B$\to$C: the working fluid, now in a thermal equilibrium state at temperature $T_{h}$, is decoupled from the hot reservoir and has its interaction strength quenched from $\chi_{h}$ to $\chi_{c}\\!<\\!\chi_{h}$, generating beneficial work $W_{1}\\!=\\!\langle\hat{H}\rangle_{\textbf{C}}\\!-\\!\langle\hat{H}\rangle_{\textbf{B}}\\!<\\!0$ done by the fluid. * (3) Thermalization with cold reservoir, C$\to$D: the working fluid is connected to a cold ($c$) reservoir at temperature $T_{c}<T_{h}$ and allowed to equilibrate at constant interaction strength $\chi_{c}$ while dumping energy in the form of heat $Q_{2}\\!=\\!\langle\hat{H}\rangle_{\textbf{D}}\\!-\\!\langle\hat{H}\rangle_{\textbf{C}}\\!<\\!0$ into the cold reservoir. * (4) Unitary compression, D$\to$A: disconnected from the cold reservoir, the working fluid has its interaction strength quenched from $\chi_{c}\\!\to\\!\chi_{h}$, with work $W_{2}\\!=\\!\langle\hat{H}\rangle_{\textbf{A}}\\!-\\!\langle\hat{H}\rangle_{\textbf{D}}\\!>\\!0$ done on the fluid, and the system returning to the initial state of the overall cycle. Such an engine cycle generates net beneficial work (done by the fluid) if the total work $W\\!=\\!W_{1}\\!+\\!W_{2}\\!<\\!0$, i.e., if $|W_{1}|>W_{2}$ (or $Q_{1}>|Q_{2}|$), with efficiency $\eta=-\frac{W}{Q_{1}}=1-\frac{|Q_{2}|}{Q_{1}},$ (1) where we used the conservation of energy $W+Q=0$, with $Q=Q_{1}+Q_{2}$ being the total heat [51]. ## 3 Work from second-order Glauber correlations. In this work, we specifically consider a sudden or instantaneous quench of the interaction strength $\chi$ in the unitary strokes (2) and (4). (For a discussion of “instantaneity” of the sudden quench, see Appendix C). Under a sudden interaction quench, the initial ($i$) and final ($f$) expectation values over field operators in the system Hamiltonian, i.e., the expectation values before and after the quench, remain unchanged as the system did not have sufficient time to evolve into a new state. Hence, the only contribution to the difference in total energy before and after the quench, $\langle\hat{H}\rangle_{f}-\langle\hat{H}\rangle_{i}$, comes from the difference between the interaction terms, $\frac{1}{2}\chi_{f}\int dz\langle\hat{\Psi}^{\dagger}\hat{\Psi}^{\dagger}\hat{\Psi}\hat{\Psi}\rangle_{f}-\frac{1}{2}\chi_{i}\int dz\langle\hat{\Psi}^{\dagger}\hat{\Psi}^{\dagger}\hat{\Psi}\hat{\Psi}\rangle_{i}$, where $\langle\hat{\Psi}^{\dagger}\hat{\Psi}^{\dagger}\hat{\Psi}\hat{\Psi}\rangle_{f}=\langle\hat{\Psi}^{\dagger}\hat{\Psi}^{\dagger}\hat{\Psi}\hat{\Psi}\rangle_{i}$ in a sudden quench, and $\hat{\Psi}^{\dagger}(z)$ and $\hat{\Psi}(z)$ represent the field creation and annihilation operators. Accordingly, the energy difference can be expressed as $\langle\hat{H}\rangle_{f}-\langle\hat{H}\rangle_{i}\\!=\\!\frac{1}{2}(\chi_{f}\\!-\\!\chi_{i})\overline{G^{(2)}_{i}}$, where we have defined the total (integrated) second-order correlation of the thermal equilibrium state $\overline{G^{(2)}_{i}}\\!=\\!\int dz\langle\hat{\Psi}^{\dagger}\hat{\Psi}^{\dagger}\hat{\Psi}\hat{\Psi}\rangle_{i}$ [30]. Identifying the $i$ and $f$ states as points B (hot, $h$) and C, or as D (cold, $c$) and A in the digram of Fig. 1, the net work of the Otto heat engine can be expressed as $W=-\frac{1}{2}(\chi_{h}-\chi_{c})\left(\overline{G^{(2)}_{h}}-\overline{G^{(2)}_{c}}\right),$ (2) whereas the efficiency, Eq. (1), of the engine as $\eta=1-\frac{\langle\hat{H}\rangle_{h}-\langle\hat{H}\rangle_{c}-\frac{1}{2}\left(\chi_{h}-\chi_{c}\right)\overline{G^{(2)}_{h}}}{\langle\hat{H}\rangle_{h}-\langle\hat{H}\rangle_{c}-\frac{1}{2}\left(\chi_{h}-\chi_{c}\right)\overline{G^{(2)}_{c}}}.$ (3) These equations allow for evaluation of the performance of the interaction- driven Otto engine under a sudden quench through solely the _equilibrium_ properties of the gas as all relevant expectation values here are with respect to $h$ (B) or $c$ (D) states. The equilibrium phase diagram of the uniform 1D Bose gas consists of several distinct regimes (connected by smooth crossovers) that can be characterized by the normalized same-point atom-atom correlation function $g^{(2)}(0)$ [29]. This local pair correlation is a thermodynamic quantity that can be calculated from the exact TBA, as well as via approximate analytic methods in six asymptotic regimes; it is shown in Fig. 5 in Appendix D. The $g^{(2)}(0)$ function and the different regimes of the 1D Bose gas can be parameterized by dimensionless interaction strength, $\gamma\\!=\\!m\chi/\hbar^{2}\rho$, and dimensionless temperature, $\tau\\!=\\!2mk_{B}T/\hbar^{2}\rho^{2}$, where $m$ is the boson mass and $\rho\\!=\\!N/L$ is the 1D density for $N$ atoms in a system of length $L$. For a uniform 1D Bose gas, the total correlation in the hot or cold thermal equilibrium state may be expressed as $\overline{G^{(2)}_{h(c)}}\\!=\\!N\rho g^{(2)}_{h(c)}(0)$ (see Appendix D). Combining this with Eq. (2), the net work per particle can be expressed as $\frac{W}{N}=-\frac{\hbar^{2}\rho^{2}}{2m}(\gamma_{h}-\gamma_{c})\left(g^{(2)}_{h}(0)-g^{(2)}_{c}(0)\right),$ (4) meaning the net work is directly proportional to the difference between atom- atom correlations of the 1D Bose gas in the hot and cold thermal equilibrium states. This simple relationship between thermodynamic work and Glauber second-order correlation function represents the key result of this work. From Eq. (4), and given that $\gamma_{h}$ is always larger than $\gamma_{c}$ (as $\chi_{h}>\chi_{c}$; see Fig. 1), we see that if the local pair correlations did not depend on the respective interaction strengths and temperatures, i.e. if they were the same, $g^{(2)}_{h}(0)\\!=\\!g^{(2)}_{c}(0)$, then the net work per particle would vanish. We therefore conclude that extracting net work ($W\\!<\\!0$) from this Otto cycle, and hence operating it as a heat engine, can only be enabled by atom-atom correlations; more specifically, the only way to extract net work is to have $g^{(2)}_{h}(0)\\!>\\!g^{(2)}_{c}(0)$ (see Fig. 5 (c) in Appendix D for an illustration). Figure 2: Performance of the interaction driven quantum Otto heat engine. Columns (a)–(c) demonstrate net work, $|W|$, and efficiency, $\eta$, as a function of the ratio of interaction strength, $\gamma_{h}/\gamma_{c}$, and temperature, $\tau_{h}/\tau_{c}$, between the hot ($h$) and cold ($c$) thermal equilibrium states. Regions corresponding to $W\\!>\\!0$, here colored grey, are outside the engine operation regime and are considered further in Figs. 3 and 4 below. The example of panel (a) is for $\gamma_{c}\\!=\\!10^{-3}$ and $\tau_{c}\\!=\\!10^{-2}$, where $\gamma_{h}/\gamma_{c}$ and $\tau_{h}/\tau_{c}$ explores the parameter range within the region II of the equilibrium phase diagram of Fig. 5 (b) of Appendix D . Similarly, panel (b) explores region IV, with $\gamma_{c}\\!=\\!1$ and $\tau\\!=\\!10$, whereas (c) explores region VI, with $\gamma_{c}\\!=\\!10$ and $\tau\\!=\\!1$. Net work, Eq. (4), and efficiency, Eq. (3), of this quantum Otto engine, calculated for simplicity using analytic approximations to the atom-atom correlation function and total energy [29, 52], are shown in Fig. 2 as a function of the ratio of interaction strengths, $\gamma_{h}/\gamma_{c}$, and temperatures, $\tau_{h}/\tau_{c}$, for three of the six asymptotic regimes (for results outside of the regimes of analytic approximation, see Appendix E). Notably, in this Otto engine cycle, for any fixed value of the temperature ratio, the interaction strength quench corresponding to maximum net work is approximately the same as that providing maximum efficiency; this occurs as, to first order, the heat intake $Q_{1}$ varies slowly with $\gamma_{h}/\gamma_{c}$, meaning $\eta\\!\propto\\!W$ (see Appendix F). The observed increase of net work and efficiency in all regimes under large temperature ratio may be attributed to the fact that the local correlation of the hot thermal state in Eq. (4) is always a monotonically increasing function of $\tau$. However, as the correlation function is also monotonically decreasing under $\gamma_{h}$, this results in no extractable net work under sufficiently large interaction strength ratios for any given temperature ratio. At a glance, one may conclude that the net work, which is enabled through the $g^{(2)}(0)$ correlation function, is maximized under the largest possible difference in correlation function, i.e. $g^{(2)}_{h}(0)\\!-\\!g^{(2)}_{c}(0)\\!\simeq\\!2$. However, to achieve this, while also guaranteeing that $\gamma_{h}\\!>\\!\gamma_{c}$, would require an unrealistically high (from practical point of view) temperature ratio to operate between regimes VI and IV, shown in Fig. 5 (b) of Appendix D. Rather, we observe that, while the $g^{(2)}(0)$ correlation function is responsible for enabling operation as a heat engine, the _magnitude_ of net work is governed more strongly by the difference in the interaction strengths, $\gamma_{h}-\gamma_{c}$, which is more susceptible to large variations in practice using, e.g., magnetic Feshbach tuning of the $s$-wave scattering length. Consequently, it is in the weakly interacting ($\gamma\ll 1$) region II, shown in Fig. 2 (a), where $\gamma_{h}-\gamma_{c}$ is small, that we observe the lowest magnitude of net work. In comparison, the magnitude of net work is largest in regime IV, shown in Fig. 2 (b), where $\gamma\\!\sim\\!1$ and hence the difference $\gamma_{h}\\!-\\!\gamma_{c}$ can also be on the order of one. The same considerations apply to the strongly interacting ($\gamma\\!\gg\\!1$) regime VI, where one can operate under the largest magnitudes of interaction strengths; however, in this regime the net work is diminished due to the vanishing of correlation itself ($g^{(2)}\\!\ll\\!1$) arising from the effect of fermionization [40]. In contrast to these observations, the efficiency of the engine, Eq. (3), is inversely dependent on the total energy of the thermal states, which is minimal in the weakly interacting low temperature regime II, which thus has the largest efficiency. ## 4 Interaction-driven Otto accelerator, heater, and refrigerator. Glauber’s $g^{(2)}(0)$ correlation function is inherently dependent on the interaction strength and temperature [29]. This implies that the condition for the Otto cycle to operate as a heat engine, $g^{(2)}_{c}(0)\\!<\\!g^{(2)}_{h}(0)$, where $\gamma_{c}\\!<\\!\gamma_{h}$, cannot hold under large quenches of interaction strength as the gas becomes increasingly fermionized (i.e., $g_{h}^{(2)}(0)\\!\to\\!0$) in the limit $\gamma\\!\to\\!\infty$ [40]. However, beyond the heat engine [E] operation regime, a further three QTM’s are thermodynamically allowed [53], namely, the accelerator [A], heater [H], and refrigerator [R]. For these QTM’s, one may define a coefficient of performance (CoP) according to [54]: $\mathrm{CoP[QTM]}=\frac{\mathrm{benefit\,of\,operation}}{\mathrm{cost\,of\,operation}}.$ (5) We now discuss the operating conditions and the CoP’s of these three additional QTM’s in detail. ### 4.1 Accelerator [A] Figure 3: Energy flow diagrams and schematics of the interaction driven Otto cycles in different regimes: (a) heat engine [E], (b) accelerator [A], (c) heater [H], and (d) refrigerator [R]. In all four regimes, the difference between the energies of the hot and cold thermal state (B and D) are the same, however the energies of the resulting state after the interaction driven work strokes can be different, leading to these four different physically valid outcomes. Panel (d) shows the layout of how the operating regimes of these different QTM’s depend on the ratio of interaction strengths, $\gamma_{h}/\gamma_{c}$, and temperatures, $\tau_{h}/\tau_{c}$, of the hot and cold thermal states in same the three asymptotic regimes (II, IV, and VI) as in Fig. 2. The respective coefficients of performance of these QTM’s within these operating boundaries are shown in Fig. 4. Figure 4: Coefficients of performance (CoP) of QTMs in the accelerator [A], heater [H], and refrigerator [R] regimes within the respective parameter ranges II, IV, and VI shown in Fig. 3 (e). The conditions of operating the Otto cycle as a thermal accelerator are given by: $W\\!>\\!0$, $Q_{1}\\!>\\!0$, $Q_{2}\\!>\\!0$, with $|Q_{2}|>|Q_{1}|$, where $W\\!=\\!0$ defines the border between the heat engine and the accelerator, whereas $Q_{1}\\!=\\!0$ defines the border between the accelerator and the next QTM, the heater; see Fig. 3, where we show the simplified schematics of all additional QTM’s compared to the heat engine from Fig. 1, which we repeat here in panel (a) for comparison. The thermal accelerator [53], shown in panel (b) of Fig. 3, enhances the natural flow of heat, taken into the working fluid from the hot reservoir, $Q_{1}$, and transferred to the cold reservoir, $Q_{2}$, which is aided by the working fluid through the investment of net mechanical work, $W>0$, in the process. According to Eq. (5), its CoP is given by: $\mathrm{CoP[A]}=-\frac{Q_{2}}{W}=1+\frac{Q_{1}}{W}>1,$ (6) where, $\displaystyle Q_{2}$ $\displaystyle=-\langle\hat{H}\rangle_{h}+\langle\hat{H}\rangle_{c}+\frac{N\hbar^{2}\rho^{2}}{2m}(\gamma_{h}-\gamma_{c})g^{(2)}_{h}(0),$ (7) $\displaystyle Q_{1}$ $\displaystyle=\langle\hat{H}\rangle_{h}-\langle\hat{H}\rangle_{c}-\frac{N\hbar^{2}\rho^{2}}{2m}(\gamma_{h}-\gamma_{c})g^{(2)}_{c}(0),$ (8) whereas work $W$ is given by Eq. (4) as before. The magnitude of CoP[A] is shown in Fig. 4 (a), where we note that, at the border between [E] and [A], the coefficient of performance diverges due to its inverse dependence on $W$. Operation of this QTM, enabled through $g^{(2)}_{h}(0)\\!<\\!g^{(2)}_{c}(0)$, additionally requires that $|W_{1}|=\langle\hat{H}\rangle_{\mathbf{B}}\\!-\\!\langle\hat{H}\rangle_{\mathbf{C}}>0$ (where $W_{1}<0$) and $W_{2}=\langle\hat{H}\rangle_{\mathbf{A}}\\!-\\!\langle\hat{H}\rangle_{\mathbf{D}}>0$, which are proportional to $g^{(2)}_{c}(0)$ and $g^{(2)}_{h}(0)$, respectively, must individually remain smaller than the energy gap, $\Delta E\\!=\\!\langle\hat{H}\rangle_{\mathbf{B}}\\!-\\!\langle\hat{H}\rangle_{\mathbf{D}}$, between the hot and cold thermal states, as shown in the cycle diagram in Fig. 3 (b). The conditions $|W_{1}|<\Delta E$ and $W_{2}<\Delta E$ may be expressed as $\displaystyle N\frac{\hbar^{2}\rho^{2}}{2m}(\gamma_{h}-\gamma_{c})g^{(2)}_{h}(0)<\Delta E,$ (9) $\displaystyle N\frac{\hbar^{2}\rho^{2}}{2m}(\gamma_{h}-\gamma_{c})g^{(2)}_{c}(0)<\Delta E,$ (10) and are equivalent to $Q_{1}\\!>\\!0$ and $Q_{2}\\!<\\!0$, respectively. ### 4.2 Heater [H] Operating the Otto cycle in the heater regime requires: $W>0,Q_{1}<0,\,Q_{2}<0$, and is shown schematically in Fig. 3 (c). This QTM utilizes mechanical work to heat up both hot and cold thermal reservoirs. The border with the accelerator regime is defined by $Q_{1}\\!=\\!0$, and the lower border with the refrigerator QTM is defined by $Q_{2}\\!=\\!0$; see Fig. 3 (e). We note that the condition $Q_{1}\\!<\\!0$ corresponds to $W_{2}\\!=\\!\langle\hat{H}\rangle_{\mathbf{A}}\\!-\\!\langle\hat{H}\rangle_{\mathbf{D}}\\!>\\!\Delta E$, opposite to the condition for the accelerator regime [A]. Additionally, for a fixed temperature ratio, $\tau_{h}/\tau_{c}$, since an arbitrarily large interaction strength quench to $\gamma_{h}\gg 1$ incurs fermionization (i.e. $g^{(2)}_{h}(0)\\!\to\\!0$), operation as a heater is inevitable as $\gamma_{h}/\gamma_{c}\to\infty$ at any fixed value of $\tau_{h}/\tau_{c}$ (see Appendix G). If one considers the benefit of operation of the heater to be heating of both reservoirs, then its CoP is trivially $\mathrm{CoP[H]}\\!=\\!-Q/W\\!=\\!1$. Instead, we define the benefit of operation of this QTM to be the heating of the hot reservoir; thus $\mathrm{CoP[H]}=-\frac{Q_{1}}{W}=1-\frac{|Q_{2}|}{W}<1,$ (11) which shown in Fig. 4 (b). One may alternatively define the benefit of the heater as heating the cold reservoir, in which case the CoP would be given by: $\mathrm{CoP[H]}^{\prime}=-\frac{Q_{2}}{W}=1-\mathrm{CoP[H]}<1.$ (12) Both definitions of $\mathrm{CoP[H]}$, however, are limited to be less than or equal to $1$ by energy conservation . ### 4.3 Refrigerator [R] The conditions of operating the Otto cycle as a refrigerator are: $W>0,\,Q_{1}<0,\,Q_{2}>0$, with $|Q_{2}|<|Q_{1}|$. The purpose of this thermal machine is to cool down the cold reservoir by extracting heat and dumping it into the hot reservoir, with the aid mechanical work done by the working fluid. The boundary between the refrigerator and the heater is defined by $Q_{2}\\!=\\!0$, see Fig. 3 (d). The CoP for the refrigerator is given by [54] $\mathrm{CoP[R]}=\frac{Q_{2}}{W}=\frac{|Q_{1}|}{W}-1,$ (13) and is shown in Fig. 4 (c); it diverges in the limit of infinitesimal quenches in both interaction strength and temperature, because the benefit of refrigeration, $Q_{2}$, vanishes slower than the cost, $W$, in these limits (see Appendix G. Further, as noted in the section addressing the heater QTM, large interaction strength quenches inevitably incur operation as a heater for any fixed temperature ratio. This implies that refrigeration occurs only over a finite parameter region, most clearly visible in regimes IV and VI of Fig. 3 (e). The role of the atom-atom local correlation function $g^{(2)}(0)$ in the thermal operation regimes and the respective boundaries of these additional QTM’s is discussed further in Appendix G. ## 5 Summary and outlook. We have proposed a sudden interaction-quench Otto cycle operating in a quantum many-body regime using a repulsive 1D Bose gas as a working fluid. Extracting net work from such a cycle in the heat engine regime was shown to be enabled by atom-atom correlations. Such correlations are characterized by Glauber’s second-order correlation function, $g^{(2)}(0)$, which is a thermodynamic quantity that can be calculated from the exact TBA solution through application of the Hellmann-Feynman theorem to the Helmholtz free energy [29]. Further, we have investigated other operational regimes of this Otto cycle, such as the thermal accelerator, heater, and refrigerator cycles, defining and examining their coefficients of performance. Though our specific results for the net work and the efficiency were calculated for a uniform 1D Bose as an example, the broad conclusions arrived at here on the basis of Eqs. (2) and (3) are applicable to any other many-body system with short-range contact interactions and should aid the tests of quantum thermodynamic concepts and realization of novel QTMs in laboratory settings. ## Acknowledgements This work was supported through Australian Research Council Discovery Project Grant No. DP190101515. Appendices In these Appendices, we briefly review the Lieb-Liniger model of the one- dimensional (1D) Bose gas with contact interactions, and discuss how the interaction quench can be achieved via changes of the strength of the transverse confinement of the 1D Bose gas. We also discuss physical considerations for the sudden quench to be effectively instantaneous. We next present the relevant analytic results for the Glauber’s second-order correlation $g^{(2)}(0)$ and the Hamiltonian energy $\langle\hat{H}\rangle$ in six different asymptotic regimes of the 1D Bose gas, used in the main text for evaluating the net work and efficiency of the Otto quantum heat engine. We also present equivalent numerical results obtained using the exact thermodynamic Bethe ansatz (TBA), and use these to explore a sudden quench Otto cycle operation outside of the analytically tractable asymptotic regimes of the 1D Bose gas. This is followed by a discussion of the maximum efficiency at maximum work. Finally, we discuss the role of the pair correlation $g^{(2)}(0)$ in the operation of other quantum thermal machines (QTMs) presented in the main text, such as accelerator, heater, and refrigerator. ## Appendix A The Lieb-Liniger model for the 1D Bose gas The Lieb-Liniger model for the uniform 1D Bose gas with repulsive contact interactions [32] is described by the follwoing second-quantized Hamiltonian $\begin{split}\begin{aligned} \hat{H}=\hat{H}^{kin}+\hat{H}^{int}=-\frac{\hbar^{2}}{2m}\int dz\hat{\Psi}^{\dagger}\frac{\partial^{2}}{\partial z^{2}}\hat{\Psi}+\frac{\chi}{2}\int dz\hat{\Psi}^{\dagger}\hat{\Psi}^{\dagger}\hat{\Psi}\hat{\Psi},\end{aligned}\end{split}$ (14) where $m$ is the atomic mass, $\chi$ is the strength of the contact interactions (see main text), and $\hat{\Psi}^{\dagger}(z)$ and $\hat{\Psi}(z)$ are the bosonic field creation and annihilation operators, respectively. We additionally highlight the separation of the Lieb-Liniger Hamiltonian into its kinetic energy, $\hat{H}^{kin}$, and interaction energy, $\hat{H}^{int}$, components, to be referred to later. Ground state solutions to this integrable model are dependent only on a single dimensionless interaction strength, $\gamma\\!=\\!m\chi/\hbar^{2}\rho$, where $\rho\\!=\\!N/L$ is the linear density for $N$ particles in a system of size $L$. Finite temperature solutions, on the other hand, can be obtained using the Yang-Yang thermodynamic Bethe ansatz (TBA) [33], and can be parameterized by an additional dimensional parameter, the dimensionless temperature $\tau\\!=\\!2mk_{B}T/\hbar^{2}\rho^{2}$ [29]. ## Appendix B Transverse Otto cycle For a magnetically trapped ultracold 1D Bose gas, the work done via transverse compression and expansion is ultimately magnetic: it is done by the magnetic field on the atomic dipole moments when $\omega_{\perp}$ is increased, or vice versa – by the atomic dipole moments on the magnetic field when $\omega_{\perp}$ is decreased. Alternatively, the change in the interaction strength $\chi$ is implemented through control over the $s$-wave scattering length $a_{s}$ via a magnetic Feshbach resonance [44], in which case the nature of the work is still magnetic. We use the term Otto cycle in the same sense as used to describe, e.g., a harmonic oscillator Otto engine [45, 46, 24, 25, 27, 6, 47, 48, 49, 50], wherein the harmonic oscillator frequency (rather than the volume of the system) is fixed as an external parameter during the thermalization strokes. In our case, it is the interaction strength that is fixed, which itself is proportional to the transverse harmonic confinement frequency of the 1D Bose gas. ## Appendix C Instantaneity of the sudden quench Realistically, a sudden quench of interaction strength from $\chi_{h(c)}$ to $\chi_{c(h)}$ would still occur over a finite duration $\Delta t$. The “instantaneity” of the quench utilized in the main text refers to the assumption that $\Delta t$ is much shorter than the characteristic time scale for longitudinal dynamics, i.e. that $\Delta t\\!\ll\\!ml_{\text{cor}}^{2}/\hbar$, where $l_{\text{cor}}$ is the characteristic short-range correlation length in the system, given, respectively, by: the healing length $l_{h}\\!=\\!\hbar/\sqrt{m\chi\rho}$ in regimes I and II; thermal phase coherence length $l_{\phi}\\!=\\!\hbar^{2}\rho/mk_{B}T$ in regime III; thermal de Broglie wavelength $\lambda_{T}\\!=\\!\sqrt{2\pi\hbar^{2}/mk_{B}T}$ in regime IV; absolute value of the 1D scattering length $|a_{1D}|=2\hbar^{2}/m\chi$ in the regime of high-temperature fermionization V; and the Fermi wavelength $\lambda_{F}\\!=\\!2/\rho$ in the Tonks-Girardeau regime of low-temperature fermionization VI. Thus, it is with respect to the _longitudinal_ dynamics that we refer to our quench as sudden. With respect to the _transverse_ dynamics, on the other hand, we are assuming that $\Delta t$ is sufficiently long ($\Delta t\gg 2\pi/\omega_{\perp}$) compared to the characteristic transverse timescale, $2\pi/\omega_{\perp}$, governed by the transverse harmonic trap frequency $\omega_{\perp}$ [43]. As a result, the quench would retain the system in the transverse ground state, and hence would not compromise the 1D character of the system. As such, the work done on (or by) the system during the unitary strokes can be regarded as transversely quasistatic. Figure 5: Atom-atom correlations, described by Glauber’s $g^{(2)}(0)$ correlation function, for the uniform 1D Bose gas evaluated using the exact Yang-Yang TBA [29, 33]. Panel a shows $g^{(2)}(0)$ as a function of the dimensionless interaction strength, $\gamma$, and temperature, $\tau$. In panel b, this is translated into a contour diagram, in which we also show the crossover boundaries (white solid and dashed lines) between the different asymptotic analytic regimes [29]. Panel (c) shows line plots of $g^{(2)}(0)$ vs $\gamma$, at different fixed values of $\tau$, together with two possible choices, D-B or $\textbf{D-B}^{\prime}$, of the thermal equilibrium operating points of the Otto cycle from Fig. 1; as we see, according to Eq. (4), operating the Otto cycle as an engine (with $W<0$) can be achieved between the points D-B ($\gamma_{c}\longleftrightarrow\gamma_{h}$), where $g^{(2)}_{c}(0)\\!<\\!g^{(2)}_{h}(0)$, but not between $\textbf{D-B}^{\prime}$ ($\gamma_{c}\longleftrightarrow\gamma^{\prime}_{h}$), where $g^{(2)}_{c}(0)\\!>\\!g^{(2)}_{h}(0)$ due to the stronger interaction quench, even though the temperature at D is still lower than at $\textbf{B}^{\prime}$. ## Appendix D Glauber’s second order correlation function The two-point correlation function may be generally defined through $g^{(2)}(z,z^{\prime})=\frac{\langle\hat{\Psi}^{\dagger}(z)\hat{\Psi}^{\dagger}(z^{\prime})\hat{\Psi}(z^{\prime})\hat{\Psi}(z)\rangle}{\rho(z)\rho(z^{\prime})}.$ (15) For a uniform (translationally invariant) system with $\rho(z^{\prime})\\!=\\!\rho(z)\\!=\\!\rho$, this $g^{(2)}(z,z^{\prime})$ depends only on the relative distance $|z-z^{\prime}|$, i.e. $g^{(2)}(z,z^{\prime})\\!=\\!g^{(2)}(|z-z^{\prime}|)$. If one is interested in the same point ($z\\!=\\!z^{\prime}$) correlation function, as utilized in the main text for calculation of the net work and efficiency of the quantum Otto cycle, this in turn becomes $g^{(2)}(0)$. The 1D Bose gas can be characterized by six distinct asymptotic regimes defined through the same-point correlation function [29], as shown in Fig. 5. The weakly interacting ($\gamma\ll 1$), low temperature ($\tau\ll 2\sqrt{\gamma}$) quasicondensate regime can be treated using the Bogoliubov theory for quasicondensates [39], and is characterised by suppressed density fluctuations, but fluctuating phase. This may be subdivided into regions dominated by quantum (I) and thermal (II) fluctuations [29]. At higher temperatures, the gas becomes nearly ideal, and can be treated using perturbation theory with respect to $\gamma$ [29, 42]. This asymptotic region may in turn be subdivided into quantum degenerate (III) and non-degenerate (IV) regimes. Finally, in the strongly interacting regime ($\gamma\gg 1$), where the Fermi-Bose gas mapping applies, the 1D Bose gas can be well approximated by a nearly ideal Fermi gas, and can be treated using perturbation theory with respect to $1/\gamma$ [29, 42]. This regime can be further subdivided into two regions corresponding to high-temperature (V) and low-temperature (VI) fermionization. In each of these asymptotic regimes, the pair correlation function $g^{(2)}(0)$ can be derived in closed approximate analytic form, where we additionally define the boundary of these regimes in terms of $\gamma$ and $\tau$: $\displaystyle\text{I}\\!:\,$ $\displaystyle g^{(2)}(0)\\!\simeq\\!1-\frac{2}{\pi}\gamma^{1/2}+\frac{\pi\,\tau^{2}}{24\,\gamma^{3/2}},\,\left[\frac{\tau}{2}\\!\ll\\!\gamma\ll 1\right],$ (16) $\displaystyle\text{II}\\!:\,\,$ $\displaystyle g^{(2)}(0)\simeq 1+\frac{\tau}{2\sqrt{\gamma}}\,\,,\quad\left[2\gamma\ll\tau\ll 2\sqrt{\gamma}\right],$ (17) $\displaystyle\text{III}\\!:\,\,$ $\displaystyle g^{(2)}(0)\simeq 2-\frac{4\gamma}{\tau^{2}},\quad\left[2\sqrt{\gamma}\ll\tau\ll 1\right],$ (18) $\displaystyle\text{IV}\\!:\,\,$ $\displaystyle g^{(2)}(0)\simeq 2-\gamma\sqrt{\frac{2\pi}{\tau}},\,\,\,\left[\tau\gg\text{max}\\{1,\gamma^{2}\\}\right],$ (19) $\displaystyle\text{V}\\!:\,\,$ $\displaystyle g^{(2)}(0)\simeq\frac{2\tau}{\gamma^{2}},\,\,\left[\frac{\pi^{2}}{(1+2/\gamma)^{2}}\ll\tau\ll\gamma^{2}\right],$ (20) $\displaystyle\text{VI}\\!:\,$ $\displaystyle g^{(2)}(0)\simeq\frac{4\pi^{2}}{3\gamma^{2}}\left(1\\!+\\!\frac{\tau^{2}}{4\pi^{2}}\right)\\!,\left[\tau\\!\ll\\!\frac{\pi^{2}}{(1+2/\gamma)^{2}},\,\gamma\\!\gg\\!1\right]\\!.$ (21) Further, we may express the total energy of system in each asymptotic regime as [52], $\displaystyle\text{I}:\,\,\,$ $\displaystyle\langle\hat{H}\rangle\\!\simeq\\!N\frac{\hbar^{2}\rho^{2}}{2m}\left(\gamma-\frac{4}{3\pi}\gamma^{3/2}+\frac{\pi}{12}\frac{\tau^{2}}{\gamma^{1/2}}\right),$ (22) $\displaystyle\text{II}:\,\,\,$ $\displaystyle\langle\hat{H}\rangle\\!\simeq\\!N\frac{\hbar^{2}\rho^{2}}{2m}\left(\gamma+\frac{\zeta(3/2)}{4\sqrt{\pi}}\tau^{3/2}+\frac{\zeta(1/2)}{2\sqrt{\pi}}\tau^{1/2}\gamma\right),$ (23) $\displaystyle\text{III}:\,\,\,$ $\displaystyle\langle\hat{H}\rangle\\!\simeq\\!N\frac{\hbar^{2}\rho^{2}}{2m}\left(\frac{1}{2}\zeta(3/2)+2\gamma-\frac{6\gamma^{2}}{\tau^{2}}\right),$ (24) $\displaystyle\text{IV}:\,\,\,$ $\displaystyle\langle\hat{H}\rangle\\!\simeq\\!N\frac{\hbar^{2}\rho^{2}}{2m}\left(\frac{\tau}{2}+2\gamma-\frac{3}{2}\sqrt{\frac{\pi}{2}}\frac{\gamma^{2}}{\tau^{1/2}}\right),$ (25) $\displaystyle\text{V}:\,\,\,$ $\displaystyle\langle\hat{H}\rangle\\!\simeq\\!N\frac{\hbar^{2}\rho^{2}}{2m}\left(\frac{\tau}{2}+\frac{1}{2}\sqrt{\frac{\pi}{2}}\tau^{1/2}-\sqrt{\frac{\pi}{2}}\frac{\tau^{1/2}}{\gamma}\right),$ (26) $\displaystyle\text{VI}:\,\,\,$ $\displaystyle\langle\hat{H}\rangle\\!\simeq\\!N\frac{\hbar^{2}\rho^{2}}{2m}\left(\frac{\pi^{2}}{3}-\frac{4\pi^{2}}{3\gamma}+\frac{\tau^{2}}{3\gamma}\right),$ (27) where $\zeta(s)$ is the Riemann zeta function of $s\\!\in\\!\mathbb{R}$. As described in the main text, the normalized local second-order correlation function may be rearranged and integrated for the total (integrated) correlation function, $\begin{split}\begin{aligned} \\!\overline{G^{(2)}}\equiv\\!\int dz\langle\hat{\Psi}^{\dagger}(z)\hat{\Psi}^{\dagger}(z)\hat{\Psi}(z)\hat{\Psi}(z)\rangle\\!=\\!\int_{0}^{L}\\!dzg^{(2)}(0)\rho^{2}.\end{aligned}\end{split}$ (28) Utilizing the linear density, $\rho\\!=\\!N/L$, this may be expressed as $\overline{G^{(2)}}\\!=\\!N\rho g^{(2)}(0)$, which is used in Eq. (4) of the main text for expressing the exact net work of the uniform 1D Bose gas in terms of $g^{(2)}(0)$. ## Appendix E Exact thermodynamic Bethe ansatz results Experimental realization of a 1D Bose gas often falls outside the asymptotic regimes where analytic approximations are applicable. In such situations, we may utilize the exact Yang-Yang thermodynamic Bethe ansatz [33, 34] to evaluate the equilibrium properties of the gas required for calculating net work and efficiency via Eqs. (2) and (3) of the main text, respectively. This is presented in Fig. 6 (a) for experimentally realistic set of system parameters that inhabit the boundary between asymptotic parameter regimes II and III (see Fig. 5). Further, one may utilize the exact TBA to confirm the results derived via approximate analytics in the main text. This is illustrated in Figs. 6 (b) and (c), where we see excellent agreement between these results when the parameters $\gamma$ and $\tau$ are sufficiently deep into the analytic asymptotic regimes. Figure 6: Performance of the sudden interaction quench quantum Otto cycle, numerically evaluated via the thermodynamic Bethe ansatz. Panel a demonstrates numerically evaluated net work and efficiency for a system with a cold thermal state defined by $\gamma_{c}\\!=\\!0.1$, $\tau_{c}\\!=\\!0.5$, lying on the border of regimes II and III (see Fig. 5), and thus lying outside the range of the analytic approximations utilized in the main text. Panel (b) is a copy of Fig. 2 (a) of the main text shown here for comparison with the results of numerical TBA evaluation of the same cycle shown in panel (c). Here, there is excellent agreement in the net work between panels (b) and (c), with small disagreement under large interaction strength and temperature ratios, as the hot thermal state is approaching the edge of the asymptotic regime where the analytic approximations become less applicable. ## Appendix F Maximum efficiency and maximum work For a fixed ratio of temperatures, it was noted in the main text that the interaction strength ratio corresponding to maximum work approximately coincides with that for maximum efficiency, which is uncommon for highly nonequilibrium engine cycles [21, 46]. In the sudden interaction-quench Otto cycle, such coincidence occurs due to the dependence of the total energy of the hot and cold thermal equilibrium states on the interaction strength. As shown in Eq. (14), the total energy may be separated into its kinetic energy, which scales predominately with temperature, and interaction energy, which scales predominately with interaction strength. Thus, for a fixed ratio of temperatures, $\tau_{h}/\tau_{c}$, the difference between the total energies of the hot and cold thermal state may be given as a sum of two terms: the first is the kinetic energy difference, determined by the temperature ratio and therefore approximately constant, the second given by the interaction energy difference, which scales with the interaction strengths, $\gamma_{h}$ and $\gamma_{c}$, of the hot and cold thermal states as $\langle\hat{H}^{int}\rangle_{h}\\!-\\!\langle\hat{H}^{int}\rangle_{c}\\!=\\!N\frac{\hbar^{2}\rho^{2}}{2m}\left(\gamma_{h}g^{(2)}_{h}(0)\\!-\\!\gamma_{c}g^{(2)}_{c}(0)\right).$ (29) However, when operating within a single asymptotic regime under a moderate quench of interaction strength, the $g^{(2)}(0)$ correlation function is only slowly varying with $\gamma$. This means, to first approximation, $g^{(2)}_{h}(0)\\!\simeq\\!g^{(2)}_{c}(0)$, which in turn transforms the interaction energy difference to $\langle\hat{H}^{int}\rangle_{h}\\!-\\!\langle\hat{H}^{int}\rangle_{c}\\!\simeq\\!N\frac{\hbar^{2}\rho^{2}}{2m}\left(\gamma_{h}\\!-\\!\gamma_{c}\right)g^{(2)}_{c}(0).$ (30) The heat intake, which is given by Eq. (8) of the main text, is therefore well approximated by $\displaystyle Q_{1}=\langle\hat{H}\rangle_{h}-\langle\hat{H}\rangle_{c}-\frac{N\hbar^{2}\rho^{2}}{2m}(\gamma_{h}-\gamma_{c})g^{(2)}_{c}(0)\simeq\langle\hat{H}^{kin}\rangle_{h}-\langle\hat{H}^{kin}\rangle_{c},$ (31) which is approximately constant for a fixed temperature ratio, as detailed above. Therefore, the efficiency, which is given by $\eta\\!=\\!W/Q_{1}$, scales predominately with $W$, hence $\eta\\!\propto\\!W$. ## Appendix G Thermal operation regimes of other QTM’s Under large interaction strength quenches, for fixed temperatures $\tau_{c}$ and $\tau_{h}$, it was noted in the main text that the heater is the inevitable mode of operation. This scenario requires the fulfilment of two conditions: first, we require $Q_{1}\\!<\\!0$, meaning the magnitude of the work out, $|W_{1}|\\!=\\!N\frac{\hbar^{2}\rho^{2}}{2m}(\gamma_{h}\\!-\\!\gamma_{c})g^{(2)}_{c}(0)$ (where $W_{1}<0$) exceeds the energy gap between the hot and cold thermal states, given by $\langle\hat{H}\rangle_{h}\\!-\\!\langle\hat{H}\rangle_{c}$. We further require $Q_{2}\\!>\\!0$, meaning that the work in, $W_{2}\\!=\\!N\frac{\hbar^{2}\rho^{2}}{2m}(\gamma_{h}\\!-\\!\gamma_{c})g^{(2)}_{h}(0)$ (where $W_{2}>0$), must remain less than this same gap (see Fig. 3 (c) of the main text). As detailed above, in Appendix F on maximum efficiency at maximum work, the total energy difference between the hot and cold thermal states for fixed temperature ratios, $\tau_{h}/\tau_{c}$, is given by a sum of the kinetic energy difference, which is approximately constant, and the interaction energy difference, where the correlation function is approximately constant in a single asymptotic regime, and hence given by Eq. (29). In contrast, for a large quench in interaction strength, the correlation function is no longer approximately constant, and $g^{(2)}_{h}(0)$ is strongly monotonically decreasing as a function of $\gamma_{h}$, i.e. $g^{(2)}_{h}(0)\\!<\\!g^{(2)}_{c}(0)$. We therefore find that the work input, $W_{2}$, exceeds the interaction energy difference given in Eq. (29), $W_{2}\propto(\gamma_{h}-\gamma_{c})g^{(2)}_{c}(0)>\gamma_{h}g^{(2)}_{h}(0)\\!-\\!\gamma_{c}g^{(2)}_{c}(0).$ (32) Further, as the kinetic energy term is approximately constant, $W_{2}$ inevitably exceeds the difference in total energy between the hot and cold thermal states due to its linear dependence on $\gamma_{h}$. Similarly, since $g^{(2)}_{h}(0)$ monotonically decreases with $\gamma_{h}$ for large quenches of interaction strength, the magnitude of the work output, $|W_{1}|\propto(\gamma_{h}-\gamma_{c})g^{(2)}_{h}(0)$, remains less than the energy gap between the hot and cold thermal states: $|W_{1}|\propto(\gamma_{h}-\gamma_{c})g^{(2)}_{h}(0)<\gamma_{h}g^{(2)}_{h}(0)\\!-\\!\gamma_{c}g^{(2)}_{c}(0).$ (33) These two conditions, taken together, imply operation as a heater under large interaction quenches. In contrast, for any fixed value of $\gamma_{c}$ and $\gamma_{h}$, an increasingly higher temperature of the hot thermal state, $\tau_{h}$, means that the corresponding correlation function, $g^{(2)}_{h}(0)$, monotonically increases towards its maximum value of $g^{(2)}_{h}(0)\\!\simeq\\!2$, which is achieved in regime IV, defined in Eq. (19). Thus, there is always a value of $\tau_{h}$ such that $g^{(2)}_{h}(0)\\!>\\!g^{(2)}_{c}(0)$, turning the Otto cycle into the engine operation regime. Finally, in the refrigerator thermal operation regime, for $\tau_{h}=\tau_{c}$ and an infinitesimal quench of interaction strength, $\gamma_{h}-\gamma_{c}=\delta\gamma$, the net work vanishes as $W\\!\propto(\gamma_{h}-\gamma_{c})$ $\times(g^{(2)}_{h}(0)-g^{(2)}_{c}(0))\propto\delta\gamma^{2}$. This occurs as the zeroth order terms in the correlation function cancel when taking their difference in a single asymptotic regime. In contrast, the heat intake, $Q_{1}$, depends on the difference in total energy, which to first order vanishes as $Q_{1}\\!\propto\\!\delta\gamma$. This results in $\mathrm{CoP[R]}\\!=\\!|Q_{1}|/W\\!-\\!1\\!\propto\delta\gamma^{-1}$, which diverges as $\delta\gamma\\!\to\\!0$, as noted in the main text. ## References * [1] S. Vinjanampathy and J. Anders, _Quantum thermodynamics_ , Contemporary Physics 57(4), 545 (2016), 10.1080/00107514.2016.1201896. * [2] R. Kosloff and A. Levy, _Quantum Heat Engines and Refrigerators: Continuous Devices_ , Annual Review of Physical Chemistry 65(1), 365 (2014), 10.1146/annurev-physchem-040513-103724. * [3] J. Roßnagel, S. T. Dawkins, K. N. Tolazzi, O. Abah, E. Lutz, F. Schmidt-Kaler and K. Singer, _A single-atom heat engine_ , Science 352(6283), 325 (2016), 10.1126/science.aad6320. * [4] D. von Lindenfels, O. Gräb, C. T. Schmiegelow, V. Kaushal, J. Schulz, M. T. Mitchison, J. Goold, F. Schmidt-Kaler and U. G. Poschinger, _Spin Heat Engine Coupled to a Harmonic-Oscillator Flywheel_ , Phys. Rev. Lett. 123, 080602 (2019), 10.1103/PhysRevLett.123.080602. * [5] J. Klatzow, J. N. Becker, P. M. Ledingham, C. Weinzetl, K. T. Kaczmarek, D. J. Saunders, J. Nunn, I. A. Walmsley, R. Uzdin and E. Poem, _Experimental Demonstration of Quantum Effects in the Operation of Microscopic Heat Engines_ , Phys. Rev. Lett. 122, 110601 (2019), 10.1103/PhysRevLett.122.110601. * [6] Q. Bouton, J. Nettersheim, S. Burgardt, D. Adam, E. Lutz and A. Widera, _A quantum heat engine driven by atomic collisions_ , Nature Communications 12(1), 2063 (2021), 10.1038/s41467-021-22222-z. * [7] R. Feynman, _There’s plenty of room at the bottom_ , In _Feynman and computation_ , pp. 63–76. CRC Press, Boca Raton, Florida, United States (2018). * [8] F. G. Brandao and M. B. Plenio, _Entanglement theory and the second law of thermodynamics_ , Nature Physics 4(11), 873 (2008), 10.1038/nphys1100. * [9] K. Funo, Y. Watanabe and M. Ueda, _Thermodynamic work gain from entanglement_ , Phys. Rev. A 88, 052319 (2013), 10.1103/PhysRevA.88.052319. * [10] R. Alicki and M. Fannes, _Entanglement boost for extractable work from ensembles of quantum batteries_ , Phys. Rev. E 87, 042123 (2013), 10.1103/PhysRevE.87.042123. * [11] S. Hilt and E. Lutz, _System-bath entanglement in quantum thermodynamics_ , Phys. Rev. A 79, 010101 (2009), 10.1103/PhysRevA.79.010101. * [12] J. Oppenheim, M. Horodecki, P. Horodecki and R. Horodecki, _Thermodynamical Approach to Quantifying Quantum Correlations_ , Phys. Rev. Lett. 89, 180402 (2002), 10.1103/PhysRevLett.89.180402. * [13] M. Perarnau-Llobet, K. V. Hovhannisyan, M. Huber, P. Skrzypczyk, N. Brunner and A. Acín, _Extractable Work from Correlations_ , Phys. Rev. X 5, 041011 (2015), 10.1103/PhysRevX.5.041011. * [14] M. Huber, M. Perarnau-Llobet, K. V. Hovhannisyan, P. Skrzypczyk, C. Klöckl, N. Brunner and A. Acín, _Thermodynamic cost of creating correlations_ , New Journal of Physics 17(6), 065008 (2015), 10.1088/1367-2630/17/6/065008. * [15] V. Narasimhachar and G. Gour, _Low-temperature thermodynamics with quantum coherence_ , Nature communications 6(1), 7689 (2015), 10.1038/ncomms8689. * [16] K. Korzekwa, M. Lostaglio, J. Oppenheim and D. Jennings, _The extraction of work from quantum coherence_ , New Journal of Physics 18(2), 023045 (2016), 10.1088/1367-2630/18/2/023045. * [17] M. Lostaglio, D. Jennings and T. Rudolph, _Description of quantum coherence in thermodynamic processes requires constraints beyond free energy_ , Nature communications 6(1), 6383 (2015), 10.1038/ncomms7383. * [18] B. d. L. Bernardo, _Unraveling the role of coherence in the first law of quantum thermodynamics_ , Phys. Rev. E 102, 062152 (2020), 10.1103/PhysRevE.102.062152. * [19] P. Kammerlander and J. Anders, _Coherence and measurement in quantum thermodynamics_ , Scientific reports 6(1), 22174 (2016), 10.1038/srep22174. * [20] P. Ćwikliński, M. Studziński, M. Horodecki and J. Oppenheim, _Limitations on the Evolution of Quantum Coherences: Towards Fully Quantum Second Laws of Thermodynamics_ , Phys. Rev. Lett. 115, 210403 (2015), 10.1103/PhysRevLett.115.210403. * [21] J. Jaramillo, M. Beau and A. del Campo, _Quantum supremacy of many-particle thermal machines_ , New Journal of Physics 18(7), 075019 (2016), 10.1088/1367-2630/18/7/075019. * [22] N. Yunger Halpern, C. D. White, S. Gopalakrishnan and G. Refael, _Quantum engine based on many-body localization_ , Phys. Rev. B 99, 024203 (2019), 10.1103/PhysRevB.99.024203. * [23] M. Beau, J. Jaramillo and A. Del Campo, _Scaling-Up Quantum Heat Engines Efficiently via Shortcuts to Adiabaticity_ , Entropy 18(5) (2016), 10.3390/e18050168. * [24] J. Li, T. Fogarty, S. Campbell, X. Chen and T. Busch, _An efficient nonlinear Feshbach engine_ , New Journal of Physics 20(1), 015005 (2018), 10.1088/1367-2630/aa9cd8. * [25] Y.-Y. Chen, G. Watanabe, Y.-C. Yu, X.-W. Guan and A. del Campo, _An interaction-driven many-particle quantum heat engine and its universal behavior_ , npj Quantum Information 5(1), 88 (2019), https://doi.org/10.1038/s41534-019-0204-5. * [26] T. Keller, T. Fogarty, J. Li and T. Busch, _Feshbach engine in the Thomas-Fermi regime_ , Physical Review Research 2(3), 033335 (2020), 10.1103/PhysRevResearch.2.033335. * [27] J. Koch, K. Menon, E. Cuestas, S. Barbosa, E. Lutz, T. Fogarty, T. Busch and A. Widera, _A quantum engine in the BEC–BCS crossover_ , Nature 621(7980), 723 (2023), https://doi.org/10.1038/s41586-023-06469-8. * [28] E. Q. Simmons, R. Sajjad, K. Keithley, H. Mas, J. L. Tanlimco, E. Nolasco-Martinez, Y. Bai, G. H. Fredrickson and D. M. Weld, _Thermodynamic engine with a quantum degenerate working fluid_ , Phys. Rev. Res. 5, L042009 (2023), 10.1103/PhysRevResearch.5.L042009. * [29] K. V. Kheruntsyan, D. M. Gangardt, P. D. Drummond and G. V. Shlyapnikov, _Pair Correlations in a Finite-Temperature 1D Bose Gas_ , Phys. Rev. Lett. 91, 040403 (2003), 10.1103/PhysRevLett.91.040403. * [30] K. V. Kheruntsyan, D. M. Gangardt, P. D. Drummond and G. V. Shlyapnikov, _Finite-temperature correlations and density profiles of an inhomogeneous interacting one-dimensional Bose gas_ , Phys. Rev. A 71, 053615 (2005), 10.1103/PhysRevA.71.053615. * [31] T. Kinoshita, T. Wenger and D. S. Weiss, _Local Pair Correlations in One-Dimensional Bose Gases_ , Phys. Rev. Lett. 95, 190406 (2005), 10.1103/PhysRevLett.95.190406. * [32] E. H. Lieb and W. Liniger, _Exact Analysis of an Interacting Bose Gas. I. The General Solution and the Ground State_ , Phys. Rev. 130, 1605 (1963), 10.1103/PhysRev.130.1605. * [33] C. N. Yang and C. P. Yang, _Thermodynamics of a One-Dimensional System of Bosons with Repulsive Delta-Function Interaction_ , Journal of Mathematical Physics 10(7), 1115 (1969), 10.1063/1.1664947. * [34] V. E. Korepin, N. M. Bogoliubov and A. G. Izergin, _Quantum Inverse Scattering Method and Correlation Functions_ , Cambridge Monographs on Mathematical Physics. Cambridge University Press, Cambridge, United Kingdom, 10.1017/CBO9780511628832 (1993). * [35] A. Görlitz, J. M. Vogels, A. E. Leanhardt, C. Raman, T. L. Gustavson, J. R. Abo-Shaeer, A. P. Chikkatur, S. Gupta, S. Inouye, T. Rosenband and W. Ketterle, _Realization of Bose-Einstein Condensates in Lower Dimensions_ , Phys. Rev. Lett. 87, 130402 (2001), 10.1103/PhysRevLett.87.130402. * [36] M. Greiner, I. Bloch, O. Mandel, T. W. Hänsch and T. Esslinger, _Exploring Phase Coherence in a 2D Lattice of Bose-Einstein Condensates_ , Phys. Rev. Lett. 87, 160405 (2001), 10.1103/PhysRevLett.87.160405. * [37] F. Schreck, L. Khaykovich, K. L. Corwin, G. Ferrari, T. Bourdel, J. Cubizolles and C. Salomon, _Quasipure Bose-Einstein Condensate Immersed in a Fermi Sea_ , Phys. Rev. Lett. 87, 080403 (2001), 10.1103/PhysRevLett.87.080403. * [38] D. S. Petrov, G. V. Shlyapnikov and J. T. M. Walraven, _Regimes of Quantum Degeneracy in Trapped 1D Gases_ , Phys. Rev. Lett. 85, 3745 (2000), 10.1103/PhysRevLett.85.3745. * [39] C. Mora and Y. Castin, _Extension of Bogoliubov theory to quasicondensates_ , Phys. Rev. A 67, 053615 (2003), 10.1103/PhysRevA.67.053615. * [40] M. Girardeau, _Relationship between Systems of Impenetrable Bosons and Fermions in One Dimension_ , Journal of Mathematical Physics 1(6), 516 (1960), 10.1063/1.1703687. * [41] T. Kinoshita, T. Wenger and D. S. Weiss, _Observation of a One-Dimensional Tonks-Girardeau Gas_ , Science 305(5687), 1125 (2004), 10.1126/science.1100700. * [42] A. G. Sykes, D. M. Gangardt, M. J. Davis, K. Viering, M. G. Raizen and K. V. Kheruntsyan, _Spatial nonlocal pair correlations in a repulsive 1d bose gas_ , Phys. Rev. Lett. 100, 160406 (2008), 10.1103/PhysRevLett.100.160406. * [43] M. Olshanii, _Atomic Scattering in the Presence of an External Confinement and a Gas of Impenetrable Bosons_ , Phys. Rev. Lett. 81, 938 (1998), 10.1103/PhysRevLett.81.938. * [44] C. Chin, R. Grimm, P. Julienne and E. Tiesinga, _Feshbach resonances in ultracold gases_ , Rev. Mod. Phys. 82, 1225 (2010), 10.1103/RevModPhys.82.1225. * [45] R. Kosloff, _A quantum mechanical open system as a model of a heat engine_ , The Journal of Chemical Physics 80(4), 1625 (1984), 10.1063/1.446862. * [46] R. Kosloff and Y. Rezek, _The Quantum Harmonic Otto Cycle_ , Entropy 19(4) (2017), 10.3390/e19040136. * [47] Y. Zheng and D. Poletti, _Work and efficiency of quantum Otto cycles in power-law trapping potentials_ , Phys. Rev. E 90, 012145 (2014), 10.1103/PhysRevE.90.012145. * [48] G. Watanabe, B. P. Venkatesh, P. Talkner and A. del Campo, _Quantum Performance of Thermal Machines over Many Cycles_ , Phys. Rev. Lett. 118, 050601 (2017), 10.1103/PhysRevLett.118.050601. * [49] O. Abah and E. Lutz, _Energy efficient quantum machines_ , Europhysics Letters 118(4), 40005 (2017), 10.1209/0295-5075/118/40005. * [50] N. M. Myers, F. J. Peña, O. Negrete, P. Vargas, G. D. Chiara and S. Deffner, _Boosting engine performance with Bose–Einstein condensation_ , New Journal of Physics 24(2), 025001 (2022), 10.1088/1367-2630/ac47cc. * [51] H. B. Callen, _Thermodynamics and an introduction to thermostatistics_ , John Wiley & Sons, Hoboken, New Jersey, 2nd ed edn., ISBN 0471862568 (1985). * [52] G. D. Rosi, R. Rota, G. E. Astrakharchik and J. Boronat, _Hole-induced anomaly in the thermodynamic behavior of a one-dimensional Bose gas_ , SciPost Phys. 13, 035 (2022), 10.21468/SciPostPhys.13.2.035. * [53] L. Buffoni, A. Solfanelli, P. Verrucchi, A. Cuccoli and M. Campisi, _Quantum measurement cooling_ , Phys. Rev. Lett. 122, 070603 (2019), 10.1103/PhysRevLett.122.070603. * [54] D. V. Schroeder, _An introduction to thermal physics_ , Oxford University Press, Oxford, ISBN 0-19-289554-0 (2020).
# Towards Immersive Generosity: The Need for a Novel Framework to Explore Large Audiovisual Archives through Embodied Experiences in Immersive Environments Giacomo Alliata Laboratory of Experimental Museology, EPFL Corresponding author<EMAIL_ADDRESS>- Rue des Jordils 41, St-Sulpice, Switzerland Sarah Kenderdine Laboratory of Experimental Museology, EPFL Lily Hibberd Laboratory of Experimental Museology, EPFL UNSW Sydney Ingrid Mason Australian National University ###### Abstract This article proposes an innovative framework to explore large audiovisual archives using Immersive Environments to place users inside a dataset and create an embodied experience. It starts by outlining the need for such a novel interface to meet the needs of archival scholars and the GLAM sector, and discusses issues in the current modes of access, mostly restrained to traditional information retrieval systems based on metadata. The paper presents the concept of “generous interfaces” as a preliminary approach to address these issues, and argues some of the key reasons why employing Immersive Visual Storytelling might benefit such frameworks. The theory of embodiment is leveraged to justify this claim, showing how a more embodied understanding of a collection can result in a stronger engagement for the public. By placing users as actors in the experience rather than mere spectators, the emergence of narrative is driven by their interactions, with benefits in terms of engagement with the public and understanding of the cultural component. The framework we propose is applied to two existing installations to analyze them in-depth and critique them, highlighting the key directions to pursue for further development. ## 1 Introduction At the beginning of his book _Zen and the Art of Motorcycle Maintenance_ , Robert [49] presents the difference felt while travelling by car compared to on a motorbike. In the first case, travelers are watching the environment through their car windows, while in the second, they are “feeling”, “living” that same environment. The humidity in the air, the wind on their faces, the vivid sound of the motor, the unconstrained visual field, all contribute to create a more embodied and immersive experience for the motorcyclists. A comparable experience is described by [59], in which one of the authors, who, after several years of looking through the windows of a computer room in passing, is one day obliged to enter and is confronted with the shock of suddenly going “inside” what had only ever been seen from the “outside”. This short story illustrates what Immersive Environments (IEs) offer to their users: the possibility to be “inside” and therefore to feel as if you were “there”, a more fully embodied experience, (more or less) completely surrounded by a virtual world. Often described as the sense of “being there”, this psychological state is sometimes also referred to as “presence” and is a concept that is one of the central features of IEs [62]. “Immersion” is arguably the second central feature of IEs: the extent to which the system can deliver “an inclusive, extensive, surrounding and vivid illusion of reality to the senses of a human participant” [[]p. 1]slater_framework_1997. Since the display size remains a critical limitation in the presentation and understanding of data visualizations [28], many large display systems (more or less immersive) have been developed in the last decades, such as wide gigapixel screens, dome-like structures, CAVEs and half-CAVEs or panoramic screens (see [24] for an historical perspective on IEs). Despite this progress, the exploration of cultural collections in IEs faces a range of limitations and dilemmas, for which a new framework is required. This article proposes that such a new paradigm is possible and that, within the project _Narratives from the Long Tail: Transforming Access to Audiovisual Archives_ , the foundations for this innovative framework are being laid and will be applied to important audiovisual collections to create new ways to explore them111Details on the archives explored can be found at https://www.futurecinema.live/archives-and-collections/.. _Narratives_ ’s goal is to produce a “groundbreaking visualization framework for interactively (re)discovering hundreds of thousands of hours of audiovisual materials” [32]. The remainder of this introduction offers a summary of the key concepts and technologies to be discussed. As the foremost mnemonic records of the 21st century, audiovisual recordings are omnipresent in our daily lives. From the second half of the 20th century, starting with the introduction of television, to today’s online sharing platforms, such as Youtube or TikTok, audiovisual recordings play an important role in the way we document, disseminate and preserve knowledge and culture. Broadcasting institutions are digitizing their collections, with examples such as the Radio Télévision Suisse (RTS) with 200,000 hours of footage [54], or the British Broadcasting Corporation (BBC) with more than a million recorded hours [66]. Furthermore, cultural video collections are useful to preserve Intangible Cultural Heritage (ICH), which has been defined as “the culture that people practise as part of their daily lives” [35] and “all [the] immaterial manifestations of culture [that] represent the variety of living heritage of humanity as well as the most important vehicle of cultural diversity” [37]. One can think of events like rituals, dance performances or festivals that audiovisual collections, due to their temporal value as well as the combination of video and audio channels, are more apt to represent than images or textual elements ever could. Audiovisual archives by themselves, however, can only offer a mostly linear, single-channel, non-interactive narrative. They need to be augmented by some sort of interface to be explored in a meaningful way. In the field of cultural heritage, the need for innovative ways to present such collections to the public is evident. A 2014 survey conducted on 1200 cultural institutions in Europe established that around 80% have digitized their collections [20], and this number has certainly increased since then. The Europeana platform222https://www.europeana.eu/en is, for instance, a key example, with more than 30 million items accessible online through a category- based interface. The majority of these institutions provide online access to (at least part of) their collections, and even though innovative forms of access can be noted [65], one can argue that these largely remain constrained to the web. Web-based systems, such as Europeana’s, are obviously powerful tools to enhance access to these collections and democratisation of culture, however they lack the immersion that IEs installations can offer. These are especially relevant when showcasing collections in public spaces such as museums or cultural institutions venues. An innovative framework to explore large digital collections is therefore necessary, in particular, to present audiovisual archives to a larger public. Within IE theory, one key aspect to consider is the way narrative can emerge in an embodied experience, what can be referred to as Immersive Visual Storytelling (IVS). This term comprises the ensemble of processes and approaches that are employed to generate narratives in immersive experiences (see [38, 36], for some examples). It is on this basis that we claim that, through this emergence of narrative, users can better understand and explore large cultural collections, by being placed “inside” the dataset [57] and being offered levels of interactivity to freely browse the collection, as well as making them actors in the experience rather than mere spectators, resulting in a more “embodied understanding” of the content displayed [27]. The next part of this paper justifies the need for a new framework to explore large audiovisual collections (and cultural collections, in general), drawing on Whitelaw’s “generous interface” theory [64] and outlining the benefits immersion can bring to the exploration of such archives. The relevant theories of embodiment will then be presented and discussed in relations to IEs and IVS with specific examples, in order to understand how “embodied understanding” works and therefore how it can be leveraged to enhance the experience. Finally, two previous works will be analyzed in-depth and critiqued in light of the proposed framework: _T_Visionarium II_ , developed at UNSW’s iCinema Research Centre in 2006, and _Jazz Luminaries_ , created in 2019 at EPFL’s Laboratory of Experimental Museology. In each case, users are immersed within large audiovisual datasets, and this results in an embodied understanding of the archives, from which narrative can emerge. ## 2 The Need for a New Framework: Towards Immersive Generosity From the perspective of scholarly archival community, it is clear that large audiovisual archives are currently lacking proper frameworks to explore them. These collections remain mostly inaccessible because cultural institutions are constrained to screening sessions where only a handful of videos can be shown at one time, without revealing the dataset in its entirety. The Digital Humanities and GLAM sector (galleries, libraries, archives, and museums) have also called for innovative forms of engagement through compelling frameworks to explore this kind of collections [21]. For one, the classical information retrieval system restrains users to classifications and tagging prior interpretations provided by curators and data managers. Furthermore, the metadata compiled prevents users from accessing visual features because images are by nature harder to verbally capture [40]. A “simple search” interface is thus insufficient to provide compelling public engagement with the diverse array of media formats held in cultural institutions [53]. The framework proposed in this article intends to solve this problem, as it provides users new ways to access large collections, thus dispensing with the need to rely on metadata and traditional cataloguing forms of description. A catalogue-like database description of an audiovisual collection cannot fully describe its content because discrete categories do not encompass the visual and temporal complexity of videos. More continuous analytics can in theory characterize visual features of a video but in practice these seem to be harder to capture as they require some sort of interpretation. One could imagine a scale from 0 to 1 measuring the visual complexity of a shot, for instance. However, a catalogue of entries between 0 and 1 would be much harder to interpret than actually watching the related videos and visually compare them, as in the SEMIA333https://bertspaan.nl/semia/#/ project. This recent initiative undertaken at the University of Amsterdam is a key example of modern web-based approaches to the exploration of audiovisual archives [40]. Figure 1 presents the main view of the application, in which 103,273 shots from 6,969 videos are spatially distributed in 2-D based on similarities on visual features: color, shape, movement or visual complexity (as in the example above). With this interface, one can easily appreciate comparable shots and grasp the full collection at a glance, an impossible task with just a list of catalogue entries. Figure 1: Main view of the SEMIA application. Shots are spatially distributed by color using a t-SNE algorithm of the similarity between all shots (credits: Bert Spaan). Furthermore, traditional access to cultural collections shapes the memories we create of the past. These memories “are not inherent in the archival stock, but are created in the context of reception, through processes of remediation and recontextualisation” [9]. This relates to the idea that traditional metadata rely on prior assumptions of their authors, as well as on the initial goal of these descriptions. For an archival entity, a catalogue is first and foremost a way to document their collections, usually connected to a database system with an information retrieval tool to (at least internally) find the relevant items. This purpose is quite different than browsing through a collection without any prior knowledge (a goal that seems to be relevant for casual users, [39] and therefore solely relying on these metadata might not be enough. Browsing is indeed “a rich and fundamental human information behaviour” [10], an iterative process based on scanning [51] or glimpsing [3] a collection of items. These processes also depend on the questions one might ask, and it has been shown that revealing a collection in its entirety through the use of spatial distributions for instance can prompt innovative questions [47]. People “browse with or without a goal in mind, and goals may change as the process unfolds” [64]. Visualization is therefore at the center of an innovative framework to explore large audiovisual collections. Visualization, as “a medium for communication (or persuasion, or engagement)” and a tool for “understanding (or problem solving, planning, orienting)” [[]pp. 1-2]scagnetti2011visual, can reveal patterns, structures, relationships in the data and prompt new enquiries. Much like Moretti’s “distant reading” approach to literature, which can disclose hidden meanings in the text [45], visualization can expose new knowledge. For design and humanities scholar Johanna Drucker, visualization “produces” knowledge through “graphical forms expressing interpretation”, and that because of the “fundamentally interpreted condition on which data is constructed” visualizations are a feature of both “knowledge production and [its] presentation” [19]. Shneiderman’s “visual information seeking” approach entails a taxonomy of tasks users might want to perform while exploring a collection: overview, zoom, filter, details-on-demand, relate, history, and extract [58]. This interaction paradigm requires surrogates in the form of previews for single items and overviews for groups of items [23], to represent the collection objects while exploring it along its latent dimensions. According to [18], however, Shneiderman’s information visualization “mantra” is “pragmatic, but highly mechanistic” and supposes users with clear goals in mind, something that might not actually hold true for casual visitors in a museum space. This ubiquitous task-based approach, widespread in the fields of human-computer interaction and information retrieval, does not meet the criteria of a more humanistic approach. It is furthermore essential to recognize the inherent reward component and creativity aspect in the action of browsing a collection [51]. While browsing, a user is the sole director of the experience (although they might be more or less consciously guided towards a certain path) and is therefore creating their own narrative. It is a very different thing to go through a curated list of items and discover the same items while autonomously exploring the collection. In the latter case, serendipity, “the fact of finding interesting or valuable things by chance” [50], plays a major role. The feeling of finding new items by chance, of “serendipitous” discoveries [63], entails a procedural emergence of narrative, driven by the user-agent of the installation. All these concepts are summarized by Whitelaw’s innovative notion of “generous interfaces” [64]. As he argues, searching requires “rich, browsable interfaces that reveal the scale and complexity of digital heritage collections”. It should be a “humanistic model of interface and interaction that emphasises exploration and interpretation over task and information retrieval”. The modes of visual storytelling offered by these generous interfaces are simultaneously “horizontal” (through the browsing features) and “vertical” (through the details-on-demand functionalities), more or less completely driven by the user. Whitelaw’s interfaces empower users to generate, to craft new knowledge, through the narrative they are creating. Psychologist Jerome Bruner discusses the importance of narrative for its fundamental role in creating and interpreting human culture [8]. He states that human beings are natural storytellers: they make sense of the world and themselves through narrative. From the time they are very young, children learn that the way to integrate their own desires with their family’s norms and rules is to construct a story about their actions. This push to construct narrative, Bruner maintains, shapes how children acquire language, and the habit persists into adulthood as a primary instrument for making meaning. These storytelling skills insure our place within human society. This point is sustained by constructionist theories [14, 48], according to which individuals do not learn by passively perceiving content but rather by actively crafting, manipulating and therefore creating new knowledge. Through their rich browsing features and interactivity, generous interfaces evidently offer ideal modes of access to large digital collections, which are far more interesting than traditional information retrieval systems. These interfaces are however mainly web-based and therefore restrained to single users in front of small and flat screens. One could argue the immersive component of IVS is completely lacking here, as well as the multi-user aspect. The framework we suggest would solve these issues through the use of Immersive Environments, relying on the concept of an immersive generosity. By transposing the generous interface concept to IEs, one must however consider how it will affect the narrative. Indeed, a story is closely correlated with the medium used to convey it. This is perfectly illustrated with cinematic adaptations of books: the overall story being told is perhaps the same, but the way it is told, its content and its intensity can greatly differ. Applying this idea to IVS, it is clear that the active role a user has in an immersive and embodied application vastly influences the way narrative can emerge, requiring a distinction between purely authorial storytelling and interactive approaches. [2] propose a useful model for this with four dimensions to characterize the narrative component of different mediums: “Contingency” (the contingency of time and space of the story being told with respect to the real time and space of the user); “Presence” (how much the user feels present in the story); “Interactivity” (the degree of controls they have on the narrative) and “Narrative Representation” (the form narrative adopts, be it through mental models for literature for instance or purely visual and aural for cinema). They argue that, when compared to the most common cases of narrative mediums, namely literature, cinema and theatre, virtual reality offers the greater contingency in time and space, the strongest feeling of “being there” and the highest degree of interactivity. Combined, these considerations imply that IVS is a form of narrative that moves beyond the Platonic concepts of “diegesis” and “mimesis” (based on an authorial view of storytelling) to the idea of “experiencing” and “creating” a story. When focusing on multi-user experiences, typically offered by large interactive systems, the concept of a user-led narrative has even greater implications for the way the other users in the interactive space experience the story being created. In Geert Mul’s work, users can simultaneously be seen as “highly productive, in that the appearance of the works changes based on their input” and “merely one in a much larger series of variables that determine the outcome of the calculation” [46]. Although these views might seem at odds, they both imply that, when IVS is user-led, it requires an external public (other users) to interpret it and appreciate it fully, through a “third-person’s perspective”. One must also remember that museums are historically social venues, and the relationships between visitors are intrinsic to the experience of exploring their collections, meaning that these principles must be applied to to the exploration of large datasets in multi- user immersive installations. In conclusion, Whitelaw’s “generous interfaces” have been highly influential as a first attempt to solve the issues outlined by archival and digital humanities scholars in the access to large audiovisual collections. Nonetheless, we argue that these interfaces would benefit from a further component of immersion, through the use of IEs and thus IVS, to create a truly embodied exploration of a cultural archive. To conceptually frame this idea, philosopher Mark Johnson’s theory of the body is leveraged in the next section to better frame how narrative can emerge from such installations through “embodied understanding” [27]. ## 3 Embodied Understanding in Immersive Environments For [27], the importance of “embodied understanding”, based on the 20th century findings of cognitive science, challenges centuries during which the body was considered less important than the mind. He provides the counter- argument that “understanding is profoundly embodied, insofar as our conceptualization and reasoning recruit sensory, motor, and affective patterns and processes to structure our understanding of, and engagement with, our world”. It is therefore clear that IVS, through its embodied approach, can indeed be beneficial to cultural institutions aiming to give the wider public meaningful access to their collections. To put these ideas in practice in IEs-based installations, one must however first appreciate how our understanding of the world is embodied. According to the field of embodied cognition, organism-environment interactions are the sources of all our human perception and understanding of the world [15]. To really understand something, we must first experience it, a complex process based on Damasio’s “homeostatis” balanced state between organism and environment that comprises both how the organism is feeling and acting as well as how the environment is structured [13]. This equilibrium is dynamic, because it evolves as the organism and the environment evolve, and is also related to the quality, the value of the experience and overall our well- being. This is why emotions are such an important element of our “embodied understanding” of the world. Neuroscientists have further shown how organisms, through the detection of “emotionally competent stimulus”, move their body- states to favorable positions for their survival and well-being [12]. Emotions are therefore at the core of understanding, something that seems quite obvious in storytelling, as any story plays with our emotions to convey its narrative. Going one step further, [26] talks of the “bodily sources of meaning”. Humans have indeed always used their bodies to express themselves, from spontaneous gestures while talking to more elaborate performances such as dance or ritual practices [25, 67]. It is thus necessary to define this “body” of ours, and relate its dimensions to IVS in the frame of the exploration of large cultural collections. Drawing on Merleau-Pointy’s _Phenomelogy of Perception_ and John Dewey’s somatic naturalism, Johnson explains how our bodies are not just “objects interacting with other objects” but are “lived”, “phenomenal bod[ies]” [43], and require at least five intertwined dimensions to be fully comprehended: the biological, ecological, phenomenological, social and cultural body. Going back to Damasio’s homeostatis state, our bodies are first and foremost “flesh-and-blood”, “functioning biological organism[s] that can perceive, move within, respond to, and transform [their] environments” [15]. Our “biological bodies” are in a continuous exchange with their environments, continuously evolving the aforementioned equilibrium, and are the locus of feelings and emotions that push us towards our physical and social well-being [12]. When being confronted with an immersive narrative experience, users are therefore first and foremost a biological body, with all their individualities and body specificities. In her interactive and immersive experience _Osmose_ 444http://www.immersence.com/osmose/ (1995), media artist Char Davies empowers visitors to automatically drive the narrative through their breathing and balance, two fundamentals and unconscious human activities. These are part of the preconscious activities [22] identifies as the “body schema” that govern our interactions with the environment. Once users start to actively engage and interact with an installation, the second dimension of the body emerges: the “phenomenological body”. Our “tactile-kinaesthetic body” [56] depends on proprioception (our feeling of our bodily posture and orientation), our kinaesthetic sensations of bodily movement and our awareness of our internal body states through our emotions and feelings [12]. In contrast to the body schema, [22] associates our more conscious activities to the “body image”: activities that comprehend the affordances of the system to explore the collection, Drucker’s “conventions of the diagrammatic knowledge form” [19]. Furthermore, this phenomenological paradigm can result in the enhancement of a cognitive operation through a shift in the nature of the task itself, where abstract operations (such as finding all the items that correspond to a certain query) can be mapped to more natural actions, the so-called “tangialities” [44]. Dario Rodighiero and his colleagues at metaLab at Harvard have pushed this dimension to the extreme of posing the entire body as the “interaction device”, a sort of “choreographic interface”. In _Surprise Machines_ 555https://dariorodighiero.com/Surprise-Machines-for-Harvard-Art-Museums, visitors explore Harvard Art Museum collections through the use of precise and choreographed gestures [52], each one mapped to a specific task that remind us of Shneiderman’s visualization mantra: “overview first, zoom and filter, then details-on-demand” [58]. Figure 2 shows the digital collection spatially distributed according to visual similarity of the different items and the gesture vocabulary defined to explore this latent space. Figure 2: _Surprise Machines_ media installation with the museum’s collection spatially distributed on the left and the gesture vocabulary on the right (credits: metaLAB (at) Harvard). In this installation, the digital collection is shaped to create a virtual environment that users can freely explore through their bodies. This intrinsic relationship to the environment is elucidated by a third dimension: the “ecological body”, which can be defined as the continuous process between our bodies and the environment we evolve within [15, 43]. The twofold phenomena of “embodied understanding’ and the “ecological body” is pushed even further in IEs such as the EPFL Laboratory of Experimental Museology’s Panorama+ or its predecessor, the UNSW’s iCinema Research Centre’s landmark system Advanced Interaction and Visualization Environment (AVIE). The Panorama+ is a 360-degree stereoscopic, interactive environment of ten meters diameter and four meters high, with five projectors and surround sound audio system, controlled by a cluster of six computers (see [42] for a more in-depth technical description of the AVIE). Its omnidirectional nature recreates a fully immersive data space that allows for both allocentric (relationships between objects) and egocentric (personal relationships to objects) cognition and spatial perspectives simultaneously [5], cited in [30]. Visitors can indeed physically represent themselves with respect to digital objects in the virtual landscape and appreciate relationships between these digital items, blurring the line between what is real and what is virtual. This phenomenon increases their sense of presence and thus enhancing the narrative component of the exploratory experience. The digital collection is no longer an ascetic database or list of catalogue entries but a fully-fledged environment that users can explore on their own terms and to which they can individually or collectively relate with other visitors, in the case of multi-user systems. It is in these multi-user experiences that the fourth dimension of the body Johnson outlines, the “social body”, is revealed. Indeed, our environment is not just physical or biological but also composed of human relationships and interactions with our peers. In the field of developmental psychology, the effect (and importance) of other people during our childhood is well-known [60], and this continues as adults, through interactions with our colleagues and friends. This is especially true in social spaces such as museums, where users are generally not alone but rather continuously confronted with the presence of others, usually strangers, who become the “spectators” in the trichotomy “system-user-spectators” [17]. Embodiment, on this basis, can thus be argued to have a “participatory” status where the user driving the experience becomes the author of narrative for a larger public. Here again, the “third-person’s perspective” Geert Mul mentioned in his work, as well as the dual view of the role of users in generative interactive pieces are crucial [46]. From the perspective of performance studies theory, “it is the ways in which the user perceives and experiences the act of interacting with the system under the potential scrutiny of spectators that greatly influences the interaction as a whole … it is precisely this awareness of the (potentiality of a) spectator that transforms the user into a performer” [11]. Similarly, the latest theoretical frameworks on creativity highlight the importance of the “public” in the creative act [61]: the interaction with a given system in this case. This creative act of interacting with an installation also has an important effect on the “embodied understanding” of the collection, as put forward by constructionist theories on learning [14, 48]. Individuals actively engaged with the knowledge they are being presented will learn more than those passively witnessing it. Bloom’s famous taxonomy of educational objectives supports this claim, since its main categories include Application, Analysis, Synthesis and Evaluation (activities that require the manipulation and creation of knowledge), in addition to Knowledge and Comprehension [6]. One of the more modern revisions of this taxonomy puts emphasis on the notion of creating new knowledge by splitting the original classification into two dimensions: the first on the actual knowledge being addressed, and the second on the cognitive processes applied to this knowledge [34]. Furthermore, empirical evidence suggests the hierarchical characteristic of these taxonomies, placing the category of Create at the top [1]. The fifth and last dimension of the body Johnson describes is the cultural one. He argues that various cultural aspects contribute to the shaping of our bodies and the way we see and relate to them. This explains why gestures and postures vary across the world, as well as our attitudes towards our environment. Cultures are enacted through rituals, practices, customs performed by humans as inherently embodied beings [26]. Therefore, in IVS, the emergent narrative will be enacted by the embodied users, attended by their prior knowledge and specific cultures. The interpretative action of the visitors mentioned before depends on the individual preconceptions people bring to the experience, and consequently the emergent narrative that results from their interactions can greatly differ. The dialogue between system and users is driven by these interpretations, users with specific backgrounds will draw connections that might appear rather strange to other visitors but nonetheless “make sense” in their specific dialogue, in their specific narrative. This fundamentally individual interpretative action reverts back to the social body and Mul’s “third-person’s perspective”, where the individual differences between users spark seemingly infinite combinations and unfolding of different narratives. Viewers’ own cultures and previous knowledge become new variables in the process of generating these narratives, both for the user actually driving the experience and for the public interpreting it. Furthermore, the cultural aspect is particularly relevant when exploring audiovisual archives, because the “immaterial manifestations of culture” (what we refer to as Intangible Cultural Heritage) can be captured and documented through videos, so that exploring such a collection amounts to exploring culture itself (or at least an aspect or portion of it). The power of IEs to “plac[e] users inside the dataset” [57] can thus immerse them in a cultural setting, with the various benefits for their experience, both for their engagement and for their learning. This full immersion in a collection creates an “embodied theatre of participation” that “permits an unprecedented level of viewer co-presence in a narrative-discovery of a cultural landscape”, facilitating “dynamic inter-actor participation and cultural learning” [29]. The need for an innovative framework having been outlined and said framework being situated within embodiment theory, this thinking will now be applied to analyze and critique two previously-built installations, to highlight the particularities and advantages of our proposition. ## 4 Critique of Interactive Installations in Light of the Proposed Framework: _T_Visionarium II_ and _Jazz Luminaries_ The need for an immersive generosity having been justified and the concept of embodied understanding explained through Johnson’s theory of the body, we now illustrate how this innovative framework we propose can be applied to previously built installations created to explore large video collections. These use cases will first be presented and then critiqued in light of the proposed framework, in an attempt to formalize it and draw conclusions on what the next stage of immersive interfaces for exploring large video collections should aim for. The two installations discussed here are _T_Visionarium II_ and _Jazz Luminaries_. _T_Visionarium II_ is part of the _T_Visionarium_ project, developed at the UNSW’s iCinema Research Centre between 2004 and 2017, resulting in three iterations of the work666See http://www.icinema.unsw.edu.au/projects/t_visionarium/project-overview/ for a project overview.. The first version was developed for an inflatable dome structure [4], while the more advanced second and third versions use the AVIE system, as it can be seen in Figure 3. The interactive and immersive installation explores 24 hours of television footage, segmented, manually annotated (based on a thesaurus) and transformed in a database of more than 20,000 clips of a few seconds each. The system is meant for a single user who navigates it with a touch tablet and with the larger audience witnessing the emergent narrative. Hundreds of clips are simultaneously playing on the 360-degree screen, and when the user-agent selects one, the digital landscape rearranges itself, mapping semantically similar clips closer to the selected item (based on the annotated metadata). The selected clips can be recombined together, rewriting the linear narrative of the original footage (already broken down by the initial segmentation) and resulting in a “recombinatory” or “transcriptive” narrative [7]. Figure 3: View of the _T_Visionarium II_ installation (credits: Sarah Kenderdine) _Jazz Luminaries_ , on the other hand, is a much more recent project. Part of the _Infinity Room II_ exhibition, the application was developed by EPFL’s Laboratory of Experimental Museology in 2019, in a full dome structure of six meters diameter. 13,000 videos of the Montreux Jazz Archive are arranged in a network where nodes represent artists and links their collaborations during the festival. The network can be navigated with a spherical controller, mimicking the structure of the dome, and when passing over a certain node, its corresponding sound excerpts rapidly plays, resulting in an acoustic search akin to radio channel surfing. Lying down under the dome, users thus explore the archive and when an artist is selected, they can choose a specific performance to finally reach a fractal view of the corresponding video (allowing users to appreciate it in spite of their relative position under the dome). As in _T_Visionarium II_ , one spectator drives the experience with the spherical controller, as shown in Figure 4, while the others appreciate the unfolding of narrative, reclined under the dome [31]. Figure 4: View of the _Jazz Luminaries_ installation (credits: Sarah Kenderdine) Both installations evidently pivot on a browsing experience browsing experience, where the user-agent explores hundreds of videos simultaneously, gradually revealing details of the collection. The omnidirectional immersion offered by the AVIE system relates to Johnson’s idea of the ecological body, as users are “inside” the dataset, affected and affecting it through their interactions. Similarly, the dome structure recalls a rich history of planetarium structures with the obvious metaphor of the sky vault, a history rich of hundreds of years where the goal is to cover the human field of view. By completely covering the surface of the screen with videos or images, users are naturally inclined to turn their head and look around, as if they were physically exploring a real room filled with archival content, echoing the ecological and phenomenological body through this kinaesthetic paradigm. The first iteration of the _T_Visionarium_ project put even a stronger emphasis on this phenomenological aspect, since the visitor wore a head-tracking system enabling the projected portion of the dome surface, calibrated to the orientation of their head. While turning around, they would therefore visually unveil the extent of the database [4]. The phenomenon of wandering around in the archive, without a clear path to follow, calls up the model of the “information flaneur” [16]. Curious, creative and critical, the goal of this information-seeker modeled on the 1840s Paris urban flaneur is not to find something in particular but rather to appreciate the collection as a whole, and to be surprised by what they have stumbled upon and ultimately simply be immersed in the archive. Furthermore, this serendipitous search paradigm relates to the idea that “information is organic” [41], and hence should be explored in an organic way. In _Jazz Luminaries_ , such a natural aspect of the data is also suggested through the acoustic search, that draws on the biological body and references the concept of “tangialities”, that are not just related to the touch but include all the five senses [44]. The social interactions that might arise from a single user driving the experience for a larger public recall the social body as well as Mul’s “third-person’s perspective”, since one could argue it is the public that is experiencing the full performance, defined by the emergent narrative that the user-agent is creating. There is, as such, no predefined narrative but an infinite combination of items sequences only bounded by viewers’ interpretative power. While browsing the archive, it is as if users were “sculpting” its contents [33], shaping the experience the way they wish, while being guided by the relationships between elements based on the metadata without being constrained by them. It is clear that such a paradigm could be augmented with modern computational approaches, allowing to generate much more intricacy and thus possibilities in the database, in particular if relying also on more visual features, such as in the SEMIA project illustrated previously [40]. Finally, the cultural body is inherent in the exploration of a cultural collection, and entails that prior individual knowledge as well as different cultures are yet another variable in the process of generating narrative. Suffice to say, that a visitor actually having attended one of the Montreux Jazz Festival concerts and then re- experiencing it reclined under the dome will have a much different experience than those discovering the Festival for the first time. Recalling Aylett and Louchart’s narrative theory, _T_Visionarium II_ and _Jazz Luminaries_ have a strong contingency in time and space: first, because the narrative is completely generative and determined by users’ interactions; secondly due to the important feeling of “being there” due to the full immersion in the archive and the idea of “sculpting” it; third, thanks to the high level of interactivity (at least for the user-agent driving the experience). In this way, IVS adopts a form of narrative based on the concept of creating a story through the millions of paths embedded in the latent structures of the collections. Through interactivity and the freedom to explore the horizontal axes of the archive, visitors obtain a strong authorial power on the narrative. In Geert Mul’s words, visitors are “highly productive, in that the appearance of the works changes based on their input” [46], and this entails the clear benefits in the learning experience maintained by constructionist theories. At the same time, it is important to ensure intuitive interaction frameworks are enabled to minimize the learning curve to adopt the system as well as not relying on users having clear goals in mind, something that cannot reliably be expected from casual visitors in a museum setting. In addition, as previously stated, museums are social venues, and interactions between visitors are welcomed and encouraged, such as viewers passing around the spherical controller in _Jazz Luminaries_. Johson’s concept of the social body is one of the reasons why multi-user shared spaces in IEs are arguably more interesting than traditional Head-Mounted Displays (HMDs) for IVS in museums spaces. Indeed, even though from a technical point of view, immersion might be higher while using HMDs, the stronger grounding in the physical reality offered by large display screen systems such as the Panorama+ / AVIE allows for a more humanistic approach to the collection, based on social proximity with other visitors. The two installations discussed, _T_Visionarium II_ (as well as its predecessor) and _Jazz Luminaries_ , have provided clear use cases to illustrate the innovative framework proposed. These projects have their limitations, however, often because they are based on traditional metadata rather than more intrinsic relationships related to strictly visual features (such as colors, visual complexity, movement…) that modern computer vision approaches can unveil. These restrictions are amongst the issues that the next generation of audiovisual archives browsers and projects like _Narratives from the Long Tail: Transforming Access to Audiovisual Archives_ are endeavouring to overcome. ## 5 Conclusion In this article, we present an innovative framework to explore large audiovisual collections through embodied experiences in immersive environments, drawing on Immersive Visual Storytelling theory, Whitelaw’s concept of generous interfaces, and Johnson’s theory of the body. The need for such a framework is highlighted by archival and humanities scholars as well as the GLAM sector. Two use cases have been analyzed in-depth and critiqued in light of our proposition, supporting our claim that a move towards a more immersive generosity will enhance the experience of visitors engaging with large cultural collections in museum settings through embodied understanding. The physicality of moving through the collection within which users are immersed, and modifying the way the system presents itself after each interaction, entails the metaphor of sculpting the data, as if it were a raw block of material offered to the viewer to create their own narrative. Furthermore, the social relationships that arise in such shared immersive spaces are key to obtaining a full picture of the embodied understanding of the kinds of cultural collections that people can experience in a museum. Nonetheless, work remains to be done in the field to improve access to further democratize these collections, defining the research that the _Narratives from the Long Tail: Transforming Access to Audiovisual Archives_ project intends to undertake as it maps such possibilities. ## 6 Acknowledgements This research is supported by the Swiss National Science Foundation through a Sinergia grant for the interdisciplinary project _Narratives from the Long Tail: Transforming Access to Audiovisual Archives_ , lead by co-author Prof. Sarah Kenderdine (grant number CRSII5_198632, see https://www.futurecinema.live/project/ for a project description). ## References * [1] Lorin W Anderson and David R Krathwohl “A Taxonomy for Learning, Teaching, and Assessing: A Revision of Bloom’s Taxonomy of Educational Objectives” Longman, 2001 * [2] R. Aylett and S. Louchart “Towards a Narrative Theory of Virtual Reality” In _Virtual Reality_ 7.1, 2003 * [3] Marcia J Bates “What is Browsing Really? A Model Drawing from Behavioural Science Research” In _Information Research - an International Electronic Journal_ 12.4 University of Sheffield, 2007 * [4] Jill Bennett “T_Visionarium : A user’s guide” University of New South Wales Press Ltd, 2008 * [5] Barry Blesser and Linda-Ruth Salter “Spaces Speak, Are You Listening? Experiencing Aural Architecture” MIT press, 2009 * [6] Benjamin Samuel Bloom “Taxonomy of Educational Objectives: The Classification of Educational Goals” In _Cognitive domain_ Longman, 1956 * [7] Neil C.M. Brown, Timothy S. Barker and Dennis Del Favero “Performing Digital Aesthetics: The Framework for a Theory of the Formation of Interactive Narratives” In _Leonardo_ 44.3, 2011, pp. 212–219 * [8] Jerome Bruner “Acts of Meaning” Harvard University Press, 1990 * [9] Dagmar Brunow “Curating Access to Audiovisual Heritage: Cultural Memory and Diversity in European Film Archives” In _Image [ &] Narrative_ 18.1 Open Humanities Press, 2017, pp. 97–110 * [10] Shan-Ju Chang and Ronald E Rice “Browsing: A Multidimensional Framework” In _Annual Review of Information Science and Technology (ARIST)_ 28 ERIC, 1993, pp. 231–76 * [11] Peter Dalsgaard and Lone Koefoed Hansen “Performing Perception—Staging Aesthetics of Interaction” In _ACM Transactions on Computer-Human Interaction (TOCHI)_ 15.3 ACM New York, NY, USA, 2008, pp. 1–33 * [12] Antonio R. Damasio “Looking for Spinoza: Joy, Sorrow, and the Feeling Brain” Houghton Mifflin Harcourt, 2003 * [13] Antonio R. Damasio “Self Comes to Mind: Constructing the Conscious” In _Brain_ , 2010 * [14] John Dewey “Democracy and Education: An Introduction to the Philosophy of Education” LondonNew York, Free Press Macmillan, 1966 * [15] John Dewey “The Later Works of John Dewey, Volume 1, 1925-1953: 1925, Experience and Nature” Carbondale: Southern Illinois University Press, 1981 * [16] Marian Dörk, Sheelagh Carpendale and Carey Williamson “The Information Flaneur: A Fresh Look at Information Seeking” In _Proceedings of the SIGCHI Conference on Human Factors in Computing Systems_ , 2011, pp. 1215–1224 * [17] Paul Dourish “Seeking a Foundation for Context-Aware Computing” In _Human–Computer Interaction_ 16.2-4 Taylor & Francis, 2001, pp. 229–241 * [18] Johanna Drucker “Performative Materiality and Theoretical Approaches to Interface” In _DHQ: Digital Humanities Quarterly_ 007.1, 2013 * [19] Johanna Drucker “Graphesis: Visual Forms of Knowledge Production” Harvard University Press Cambridge, MA, 2014 * [20] ENUMERATE “Survey Report on Digitisation in European Cultural Heritage Institutions 2014”, 2014 URL: http://www.enumerate.eu/fileadmin/ENUMERATE/documents/ENUMERATE-Digitisation-Survey-2014.pdf * [21] Giovanna Fossati “Found Footage Filmmaking, Film Archiving and New Participatory Platforms” In _Found Footage. Cinema Exposed. Amsterdam: Amsterdam University Press/EYE Film Institute Netherlands_ , 2012, pp. 177–184 * [22] Shaun Gallagher “How the Body Shapes the Mind” Clarendon Press, 2006 * [23] Stephan Greene, Gary Marchionini, Catherine Plaisant and Ben Shneiderman “Previews and Overviews in Digital Libraries: Designing Surrogates to Support Visual Information Seeking” In _Journal of the American Society for Information Science_ 51.4, 2000, pp. 380–393 * [24] Erkki Huhtamo “Illusions in Motion: Media Archaeology of the Moving Panorama and Related Spectacles” MIT Press, 2013 * [25] Mark Johnson “The Meaning of the Body: Aesthetics of Human Understanding” University of Chicago Press, 2007 * [26] Mark Johnson “What Makes a Body?” In _The Journal of Speculative Philosophy_ 22.3, 2008, pp. 159–169 * [27] Mark Johnson “Embodied Understanding” In _Frontiers in Psychology_ 6 Frontiers, 2015, pp. 875 * [28] David J Kasik et al. “Data Transformations and Representations for Computation and Visualization” In _Information Visualization_ 8.4 SAGE Publications Sage UK: London, England, 2009, pp. 275–285 * [29] Sarah Kenderdine “Somatic Solidarity, Magical Realism and Animating Popular Gods: Place-Hampi ”where intensities are felt”” In _2007 11th International Conference Information Visualization (IV ’07)_ , 2007, pp. 402–408 DOI: 10.1109/IV.2007.103 * [30] Sarah Kenderdine “Embodiment, Entanglement, and Immersion in Digital Cultural Heritage” In _A New Companion to Digital Humanities_ John Wiley & Sons, Ltd, 2015, pp. 22–41 DOI: 10.1002/9781118680605.ch2 * [31] Sarah Kenderdine “Experimental Museology: Immersive Visualisation and Cultural (Big) Data” In _Experimental Museology_ , 2021, pp. 15 * [32] Sarah Kenderdine, Ingrid Mason and Lily Hibberd “Computational Archives for Experimental Museology” In _International Conference on Emerging Technologies and the Digital Transformation of Museums and Heritage Sites_ , 2021, pp. 3–18 Springer * [33] Sarah Kenderdine, Jeffrey Shaw and Tobias Gremmler “Cultural Data Sculpting: Omnidirectional Visualization for Cultural Datasets” In _Knowledge Visualization Currents: From Text to Art to Culture_ London: Springer, 2013, pp. 199–220 * [34] David R Krathwohl “A Revision of Bloom’s Taxonomy: An Overview” In _Theory into practice_ 41.4 Taylor & Francis, 2002, pp. 212–218 * [35] Richard Kurin “Safeguarding Intangible Cultural Heritage in the 2003 UNESCO Convention: a critical appraisal” In _Museum International_ 56.1-2, 2004, pp. 66–77 * [36] Kris Layng et al. “CAVE: Making Collective Virtual Narrative: Best Paper Award” In _Leonardo_ 52.4 MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info …, 2019, pp. 349–356 * [37] Federico Lenzerini “Intangible Cultural Heritage: The Living Culture of Peoples” In _European Journal of International Law_ 22.1, 2011, pp. 101–120 DOI: 10.1093/ejil/chr006 * [38] Joan Llobera, Kristopher J Blom and Mel Slater “Telling Stories within Immersive Virtual Environments” In _Leonardo_ 46.5 MIT Press, 2013, pp. 471–476 * [39] Irene Lopatovska, Iris Bierlein, Heather Lember and Eleanor Meyer “Exploring Requirements for Online Art Collections” In _Proceedings of the American Society for Information Science and Technology_ 50.1 Wiley Online Library, 2013, pp. 1–4 * [40] Eef Masson, Christian Gosvig Olesen, Nanne Noord and Giovanna Fossati “Exploring Digitised Moving Image Collections: The SEMIA Project, Visual Analysis and the Turn to Abstraction” In _DHQ: Digital Humanities Quarterly_ , 2020 * [41] Helen McCorry “Museums, the Web and the Serendipity Facilitator” In _MDA Information_ , 2001, pp. 133–135 * [42] Matthew McGinity et al. “AVIE: a Versatile Multi-User Stereo 360° Interactive VR Theatre” Association for Computing Machinery, 2007 * [43] Maurice Merleau-Ponty “Phenomenology of Perception” Routledge, 2013 * [44] Slavko Milekic “Towards Tangible Virtualities: Tangialities” ERIC, 2002 * [45] Franco Moretti “Distant Reading” Verso Books, 2013 * [46] Geert Mul and Eef Masson “Data-Based Art, Algorithmic Poetry: Geert Mul in Conversation with Eef Masson” In _TMG Journal for Media History_ 21.2 Netherlands Institute for SoundVision, 2018 * [47] Christian Gosvig Olesen et al. “Data-driven Research for Film History: Exploring the Jean Desmet Collection” In _Moving Image: The Journal of the Association of Moving Image Archivists_ 16.1 JSTOR, 2016, pp. 82–105 * [48] Jean Piaget “To Understand is to Invent: The Future of Education”, 1973 * [49] Robert M. Pirsig “Zen and the Art of Motorcycle Maintenance” William MorrowCompany, 1974 * [50] Cambridge University Press “Definition of serendipity from the Cambridge Advanced Learner’s Dictionary and Thesaurus”, 2021 URL: https://dictionary.cambridge.org/dictionary/english/serendipity * [51] Ronald E Rice, Maureen McCreadie and Shan-Ju L Chang “Accessing and Browsing Information and Communication” MIT Press, 2001 * [52] Dario Rodighiero et al. “Surprise Machines: Revealing Harvard Art Museums’ Image Collection” In _Information Design Journal_ , forthcoming * [53] Katja Rogers, Uta Hinrichs and Aaron Quigley “It doesn’t Compare to Being There: In-situ vs. Remote Exploration of Museum Collections”, 2014 * [54] RTSArchives “Le nouveau site RTSarchives”, 2018 URL: https://www.rts.ch/archives/5919889-le-nouveau-site-rtsarchives.html * [55] G Scagnetti “Visual Epistemology for Design in Complex Systems, lecture at Tecnológica de Monterrey: Academic leaders’ program”, 2011 * [56] Maxine Sheets-Johnstone “The primacy of Movement” John Benjamins Publishing, 2011 * [57] Haifeng Shen et al. “Information Visualisation Methods and Techniques: State-of-the-art and Future Directions” In _Journal of Industrial Information Integration_ 16 Elsevier, 2019, pp. 100–102 * [58] Ben Shneiderman “The Eyes Have It: A Task by Data Type Taxonomy for Information Visualizations” In _The Craft of Information Visualization_ Morgan Kaufmann, 2003, pp. 364–371 * [59] Mel Slater and Sylvia Wilbur “A Framework for Immersive Virtual Environments (FIVE): Speculations on the Role of Presence in Virtual Environments” In _Presence: Teleoperators and Virtual Environments_ 6.6, 1997, pp. 603–616 DOI: 10.1162/pres.1997.6.6.603 * [60] Daniel N Stern “The Interpersonal World of the Infant: A View from Psychoanalysis and Developmental Psychology” Routledge, 2018 * [61] Robert J Sternberg and Sareh Karami “An 8P Theoretical Framework for Understanding Creativity and Theories of Creativity” In _The Journal of Creative Behavior_ 56.1 Wiley Online Library, 2022, pp. 55–78 * [62] Jonathan Steuer “Defining Virtual Reality: Dimensions Determining Telepresence” In _Journal of communication_ 42.4 Wiley Online Library, 1992, pp. 73–93 * [63] Elaine G Toms “Serendipitous Information Retrieval” In _DELOS_ , 2000 * [64] Mitchell Whitelaw “Generous Interfaces for Digital Cultural Collections” In _DHQ: Digital Humanities Quarterly_ , 2015 * [65] Florian Windhager et al. “Visualization of Cultural Heritage Collection Data: State of the Art and Future Challenges” In _IEEE Transactions on Visualization and Computer Graphics_ 25.6, 2019, pp. 2311–2330 DOI: 10.1109/TVCG.2018.2830759 * [66] R. Wright “The Future of Television Archives - Digital Preservation Coalition”, 2017 URL: https://www.dpconline.org/blog/wdpd/the-future-of-television-archives * [67] Tom Ziemke, Jordan Zlatev and Roslyn M Frank “Body, Language, and Mind” Walter de Gruyter, 2007
Remotely Operating a Single-Point LDV System to Acquire 2D Measurements Alec Ikei, Dr. Kaushik Sampath Acoustic Signal Processing and Systems Branch Acoustics Division January 27, 2021 Distribution A: Approved for public release: distribution unlimited NRL Memorandum Report01/01/2019 - 01/07/2021 Alec K. Ikei, Dr. Kaushik Sampath U.S Naval Research Laboratory 4555 Overlook Ave SW Washington, DC 20375 U U U UU Alec K. Ikei (202) - 404 - 4816 # Remote Operation of a Single-Point LDV System to Acquire 2D Measurements Alec K. Ikei Dr. Kaushik Sampath Due to the unprecedented increase in telework requirements, the motivation to further automate and remotely control experiments has become apparent. This work documents the technical development of creating a two-dimensional (2D) Laser Doppler Vibrometry (LDV) measurement using a single-point LDV system through an automated and remotely controllable process. This report aims to assist in rapid development of setups for similar use cases. The setup described is also modular, and has been used to analyze the modal response of samples actuated through air-based acoustic signals as well as those mechanically induced. ###### Contents 1. 1 Introduction 1. 1.1 Motivation 2. 1.2 Background 3. 1.3 A Brief Overview of This Work 1. 1.3.1 Input Excitation 2. 1.3.2 Output Measurement 3. 1.3.3 Data Analysis 2. 2 Equipment and Connections 3. 3 Data Acquisition Using LabVIEW 1. 3.1 FPGA 2. 3.2 RT Controller 3. 3.3 Desktop PC 4. 4 Data Analysis in MATLAB 1. 4.1 Loading and Processing Data 2. 4.2 Using FFT Data to Perform Modal Analysis 5. 5 Discussion 6. 6 Conclusion 7. A LabVIEW Data Acquisition Code 8. B MATLAB Parallel Processing 2D LDV Code 9. C MATLAB Custom Function: func_SSA 10. D MATLAB Custom Function: importVelocity Due to the unprecedented increase in telework requirements, the motivation to further automate and remotely control experiments has become apparent. This work documents the technical development of creating a two-dimensional (2D) Laser Doppler Vibrometry (LDV) measurement using a single-point LDV system through an automated and remotely controllable process. This report aims to assist in rapid development of setups for similar use cases. A key achievement of this work is the complete control integration of various types of hardware and communication protocols, i.e. Field-Programmable Gate Array (FPGA), function generator, translation stage, LDV and accelerometer. While several experiments may be ’executed’ remotely once the equipment parameters have been optimized in person, in this work, each tunable parameter of all the hardware can be adjusted remotely as well, therefore, eliminating that extra requirement. The setup described is also modular, and has been used to analyze the modal response of samples actuated through air-based acoustic signals as well as those induced mechanically. In the setup described in this work, an arbitrary waveform is set in a function generator, which is then amplified and played through a speaker in an opened transmission loss tube. The acoustic signal from the speaker travels through the air in the tube to vibrate the sample surface. The LDV is mounted on a 2D motorized platform, which is controlled through LabVIEW on a desktop computer. The LDV continuously sends velocity and signal strength data to a Compact Rio (cRio) data acquisition device, which records data upon trigger from the the function generator. The recorded data is then streamed and saved on the desktop computer. The Fourier transform of each measurement location is calculated in MATLAB, generating a 3D data array, consisting of two spatial dimensions and one dimension in frequency. A single frequency slice of data is taken from this array, which gives the amplitude of a 2D surface at a single frequency, also called a mode shape. These mode shapes can then be compared to expected results from COMSOL simulations. ## Chapter 1 Introduction ### 1.1 Motivation The current COVID-19 pandemic has required adaptation to a rapidly changing situation and long-term telework. To continue to do bench work, it makes sense to automate processes that are able to be automated, and allow for remote access and control of the testing environment. For example, one of our experiments required finely spaced vibration measurements across the surface of a sample, with various input excitations, measurement sampling rates and spacing. With only a single-point LDV system and a data acquisition system (DAQ), it would take many hours to take the measurements while also increasing the chance of human error. Instead of the experiment operator manually changing the excitation and measurement parameters, it is more efficient and convenient to automate, remotely control and monitor the experiment. Therefore, the function generator parameter setup was automated and integrated with the pre-pandemic point-scanning setup. An LDV system is widely used to measure the response of a system/sample to an excitation. Such experiments involve three broad stages: (1) sample fabrication and mounting, (2) excitation and (3) measurement scans. The sample preparation is entirely problem-specific and is usually not integrable with the latter two. A variety of different excitation mechanisms/inputs maybe employed for experiments. In our lab, we perform vibrational or acoustic excitations. Typically, once the sample is fabricated and mounted to the excitation mechanism in a location accessible to the LDV for scanning, it still requires a significant (at-work) time commitment to synchronize and optimize the excitation and LDV settings before the experiment can begin. Here, our remote operation of the LDV system integrates all the components of the excitation mechanism as well, allowing the user to remotely control excitation and LDV settings to optimize them and run the measurement scan. This includes control of the function generator, thereby adjusting the output of a vibrational exciter or sound tube. Commercial solutions to just the scanning stage of the problem exist for some specific applications, such as the 2D and 3D LDV systems sold by various companies. However, higher dimensional systems can cost on the order of 10 times more than a single-point system for each added dimension of measurement capability. For many lab setups, this may not be feasible, and in many situations do not allow for enough adaptability and control of the scan parameters. The methods described here can be reproduced at lower cost and less administrative time used to purchase a 2D LDV while still producing high quality results, in addition to customization and integration with the testing environment. Another crucial advantage of the current approach is that a vendor-built custom 2D scanning system is wedded to just the LDV measurement, and has no use during the machine’s down-time. Whereas, as a consequence of the present work, the 2D motorized stage can be used modularly for a wide range of other experiments, and the automation/remote-control components carry forward to several other case scenarios. For example, the motorized stage and function generator can be placed in an acoustic (hydrophone) scan experiment and moved back to the LDV setup as needed. ### 1.2 Background Historically, in the Acoustic Signal Processing and Systems Branch (Code 7160), we performed experiments using scanning-type measurements for underwater acoustic tests, using an automated system to move and acquire data from a hydrophone in 2D and 3D scans. The code used to run the underwater tests was modified to run on different hardware and connect to different systems, but the idea behind the measurement remains the same. By moving the data acquisition location around, the data collected represents the same data that would have been collected if using a large array of measurement devices. Conversely, the excitation source can be moved around with a constant data acquisition location, which would emulate having an array of sources. The data collected at each measurement point is usually then processed in MATLAB to extract binned time series data or frequency domain data. These analyses can then be represented as a propagating wave or modal shape, respectively. As additive manufacturing and featuring capabilities achieve finer resolutions, the demand for higher frequency scans has only been increasing. Moreover, an interest in studying non-linear acoustic or transient phenomena also requires large, dense collections of data points and high- frequency scans. These result in a substantial increase in the amount of data acquired per scan, necessitating parallel processing of the data processing step. ### 1.3 A Brief Overview of This Work #### 1.3.1 Input Excitation The testing setup described in this work sets an arbitrary waveform in a function generator, which is then amplified and played through a speaker in an opened transmission loss tube. The acoustic signal from the speaker travels through the air in the tube to vibrate the sample surface. The transmission tube can be replaced by a vibration exciter, if the desired actuation is mechanical rather than pressure based. #### 1.3.2 Output Measurement The LDV is mounted on a 2D motorized platform, which is controlled through LabVIEW on a desktop computer. The LDV continuously sends velocity and signal strength data to a cRio data acquisition device, which records data upon trigger from the the function generator. At the same time, the data from the accelerometer is also collected, so that the output can be normalized for each measurement location. The recorded data is then streamed and saved on the desktop computer. Since the controls for this setup are all on the desktop personal computer (PC), this allows remote desktop users to modify the input excitation, linear stages and DAQ parameters while teleworking. #### 1.3.3 Data Analysis The Fourier transform of each measurement location is calculated in MATLAB, generating a 3D data array. A single frequency slice of data is taken from this array, which gives the amplitude of a 2D surface at a single frequency, also called a mode shape. These mode shapes can then be further analyzed and compared to expected results in COMSOL simulations, as seen in Fig. 1.1. Figure 1.1: Vibrational modes of a circular elastomeric plate, produced using the setup described in this work and compared with a COMSOL model. Figure reused with permission wissman2019soft wissman2019liquid. ## Chapter 2 Equipment and Connections The desktop PC was connected through a Universal Serial Bus (USB) cable to the function generator. The signal output of the function generator was connected to the input of the amplifier through a Bayonet Neill-Concelman (BNC) cable. The output of the amplifier was connected to the input of the tube through banana cables. The impedance/sound tube kit can be substituted with a vibrational exciter when mechanical actuation is required instead. The PC was connected through RS-232 to the 2D stage controller. The LDV was mounted on a custom mounting bracket to the stage. The PC was connected to the DAQ through an Ethernet cable, and the DAQ modules were inserted into the DAQ chassis. The signal strength and velocity outputs of the LDV were connected to the analog input of the DAQ. The trigger signal from the function generator was connected through BNC to the digital module on the DAQ as well as the anti-drift input on the LDV controller. This causes the voltage output of the LDV to be set to 0V at each trigger, which helps to avoid the signal exceeding the maximum allowed by the LDV controller (10V). The trigger is also used by the DAQ to determine when to start recording the signal for each measurement location. The equipment and connection schematics are illustrated in Fig. 2.1, and their make and model are listed in Table 2.1. Equipment | Model | Manufacturer ---|---|--- Function Generator | 33500B | Agilent Technologies Power Amplifier | Type 2718 | Brüel & Kjær (B&K) Transmission Loss Tube Kit | Type 4206 | Brüel & Kjær (B&K) 2D Linear Stage | Bi-Slide | Velmex Inc. Single-Point LDV | CLV-2534 | Polytec Gmbh DAQ Chassis | cRio-9035 | National Instruments Analog Input Module | NI-9223 | National Instruments Digital I/O Module | NI-9402 | National Instruments Table 2.1: List of equipment and their respective makes and models, used to take a 2D LDV scan of the elastomeric plate seen in Fig 1.1. Figure 2.1: Equipment connections used to perform 2D LDV scan of the elastomeric plate seen in Fig 1.1. ## Chapter 3 Data Acquisition Using LabVIEW The description here assumes basic familiarity with LabVIEW. The code is run from several systems: the desktop PC, the Real-Time (RT) controller, and the FPGA inside of the controller chassis. The LabView code uses the FPGA and the C-series DAQ modules on the cRio to take data, passes it along through the RT controller and then the PC. In addition to the base LabVIEW program, the Embedded Control Suite is required to operate the FPGA. A flow chart showing the communication logic is displayed in Fig. A.1. LabVIEW block diagrams are shown in Appendix A. ### 3.1 FPGA In the FPGA code, the digital input is constantly read until it detects a rising edge from the function generator trigger. It then takes the analog voltage time series through a for-loop, and passes the values into a Direct Memory Access (DMA) first-in first-out (FIFO) channel. DMA FIFO channels are a way to communicate data at high sampling rates between the FPGA and the RT controller. The number of iterations of the for-loop and the timing between iterations is set by a control. Control values are modifiable from the RT code. The LDV signal strength is passed into an indicator. The DMA FIFO is also checked if it overflowed, and this value is passed onto an indicator. The overflow should be incorporated with a feedback node, so that successive reads of the overflow indicator do not reset the value. The FPGA code can be seen in Figure A.3. ### 3.2 RT Controller The RT controller opens the FPGA program, and sets the FPGA control values using the “Read/Write Control” function. The RT controller waits until the 2D stage has stopped, which is determined by a Boolean shared variable. Before and after data acquisition, the RT controller code checks if the 2D stage has stopped moving, and tells the PC that it is done acquiring data for that sampling location through Boolean shared variables. The FPGA code is then started, and the data is read from the DMA FIFO through a while loop. The while loop queries the DMA FIFO to check if there are elements stored on it. If there are, it reads the elements and places it in an array. The array is appended to previous iterations of the while loop. When the while loop is done, the array contains data from a single average. The for loop that surrounds the while loop repeats the single average process until it contains the data from the number of averages requested, as seen in Fig. A.4. ### 3.3 Desktop PC The desktop PC controls the 2D stage and the function generator, and coordinates with the RT controller to acquire data between movements of the LDV. In the first section of the PC code, the 2D linear stage is initialized, which gives the stage controller an origin to reference. The initialization sub-function of the LabVIEW Virtual Instrument (sub VI) is custom made, based on communication protocols supplied by the manufacturer. The function generator parameters and the arbitrary waveform file are read into a custom made sub VI, which utilizes drivers from the manufacturer. This is shown in Fig. A.5. In the next selected snippet, the PC code creates a raster grid based on user input in Fig. A.6. The raster grid is iterated through nested for loops, reading and saving data each time to the hard drive, as seen in Fig. A.7. The PC tells the RT controller when the movement is done through a Boolean shared variable, and waits for the data acquisition to finish before reading the data. Once it finishes saving, it moves on to the next iteration of the nested for loop. When the nested for loop is done, it tells the cRio that the scan is done, and closes the communication port for the function generator. Failure to close ports properly as mentioned here, will lead to communication issues in subsequent operation, which can usually be remedied by resetting the device. Some of the default settings on the function generator are not suited for our application. For example, the function generator applies a low pass filter at 17kHz, which limits the types of response that can be induced in the sample. By including code that can modify parameters on the function generator, these settings can be turned off, modified and automated, allowing for different parameter sets to be used when acquiring multiple data sets. ## Chapter 4 Data Analysis in MATLAB ### 4.1 Loading and Processing Data The data taken during the experiment was saved in a folder as separate comma delimited text files. In MATLAB, the working directory was changed to the folder containing the data files. Since the script described here used parallel processing to increase the processing speed, the MATLAB parallel processing toolbox is required to run it. In the script, the file names in the directory are read into an array, and the X and Y values are extracted from their names into two 1D arrays. The frequency array is calculated based on the sampling rate and the amount of zero padding that is desired. Zero padding refers to a signal processing technique in which zeros are added to the original time series data, which decreases the discrete frequency step size in the corresponding fast Fourier Transform (FFT). While increasing the zero padding of a data set provides an increasingly small frequency step size, the resulting FFT calculations taking increasingly more time to calculate, and therefore there is an inherent trade-off between the acceptable processing time and the minimum frequency step size achievable. A parallelized for (parfor) loop is then used to read the data and perform an FFT on all the files in the directory, resulting in a 2D array (each column is the FFT of a different position, and each row represents a different frequency). ### 4.2 Using FFT Data to Perform Modal Analysis In the next section of the script, a new directory is created, which is used later to save the figure files in. The frequency limit of the mode shapes to be displayed are set here, because the frequency range of interest is usually less than what the entire FFT contains. Using a parfor loop again, the images are generated from the FFT data. The X and Y arrays are used to index the FFT array, and the iteration counter of the parfor loop is used to select the row, which generates a 2D array of amplitudes that represent a single frequency. The 2D array is then plotted as a color map or a 3D mesh figure. Using this code, 500 images containing 143 by 136 spatial points (representing 19,448 sampling locations) were created in about 2 minutes on a computer with a 16 core processor. Further optimization can be carried out by doing parallel analysis on a Graphics Processing Unit (GPU) rather than on a Central Processing Unit (CPU) as was done here, which would allow many additional parallel processes, resulting in lower computation time. In comparison, without a parallel processing approach the code would have to run on a single processor, which would likely take approximately 16 times as long, saving half an hour for a relatively small subset of the data. With larger datasets or more detailed binning of the Fourier transform, the time saved would likely increase linearly, with a similar factor of performance improvement. The MATLAB code is included in Appendix B, and the custom functions used are included in Appendix C and D. ## Chapter 5 Discussion The method described here shows that a 2D LDV measurement can be done well with a single-point LDV and easily obtained components. The modular nature of the setup allows it to be used for different applications; to actuate a thin elastomeric plate as seen here through pressure waves in air, or through mechanical means to vibrate more rigid samples, using a vibration exciter and a stinger. In addition, some components can be swapped out for cheaper parts, especially the DAQ. A simple FPGA based DAQ card can be obtained on the order of $100, which is two orders of magnitude cheaper than a cRio. Uniquely for NRL, and perhaps many other research laboratories, FPGA DAQs are usually present with a lot of down time. The modularity of this solution can be extended to various existing hardware. Conversely, to acquire data at faster speeds, better damping on the linear stages and increasing its rigidity can be done to decrease the time needed for each sampling location. The number of averages and the number of time samples can also be modified to increase acquisition speed. However, this can affect the quality of the data produced, since noise generally decreases with the square root of the number of averages taken. Another important factor to consider is the sample mounting. Poorly mounted samples can lead to poor actuation of the sample. The sample shown in Fig. 1.1 has the elastomer bonded with an acrylic frame, which helps to ensure more uniform actuation. Especially in mechanical actuation (e.g. actuation of more solid objects, like plastic or metal plates), the contact between the vibration exciter, the stinger, and the sample must be snug. The signal strength value from the LDV system can also be used to filter out low signal sampling locations. In many cases, signal strength varies arbitrarily due to surface roughness effects, and more often a region of good signal strength maybe found infinitesimally close to a spot with poor signal strength. Therefore, for these scenarios, the LabVIEW code could also be modified to step a small amount in the X or Y direction upon measurement of a poor signal location. If these approaches are not feasible for a particular use case, then the data can also be spatially filtered to smooth out the figure surface. ## Chapter 6 Conclusion To summarize the process described in this work, a sample is vibrated using an arbitrary waveform played through a speaker. A 2D linear stage is used to move a single-point LDV, which takes the velocity time series of the sample’s surface. The time series is saved to individual text files, and the FFT is taken of each file. For each frequency bin in the FFT, the sampling position is used to assign its position in a 2D mesh. The mesh is then plotted either as a surface plot or as a colorplot, resulting in a mode shape plot. By using parallel processing, this process is sped up significantly, resulting in large time savings which increase with the complexity and size of the analysis. The modular nature of this setup also allows for modification and therefore versatility for different use cases and budgets. In addition to the time savings and low cost, the automation and remote operation of its sub components makes this setup even more appealing in times where lab access is restricted. As such, this report serves to document the technical process used to develop the measurement system to accelerate the development of similar setups and preserve institutional knowledge. The extent of remote operation can be further improved from this setup by employing a third-axis translation stage (in the direction of the LDV scanning beam). For optimal LDV signal strength, it requires the sample to be separated at discretized distances from the beam output location. The third stage can help achieve this remotely, and eliminate the need for in-person adjustment of the sample and/or LDV. Furthermore, when samples have a curvature, thereby varying that distance during a scan, the third axis translation stage can be used to dynamically adjust the LDV separation based on each sampling location to provide the best signal strength throughout the sample. The high level of detail and discussion of the code used are included to help unfamiliar engineers to be brought up to speed quickly. The LabVIEW and MATLAB codes used are shown in the appendices below, to give further guidance. ###### Acknowledgements. The LabVIEW code displayed here was built upon work performed by Dr. Michael Nicholas (NRL, Code 7165). We thank Wissman et. al. for permission to reuse their figure. We also thank Dr. Matthew Guild (NRL, Code 7165) and Dr. David Calvo (NRL, Code 7165) for their time spent reviewing this manuscript. ## References * Wissman et al. (2019a) J. Wissman, K. Sampath, A. Ikei, K. B. Özütemiz, C. Majidi, and C. A. Rohde, “Soft-matter pressure sensors for turbulence detection,” Proceedings of the Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems 2019, volume 10970 (International Society for Optics and Photonics), 2019a, p. 109702D. * Wissman et al. (2019b) J. P. Wissman, K. Sampath, A. Ikei, K. B. Özütemiz, M. Carmel, and C. Rohde, “Liquid metal-based resistive membranes for flow acoustics detection,” _The Journal of the Acoustical Society of America_ 146(4), 2997–2997 (2019b). ## Appendix A LabVIEW Data Acquisition Code Figure A.1: Communication flow chart between systems running LabVIEW code. Figure A.2: Organization of LabVIEW files. Opening the ”.lvproj” file opens this LabVIEW project window. Figure A.3: FPGA code, located under cRio $\rightarrow$ Chassis $\rightarrow$ FPGA Target. This code reads the digital input and moves to the second panel after receiving a rising edge. The second panel contains a for loop that collects the analog voltage signal and passes the value into the DMA FIFO at specifically timed intervals. If the FIFO overflows, it sets a boolean indicator to true. The signal strength of the LDV is also recorded as a single value. The controls can be adjusted from the RT controller code. Figure A.4: Selected portion of the Real-Time Controller Code, located under cRIO. This code runs the FPGA code, then reads the LDV strength into a shared variable. The DMA FIFO for each channel is queried. If there are elements remaining, they are read and passed out of the while loop into the for loop. The for loop continues until the number of averages taken is complete. The data is then averaged and passed into shared variables and the front panel display. Figure A.5: First part of the PC code, located under My Computer. This section initializes the 2D stage, so that it has an origin to reference. The waveform generator is also loaded with the arbitrary waveform. Figure A.6: Second part of the PC code, located under My Computer. This section generates the raster grid based on user input, and then moves the stage, iterating through nested for loops. Figure A.7: Next important part of the PC code, located under My Computer. This section reads the data stored int he shared variables, and saves them to text files. The precision is set by the format string ”%.6f”. ## Appendix B MATLAB Parallel Processing 2D LDV Code %%Attribution %Code written by Alec Ikei and Kaushik Sampath, November 2020 %Included in 2D LDV US Naval Research Laboratory Memorandum Report %% Read Data and Parallel Calculate FFT % get directory list clc; clear; close all; folder = uigetdir(); %prompts user to select folder containing data fileList = dir(fullfile(folder, ’*.tsv’)); %only lists .tsv files fileNames=strcat(folder,’\’,{fileList.name}); %extract the names of the files and %put them into a n x 1 cell array %User inputs samplingRate=100000; %sampling rate used in Hz padMultiplier=1; %number of TS lengths of 0s to pad minFreq=100; maxFreq=20000; %set frequency limits for slicing signalCutoff=0.05; %end of user inputs x=zeros(length(fileNames),1); y=x;z=x; badpxList=x; %length(a)=number of files badpx=0; %# of low signal str pixels %calculate the frequency vector for just one timeseries velocityArray=importVelocity(fileNames{1}); %get LDV data for particular %spatial location localfft=func_SSA(velocityArray,padMultiplier); %calculate FFT vector numofBins=length(velocityArray)/2*padMultiplier; %calculate # frequency bins localfreqBinSize=samplingRate*0.5/numofBins; %calculate bin width freqArray=0:localfreqBinSize:samplingRate/2; %generate frequency array %locate index of max frequency dist = abs(freqArray - maxFreq); minDist = min(dist); maxFreqidx = find(dist == minDist); %locate index of min frequency dist = abs(freqArray - minFreq); minDist = min(dist); minFreqidx = find(dist == minDist); %Pre-allocate variable for FFT array fftArray=zeros(length(localfft(minFreqidx:maxFreqidx)),length(fileNames)); Ψ%for FFT in columns parfor m=1:length(fileNames) %parallel computation of FFTs m %shows current iteration in command window indexX=regexp(fileNames{m},’X_’); indexY=regexp(fileNames{m},’Y2_’); indexS=regexp(fileNames{m},’Strength’); indexP=regexp(fileNames{m},’.tsv’);%location from filename if str2double(fileNames{m}(indexS+8:indexP-1))>signalCutoff Ψ% only use data if signal strength is good x(m)=str2double(fileNames{m}(indexX+2:indexY-1)); ΨΨ% X coordinate from filename y(m)=str2double(fileNames{m}(indexY+3:indexS-1)); ΨΨ% Y2 coordinate velocityArray=importVelocity(fileNames{m}); ΨΨ%LDV data for particular spatial location localfft=func_SSA(velocityArray,padMultiplier); ΨΨ%calculate FFT fftArray(:,m)=localfft(minFreqidx:maxFreqidx); ΨΨ%place relevant FFT data in columns else badpx=badpx+1; %count number of excluded pixels badpxList(m)=1; %list bad pixels end end x(badpxList>0)=[];y(badpxList>0)=[];fftArray(:,badpxList>0)=[]; %remove bad pixels fftArrayMax=max(max(abs(fftArray))); %global maximum amplitude disp(’done reading data’); %% Create color plot of each slice and save as .png pngfolder=strcat(folder,’\pngs’);figfolder=strcat(folder,’\figs’); %folders for images mkdir(pngfolder); mkdir(figfolder); xx=unique(x,’rows’);yy=unique(y,’rows’); %remove duplicate coordinates %creates a 2d set of points to assign the sampNormPlt values to [X,Y]=meshgrid(min(xx):xx(2)-xx(1):max(xx),min(yy):yy(2)-yy(1):max(yy)); %pre-allocate space to save all surface data surfaceData=zeros(length(minFreqidx:maxFreqidx),length(yy),length(xx)); currentFreq=minFreqidx:maxFreqidx; parfor k=1:length(currentFreq) %must start at 1, since parfor f=scatteredInterpolant(x,y,fftArray(currentFreq(k),:)’); Ψ%create function from data Z=f(X,Y)-mean(f(X,Y)); %create surface values from function, minus DC offset surf(X,Y,real(Z),’LineStyle’,’none’); %create surface plot view(2); %flat view imageName=strcat(num2str(freqArray(currentFreq(k))), ’ Hz’); print(’-dpng’,’-r75’,strcat(pngfolder,’\’,imageName)); Ψ%convert current fig to png saveas(gcf, strcat(figfolder,’\’,imageName)); Ψ%saves the current figure as Matlab file end ## Appendix C MATLAB Custom Function: func_SSA This function takes the single sided amplitude complex Fourier transform. It also can zero pad the timeseries input, to change the frequency binning. Used in the parallel processing 2D LDV code. function [SSA,Phase] = func_SSA(X,k) %k=1; changed from kaushiks code % Sampling Length L = length(X); % Pad with a length of k times L Lpad = k*L;X1 = zeros(Lpad,1); X1(Lpad/2-L/2+1:Lpad/2+L/2)=X; X=X1; % Compute FFT FFT1 = fft(X); Phase=imag(FFT1); % Single Sided Amplitude Spectrum FFT2 = FFT1/Lpad; SSA = FFT2(1:Lpad/2+1,:); SSA(2:end-1,:) = 2*SSA(2:end-1,:); SSA = SSA/(L/Lpad); ## Appendix D MATLAB Custom Function: importVelocity This function reads the text data file and converts it into a vector. Used in the parallel processing 2D LDV code. function velocityArray = importVelocity(filename, dataLines) if nargin < 2 dataLines = [1, Inf]; end %% Setup the Import Options opts = delimitedTextImportOptions("NumVariables", 2); % Specify range and delimiter opts.DataLines = dataLines; opts.Delimiter = "\t"; % Specify column names and types opts.VariableNames = ["VarName1", "Var2"]; opts.SelectedVariableNames = "VarName1"; opts.VariableTypes = ["double", "string"]; opts = setvaropts(opts, 2, "WhitespaceRule", "preserve"); opts = setvaropts(opts, 2, "EmptyFieldRule", "auto"); opts.ExtraColumnsRule = "ignore"; opts.EmptyLineRule = "read"; % Import the data tbl = readtable(filename, opts); %% Convert to output type velocityArray = tbl.VarName1; end
# ORIENT: A Priority-Aware Energy-Efficient Approach for Latency-Sensitive Applications in 6G Masoud Shokrnezhad1, and Tarik Taleb1, 2 1Oulu University, Oulu, Finland; {masoud.shokrnezhad<EMAIL_ADDRESS> 2Ruhr University Bochum, Bochum, Germany<EMAIL_ADDRESS> ###### Abstract Anticipation for 6G’s arrival comes with growing concerns about increased energy consumption in computing and networking. The expected surge in connected devices and resource-demanding applications presents unprecedented challenges for energy resources. While sustainable resource allocation strategies have been discussed in the past, these efforts have primarily focused on single-domain orchestration or ignored the unique requirements posed by 6G. To address this gap, we investigate the joint problem of service instance placement and assignment, path selection, and request prioritization, dubbed PIRA. The objective function is to maximize the system’s overall profit as a function of the number of concurrently supported requests while simultaneously minimizing energy consumption over an extended period of time. In addition, end-to-end latency requirements and resource capacity constraints are considered for computing and networking resources, where queuing theory is utilized to estimate the Age of Information (AoI) for requests. After formulating the problem in a non-linear fashion, we prove its NP-hardness and propose a method, denoted ORIENT. This method is based on the Double Dueling Deep Q-Learning (D3QL) mechanism and leverages Graph Neural Networks (GNNs) for state encoding. Extensive numerical simulations demonstrate that ORIENT yields near-optimal solutions for varying system sizes and request counts. ###### Index Terms: 6G, Resource Allocation, Energy Consumption, Service Placement and Assignment, Path Selection, Prioritization, E2E Latency, Age of Information (AoI), Reinforcement Learning, Q-Learning, and Graph Neural Networks (GNNs). ## I Introduction The advent of the 6th generation of telecommunication systems (6G) signifies a pivotal era marked by unparalleled connectivity and technological advancements. With ultra-low End-to-End (E2E) latency (less than $1$ milisecond), exceeding $1$ terabit per second peak data rates, and ultra-high reliability surpassing $99.99999\%$ [1], 6G promises to revolutionize industries such as holographic telepresence utilizing extended reality [2], dynamic metaverse empowered by semantic communications [3], and quantum networking [4]. However, achieving these capabilities raises substantial energy consumption concerns for both computing and networking resources. Presently, these resources consume around $200$ terawatt-hours of electricity annually, approximately $1\%$ of the global total [5]. Many quality-sensitive applications may require uploading up to $50\%$ of data to computing facilities for processing [6], adding even more strain to computing and networking resources. Moreover, the projected surge in 6G-connected devices and global data exacerbates the energy consumption challenge, underscoring the need for sustainable solutions. In order to realize a 6G-enabled future, it may be necessary to create novel resource orchestration mechanisms to address impending energy challenges. The subject has been extensively studied in the literature. Xuan et al. [7] addressed the Service Function Chaining (SFC) problem with the objective of minimizing energy consumption by proposing an algorithm based on multi-agent Reinforcement Learning (RL) and a self-adaptive division strategy. Solozabal et al. [8] investigated the same problem and proposed a single-agent solution. Other authors have also examined the SFC problem. By proposing a sampling- based Markov approximation method, Pham et al. [9] solved the problem in an effort to minimize operational and traffic energy consumption. Santos et al. [10] developed two policy-aware RL algorithms based on actor-critic and proximal policy optimization to maximize availability while minimizing energy consumption. Reducing energy consumption was considered in the Service Function (SF) placement problem as well. Sasan et al. [11] presented a heuristic algorithm to tackle the joint problem of network slicing, path selection, and SF placement, with the objective of maximizing user acceptance while minimizing energy consumption. Farhoudi [12] and He et al. [13] investigated a comparable problem and proposed RL-based solutions, taking into account the dynamic nature of service requests and overall cost considerations (including operation, deployment, and transmission), respectively. While effective in specific contexts, the mentioned methods may not be suitable for 6G systems. These approaches prioritize energy efficiency over maximizing device support, whereas achieving an E2E efficient solution requires holistic management of computing and networking resources, considering the stringent Quality of Service (QoS) demands of 6G. Furthermore, certain studies overlook or oversimplify critical network parameters like latency, which contradicts the intricate and dynamic requirements of 6G systems. This paper addresses this gap by investigating the joint problem of allocating computing and networking resources (service instance placement and assignment, path selection, and request prioritization), termed PIRA. The objective is to optimize the system’s overall profit (as a function of supported concurrent requests) while minimizing energy consumption over time, accounting for E2E latency and resource capacity constraints. The $M/M/1$ queuing model is employed to accurately evaluate request latency on compute nodes and network devices. To solve this problem, we propose ORIENT, an approach leveraging Double Dueling Deep Q-Learning (D3QL) reinforced by Graph Neural Networks (GNNs). This hybrid method effectively encodes the system state and facilitates the identification of near-optimal solutions. The remainder of this paper is organized as follows. Section II introduces the system model. PIRA is defined and formulated in Section III, and ORIENT is presented in Section IV. Finally, numerical results are illustrated in Section V, followed by concluding remarks in Section VI. ## II System Model As shown in Fig. 1, the following is an explanation of the two main components of the system: resources and requests. Figure 1: The system model, including network devices and distributed compute nodes facilitating holographic telepresence services for end users. ### II-A Resources The 6G system examined in this paper is an integrated infrastructure of computing and networking resources comprised of $\mathcal{N}$ network devices and $\mathcal{V}$ compute nodes (radio resources are excluded) [14]. $\mathbb{N}=\\{n|0\leq n\leq\mathcal{N}\\}$ is the set of network devices, and $\mathbb{V}=\\{v|0\leq v\leq\mathcal{V}\\}$ denotes the set of compute nodes. Compute nodes are connected through network devices via $\mathcal{P}$ paths, the set of which is denoted by $\mathbb{P}=\\{p|0\leq p\leq\mathcal{P}\\}$, and the immediate network device of compute node $v$ is indicated by $n_{v}$. Each path $p$ contains a set of links $\mathbb{L}_{p}\subset\mathbb{L}$, where $\mathbb{L}=\\{l:(n,n^{\prime})|n,n^{\prime}\in\mathbb{N}\\}$ is the set of all network links, and $\mathcal{L}$ is its size. Network devices and compute nodes are priority-aware, i.e., $\mathbb{K}=\\{k|0\leq k\leq\mathcal{K}\\}$ is regarded as the set of permissible priority levels (where lower levels indicate higher priorities), and the resources in both domains are virtually partitioned, isolated, and guaranteed for each priority level $k$. Note that higher priorities receive a larger share of available resources than lower priorities. To evaluate the performance of allocated resources, we will employ the $M/M/1$ queuing model for each priority level on network devices and compute nodes assuming that this theory’s stability requirements are met and that all queues are independent. The service rate allocated to priority level $k$ on network device $n$ is $\widehat{\mathcal{B}_{n,k}}$, and those packets leaving this queue will be forwarded through their corresponding link, let’s call it $l$, allocating $\widehat{\mathcal{B}_{l,k}}$ bandwidth. Note that the overall capacity of this network device and link is constrained by $\widehat{\mathcal{B}_{n}}$ and $\widehat{\mathcal{B}_{l}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>respectively. Similarly, the requests of priority level $k$ will be served on compute node $v$ leveraging a queue with dedicated service rate $\widehat{\mathcal{C}_{v,k}}$, and this node is equipped with computing resources limited to a predefined capacity threshold, dubbed $\widehat{\mathcal{C}_{v}}$. In addition, the energy consumption for transmitting bandwidth units over network device $n$ is $\widetilde{\mathcal{E}_{n}}$, and compute node $v$ consumes $\widetilde{\mathcal{E}_{v}}$ energy per capacity unit and $\overline{\mathcal{E}_{v}}$ energy when its state changes (from the idle mode to the operation mode or vice versa). ### II-B Requests This paper investigates the system for $\mathcal{T}$ time slots while a set of $\mathcal{R}_{t}$ requests, denoted by $\mathbb{R}_{t}=\\{r|0\leq r\leq\mathcal{R}_{t}\\}$, arrives at time slot $t\in\mathbb{T}=\\{t|1\leq t\leq\mathcal{T}\\}$. The set of all requests is $\mathbb{R}=\\{\mathbb{R}_{t}|1\leq t\leq\mathcal{T}\\}$, and $\mathcal{R}$ represents the number of all requests. Each request $r$ enters the system through an edge network device, denoted by $n_{r}$ and referred to as its Point of Arrival (PoA), and orders service $s_{r}$ from the set of obtainable service instances, that is $\mathbb{I}=\\{i|0\leq i\leq\mathcal{I}\\}$, where instance $i$ provides service $s_{i}$. In order to successfully fulfill each request, one instance of its target service must be replicated on one of the compute nodes in order to receive the request, process it, and return it to its entry point so that it can be delivered to the end user. $\widehat{\mathcal{C}_{i{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>represents the maximum capacity of instance $i$. To fulfill each user’s request, its QoS requirements must be met, including the minimum service capacity and network bandwidth, as well as the maximum tolerable E2E latency, denoted by $\widecheck{\mathcal{C}_{r{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>$\widecheck{\mathcal{B}_{r}}$, and $\widecheck{\mathcal{D}_{r}}$, respectively. Besides, the maximum permissible packet size for request $r$ is $\widehat{\mathcal{H}_{r}}$. If request $r$ is successfully completed, the system will achieve a profit, that is $\gamma_{r}$. Note that the arrival rate for each queue is determined by a Poisson process and is assumed to be the sum of $\widecheck{\mathcal{C}_{r{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>(for compute queues) and $\widecheck{\mathcal{B}_{r}}$ (for network queues) for all requests assigned to that queue, respectively. ## III Problem Definition This section discusses the joint problem of instance placement and assignment, request prioritization, and path selection for integrated compute-network infrastructures to maximize the overall profit of the system while minimizing its energy consumption. In this section, the constraints and objective function are formulated, followed by the problem statement as a Mixed-Integer Non-Linear Programming (MINLP) formulation and its complexity analysis. ### III-A Instance Orchestration Constraints Constraints C1-C6 assign requests to instances and place them on compute nodes while maintaining the capacity constraints of instances and compute nodes. Considering that $\ddot{\mathcal{I}}^{t}_{r,i}$ is a binary variable whose value is $1$ if request $r$ is assigned to instance $i$ at time slot $t$, C1 ensures that each request $r$ is assigned to no more than one instance of its service for each time slot $t$. C2 defines a new binary variable, $\dot{\mathcal{I}}^{t}_{i}$, which indicates that whether instance $i$ is activated at time slot $t$. If $\sum_{\mathbb{R}_{t}}\ddot{\mathcal{I}}^{t}_{r,i}$ is equal to or greater than $1$ (i.e., at least one request is assigned to instance $i$), $(\sum_{\mathbb{R}_{t}}\ddot{\mathcal{I}}^{t}_{r,i})/\mathcal{R}_{t}$ will be a small number (between $0$ and $1$) and $\sum_{\mathbb{R}_{t}}\ddot{\mathcal{I}}^{t}_{r,i}$ will be a large number, so $\dot{\mathcal{I}}^{t}_{i}$ will be set to $1$. Otherwise, both sides of the equation will equal $0$, causing $\dot{\mathcal{I}}^{t}_{i}$ to also equal $0$. C3 ensures that each activated instance is assigned to exactly one compute node, where $\ddot{\mathcal{G}}^{t}_{i,v}$ is a binary variable indicating the compute node of instance $i$ at time slot $t$. Similar to C2, C4 defines $\dot{\mathcal{G}}^{t}_{v}$ as a binary variable to determine whether compute node $v$ should be activated at time slot $t$. Then, it must be assured that assigned requests do not exceed the capacity limitations of instances and compute nodes (C5 and C6). $\displaystyle\sum\nolimits_{\mathbb{I}|s_{i}=s_{r}}\ddot{\mathcal{I}}^{t}_{r,i}\leq 1\quad\forall t,r\in\mathbb{T},\mathbb{R}_{t}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}ttttttttttttttttttttttttttttttttttt}$ (C1) $\displaystyle\frac{1}{\mathcal{R}_{t}}\cdot\sum\nolimits_{\mathbb{R}_{t}}\ddot{\mathcal{I}}^{t}_{r,i}\leq\dot{\mathcal{I}}^{t}_{i}\leq\sum\nolimits_{\mathbb{R}_{t}}\ddot{\mathcal{I}}^{t}_{r,i}\quad\forall t,i\in\mathbb{T},\mathbb{I}$ (C2) $\displaystyle\sum\nolimits_{\mathbb{V}}\ddot{\mathcal{G}}^{t}_{i,v}=\dot{\mathcal{I}}^{t}_{i}\quad\forall t,i\in\mathbb{T},\mathbb{I}$ (C3) $\displaystyle\frac{1}{\mathcal{I}}\cdot\sum\nolimits_{\mathbb{I}}\ddot{\mathcal{G}}^{t}_{i,v}\leq\dot{\mathcal{G}}^{t}_{v}\leq\sum\nolimits_{\mathbb{I}}\ddot{\mathcal{G}}^{t}_{i,v}\quad\forall t,v\in\mathbb{T},\mathbb{V}$ (C4) $\displaystyle\sum\nolimits_{\mathbb{R}_{t}}\widecheck{\mathcal{C}_{r}}\cdot\ddot{\mathcal{I}}^{t}_{r,i}\leq\widehat{\mathcal{C}_{i{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>t,i\in\mathbb{T},\mathbb{I}$ (C5) $\displaystyle\sum\nolimits_{\mathbb{I}}\widehat{\mathcal{C}_{i{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}.}}}\cdot\ddot{\mathcal{G}}^{t}_{i,v}\leq\widehat{\mathcal{C}_{v}}\quad\forall v,t\in\mathbb{V},\mathbb{T}$ (C6) ### III-B Path Selection Constraints Constraints C7-C9 ensure that an E2E path is selected for each request considering the capacity constraints of network links and the traffic pattern, where packets of each request enter the network through its PoA and, after visiting its assigned instance, are returned to the same PoA to be handed off to the corresponding end user. C7 determines the allocated path for each request $r$, ensuring that it originates and terminates at $n_{r}$ and traverses the network device directly connected to the compute node hosting the instance assigned to the request. In this constraint, $\ddot{f}^{t}_{r,p}$ is a binary variable that represents the assigned path of request $r$ at time slot $t$. Finally, C8 and C9 maintain the maximum capacity of network links and devices. $\displaystyle\sum\nolimits_{\mathbb{P}|n_{r}\&n_{v}\in p}\ddot{f}^{t}_{r,p}=\ddot{\mathcal{I}}^{t}_{r,i}\cdot\ddot{\mathcal{G}}^{t}_{i,v}\quad\forall t,r,i,v,\in\mathbb{T},\mathbb{R},\mathbb{I},\mathbb{V}$ (C7) $\displaystyle\sum\nolimits_{\mathbb{R}_{t},\mathbb{P}|l\in\mathbb{L}_{p}}\widecheck{\mathcal{B}_{r}}\cdot\ddot{f}^{t}_{r,p}\leq\widehat{\mathcal{B}_{l}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>t,l\in\mathbb{T},\mathbb{L}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}tttttttttttttttttttttttttt}$ (C8) $\displaystyle\sum\nolimits_{\mathbb{R}_{t},\mathbb{P}|n\in\mathbb{L}_{p}}\widecheck{\mathcal{B}_{r}}\cdot\ddot{f}^{t}_{r,p}\leq\widehat{\mathcal{B}_{n}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>t,n\in\mathbb{T},\mathbb{N}$ (C9) ### III-C Request Prioritization Constraints $\displaystyle\sum\nolimits_{\mathbb{K}}\ddot{\varrho}^{t}_{r,k}=\sum\nolimits_{\mathbb{I}}\ddot{\mathcal{I}}^{t}_{r,i}\quad\forall t,r\in\mathbb{T},\mathbb{R}_{t}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}ttttttttttttttttttttttttttttttt}$ (C10) $\displaystyle\sum\nolimits_{\mathbb{R}_{t},\mathbb{I}}\widecheck{\mathcal{C}_{r}}\cdot\ddot{\varrho}^{t}_{r,k}\cdot\ddot{\mathcal{I}}^{t}_{r,i}\cdot\ddot{\mathcal{G}}^{t}_{i,v}<\widehat{\mathcal{C}_{v,k}}\quad\forall t,k,v\in\mathbb{T},\mathbb{K},\mathbb{V}$ (C11) $\displaystyle\sum\nolimits_{\mathbb{R}_{t},\mathbb{P}|l\in\mathbb{L}_{p}}\widecheck{\mathcal{B}_{r}}\cdot\ddot{\varrho}^{t}_{r,k}\cdot\ddot{f}^{t}_{r,p}<\widehat{\mathcal{B}_{l,k}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>t,k,l\in\mathbb{T},\mathbb{K},\mathbb{L}$ (C12) $\displaystyle\sum\nolimits_{\mathbb{R}_{t},\mathbb{P}|n\in\mathbb{L}_{p}}\widecheck{\mathcal{B}_{r}}\cdot\ddot{\varrho}^{t}_{r,k}\cdot\ddot{f}^{t}_{r,p}<\widehat{\mathcal{B}_{n,k}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>t,k,n\in\mathbb{T},\mathbb{K},\mathbb{N}$ (C13) To maintain integrity, it’s crucial to prevent any overuse of resources allocated to each priority level. Given that $\ddot{\varrho}^{t}_{r,k}$ is the priority of request $r$ at time slot $t$, C10 promises that the request’s priority is determined if an instance is assigned to serve it. Then, C11 to C13 satisfy the capacity constraints of priority queues on compute nodes and network resources. ### III-D Latency Constraints Each packet has to wait for three sources of latency through its request’s assigned E2E path in the system: 1) the service latency experienced at the network devices included in the path, 2) the transmission latency over the network links of the path, and 3) the service latency at the assigned compute node. Since the average latency of a packet in a $M/M/1$ queue is equal to $1/(\mu-\lambda)$ when the arrival rate is $\lambda$ and the service rate is $\mu$, the average latency experienced by the packets of request $r$ at network device $n$ allocated to priority level $k$ during time slot $t$ can be calculated as C14. In this constraint, the numerator will be $0$ for network devices and priority levels that have not been allocated to request $r$, causing $\ddot{\mathcal{D}}^{t}_{r,n,k}$ to equal $0$. Otherwise, the numerator will be $1$, and the latency will be determined following the adopted queuing theorem with the arrival rate of the queue set to the overall bandwidth of requests assigned to priority level $k$ and traversing network device $n$. C15 follows the same logic to calculate the average latency of request $r$ allocated to priority level $k$ at time slot $t$ on compute node $v$. C16 calculates the transmission latency of request $r$ over link $l$ at time slot $t$, considering its priority level and maximum packet size, if the link is part of the path assigned to the request. Otherwise, the latency will be $0$. Finally, C17 determines the Age of Information (AoI), followed by C18, which ensures the maximum acceptable latency requirement of requests. $\displaystyle\ddot{\mathcal{D}}^{t}_{r,n,k}=\frac{\sum\nolimits_{\mathbb{P}|n\in\mathbb{L}_{p}}\ddot{\varrho}^{t}_{r,k}\cdot\ddot{f}^{t}_{r,p}}{\widehat{\mathcal{B}_{n,k}}-\sum\nolimits_{\mathbb{R}_{t},\mathbb{P}|n\in\mathbb{L}_{p}}\widecheck{\mathcal{B}_{r^{\prime}}}\cdot\ddot{\varrho}^{t}_{r^{\prime},k}\cdot\ddot{f}^{t}_{r^{\prime},p}}\quad\begin{aligned} &\forall t,r,k,n\in\\\ &\mathbb{T},\mathbb{R}_{t},\mathbb{K},\mathbb{N}\\\ \end{aligned}$ (C14) $\displaystyle\ddot{\mathcal{D}}^{t}_{r,v,k}=\frac{\sum\nolimits_{\mathbb{I}}\ddot{\varrho}^{t}_{r,k}\cdot\ddot{\mathcal{I}}^{t}_{r,i}\cdot\ddot{\mathcal{G}}^{t}_{i,v}}{\widehat{C_{v,k}}-\sum\nolimits_{\mathbb{R}_{t},\mathbb{I}}\widecheck{\mathcal{C}_{r^{\prime}}}\cdot\ddot{\mathcal{I}}^{t}_{r^{\prime},i}\cdot\ddot{\mathcal{G}}^{t}_{i,v}}\quad\begin{aligned} &\forall t,r,k,v\in\\\ &\mathbb{T},\mathbb{R}_{t},\mathbb{K},\mathbb{V}\\\ \end{aligned}$ (C15) $\displaystyle\ddot{\mathcal{D}}^{t}_{r,l,k}=\frac{\widehat{\mathcal{H}_{r}}}{\widehat{\mathcal{B}_{l,k}}}\cdot\sum\nolimits_{\mathbb{P}|l\in\mathbb{L}_{p}}\ddot{\varrho}^{t}_{r,k}\cdot\ddot{f}^{t}_{r,p}\quad\forall t,r,l,k\in\mathbb{T},\mathbb{R}_{t},\mathbb{L},\mathbb{K}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}tttt}$ (C16) $\displaystyle\ddot{\mathcal{D}}^{t}_{r}=\sum\nolimits_{\mathbb{N},\mathbb{K},\mathbb{V},\mathbb{L}}(\ddot{\mathcal{D}}^{t}_{r,n,k}+\ddot{\mathcal{D}}^{t}_{r,v,k}+\ddot{\mathcal{D}}^{t}_{r,l,k})\quad\forall t,r\in\mathbb{T},\mathbb{R}_{t}$ (C17) $\displaystyle\ddot{\mathcal{D}}^{t}_{r}\leq\widecheck{\mathcal{D}_{r}}\quad\forall t,r\in\mathbb{T},\mathbb{R}_{t}$ (C18) ### III-E Objective Function The objective is to maximize the overall profit while minimizing the energy consumption of resources, that is: $\displaystyle\sum\nolimits_{\mathbb{T},\mathbb{R}_{t}}(\gamma_{r}\cdot\sum\nolimits_{\mathbb{I}}\ddot{\mathcal{I}}^{t}_{r,i})-\alpha\cdot(\sum\nolimits_{\mathbb{N}}\ddot{\mathcal{E}}_{n}+\sum\nolimits_{\mathbb{V}}\ddot{\mathcal{E}}_{v}),{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}ttttttttttttt}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}ttt}$ (OF) where $\sum_{\mathbb{I}}\ddot{\mathcal{I}}^{t}_{r,i}$ is $1$ if request $r$ is supported at time slot $t$, $\alpha$ is a small positive number, and $\ddot{\mathcal{E}}_{v}$ and $\ddot{\mathcal{E}}_{n}$ represent, respectively, the total energy consumption of compute node $v$ and network device $n$. Note that $\alpha$ must be set such that the total profit exceeds the total amount of energy consumed. Otherwise, supporting requests would result in a negative objective function value, and the only optimal solution would be to support no requests, making OF equal to $0$. To determine $\ddot{\mathcal{E}}_{n}$, where the only source of energy consumption is transmitting requests’ data, the following equations are employed: $\displaystyle\ddot{\mathcal{E}}_{n}=\widetilde{\mathcal{E}_{n}}\cdot\sum\nolimits_{\mathbb{T},\mathbb{R}_{t},\mathbb{P}|n\in\mathbb{L}_{p}}\widecheck{\mathcal{B}_{r}}\cdot\ddot{f}^{t}_{r,p}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}ttttttttttttttttttttttttttttttttttt}$ (1) To calculate $\ddot{\mathcal{E}}_{v}$, it should be noted that the energy consumed on each compute node has two primary sources: 1) the energy consumed to service each unit of requests’ data, and 2) the energy consumed during booting up or shutting down the compute node. Consequently, $\ddot{\mathcal{E}}_{v}$ for each $v\in\mathbb{V}$ is: $\displaystyle\widetilde{\mathcal{E}_{v}}\cdot\sum\nolimits_{\mathbb{T},\mathbb{R}_{t},\mathbb{I}}\widecheck{\mathcal{C}_{r}}\cdot\ddot{\mathcal{I}}^{t}_{r,i}\cdot\ddot{\mathcal{G}}^{t}_{i,v}+\overline{\mathcal{E}_{v}}\cdot\sum\nolimits_{\mathbb{T}|0\leq t<\mathcal{T}}\dot{\mathcal{G}}^{t}_{v}\oplus\dot{\mathcal{G}}^{t+1}_{v}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}ttttttttt}$ (2) In this equation, $\ddot{\mathcal{I}}^{t}_{r,i}\cdot\ddot{\mathcal{G}}^{t}_{i,v}$ equals $1$ if compute node $v$ is selected as the host for request $r$; therefore, the first part of the equation calculates the total energy consumption of compute node $v$ to service its assigned requests. Assuming $\dot{\mathcal{G}}^{0}_{v}$ equals $0$, $\dot{\mathcal{G}}^{t}_{v}\oplus\dot{\mathcal{G}}^{t+1}_{v}$ is equal to $1$ in the second part if and only if $\dot{\mathcal{G}}^{t}_{v}$ and $\dot{\mathcal{G}}^{t+1}_{v}$ differ, representing a boot-up or shutdown for compute node $v$. Then, the second part demonstrates the total energy consumption of state transitions in compute note $v$ within $\mathbb{T}$. ### III-F Problem Considering the constraints and the objective function, the Problem of Integrated Resource Allocation (PIRA) is: $\footnotesize\text{PIRA: }\max\text{ OF}\text{ s.t. }\text{C1 - C18.}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}tttttttttttttttttttttttttttttttttttttttttt}$ (3) The optimal solution involves assigning requests with stringent latency requirements to high-priority queues and keeping those queues as empty as possible to minimize latency. Resource selection should aim to minimize activated resources and consider future requests to reduce energy consumption through fewer start-ups and shutdowns, improving energy efficiency. ### III-G Complexity Analysis The problem defined in (3) is an extended version of the Multi-Dimensional Knapsack (MDK) problem. Assume the problem is relaxed and reformulated specifically for time slot $t$ as the problem of maximizing profit and minimizing energy consumption while the only decision is to assign requests to instances concerning only their capacity constraints. Since the MDK problem is NP-hard and this relaxed version is an MDK problem with at least $\mathcal{R}_{t}$ items and $\mathcal{I}$ knapsacks, PIRA is at least as difficult as the MDK problem and is also NP-hard. ## IV ORIENT This section proposes an RL-based priORIty-aware Energy-efficieNt laTency- sensitive resource allocation approach (ORIENT) to find near-optimal solutions for PIRA. Subsequently, the learning mechanism is elaborated upon, followed by an explanation of the agent’s design, and concluding with a description of the algorithm. ### IV-A Learning Mechanism Given the continuous operation of the system defined in this paper and the recurring necessity for consistent resource allocation decisions at each time slot of PIRA, the adoption of RL presents itself as a viable means to enhance decision-making proficiency to solve it. Within the framework of RL, an agent undergoes a process of learning by means of trial and error at each step (here, time slot), with the primary aim of optimizing a specific decision- making problem. The system’s designer defines a reward function in alignment with the objectives of the problem. By learning and following the optimal strategy derived from this reward function, the agent aims to maximize cumulative discounted rewards, regardless of the initial state. Among various RL-based algorithms, Q-Learning stands out as widely acknowledged. In Q-Learning, every state-action pair is associated with a numeric value referred to as the Q-value, where the agent selects the action with the maximum Q-value at each step. In Deep Q-Learning (DQL), a Deep Neural Network (DNN) serves as the approximator for these Q-values. In this arrangement, the state and action are presented as inputs, and the DNN-based $Q$-function encompassing all feasible actions, denoted by $Q(s,.;\boldsymbol{\mathcal{W}})$, is generated as the output and systematically updated over time according to the following equation: $\footnotesize\boldsymbol{\mathcal{W}}^{t+1}=\boldsymbol{\mathcal{W}}^{t}+\sigma[Y^{t}-Q(\boldsymbol{S}^{t},a^{t};\boldsymbol{\mathcal{W}}^{t})]\nabla_{\boldsymbol{\mathcal{W}}^{t}}\cdot Q(\boldsymbol{S}^{t},a^{t};\boldsymbol{\mathcal{W}}^{t}){\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}tttttttt}$ (4) In this equation, $\boldsymbol{\mathcal{W}}$ is the set of DNN weights, $\sigma$ is a scalar step size, $\boldsymbol{S}^{t}$ and $a^{t}$ are the agent’s state and action at time slot $t$, and $Y^{t}$ (also known as the target) shows the maximum value expected to be achieved by following $a^{t}$ at $\boldsymbol{S}^{t}$. $Y^{t}$ is the only variable that must be estimated in this equation, and in Double DQL (DDQL), where the selection and evaluation processes are decoupled, it can be expressed as follows: $\footnotesize Y^{t}=r^{t+1}+\gamma\;\widehat{Q}(\boldsymbol{S}^{t+1},a^{\prime},\boldsymbol{\mathcal{W}}^{t-}),{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}tttttttttttttttttttttttttttttttttttt}$ (5) where $r^{t+1}$ is the earned reward at time slot $t+1$, $\gamma\in[0,1]$ is a discount factor that balances the importance of immediate and future rewards, $a^{\prime}=\text{argmax}_{a\in\boldsymbol{\mathcal{A}}}Q(\boldsymbol{S}^{t+1},a,\boldsymbol{\mathcal{W}}^{t})$, and $\boldsymbol{\mathcal{A}}$ is the set of actions. In this equation, $\boldsymbol{\mathcal{W}}$ represents the set of weights for the main $Q$ and is updated in each step, whereas $\boldsymbol{\mathcal{W}}^{-}$ is for the target $\widehat{Q}$ and is replaced with the weights of the main network every $t$ steps. In other words, $\widehat{Q}$ remains a periodic copy of $Q$. Furthermore, we augment DDQL by incorporating the dueling concept introduced by Wang et al. [15]. Unlike conventional DDQL, which directly approximates Q-values using DNNs, this method initially computes separate estimators for state values ($\psi$) and action advantages ($\varphi$). Q-values are then derived from these estimators, as illustrated below: $\footnotesize Q(\boldsymbol{S}^{t},a^{t},\boldsymbol{\mathcal{W}}^{t})=\psi(\boldsymbol{S}^{t},\boldsymbol{\mathcal{W}}^{t})\Bigg{(}\varphi(\boldsymbol{S}^{t},a^{t},\boldsymbol{\mathcal{W}}^{t})-\frac{\Phi}{\left|\boldsymbol{\mathcal{A}}\right|}\Bigg{)}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}ttttttttttttt}$ (6) where $\Phi=\sum_{\boldsymbol{\mathcal{A}}}\varphi(\boldsymbol{S}^{t},a^{\prime},\boldsymbol{\mathcal{W}}^{t})$. The primary benefit is the ability to generalize learning across actions without modifying the learning algorithm, which improves policy evaluation in the presence of numerous actions with similar state values. As a result of combining the Dueling technique and DDQL, we can expect that the resultant D3QL agent will outperform its predecessors. To bolster the effectiveness and resilience of D3QL, observed transitions are archived in a memory bank known as the experience memory. The learning process entails randomly selecting transitions from this repository [16]. ### IV-B Agent Customization The first step toward exploiting D3QL to solve PIRA is to define the agent’s action space, state space, and reward. #### Action Space We define the action space as set $\boldsymbol{\mathcal{A}}=\\{a:(i,v,p,k)|i,v,p,k\in\mathbb{I},\mathbb{V},\mathbb{P},\mathbb{K}\\}$. During each time slot and for every request, a specific action must be executed to finalize the resource allocation pertaining to that request. #### State Space For encoding the system’s state, an architecture involving aggregation GNN layers is employed, constructing an aggregation sequence across all compute nodes, iteratively facilitating information exchange with neighboring nodes. Therefore, at time slot $t$ when request $r$ is on the verge of receiving service, the system state is denoted as $\boldsymbol{S}^{t}(r)=\\{\boldsymbol{S}^{t}_{\mathbb{V}}(r),\boldsymbol{S}^{t}_{\mathbb{P}}(r)\\}$ and can be formally defined as: $\displaystyle\boldsymbol{S}^{t}_{\mathbb{V}}(r)=\Bigg{\\{}\Big{[}[\widehat{\mathcal{C}^{t}_{v,k}}-\widecheck{\mathcal{C}_{r{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}.}}}]_{\mathbb{K}},[\ddot{\mathcal{D}}^{t}_{r,v,k}-\widecheck{\mathcal{D}_{r{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}.}}}]_{\mathbb{K}},\widetilde{\mathcal{E}_{v}},\overline{\mathcal{E}_{v}}\cdot(1-\dot{\mathcal{G}}^{t}_{v})\Big{]}_{\mathbb{V}}\Bigg{\\}},$ (7) $\displaystyle\boldsymbol{S}^{t}_{\mathbb{P}}(r)=\Bigg{\\{}\Big{[}\big{[}\wedge_{\mathbb{L}_{p}}-\widecheck{\mathcal{B}_{r{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}.}}}\big{]}_{\mathbb{K}},[\ddot{\mathcal{D}}^{t}_{r}-\ddot{\mathcal{D}}^{t}_{r,v,k}-\widecheck{\mathcal{D}_{r{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}.}}}]_{\mathbb{K}},\widetilde{\mathcal{E}_{n}}\Big{]}_{\mathbb{P}}\Bigg{\\}},{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color@gray@fill{1}ttttttttt}$ (8) where $\wedge_{\mathbb{L}_{p}}=\min\nolimits_{\mathbb{L}|n,l\in\mathbb{L}_{p}}\\{\widehat{\mathcal{B}^{t}_{n,k}},\widehat{\mathcal{B}^{t}_{l,k}}\\}$. $\boldsymbol{S}^{t}_{\mathbb{V}}(r)$ and $\boldsymbol{S}^{t}_{\mathbb{P}}(r)$ represent the embeddings of compute nodes and network paths, respectively, which function as inputs for the GNN layers. These embeddings encompass the remaining resource capacity and the anticipated latency when request $r$ is allocated to them, as well as their associated energy consumption. #### Reward Since the agent is designated to maximize OF, the reward should be engineered to reinforce the support of high-profit requests while selecting resources with low energy consumption. This goal is satisfied in (9), that is: $\footnotesize r^{t+1}=\left\\{\begin{array}[]{ll}0&\mbox{otherwise}\\\ \dfrac{\mathcal{M}}{\text{OF}\big{(}\boldsymbol{S}^{t}(r),a^{t}\big{)}-\min_{\boldsymbol{\mathcal{A}}}\text{OF}\big{(}\boldsymbol{S}^{t}(r),a^{\prime}\big{)}}&r\text{ is met}{\color[rgb]{1,1,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,1,1}\pgfsys@color@gray@stroke{1}\pgfsys@color<EMAIL_ADDRESS>(9) where $\mathcal{M}=\max_{\boldsymbol{\mathcal{A}}}\text{OF}\big{(}\boldsymbol{S}^{t}(r),a^{\prime}\big{)}-\min_{\boldsymbol{\mathcal{A}}}\text{OF}\big{(}\boldsymbol{S}^{t}(r),a^{\prime}\big{)}$, $\max\text{/}\min_{\boldsymbol{\mathcal{A}}}\text{OF}\big{(}\boldsymbol{S}^{t}(r),a^{\prime}\big{)}$ is the maximum/minimum profit that can be achieved by allocating the available resources at time slot $t$ to request $r$ without considering any constraints or requirements, and $\text{OF}\big{(}\boldsymbol{S}^{t}(r),a^{t}\big{)}$ is the profit of the allocation provided by the agent. If the action fails to meet the requirements of the request, it results in a reward of $0$. Conversely, actions that yield greater profits correspond to higher rewards. TABLE I: Simulation Parameters. Parameter | Value ---|--- number of priority levels | $4$ resource capacity bounds | $\sim\mathcal{U}\\{250,300\\}$ mbps Instance capacity bound | $20$ mbps energy consumptions per capacity unit | $\sim\mathcal{U}\\{10,20\\}$ energy consumptions per state transition | $\sim\mathcal{U}\\{100,200\\}$ capacity requirement per request | $\sim\mathcal{U}\\{4,8\\}$ mbps bandwidth requirement per request | $\sim\mathcal{U}\\{2,10\\}$ mbps latency requirement per request | $\sim\mathcal{U}\\{1,3\\}$ ms packet size per request | $1$ profit per request ($\gamma_{r}$) | $\mathcal{U}\\{5,15\\}$ Input: $\mathcal{T}$, $\epsilon^{\prime}$, and $\widetilde{\epsilon}$ 1 $\boldsymbol{\Omega}\leftarrow\emptyset$, $\boldsymbol{\mathcal{W}}\leftarrow\mathbf{0}$, $\boldsymbol{\mathcal{W}^{-}}\leftarrow\mathbf{0}$, $\epsilon\leftarrow 1$, $memory\leftarrow\\{\\}$ 2 for _each $t$ in $[0:\mathcal{T}]$_ do 3 if _new request $r$ is arrived_ then 4 calculate $\boldsymbol{S}^{t}(r)=\\{\boldsymbol{S}^{t}_{\mathbb{V}}(r),\boldsymbol{S}^{t}_{\mathbb{P}}(r)\\}$ 5 $\zeta\leftarrow$ generate a random number from $[0:1]$ 6 if _$\zeta >\epsilon$_ then 7 $a^{t}=(i,v,p,k)\leftarrow$ argmax${}_{\boldsymbol{\mathcal{A}}}Q(\boldsymbol{S}^{t},a^{\prime},\boldsymbol{\mathcal{W}}^{t})$ 8 9 else 10 select a random $a^{t}=(i,v,p,k)$ from $\boldsymbol{\mathcal{A}}$ 11 calculate $r^{t+1}$ 12 if _$r^{t+1} >0$_ then 13 Establish request $r$ connection based on $a^{t}$ 14 15 $memory\leftarrow memory\cup\\{(\boldsymbol{S}^{t},a^{t},r^{t+1})\\}$ 16 choose a batch of samples from $memory$ 17 train the agent 18 if _$\epsilon >\widetilde{\epsilon}$_ then 19 $\epsilon\leftarrow\epsilon-\epsilon^{\prime}$ 20 $\boldsymbol{\Omega}\leftarrow\boldsymbol{\Omega}\cup\\{(t,r,a^{t})\\}$ return $\boldsymbol{\Omega}$ Algorithm 1 ORIENT ### IV-C ORIENT’s Algorithm ORIENT is detailed in Algorithm 1. In this algorithm, $\epsilon^{\prime}$ and $\widetilde{\epsilon}$ are small positive integers that control the $\epsilon$-greedy mechanism. During each time slot $t$, the agent receives notifications of new request arrivals ($r$), and it computes the state based on request $r$ requirements and the current system state. The action is then chosen using an $\epsilon$-greedy policy, which follows the evaluation function of the corresponding agent with probability $(1-\epsilon)$ and selects a random action with probability $\epsilon$. Subsequently, the reward is calculated, and if it exceeds $0$, it indicates that $a^{t}$ is feasible and meets request $r$’s QoS requirements, enabling its connection based on $a^{t}$ allocations. Finally, the experience memory is updated, samples are drawn from the memory bank, and the agent undergoes training. during the training process, $\epsilon$ decreases from $1$ to $\widetilde{\epsilon}$. The algorithm yields $\boldsymbol{\Omega}$ as the history of allocations. ## V Performance Evaluation Figure 2: The mean energy consumption of supported requests and the total profit vs. the system size (A & B) and the number of all requests (C & D). In this section, we present numerical results based on the system model parameters listed in Table I. Other parameters can be chosen arbitrarily so long as the logic outlined throughout the paper remains valid. To evaluate the efficiency of ORIENT, we conduct a comparative analysis with OPT, D3QL, and RND. OPT represents the optimal solution for PIRA, obtained through the use of CPLEX 12.10. D3QL, on the other hand, bears similarities to the method outlined in Algorithm 1, but employs exclusively simple linear layers in its DNN, without utilizing any GNNs. This approach forms the foundation for several related studies, including A-DDPG [13] and MDRL-SaDS [7], both of which are RL methods designed to enhance the utility of individual requests by considering factors such as resource cost, required bandwidth, and E2E path latency. Lastly, RND represents a random allocation strategy, where resources are allocated to active requests without considering any constraints. The results are depicted in Fig. 2, where subfigures A and B represent the average energy consumption per request and total profit across various system sizes, with a constant of $300$ active requests. Here, incrementing the system size entails the creation of a new system graph, incorporating $\mathcal{N}+1$ network devices and $\mathcal{V}+1$ compute nodes. Particularly, from $10$ to $13$, resources with a significant energy consumption are introduced into the graph. Between $14$ and $17$, resources with a moderate energy consumption are added to the graph, and the remaining resources included from $18$ to $21$ are characterized by low energy consumption. Subfigures C and D display similar quality metrics, but for different numbers of requests, while keeping the system size fixed at $12$ with equal numbers of resources ($4$) from each level of energy consumption. Within the subfigures, it is evident that OPT serves as an upper performance bound, while RND serves as the lower bound. Furthermore, when all resources exhibit high energy consumption rates or are fully occupied (with $\mathcal{N}\leq 13$ in A and $\mathcal{R}\geq 200$ in C), RND shows a similar energy consumption pattern to D3QL-based techniques, but its support rate is limited due to the absence of intelligence and feasibility checks. In contrast, ORIENT excels in both scenarios. As demonstrated in A and B, it achieves near-optimal results by prioritizing high-capacity resources with minimal energy consumption, especially when multiple choices are available for each request ($\mathcal{N}\geq 13$). Similarly, regardless of whether all requests can be supported (C and D, with $\mathcal{R}\leq 200$), near-optimal solutions are consistently attained. However, D3QL exhibits less efficiency and stability compared to ORIENT, mainly due to its inferior state decoding capability. ## VI Conclusion In this paper, we examined the joint problem of service instance placement and assignment, path selection, and request prioritization, dubbed PIRA, with the objective of maximizing the overall profit of the system (as a function of the number of supported concurrent requests) while minimizing the overall energy consumption over a continuous period of time, taking into account E2E latency and resource capacity constraints. This problem was formulated as a MINLP problem, its complexity was analyzed, and it was demonstrated that it is an NP-hard problem. Subsequently, a technique named ORIENT was introduced to address the problem in a near-optimal manner, utilizing a GNN-empowered D3QL strategy. The effectiveness of the suggested technique was validated through numerical results. As potential future work, our intention is to tackle the problem in the context of dynamic environments characterized by temporal/spatial fluctuations in requests and resources. ## Acknowledgment It is partially supported by the European Union’s Horizon 2020 Research and Innovation Program through the aerOS project under Grant No. 101069732; the Business Finland 6Bridge 6Core project under Grant No. 8410/31/2022; the European Union’s HE research and innovation program HORIZON-JUSNS-2022 under the 6GSandbox project (Grant No. 101096328); and the Research Council of Finland 6G Flagship Programme under Grant No. 346208. This research was also conducted at ICTFICIAL Oy. The paper reflects only the authors’ views, and the commission bears no responsibility for any utilization of the information contained herein. ## References * [1] M. Latva-aho, “Key Drivers and Research Challenges for 6G Ubiquitous Wireless Intelligence,” 2019, publisher: University of Oulu. * [2] H. Yu _et al._ , “Toward 6G-Based Metaverse: Supporting Highly-Dynamic Deterministic Multi-User Extended Reality Services,” _IEEE Network_ , vol. 37, no. 4, pp. 30–38, 2023. * [3] H. Mazandarani _et al._ , “A Semantic-Aware Multiple Access Scheme for Distributed, Dynamic 6G-Based Applications,” _arXiv preprint arXiv:2401.06308_ , 2024. * [4] J. Prados-Garzon _et al._ , “Deterministic 6GB-Assisted Quantum Networks with Slicing Support: A New 6GB Use Case,” _IEEE Network_ , 2023. * [5] Z. Yang _et al._ , “Increasing the Energy Efficiency of a Data Center Based on Machine Learning,” _Journal of Industrial Ecology_ , vol. 26, no. 1, pp. 323–335, 2022. * [6] “Cisco Annual Internet Report (2018–2023) White Paper.” * [7] H. Xuan _et al._ , “Multi-Agent Deep Reinforcement Learning Algorithm with Self-adaption Division Strategy for VNF-SC Deployment in SDN/NFV-Enabled Networks,” _Applied Soft Computing_ , vol. 138, p. 110189, May 2023. * [8] R. Solozabal _et al._ , “Virtual Network Function Placement Optimization With Deep Reinforcement Learning,” _IEEE Journal on Selected Areas in Communications_ , vol. 38, no. 2, pp. 292–303, Feb. 2020\. * [9] C. Pham _et al._ , “Traffic-Aware and Energy-Efficient VNF Placement for Service Chaining: Joint Sampling and Matching Approach,” _IEEE Trans. on Services Computing_ , vol. 13, no. 1, pp. 172–185, Jan. 2020. * [10] G. L. Santos _et al._ , “Availability-Aware and Energy-Aware Dynamic SFC Placement using Reinforcement Learning,” _The Journal of Supercomputing_ , vol. 77, no. 11, Nov. 2021. * [11] Z. Sasan _et al._ , “Joint Network Slicing, Routing, and In-Network Computing for Energy-Efficient 6G,” _arXiv preprint arXiv:2401.06306_ , 2024\. * [12] M. Farhoudi _et al._ , “Qos-Aware Service Prediction and Orchestration in Cloud-Network Integrated Beyond 5G,” _arXiv preprint arXiv:2309.10185_ , 2023. * [13] N. He _et al._ , “Leveraging Deep Reinforcement Learning With Attention Mechanism for Virtual Network Function Placement and Routing,” _IEEE Trans. on Parallel and Distributed Systems_ , vol. 34, no. 4, Apr. 2023. * [14] M. Shokrnezhad _et al._ , “Double Deep Q-Learning-based Path Selection and Service Placement for Latency-Sensitive Beyond 5G Applications,” _IEEE Transactions on Mobile Computing_ , pp. 1–14, 2023. * [15] Z. Wang _et al._ , “Dueling Network Architectures for Deep Reinforcement Learning,” in _Proceedings of The 33rd International Conference on Machine Learning_ , vol. 48, Jun. 2016, pp. 1995–2003. * [16] V. Mnih _et al._ , “Human-level Control through Deep Reinforcement Learning,” _Nature_ , vol. 518, no. 7540, pp. 529–533, Feb. 2015.
# Means refinements via convexity M. Sababheh Department of Basic Sciences, Princess Sumaya University For Technology, Al Jubaiha, Amman 11941, Jordan<EMAIL_ADDRESS><EMAIL_ADDRESS> ###### Abstract. The main goal of this article is to find the exact difference between a convex function and its secant, as a limit of positive quantities. This idea will be expressed as a convex inequality that leads to refinements and reversals of well established inequalities treating different means. The significance of these inequalities is to write one inequality that brings together and refine almost all known inequalities treating the arithmetic, geometric, harmonic and Heinz means, for numbers and operators. ###### Key words and phrases: convex functions, means inequalities, norm inequalities. ###### 2010 Mathematics Subject Classification: 15A39, 15B48, 26D15, 26B25, 47A30, 47A63. ## 1\. introduction Convex functions and their inequalities have played a major role in the study of various topics in Mathematics; including applied Mathematics, Mathematical Analysis and Mathematical Physics. Means and their comparison is indeed an important application of convexity. Recall that a function $f:\mathbb{I}\to\mathbb{R}$, defined on a real interval $\mathbb{I}$, is said to be convex if $f(\alpha x_{1}+\beta x_{2})\leq\alpha f(x_{1})+\beta f(x_{2})$, when $x_{1},x_{2}\in\mathbb{I}$ and $\alpha,\beta\geq 0$ satisfying $\alpha+\beta=1.$ On the other hand, $f:\mathbb{I}\to\mathbb{R}^{+}$ is said to be log-convex if $g(x)=\log f(x)$ is convex, or equivalently if $f(\alpha x_{1}+\beta x_{2})\leq f^{\alpha}(x_{1})f^{\beta}(x_{2})$ for the above parameters. Speaking of means, the comparison between the weighted arithmetic, geometric and harmonic means is an immediate consequence of convexity or log-convexity of the functions $x\nabla_{t}y=(1-t)x+ty,x\\#_{t}y=x^{1-t}y^{t}$ and $x!_{t}y=((1-t)x^{-1}+ty^{-1})^{-1},x,y>0,$ defined for $0\leq t\leq 1$. Adopting these notations, we drop $t$ when $t=\frac{1}{2}.$ Convexity of the function $f(t)=x\\#_{t}y$ implies the well known Young’s inequality $x\\#_{t}\leq x\nabla_{t}y.$ On the other hand, convexity of the function $g(t)=x!_{t}y$ implies the arithmetic-harmonic mean inequality $x!_{t}y\leq x\nabla_{t}y$, while log-convexity of $g$ implies the geometric- harmonic mean inequality $x!_{t}y\leq x\\#_{t}y.$ These inequalities, though very simple, have some significant applications. For example, the above Young’s inequality implies the celebrated Holder’s inequality $\|fg\|_{1}\leq\|f\|_{p}\|g\|_{q}$ for $f\in L^{p}(X)$ and $g\in L^{q}(X)$, for the conjugate exponents $p,q$, where $X$ is some measure space. Among the most interesting applications of the above mean inequalities is the possible comparison between operators acting on a finite dimensional Hilbert space $H$. In the sequel, $\mathbb{M}_{n}$ will denote the space of operators acting on an $n-$deimentional Hilbert space $H$, $\mathbb{M}_{n}^{+}$ will denotes the cone of semi positive operators in $\mathbb{M}_{n}$ while $\mathbb{M}_{n}^{++}$ will denotes the cone of strictly positive operators in $\mathbb{M}_{n}.$ Then the above numerical inequalities have their operator versions such as $A\\#_{t}B\leq A\\#_{t}B$, where $A,B\in\mathbb{M}_{n}^{++},A\nabla_{t}B=(1-t)A+tB$ and $A\\#_{t}B=A^{\frac{1}{2}}\left(A^{-\frac{1}{2}}BA^{-\frac{1}{2}}\right)^{t}A^{\frac{1}{2}}.$ In this context, we say that $A\leq B$ for two self-adjoint operators $A$ and $B$ if $B-A\in\mathbb{M}_{n}^{+}.$ Obtaining the operator versions from the corresponding numerical versions can be done in different approaches, among which is the application of the following lemma [2]. ###### Lemma 1.1. Let $X\in\mathcal{M}_{n}$ be self-adjoint and let $f$ and $g$ be continuous real valued functions such that $f(t)\geq g(t)$ for all $t\in{\text{Sp}}(X),$ the spectrum of $X$. Then $f(X)\geq g(X).$ Recent studies of the topic have investigated possible refinements of the above inequalities, where adding a positive term to the left side becomes possible. This idea has been treated in [3, 5, 6, 7, 9, sabjmaa, 10, 11, 12], where not only refinements have been investigated, but reversed versions and much more have been discussed. Keeping our paper concise, we will not go through the exact results done in the above references now, however we will comment later how the results in this paper generalize almost all results in these references, regarding the refinements and the reverses of the above mean inequalities. The main goal of this article is to avoid dealing with the specific means, and to treat a general convexity argument that leads to these refinements. In particular, we prove that for certain positive quantities $A_{j}(\nu)\Delta_{j}f(\nu;a,b),$ we have $\displaystyle f\left((1-\nu)a+\nu b\right)+\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;a,b)\leq(1-\nu)f(a)+\nu f(b),N\in\mathbb{N},$ for the convex function $f:[a,b]\to\mathbb{R}$. This provides $N$ refining terms of the inequality $f\left((1-\nu)a+\nu b\right)\leq(1-\nu)f(a)+\nu f(b)$, which follows from convexity of $f$. Furthermore, we prove a reversed version and we prove that as $N\to\infty$ the above inequality becomes an equality. As a natural consequence, we obtain some refinements and reverses for log-convex functions. As we will see, the above inequality and its consequences happen to be generalizations that imply almost all inequalities in the references [3, 5, 6, 9, sabjmaa, 10, 11, 12]. This is our main motivation behind this work; to find a formula that implies and generalizes all other formulae and hence, to enhance our understanding of these inequalities. We remark that the proof of the first main result in this work is inspired by our recent work in [sabjmaa]. ## 2\. main results For the rest of the paper, the following notations will be adopted. For $0\leq\nu\leq 1$ and $j\in\mathbb{N}$, let $\left\\{\begin{array}[]{cc}k_{j}(\nu)=[2^{j-1}\nu],r_{j}(\nu)=[2^{j}\nu]\;{\text{and}}\\\ A_{j}(\nu)=(-1)^{r_{j}(\nu)}2^{j-1}\nu+(-1)^{r_{j}(\nu)+1}\left[\frac{r_{j}(\nu)+1}{2}\right]\end{array}\right..$ (2.1) Moreover, if $f:[a,b]\to\mathbb{R}$ is any function, define $\displaystyle\Delta_{j}f(\nu;a,b)$ $\displaystyle=$ $\displaystyle f\left(\left(1-\frac{k_{j}(\nu)}{2^{j-1}}\right)a+\frac{k_{j}(\nu)}{2^{j-1}}b\right)+f\left(\left(1-\frac{k_{j}(\nu)+1}{2^{j-1}}\right)a+\frac{k_{j}(\nu)+1}{2^{j-1}}b\right)$ (2.2) $\displaystyle-$ $\displaystyle 2f\left(\left(1-\frac{2k_{j}(\nu)+1}{2^{j}}\right)a+\frac{2k_{j}(\nu)+1}{2^{j}}b\right),0\leq\nu\leq 1.$ ### 2.1. Convex functions We discuss first the inequalities that govern convex functions, then we apply these inequalities to log-convex functions. ###### Lemma 2.1. If $f:[a,b]\to\mathbb{R}$ is convex, then $\Delta_{j}f(\nu;a,b)\geq 0$ for $j\in\mathbb{N}$ and $0\leq\nu\leq 1.$ ###### Proof. Letting $x_{j}(\nu)=\left(1-\frac{k_{j}(\nu)}{2^{j-1}}\right)a+\frac{k_{j}(\nu)}{2^{j-1}}b,y_{j}(\nu)=\left(1-\frac{k_{j}(\nu)+1}{2^{j-1}}\right)a+\frac{k_{j}(\nu)+1}{2^{j-1}}b$ and $z_{j}(\nu)=\left(1-\frac{2k_{j}(\nu)+1}{2^{j}}\right)a+\frac{2k_{j}(\nu)+1}{2^{j}}b,$ it is easy that $z_{j}(\nu)=\frac{x_{j}(\nu)+y_{j}(\nu)}{2}.$ The $\Delta_{j}f(\nu;a,b)=f(x_{j}(\nu))+f(y_{j}(\nu))-2f(z_{j}(\nu))\geq 0,$ by convexity of $f$. ∎ ###### Remark 2.2. When $f:[a,b]\to\mathbb{R}$, we adopt the convention that $f(x)=0$ for $x\not\in[a,b].$ This convention will be needed, for example, in the next lemma, when $N=1$ and $\nu=1.$ ###### Lemma 2.3. Let $f:[0,1]\to\mathbb{R}$ be a function and let $N\in\mathbb{N}$. Then $\displaystyle(1-\nu)f(0)$ $\displaystyle+$ $\displaystyle\nu f(1)-\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;0,1)$ $\displaystyle=$ $\displaystyle\left([2^{N}\nu]+1-2^{N}\nu\right)f\left(\frac{[2^{N}\nu]}{2^{N}}\right)+\left(2^{N}\nu-[2^{N}\nu]\right)f\left(\frac{[2^{N}\nu]+1}{2^{N}}\right).$ ###### Proof. We proceed by induction on $N$. When $N=1$ and $0\leq\nu<\frac{1}{2},$ $r_{1}(\nu)=0$ and $k_{1}(\nu)=0$. Hence $A_{1}(\nu)=\nu$ and $\Delta_{1}f(\nu;a,b)=f(a)+f(b)-2f\left(\frac{a+b}{2}\right).$ Then direct computations show the result. Now if $\frac{1}{2}\leq\nu<1$, then $r_{1}(\nu)=1$ and $k_{1}(\nu)=0$, hence $A_{1}(\nu)=1-\nu$ and $\Delta_{1}f(\nu;a,b)=f(a)+f(b)-2f\left(\frac{a+b}{2}\right).$ Again, direct computations show the result. When $\nu=1,$ the result follows immediately. Now assume that (LABEL:exact_difference) is true for some $N\in\mathbb{N}$. We assert its truth for $N+1.$ Notice that, using the inductive step, $\displaystyle(1-\nu)f(0)$ $\displaystyle+$ $\displaystyle\nu f(1)-\sum_{j=1}^{N+1}A_{j}(\nu)\Delta_{j}f(\nu;0,1)$ (2.4) $\displaystyle=$ $\displaystyle(1-\nu)f(0)+\nu f(1)-\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;0,1)-A_{N+1}(\nu)\Delta_{N+1}f(\nu;0,1)$ $\displaystyle=$ $\displaystyle\left([2^{N}\nu]+1-2^{N}\nu\right)f\left(\frac{[2^{N}\nu]}{2^{N}}\right)+\left(2^{N}\nu-[2^{N}\nu]\right)f\left(\frac{[2^{N}\nu]+1}{2^{N}}\right)$ $\displaystyle-$ $\displaystyle\left((-1)^{[2^{N+1}\nu]}2^{N}\nu+(-1)^{[2^{N+1}\nu]+1}\left[\frac{[2^{N+1}\nu]+1}{2}\right]\right)\times$ $\displaystyle\times$ $\displaystyle\left(f\left(\frac{[2^{N}\nu]}{2^{N}}\right)+f\left(\frac{[2^{N}\nu]+1}{2^{N}}\right)-2f\left(\frac{2[2^{N}\nu]+1}{2^{N+1}}\right)\right).$ Now we treat two cases. Case I If $[2^{N+1}\nu]$ is odd, then we easily see that $[2^{N}\nu]=\frac{[2^{N+1}\nu]-1}{2}.$ Therefore, $f\left(\frac{[2^{N}\nu]+1}{2^{N}}\right)=f\left(\frac{[2^{N+1}\nu]+1}{2^{N+1}}\right)\;{\text{and}}\;f\left(\frac{2[2^{N}\nu]+1}{2^{N+1}}\right)=f\left(\frac{[2^{N+1}\nu]}{2^{N+1}}\right).$ Substituting these values in (2.4) and simplifying imply $\displaystyle(1-\nu)f(0)+\nu f(1)-\sum_{j=1}^{N+1}A_{j}(\nu)\Delta_{j}f(\nu;0,1)$ $\displaystyle=$ $\displaystyle\left(2^{N+1}\nu-[2^{N+1}\nu]\right)f\left(\frac{[2^{N+1}\nu]+1}{2^{N+1}}\right)+\left([2^{N+1}\nu]+1-2^{N+1}\nu\right)f\left(\frac{[2^{N+1}\nu]}{2^{N+1}}\right),$ which completes the proof, when $[2^{N+1}\nu]$ is odd. Case II If $[2^{N+1}\nu]$ is even, then $2[2^{N}\nu]=[2^{N+1}\nu]$ and $f\left(\frac{[2^{N}\nu]}{2^{N}}\right)=f\left(\frac{[2^{N+1}\nu]}{2^{N+1}}\right)\;{\text{and}}\;f\left(\frac{2[2^{N}\nu]+1}{2^{N+1}}\right)=f\left(\frac{[2^{N+1}\nu]+1}{2^{N+1}}\right).$ Substituting these values in (2.4) and simplifying imply $\displaystyle(1-\nu)f(0)+\nu f(1)-\sum_{j=1}^{N+1}A_{j}(\nu)\Delta_{j}f(\nu;0,1)$ $\displaystyle=$ $\displaystyle\left(2^{N+1}\nu-[2^{N+1}\nu]\right)f\left(\frac{[2^{N+1}\nu]+1}{2^{N+1}}\right)+\left([2^{N+1}\nu]+1-2^{N+1}\nu\right)f\left(\frac{[2^{N+1}\nu]}{2^{N+1}}\right).$ This completes the proof. ∎ ###### Corollary 2.4. Let $f:[0,1]\to\mathbb{R}$ be convex and let $N\in\mathbb{N}$. Then $f(\nu)+\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;0,1)\leq(1-\nu)f(0)+\nu f(1).$ (2.5) ###### Proof. From Lemma 2.3 and convexity of $f$, we have $\displaystyle(1-\nu)f(0)$ $\displaystyle+$ $\displaystyle\nu f(1)-\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;0,1)$ $\displaystyle=$ $\displaystyle\left([2^{N}\nu]+1-2^{N}\nu\right)f\left(\frac{[2^{N}\nu]}{2^{N}}\right)+\left(2^{N}\nu-[2^{N}\nu]\right)f\left(\frac{[2^{N}\nu]+1}{2^{N}}\right)$ $\displaystyle\geq$ $\displaystyle f\left(\left([2^{N}\nu]+1-2^{N}\nu\right)\frac{[2^{N}\nu]}{2^{N}}+\left(2^{N}\nu-[2^{N}\nu]\right)\frac{[2^{N}\nu]+1}{2^{N}}\right)$ $\displaystyle=$ $\displaystyle f(\nu).$ This completes the proof. ∎ Now our first main result in its general form can be stated as follows. ###### Theorem 2.5. Let $f:[a,b]\to\mathbb{R}$ be convex. Then for each $N\in\mathbb{N}$ and $0\leq\nu\leq 1,$ we have $\displaystyle f\left((1-\nu)a+\nu b\right)+\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;a,b)\leq(1-\nu)f(a)+\nu f(b).$ (2.6) ###### Proof. For the given $f$, define $g:[0,1]\to\mathbb{R}$ by $g(x)=f((1-x)a+xb).$ Then $g$ is convex on $[0,1].$ Applying Corollary 2.4 on the function $g$ implies the result. ∎ ###### Remark 2.6. We remark that a negative version of the above theorem has been recently shown in [8]. Namely, it was proved $\displaystyle(1+\nu)f(a)-\nu f(b)$ $\displaystyle+\sum_{j=1}^{N}2^{j}\nu\left[\frac{f(a)+f\left(\frac{(2^{j-1}-1)a+b}{2^{j-1}}\right)}{2}-f\left(\frac{(2^{j}-1)a+b}{2^{j}}\right)\right]$ $\displaystyle\leq f\left((1+\nu)a-\nu b\right),\nu\geq 0,a<b,$ (2.7) for the convex function $f:\mathbb{R}\to\mathbb{R}.$ However, the method of proof is considerably easier than the above proofs and the applications are different. Our next step is to prove a reversed version of (2.6). ###### Theorem 2.7. Let $f:[a,b]\to\mathbb{R}$ be convex and let $N\in\mathbb{N}$. Then for $0\leq\nu\leq\frac{1}{2},$ we have $\displaystyle f\left((1-\nu)a+\nu b\right)$ $\displaystyle+$ $\displaystyle(1-A_{1}(\nu))\Delta_{1}f(\nu;a,b)$ $\displaystyle\geq$ $\displaystyle(1-\nu)f(a)+\nu f(b)+\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f\left(1-2\nu;\frac{a+b}{2},b\right).$ On the other hand, if $\frac{1}{2}\leq\nu\leq 1,$ we have $\displaystyle f\left((1-\nu)a+\nu b\right)$ $\displaystyle+$ $\displaystyle(1-A_{1}(\nu))\Delta_{1}f(\nu;a,b)$ $\displaystyle\geq$ $\displaystyle(1-\nu)f(a)+\nu f(b)+\sum_{j=1}^{N}A_{j}(2-2\nu)\Delta_{j}f\left(2-2\nu;a,\frac{a+b}{2}\right).$ ###### Proof. For $0\leq\nu\leq\frac{1}{2},$ we have $\displaystyle f\left((1-\nu)a+\nu b\right)+(1-A_{1}(\nu))\Delta_{1}f(\nu;a,b)-\left((1-\nu)f(a)+\nu f(b)\right)$ $\displaystyle=$ $\displaystyle 2\nu f\left(\frac{a+b}{2}\right)+(1-2\nu)f(b)+f\left((1-\nu)a+\nu b\right)-2f\left(\frac{a+b}{2}\right)$ $\displaystyle\geq$ $\displaystyle\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f\left(1-2\nu;\frac{a+b}{2},b\right)+f\left(2\nu\frac{a+b}{2}+(1-2\nu)b\right)$ $\displaystyle+f\left((1-\nu)a+\nu b\right)-2f\left(\frac{a+b}{2}\right)$ $\displaystyle=$ $\displaystyle\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f\left(1-2\nu;\frac{a+b}{2},b\right)+f\left(\nu a+(1-\nu)b\right)$ $\displaystyle+f\left((1-\nu)a+\nu b\right)-2f\left(\frac{a+b}{2}\right)$ $\displaystyle\geq$ $\displaystyle\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f\left(1-2\nu;\frac{a+b}{2},b\right),$ where the last line follows from convexity of $f$, where one has $\displaystyle f\left(\nu a+(1-\nu)b\right)+f\left((1-\nu)a+\nu b\right)$ $\displaystyle\geq 2f\left(\frac{\nu a+(1-\nu)b+(1-\nu)a+\nu b}{2}\right)=2f\left(\frac{a+b}{2}\right).$ This completes the proof for $0\leq\nu\leq\frac{1}{2}.$ Similar computations imply the desired inequality for $\frac{1}{2}\leq\nu\leq 1.$ ∎ In fact, the above reversed version turns out to be equivalent to convexity. ###### Proposition 2.8. Let $f:\mathbb{I}\to\mathbb{R}$ be a function defined on the interval $\mathbb{I}$. Assume that for all $a<b$ in $\mathbb{I}$ and all $0\leq\nu\leq 1$, we have $f\left((1-\nu)a+\nu b\right)+(1-A_{1}(\nu))\Delta_{1}f(\nu;a,b)\geq(1-\nu)f(a)+\nu f(b),$ (2.8) then $f$ is convex on $\mathbb{I}.$ ###### Proof. Observe that when $0\leq\nu\leq\frac{1}{2},$ (2.8) is equivalent to $f\left((1-\nu)a+\nu b\right)+(1-\nu)\left(f(a)+f(b)-2f\left(\frac{a+b}{2}\right)\right)\geq(1-\nu)f(a)+\nu f(b),$ or $f\left(\frac{a+b}{2}\right)\leq\frac{1}{2-2\nu}f\left((1-\nu)a+\nu b\right)+\frac{1-2\nu}{2-2\nu}f(b).$ (2.9) On the other hand, if $\frac{1}{2}\leq\nu\leq 1,$ (2.8) is equivalent to $f\left(\frac{a+b}{2}\right)\leq\frac{2\nu-1}{2\nu}f(a)+\frac{1}{2\nu}f\left((1-\nu)a+\nu b\right).$ (2.10) Let $x_{1}<x_{2}\in\mathbb{I}$ and let $0<\lambda<1.$ We assert that $f((1-\lambda)x_{1}+\lambda x_{2})\leq(1-\lambda)f(x_{1})+\lambda f(x_{2}).$ If $0<\lambda\leq\frac{1}{2},$ let $\nu=\frac{1-2\lambda}{2(1-\lambda)},a=(2-2\lambda)x_{1}+(2\lambda-1)x_{2}\;{\text{and}}\;b=x_{2}.$ Then one can easily check that when $0<\lambda\leq\frac{1}{2},$ we have $0<\nu\leq\frac{1}{2}$ and $a<b$. With these choices, we have $\frac{a+b}{2}=(1-\lambda)x_{1}+\lambda x_{2}\;{\text{and}}\;(1-\nu)a+\nu b=x_{1}.$ Substituting these quantities in (2.9) implies $f((1-\lambda)x_{1}+\lambda x_{2})\leq(1-\lambda)f(x_{1})+\lambda f(x_{2}).$ This proves the desired inequality for $0<\lambda\leq\frac{1}{2}.$ Now if $\frac{1}{2}\leq\lambda<1,$ let $\nu=\frac{1}{2\lambda},a=x_{1}\;{\text{and}}\;b=(1-2\lambda)x_{1}+2\lambda x_{2}.$ With these choices, we have $\frac{1}{2}<\nu\leq 1$ and $a<b$. Now substituting these quantities in (2.10) implies the desired inequality for $\frac{1}{2}\leq\lambda<1.$ This completes the proof. ∎ As for the geometric meaning of these refinements, it turns out we are dealing with the interpolation of the function $f$ over the dyadic partition. ###### Proposition 2.9. Let $f:[0,1]\to\mathbb{R}$ be any function, and let $N\in\mathbb{N}$. Then, if $\nu_{i}=\frac{i}{2^{N}}$ for some $i=0,1,\cdots,2^{N},$ we have $f(\nu_{i})+\sum_{j=1}^{N}A_{j}(\nu_{i})\Delta_{j}f(\nu_{i};0,1)=(1-\nu_{i})f(0)+\nu_{i}f(1).$ (2.11) ###### Proof. Observe that when $\nu_{i}=\frac{i}{2^{N}},$ we have $[2^{N}\nu_{i}]=2^{N}\nu_{i}=i.$ From Lemma LABEL:exact_difference, we have $\displaystyle(1-\nu_{i})f(0)$ $\displaystyle+$ $\displaystyle\nu_{i}f(1)-\sum_{j=1}^{N}A_{j}(\nu_{i})\Delta_{j}f(\nu_{i};0,1)$ $\displaystyle=$ $\displaystyle\left([2^{N}\nu_{i}]+1-2^{N}\nu_{i}\right)f\left(\frac{[2^{N}\nu_{i}]}{2^{N}}\right)+\left(2^{N}\nu_{i}-[2^{N}\nu_{i}]\right)f\left(\frac{[2^{N}\nu_{i}]+1}{2^{N}}\right)$ $\displaystyle=$ $\displaystyle f\left(\frac{i}{2^{N}}\right)=f(\nu_{i}).$ This completes the proof. ∎ ###### Proposition 2.10. Let $f:[0,1]\to\mathbb{R}$ be a given function. If $f$ is continuous, then $f(\nu)+\lim_{N\to\infty}\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;0,1)=(1-\nu)f(0)+\nu f(1),$ (2.12) uniformly in $\nu\in[0,1].$ ###### Proof. Let $N\in\mathbb{N}$ and define the function $g_{N}(\nu)=\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;0,1).$ From Proposition 2.9, we have $g(\nu_{i})=(1-\nu_{i})f(0)+\nu_{i}f(1)-f(\nu_{i}),$ when $\nu_{i}=\frac{i}{2^{N}}$ for some $i=1,\cdots,2^{N}.$ Noting the definitions of $A_{j}$ and $\Delta_{j}f$, one can easily see that $g_{N}$ is linear on each dyadic interval $I_{i}:=\left[\frac{i}{2^{N}},\frac{i+1}{2^{N}}\right],i=0,\cdots,2^{N}-1.$ Now since $g_{N}$ is linear on $I_{i}$ and $g_{N}$ coincides with the continuous function $h(\nu):=(1-\nu)f(0)+\nu f(1)-f(\nu),$ it follows that $g_{N}$ is the linear interpolation of $h$ at the dyadic partition of $[0,1].$ Since $f$ is continuous, it follows that $g_{N}\to h$ uniformly, completing the proof. ∎ ### 2.2. Log-Convex function The proof of the following result follows from Theorem 2.5 on replacing $f$ by $\log f.$ ###### Corollary 2.11. Let $f:[a,b]\to(0,\infty)$ be log-convex. Then for $0\leq\nu\leq 1$ and $N\in\mathbb{N}$, we have $f\left((1-\nu)a+\nu b\right)\prod_{j=1}^{N}\left(\frac{f(x_{j}(\nu))f(y_{j}(\nu))}{f^{2}(z_{j}(\nu))}\right)^{A_{j}(\nu)}\leq f^{1-\nu}(a)f^{\nu}(b),$ (2.13) where $x_{j}(\nu),y_{j}(\nu)$ and $z_{j}(\nu)$ are as in the proof of Lemma 2.1. On the other hand, applying Theorem 2.7 implies the following. ###### Corollary 2.12. Let $f:[a,b]\to(0,\infty)$ be log-convex. Then for $0\leq\nu\leq\frac{1}{2}$ and $N\in\mathbb{N}$, we have $f\left((1-\nu)a+\nu b\right)\left(\frac{f(a)f(b)}{f^{2}(\frac{a+b}{2})}\right)^{1-A_{1}(\nu)}\geq f^{1-\nu}(a)f^{\nu}(b)\prod_{j=1}^{N}\left(\frac{f(t_{j}(\nu))f(u_{j}(\nu))}{f^{2}(w_{j}(\nu))}\right)^{A_{j}(1-2\nu)},$ (2.14) where $t_{j}(\nu),u_{j}(\nu)$ and $w_{j}(\nu)$ are obtained from the above $x_{j}(\nu),y_{j}(\nu)$ and $z_{j}(\nu)$ on replacing $(\nu,a,b)$ by $\left(1-2\nu,\frac{a+b}{2},b\right).$ On the other hand, if $\frac{1}{2}\leq\nu\leq 1,$ we have $f\left((1-\nu)a+\nu b\right)\left(\frac{f(a)f(b)}{f^{2}(\frac{a+b}{2})}\right)^{1-A_{1}(\nu)}\geq f^{1-\nu}(a)f^{\nu}(b)\prod_{j=1}^{N}\left(\frac{f(t_{j}(\nu))f(u_{j}(\nu))}{f^{2}(w_{j}(\nu))}\right)^{A_{j}(2-2\nu)},$ (2.15) where $t_{j}(\nu),u_{j}(\nu)$ and $w_{j}(\nu)$ are obtained from the above $x_{j}(\nu),y_{j}(\nu)$ and $z_{j}(\nu)$ on replacing $(\nu,a,b)$ by $\left(2-2\nu,a,\frac{a+b}{2}\right).$ The following is a squared additive version for log-convex functions. This inequality will help prove some squared versions of certain means. ###### Theorem 2.13. Let $f:[a,b]\to[0,\infty)$ be log-convex. Then for $0\leq\nu\leq 1$ and $N\geq 2$, we have $\displaystyle f^{2}((1-\nu)a+\nu b)$ $\displaystyle+$ $\displaystyle A_{1}^{2}(\nu)\Delta_{1}f^{2}(\nu;a,b)+\sum_{j=2}^{N}A_{j}(\nu)\Delta_{j}f^{2}(\nu;a,b)$ $\displaystyle\leq$ $\displaystyle\left((1-\nu)f(a)+\nu f(b)\right)^{2}.$ ###### Proof. We prove the result for $0\leq\nu\leq\frac{1}{2}.$ Since $f$ is log-convex, it follows that $g=f^{2}$ is log-convex too, and hence is convex. Therefore, Theorem 2.5 implies $g((1-\nu)a+\nu b)+\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}g(\nu;a,b)\leq(1-\nu)g(a)+\nu g(b),$ which implies, for $0\leq\nu\leq\frac{1}{2},$ $\displaystyle f^{2}((1-\nu)a+\nu b)$ $\displaystyle+$ $\displaystyle\nu^{2}\Delta_{1}f^{2}(\nu;a,b)+\sum_{j=2}^{N}A_{j}(\nu)\Delta_{j}f^{2}(\nu;a,b)$ (2.16) $\displaystyle\leq$ $\displaystyle\left((1-\nu)f(a)+\nu f(b)\right)^{2}+H(\nu;a,b),$ where $\displaystyle H(\nu;a,b)$ $\displaystyle=$ $\displaystyle(1-\nu)f^{2}(a)+\nu f^{2}(b)+\nu^{2}\Delta_{1}f^{2}(\nu;a,b)-\nu\Delta_{1}f^{2}(\nu;a,b)$ $\displaystyle-\left((1-\nu)f(a)+\nu f(b)\right)^{2}$ $\displaystyle=$ $\displaystyle 2\nu(1-\nu)\left(f^{2}\left(\frac{a+b}{2}\right)-f(a)f(b)\right)$ $\displaystyle\leq$ $\displaystyle 0,$ where the last inequality follows from log-convexity of $f$. Since $H(\nu;a,b)\leq 0$, it follows from (2.16) that $\displaystyle f^{2}((1-\nu)a+\nu b)$ $\displaystyle+$ $\displaystyle\nu^{2}\Delta_{1}f^{2}(\nu;a,b)+\sum_{j=2}^{N}A_{j}(\nu)\Delta_{j}f^{2}(\nu;a,b)$ $\displaystyle\leq$ $\displaystyle\left((1-\nu)f(a)+\nu f(b)\right)^{2}.$ Similar computations imply the result for $\frac{1}{2}\leq\nu\leq 1.$ ∎ Then reversed squared versions maybe obtained in a similar way from Theorem 2.7 as follows. ###### Theorem 2.14. Let $f:[a,b]\to[0,\infty)$ be log-convex and $N\in\mathbb{N}$. If $0\leq\nu\leq\frac{1}{2},$ we have $\displaystyle f^{2}((1-\nu)a+\nu b)$ $\displaystyle+$ $\displaystyle(1-\nu)^{2}\Delta_{1}f^{2}(\nu;a,b)+2\nu(1-\nu)\left(f(a)f(b)-f^{2}\left(\frac{a+b}{2}\right)\right)$ $\displaystyle\geq$ $\displaystyle\left((1-\nu)f(a)+\nu f(b)\right)^{2}$ $\displaystyle+\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f^{2}\left(1-2\nu;\frac{a+b}{2},b\right).$ If $\frac{1}{2}\leq\nu\leq 1,$ we have $\displaystyle f^{2}((1-\nu)a+\nu b)$ $\displaystyle+$ $\displaystyle\nu^{2}\Delta_{1}f^{2}(\nu;a,b)+2\nu(1-\nu)\left(f(a)f(b)-f^{2}\left(\frac{a+b}{2}\right)\right)$ $\displaystyle\geq$ $\displaystyle\left((1-\nu)f(a)+\nu f(b)\right)^{2}$ $\displaystyle+\sum_{j=1}^{N}A_{j}(2-2\nu)\Delta_{j}f^{2}\left(2-2\nu;a,\frac{a+b}{2}\right).$ ## 3\. Application ### 3.1. Refinements of means inequalities In this section we present some interesting applications of the above inequalities. The first result is the following refinement of Young’s inequality. ###### Corollary 3.1. Let $x,y>0,N\in\mathbb{N}$ and $0\leq\nu\leq 1.$ Then $x\\#_{\nu}y+\sum_{j=1}^{N}A_{j}(\nu)\left(\sqrt[2^{j}]{y^{k_{j}(\nu)}x^{2^{j-1}-k_{j}(\nu)}}-\sqrt[2^{j}]{y^{k_{j}(\nu)+1}x^{2^{j-1}-k_{j}(\nu)-1}}\right)^{2}\leq x\nabla_{\nu}y.$ (3.1) ###### Proof. This follows from Theorem 2.5, on letting $f(t)=x^{1-t}y^{t},a=0,b=1.$ Then $f$ is convex. Moreover, direct computations show that $\Delta_{j}f(\nu;0,1)=\left(\sqrt[2^{j}]{y^{k_{j}(\nu)}x^{2^{j-1}-k_{j}(\nu)}}-\sqrt[2^{j}]{y^{k_{j}(\nu)+1}x^{2^{j-1}-k_{j}(\nu)-1}}\right)^{2}.$ ∎ The above theorem has been recently proved in [sabjmaa] as a refinement of Young’s inequality. This inequality refines the corresponding refinements appearing in [5] and [11], where the inequality was proved only for $N=1,2.$ On the other hand, letting $f(t)=x!_{t}y$, the weighted harmonic mean, we obtain the following refinement of the arithmetic-harmonic mean inequality. ###### Corollary 3.2. Let $x,y>0,N\in\mathbb{N}$ and $0\leq\nu\leq 1.$ Then $x!_{\nu}y+\sum_{j=1}^{N}A_{j}(\nu)\left(x!_{\alpha_{j}(\nu)}y+x!_{\beta_{j}(\nu)}y-2x!_{\gamma_{j}(\nu)}y\right)\leq x\nabla_{\nu}y,$ (3.2) where $\alpha_{j}(\nu)=\frac{[2^{j-1}\nu]}{2^{j-1}},\beta_{j}(\nu)=\frac{[2^{j-1}\nu]+1}{2^{j-1}}$ and $\gamma_{j}(\nu)=\frac{\alpha_{j}(\nu)+\beta_{j}(\nu)}{2}.$ This inequality is a significant refinement of the corresponding inequality in [12], where the inequality was proved only for $N=1.$ Now noting log-convexity of the function $t\mapsto x!_{t}y$ on $[0,1],$ and applying Corollary 2.11, we get the following multiplicative refinement of the geometric-harmonic mean inequality. ###### Corollary 3.3. Let $x,y>0,N\in\mathbb{N}$ and $0\leq\nu\leq 1$, we have $x!_{\nu}y\prod_{j=1}^{N}\left(\frac{(x!_{\alpha_{j}(\nu)}y)(x!_{\beta_{j}(\nu)}y)}{(x!_{\gamma_{j}(\nu)}y)^{2}}\right)^{A_{j}(\nu)}\leq x\\#_{\nu}y,$ where $\alpha_{j}(\nu)=\frac{[2^{j-1}\nu]}{2^{j-1}},\beta_{j}(\nu)=\frac{[2^{j-1}\nu]+1}{2^{j-1}}$ and $\gamma_{j}(\nu)=\frac{\alpha_{j}(\nu)+\beta_{j}(\nu)}{2}.$ When $N=1$, Corollary 3.3, reduces to $(x!_{\nu}y)\left(\frac{x\nabla y}{x\\#y}\right)^{2\nu}\leq x\\#_{\nu}y,0\leq\nu\leq\frac{1}{2}$ and $(x!_{\nu}y)\left(\frac{x\nabla y}{x\\#y}\right)^{2(1-\nu)}\leq x\\#_{\nu}y,\frac{1}{2}\leq\nu\leq 1.$ The constant $\left(\frac{x\nabla y}{x\\#y}\right)^{2}$ appearing in these inequalities is called the Kantorovich constant, and has appeared in recent refinements of these mean inequalities. One can see [6] as a recent reference treating some inequalities using this constant. As for the squared version, applying Theorem 2.13 to the log-convex functions $t\mapsto x\\#_{t}y$ and $t\mapsto x!_{t}y$ implies the following. The first inequality refines the corresponding results in [3] and [11], while the other inequality is new. ###### Corollary 3.4. Let $x,y>0,0\leq\nu\leq 1,N\geq 2$ and $\alpha_{j}(\nu)=\frac{k_{j}(\nu)}{2^{j-1}},\beta_{j}(\nu)=\frac{k_{j}(\nu)+1}{2^{j-1}}\;{\text{and}}\;\gamma_{j}(\nu)=\frac{\alpha_{j}(\nu)+\beta_{j}(\nu)}{2}.$ Then $\displaystyle\left(x\\#_{\nu}y\right)^{2}$ $\displaystyle+$ $\displaystyle A_{1}^{2}(\nu)(x-y)^{2}+\sum_{j=2}^{N}A_{j}(\nu)\left(x^{1-\alpha_{j}(\nu)}y^{\alpha_{j}(\nu)}-x^{1-\beta_{j}(\nu)}y^{\beta_{j}(\nu)}\right)^{2}$ $\displaystyle\leq$ $\displaystyle(x\nabla_{\nu}y)^{2},$ and $\displaystyle\left(x!_{\nu}y\right)^{2}$ $\displaystyle+$ $\displaystyle 2A_{1}^{2}(\nu)(x^{2}\nabla y^{2}-(x!y)^{2})+\sum_{j=2}^{N}A_{j}(\nu)\left((x!_{\alpha_{j}(\nu)}y)^{2}+(x!_{\beta_{j}(\nu)}y)^{2}-2(x!_{\gamma_{j}(\nu)}y)^{2}\right)$ $\displaystyle\leq$ $\displaystyle(x\nabla_{\nu}y)^{2}.$ ### 3.2. Reversed Version Applying Theorem 2.7 to the function $f(t)=x\\#_{t}y$ implies the following reversed version of Young’s inequality. ###### Corollary 3.5. For $x,y>0$, let $f(t)=x\\#_{t}y$ and let $N\in\mathbb{N}$. If $0\leq\nu\leq\frac{1}{2},$ we have $\displaystyle x\\#_{\nu}y+(1-\nu)(\sqrt{x}-\sqrt{y})^{2}\geq x\nabla_{\nu}y+\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f\left(1-2\nu;\frac{1}{2},1\right).$ On the other hand, if $\frac{1}{2}\leq\nu\leq 1,$ we have $\displaystyle x\\#_{\nu}y+\nu(\sqrt{x}-\sqrt{y})^{2}\geq x\nabla_{\nu}y+\sum_{j=1}^{N}A_{j}(2-2\nu)\Delta_{j}f\left(2-2\nu;0,\frac{1}{2}\right).$ These inequalities refine those in [5] and [11]. Then an arithmetic-harmonic reversed version maybe obtained by applying Theorem 2.7 to the function $f(t)=x!_{t}y$ as follows. ###### Corollary 3.6. For $x,y>0$, let $f(t)=x\\#_{t}y$ and let $N\in\mathbb{N}$. If $0\leq\nu\leq\frac{1}{2},$ we have $\displaystyle x!_{\nu}y+(1-\nu)\left(x+y-2x!y\right)\geq x\nabla_{\nu}y+\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f\left(1-2\nu;\frac{1}{2},1\right).$ On the other hand, if $\frac{1}{2}\leq\nu\leq 1,$ we have $\displaystyle x!_{\nu}y+\nu(x+y-2x!y)\geq x\nabla_{\nu}y+\sum_{j=1}^{N}A_{j}(2-2\nu)\Delta_{j}f\left(2-2\nu;0,\frac{1}{2}\right).$ These inequalities refine those in [6]. Similarly, noting log-convexity of the function $f(t)=x!_{t}y$, we may apply Corollary 2.12 to obtain reversed multiplicative version of the harmonic- geometric mean inequality. We leave the application to the reader. Following the same guideline, we may obtain reversed squared versions by applying Theorem 2.14 to the functions $t\mapsto x\\#_{t}y$ and $x\mapsto x!_{t}y.$ Observe that when $f(t)=x\\#_{t}y$ we have $f(a)f(b)-f^{2}\left(\frac{a+b}{2}\right)=0.$ Therefore, applying Theorem 2.14 implies the following inequalities, which refine the corresponding inequalities in [3] and [11]. ###### Corollary 3.7. Let $x,y>0$ and $N\geq 1.$ If $0\leq\nu\leq 1,$ we have $\displaystyle\left(x\\#_{\nu}y\right)^{2}+(1-\nu)^{2}(x-y)^{2}\geq(x\nabla_{\nu}y)^{2}+\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f^{2}\left(1-2\nu;\frac{1}{2},1\right).$ If $\frac{1}{2}\leq\nu\leq 1,$ we have $\displaystyle\left(x\\#_{\nu}y\right)^{2}+\nu^{2}(x-y)^{2}\geq(x\nabla_{\nu}y)^{2}+\sum_{j=1}^{N}A_{j}(2-2\nu)\Delta_{j}f^{2}\left(1-2\nu;0,\frac{1}{2}\right).$ Now letting $g(t)=x!_{t}y$ we obtain the following new inequalities for the arithmetic-harmonic means. ###### Corollary 3.8. Let $x,y>0$ and $N\geq 1.$ If $0\leq\nu\leq 1,$ we have $\displaystyle\left(x!_{\nu}y\right)^{2}$ $\displaystyle+$ $\displaystyle 2(1-\nu)^{2}(x^{2}\nabla y^{2}-(x!y)^{2})+2\nu(1-\nu)\left(xy-\left(\frac{2xy}{x+y}\right)^{2}\right)$ $\displaystyle\geq$ $\displaystyle(x\nabla_{\nu}y)^{2}+\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}g^{2}\left(1-2\nu;\frac{1}{2},1\right).$ If $\frac{1}{2}\leq\nu\leq 1,$ we have $\displaystyle\left(x!_{\nu}y\right)^{2}$ $\displaystyle+$ $\displaystyle 2\nu^{2}(x^{2}\nabla y^{2}-(x!y)^{2})+2\nu(1-\nu)\left(xy-\left(\frac{2xy}{x+y}\right)^{2}\right)$ $\displaystyle\geq$ $\displaystyle(x\nabla_{\nu}y)^{2}+\sum_{j=1}^{N}A_{j}(2-2\nu)\Delta_{j}g^{2}\left(1-2\nu;0,\frac{1}{2}\right).$ ### 3.3. Some $L^{p}$ inequalities Let $(X,\mathcal{M},\mu)$ be a measure space, and let $0<p<q<r.$ Then $L^{p}\cap L^{r}\subset L^{q}$ and $\|f\|_{q}\leq\|f\|_{p}^{\nu}\|f\|_{r}^{1-\nu},\;{\text{where}}\;f\in L^{p}\cap L^{r}\;{\text{and}}\;\nu=\frac{q^{-1}-r^{-1}}{p^{-1}-r^{-1}}.$ This inequality can be modified using Corollary 2.11 and a reversed version can be obtained using Corollary 2.12. ###### Proposition 3.9. Let $(X,\mathcal{M},\mu)$ be a measure space, $0<p<q<r$ and $\nu$ be as above. If $f\in L^{p}\cap L^{r}$ and $N\in\mathbb{N}$, then we have $\|f\|_{q}\prod_{j=1}^{N}\left(\frac{\|f\|_{x_{j}^{-1}(\nu)}\|f\|_{y_{j}^{-1}(\nu)}}{\|f\|^{2}_{z_{j}^{-1}(\nu)}}\right)^{A_{j}(\nu)}\leq\|f\|_{p}^{\nu}\|f\|_{r}^{1-\nu}.$ ###### Proof. It is easy to check that the function $h(t)=\|f\|_{1/t}$ is log-convex on $[r^{-1},p^{-1}]$. Then direct application of Corollary 2.11 implies the result. ∎ In particular, when $N=1$, the above proposition implies $\|f\|_{q}\leq\left\\{\begin{array}[]{cc}\|f\|_{r}^{1-2\nu}\|f\|^{2\nu}_{\frac{2pr}{p+r}},&0\leq\nu\leq\frac{1}{2}\\\ \|f\|_{p}^{2\nu-1}\|f\|^{2-2\nu}_{\frac{2pr}{p+r}},&\frac{1}{2}\leq\nu\leq 1\end{array}\right\\}\leq\|f\|_{p}^{\nu}\|f\|_{r}^{1-\nu}.$ The condition $0\leq\nu\leq\frac{1}{2}$ can be interpreted as $\frac{2pr}{p+r}\leq q\leq r$, while $\frac{1}{2}\leq\nu\leq 1$ means $p\leq q\leq\frac{2pr}{p+r}.$ Moreover, a reversed version maybe obtained using Corollary 2.12. ###### Proposition 3.10. Let $(X,\mathcal{M},\mu)$ be a measure space, $0<p<q<r$ and $\nu=\frac{q^{-1}-r^{-1}}{p^{-1}-r^{-1}}$. If $\frac{2pr}{p+r}\leq q\leq r$, $f\in L^{p}\cap L^{r}$ and $N\in\mathbb{N}$, then $\|f\|_{q}\geq\|f\|^{2-2\nu}_{\frac{2pr}{p+r}}\|f\|_{p}^{2\nu-1}\prod_{j=1}^{N}\left(\frac{\|f\|_{t_{j}^{-1}(\nu)}\|f\|_{u_{j}^{-1}(\nu)}}{\|f\|^{2}_{w_{j}^{-1}(\nu)}}\right)^{A_{j}(1-2\nu)}\geq\|f\|^{2-2\nu}_{\frac{2pr}{p+r}}\|f\|_{p}^{2\nu-1},$ where $t_{j},u_{j}$ and $z_{j}$ are obtained from $x_{j},y_{j}$ and $z_{j}$ by replacing $(\nu,a,b)$ with $\left(1-2\nu,\frac{p+r}{2pr},p^{-1}\right).$ On the other hand, if $p\leq q\leq\frac{2pr}{p+r},$ then $\|f\|_{q}\geq\|f\|^{2\nu}_{\frac{2pr}{p+r}}\|f\|_{r}^{1-2\nu}\prod_{j=1}^{N}\left(\frac{\|f\|_{t_{j}^{-1}(\nu)}\|f\|_{u_{j}^{-1}(\nu)}}{\|f\|^{2}_{w_{j}^{-1}(\nu)}}\right)^{A_{j}(1-2\nu)}\geq\|f\|^{2\nu}_{\frac{2pr}{p+r}}\|f\|_{r}^{1-2\nu},$ where $t_{j},u_{j}$ and $z_{j}$ are obtained from $x_{j},y_{j}$ and $z_{j}$ by replacing $(\nu,a,b)$ with $\left(2-2\nu,r^{-1},\frac{p+r}{2pr}\right).$ Propositions 3.9 and 3.10 have been obtained using log-convexity of the function $h(t)=\|f\|_{t^{-1}}.$ In fact, noting log-convexity of the function $h(t)=\|f\|_{t}^{t}$, we obtain the same results! This is due to the equivalence of log-convexity of the functions $t\mapsto\|f\|_{t^{-1}}$ and that of $\|f\|_{t}^{t}.$ We refer the reader to [9] where these relations between the different log-convex function criteria have been discussed. The celebrated three lines lemma of Hadamard states the following. ###### Lemma 3.11. Let $\mathbb{D}=\\{z\in\mathbb{C}:0\leq\Re z\leq 1\\}$ and let $\varphi:\mathbb{D}\to\mathbb{C}$ be continuous on $\mathbb{D}$ and analytic in the interior of $\mathbb{D}$. Then the function $f:[0,1]\to\mathbb{R}$ defined by $f(x)=\sup_{y}|\varphi(x+iy)|$ is log-convex. This lemma is an extremely useful tool in the theory of complex functions. In particular, this lemma becomes handy in proving different interpolation versions of bounded linear operators between $L^{p}$ spaces. Log-convexity implied by Lemma 3.11 allows us to apply our refined and reversed versions for log-convex functions. In the following proposition, we present one term refinement and reverse. ###### Proposition 3.12. Let $\mathbb{D}=\\{z\in\mathbb{C}:0\leq\Re z\leq 1\\}$ and let $\varphi:\mathbb{D}\to\mathbb{C}$ be continuous on $\mathbb{D}$ and analytic in the interior of $\mathbb{D}$. Then the function $f:[0,1]\to\mathbb{R}$ defined by $f(x)=\sup_{y}|\varphi(x+iy)|$ satisfies the following $f(x)\leq\left\\{\begin{array}[]{cc}f^{1-2x}(0)f^{2x}\left(\frac{1}{2}\right),&0\leq x\leq\frac{1}{2}\\\ f^{2x-1}(1)f^{2-2x}\left(\frac{1}{2}\right),&\frac{1}{2}\leq x\leq 1\end{array}\right.$ and $f(x)\geq\left\\{\begin{array}[]{cc}f^{2x-1}(1)f^{2-2x}\left(\frac{1}{2}\right),&0\leq x\leq\frac{1}{2}\\\ f^{1-2x}(0)f^{2x}\left(\frac{1}{2}\right),&\frac{1}{2}\leq x\leq 1\end{array}\right..$ ### 3.4. Operator versions The following theorem provides a refinement of the well known Heinz inequality and its reverse. The proof follows immediately noting convexity of the Heinz means, see [1]. ###### Theorem 3.13. For $A,B\in\mathbb{M}_{n}^{+},X\in\mathbb{M}_{n},0\leq\nu\leq 1$ and any unitarily invariant norm $\||\;\;\||$, let $f(\nu)=\||A^{\nu}XB^{1-\nu}+A^{1-\nu}XB^{\nu}\||.$ Then we have the following refinement of Heinz inequality $\displaystyle\||A^{\nu}XB^{1-\nu}+A^{1-\nu}XB^{\nu}\||+\sum_{j=1}^{N}A_{j}(\nu)\Delta_{j}f(\nu;0,1)\leq\||AX+XB\||.$ Moreover, if $0\leq\nu\leq\frac{1}{2},$ we have $\displaystyle\||A^{\nu}XB^{1-\nu}+A^{1-\nu}XB^{\nu}\||$ $\displaystyle+$ $\displaystyle 2(1-\nu)\left(\||AX+XB\||-\||\sqrt{A}X\sqrt{B}\||\right)$ $\displaystyle\geq$ $\displaystyle\||AX+XB\||+\sum_{j=1}^{N}A_{j}(1-2\nu)\Delta_{j}f\left(1-2\nu;\frac{1}{2},1\right).$ On the other hand, if $\frac{1}{2}\leq\nu\leq 1,$ we have $\displaystyle\||A^{\nu}XB^{1-\nu}+A^{1-\nu}XB^{\nu}\||$ $\displaystyle+$ $\displaystyle 2\nu\left(\||AX+XB\||-\||\sqrt{A}X\sqrt{B}\||\right)$ $\displaystyle\geq$ $\displaystyle\||AX+XB\||+\sum_{j=1}^{N}A_{j}(2-2\nu)\Delta_{j}f\left(2-2\nu;0,\frac{1}{2}\right).$ In [9], it is shown that for $A,B\in\mathbb{M}_{n}^{+}$ and $X\in\mathbb{M}_{n},$ the functions $t\to\||A^{t}XB^{1-t}\||,t\to\||A^{t}XB^{1-t}\||\;\||A^{1-t}X^{B}{t}\||,t\to{\rm{tr}}(A^{t}XB^{1-t}X^{*})$ are log-convex on $[0,1].$ Therefore, we may apply Corollaries 2.11 and 2.12 to obtain refinements and reversed versions for such functions. For the $\|\;\;\|_{2}$ norm, we can prove log convexity of the Heinz means, which allows us to obtain further refinements of the Heinz inequality by applying Corollaries 2.11 and 2.12. ###### Proposition 3.14. Let $A,B\in\mathbb{M}_{n}^{+}$ and $X\in\mathbb{M}_{n}$, and define $f(\nu)=\|A^{\nu}XB^{1-\nu}+A^{1-\nu}XB^{\nu}\|_{2}.$ Then $f$ is log-convex on $[0,1]$. ###### Proof. Since $A,B\in\mathbb{M}_{n}^{+}$, there are diagonal matrices $D_{1}:={\text{diag}}(\lambda_{i}),D_{2}:={\text{diag}}(\mu_{i})$ and unitarily matrices $U,V$ such that $\lambda_{i},\mu_{i}\geq 0,$ $A=UD_{1}U^{*}$ and $B=VD_{2}V^{*}.$ Letting $Y=U^{*}XV,$ we have $A^{\nu}XB^{1-\nu}+A^{1-\nu}XB^{\nu}=U(\lambda_{i}^{\nu}y_{ij}\mu_{j}^{1-\nu}+\lambda_{i}^{1-\nu}y_{ij}\mu_{j}^{\nu})V^{*}.$ Since $\|\;\;\|_{2}$ is a unitarily invariant norm, we have $\displaystyle f^{2}(\nu)$ $\displaystyle=$ $\displaystyle\|U(\lambda_{i}^{\nu}y_{ij}\mu_{j}^{1-\nu}+\lambda_{i}^{1-\nu}y_{ij}\mu_{j}^{\nu})V^{*}\|_{2}^{2}$ $\displaystyle=$ $\displaystyle\sum_{i,j}\left(\lambda_{i}^{\nu}\mu_{j}^{1-\nu}+\lambda_{i}^{1-\nu}\mu_{j}^{\nu}\right)^{2}|y_{ij}|^{2}.$ Notice that each summand is log-convex, being the square of a log-convex function. This implies that $f^{2}$ is log-convex. Consequently, $f$ is log- convex. ∎ Letting $f(\nu)=\|A^{\nu}XB^{1-\nu}+A^{1-\nu}XB^{\nu}\|_{2}$ and applying Theorem 2.13 imply the following squared version of Heinz inequality. ###### Corollary 3.15. Let $A,B\in\mathbb{M}_{n}^{+},X\in\mathbb{M}_{n},0\leq\nu\leq 1$ and $N\geq 2.$ Then $\displaystyle\|A^{\nu}XB^{1-\nu}+A^{1-\nu}XB^{\nu}\|_{2}^{2}+2A_{1}^{2}(\nu)\left(\|AX+XB\|_{2}^{2}-2\|A^{\frac{1}{2}}XB^{\frac{1}{2}}\|_{2}^{2}\right)$ $\displaystyle\hskip 8.5359pt+\sum_{j=2}^{N}A_{j}(\nu)\Delta_{j}f^{2}(\nu;0,1)\leq\|AX+XB\|_{2}^{2}.$ We leave the application of Corollaries 2.11 and 2.12 to the reader. Further operator versions maybe obtained using Lemma 1.1. The following operator versions refine the corresponding results in [5] and [11]. ###### Proposition 3.16. Let $A,B\in\mathbb{M}_{n}^{++}$ and $0\leq\nu\leq 1.$ Then for $\alpha_{j}(\nu)=\frac{k_{j}(\nu)}{2^{j-1}},\beta_{j}(\nu)=\frac{k_{j}(\nu)+1}{2^{j-1}},\gamma_{j}(\nu)=\frac{\alpha_{j}(\nu)+\beta_{j}(\nu)}{2}$ and $N\in\mathbb{N}$, we have $\displaystyle A\\#_{\nu}B+\sum_{j=1}^{N}A_{j}(\nu)\left(A\\#_{\alpha_{j}(\nu)}+A\\#_{\beta_{j}(\nu)}-2A\\#_{\gamma_{j}(\nu)}B\right)\leq A\nabla_{\nu}B.$ ###### Proof. In Corollary 3.1, let $x=1$, expand the summand and apply Lemma 1.1 with $y$ replaced by $X=A^{-\frac{1}{2}}BA^{-\frac{1}{2}}.$ Then the result follows upon conjugating both sides with $A^{\frac{1}{2}}.$ ∎ In a similar way one may obtain reversed versions by applying Corollary 3.5. This provides refinements of the reversed versions of [11]. The following is an operator arithmetic-harmonic version, refining the corresponding results in [12]. ###### Proposition 3.17. Let $A,B\in\mathbb{M}_{n}^{++}$ and $0\leq\nu\leq 1.$ Then for $\alpha_{j}(\nu)=\frac{k_{j}(\nu)}{2^{j-1}},\beta_{j}(\nu)=\frac{k_{j}(\nu)+1}{2^{j-1}},\gamma_{j}(\nu)=\frac{\alpha_{j}(\nu)+\beta_{j}(\nu)}{2}$ and $N\in\mathbb{N}$, we have $\displaystyle A!_{\nu}B+\sum_{j=1}^{N}A_{j}(\nu)\left(A!_{\alpha_{j}(\nu)}+A!_{\beta_{j}(\nu)}-2A!_{\gamma_{j}(\nu)}B\right)\leq A\nabla_{\nu}B.$ The proof follows immediately on applying Lemma 1.1 together with Corollary 3.2. On the other hand, applying Corollary 3.6 implies the following refinement of the corresponding inequalities in [6]. ###### Proposition 3.18. Let $A,B\in\mathbb{M}_{n}^{++}$ and $N\in\mathbb{N}$. If $0\leq\nu\leq 1,$ we have $\displaystyle A!_{\nu}B$ $\displaystyle+$ $\displaystyle(1-\nu)(A+B-2A!B)$ $\displaystyle\geq$ $\displaystyle A\nabla_{\nu}B+\sum_{j=1}^{N}A_{j}(1-2\nu)\left(A!_{\alpha_{j}(\nu)}B+A!_{\beta_{j}(\nu)}B-2A!_{\gamma_{j}(\nu)}B\right),$ where $\alpha_{j}(\nu)=\frac{1}{2}\left(1-\frac{k_{j}(1-2\nu)}{2^{j-1}}\right)+\frac{k_{j}(1-2\nu)}{2^{j-1}},$ $\beta_{j}(\nu)=\frac{1}{2}\left(1-\frac{k_{j}(1-2\nu)+1}{2^{j-1}}\right)+\frac{k_{j}(1-2\nu)+1}{2^{j-1}}$ and $\gamma_{j}(\nu)=\frac{\alpha_{j}(\nu)+\beta_{j}(\nu)}{2}.$ On the other hand, if $\frac{1}{2}\leq\nu\leq 1,$ we have $\displaystyle A!_{\nu}B$ $\displaystyle+$ $\displaystyle\nu(A+B-2A!B)$ $\displaystyle\geq$ $\displaystyle A\nabla_{\nu}B+\sum_{j=1}^{N}A_{j}(2-2\nu)\left(A!_{\alpha_{j}(\nu)}B+A!_{\beta_{j}(\nu)}B-2A!_{\gamma_{j}(\nu)}B\right),$ where $\alpha_{j}(\nu)=\frac{k_{j}(2-2\nu)}{2^{j-1}},$ $\beta_{j}(\nu)=\frac{k_{j}(2-2\nu)+1}{2^{j-1}}$ and $\gamma_{j}(\nu)=\frac{\alpha_{j}(\nu)+\beta_{j}(\nu)}{2}.$ The following is an interesting one-term multiplicative refinement of the operator geometric-harmonic mean inequality. ###### Theorem 3.19. Let $A,B\in\mathbb{M}_{n}^{++}$ and $0\leq\nu\leq 1.$ Then $\displaystyle(A!_{\nu}B)\left(\frac{A^{-1}B+2I+B^{-1}A}{4}\right)^{r}\leq A\\#_{\nu}B,$ where $r=\min\\{\nu,1-\nu\\}.$ ###### Proof. We prove the desired inequality for $0\leq\nu\leq\frac{1}{2}.$ In Corollary 3.3, let $N=1$ and $x=1$, to get $(1!_{\nu}y)\left(\frac{1+y}{2\sqrt{y}}\right)^{2\nu}\leq 1\\#_{\nu}y,$ or $\displaystyle\frac{1}{4^{\nu}}\left((1-\nu)+\nu y^{-1}\right)^{-1}\left(y+2+y^{-1}\right)^{\nu}\leq y^{\nu}.$ (3.3) Let $X=A^{-\frac{1}{2}}BA^{-\frac{1}{2}}$ and apply Lemma 1.1. The left hand side of (3.3) becomes $\displaystyle\frac{1}{4^{\nu}}\left((1-\nu)I+\nu A^{\frac{1}{2}}B^{-1}A^{\frac{1}{2}}\right)^{-1}\left(A^{-\frac{1}{2}}BA^{-\frac{1}{2}}+2I+A^{\frac{1}{2}}B^{-1}A^{\frac{1}{2}}\right)^{\nu}$ (3.4) $\displaystyle=$ $\displaystyle\frac{1}{4^{\nu}}\left[A^{-\frac{1}{2}}(A!_{\nu}B)A^{-\frac{1}{2}}\right]\left[A^{\frac{1}{2}}\left(A^{-1}B+2I+B^{-1}A\right)^{\nu}A^{-\frac{1}{2}}\right]$ $\displaystyle=$ $\displaystyle A^{-\frac{1}{2}}(A!_{\nu}B)\left(\frac{A^{-1}B+2I+B^{-1}A}{4}\right)^{\nu}A^{-\frac{1}{2}}.$ On the other hand, the right hand side of (3.3) is simply $\left(A^{-\frac{1}{2}}BA^{-\frac{1}{2}}\right)^{\nu}.$ This together with (3.4) imply the desired inequality, upon conjugating both sides with $A^{\frac{1}{2}}.$ This completes the proof. ∎ ## References * [1] R. Bhatia, Matrix analysis, Springer-Verlag, New York, 1997. * [2] T. Furuta, J. Micic Hot and J. Pecaric, Mond-Pecaric Method in Operator Inequalities, Element, Zagreb, 2005. * [3] O. Hirzallah and F. Kittaneh, Matrix Young inequalities for the Hilbert-Schmidt norm, Linear Algebra Appl. 308 (2000), 77–84. * [4] F. Kittaneh, On the convexity of the Heinz mean, Integr. Equ. Oper. Theory 68 (2010), 519–527. * [5] F. Kittaneh and Y. Manasrah, Improved Young and Heinz inequalities for matrices, J. Math. Anal. Appl. 36 (2010), 262–269. * [6] W. Liao and J. Wu, Reverse arithmetic-harmonic mean and mixed mean operator inequalities, J. Inequal. Appl., 2015:215. * [7] Y. Manasrah and F. Kittaneh, A generalization of two refined Young inequalities, Positivity 19 (2015), 757–768. * [8] M. Sababheh, Convex functions and means of matrices, Math. Ineq. Appl., accepted. * [9] M. Sababheh, Log and harmonically log-convex functions related to matrix norms, Operators and Matrices, in press. * [10] M. Sababheh, Integral inequalities of the Heinz means as convex functions, J. Math. Ineq., 10 (2) (2016), 313–325. * [11] J. Zhao and J. Wu, Operator inequalities involving improved Young and its reverse inequalities, J. Math. Anal. Appl. 421 (2015) 1779–1789. * [12] H. Zuo, G. Shi and M. Fujii, Refined Young inequality with Kantorovich constant, J. Math. Inequal. 5(4)(2011), 551-556 .
# Spin Hall Nano-Oscillator Empirical Electrical Model for Optimal On-chip Detector Design Rafaella Fiorelli<EMAIL_ADDRESS>Instituto de Microelectrónica de Sevilla, CSIC and Universidad de Sevilla, Sevilla, 41092. Mona Rajabali NanOsc AB, Kista, 16440 Sweden. Roberto Méndez-Romero Instituto de Microelectrónica de Sevilla, CSIC and Universidad de Sevilla, Sevilla, 41092. Akash Kumar Applied Spintronics Group, Department of Physics, University of Gothenburg, 41296 Gothenburg, Sweden. Research Institute of Electrical Communication and Center for Science and Innovation in Spintronics, Tohoku University, 2-1-1 Katahira, Aoba-ku, Sendai 980-8577 Japan. Artem Litvinenko Applied Spintronics Group, Department of Physics, University of Gothenburg, 41296 Gothenburg, Sweden. Teresa Serrano-Gotarredona Instituto de Microelectrónica de Sevilla, CSIC and Universidad de Sevilla, Sevilla, 41092. Farshad Moradi Integrated Circuits and Electronics Lab (ICELab), Electrical and Computer Engineering Department, Aarhus University, 8200 Aarhus N, Denmark. Johan Åkerman Applied Spintronics Group, Department of Physics, University of Gothenburg, 41296 Gothenburg, Sweden. Research Institute of Electrical Communication and Center for Science and Innovation in Spintronics, Tohoku University, 2-1-1 Katahira, Aoba-ku, Sendai 980-8577 Japan. Bernabé Linares-Barranco Instituto de Microelectrónica de Sevilla, CSIC and Universidad de Sevilla, Sevilla, 41092. Eduardo Peralías Instituto de Microelectrónica de Sevilla, CSIC and Universidad de Sevilla, Sevilla, 41092. ###### Abstract As nascent nonlinear oscillators, nano-constriction spin Hall nano-oscillators (SHNOs) represent a promising potential for integration into more complicated systems such as neural networks, magnetic field sensors, and radio frequency (RF) signal classification, their tunable high-frequency operating regime, easy synchronization, and CMOS compatibility can streamline the process. To implement SHNOs in any of these networks, the electrical features of a single device are needed before designing the signal detection CMOS circuitry. This study centers on presenting an empirical electrical model of the SHNO based on a comprehensive characterization of the output impedance of a single SHNO, and its available output power in the range of 2-10 GHz at various bias currents. ## I Introduction As complementary metal-oxide-semiconductor (CMOS) technology is reaching its physical limit, other technologies are about to thrive. In particular, nonlinear spiking and oscillatory spintronic devices exhibit tremendous potential in leading-edge areas such as emulating the spiking behaviors of neurons Torrejon _et al._ (2017); Romera _et al._ (2018); Zahedinejad _et al._ (2020); Houshang _et al._ (2022); Kumar _et al._ (2023a), RF signal classification Ross _et al._ (2023), ultra-fast microwave spectral analysis Litvinenko _et al._ (2022), and the possibility of implementing highly accelerated neuromorphic computing systems Müller _et al._ (2022); González _et al._ (2024). Among the family of spintronic microwave oscillators, nano- constriction spin Hall nano-oscillators (SHNOs) are simple heavy metal/ferromagnet bilayers (HM/FM) through which the pure spin current is produced by passing DC bias current ($I_{B}$) and leads to a steady-state precession, known as auto-oscillation Demidov _et al._ (2014); Chen _et al._ (2016); Behera _et al._ (2024). Considering their facile fabrication, broad frequency tunability Zahedinejad _et al._ (2018), easy injection lockingRajabali _et al._ (2023), robust mutual synchronization Kumar _et al._ (2023b), and individual tunability Muralidhar _et al._ (2022); Khademi _et al._ (2024), the SHNOs are particularly promising for various applications from magnetic field sensors Xie, Cheung, and Fuchs (2023) to neural network integration Sethi _et al._ (2023); Kumar _et al._ (2024); González _et al._ (2024). When implementing an SHNO-based network, it is essential to ascertain the oscillation status of the SHNO (e.g., firing/non-firing in case of a neural network). However, due to the relatively low power and high noise of the output oscillating signal Litvinenko _et al._ (2023), the CMOS detector and the SHNO necessitate achieving a good impedance matching Bendjeddou _et al._ (2023). If not, reflecting a significant share of the available SHNO output power makes it more challenging to detect the SHNO signal, as confirmed in Fiorelli _et al._ (2023). In the realm of SHNO output signal detection, fully integrated approaches not only reduce detector size and power consumption but also are more adaptable for the joint integration of SHNO and detectors; whether homogeneous or heterogeneous. The selection of the optimal technology and detector’s architecture is dependent on the SHNO’s electrical characteristics, which include its working frequency, output power, noise, and output impedance Fiorelli _et al._ (2023). Hence, an accurate electrical model of the SHNO is mandatory. This model must include information about the SHNO: (i) output power, (ii) noise levels, and (iii) output impedance. This information allows for knowing the SHNO signal-to-noise (SNR) ratio and therefore deduce the maximum accepted noise level of the receiver so that the SHNO signal is not masked by its noise. Moreover, an optimized design of the electrical circuitry that senses the SHNO output signal requires information on the SHNO output impedance to address an excellent impedance coupling between this detector and the SHNO, minimizing the power losses. It is known that the SHNO output power increases as its oscillation frequency goes up, which in turn, allows a quicker classification of its state Fulara _et al._ (2019). But this increase in frequency must be accompanied by a guarantee of observation of the oscillation by the integrated detector, whose implementation is limited to its (i) maximum detection frequency and, (ii) minimum practicable bandwidth (BW) associated with the SNR. In this study, we demonstrate that excellent trade-offs in terms of SHNO output power, detector’s feasibility, and performance, can be achieved by selecting the optimal oscillating frequency around 6 GHz. Furthermore, the empirical electrical model of the SHNO is extracted and presented for the frequency range of 2-10 GHz. The SNR will not be notably favorable when the detector possesses a broad BW Pozar (2011) (i.e., exceeding 100 MHz as noted later in Section III-A), which is likely to be the case in fully integrated detectors. Hence, achieving the desired signal matching and developing an accurate electrical AC SHNO model is critical if we aim to leverage the full potential of the available SNR. Although other types of nano-oscillators were modeled electrically Bendjeddou _et al._ (2023), there is a notable absence of an AC SHNO model that comprehensively describes its output impedance, power characteristics, and noise equivalent output. For the first time, this paper introduces an AC empirical electrical model for an SHNO derived from experimental measurements. This AC evaluation holds immense significance since only considering the DC range characterization fails to provide sufficient insights for an optimal signal detector design, as in Fiorelli _et al._ (2023). This AC empirical model is developed by the data captured from two sets of experiments that measure the SHNO load output power, $P_{L}$, and the output impedance, $Z_{SHNO}$. Figure 1: (a) Fabricated SHNO with GSG CPW. (b) Magnified scheme of SHNO geometry and the contact pads. The inset displays the detail of the heterostructure on which SHNO is fabricated. (c) The SEM top-view image of a single 180 nm wide SHNO. This paper is outlined as follows. Section II presents fabrication details and the experimental setup to measure the SHNO impedance and power. The measurement results are displayed in Section III, whereas the empirical electrical model is deployed in Section IV. Finally, Section V concludes our approach to develop this empirical electrical model for SHNOs for an optimal on-chip detector design. ## II Experimental part The 180-nm wide SHNOs, used in the measurements to extract the empirical electrical model, were fabricated on heterostructures consisting of W(5nm)/Py(5nm)(Ni80Fe20)/Al2O3 (4nm) Kumar _et al._ (2023b),Litvinenko _et al._ (2023). The fabrication process involves DC/RF magnetron sputtering of the stack at room temperature onto a high-resistance silicon substrate (20$\times$20 mm2) with a base pressure of less than 3$\times$10-8 Torr. The heterostructure is then patterned into 8$\times$12 $\mu$m2 rectangles with bow-tie shaped NCs using a Raith EBPG 5200 electron beam lithography (EBL) system followed by Ar-ion etching. Optical lithography is used to define the ground-signal-ground (GSG) coplanar waveguide (CPW), which is followed by a deposition and liftoff process using a bi-layer of Cu(500nm)/Pt(20nm) as top contacts (Fig. 1(a)). A magnified scheme of the SHNO is presented in Fig. 1(b), along with the SHNO layer arrangement details in the cross-sectional view (inset). Furthermore, a top-view scanning electron microscopy (SEM) image of the fabricated SHNO is seen in Fig. 1(c) [more details of fabrication can be found in Kumar _et al._ (2022)]. Figure 2: Schematic and setup photo for Experiment 1 (a-b) and Experiment 2 (c-d). ### II.1 Impedance setup Experiment 1 focuses on the impedance measurement, with the corresponding setup scheme and photo presented in Fig. 2(a) and (b), respectively. This measurement is conducted using an Agilent N5230A Vector Network Analyzer (VNA). The SHNO is biased via the Mini-circuits ZX85-12G+ bias-tee. Its output is connected to the VNA and evaluated between 2-10 GHz, under the values of $I_{B}=\\{2.0,2.25,2.35,2.45\\}~{}mA$. To do impedance measurements over a biased device-under-test (DUT), all external electromagnetic signals of the DUT are voided except the one that injects into the DUT node where the impedance is to be known. Hence, this test is performed in the absence of an external magnetic field. Furthermore, although the frequency of interest to obtain the model is around 6 GHz, the wide range of frequencies of [2,10] GHz was evaluated to assess its behavior, ensuring both a smooth behavior at high frequencies and a tendency to the expected values at the DC level. The accuracy of the impedance measurement relies on the careful calibration of the entire setup to remove the effect of cables C1, C2, bias-tee, and picoprobe (Fig. 2(a)). We made use of the Cascade Microtech GSG 150 $\mu$m-pitch P/N 101-190 impedance standard substrate to carry out this calibration, obtaining excellent results with a measured voltage standing wave ratio lower than 1.1 at [2,10] GHz range. ### II.2 Power and noise setup Experiment 2 aims to find the maximum SHNO output power ( $P_{meas\\_peak}$) around 6 GHz, using a spectrum analyzer. Its setup and photo are given in Fig. 2(c) and (d), respectively. The GSG picoprobe (GGB Industries) conveys the input $I_{B}$ and the output generated signal between the SHNO and a bias-tee (MITEQ BT4000). The auto-oscillation is driven by $I_{B}$ running to the sample through the bias-tee DC port. Then, the generated SHNO RF signal goes to a low noise amplifier (LNA, B&Z BZ0218A) with a transducer power gain $G_{T}=23~{}dB$ and eventually reaches the Spectrum Analyzer (SA, R&S FSV 40 GHz), where the spectra data is captured. The optimum magnetic field $\overrightarrow{B}$ is obtained by preliminary field scans at several values of $I_{B}$. At the chosen magnetic field, the signal will then be recorded while increasing $I_{B}$ to reach a maximum power value such that a further increase in $I_{B}$ does not substantially increase the measured power, $P_{meas}$. Additionally, the second outcome of the experiment pertains to the noise floor level, which will be useful in further calculations. As a result, the sample is subjected to an external magnetic field ($|B|=0.68~{}T$) at out-of-plane ($\theta=84^{\circ}$) and in-plane ($\varphi=22^{\circ}$) angles. The value $I_{B}=2.45~{}mA$ above which $P_{meas}$ does not vary substantially is considered the one that generates $P_{meas\\_peak}$. ## III Results and discussion ### III.1 Impedance results Impedance measurement results are deployed in Fig. 3. In the RF/MW frequencies, the real part of $Z_{SHNO}$, $Re(Z_{SHNO})$, decreases almost linearly with frequency, reaching approximately 350 $\Omega$ at 6 GHz, and 250 $\Omega$ at 10 GHz (Fig. 3(a)). This result is promising since making the SHNO work at higher frequencies implies a lower $Re(Z_{SHNO})$, simplifying the detector design and allowing it to reach optimal matching with the detector. In addition, Fig. 3(a) verifies that $Re(Z_{{SHNO}})$ increases as $I_{B}$ increases. Figure 3: Real and imaginary part of $Z_{SHNO}$ for four $I_{B}$ (a,b), and four SHNO samples at $I_{B}=2.45~{}mA$ (c,d). Fig. 3(b) depicts the imaginary part $Z_{SHNO}$, $Im(Z_{SHNO})$, which diminishes as $I_{B}$ decreases for the whole frequency range. In particular, we observe a capacitive behavior over the entire frequency span under study. Finally, we measure the values of $Re(Z_{SHNO})$ and $Im(Z_{SHNO})$ for four SHNO identical samples to evaluate its dispersion. The findings, shown in Fig. 3(c) and (d), yield a variation of 5% and 2.5% around 6 GHz, respectively. ### III.2 Power and noise results The captured spectrum corresponding to $I_{B}=2.45~{}mA$ is illustrated in Fig. 4, in the range of [6.0,6.5] GHz. Using a BW of 1 MHz, the maximum measured power is $P_{meas\\_peak}=-55~{}dBm$ at 6.25 GHz. This plot also yields noise floor power in the order of $P_{meas\\_noise\\_floor}\approx-80~{}dBm$. Then, an estimation of SHNO output signal power will be about $P_{L}\approx P_{meas\\_peak}-G_{T}=-79~{}dBm$, and an estimation of the power spectral density (PSD) of the signal at the SHNO output is shown in the inset of Fig. 4 with a noise floor power density of $P_{noise\\_floor}\approx P_{meas\\_noise\\_floor}-10~{}log_{10}(BW)-G_{T}=-163~{}dBm/Hz$. Figure 4: SHNO output spectrum for $I_{B}=2.45~{}mA$ with BW=1 MHz, considering noise levels associated with a detector with BW = [1, 100, 500] MHz. Inset: PSD for $I_{B}$ = [$2.0~{}mA$ (blue), $2.25~{}mA$ (orange), $2.35~{}mA$ (yellow), $2.45~{}mA$ (purple)]. The feasibility of the detection strongly depends on the detector BW, which is reflected in the SNR; for instance, SNR=$\\{25dB_{@1MHz},5dB_{@100MHz},-2dB_{@500MHz}\\}$. Given that the BW is anticipated to be no less than 100 MHz in a monolithically integrated CMOS detector, we expect a low peak PSD level. This low SNR necessitates maximizing the available power at the detector input, highlighting the critical role of impedance matching. ## IV Electrical model The information collected with Experiment 1 and Experiment 2 allows us to develop the SHNO Thévenin model, as seen in Fig. 5(a). It comprises the Thévenin voltage, $V_{SHNO}$, the equivalent output noise voltage $V_{n}$, and the equivalent output impedance $Z_{SHNO}$. The equations that describe this model. presented next, are derived from classical circuits’ theory Alexander and Sadiku (2016). The amplitude of the Thévenin’s voltage, $|V_{SHNO}|$ and $Z_{SHNO}$, are related to the estimated output peak power, $P_{L}(\approx-79dBm)$, over the load impedance $Z_{L}$, by the following expression: $\left|V_{SHNO}\right|=\left|Z_{SHNO}+Z_{L}\right|\sqrt{\frac{2P_{L}}{Re(Z_{L})}}$ (1) where $Z_{L}$ is the impedance seen at the output of the SHNO of the experimental setup shown in Fig. 2(c), ideally, $Z_{L}=50\Omega$. Figure 5: (a) SHNO empirical electrical model. (b) Real and (c) imaginary part of $Z_{SHNO}$ measured at $I_{B}=2.45mA$ (red line) and $Z_{SHNO}(\omega)^{model}$ (black). The proposed impedance model is presented in the inset of Fig. 5(a), being consistent with the device impedance measurements in DC. It comprises a resistor $R_{SHNO}$ in parallel with a capacitor $C_{SHNO}$, and where $Z_{SHNO}(\omega)^{model}=\frac{1}{1/R_{SHNO}+j\omega C_{SHNO}}$ (2) The values of $R_{SHNO}$ and $C_{SHNO}$ are listed in the table embedded in Fig. 5(a), which are valid over the whole frequency range of [2,10] GHz (for $I_{B}=2.45~{}mA$). This is reflected in Fig. 5(b-c), where the real and imaginary parts of $Z_{SHNO}$ are correctly described over the whole frequency span with the black curves of $Z_{SHNO}(\omega)^{model}$. The root-mean-square of noise voltage source, $v_{n}$, is obtained by using (1), and substituting $2P_{L}$ for the estimated floor noise power, $P_{{noise\\_floor}}(\approx-163dBm/Hz)$. Finally, to complete the empirical model, we bring the available power gain of the SHNO, $P_{av,SHNO}$. As it is known, this is the maximum power of the SHNO that can be delivered to the load, and it occurs when the load is conjugately matched to the SHNO output impedance. It does not depend on the load and gives the designer a maximum value of power that could be delivered to the load, although this is never feasible in practice. This expression is, $P_{av,SHNO}=\frac{|V_{SHNO}|^{2}}{8Re(Z_{SHNO})}$ (3) The numerical values of the empirical electrical model are listed in the table of Fig. 5(a). They have been obtained from the experimental quantities presented in subsections IIIA and IIIB and the expressions (1)-(3). To conclude this section, we provide a practical example of how critical is to provide a good SHNO electrical model in designing an associated signal detector with an input impedance $Z_{det}$. Assuming the SHNO device interfaces with a noiseless detector featuring a practical BW of several hundred MHz, the detector output is anticipated to exhibit a maximum SNR of around 3 dB (Refer to Fig. 4). Figure 6: SNR at the input of the detector: (a) versus the real parts of $Z_{det}$ and $Z_{SHNO}$ impedances with opposite imaginary parts, and (b) the imaginary part of $Z_{SHNO}+Z_{det}$ for different $Re(Z_{det})$ at $R_{SHNO}=350\Omega$. Owing to the high value of $R_{SHNO}(\approx 350\Omega$ at 6 GHz), the main challenge of achieving the matching condition lies in equalizing $Re({Z_{det}})$ to $R_{SHNO}$. This is particularly troublesome when attempting on-chip detectors, to achieve a joint integration of multiple units. To illustrate this challenge, Fiorelli _et al._ (2023) has reported the difficulty of designing on-chip detectors in CMOS technologies at 4.7 GHz, with $Re(Z_{det})$ above 100 $\Omega$. Presuming the input impedance of the detector, $Z_{det}$, can be complex, and considering the maximum power-transfer theorem, the seek of maximum power transfer from the SHNO to the detector implies that $Z_{det}=Z_{SHNO}^{*}$. This is especially desirable due to the very low value of $P_{av,SHNO}$. To elaborate on the effect of the mismatch between the SHNO and the detector in the SNR of the system, with a maximum value of SNR = +3 dB, one can initially consider a partial conjugation such that only the imaginary components of the impedance are matched, i.e. $Im(Z_{det})=-Im(Z_{shno})$. When both $Re(Z_{det})$ and $Re(Z_{SHNO})$ depart from equality, the SNR falls below 0 (see Fig. 6(a)). For example, at $Re(Z_{SHNO})=350\Omega$ (red dash line), for $Re(Z_{det})<100\Omega$, the SNR is below 1.5 dB, and thus eliminating most of the margin that guarantees the detection of the oscillation signal. Let’s now consider the case when the net imaginary component, $\rho=|Im(Z_{SHNO}+Z_{det})|$ departs from the ideal condition ($\rho=0$). For $\rho=$[0,300]$~{}\Omega$ and $Re(Z_{det})$=[25,350] $\Omega$ we obtain the plot shown in Fig. 6(b). For instance, when $\rho>$100 $\Omega$ and $Re(Z_{det})<$ 200 $\Omega$, the deviation from matching conditions causes SNR to drop below 0 dB and therefore making it impossible to detect the oscillation signal. In other words, to ensure signal detection, it’s crucial to design the detector with the goal of minimizing the net imaginary component, ideally approaching $\rho=0$, while simultaneously optimizing $Re(Z_{det})$ to closely match the oscillator resistance. By establishing these key criteria, we proposed a practical solution for on-chip detection of the SHNO signal, offering insights for future implementations of this nonlinear oscillator in complex networks. ## V Conclusions This paper presents an electrical empirical model of the SHNO working in the 6-GHz range, which is drawn from a comprehensive study of the SHNO output impedance and its output power and noise levels shown at the SHNO signal detector. From the results of the study, especially due to the high $Re(Z_{SHNO})$ values and the non-negligible capacitive effect, it is clear that there is a need to provide an empirical electrical model to the designer of the fully integrated detector, and thus make the discrimination of the SHNO operating state feasible. ## Acknowledgement This work was supported in part by the Horizon 2020 Research and Innovation Program No. 899559 “SpinAge”, DOI 10.3030/899559. ## References * Torrejon _et al._ (2017) J. Torrejon, M. Riou, F. A. Araujo, S. Tsunegi, G. Khalsa, D. Querlioz, P. Bortolotti, V. Cros, K. Yakushiji, A. Fukushima, H. Kubota, S. Yuasa, M. D. Stiles, and J. Grollier, Nature 547, 428 (2017). * Romera _et al._ (2018) M. Romera, P. Talatchian, S. Tsunegi, F. A. Araujo, V. Cros, P. Bortolotti, J. Trastoy, K. Yakushiji, A. Fukushima, H. Kubota, S. Yuasa, M. Ernoult, D. Vodenicarevic, T. Hirtzlin, N. Locatelli, D. Querlioz, , and J. Grollier, Nature 563, 230 (2018). * Zahedinejad _et al._ (2020) M. Zahedinejad, A. A. Awad, S. Muralidhar, R. Khymyn, H. Fulara, H. Mazraati, M. Dvornik, and J. Åkerman, Nat. Nano. 15, 47 (2020). * Houshang _et al._ (2022) A. Houshang, M. Zahedinejad, S. Muralidhar, J. Chȩciński, R. Khymyn, M. Rajabali, H. Fulara, A. A. Awad, M. Dvornik, and J. Åkerman, Phys. Rev. Appl. 17, 014003 (2022). * Kumar _et al._ (2023a) A. Kumar, A. Litvinenko, N. Behera, A. A. Awad, R. Khymyn, and J. Åkerman, arXiv preprint arXiv:2312.09656 (2023a). * Ross _et al._ (2023) A. Ross, N. Leroux, A. De Riz, D. Marković, D. Sanz-Hernández, J. Trastoy, P. Bortolotti, D. Querlioz, L. Martins, L. Benetti, _et al._ , Nature Nanotechnology 18, 1273 (2023). * Litvinenko _et al._ (2022) A. Litvinenko, A. Sidi El Valli, V. Iurchuk, S. Louis, V. Tyberkevych, B. Dieny, A. N. Slavin, and U. Ebels, Nano Lett. 22, 1874 (2022). * Müller _et al._ (2022) E. Müller, E. Arnold, O. Breitwieser, M. Czierlinski, A. Emmel, J. Kaiser, C. Mauch, S. Schmitt, P. Spilger, R. Stock, Y. Stradmann, J. Weis, A. Baumbach, S. Billaudelle, B. Cramer, F. Ebert, J. Göltz, J. Ilmberger, V. Karasenko, M. Kleider, A. Leibfried, C. Pehle, and J. Schemmel, Frontiers in Neuroscience 16 (2022), 10.3389/fnins.2022.884128. * González _et al._ (2024) V. H. González, A. Litvinenko, A. Kumar, R. Khymyn, and J. Åkerman, arXiv preprint arXiv:2403.13564 (2024). * Demidov _et al._ (2014) V. Demidov, S. Urazhdin, A. Zholud, A. Sadovnikov, and S. Demokritov, Appl. Phys. Lett. 105, 172410 (2014). * Chen _et al._ (2016) T. Chen, R. K. Dumas, A. Eklund, P. K. Muduli, A. Houshang, A. A. Awad, P. Dürrenfeld, B. G. Malm, A. Rusu, and J. Åkerman, Proceedings of the IEEE 104, 1919 (2016). * Behera _et al._ (2024) N. Behera, A. K. Chaurasiya, V. H. González, A. Litvinenko, L. Bainsla, A. Kumar, R. Khymyn, A. A. Awad, H. Fulara, and J. Åkerman, Advanced Materials 36, 2305002 (2024). * Zahedinejad _et al._ (2018) M. Zahedinejad, H. Mazraati, H. Fulara, J. Yue, S. Jiang, A. A. Awad, and J. Åkerman, Appl. Phys. Lett. 112, 132404 (2018). * Rajabali _et al._ (2023) M. Rajabali, R. Ovcharov, R. Khymyn, H. Fulara, A. Kumar, A. Litvinenko, M. Zahedinejad, A. Houshang, A. A. Awad, and J. Åkerman, Phys. Rev. Appl. 19, 034070 (2023). * Kumar _et al._ (2023b) A. Kumar, H. Fulara, R. Khymyn, A. Litvinenko, M. Zahedinejad, M. Rajabali, X. Zhao, N. Behera, A. Houshang, A. A. Awad, and J. Åkerman, Nano Lett. 23, 6720 (2023b). * Muralidhar _et al._ (2022) S. Muralidhar, A. Houshang, A. Alemán, R. Khymyn, A. A. Awad, and J. Åkerman, Appl. Phys. Lett. 120, 262401 (2022). * Khademi _et al._ (2024) M. Khademi, A. Kumar, M. Rajabali, S. P. Dash, and J. Åkerman, IEEE Electron Device Lett. 45, 268 (2024). * Xie, Cheung, and Fuchs (2023) Y. Xie, H. F. H. Cheung, and G. D. Fuchs, arXiv preprint arXiv:2303.02478 (2023), https://doi.org/10.48550/arXiv.2303.02478. * Sethi _et al._ (2023) P. Sethi, D. Sanz-Hernández, F. Godel, S. Krishnia, F. Ajejas, A. Mizrahi, V. Cros, D. Marković, and J. Grollier, Physical Review Applied 19, 064018 (2023). * Kumar _et al._ (2024) A. Kumar, A. K. Chaurasiya, V. H. González, N. Behera, R. Khymyn, A. A. Awad, and J. Åkerman, arXiv:2402.00586 (2024). * Litvinenko _et al._ (2023) A. Litvinenko, A. Kumar, M. Rajabali, A. A. Awad, R. Khymyn, and J. Åkerman, Appl. Phys. Lett. 122 (2023). * Bendjeddou _et al._ (2023) I. Bendjeddou, M. J. Garcia, A. S. el Valli, A. Litvinenko, V. Cros, U. Ebels, A. Jenkins, R. Ferreira, R. Dutra, D. Morche, E. Pistono, S. Bourdel, Y. Le Guennec, and F. Podevin, IEEE Transactions on Microwave Theory and Techniques 71, 2771 (2023). * Fiorelli _et al._ (2023) R. Fiorelli, E. Peralías, R. Méndez-Romero, M. Rajabali, A. Kumar, M. Zahedinejad, J. Åkerman, F. Moradi, T. Serrano-Gotarredona, and B. Linares-Barranco, Electronics 12, 230 (2023). * Fulara _et al._ (2019) H. Fulara, M. Zahedinejad, R. Khymyn, A. Awad, S. Muralidhar, M. Dvornik, and J. Åkerman, Science advances 5, eaax8467 (2019). * Pozar (2011) D. M. Pozar, _Microwave Engineering_ (Wiley, USA, 2011). * Kumar _et al._ (2022) A. Kumar, M. Rajabali, V. H. González, M. Zahedinejad, A. Houshang, and J. Åkerman, Nanoscale 14, 1432 (2022). * Alexander and Sadiku (2016) C. Alexander and M. Sadiku, _Fundamentals of electric circuits, 6th Edition_ (McGraw Hill, 2016).
††thanks: These authors contributed equally to this work††thanks: These authors contributed equally to this work # Contactless Interfacial Rheology: Probing Shear at Liquid-Liquid Interfaces without an Interfacial Geometry via Fluorescence Microscopy Iain Muntz SUPA School of Physics and Astronomy, The University of Edinburgh, Edinburgh, EH9 3FD, Scotland, United Kingdom Department of Bionanoscience, Kavli Institute of Nanoscience Delft, Delft University of Technology, Van der Maasweg 9, 2629 HZ Delft, The Netherlands James A. Richards SUPA School of Physics and Astronomy, The University of Edinburgh, Edinburgh, EH9 3FD, Scotland, United Kingdom Edinburgh Complex Fluids Partnership, The University of Edinburgh, Edinburgh EH9 3FD, United Kingdom Sam Brown SUPA School of Physics and Astronomy, The University of Edinburgh, Edinburgh, EH9 3FD, Scotland, United Kingdom Andrew B. Schofield SUPA School of Physics and Astronomy, The University of Edinburgh, Edinburgh, EH9 3FD, Scotland, United Kingdom Marcel Rey SUPA School of Physics and Astronomy, The University of Edinburgh, Edinburgh, EH9 3FD, Scotland, United Kingdom Department of Physics, University of Gothenburg, Gothenburg, Sweden Job H. J. Thijssen<EMAIL_ADDRESS>SUPA School of Physics and Astronomy, The University of Edinburgh, Edinburgh, EH9 3FD, Scotland, United Kingdom ###### Abstract Interfacial rheology is important for understanding properties such as Pickering emulsion or foam stability. Currently, the response is measured using a probe directly attached to the interface. This can both disturb the interface and is coupled to flow in the bulk phase, limiting its sensitivity. We have developed a contactless interfacial method to perform interfacial shear rheology on liquid/liquid interfaces with no tool attached directly to the interface. This is achieved by shearing one of the liquid phases and measuring the interfacial response via confocal microscopy. Using this method we have measured steady shear material parameters such as interfacial elastic moduli for interfaces with solid-like behaviour and interfacial viscosities for fluid-like interfaces. The accuracy of this method has been verified relative to a double-wall ring geometry. Moreover, using our contactless method we are able to measure lower interfacial viscosities than those that have previously been reported using a double-wall ring geometry. A further advantage is the simultaneous combination of macroscopic rheological analysis with microscopic structural analysis. Our analysis directly visualizes how the interfacial response is strongly correlated to the particle surface coverage and their interfacial assembly. Furthermore, we capture the evolution and irreversible changes in the particle assembly that correspond with the rheological response to steady shear. ## I Introduction Interfacial rheometry is essential when characterising systems with large interfacial area, such as emulsions or foams [1, 2, 3]. These systems are ubiquitous in industries such as pharmaceuticals, cosmetics and foodstuffs [4, 5, 6, 7, 8, 9, 10]. In order to probe the rheological properties, one can use shear rheology [11, 12, 13, 14, 15], dilational rheology [16, 17, 18, 19], or simultaneously image the interface as shear is applied to connect the rheological properties to the interfacial microstructure [20, 21]. Previous work on interfacial shear rheology has used probes which directly attach to an oil-water or air-water interface, such as the magnetic rod interfacial stress rheometer [22, 23, 24], or the double-wall ring (DWR) geometry attached to a rotational rheometer [11, 20]. These experimental setups are both based on the maximisation of the ratio of the surface force to the sub-phase drag, the Boussinesq number [1, 25]: $\mathrm{Bo}=\frac{\eta^{s}}{\eta l},$ (1) where $\eta^{s}$ is the surface viscosity, $\eta$ is the sub-phase viscosity and $l$ is a characteristic length scale roughly equal to the ratio of contact area to contact perimeter. In order to accurately measure the surface properties without unintentionally probing the sub-phase, this ratio must be maximised for the surface to contribute at least an order of magnitude more than the bulk. Of the two setups mentioned, the magnetic rod has the larger Bo, while both have a Bo an order of magnitude larger than that of a rotating disk, due to a much smaller $l$ [26]. This maximisation of Bo can be considered as optimising the interface-to-bulk signal to noise ratio. Even though the magnetic rod set up has higher sensitivity, DWR has the advantages of using a conventional rotational rheometer combined with a larger dynamic range [11]. In our work, we take a different approach which makes consideration of Bo less tangible as we have no contact area or contact perimeter. Rather than affixing a probe directly to the interface, we shear the upper phase, indirectly deforming the interface, and measure the response using confocal microscopy: a fundamentally different approach. Using this contactless technique, we investigate the efficacy of this method by studying a jammed core-shell PNIPAM-SiO2–laden interface labelled with tracer particles, which we compare directly to DWR measurements. We then demonstrate the advantages of this technique by looking at a weakly interacting system of interfacially adsorbed colloidal particles. This system has been studied previously using direct probe techniques [14, 27], and considering the interparticle interactions [28]. Our technique has two main advantages: (i) the liquid-liquid interface we probe is not disturbed by a large probe immersed therein, and (ii) this setup models general applications of these large interfacial area systems, where interfacial shear is applied indirectly via the continuous phase. A clear example of this second point is in the application of skin creams, where the continuous phase is sheared, which indirectly deforms the large area of interface of the dispersed phase. Notably, the equipment required to perform these measurements is relatively common. While we use confocal microscopy coupled to a stress-controlled rheometer, reflection or fluorescence microscopy and a fixed-rate motor should suffice. We show that our technique can measure surfaces with lower viscosities than have been measured before using a DWR geometry, due to the inherent sensitivity of the technique arising from the absence of direct sub-phase drag. Finally, our setup lends itself to simultaneous structural analysis, which we show is key to understanding the rheological properties of a particle-laden interface. ## II Materials and Methods ### II.1 Materials All chemicals were obtained from commercial sources and used as received if not stated otherwise. N,N’-Methylenebis(acrylamide) (BIS; 99 $\%$, Sigma Aldrich), ethanol (EtOH, 99.9 $\%$, Sigma Aldrich), ammonium persulfate (APS; 98 $\%$ Sigma Aldrich), tetraethyl orthosilicate (TEOS; 98 $\%$, Sigma Aldrich), ammonium hydroxide solution (28-30 $\%$ NH3 basis, Sigma Aldrich), (3-(trimethoxysilyl)propyl methacrylate (MPS; 98 $\%$, Sigma Aldrich) and isopropyl alcohol (IPA, $>99.8$ $\%$, Sigma Aldrich), were used as received. N-Isopropylacrylamide (NIPAM; 97 $\%$, Sigma Aldrich) was purified by recrystallization from hexane (95 $\%$, Sigma Aldrich). Water was distilled and deionized ($18\text{\,}\mathrm{M\SIUnitSymbolOhm}\text{\,}\mathrm{cm}$) and n-dodecane (Acros organics, 99% pure) was filtered three times through a column of alumina (Sigma-Aldrich, activated) to remove polar impurities following a standard procedure [29]. Red fluorescent carboxyl-functionalized polystyrene (PS) particles ($2\text{\,}\mathrm{\SIUnitSymbolMicro m}$ diameter, Thermo Fisher) were cleaned twice via centrifugation and redispersion in water/ethanol (1:1). ### II.2 Synthesis and Characterisation #### II.2.1 PNIPAM-SiO2 core-shell particles Poly(N-isopropylacrylamide)(PNIPAM)-SiO2 core-shell particles were obtained by growing a PNIPAM shell onto the silica cores via a batch surfactant-free precipitation polymerization as described in previous work [30]. First, colloidal silica particles used as cores with a diameter of $160(10)\text{\,}\mathrm{nm}$ were prepared according to a modified Stöber process[31]. In a round bottom flask, 250 mL EtOH , 12.5 mL deionised water and 25 mL NH3 (aq) were stirred together. 18.75 mL of TEOS was stirred in 75 mL EtOH and both solutions were heated to $50\text{\,}\mathrm{\SIUnitSymbolCelsius}$ and equilibrated for $30\text{\,}\mathrm{min}$. Next, the TEOS solution was quickly added to the first mixture under heavy stirring. We let the reaction proceed for $2\text{\,}\mathrm{d}$ at $50\text{\,}\mathrm{\SIUnitSymbolCelsius}$. The suspension was functionalised without any further purification by adding $102.7\text{\,}\mathrm{\SIUnitSymbolMicro l}$ MPS. We allowed the reaction mixture to stir at room temperature for at least $1\text{\,}\mathrm{d}$ and then boiled it for $1\text{\,}\mathrm{h}$ to ensure successful functionalisation. Afterwards, we purified the particles by centrifugation and redispersed them three times in ethanol and three times in Milli-Q water. In a 500 mL three-neck round bottom flask, 282.9 mg NIPAM and 19.3 mg BIS ($5\,{\rm mol.\,\%}$) were dissolved in 47 mL Milli-Q water. We added the 2.591 g aqueous SiO2 core dispersion (6.6 ${\rm wt}\,\%$). The solution was heated to $80\text{\,}\mathrm{\SIUnitSymbolCelsius}$ and purged with nitrogen. After equilibration for $30\text{\,}\mathrm{min}$, a balloon filled with nitrogen was used to keep the nitrogen atmosphere. Subsequently, 11 mg APS was rapidly added to initiate the reaction. We let the reaction proceed for 4 h, and after it cooled down, we purified the suspension 6$\times$ by centrifugation and redispersion in deionised water. The hydrodynamic diameter at $20\text{\,}\mathrm{\SIUnitSymbolCelsius}$ was determined by dynamic light scattering (Malvern Zetasizer Nano-ZS) to $525(53)\text{\,}\mathrm{nm}$. #### II.2.2 PMMA particles Poly(methyl methacrylate) (PMMA) particles, stabilized by poly(lauryl methacrylate), were used as the hydrophobic system. To synthesize these, the poly(lauryl methacrylate) stabilizer was fabricated first following the recipe in Ref. 32, Sec. 3.9.1, and it was kept as a 40% solution in dodecane. To make the particles, a mixture was created that contained 2.1% w/w poly(lauryl methacrylate) stabilizer, 41.2% w/w methyl methacrylate, 0.84% methacrylic acid, 11% butyl acetate, 29.6% hexane, 14.2% dodecane, 0.21% octyl mercaptan and 0.47% of the dye NBD-MAA (7-nitrobenzo-2-oxa-1,3-diazole-methyl methacrylate), whose preparation can be found in Ref. 33. This mixture was placed in a 3-necked round-bottomed flask with a condenser attached, brought under a nitrogen atmosphere, stirred at 350 rpm and heated to 80°C before 0.4% w/w of the initiator azo-bis-isobutyronitrile was added to start the polymerization reaction which was left to proceed for 6 hours. The resultant particles were filtered through glass wool to remove any coagulum present. The particles were qualitatively inspected using scanning electron microscopy, and sized by static light scattering to find a diameter of $3.0\text{\,}\mathrm{\SIUnitSymbolMicro m}$ with a dispersity of 5%. The particles were cleaned by repeated centrifugation (5$\times$) in $n$-hexane followed by repeated centrifugation (5$\times$) in $n$-dodecane. The particles were kept as a dispersion in $n$-dodecane and sonicated for 30 minutes before dilution, followed by a further 30 minutes of sonication before use to minimise the number of aggregates in bulk. ### II.3 Contactless methods Figure 1: Interfacial shear geometries. (a) Schematic setup for contactless interfacial rheology using a parallel-plate geometry rotated at fixed angular velocity, $\omega$. Interface imaging via confocal microscope. Dimensions: pinned interface ring radius, $r_{r}=10$ mm; water sub-phase height, $h_{w}=3$ mm; and oil depth, $h_{o}$. $\omega$ and $h_{o}$ vary. $r_{\rm out}$ distance from imaging region to outer edge varies with experiment from $4\text{\,}\mathrm{mm}6\text{\,}\mathrm{mm}$. (b) Creep-recovery protocol for contactless method. Recording (dashed lines) begins before applied $\omega$ (solid line) to $t_{\omega}=$30\text{\,}\mathrm{s}60\text{\,}\mathrm{s}$$, followed by recovery to recording end. Multiple steps with increasing $\omega$. (c) Schematic of double-wall ring geometry. Radii: ring, inner and outer, $R_{r,i}=$35.5\text{\,}\mathrm{mm}$$ and $R_{r,o}=$34.5\text{\,}\mathrm{mm}$$; trough, $R_{i}=$31\text{\,}\mathrm{mm}$$ and $R_{o}=$39.5\text{\,}\mathrm{mm}$$. Ring width, $l=R_{r,o}-R_{r,i}=$1\text{\,}\mathrm{mm}$$ for Bo, Eq. (1). Torque, $T(t)$, and angle, $\theta(t)$, to calculate interfacial stress, $\sigma^{s}$, and strain, $\gamma^{s}_{0}$. (d) Increasing logarithmic oscillatory strain amplitude sweep. Interfaces were prepared in a custom made polytetrafluoroethylene (PTFE) cup with an aluminium ring insert to pin a flat interface. Fig. 1(a) shows a schematic representation of the cup. The PTFE cup’s inner radius ($r_{c}$) is $21\text{\,}\mathrm{m}\mathrm{m}$, the aluminium ring’s inner radius ($r_{r}$) is $10\text{\,}\mathrm{m}\mathrm{m}$. The aluminium ring has a height of $3\text{\,}\mathrm{m}\mathrm{m}$ and its edge was roughened using silicon carbide sandpaper to allow pinning of the interface. The PTFE cup was filled with water, pinned at the edge of the aluminium ring. As the first interface, we choose a monolayer of PNIPAM-SiO2 core-shell particles mixed with fluorescent tracer particles at a fixed surface pressure of $24\text{\,}\mathrm{mN}\text{\,}{\mathrm{m}}^{-1}$. We first created a suspension by mixing $800\text{\,}\mathrm{\SIUnitSymbolMicro l}$ core-shell particles (0.1 ${\rm wt}\,\%$) and $100\text{\,}\mathrm{\SIUnitSymbolMicro l}$ fluorescent PS microspheres (1 ${\rm wt}\,\%$) with $100\text{\,}\mathrm{\SIUnitSymbolMicro l}$ IPA as a spreading agent. The PNIPAM-SiO2 core-shell particles are smaller and able to spread and extend once adsorbed to liquid interfaces [34, 35]. Therefore, they occupy more interfacial area compared to the PS particles. Further, we should point out that PNIPAM-based microgels are known to adsorb onto PS particles [36] and accumulate around them when confined at liquid interfaces [37]. Thus, we expect the PS particles to be integrated within the PNIPAM-SiO2 interface and only minimally influence the rheological response. We then spread the mixed suspension on a Langmuir trough and measure the surface pressure using the Wilhelmy method. We determined that $3.72\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{l}\mathrm{/}\mathrm{c}\mathrm{m}^{2}$ of prepared suspension is required to obtain a surface pressure of $24\text{\,}\mathrm{mN}\text{\,}{\mathrm{m}}^{-1}$. Thus, $11.7\text{\,}\mathrm{\SIUnitSymbolMicro l}$ was pipetted onto the air-water interface pinned by the aluminum ring. Notably, in a control experiment, we could directly verify the surface pressure of the interface within the cell using the Wilhelmy method to be $24\text{\,}\mathrm{mN}\text{\,}{\mathrm{m}}^{-1}$. Lastly, $1.5\text{\,}\mathrm{ml}$ of dodecane was carefully pipetted on top of the interface to give a depth of the oil phase ($h_{o}$) of $1.6\text{\,}\mathrm{mm}$. As a second interface, we choose hydrophobic PMMA colloidal particles. The aluminum ring was again filled with water and $3\text{\,}\mathrm{ml}$ of a dilute 0.005 ${\rm vol.}\,\%$ PMMA-in-oil dispersion was pipetted onto the water sub-phase ($h_{o}=$2.2\text{\,}\mathrm{mm}$$). A 0.005 ${\rm vol.}\,\%$ dispersion generally leads to a low volume fraction interface, however there is large variability in the final surface fraction for the same initial volume fraction. To achieve higher surface fractions a higher initial volume fraction was used. During equilibration for 1 hour, the PMMA particles sedimented to the oil-water interface and formed a monolayer. The surface fraction, $\phi$, was adjusted by either adjusting the amount of deposited oil dispersion or by adjusting the particle concentration. Surface fractions were measured from microscopy images using a pixel counting method determining the fraction of foreground (particle) pixels to total pixels after performing a thresholding procedure. This measurement was made over multiple frames and the final value for surface fraction was determined as the mean through one rotation of the interface. We note that this approach likely overestimates the actual $\phi$, see Appendix A. However, this simple approach allows systematic comparison between different surface coverages. Connected pixel clusters can then be identified to assess interface homogeneity, or aggregation state, via the dispersity [14], $\mathcal{D}=\frac{s}{\langle A\rangle}\,,$ (2) with $\langle A\rangle$ the average cluster size and $s$ the standard deviation. These dispersities were measured from relatively zoomed out images. A 25 mm diameter parallel-plate geometry was attached to the oil-air interface in the centre of the PTFE cup. Fixed rotation speeds, $\omega$, were then applied (MCR 301, Anton Paar), shearing the upper oil phase. Using a rheoimaging setup, as described by Besseling _et al._ [38] (although our rheometer setup lies directly on top of the confocal, as described by Dutta _et al._ [39], providing greater stability), rheometry was conducted while the interface was simultaneously imaged using a Leica SP8 confocal microscope with a 10$\times$ / 0.3 NA air-immersion objective, at $1024\\!\times\\!1024$ px2 (PNIPAM-SiO2) or $512\\!\times\\!512$ px2 (PMMA) ($$932\text{\,}\mathrm{\SIUnitSymbolMicro m}$\\!\times\\!$932\text{\,}\mathrm{\SIUnitSymbolMicro m}$$). The imaging setup was such that the motion of the interface under shear was horizontally oriented. Velocimetry of the confocal images was performed using C code written in house. This splits the images into 10 equally spaced horizontal bands, i.e. the top 10% of the image, the second 10% of the image, etc. Each band is offset horizontally by a well defined distance. The Pearson correlation coefficient of this new offset band with the same band in the previous frame is calculated [40]. The distance moved between that frame and the previous is then the horizontal offset which maximises this correlation, over time this gives the net displacement. Note that, as we use fluorescence microscopy, and only the particles are fluorescently labelled, the strain we measure is the strain of the (interfacial) colloidal particles in the field of view of the microscope. The interfacial strain is calculated as $x/r_{\rm out}$ for each band and averaged, where $x$ is the measured displacement of the interface and $r_{\rm out}$ is the distance from the measurement to the outer, pinned wall; $r_{\rm out}$ varies for each experiment and is measured in situ, it is always approximately $6\text{\,}\mathrm{mm}$. To probe the yielding and flow of the interfaces, a creep-recovery protocol is used [41]. Fixed rotation rates were set for $t_{\omega}=$60\text{\,}\mathrm{s}$$ (PNIPAM-SiO2) or $120\text{\,}\mathrm{s}$ (PMMA), applying stress to the interface, before a further period of fixing the rotation of the rheometer to 0, $60\text{\,}\mathrm{s}$ (PNIPAM-SiO2) or $30\text{\,}\mathrm{s}$ (PMMA), allowing the interface to relax, Fig. 1(b). The two steps allow separate measurement of the elastic response and the plastic, or irrecoverable, response. The (elastic) recoverable strain ($\gamma^{s}_{\rm rec}$), is given by the recoil from the peak strain at the end of the applied rotation to the end of the recovery step. The (plastic) irrecoverable shear rate, $\dot{\gamma}^{s}$, can be calculated from the total change in strain from the start of applied rotation to after recovery over the time of the applied shear. Alternatively, for faster relaxing interfaces (PMMA) $\dot{\gamma}^{s}$ can be calculated from the average shear rate over a $45\text{\,}\mathrm{s}$ window towards the end of the applied rotation. This sequence is repeated at multiple increasing rotation rates to determine the stress-dependent response of the interfaces. ### II.4 Double-wall ring geometry Conventional interfacial shear rheometry was performed using a double-wall ring geometry, Fig. 1(c), connected to a stress-controlled rotational rheometer (TA Instruments, DHR-2). This consists of a Platinum/Iridium ring (diamond cross-section with inner/outer radius $R_{r,i/o}=34.5/$35.5\text{\,}\mathrm{mm}$$ and hence width $l=$1\text{\,}\mathrm{mm}$$) inside a ring-shaped polyoxymethylene trough (inner/outer radius $R_{i/o}=31/$39.5\text{\,}\mathrm{mm}$$). All surfaces were cleaned multiple times with ethanol and deionised water. To form an interface the trough is first filled with the water sub-phase until level and pinned at the edges of the trough. $70\text{\,}\mathrm{\SIUnitSymbolMicro l}$ of the mixed PNIPAM-SiO2 and PS tracer particle dispersion in a spreading solvent were then carefully pipetted onto the air-water interface. The ring is then lowered until pinned and level at the interface, before dodecane is pipetted on top to cover the ring. The ring is therefore in direct contact with both the interface and the sub-phase. Figure 2: Thin ring element of oil-water interface, radius $r$ and width d$r$. The interfacial stress, $\sigma^{s}$, can be found by considering the torque balance of bulk oil flow drag and interfacial stress gradient, d$\sigma^{s}$, with rotation rates (interface, $\omega_{i}$, and top of oil phase, $\omega$), Sec. II.5. Oscillatory strain amplitude sweeps were performed using controlled strain at $f=$0.2\text{\,}\mathrm{Hz}$$ in the low-frequency response region, following previous protocols for microgel-laden interfaces [16] with one equilibration cycle and six measurement cycles per point, Fig. 1(d). Strain was increased logarithmically at 20 points/decade from 0.001 to 1.0 strain amplitude and we report the strain-dependent elastic ($G^{s\prime}$) and loss ($G^{s\prime\prime}$) moduli from the primary Fourier components. The sinusoidal oscillation of the ring, $\theta(t)=\theta_{0}\sin(2\pi ft)$, is converted into an interfacial strain, $\gamma^{s}_{0}=\theta_{0}[(1-(R_{r,o}/R_{o})^{2})^{-1}+((R_{r,i}/R_{i})^{2}-1)^{-1}]$ using the 2D-equivalent expressions for a Couette cylinder at the position of the ring [42]. The strain can be approximated as $\gamma^{s}_{0}\approx\theta R_{r,o}/(R_{o}-R_{r,o})$ 111This can be recovered by factorising the denominator, $\gamma^{s}_{0}=\theta_{0}[R_{o}^{2}/[(R_{o}-R_{r,o})(R_{o}+R_{r,o})]+R_{i}^{2}/[(R_{r,i}-R_{i})(R_{r,i}+R_{i})]$, and approximating to first order by taking $R_{i/o}+R_{r,i/o}\approx 2R_{r,i/o}$, $R_{r,i}\approx R_{r,o}$ and $R_{o}-R_{r,o}\approx R_{r,i}-R_{i}$., i.e. the displacement of the ring divided by the distance from the ring to the outer pinned interface, analogous with the expression for the contactless geometry. As the ring is in direct contact with the interface, the measured torque can be converted straight to an interfacial stress, $\sigma^{s}=T/2\pi(R_{r,i}^{2}+R_{r,o}^{2})$. All reported data is at high Bo, such that sub-phase drag correction is not applied, and low raw-phase angle, where geometry inertia does not dominate. Due to noise from the strain-control feedback loop, torque resolution is limited to $0.07\text{\,}\mathrm{\SIUnitSymbolMicro N}\text{\,}\mathrm{m}$[44] and we highlight or truncate data below this threshold. ### II.5 Measuring stress in a contactless geometry Figure 3: Contactless interfacial rheometry of PNIPAM-SiO2–laden interfaces. (a) Creep-recovery protocol, applied rotation rate, $\omega=6.31$ rpm, with time, $t$. Stress, $\sigma^{s}=$3.2\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, for $\omega_{i}\approx 0$, Eq. (9). Highlighted $t$: (i) dashed (blue), start of applied $\omega$ at $t=$0\text{\,}\mathrm{s}$$; (ii) dot-dashed (grey), end of applied $\omega$ at $t_{\omega}=$60\text{\,}\mathrm{s}$$; and (iii) dotted (orange), end of recorded recovery at $t=$120\text{\,}\mathrm{s}$$. (b) Measured interface response at highlighted $t$. Zoomed-in confocal microscopy at $t$ in (a). Outline, distinctive particle cluster showing interface motion along $x$, axis rotated to match (c). (c) Extracted time-dependent strain, $\gamma^{s}(t)$. Horizontal lines trace strain from: (i) dashed, (ii) dot- dashed and (iii) dotted. Arrows: recoverable (elastic) strain, $\gamma^{s}_{\rm rec}=\gamma^{s}($60\text{\,}\mathrm{s}$)-\gamma^{s}($120\text{\,}\mathrm{s}$)$; and irrecoverable (plastic) strain, $\gamma^{s}_{\rm irr}=\gamma^{s}($120\text{\,}\mathrm{s}$)-\gamma^{s}($0\text{\,}\mathrm{s}$)$. In contrast to the DWR geometry, the interface is not in direct contact with the geometry in the contactless method. Therefore, the stress cannot be directly measured by the rheometer. To make meaningful statements on the rheological properties of the interface, we must first find the interfacial stress applied at the interface. It is well known that for a parallel-plate setup, the stress is independent of the height through the sample upon reaching a steady state [42]. The timescale for momentum diffusion and to reach steady state is up to $\approx 0.3h_{o}^{2}\rho_{o}/\eta_{o}=\mathcal{O}($1\text{\,}\mathrm{s}$)$ for our setup [45], where $\rho_{o}$ is the oil density; this is far below the creep step duration. Considering the upper phase as a Newtonian fluid allows us to, therefore, find the stress on the interface from the upper fluid using the applied rotational speed of the geometry. If the rheometer is rotated at a fixed angular velocity of $\omega$ then the shear rate of the upper fluid at a radius $r$ (Fig. 2) of the parallel-plate geometry is given by $\dot{\gamma}=\frac{\omega r}{h_{o}},$ (3) where $h_{o}$ is the oil phase depth. This definition relies on the interface having zero speed. Therefore, this is a first approximation if the speed of the geometry is much larger than the speed of the interface. If this is not the case, then we can reduce $\omega$ by the angular speed of the interface at $r$, $\omega_{i}$, giving, $\dot{\gamma}=\frac{(\omega-\omega_{i})r}{h_{o}}.$ (4) The stress induced is given by the product of $\dot{\gamma}$ with the upper phase bulk viscosity, $\eta_{o}$, i.e., $\sigma=\frac{\eta_{o}(\omega-\omega_{i})r}{h_{o}}.$ (5) To convert this bulk stress into an interfacial stress we consider the torque applied from the bulk on an area element of a ring in the interface at $r$ of width ${\rm d}r$, Fig. 2. We can write the torque element, d$T$, as a product of the force element, $\sigma{\rm d}A$, and the radius, where d$A=2\pi r{\rm d}r$ is the area of the infinitesimal ring, ${\rm d}T=\sigma 2\pi r^{2}{\rm d}r=2\pi\frac{\eta_{o}(\omega-\omega_{i})}{h_{o}}r^{3}{\rm d}r.$ (6) This torque then gives rise to the interfacial stress, $\sigma^{s}$. With $T(r)=2\pi r^{2}\sigma^{s}(r)$ the interfacial torque as the product of $\sigma^{s}$, radius and perimeter, we can write the torque balance $\begin{split}{\rm d}T=&T(r+{\rm d}r)-T(r)\\\ \approx&2\pi\left[(r+{\rm d}r)^{2}\sigma^{s}(r+{\rm d}r)-r^{2}\sigma^{s}(r)\right]\\\ \approx&2\pi r\left[r{\rm d}\sigma^{s}+2\sigma^{s}{\rm d}r\right]\end{split}$ (7) where second order differential terms, e.g., $({\rm d}r)^{2}$, have been dropped. Equating Eqs (6) and (7) and rearranging, $\frac{{\rm d}\sigma^{s}}{{\rm d}r}+\frac{2\sigma^{s}}{r}=\frac{\eta_{o}(\omega-\omega_{i})}{h_{o}}r.$ (8) Figure 4: Strain response, $\gamma^{s}(t)$, of PNIPAM-SiO2–laden interface for contactless rheology protocol, Fig. 3(a), showing transition from elastic dominated response to plastic flow with increasing rotation rate, $\omega$. (a) Low $\omega=$1\text{\,}\mathrm{r}\mathrm{p}\mathrm{m}8\text{\,}\mathrm{r}\mathrm{p}\mathrm{m}$$ (dark to light), see legend, with dominant elastic response, $\gamma^{s}_{\rm rec}$, after $t_{\omega}=$60\text{\,}\mathrm{s}$$. Shading, error from standard deviation in image correlation analysis bands. (b) Moderate $\omega$ around yielding with rising plastic response. (c) Strain response at high $\omega$ dominated by irreversible plastic flow, $\dot{\gamma}^{s}$ (see text for details). Equation (8) can be readily solved to give $\sigma^{s}=\frac{\eta_{o}(\omega-\omega_{i})}{4h_{o}}r^{2}$ (9) if $\frac{{\rm d}\omega_{i}}{{\rm d}r}\simeq 0$, so there is no interfacial shear banding, for example; this aligns with our confocal-microscopy observations, e.g., $\omega_{i}$ is constant in time and across the field of view (within error). It is evident this expression yields the correct dimensions for interfacial stress as Pa m, as well as physically reasonable dependencies on viscosity, applied rotation, oil phase height, and radius. In order to vary the interfacial stress, we vary $\omega$ while observing at a fixed $r$, a practically simpler approach. Finally, we need to ensure that we are measuring surface rather than bulk sub- phase properties. Typically, this can be checked via Bo [Eq. (1)], but this assumes a probe in contact with the interface. For our contactless setup, as we measure interfacial strain via the motion of the interface itself, we focus instead on the question: is the force on the _interface_ dominated by surface or bulk sub-phase viscosity? This effect is included in sub-phase drag corrections [46], but is typically neglected in Bo as drag on the probe dominates. For our contactless geometry, the stress from the sub-phase can be crudely approximated as parallel-plate flow with $\omega_{i}$, giving an equivalent to Bo, $\mathrm{Bo}^{*}=\frac{\sigma^{s}}{\sigma^{s}_{\rm drag}}=\frac{\eta_{o}h_{w}}{\eta_{w}h_{o}}\frac{\omega-\omega_{i}}{\omega_{i}}\,\mathrm{for}~{}\sigma^{s}_{\rm drag}\approx\frac{\eta_{w}\omega_{i}}{4h_{w}}r^{2},$ (10) where $\eta_{w}$ and $h_{w}$ are the water sub-phase viscosity and depth, Fig. 1(a). The first term in Bo∗ is $\mathcal{O}(1)$ for our bulk phases and dimensions, and likely most common uses, but the second term can be arbitrarily large as $\omega_{i}\rightarrow 0$. This makes the technique suitable for measuring weak interface yielding, a fact that arises from decoupling stress application and interface motion by not having a direct probe. For interfaces with significant $\omega_{i}$, Bo∗ can drop and this is discussed where relevant. ## III Results and Discussion ### III.1 Elastic PNIPAM-SiO2–laden interface To establish the validity of our novel contactless interfacial rheometric technique we begin by measuring a highly elastic PNIPAM-SiO2–laden interface, which can also be studied by conventional DWR interfacial rheometry. #### III.1.1 Contactless rheometry Using the contactless setup, we perform a creep-recovery test with increasing rotation rates, $\omega$, from 0.1 to 400 revolutions per minute (rpm) with logarithmic spacing at 5 pts/decade and higher resolution (20 pts/decade) where behaviour is observed to be changing from the confocal microscopy recordings. At each step, while recording, $\omega$ is applied for $60\text{\,}\mathrm{s}$ before zero rotation rate is set for a further $60\text{\,}\mathrm{s}$, Fig. 3(a). Imaging tracer particles embedded in the interface throughout these steps gives the resulting deformation of the interface. This is illustrated by following a distinctive cluster of particles in magnified images, Fig. 3(b). From the start of applied rotation (i) to when rotation is stopped (ii), the interface first moves along the flow direction, $x$. After cessation and relaxation the interface recoils backwards, along the previous flow direction, until the recovery step ends, Fig. 3(b)(ii) to (iii). Over time, this particle motion can be seen as tracing out the strain response of the interface, $\gamma^{s}(t)=x(t)/r_{\rm out}$, lines from (b)(i)–(iii) to (c). At low applied $\omega$, e.g., $6.31$ rpm (Fig. 3), there is an initial jump in $\gamma^{s}$ as $\omega$ is applied before a further slow increase over $t_{\omega}$ [the creep step length, Fig. 1(d)], see Fig. 3(c). At the cessation of applied flow there is a rapid recoil, followed by a slower further relaxation, approaching a constant value. From the strain profile, $\gamma^{s}(t)$, we extract two quantities: the recoverable strain as the decrease from the peak to the final strain [dot-dashed to dotted lines, (ii) to (iii)], and the irrecoverable strain, $\gamma^{s}_{\rm irr}$, as the increase from the initial strain, at the start of shear, to the final strain [dashed to dotted lines, (i) to (iii)]. A strong initial deformation and near complete elastic recovery can be seen over a range of $\omega\lesssim 8$ rpm, Fig. 4(a), with both steps growing with $\omega$ (dark to light). As $\omega$ increases further, up to $20$ rpm [Fig. 4(b) (dark to light)], $\gamma^{s}_{\rm rec}$ increases proportionally; $\gamma^{s}_{\rm irr}$ also begins to slowly increase as the interface does not fully recover. At higher $\omega$ still, $\gtrsim 25$ rpm [Fig. 4(c)] there is a clear change in behaviour, as $\gamma^{s}_{\rm irr}$ rapidly increases while the elastic recovery remains unchanged. During applied rotation, the strain is linear in time, giving a well-defined interfacial shear rate, $\dot{\gamma}^{s}=\gamma^{s}_{\rm irr}/t_{\omega}$. This behaviour is indicative of yielding in a creep-recovery test [41]. Using $\gamma^{s}_{\rm rec}$ and $\dot{\gamma}^{s}$ alongside the calculated interfacial stress ($\sigma^{s}$), Eq. (9), the rheological response can be quantified. The elastic response, $\sigma^{s}(\gamma^{s}_{\rm rec})$, shows three regimes, Fig. 5(a) [solid circles]. At the lowest stresses, $\sigma^{s}<${10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, no clear trend is observed. This (shaded) region lies below a minimum strain, $\gamma_{\rm rec}^{s,\min}\approx 2\times 10^{-3}$, set by noise, e.g., vibration or drift. With increasing stress, $\gamma^{s}_{\rm rec}$ increases linearly until $\sigma^{s}=$5\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$. Above this the elastic recovery appears constant, but noisy (i.e. spatially heterogeneous across the field of view). Within the linear region an elastic constant, $G^{s\prime}=\sigma^{s}/\gamma^{s}_{\rm rec}=$4.2(1)\text{\times}{10}^{-4}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$ can be fitted (dashed line). The interface is well described as a linear elastic solid below $\sigma^{s}=$5\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, this can be emphasised by plotting on linear axes, Fig. 5(b). However, this is only one side of the measured response. The stress-dependent plastic flow, $\sigma^{s}(\dot{\gamma}^{s})$, further illuminates the change around $\sigma^{s}\approx$5\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, Fig. 5(c) (squares). Below this threshold $\dot{\gamma}^{s}$ is limited, but on further increase the interface begins to flow $\dot{\gamma}^{s}\sim\mathcal{O}($0.1\text{\,}{\mathrm{s}}^{-1}$)$. The transition point, or yield stress ($\sigma^{s}_{y}$) can be quantified by fitting a simple piecewise function to these data. We model the interface as a Bingham fluid, one of the simplest models capturing suitable non-Newtonian behaviour, described by: $\begin{array}[]{lr}\dot{\gamma}^{s}=0&:\sigma^{s}<\sigma^{s}_{y}\\\ \sigma^{s}=\sigma^{s}_{y}+\eta^{s}\dot{\gamma}^{s}&:\sigma^{s}\geq\sigma^{s}_{y}.\end{array}$ (11) Below $\sigma^{s}_{y}$ there is no flow, above this the excess stress leads to a shear rate set by the interfacial viscosity, $\eta^{s}$. From this we can obtain both a yield stress and an interfacial viscosity from the contactless technique, $\sigma^{s,\rm Contactless}_{y}=$3.5(1)\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, and $\eta^{s}=$9.4(1)\text{\times}{10}^{-4}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}\text{\,}\mathrm{s}$$. While the Bingham model is appropriate to capture a clear yield stress transition, in general more complex interfacial yielding behaviour is often observed [47, 48]. Figure 5: PNIPAM-SiO2–laden interface rheology. (a) Elastic response, stress ($\sigma^{s}$) vs strain. Points: solid, contactless [$\sigma^{s}$ from Eq. (9) and strain, $\gamma^{s}_{\rm rec}$, from relaxation, Fig. 4]; open, DWR elastic stress, [$\sigma^{s}=\gamma^{s}_{0}G^{s\prime}$ for strain amplitude, $\gamma^{s}_{0}$]. Minimum limits: shading, contactless strain, $\gamma_{\rm rec}^{s,\min}\\!\approx\\!0.002$; vertical dotted line, DWR torque at $\gamma_{0}^{s,\min}$. Fit lines: dashed, contactless elastic response, $G^{s\prime}\\!=\\!\sigma^{s}/\gamma^{s}_{\rm rec}\\!=\\!$4.2(1)\text{\times}{10}^{-4}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$; horizontal dotted, yield stress from (c). Fit of linear elastic response for DWR in Fig. 6. (b) Linear plot of elastic response before yielding, symbols as in (a). (c) Contactless viscous response. Points, stress vs shear rate, $\dot{\gamma}^{s}=\gamma^{s}_{\rm irr}/t_{\omega}$ from irrecoverable strain, Fig. 4(c). Lines: bold dashed, Bingham plastic fit, Eq. 11; fine dashed, yield stress, $\sigma_{y}^{s,\rm Contactless}\\!=\\!$3.5\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$; and dotted, DWR yield stress, $\sigma_{y}^{s,\rm DWR}\\!=\\!$4.8\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$. #### III.1.2 Comparison to DWR rheometry Figure 6: DWR rheology of a PNIPAM-SiO2–laden interface. Elastic ($G^{s\prime}$, dark circles) and loss ($G^{s\prime\prime}$, light squares) moduli vs strain amplitude, $\gamma^{s}_{0}$. Shading, below instrument resolution. Lines: dotted, $\bar{G^{s\prime}}=$3.7(4)\text{\times}{10}^{-4}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, mean of $G^{s\prime}(\gamma^{s}_{0}\leq 0.05)$; dashed, yielding at $G^{s\prime}=G^{s\prime\prime}$, $\sigma^{s,\rm DWR}_{y}=\gamma^{s}_{0}\sqrt{(G^{s\prime})^{2}+(G^{s\prime\prime})^{2}}=$4.8\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$ As a strong and highly elastic interface we can directly compare the results of the contactless geometry to DWR interfacial rheology for the same PNIPAM- SiO2 at equal surface pressures. As the torque resolution for oscillatory tests is finer than steady shear for rotational rheometers, an oscillatory strain amplitude sweep is performed. This measures the strain-dependent elastic and loss moduli, Fig. 6 (symbols). At low $\gamma^{s}_{0}<0.1$, the elastic modulus is higher than the loss modulus, with $G^{s\prime}$ only weakly decreasing once above the torque resolution (shaded region), indicative of a solid elastic material as the stress in phase with the strain. With increasing $\gamma^{s}_{0}$, $G^{s\prime}$ begins to drop while $G^{s\prime\prime}$ remains near constant. At $\gamma^{s}_{0}=0.23$ the moduli become equal, $G^{s\prime}=G^{s\prime\prime}$; above this point $G^{s\prime}$ continues to sharply drop, while $G^{s\prime\prime}$ weakly decreases. In this region the stress is in phase with the shear rate, i.e. liquid-like. Therefore, with increasing strain amplitude the interface yields from a solid- like state to a liquid-like state where $G^{s\prime}=G^{s\prime\prime}$. This stress, $\sigma^{s,\rm DWR}_{y}=$4.8\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, is an operative yield stress at a finite frequency and shear rate, in contrast to the ‘static’ measurement of creep recovery [49]. The comparison of $\sigma_{y}^{s,\rm DWR}$ [Fig. 5(c) (dotted line)] with the contactless rheology value, $\sigma_{y}^{s,\rm Contactless}$ (fine dashed line), finds similar values, with only a 30% difference, 4.8 vs $3.5\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$. This is within the expected variation for different measurement protocols of a non- linear property 222Oscillatory yielding is ambiguous, as it can be gradual, with multiple definitions; $G^{s\prime}=G^{s\prime\prime}$ is an upper (over)estimate [49]. E.g., a tangent analysis appears to give better agreement with the contactless method, but requires extrapolation from data below instrument resolution., e.g., in colloidal gels or glasses [51]. The yield strains, $\gamma^{s}_{0}=0.23$ and $\gamma^{s}_{\rm rec}=0.1$, are also comparable, Fig. 5(a), but greater than previously observed for microgel-laden [16] or amorphous jammed interfaces [52]. Above yielding we are not able to compare DWR and contactless measurements, as the applied shear rates for DWR ($\dot{\gamma}^{s}=2\pi f\gamma^{s}_{0}\gtrsim$0.25\text{\,}{\mathrm{s}}^{-1}$$) are larger than those in the contactless geometry. To yield at comparable $\dot{\gamma}^{s}\approx$0.01\text{\,}{\mathrm{s}}^{-1}$$ would require $f\approx$0.01\text{\,}\mathrm{Hz}$$, resulting in infeasibly long experiments. Figure 7: PMMA-laden interface at low surface coverage. (a) Fluorescent confocal micrographs of PMMA particles (white) at an oil-water interface with $\phi=31.0\%$. (b) Corresponding strain vs time, $\gamma^{s}(t)$, at low imposed rheometer rotation rates, $\omega$, see inset legend. $r_{\rm out}=$6.6\text{\,}\mathrm{mm}$$. $\omega$ applied from $t\gtrsim$5\text{\,}\mathrm{s}$$ for $120\text{\,}\mathrm{s}$, followed by $30\text{\,}\mathrm{s}$ further recording. (c) Corresponding high imposed $\omega$ response. In contrast to the yield stress, the linear elastic behaviour is well-defined. Comparing the DWR elastic modulus below yielding, Fig. 5(a) (open circles), and the contactless elastic modulus (filled circles) we see excellent agreement. Fitting over the linear response region ($\gamma^{s}_{0}\leq 0.05$) the average elastic modulus from DWR is only 12% lower than the contactless measurement, $\bar{G^{s\prime}}=3.7(4)$ vs $4.2(1)\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$, demonstrating that they measure equal quantities within error. As yielding is approached, the DWR response appears to soften while the contactless measurement remains linear, Fig. 5(b). However, the oscillatory strain amplitude will contain both the recoverable elastic strain and plastic flow [53], in contrast to a creep-recovery measurement that separates the terms. So, approaching yielding, where plastic flow gradually begins [Fig. 5(c)] they are no longer directly comparable. The comparable $\sigma^{s}_{y}$ values and equal elastic moduli demonstrate that the contactless geometry accurately measures well-defined interfacial rheological properties. ### III.2 PMMA particle laden interface #### III.2.1 Low surface coverage We now turn to an interface laden with solid colloidal hydrophobic PMMA particles, which are typically challenging to measure using conventional interfacial rheometric techniques [14]. First, we investigate interfaces with low surface fraction, e.g. $\phi=31$% [Fig. 7(a)]. Interestingly, even though the particles are hydrophobic sterically stabilised nearly hard-sphere particles, they exhibit long-range repulsion when confined at an oil-water interface and assemble predominantly in a non-close packed arrangement [28]. Some aggregation is present due to attractive capillary and Van der Waals forces, Fig. 7(a). While capillary forces should be negligible, due to a vanishingly small Bond number and the use of spherical particles [4], there will be a certain particle roughness from the variable length of the steric stabiliser “hairs”. This may cause contact line undulations that lead to short-range capillary attraction [54, 55]. For this particle system, strain vs time plots show smooth flow with a constant shear rate over the duration of the creep step, Fig. 7(b). As expected, at larger imposed rotation rates from the rheometer a larger interfacial shear rate is measured in response, Fig. 7(c). Note that, when shear starts the interface appears to immediately (within temporal resolution of the analysis method) begin flowing at a constant shear rate. Similarly, when the shear ends the interface immediately stops flowing. This implies that the interface response in this regime is purely viscous, with no measurable elasticity. To quantify the rheology of these particle-laden interfaces, we plot stress vs shear rate, $\sigma^{s}(\dot{\gamma}^{s})$, with $\dot{\gamma}^{s}$ defined from the slope of $\gamma^{s}(t)$, Sec. II, due to the immediate response. As expected, for relatively low $\phi$ we observe a Newtonian response, Fig. 8(a) [dot-dashed (orange) line]. We can therefore assign a constant $\eta^{s}=$4.43(9)\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}\text{\,}\mathrm{s}$$ for the interfacial viscosity of this interface. This $\eta^{s}$ is comparable to that measured using a magnetic rod interfacial rheometer on a similar system [56]. Figure 8: Newtonian rheological behaviour of low surface coverage interfaces with varying parameters. (a) Stress–shear-rate behaviour of three interfaces with varying oil heights ($h_{o}$), surface coverages ($\phi$) and aggregation states ($\mathcal{D}$). Points, measured data. Lines, fits to constant viscosity: (i) $4.1(5)\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}\text{\,}\mathrm{s}$, solid (green); (ii) $2.10(10)\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}\text{\,}\mathrm{s}$, dashed (blue); (iii) $4.43(9)\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}\text{\,}\mathrm{s}$, dot-dashed (orange), also in Fig. 7. (b) Interface images: (i) $\phi=38\%$, $h_{o}=$2.9\text{\,}\mathrm{mm}$$ and $\mathcal{D}=24.7$; (ii) $\phi=29\%$, $h_{o}=$1.4\text{\,}\mathrm{mm}$$ and $\mathcal{D}=23.4$; and, (iii) $\phi=31\%$, $h_{o}=$2.2\text{\,}\mathrm{mm}$$ and $\mathcal{D}=2.49$. NB: surface fractions are determined from a rotational average, so single images are not wholly representative. ##### Effect of particle assembly. Next, we have performed repeats of these experiments while varying the oil phase thickness to test the robustness of our technique, Fig. 8. Confocal images prior to shearing reveal that the partial presence of aggregates, with particles in direct contact, varies between samples, Fig. 8(b). The aggregation state is characterized by the dispersity $\mathcal{D}$, Eq. (2), where a low $\mathcal{D}$ corresponds to a homogeneous particle distribution and a high $\mathcal{D}$ to an aggregated assembly. We will discuss the particle assembly’s influence on the rheology below. First, we observe a Newtonian response for all three oil thicknesses, Fig. 8(a). This suggests that, as expected, the height of the oil phase does not seem to have a substantial effect, i.e. the rheology of the interface is not expected to depend on the depth of the bulk phases. When comparing samples with different $\phi$ and aggregation states, we observe the following trends. First, an increase in $\phi$ leads to an increase in interfacial viscosity. For example, when comparing interfaces with similar aggregation states ($\mathcal{D}=24.7$ and $\mathcal{D}=23.4$), but different $\phi$, the interfacial viscosity at 38%, (green) solid line, is higher than at 29%, (blue) dashed, 4.1 vs $2.1\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}\text{\,}\mathrm{s}$. This trend is to be expected, as we later demonstrate elastic responses for high surface fractions of 56.7%, Sec. III.2.2. Second, the aggregation state seemingly affects the interfacial viscosity. When comparing samples with similar $\phi$ (29% and 31%), but different aggregation states ($\mathcal{D}=23.4$ and $\mathcal{D}=2.49$ respectively), we measure a significant difference in $\eta^{s}$ (2.1 vs $4.4\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}\text{\,}\mathrm{s}$), cf. (blue) dashed and (orange) dot-dashed lines. This suggests that the aggregation of PMMA particles at liquid interfaces leads to lower interfacial viscosities compared to more homogeneously distributed particles. Our observation that surface coverage and aggregation state make a significant difference to measured interfacial viscosity demonstrates the utility of the inherent combination of rheometric measurement with simultaneous imaging for the contactless geometry. It also aligns with previous reports; for example, Reynaert, Moldenaers, and Vermant [56] have shown that the complex surface viscosity magnitude increases with the surface fraction of weakly aggregated polystyrene particles at a water-oil interface. They also show that, for surface coverages below 80%, as studied here, the interfacial viscosity of aggregated particles at a water-oil interface is lower than that for stable particles. Note that aggregation _can_ lead to higher interfacial viscosities at high(er) surface coverage, see also, e.g., Ref. 21. Notably, the results of varying the thickness of the oil phase also imply that edge effects, i.e. deviations from our assumed parallel-plate flow field, Sec. II.5, do not play a substantial role in our setup. This is consistent with the agreement in the measured linear elastic modulus for PNIPAM-SiO2–laden interfaces between DWR and contactless methods, Sec. III.1.2. A substantial part of the top surface in our set-up is an oil-plate interface, but the outer edge of it is an oil-air interface i.e. the sample cup is $21\text{\,}\mathrm{mm}$ radius at the top, whereas the parallel-plate geometry attached to the oil-air interface is $12.5\text{\,}\mathrm{mm}$ radius, leaving a radial gap of $9\text{\,}\mathrm{mm}$ between the parallel- plate geometry and the inner wall of the PTFE cup. To mitigate any edge effects, we measure interfacial strains at a distance of about 4 mm $\ll$ 12.5 mm from the rotational axis. The results in Fig. 8, i.e. that the interfacial viscosities do not differ strongly with oil-phase thickness, imply that edge effects do not substantially affect our results. However, we should note that even in the contactless method, at low $\phi$, Fig. 7, such low interfacial viscosities mean that measurements are at a moderate Bo, around $\mathcal{O}(10)$. This suggests that for precise and absolute determination of $\eta^{s}$ in this regime a sub-phase drag correction may still be necessary. ##### Irreversible effect of shear. Next, we take advantage of the simultaneous confocal imaging to observe the evolution in particle assembly upon shearing. Before shearing, the particles are mostly in a non-close packed arrangement with partial aggregation ($\mathcal{D}=2.49$), which we attribute to attractive capillary and Van der Waals forces, Fig. 9(a). Upon mild shearing ($\omega\leq 6.3$ rpm), the arrangement is preserved and only minor changes in aggregation state are observed, Fig. 9(b) and (c). We see from images taken after high shear is applied, $\omega\gtrsim 10$ rpm corresponding to $\sigma^{s}=$6.6\text{\times}{10}^{-7}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$ [Fig. 9(d) and (e)], that the interface changes to an inhomogeneous structure with most particles forming one large aggregate percolating across the region imaged combined with an increase in dispersity to $\mathcal{D}=13.5$ in the final state. Importantly, the aggregation appears irreversible and persists even when higher shear rates are applied, Fig. 9(f). Figure 9: Structural evolution of PMMA particles at an oil-water interface with 31% surface coverage via fluorescent confocal micrographs with increasing applied rotation rate, $\omega$, and, hence, interfacial stress, $\sigma^{s}$. Interface corresponding to (orange) dot-dashed line in Fig. 8(a). (a) After low stress, $\omega=0.05$ rpm. Scale bar $200\text{\,}\mathrm{\SIUnitSymbolMicro m}$. (b) $\omega=4.0$ rpm. (c) $\omega=6.3$ rpm. (d) Aggregation threshold, $\omega=10$ rpm or $\sigma^{s}=$6.6\text{\times}{10}^{-7}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$. (d) Continued aggregation, $\omega=16$ rpm. (e) Highest $\omega=25$ rpm. Figure 10: PMMA-laden interface at high surface coverage. (a) Fluorescent confocal micrographs of PMMA particles (white) at an oil-water interface, $\phi=56.7\%$. (b) Corresponding strain vs time at low imposed rheometer rotation rates, $\omega$, see inset legend. $r_{\rm out}=$6.6\text{\,}\mathrm{mm}$$. $\omega$ applied from $t\gtrsim$5\text{\,}\mathrm{s}$$ for $120\text{\,}\mathrm{s}$, followed by $30\text{\,}\mathrm{s}$ further recording. (c) Corresponding high imposed $\omega$ response. We have seen that applying shear leads to considerable aggregation in this system. In order for aggregation to occur the shear force must exceed the maximum repulsive force between these particles. It has been observed that these particles have an interaction that can be described by a repulsive screened Coulomb potential [28]. In order to overcome this repulsion we assume that they must overcome the maximum repulsive force. This maximum force can be found to be $9.88\text{\times}{10}^{-13}\text{\,}\mathrm{N}$, where we have rescaled the parameters from Ref. 28, as in this work we use larger particles. This force can then be converted to an interfacial stress by dividing by the particle diameter to give a critical aggregation stress of $3.3\text{\times}{10}^{-7}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$. Experimentally, we found that aggregation occurs starting at a stress of $6.6\text{\times}{10}^{-7}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$, which is in reasonable agreement. This agreement between experiment and prediction lends further confidence to the use of Eq. (9) when calculating interfacial stress in our unique geometry. In order to aggregate, the applied stress must also overcome the steric repulsion, however the steric barrier is considerably smaller (for a similar particle) than the electrostatic barrier [28, 57]. Once these particles come into close contact, attractive capillary forces and/or van der Waals forces are large enough such that this is irreversible. #### III.2.2 High surface coverage Now, we investigate the same PMMA particles at higher surface fractions of 56.7%, Fig. 10(a), where the particles assemble into a percolated aggregated structure. We observe markedly different behaviour, with elastic behaviour being evident from the strain vs time plots, Fig. 10(b)–(c). Focussing on the low stress behaviour, Fig. 10(b), an initial jump to a higher strain is observed, indicative of an elastic material. There is then some erratic motion in the direction of shear (i.e. the strain is always positive), indicating that there is some frustrated motion and rearrangements of the interfacial structure [58]. While the initial elastic response is difficult to measure precisely due to background noise in the flow, upon cessation of shear the interface clearly recoils, allowing the elastic strain to be readily measured [59]. This statement becomes even more apparent when looking at higher applied stresses, Fig. 10(c), as at these large stresses the flowing behaviour completely dominates the strain response and the initial elastic jump is barely visible in the data. However, once the shear has been stopped, the elastic recoil is clear. At the end of shear at low rotation rate we sometimes observe motion, e.g., Fig. 10(b) (solid line), perhaps due to thermal gradients or air flow — note, however, that the shear rates are small. At higher $\phi$ we observe a more complex rheological response. By plotting the plastic response, stress vs shear rate, we can infer that the particle- laden interface behaves as a yield stress fluid, Fig. 11(a), as has been observed previously using the DWR geometry [14]. By fitting a Bingham plastic model, Eq. (11), we measure a yield stress of $1.05(15)\text{\times}{10}^{-7}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$. The effective interfacial viscosity ($2.16(14)\text{\times}{10}^{-5}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}\text{\,}\mathrm{s}$) is, as expected, larger than that measured at the lower $\phi$, Fig. 8. We feel that this is an appropriate model, as the parameter which we are most interested in comparing to data in the published literature is the yield stress. The measured $\sigma^{s}_{y}$ is, however, an order of magnitude lower than the yield stress quoted in Ref. 14. As close agreement between the techniques is found for a PNIPAM-SiO2–laden interface, Fig. 5(c), this suggests that the different surface coverages may not be comparable, 56.7% here vs 74% in Ref. 14. Moreover, there is a difference in PMMA stabilizer, poly(12-hydroxystearic acid) in Ref. 14 and poly(lauryl methacrylate) here, though that only makes a small difference in contact angle and a relatively small difference in interaction potential [28]. When plotting the elastic response at higher $\phi$, stress vs recoverable strain [Fig. 11(b)], the response is initially linear, with $\sigma^{s}$ and $\gamma^{s}_{\rm rec}$ proportional up to a strain of 0.03. We fit a linear dependence of the elastic strain response to the imposed shear stress in the low-strain regime [inset (orange) points]. This modelled Hookean behaviour gives us a shear modulus of $3.16(16)\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$; this is a factor $10\times$ lower in stiffness compared to the PNIPAM-SiO2 interface, Fig. 5(a), and over a smaller linear region, leading to a far weaker interface ($\sim 30\times$). Previously reported measurements on a similar particle-laden interface [14] found interfacial moduli of $\sim$2\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, which is consistent with our measurements here being at a slightly lower surface coverage, similar to the difference in $\sigma^{s}_{y}$. Figure 11: Rheological behaviour of an oil-water interface laden with PMMA particles at $\phi=56.7\%$. (a) Viscous response to applied stress, i.e. after initial elastic response, but before recoil. Points, data; line, fit to Bingham fluid model, Eq. (11). (b) Elastic response measured from recoil. Points, data; line, linear elastic fit to low strain, $\leq 0.03$ (orange and inset), $G^{s\prime}=$3.16(16)\text{\times}{10}^{-6}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$. Finally, we look at the high-stress elastic response, which suggests a complex structural evolution. At $\gamma^{s}_{\rm rec}\gtrsim 0.03$, Fig. 11(b) (blue), the recoverable strain initially remains unchanged with $\sigma^{s}$, no longer increasing linearly. The shift in response at $\sigma^{s}\\!\approx\\!$1\text{\times}{10}^{-7}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$$, correlates well with the yield stress, Fig. 11(a). So, just above $\sigma^{s}_{y}$ a fixed maximum $\gamma^{s}_{\rm rec}$ can be stored, similar to the PNIPAM-SiO2 interface, Fig. 5(a). However, as $\sigma^{s}$ increases further, $\gamma^{s}_{\rm rec}$ begins to increase more rapidly, i.e. strain softening. Qualitatively, this aligns with a variety of literature results. For example, Reynaert, Moldenaers, and Vermant [56] measured the surface elastic modulus for polystyrene particle aggregates at a water-oil interface, which decreased with strain amplitude. Zhang _et al._ [60] used large amplitude oscillatory strain rheology and observed strain softening for weakly attractive silica nanoparticles at an air-aqueous interface. Finally, Orsi _et al._ [61] used an interfacial shear rheometer on gold nanoparticles at an air-water interface and also observed strain softening, which they attributed to breaking of weak bonds in a 2D gel. This suggests that the interface evolves above yielding, notably, in the stress range for aggregation at low $\phi$, $3.3\text{\times}{10}^{-7}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$, which should be $\phi$ independent. It is then possible that the strain softening is driven by aggregation, consistent with aggregation weakening interfaces at all but the highest $\phi$ (and lowering $\eta^{s}$, Sec. III.2.1). ### III.3 Limits and comparison of techniques Our results, and the comparison to 1) DWR results using the same PNIPAM-SiO2 system, Fig. 5, and 2) similar colloidal particle laden interfaces using a DWR [14] or magnetic rod [56], suggest that our setup represents a useful addition to the field of interfacial rheology. In comparable situations, our results are highly consistent with results using conventional probes directly attached to the interface, e.g., the elastic modulus of the PNIPAM-SiO2 interface, or within expected variation due to differing methods or interface preparation, i.e. the yield stress for both a PNIPAM-SiO2 or PMMA-particle laden system. Crucially, our setup allows us to both measure viscosities at lower $\phi$ than have been previously observed using the DWR, and to probe static yielding at lower stresses. The lower stress limit can be estimated using Bo${}^{*}\approx 10$, Eq. (10), alongside an estimate of the minimum interfacial shear rate set by the imaging setup resolution. With a minimum resolvable strain of $\gamma_{\rm rec}^{s,\min}=0.002$, Fig. 5(a), over a $120\text{\,}\mathrm{s}$ experiment the minimum shear rate is $\dot{\gamma}^{s,\min}\sim$2\text{\times}{10}^{-5}\text{\,}{\mathrm{s}}^{-1}$$, or a rotation rate $\omega_{i}^{\min}=\dot{\gamma}^{s}r_{\rm out}/(r_{r}-r_{\rm out})\approx$2.5\text{\times}{10}^{-5}\text{\,}\mathrm{rad}\text{\,}{\mathrm{s}}^{-1}$$. For pre-factors in Eq. (10) $\approx 1$, this sets $\omega^{\min}\approx 0.002$ rpm. The minimum rotation rate then sets a lower stress limit, Eq. (9), $\sigma^{s,\min}=\mathcal{O}(${10}^{-8}\text{\,}\mathrm{Pa}\text{\,}\mathrm{m}$)$, comparable to more sensitive interfacial techniques, e.g., a micro-needle [24, 44]. To specifically probe low stresses, $\dot{\gamma}^{s,\min}$, and so $\sigma^{s,\min}$, could be further optimised by using a high magnification or longer imaging period, together with minimisation of noise (e.g., thermal gradients and vibrations). Most remarkably, our contactless technique retains the maximum stresses, Fig. 5(c), attainable for a DWR using a closed feedback loop [44], and hence has a dynamic range that spans the majority of interfacial shear rheometry methods. This wide range of measurable stress combined with in situ determination of interfacial characteristics, either surface pressure via a Wilhelmy plate or surface fraction via imaging, opens up the contactless technique to multiple future applications. ## IV Conclusion In this work, we have developed a contactless method to perform interfacial shear rheology on liquid-liquid interfaces without an interfacial geometry. The shear is applied to the continuous phase using a rotational rheometer and indirectly deforming the interface and the surface response is measured via confocal microscopy, either of a fluorescent particle-laden interface or via tracer particles embedded in the interface, enabling the measurement of a broad range of interfaces formed from, e.g., proteins, polymers or molecular surfactants [8]. While we use a confocal microscope and stress-controlled rheometer, the same results should be achievable using any fixed-rate motor and reflection or fluorescence microscopy with sufficient resolution and frame rate. This enhances the applicability of this method as only relatively common equipment is required. The method has been verified using a PNIPAM-SiO2–laden interface measured using both our novel contactless geometry and a conventional DWR method, with equal elastic moduli found and comparable yield stress values. Our contactless setup allows us to both measure interfacial viscosities at lower surface fractions than have been previously observed using the DWR, owing to the high sensitivity achieved by having no probe attached directly to the interface, while maintaining the ability to apply large interfacial stresses. Additionally, we have linked the rheological behaviour to the structural behaviour of PMMA particle interfaces with different initial conditions. At low surface coverage, the interface behaves as a two-dimensional Newtonian fluid and is subject to aggregation above a certain shear threshold. At higher surface coverage the interface begins to behave as an elastic sheet with a measurable shear modulus, up to a yield stress where the interface begins to flow. In addition, our results suggest that both surface coverage and interfacial particle aggregation state affect the rheology of the interface, in line with results in the literature. This work has focussed on the motion of the particles in the plane of the interface under steady shear. As the setup presented here does not have a probe attached to the liquid interface, the effect of the interface on how shear is propagated from the oil to the water phase can now be studied. This would facilitate observation of how the inside of an emulsion droplet is influenced by shear of the continuous phase, thereby greatly increasing the understanding and predictability of the flow behaviour of these systems, which are encountered ubiquitously in many formulation applications. ## Acknowledgments IM acknowledges studentship funding from the EPSRC Centre for Doctoral Training in Condensed Matter Physics (CM-DTC, EP/L015110/1). SB acknowledges studentship funding from the EPSRC Centre for Doctoral Training in Soft Matter and Functional Interfaces (SOFI-CDT, EP/L015536/1). MR acknowledges funding from the Marie Sklodowska-Curie Individual Fellowship (Grant No. 101064381). The authors acknowledge R. O’Neill, R. Van Hooghten, Damian Renggli and Jan Vermant for useful discussions, A. Garrie for the PTFE cup and aluminium ring, J. Royer for assistance with the rheoimaging setup, and M. Hermes for the velocimetry C code. For the purpose of open access, the authors have applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising from this submission. ## Author declarations ### Conflicts of Interest The authors declare that they have no conflicts of interest. ### Data Availability The data that support the findings of this study are openly available in Edinburgh DataShare at https://doi.org/10.7488/ds/3759. * ## Appendix A Surface Coverage Measurements As small differences in surface fraction can cause significant differences in rheological response, and surface fractions for non-close packed PMMA monolayers formed by sedimentation are challenging to reproduce, the surface coverage $\phi$ is determined for each interface prepared. First, microscopy images are analysed, the magnification level being a balance between individual particle resolution, which improves upon magnification, and statistics of particle counting, which decreases upon magnification. Examples are shown in Fig. 8(b). Surface fractions are quoted in terms of a pixel counting method, as described in the main text. To determine how accurately surface coverage can be defined, we compare the pixel fraction method with a particle counting method. Here, we can count the number of particles and use knowledge of particle size and image size to calculate surface fraction. For the sample in Fig. 7(a), measurements of surface fraction via the pixel fraction and the particle counting methods give respectively 31.0% and 23.0% while for the sample in Fig. 10(a) these give respectively 56.7% and 46.1%. Note the significant difference in these measurements, highlighting the challenge in defining the surface coverage. With perfect particle resolution, the particle counting method should yield the correct answer for that particular region of the interface, however, perfect particle resolution is rarely achieved (especially at high $\phi$), for instance because an aggregate could be mistaken for one particle. The pixel fraction method is also flawed in that it assumes a direct match between area of emitted light with area of the particle. This however is not true due to the point spread function of the imaging setup, a question over whether the particles are exactly in the focal plane, and the brightness of the fluorophore itself, among other considerations. Comparing these images with other, similar images taken using the same imaging setup allows us to have an estimate for the surface fraction and certainly allows us to observe trends in flow behaviour as surface fraction changes. It is also worth noting in Fig. 10(a) there appears to be two types of particle with different intensity levels. There are a few possible reasons: firstly, dispersity in particle size, fluorophore intensity, or contact angle [62] may lead to this effect. This aligns with some particles in Fig. 7(a) appearing smaller but with less appreciable change in intensity, presumably as the excitation signal is at saturation in these imaging conditions. Secondly, the particles, while in close contact, may be pushed out of the surface leading to variations in their vertical $z$-position. This, however, is unlikely as the $z$ resolution is $\gg$ particle diameter [63], $\delta z=\frac{0.88\lambda_{\rm exc}}{1-\sqrt{(n^{2}-{\rm NA}^{2})}}=$9.3\text{\,}\mathrm{\SIUnitSymbolMicro m}$,$ (12) for excitation wavelength $\lambda_{\rm exc}=$488\text{\,}\mathrm{nm}$$, dry objective refractive index $n=1$, and numerical aperture ${\rm NA}=0.3$. Finally, aggregates might cause a (local) curvature in the liquid-liquid interface, which could lead to some particles remaining interfacial but moving relative to the imaging plane, which would result in particles appearing smaller and/or less bright. These possible artefacts highlight the challenge in precisely determining the surface fraction from images. However, the pixel counting method gives a consistent result that allows a systematic comparison with increasing surface coverage of particles, Sec. III.2, and a reasonable comparison with the literature results of e.g., Ref. 14. ## References ## References * Fuller and Vermant [2012] G. G. Fuller and J. Vermant, “Complex fluid-fluid interfaces: Rheology and structure,” Annu. Rev. Chem. Biomol. Eng. 3, 519–543 (2012). * Thijssen and Vermant [2018] J. H. J. Thijssen and J. Vermant, “Interfacial rheology of model particles at liquid interfaces and its relation to (bicontinuous) Pickering emulsions,” J. Phys. Condens. Matter 30, 023002 (2018). * Derkach, Krägel, and Miller [2009] S. R. Derkach, J. Krägel, and R. Miller, “Methods of measuring rheological properties of interfacial layers (experimental methods of 2D rheology),” Colloid Journal 71, 1–17 (2009). * Binks and Horozov [2006] B. P. Binks and T. S. Horozov, eds., _Colloidal Particles at Liquid Interfaces_ (Cambridge University Press, Cambridge, 2006). * Tavacoli, Thijssen, and Clegg [2015] J. W. Tavacoli, J. H. J. Thijssen, and P. S. Clegg, “Chapter 6. bicontinuous emulsions stabilized by colloidal particles,” in _Particle-Stabilized Emulsions and Colloids: Formation and Applications_ (Royal Society of Chemistry, 2015) pp. 129–168. * Leal-Calderon and Schmitt [2008] F. Leal-Calderon and V. Schmitt, “Solid-stabilized emulsions,” Curr. Opin. Colloid Interface Sci. 13, 217–227 (2008). * Hunter _et al._ [2008] T. N. Hunter, R. J. Pugh, G. V. Franks, and G. J. Jameson, “The role of particles in stabilising foams and emulsions,” Adv. Colloid Interface Sci. 137, 57–81 (2008). * Tcholakova, Denkov, and Lips [2008] S. Tcholakova, N. D. Denkov, and A. Lips, “Comparison of solid particles, globular proteins and surfactants as emulsifiers,” Phys. Chem. Chem. Phys. 10, 1608–1627 (2008). * Fernandez-Rodriguez, Martín-Molina, and Maldonado-Valderrama [2021] M. A. Fernandez-Rodriguez, A. Martín-Molina, and J. Maldonado-Valderrama, “Microgels at interfaces, from mickering emulsions to flat interfaces and back,” Adv. Colloid Interface Sci. 288, 102350 (2021). * Perrin _et al._ [2020] L. Perrin, G. Gillet, L. Gressin, and S. Desobry, “Interest of Pickering emulsions for sustainable micro/nanocellulose in food and cosmetic applications,” Polymers 12, 1–14 (2020). * Vandebril _et al._ [2010] S. Vandebril, A. Franck, G. G. Fuller, P. Moldenaers, and J. Vermant, “A double wall-ring geometry for interfacial shear rheometry,” Rheol. Acta 49, 131–144 (2010). * Masschaele, Fransaer, and Vermant [2011] K. Masschaele, J. Fransaer, and J. Vermant, “Flow-induced structure in colloidal gels: direct visualization of model 2D suspensions,” Soft Matter 7, 7717 (2011). * Keim and Arratia [2013] N. C. Keim and P. E. Arratia, “Yielding and microstructure in a 2D jammed material under shear deformation,” Soft Matter 9, 6222 (2013). * Van Hooghten _et al._ [2017] R. Van Hooghten, V. E. Blair, A. Vananroye, A. B. Schofield, J. Vermant, and J. H. J. Thijssen, “Interfacial rheology of sterically stabilized colloids at liquid interfaces and its effect on the stability of Pickering emulsions,” Langmuir 33, 4107–4118 (2017). * Krägel and Derkach [2010] J. Krägel and S. R. Derkach, “Interfacial shear rheology,” Curr. Opin. Colloid Interface Sci. 15, 246–255 (2010). * Brugger, Vermant, and Richtering [2010] B. Brugger, J. Vermant, and W. Richtering, “Interfacial layers of stimuli-responsive poly-(N-isopropylacrylamide-co- methacrylicacid) (PNIPAM-co-MAA) microgels characterized by interfacial rheology and compression isotherms,” Phys. Chem. Chem. Phys. 12, 14573–14578 (2010). * Razavi _et al._ [2015] S. Razavi, K. D. Cao, B. Lin, K. Y. C. Lee, R. S. Tu, and I. Kretzschmar, “Collapse of particle-laden interfaces under compression: Buckling vs particle expulsion,” Langmuir 31, 7764–7775 (2015). * Garbin _et al._ [2015] V. Garbin, I. Jenkins, T. Sinno, J. C. Crocker, and K. J. Stebe, “Interactions and stress relaxation in monolayers of soft nanoparticles at fluid-fluid interfaces,” Phys. Rev. Lett. 114, 108301 (2015). * Ravera, Loglio, and Kovalchuk [2010] F. Ravera, G. Loglio, and V. I. Kovalchuk, “Interfacial dilational rheology by oscillating bubble/drop methods,” Curr. Opin. Colloid Interface Sci. 15, 217–228 (2010). * Barman and Christopher [2014] S. Barman and G. F. Christopher, “Simultaneous interfacial rheology and microstructure measurement of densely aggregated particle laden interfaces using a modified double wall ring interfacial rheometer,” Langmuir 30, 9752–9760 (2014). * Barman and Christopher [2016] S. Barman and G. F. Christopher, “Role of capillarity and microstructure on interfacial viscoelasticity of particle laden interfaces,” J. Rheol. 60, 35–45 (2016). * Brooks _et al._ [1999] C. F. Brooks, G. G. Fuller, C. W. Frank, and C. R. Robertson, “Interfacial stress rheometer to study rheological transitions in monolayers at the air-water interface,” Langmuir 15, 2450–2459 (1999). * Reynaert _et al._ [2008] S. Reynaert, C. F. Brooks, P. Moldenaers, J. Vermant, and G. G. Fuller, “Analysis of the magnetic rod interfacial stress rheometer,” J. Rheol. 52, 261–285 (2008). * Tajuelo _et al._ [2015] J. Tajuelo, J. M. Pastor, F. Martínez-Pedrero, M. Vázquez, F. Ortega, R. G. Rubio, and M. A. Rubio, “Magnetic microwire probes for the magnetic rod interfacial stress rheometer,” Langmuir 31, 1410–1420 (2015). * Jaensson, Anderson, and Vermant [2021] N. O. Jaensson, P. D. Anderson, and J. Vermant, “Computational interfacial rheology,” J. Non-Newt. Fluid Mech. 290, 104507 (2021). * Kim _et al._ [2011] K. Kim, S. Q. Choi, J. A. Zasadzinski, and T. M. Squires, “Interfacial microrheology of DPPC monolayers at the air-water interface,” Soft Matter 7, 7782–7789 (2011). * Mears, Muntz, and Thijssen [2020] R. Mears, I. Muntz, and J. H. J. Thijssen, “Surface pressure of liquid interfaces laden with micron-sized particles,” Soft Matter 16, 9347–9356 (2020). * Muntz _et al._ [2020] I. Muntz, F. Waggett, M. Hunter, A. B. Schofield, P. Bartlett, D. Marenduzzo, and J. H. J. Thijssen, “Interaction between nearly hard colloidal spheres at an oil-water interface,” Phys. Rev. Research 2, 023388 (2020). * Goebel and Lunkenheimer [1997] A. Goebel and K. Lunkenheimer, “Interfacial tension of the water/n-alkane interface,” Langmuir 13, 369–372 (1997). * Ciarella _et al._ [2021] S. Ciarella, M. Rey, J. Harrer, N. Holstein, M. Ickler, H. Löwen, N. Vogel, and L. M. C. Janssen, “Soft particles at liquid interfaces: From molecular particle architecture to collective phase behavior,” Langmuir 37, 5364–5375 (2021). * Sing _et al._ [2018] J. Sing, J. Tang, R. S. Bader, E. S. A. Goerlitzer, J. F. Wendisch, G. R. Bourret, M. Rey, and N. Vogel, “Surface patterning with SiO2@PNiPAm core – shell particles,” ACS Omega 3, 12089–12098 (2018). * Barrett [1974] K. E. J. Barrett, ed., _Dispersion Polymerization in Organic Media_ (Wiley, New York, 1974). * Jardine and Bartlett [2002] R. S. Jardine and P. Bartlett, “Synthesis of non-aqueous fluorescent hard-sphere polymer colloids,” Colloids Surf. A 211, 127–132 (2002). * Rey _et al._ [2020a] M. Rey, M. A. Fernandez-Rodriguez, M. Karg, L. Isa, and N. Vogel, “Poly- N-isopropylacrylamide nanogels and microgels at fluid interfaces,” Acc. Chem. Res. 53, 414–424 (2020a). * Rauh _et al._ [2017] A. Rauh, M. Rey, L. Barbera, M. Zanini, M. Karg, and L. Isa, “Compression of hard core–soft shell nanoparticles at liquid–liquid interfaces: influence of the shell thickness,” Soft Matter 13, 158–169 (2017). * Rey _et al._ [2020b] M. Rey, M. J. Uttinger, W. Peukert, J. Walter, and N. Vogel, “Probing particle heteroaggregation using analytical centrifugation,” Soft Matter 16, 3407–3415 (2020b). * Rey _et al._ [2017] M. Rey, A. D. Law, D. M. A. Buzza, and N. Vogel, “Anisotropic self-assembly from isotropic colloidal building blocks,” J. Am. Chem. Soc. 139, 17464–17473 (2017). * Besseling _et al._ [2009] R. Besseling, L. Isa, E. R. Weeks, and W. C. K. Poon, “Quantitative imaging of colloidal flows,” Adv. Colloid Interface Sci. 146, 1–17 (2009). * Dutta _et al._ [2013] S. K. Dutta, A. Mbi, R. C. Arevalo, and D. L. Blair, “Development of a confocal rheometer for soft and biological materials,” Rev. Sci. Instrum. 84, 063702 (2013). * Hermes and Clegg [2013] M. Hermes and P. S. Clegg, “Yielding and flow of concentrated Pickering emulsions,” Soft Matter 9, 7568 (2013). * Nguyen and Boger [1992] Q. D. Nguyen and D. V. Boger, “Measuring the flow properties of yield stress fluids,” Annu. Rev. Fluid Mech. 24, 47–88 (1992). * Macosko [1994] C. W. Macosko, _Rheology: Principles, Measurements, and Applications_ (VCH, New York, 1994) p. 217. * Note [1] This can be recovered by factorising the denominator, $\gamma^{s}_{0}=\theta_{0}[R_{o}^{2}/[(R_{o}-R_{r,o})(R_{o}+R_{r,o})]+R_{i}^{2}/[(R_{r,i}-R_{i})(R_{r,i}+R_{i})]$, and approximating to first order by taking $R_{i/o}+R_{r,i/o}\approx 2R_{r,i/o}$, $R_{r,i}\approx R_{r,o}$ and $R_{o}-R_{r,o}\approx R_{r,i}-R_{i}$. * Renggli _et al._ [2020] D. Renggli, A. Alicke, R. H. Ewoldt, and J. Vermant, “Operating windows for oscillatory interfacial shear rheology,” J. Rheol. 64, 141–160 (2020). * Oza and Venerus [2021] A. U. Oza and D. C. Venerus, “The dynamics of parallel-plate and cone–plate flows,” Phys. Fluids 33, 023102 (2021). * Verwijlen _et al._ [2011] T. Verwijlen, P. Moldenaers, H. A. Stone, and J. Vermant, “Study of the flow field in the magnetic rod interfacial stress rheometer,” Langmuir 27, 9345–9358 (2011). * Erni and Parker [2012] P. Erni and A. Parker, “Nonlinear viscoelasticity and shear localization at complex fluid interfaces,” Langmuir 28, 7757–7767 (2012). * Truzzolillo _et al._ [2016] D. Truzzolillo, H. Sharaf, U. Jonas, B. Loppinet, and D. Vlassopoulos, “Tuning the structure and rheology of polystyrene particles at the air–water interface by varying the ph,” Langmuir 32, 6956–6966 (2016). * Dinkgreve _et al._ [2016] M. Dinkgreve, J. Paredes, M. M. Denn, and D. Bonn, “On different ways of measuring “the” yield stress,” J. Non-Newt. Fluid Mech. 238, 233–241 (2016). * Note [2] Oscillatory yielding is ambiguous, as it can be gradual, with multiple definitions; $G^{s\prime}=G^{s\prime\prime}$ is an upper (over)estimate [49]. E.g., a tangent analysis appears to give better agreement with the contactless method, but requires extrapolation from data below instrument resolution. * Pham _et al._ [2008] K. N. Pham, G. Petekidis, D. Vlassopoulos, S. U. Egelhaaf, W. C. K. Poon, and P. N. Pusey, “Yielding behavior of repulsion-and attraction-dominated colloidal glasses,” J. Rheol. 52, 649–676 (2008). * Galloway, Jerolmack, and Arratia [2020] K. L. Galloway, D. J. Jerolmack, and P. E. Arratia, “Quantification of plasticity via particle dynamics above and below yield in a 2D jammed suspension,” Soft Matter 16, 4373–4382 (2020). * Kamani, Donley, and Rogers [2021] K. Kamani, G. J. Donley, and S. A. Rogers, “Unification of the rheological physics of yield stress fluids,” Phys. Rev. Lett. 126, 218002 (2021). * Stamou, Duschl, and Johannsmann [2000] D. Stamou, C. Duschl, and D. Johannsmann, “Long-range attraction between colloidal spheres at the air-water interface: The consequence of an irregular meniscus,” Phys. Rev. E 62, 5263–5272 (2000). * Van Hooghten _et al._ [2013] R. Van Hooghten, L. Imperiali, V. Boeckx, R. Sharma, and J. Vermant, “Rough nanoparticles at the oil-water interfaces: Their structure, rheology and applications,” Soft Matter 9, 10791–10798 (2013). * Reynaert, Moldenaers, and Vermant [2007] S. Reynaert, P. Moldenaers, and J. Vermant, “Interfacial rheology of stable and weakly aggregated two-dimensional suspensions,” Phys. Chem. Chem. Phys. 9, 6463–6475 (2007). * Mewis and Vermant [2000] J. Mewis and J. Vermant, “Rheology of sterically stabilized dispersions and lattices,” Prog. Org. Coat. 40, 111–117 (2000). * Weeks and Weitz [2002] E. R. Weeks and D. A. Weitz, “Properties of cage rearrangements observed near the colloidal glass transition,” Phys. Rev. Lett. 89, 095704 (2002). * Imperiali _et al._ [2012] L. Imperiali, K. H. Liao, C. Clasen, J. Fransaer, C. W. Macosko, and J. Vermant, “Interfacial rheology and structure of tiled graphene oxide sheets,” Langmuir 28, 7990–8000 (2012). * Zhang _et al._ [2016] H. Zhang, K. Yu, O. J. Cayre, and D. Harbottle, “Interfacial particle dynamics: One and two step yielding in colloidal glass,” Langmuir 32, 13472–13481 (2016). * Orsi _et al._ [2012] D. Orsi, G. Baldi, P. Cicuta, and L. Cristofolini, “On the relation between hierarchical morphology and mechanical properties of a colloidal 2D gel system,” Colloids Surf. A 413, 71–77 (2012). * Isa _et al._ [2011] L. Isa, F. Lucas, R. Wepf, and E. Reimhult, “Measuring single-nanoparticle wetting properties by freeze-fracture shadow-casting cryo-scanning electron microscopy,” Nature Commun. 2, 438 (2011). * Cole, Jinadasa, and Brown [2011] R. W. Cole, T. Jinadasa, and C. M. Brown, “Measuring and interpreting point spread functions to determine confocal microscope resolution and ensure quality control,” Nat. Protoc. 6, 1929–1941 (2011).
# Quantum Configuration and Phase Spaces: Finsler and Hamilton Geometries Saulo Albuquerque<EMAIL_ADDRESS>Physics Department, Federal University of Paraíba, Caixa Postal 5008, 58059-900, João Pessoa, PB, Brazil. Valdir B. Bezerra<EMAIL_ADDRESS>Physics Department, Federal University of Paraíba, Caixa Postal 5008, 58059-900, João Pessoa, PB, Brazil. Iarley P. Lobo<EMAIL_ADDRESS>Department of Chemistry and Physics, Federal University of Paraíba, Rodovia BR 079 - Km 12, 58397-000 Areia-PB, Brazil. Physics Department, Federal University of Lavras, Caixa Postal 3037, 37200-000 Lavras-MG, Brazil. Gabriel Macedo<EMAIL_ADDRESS>Physics Department, Federal University of Paraíba, Caixa Postal 5008, 58059-900, João Pessoa, PB, Brazil. Pedro H. Morais<EMAIL_ADDRESS>Physics Department, Federal University of Paraíba, Caixa Postal 5008, 58059-900, João Pessoa, PB, Brazil. Ernesto Rodrigues<EMAIL_ADDRESS>Physics Department, Federal University of Paraíba, Caixa Postal 5008, 58059-900, João Pessoa, PB, Brazil. Luis C. N. Santos<EMAIL_ADDRESS>Physics Department, Federal University of Paraíba, Caixa Postal 5008, 58059-900, João Pessoa, PB, Brazil. Gislaine Varão<EMAIL_ADDRESS>Physics Department, Federal University of Paraíba, Caixa Postal 5008, 58059-900, João Pessoa, PB, Brazil. ###### Abstract In this paper, we review two approaches that can describe, in a geometrical way, the kinematics of particles that are affected by Planck-scale departures, named Finsler and Hamilton geometries. By relying on maps that connect the spaces of velocities and momenta, we discuss the properties of configuration and phase spaces induced by these two distinct geometries. In particular, we exemplify this approach by considering the so-called $q$-de Sitter-inspired modified dispersion relation as a laboratory for this study. We finalize with some points that we consider as positive and negative ones of each approach for the description of quantum configuration and phases spaces. ## I Introduction Since the original works by Bronstein Bronstein:2012zz that demonstrated uncertainty in the localization of events when geometrical degrees of freedom are quantized, it has been argued that attempts to formulate quantum gravity in a differentiable manifold endowed with smooth geometric quantities would not be an interesting path to follow if one aims to pursue a fundamental approach to this problem. Attempts in this direction have accumulated over the years, having prominent representatives such as loop quantum gravity (LQG) Rovelli:2008zza and causal dynamical triangulation Ambjorn:2010rx . These approaches to quantum gravity predict or describe several effects that should be manifest at the Planckian regime of length and energy, such as the discretization of geometry, which requires a language that obviously departs from the usual Riemannian construction of general relativity. Despite the elegance of such approaches, with current technology we are far from being able to concretely address the regime in which such discretization would become evident. Nevertheless, the notion that spacetime could effectively behave like a medium formed by ”atoms of space” has led to a rich phenomenological approach to quantum gravity, which by encoding generic departures from relativistic equations, can describe common predictions expected to be present at an intermediate stage between classical and quantum gravity. Such an approach is encompassed in the area of quantum gravity phenomenology, which addresses a myriad of effects beyond the one described in this paragraph, as can be seen in Ref. Amelino-Camelia:2008aez , and in particular, has found in multimessenger astronomy a fruitful environment to be explored Addazi:2021xuf . Usually, the regime, in which this idea is considered, is the regime, in which the test particle approximation is valid consisting of the approximation, in which one would have simultaneously faint gravitational and quantum effects, described by the limits of the gravitational constant, $G\rightarrow 0$, and the reduced Planck’s constant, $\hbar\rightarrow 0$, however, with the Planck energy, $E_{P}=\sqrt{c^{5}\hbar/G}$, being finite, with $c$ the speed of light. This deformed ”Minkowski limit,” which presents departures from Minskowski spacetime’s structure has been suggested by various quantum gravity proposals, such as the linearization of the hypersurface deformation algebra inspired by LQG Amelino-Camelia:2016gfx ; Brahma:2016tsq ; Brahma:2018rrg and non-commutative geometry Majid:1994cy ; Lukierski:1991pn ; Lukierski:1992dt ; Majid:1996kd (for more details on this Miskowski limit, see Section 3.1.1 of Ref. Amelino-Camelia:2008aez , and for more references on other theoretical approaches in which such limit emerges, please refer to Section 2.2 of Ref. Addazi:2021xuf ). It is expected that the path between the differentiable Riemannian description of special (and general) relativity and the complete quantum gravity theory should pass through an intermediate regime, in which one has departures from the Riemannian character of spacetime but still has geometric features that could describe a bottom-up phenomenology. Furthermore, geometry plays an important role in the description of principles that have guided the developments of relativistic theories; for example, the principle of covariance is manifest through the use of tensorial equations of motion, the local relativity principle is a physical manifestation of having local equations of motion invariant under the Poincaré group (which is the group of isometries of Minkowski space), the equivalence principle of general relativity is manifest in the fact that the motion of free particles is realized through geodesics, and the clock postulate can be expressed by stating that an observer measures its proper time by the arc-length of its own trajectory. An important part of quantum gravity phenomenology is devoted to the question of whether, in the aforementioned Minkowski limit, the Lorentz invariance, and consequently, the local relativity principle, is preserved or broken due to Planck-scale effects Amelino-Camelia:2010lsq . As is known, a length/energy scale is not invariant under Lorentz transformations, which implies that either a quantum gravity scale breaks the equivalence of inertial frames in the aforementioned Minkowski limit, or the Lorentz or Poincaré group only describes a low energy/large distance approximation of a deeper transformation between inertial frames. The former possibility is known as a Lorentz invariance violation (LIV) scenario Mattingly:2005re ; Liberati:2013xla , and the latter is known as doubly (or deformed) special relativity (DSR) Amelino- Camelia:2000stu ; Magueijo:2001cr . As the geometrization of special relativity, due to Minkowski, paved the way to more fundamental descriptions of nature, we shall follow a similar path, but of geometrizing DSR. Geometric descriptions of DSR through non-commutative geometry are known Majid:1994cy ; Lukierski:1991pn ; Lukierski:1992dt ; Majid:1996kd , but we revise some continuous, differentiable ways of exploring non-Riemannian degrees of freedom and the possibilities for preserving the aforementioned principles. This way, we critically analyze two extensions of Riemannian geometry that are capable of describing aspects of an emergent ”quantum configuration and phase spaces” that preserve the intuition of those principles: they are Finsler and Hamilton geometries. Finsler geometry originally is related to the space of events and velocities (for this reason we refer to a quantum configuration space), and Hamilton geometry originally described the space of events and momenta (for this reason, we call it a quantum phase space). In this paper, we revise the phenomenological opportunities that emerge from these approaches and the interplay between them. We also condensate the utility of each of these geometries and their limitations in the current scenario. We should also stress that the approaches described in this review, refer to configuration and phase spaces probed by a single particle. The geometry probed by a multi-particle system and its interplay with Finsler and Hamilton languages (or even geometries that go beyond them) should still be further explored, in which, possibly the intuition gained from the relative locality framework Amelino-Camelia:2011lvm would play a prominent role in this approach. The paper is organized as follows. Section II revisits the origin of the idea of describing the effective spacetime probed by a particle that propagates through a modified dispersion relation (MDR) by the proposal of rainbow metrics. Section III revisits how this general idea is realized by the use of Finsler geometry in the tangent bundle, whose dual version in the cotangent bundle is discussed in Section IV, which is illustarated by considering the curved non- trivial case of $q$-de Sitter-inspired Finsler geometry. Section V considers the situation of deriving the geometry of the cotangent bundle, and, in Section VI, its dual tangent bundle formalization defined by Hamilton geometry is considered, which is illustrated by the $q$-de Sitter case. In Section VII, we comparatively discuss these two approaches and highlight points that we consider as useful as well as their limitations. Finally, some important remarks are drawn in Section VIII. Throughout the paper, a system of units with $c=\hbar=1$ is used, so that the Planck length is the inverse of the Planck energy: $\sqrt{G}=\ell=E_{\text{P}}^{-1}$. ## II Preliminaries on Rainbow Geometries As described above, over the years, the intuition that spacetime would behave like material media, where instead of atoms of matter, one would have atoms of spacetime, has been solidified through some approaches of quantum gravity. Just as occurs in matter, in which one does not need to know the specific details of the granular structure of a given medium to study the propagation of particles through it, in spacetime, one can build phenomenology-inspired ways of modeling how elementary particles interact with discrete gravitational degrees of freedom while traveling through space, a so-called ”in-vaccuum dispersion.” One could say that the most popular way of doing this is through the assumption that particles would obey a modified dispersion relation, whose corrections are given perturbatively by powers of the quantum gravity scale, which we could assume as being in the order of Planck units. The dispersion relation furnishes the group velocity of waves and defines the trajectory that on-shell particles follow from the Hamilton equations. Actually, when the interplay between the presence of amplifiers of observables and the uncertainties of observations allows us to constrain this parameter at a level close to its Planckian version, we say that we are at Planck-scale sensitivity Amelino-Camelia:2008aez . Such behavior also happens in meta-materials proutorov2018finsler , in which it is possible to describe the motion of particles through it by geodesics in a given geometry; it also appears in the motion of a charged particle in a pre-metric formulation of electromagnetism hehl2003foundations , in the description of seismic waves vcerveny2002fermat , etc.; for a review, see Ref. Pfeifer:2019wus . Additionally, one could wonder if the motion of particles, determined by Planck-scale modified dispersion relations, could also be described by geodesics of a non-Riemannian geometry. Besides, the dispersion relation itself is usually determined by the norm of the $4$-momentum measured by a Riemannian metric, which also determines the symmetries observed by measurements in that spacetime. This intuition was early realized by the so-called ”rainbow geometries” Magueijo:2002xx , idealized by João Magueijo and Lee Smolin which aimed to extend the DSR formulation proposed by them in Ref. Magueijo:2001cr to curved spacetimes. In that case, the way found to express local modified dispersion relations through a norm, consisted in absorbing functions of the particle’s energy divided by Planck energy, $\epsilon=E/E_{\text{P}}$, such as $f(\epsilon)$ and $g(\epsilon)$, which would appear in the MDR that follows: $m^{2}=f^{2}(\epsilon)E^{2}-g^{2}(\epsilon)|\vec{p}|^{2}\,,$ (1) (with the three-momentum $\vec{p}$) into the definition of new spacetime tetrads, $\tilde{e}_{(0)}^{\quad\mu}(\epsilon)=f(\epsilon)e_{(0)}^{\quad\mu}$ and $\tilde{e}_{(I)}^{\quad\mu}(\epsilon)=g(\epsilon)e_{(I)}^{\quad\mu}$, such that the MDR reads $m^{2}=\eta^{AB}\tilde{e}_{(A)}^{\quad\mu}\tilde{e}_{(B)}^{\quad\nu}p_{\mu}p_{\nu}=\tilde{g}^{\mu\nu}(\epsilon)p_{\mu}p_{\nu}\,,$ (2) where $g^{\mu\nu}(\epsilon)=\eta^{AB}\tilde{e}_{(A)}^{\quad\mu}\tilde{e}_{(B)}^{\quad\nu}$ is the rainbow metric, $\eta^{AB}$ is the Minkowski metric $\text{diag}(+---)$ , Greek letters denote four-dimensional indices and take on the values 0 (time) 1, 2, and 3 (space), low-case Latin letters denote the space indices, and $p_{\mu}$ is the 4-momentum. This description would imply that when an observer uses the motion of a particle with energy $E$ to probe spacetime, then the line element assigned to that spacetime is the following: $ds^{2}=\tilde{g}_{\mu\nu}dx^{\mu}dx^{\nu}=\frac{g_{00}}{f^{2}(\epsilon)}(dx^{0})^{2}+\frac{g_{ij}}{g^{2}(\epsilon)}dx^{i}dx^{j}+2\frac{g_{0i}}{f(\epsilon)g(\epsilon)}dx^{0}dx^{i}\,,$ (3) where $g_{\mu\nu}$ is the metric found from undeformed tetrads. Thus, in a nutshell, one identifies the rainbow functions, $f$ and $g$, from a MDR that is usually inspired by fundamental theories of quantum gravity or by phenomenological intuition; then, one uses $\tilde{g}_{\mu\nu}$ as an input into the classical gravitational field equations. Considering modifications of the stress-energy tensor due to the rainbow functions, one derives what should be $g_{\mu\nu}$ (since $f$ and $g$ are determined a priori). Usually, this procedure gives that $g_{\mu\nu}$ is the Riemannian metric found from the usual gravitational field equations. Therefore, this approach gives basically the usual metric components of a given theory, just modified by factors of the rainbow functions as in Equation (3). Effective energy-dependent spacetimes have emerged in different approaches to the description of the quantization of gravitational/geometric degrees of freedom Weinfurtner:2008if ; Assanioussi:2014xmz ; Olmo:2011sw . Along this line of research, Magueijo-Smolin’s proposal has been applied in a myriad of contexts, such as in black hole physics Ling:2005bp ; Lobo:2021bag , cosmology Gorji:2016laj , wormholes Garattini:2015pmo ; Amirabi:2018ncf , cosmic strings Bezerra:2019vrz , disformal geometries Carvalho:2015omv ; Lobo:2017bfh , and electrostatic self-interaction of charged particles Santos:2019 . However, despite its range of applicability and utility in furnishing intuition about extreme scenarios, this approach presents some conceptual and technical limitations that seem unavoidable, such as the lack of a rigorous mathematical framework in which this idea is formulated or the imposition of a preferred vielbein in which the particle’s energy is measured, which seems in contradiction with the local DSR intention of this proposal. As shown below, the solution to these problems is actually coincident, and the search for a rigorous mathematical formulation for these ideas will be responsible for giving a framework, in which proper physical questions can be answered and novel phenomenological opportunities to born. The main issue here is what is the proper formulation of a geometry that should not only depend on spacetime points, but also should carry energy dependence of the particle itself that probes this spacetime. This paper deals with the two main proposals—Finsler and Hamilton geometries— solving some of the raised problems and also discusses limitations on their owns. ## III Geometry of the Tangent Bundle: Finsler Geometry The 1854 Habilitation Dissertation by Bernhard Riemann presents the germ of the idea behind what would later be called Finsler geometry. In the second part of the dissertation, it is said (see Ref. riemann , p. 35): > ”For Space, when the position of points is expressed by rectilinear co- > ordinates, $ds=\sqrt{\sum(dx)^{2}}$; Space is therefore, included in this > simplest case. The next case in simplicity includes those manifoldnesses in > which the line-element may be expressed as the fourth root of a quartic > differential expression. The investigation of this more general kind would > require no really different principles, but would take considerable time and > throw little new light on the theory of space, especially as the results > cannot be geometrically expressed; I restrict myself, therefore, to those > manifoldnesses in which the line-element is expressed as the square root of > a quadric differential expression.” The exploration of such more general cases of line elements will be done only 64 years later, in 1918, in the Ph.D. thesis of Paul Finsler book:1130902 , where at least from the metric point of view, the distance between points is (”distance”), please confirm. measured by a 1-homogeneous function (homogeneous with the degree of 1) Such a metric tensor would be defined in the tangent bundle of the base manifold, since it would depend not only on the manifold points, but also on a direction, which is a manifestation of the non- Pythagorean nature of this space. Later on, the issue of non-linear connections was further developed and incorporated as a fundamental structure for the dynamical description of Finsler spaces (for a historical perspective on Finsler geometry, we refer the reader to the Preface of Ref. bao2000introduction and references therein). The case of pseudo-Finsler geometries, as an arena for describing spacetime, has been recently discussed Hohmann:2021zbt ; Bernal:2020bul , where, for instance, different definitions are presented and important theorems regarding its causal structure among other issues are being derived Minguzzi:2014aua . In Section II, a glimpse of the non-Riemannian nature of spacetime was notified emerging as a manifestation of the quantization of gravitational degrees of freedom. Actually, as one can anticipate, the non-quadratic, i.e., non-Pythagorean nature of a dispersion relation is connected to a possible Finslerian nature of spacetime through an intermediate step that connects the kinematics of particles in a Hamiltonian to a Lagrangian formulation. Actually, the MDR corresponds to a Hamiltonian constraint, which physical particles supposedly obey, the way that the trajectories of free particles, induced by such a deformed Hamiltonian, capture the propagation of a particle through a quantized spacetime. For this reason, the Helmholtz action, associated with such a particle, is naturally given by the functional, $S[x,p,\lambda]=\int d\mu(\dot{x}^{\alpha}p_{\alpha}-\lambda f(H(x,p),m))\,,$ (4) where the dot denotes differentiation with respect to the parameter $\mu$, $p_{\mu}$ is the particle’s momenta, $f$ is a function that is null if the dispersion relation is satisfied, namely, $H(x,p)=m$, and $\lambda$ is a Lagrange multiplier. This is a premetric formulation that is actually defined in the space $T^{*}M\times{\mathbb{R}}$, where $T^{*}M$ is the phase space of analytical mechanics or cotangent bundle. In order to find an arc-length, and consequently, a geometric structure, one needs to calculate an equivalent Lagrangian defined in the configuration space or the tangent bundle $TM$ described by points and velocities (such an observation was firstly presented in Ref. Girelli:2006fw ). The algorithm for doing so is as follows Lobo:2020qoa : 1. 1. variation with respect to $\lambda$ enforces the dispersion relation; 2. 2. variation with respect to $p_{\mu}$ yields an equation $\dot{x}^{\mu}=\dot{x}^{\mu}(p,\lambda)$, which must be inverted to obtain $p_{\mu}(x,\dot{x},\lambda)$ to eliminate the momenta $p_{\mu}$ from the action; 3. 3. using $p_{\mu}(x,\dot{x},\lambda)$ in the dispersion relation, one can solve for $\lambda(x,\dot{x})$; and 4. 4. finally, the desired length measure is obtained as $S[x]=S[x,p(x,\dot{x},\lambda(x,\dot{x})),\lambda(x,\dot{x})]_{H}$. This is a Legendre transformation, whose conditions of existence and capability of providing a physical framework are discussed in Refs. Raetzel:2010je ; Rodrigues:2022mfj . These formal conditions are always guaranteed when one considers deformations at the perturbative level. This is crucial because the following algorithm cannot be applied in practice if it is not possible to invert the velocity function to find the momenta as a function of the other variables. In general, this cannot be done, especially for complicated dispersion relations, such as those that depend on sums of hyperbolic functions Gubitosi:2011hgc . Anyway, since quantum gravity phenomenology is usually concerned with first order effects, which are those attainable by experiments nowadays, we shall concentrate on the perturbative level in order to derive our conclusions. For example, if this algorithm is applied to a Hamiltonian of the form, $H(x,p)=g(p,p)+\varepsilon h(x,p)\,,$ (5) where $g(p,p)=g^{ab}(x)p_{a}p_{b}$ is an undeformed dispersion relation, $h(x,p)$ is a function of spacetime points and momenta that depends on the model under consideration, and $\varepsilon$ is the perturbation parameter that is usually a function of the energy scale of the deformation (such as the Planck or quantum gravity length scale). As shown in Ref. Lobo:2020qoa , after the Legendre transformation, the equivalent action takes the form, $S[x]=m\int d\mu\sqrt{g(\dot{x},\dot{x})}\left(1-\varepsilon\frac{h(x,\bar{p}(x,\dot{x}))}{2m^{2}}\right)\,,$ (6) where $\bar{p}_{a}(x,\dot{x})=m{\dot{x}_{a}}/{\sqrt{g(\dot{x},\dot{x})}}$. In particular, when $h$ is a polynomial function of momenta as (the index is shifted: $n\rightarrow n+2$, in comparison with Ref. Lobo:2020qoa , such that now $n$ corresponds to the power of Planck length in the MDR), $h(x,p)=h^{\mu_{1}\mu_{2}....\mu_{n+2}}(x)p_{\mu_{1}}p_{\mu_{2}}...p_{\mu_{n+2}}\,,$ (7) and $\varepsilon=\ell^{n}$, one finds an action of the form, $S[x]=m\int d\mu\sqrt{g(\dot{x},\dot{x})}\left(1-(\ell m)^{n}\frac{h_{\mu_{1}\mu_{2}....\mu_{n+2}}(x)\dot{x}^{\mu_{1}}\dot{x}^{\mu_{2}}...\dot{x}^{\mu_{n+2}}}{2g(\dot{x},\dot{x})^{\frac{n+2}{2}}}\right)\,,$ (8) where we lowered the indices of $h$ with the components of $g$. The connection between the mechanics of free particle and geometry takes place when the above expression is identified with the arc-length functional, $s[x]$, of a given geometry, i.e., $s[x]=S[x]/m$. Such an identification makes sense if we want to state that the trajectories of free particles are extremizing curves or geodesics in a given geometry, it is related to the preservation of the equivalence principle even in this Planck-scale deformed scenario. In this case, the spacetime in which a particle propagates by a MDR is described by an arc-length functional that generalizes the one of Riemannian geometry and is given by a function $F(x,\dot{x})$ that is 1-homogeneous in the velocity $\dot{x}$, such that the arc-length is indeed parametrization invariant, as it must be: $s[x]=\int F(x,\dot{x})d\mu\,.$ (9) Actually, this is the kind of scenario envisaged by Riemann in his dissertation, and explored by Finsler, that emerges here quite naturally. There are some definitions of a pseudo-Finsler spacetime in the literature, but we rely on that given in Ref. Hohmann:2021zbt (the differences in comparison to other definitions are discussed in Ref. Hohmann:2021zbt ). First of all, we are going to work with a smooth manifold, $M$, endowed with a real valued positive function $L$ that takes values on the tangent bundle $TM$, described by coordinates $(x,y)$, where $\\{x^{\mu}\\}$ are spacetime coordinates and $\\{y^{\mu}\\}$ refer to vector or velocity coordinates. Actually, we shall need the slit tangent bundle $\widetilde{TM}=TM/\\{0\\}$, in which we remove the zero section, and we also need the projection $\pi:TM\rightarrow M$. A conic subbundle is a submanifold ${\cal D}\subset\widetilde{TM}$ such that $\pi({\cal D})=M$ and with the conic property that states that if $(x,y)\in{\cal D}\Rightarrow(x,\lambda y)\in{\cal D},\,\forall\lambda>0$. In a nutshell, a Finsler spacetime is a triple $(M,{\cal D},L)$, where $L:{\cal D}\rightarrow{\mathbb{R}}$ is a smooth function satisfying the conditions: 1. 1. positive 2-homogeneity: $L(x,\alpha y)=\alpha^{2}L(x,y),\,\forall\alpha>0$; 2. 2. at any $(x,y)\in{\cal D}$ and in any chart of $\widetilde{TM}$, the following Hessian (metric) is non-degenerate: $g_{\mu\nu}(x,y)=\frac{1}{2}\frac{\partial^{2}}{\partial y^{\mu}\partial y^{\nu}}L(x,y)\,;$ (10) 3. 3. the metric $g_{\mu\nu}$ has a Lorentzian signature. The function $L$ is actually the square of the Finsler function, $L(x,y)=F^{2}(x,y)$, and from it the Finsler arc-length is defined as given in Equation (9). Condition $1$ above guarantees that Equation (9) does not depend on the parametrization used to describe the curve and that using Euler’s theorem for homogeneous functions, this expression can be cast as $s[x]=\int\sqrt{g_{\mu\nu}(x,\dot{x})\dot{x}^{\mu}\dot{x}^{\nu}}d\mu\,.$ (11) From a coordinate transformation, $\displaystyle\tilde{x}^{\mu}=\tilde{x}^{\mu}(x)\,,$ (12) $\displaystyle\tilde{y}^{\mu}=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\nu}}y^{\nu}\,,$ (13) the functions $g_{\mu\nu}$ transform according to $\tilde{g}_{\mu\nu}(\tilde{x},\tilde{y})=\frac{\partial x^{\alpha}}{\partial\tilde{x}^{\mu}}\frac{\partial x^{\beta}}{\partial\tilde{x}^{\nu}}g_{\alpha\beta}(x,y)\,.$ (14) Due the property 14, $g_{\mu\nu}$ is referred here as the components of a distinguished tensor field (or $d$-tensor field) on the manifold $\widetilde{TM}$, which follows the notation adopted in Ref. Miron:1994nvt . The extremization of the arc-lenght functional (9) gives the following geodesic equation, $\frac{d^{2}x^{\mu}}{d\mu^{2}}+2G^{\mu}(x,\dot{x})=2\frac{dF}{d\mu}\frac{\partial F}{\partial\dot{x}^{\mu}}\,,$ (15) where $G^{\mu}=G^{\mu}(x,\dot{x})$ are the spray coefficients spray-finsler and are given in terms of the Christoffel symbols, $\gamma^{\alpha}_{\mu\nu}$, of the metric $g_{\mu\nu}$: $\displaystyle G^{\alpha}(x,\dot{x})=\frac{1}{2}\gamma^{\alpha}_{\mu\nu}(x,\dot{x})\dot{x}^{\mu}\dot{x}^{\nu}\,,$ (16) $\displaystyle\gamma^{\alpha}_{\mu\nu}(x,\dot{x})=\frac{1}{2}g^{\alpha\beta}\left(\frac{\partial g_{\mu\beta}}{\partial x^{\nu}}+\frac{\partial g_{\nu\beta}}{\partial x^{\mu}}-\frac{\partial g_{\mu\nu}}{\partial x^{\beta}}\right)\,.$ (17) If we choose the arc-length parametrization, i.e., the one in which $F=1$, we have a sourceless geodesic equation. This expression means that the trajectories generated by a MDR of the form $H(x,\dot{x})=m^{2}$ are, actually, geodesics of a Finsler metric. The presence of spray coefficients allows us to construct another quite a useful quantity, the so-called Cartan non-linear connection, given by (in this paper, we interchange the notation $\dot{x}\leftrightarrow y$ freely) $N^{\mu}{}_{\nu}(x,y)=\frac{\partial}{\partial y^{\nu}}G^{\mu}(x,y)\,,$ (18) that transforms according to $\tilde{N}^{\mu}{}_{\nu}=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\alpha}}\frac{\partial x^{\beta}}{\partial\tilde{x}^{\nu}}N^{\alpha}{}_{\beta}-\frac{\partial^{2}\tilde{x}^{\mu}}{\partial x^{\alpha}\partial x^{\beta}}\frac{\partial x^{\beta}}{\partial\tilde{x}^{\nu}}y^{\alpha}\,.$ (19) The introduction of this quantity allows us to introduce a useful basis of the tangent space of the tangent bundle at each point. In fact, since according to the coordinate transformation (12) and (13), the usual coordinate basis transforms as $\displaystyle\frac{\partial}{\partial\tilde{x}^{\mu}}=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}\frac{\partial}{\partial x^{\nu}}+\frac{\partial^{2}x^{\nu}}{\partial\tilde{x}^{\mu}\partial\tilde{x}^{\alpha}}\frac{\partial\tilde{x}^{\alpha}}{\partial x^{\beta}}y^{\beta}\frac{\partial}{\partial y^{\nu}}\,,$ (20) $\displaystyle\frac{\partial}{\partial\tilde{y}^{\mu}}=\frac{\partial y^{\nu}}{\partial\tilde{y}^{\mu}}\frac{\partial}{\partial y^{\nu}}\,.$ (21) In addition, a non-linear connection allows us to define the following frame: $\displaystyle\frac{\delta}{\delta x^{\mu}}=\delta_{\mu}=\frac{\partial}{\partial x^{\mu}}-N^{\nu}{}_{\mu}\frac{\partial}{\partial y^{\nu}}\,,$ (22) $\displaystyle\dot{\partial}_{\mu}=\frac{\partial}{\partial y^{\mu}}\,.$ (23) Due to the transformation properties of the non-linear connection, this basis transforms as $\displaystyle\tilde{\delta}_{\mu}=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}\delta_{\nu}\,,$ (24) $\displaystyle\tilde{\dot{\partial}}_{\mu}=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}\dot{\partial}_{\nu}\,.$ (25) This means that one is able to split the tangent space of the tangent bundle into horizontal, $HTM=\text{span}\\{\delta_{\mu}\\}$, and vertical, $VTM=\text{span}\\{\dot{\partial}_{\mu}\\}$, spaces, such that $T\widetilde{TM}=HTM\oplus VTM$ in each point $(x,y)$. Similarly, the same reasoning applies to the cotangent space; i.e., we split $T^{*}\widetilde{TM}=H^{*}TM\oplus V^{*}TM$ spanned as $H^{*}TM=\text{span}\\{dx^{\mu}\\}$ and $V^{*}TM=\text{span}\\{\delta y^{\mu}\\}$, where $\displaystyle\delta y^{\mu}=dy^{\mu}+N^{\mu}{}_{\nu}dx^{\nu}\,,$ (26) which transforms as $\displaystyle d\tilde{x}^{\mu}=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\nu}}dx^{\nu}\,,$ (27) $\displaystyle\delta\tilde{y}^{\mu}=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\nu}}\delta y^{\nu}\,.$ (28) Such a decomposition of the tangent and cotangent vector spaces implies that a vector $X$ and a $1$-form $\omega$ with horizontal and vertical terms can read as $\displaystyle X=X^{\mu}\delta_{\mu}+\dot{X}^{\mu}\dot{\partial}_{\mu}=X^{H}+X^{V}\,,$ (29) $\displaystyle\omega=\omega_{\mu}dx^{\mu}+\dot{\omega}_{\mu}\delta y^{\mu}=\omega^{H}+\omega^{V}\,.$ (30) Endowed with this basis, the metric $\mathbb{G}(x,y)$ of the configuration space is described by the so-called Sasaki-Matsumoto lift of the metric $g_{\mu\nu}$: $\mathbb{G}(x,y)=g_{\mu\nu}(x,y)dx^{\mu}\otimes dx^{\nu}+g_{\mu\nu}(x,y)\delta y^{\mu}\otimes\delta y^{\nu}\,.$ (31) ###### Definition III.1 A tensor field $T$ of type $(m+n,p+q)$ on the manifold $\widetilde{TM}$ is called a distinguished tensor field (or $d$-tensor field) if it has the property $\displaystyle T\left(\overset{1}{\omega},...,\overset{m}{\omega},\overset{1}{\tau},...,\overset{n}{\tau},\underset{1}{X},...,\underset{p}{X},\underset{1}{Y},...,\underset{q}{Y}\right)=T\left(\overset{1}{\omega}{}^{H},...,\overset{m}{\omega}{}^{H},\overset{1}{\tau}{}^{V},...,\overset{n}{\tau}{}^{V},\underset{1}{X}{}^{H},...,\underset{p}{X}{}^{H},\underset{1}{Y}{}^{V},...,\underset{q}{Y}{}^{V}\right).$ (32) This definition implies that one can write a $d$-tensor $T$ in the preferred frame as $\displaystyle T=T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}}\frac{\delta}{\delta x^{\mu_{1}}}$ $\displaystyle\otimes...\otimes\frac{\delta}{\delta x^{\mu_{m}}}\otimes\frac{\partial}{\partial y^{\nu_{1}}}\otimes...\otimes\frac{\partial}{\partial y^{\nu_{n}}}$ $\displaystyle\otimes dx^{\alpha_{1}}\otimes...\otimes dx^{\alpha_{p}}\otimes\delta y^{\beta_{1}}\otimes...\otimes\delta y^{\beta_{q}}\,,$ (33) and that it transforms according to the rule, $\displaystyle\widetilde{T}^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}}$ (34) $\displaystyle=\frac{\partial\tilde{x}^{\mu_{1}}}{\partial x^{\epsilon_{1}}}...\frac{\partial\tilde{x}^{\mu_{m}}}{\partial x^{\epsilon_{m}}}\frac{\partial\tilde{x}^{\nu_{1}}}{\partial x^{\lambda_{1}}}...\frac{\partial\tilde{x}^{\nu_{n}}}{\partial x^{\lambda_{n}}}\frac{\partial x^{\gamma_{1}}}{\partial\tilde{x}^{\alpha_{1}}}...\frac{\partial x^{\gamma_{p}}}{\partial\tilde{x}^{\alpha_{p}}}\frac{\partial x^{\rho_{1}}}{\partial\tilde{x}^{\beta_{1}}}...\frac{\partial x^{\rho_{q}}}{\partial\tilde{x}^{\beta_{q}}}T^{\epsilon_{1}..\epsilon_{n}\lambda_{1}...\lambda_{m}}{}_{\gamma_{1}...\gamma_{p}\rho_{1}...\rho_{q}}\,.$ An example of $d$-tensor field is the metric whose components are given by Equation (14). ### III.1 $N$-Linear Connection Given a linear connection, $D$, on the manifold $\widetilde{TM}$, if it preserves the parallelism of the horizontal and vertical spaces, i.e., if it can be written as $D_{\delta_{\nu}}\delta_{\mu}=L^{\alpha}_{\mu\nu}\delta_{\alpha}\,,\qquad D_{\delta_{\nu}}\dot{\partial}_{\alpha}=L^{\mu}_{\alpha\nu}\dot{\partial}_{\mu}\,,$ (35) $D_{\dot{\partial}_{\nu}}\delta_{\mu}=C^{\alpha}_{\mu\nu}\delta_{\alpha}\,,\qquad D_{\dot{\partial}_{\nu}}\dot{\partial}_{\mu}=C^{\alpha}_{\mu\nu}\dot{\partial}_{\alpha}\,,$ (36) then is called an $N$-linear connection. Let us consider a coordinate change; thus, the coefficients (35) and (36) transform as $\displaystyle\tilde{L}^{\alpha}_{\mu\nu}=\frac{\partial\tilde{x}^{\alpha}}{\partial x^{\beta}}\frac{\partial x^{\lambda}}{\partial\tilde{x}^{\mu}}\frac{\partial x^{\epsilon}}{\partial\tilde{x}^{\nu}}L^{\beta}_{\lambda\epsilon}+\frac{\partial^{2}x^{\beta}}{\partial\tilde{x}^{\mu}\partial\tilde{x}^{\nu}}\frac{\partial\tilde{x}^{\alpha}}{\partial x^{\beta}}\,,$ (37) $\displaystyle\tilde{C}^{\alpha}_{\mu\nu}=\frac{\partial\tilde{x}^{\alpha}}{\partial x^{\beta}}\frac{\partial x^{\lambda}}{\partial\tilde{x}^{\mu}}\frac{\partial x^{\epsilon}}{\partial\tilde{x}^{\nu}}C^{\beta}_{\lambda\epsilon}\,.$ (38) Endowed with these coefficients, the derivative of a $d$-tensor can be decomposed into a horizontal and a vertical parts, such that one can apply the covariant derivative of a tensor $T$ of type $(m+n,p+q)$ in the direction of a vector $X$ as a direction of a vector $X$ as $\displaystyle D_{X}T=D_{X^{H}}T+D_{X^{V}}T$ $\displaystyle=\left(T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}|\epsilon}X^{\epsilon}+T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}||\epsilon}\dot{X}^{\epsilon}\right)\frac{\delta}{\delta x^{\mu_{1}}}\otimes...\otimes\frac{\delta}{\delta x^{\mu_{m}}}$ $\displaystyle\otimes\frac{\partial}{\partial y^{\nu_{1}}}\otimes...\otimes\frac{\partial}{\partial y^{\nu_{n}}}\otimes dx^{\alpha_{1}}\otimes...\otimes dx^{\alpha_{p}}\otimes\delta y^{\beta_{1}}\otimes...\otimes\delta y^{\beta_{q}}\,,$ (39) where $\displaystyle T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}|\epsilon}$ (40) $\displaystyle=\frac{\delta}{\delta x^{\epsilon}}T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}}+L^{\mu_{1}}_{\gamma\epsilon}T^{\gamma...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}}+...-L^{\gamma}_{\alpha_{1}\epsilon}T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\gamma...\alpha_{p}\beta_{1}...\beta_{q}}\,,$ $\displaystyle T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}||\epsilon}$ (41) $\displaystyle=\frac{\partial}{\partial y^{\epsilon}}T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}}+C^{\mu_{1}}_{\gamma\epsilon}T^{\gamma...\mu_{m}\nu_{1}...\nu_{n}}{}_{\alpha_{1}...\alpha_{p}\beta_{1}...\beta_{q}}+...-C^{\gamma}_{\alpha_{1}\epsilon}T^{\mu_{1}...\mu_{m}\nu_{1}...\nu_{n}}{}_{\gamma...\alpha_{p}\beta_{1}...\beta_{q}}\,,$ and the property that the covariant derivative is linear in the direction $X$ is used. The triple $D\Gamma(N,L,C)$ describes the parallel transport and decomposition of the tangent and cotangent spaces of the tangent bundle into horizontal and vertical spaces. At this point, we need to comment on some remarkable $N$-linear connections that are considered in the literature. The first connection is the metrical Cartan connection, $C\Gamma(N^{\mu}{}_{\nu},L^{\alpha}_{\mu\nu},C^{\alpha}_{\mu\nu})$. In this case, $N^{\mu}{}_{\nu}$ is given by the canonical Cartan non-linear connection, defined by the spray coefficients (18). The coefficients $L^{\alpha}_{\mu\nu}$ and $C^{\alpha}_{\mu\nu}$ are given, respectively, by $\displaystyle L^{\alpha}_{\mu\nu}=\frac{1}{2}g^{\alpha\beta}\left(\frac{\delta g_{\mu\beta}}{\delta x^{\nu}}+\frac{\delta g_{\nu\beta}}{\delta x^{\mu}}-\frac{\delta g_{\mu\nu}}{\delta x^{\beta}}\right)\,,$ (42) $\displaystyle C^{\alpha}_{\mu\nu}=\frac{1}{2}g^{\alpha\beta}\left(\frac{\delta g_{\mu\beta}}{\delta y^{\nu}}+\frac{\delta g_{\nu\beta}}{\delta y^{\mu}}-\frac{\delta g_{\mu\nu}}{\delta y^{\beta}}\right)\,.$ (43) This connection is metrical (i.e., without non-metricity tensors) considering both horizontal and vertical covariant derivatives of the Finsler metric. Besides, the Berwald connection is given by the triple $B\Gamma(N^{\mu}{}_{\nu},\partial N^{\alpha}{}_{\mu}/\partial y^{\nu},0)$ and presents horizontal and vertical non-metricities. The Chern–Rund connection, $R\Gamma(N^{\mu}{}_{\nu},L^{\alpha}_{\mu\nu},0)$, is horizontally metrical, but represents vertical non-metricity. Additionally, the Hashiguchi connection, $H\Gamma(N^{\mu}{}_{\nu},\partial N^{\alpha}{}_{\mu}/\partial y^{\nu},C^{\alpha}_{\mu\nu})$, represents horizontal non-metricity, but it is vertically metrical. In these expressions, $N$ is the canonical Cartan non- linear connection (18), $L$ is given by Equation (42), and $C$ is given by Equation $\eqref{c-cartan}$. ### III.2 Symmetries Geometrical language naturally realizes the concept of symmetry of physical equations. General relativity given in terms of Riemannian geometry encompasses the invariance under general coordinate transformations, and the isometries of the Minkowski space describe the Poincaré transformations (actually, one can further apply this technique for maximally symmetric spaces, including de Sitter and anti-de Sitter ones). Finsler geometry, as we have been using, allows us to go beyond this scope and to define deformed Lorentz/Poincaré transformations that present Planck scale corrections even in the presence of a local modified dispersion relation. One can see how this will naturally emerge, since the invariance of the arc-length (9) is compatible with the invariance of the action in the Hamiltonian formulation (4), from which such an arc-length was derived. This idea was firstly noticed in Ref. Girelli:2006fw and later explicitly explored in Refs. Amelino- Camelia:2014rga ; Lobo:2016xzq . The master equation for this purpose is the one that follows from the invariance of the Finslerian interval $ds^{2}$, as done in Appendix A of Ref. Amelino-Camelia:2014rga . From this invariance, the Finslerian killing equation for the killing vector was found, with components $\xi^{\alpha}$, which should be solved in order to derive the deformed symmetries in the DSR context, $\xi^{\alpha}\partial_{\alpha}g_{\mu\nu}+g_{\alpha\nu}\partial_{\mu}\xi^{\alpha}+g_{\mu\alpha}\partial_{\nu}\xi^{\alpha}+y^{\alpha}\partial_{\alpha}\xi^{\beta}\dot{\partial}_{\beta}g_{\mu\nu}=0\,.$ (44) ### III.3 Finsler–q-de Sitter (Tangent Bundle Case) As an example that presents a non-trivial non-linear connection, we shall consider the case of a Finsler geometry inspired by the so-called $q$-de Sitter deformed relativity. This case has been previously studied in the literature, e.g., in Refs. Barcaroli:2015eqe ; Letizia:2016lew ; Lobo:2016xzq ; Lobo:2016lxm , and can be described by an algebra that deforms the one of Poincaré in a way that gives the de Sitter symmetry when a quantum gravity parameter goes to zero, and on the other hand, gives the so-called $\kappa$-Poincaré algebra (that deforms the Poincaré one by an energy scale parameter, supposedly the Planck energy) when the de Sitter curvature parameter goes to zero. Therefore, it corresponds to an authentic realization of a deformed relativity scenario, even in the presence of what can be interpreted as spacetime curvature. In this subsection, we initially consider results that were originally presented in Ref. Letizia:2016lew in $1+1$ dimensions. The MDR related to this algebra (in a given basis) can be perturbed to first order in the Planck length and de Sitter curvature parameters $\ell$ and $H$, respectively, as ${\cal H}(x,p)=p_{0}^{2}-p_{1}^{2}(1+\ell p_{0})(1-2Hx^{0})\,.$ (45) By using the action given by Equation (4) and the algorithm that follows it, the following Finsler function can be obtained: $F(x,\dot{x})=\sqrt{(\dot{x}^{0})^{2}-(1-2Hx^{0})(\dot{x}^{1})^{2}}+\ell\frac{m}{2}\frac{(1-2Hx^{0})\dot{x}^{0}(\dot{x}^{1})^{2}}{(\dot{x}^{0})^{2}-(1-2Hx^{0})(\dot{x}^{1})^{2}}\,,$ (46) from which the Finsler metric can be found from Equation (10): $\displaystyle g_{\mu\nu}^{F}(x,\dot{x})=\begin{pmatrix}1+\frac{3a^{4}m\ell\dot{x}^{0}(\dot{x}^{1})^{4}}{2[(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}]^{5/2}}&\frac{m\ell a^{4}(\dot{x}^{1})^{3}[a^{2}(\dot{x}^{1})^{2}-4(\dot{x}^{0})^{2}]}{2[(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}]^{5/2}}\\\ \frac{m\ell a^{4}(\dot{x}^{1})^{3}[a^{2}(\dot{x}^{1})^{2}-4(\dot{x}^{0})^{2}]}{2[(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}]^{5/2}}&-a^{2}+\frac{m\ell a^{2}(\dot{x}^{0})^{3}[2(\dot{x}^{0})^{2}+a^{2}(\dot{x}^{1})^{2}]}{2[(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}]^{5/2}}\end{pmatrix}\,,$ (47) where $a=a(t)=e^{Ht}=1+Ht+{\cal O}(H^{2})$ (in this paper, the terms that grow with higher orders of $H$ and $\ell$ are discarded). The geodesic equation is found from the extremization of the Finsler arc-length defined by $F$, from which Christoffel symbols and spray coefficients can be calculated. Actually, the $\gamma^{\alpha}_{\mu\nu}(x,\dot{x})$ are given, for an arbitrary parametrization, by the set of Equations (44) of Ref. Letizia:2016lew , from which the spray coefficients are given by $\displaystyle G^{0}(x,\dot{x})=\frac{1}{8}a^{2}H(\dot{x}^{1})^{2}$ $\displaystyle\left[4-\frac{\ell m\dot{x}^{0}}{\left[(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}\right]^{7/2}}\left(-28a^{6}(\dot{x}^{1})^{6}+12a^{2}(\dot{x}^{0})^{4}(\dot{x}^{1})^{2}\right.\right.$ $\displaystyle\left.\left.+a^{2}\left(17a^{2}+28\right)(\dot{x}^{0})^{2}(\dot{x}^{1})^{4}+16(\dot{x}^{0})^{6}\right)\right]\,,$ (48) $\displaystyle G^{1}(x,\dot{x})=H\dot{x}^{0}\dot{x}^{1}+\ell$ $\displaystyle\left[\frac{a^{2}Hm(\dot{x}^{1})^{3}\left(a^{6}(\dot{x}^{1})^{6}-6a^{4}(\dot{x}^{0})^{2}(\dot{x}^{1})^{4}+3a^{2}(\dot{x}^{0})^{4}(\dot{x}^{1})^{2}-28(\dot{x}^{0})^{6}\right)}{4\left((\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}\right)^{7/2}}\right]\,.$ (49) As can be seen, these coefficients are $2$-homogeneous in the velocities, as expected. The Cartan non-linear connection coefficients read: $\displaystyle N^{0}{}_{0}(x,\dot{x})=$ $\displaystyle\frac{H\ell m(\dot{x}^{1})^{4}\left(-28(\dot{x}^{1})^{6}-33(\dot{x}^{1})^{4}(\dot{x}^{0})^{2}+240(\dot{x}^{1})^{2}(\dot{x}^{0})^{4}+136(\dot{x}^{0})^{6}\right)}{8\left((\dot{x}^{0})^{2}-(\dot{x}^{1})^{2}\right)^{9/2}}\,,$ (50) $\displaystyle N^{0}{}_{1}(x,\dot{x})=$ $\displaystyle H\dot{x}^{1}-\frac{H\ell m\dot{x}^{1}\dot{x}^{0}}{8\left((\dot{x}^{0})^{2}-(\dot{x}^{1})^{2}\right)^{9/2}}\left(28(\dot{x}^{1})^{8}-179(\dot{x}^{1})^{6}(\dot{x}^{0})^{2}+306(\dot{x}^{1})^{4}(\dot{x}^{0})^{4}\right.$ $\displaystyle\left.+128(\dot{x}^{1})^{2}(\dot{x}^{0})^{6}+32(\dot{x}^{0})^{8}\right),$ $\displaystyle N^{1}{}_{0}(x,\dot{x})=$ $\displaystyle H\dot{x}^{1}+\frac{H\ell m(\dot{x}^{1})^{3}\dot{x}^{0}\left(5(\dot{x}^{1})^{6}+18(\dot{x}^{1})^{4}(\dot{x}^{0})^{2}+159(\dot{x}^{1})^{2}(\dot{x}^{0})^{4}+28(\dot{x}^{0})^{6}\right)}{4\left((\dot{x}^{0})^{2}-(\dot{x}^{1})^{2}\right)^{9/2}}\,,$ (51) $\displaystyle N^{1}{}_{1}(x,\dot{x})=$ $\displaystyle H\dot{x}^{0}-\frac{H\ell m(\dot{x}^{1})^{2}\left(2(\dot{x}^{1})^{8}-9(\dot{x}^{1})^{6}(\dot{x}^{0})^{2}+36(\dot{x}^{1})^{4}(\dot{x}^{0})^{4}+97(\dot{x}^{1})^{2}(\dot{x}^{0})^{6}+84(\dot{x}^{0})^{8}\right)}{4\left((\dot{x}^{0})^{2}-(\dot{x}^{1})^{2}\right)^{9/2}}\,,$ (52) where the worldlines are autoparallel curves of this non-linear connection. Let us note that some terms of the connection are only present due to the coupling between the spacetime curvature parameter, $H$, and the one that gives a non-trivial velocity space, $\ell$. Some curvature-triggered effects in quantum gravity have been recently analyzed Amelino-Camelia:2020bvx . Endowed with these coefficients, the preferred frames that induce the horizontal and vertical decomposition can be immediately found, in addition the $N$-linear connection coefficients $L^{\alpha}_{\mu\nu}$ and $C^{\alpha}_{\mu\nu}$, as discussed in Section II. Till now, only kinematical properties were discussed, but the choice of the given connection should be given either by physical conditions imposed on the dynamics of the spacetime or by possible effective gravitational field equations for a quantum configuration space. To finalize this Section, let us discuss the symmetries of the spacetime. A deep analysis of the killing vectors of the $H\rightarrow 0$ limit of this Finsler framework was carried out in Ref. Barcaroli:2015eqe . Even in that simplified scenario, the equations are quite lengthy which we omit here. However, some properties should be mentioned. Firstly, the transformations generated by the killing vectors seem to not exactly preserve the line element, but contribute with a term that is given by a total derivative in the action parameter; therefore, the kinematical results of these two line elements coincide. Secondly, the results found are compatible with the $\kappa$-Poincaré scenario that inspired this approach. From the Finsler perspective, it is possible to derive more general results, but they reduces to those of the bicrossproduct basis of $\kappa$-Poincaré by an appropriate choice of free functions and parameters. The third point is that a finite version of transformations that preserve the $\kappa$-Poincaré dispersion relation was recently made in Ref. Lobo:2021yem through an alternative approach, which does not rely on the killing vectors but is determined by the Finsler function and the definition of momentum (explored in Section IV below); however, a complete integration of the finite isometry and a comparison between these approaches is still missing in the literature. To finalize, the case of $H\neq 0$ was investigated in Ref. Lobo:2016xzq , but in conformal coordinates (which are not the ones that are considered in this application), and was not done in so much detail as the flat case, but a generator of the corresponding curved boost transformation was made explicit in Equation (25) of Ref. Lobo:2016xzq . ## IV The Cotangent Bundle Version of Finsler Geometry As was discussed in Ref. Girelli:2006fw , by mapping the velocity of the particle to its momentum, it is possible to find the version of the Finsler metric defined in the cotangent bundle or phase space. Already from the definition of the 4-momentum, $p_{\mu}=m\frac{\partial F}{\partial y^{\mu}}\,,$ (53) when it is possible to invert this expression to find $y=y(p)$, one can substitute this result in the Finsler metric as $h^{F}_{\mu\nu}(x,p)=g_{\mu\nu}^{F}(x,y(p))$. This metric is defined on the slit cotangent bundle, $\widetilde{T^{*}M}=T^{*}M/\\{0\\}$, where we also remove the zero section in each spacetime point for the same technical reasons as discussed in Section III above. Since the quantities are now defined in the cotangent bundle, we need to also address some issues that were raised in Section III concerning the tangent bundle. This Section’s notation is applied according to Ref. Miron:1994nvt . For instance, under a change of coordinates, the spacetime and momentum variables transformed according to $\displaystyle\tilde{x}^{\mu}$ $\displaystyle=\tilde{x}^{\mu}(x)\,,$ (54) $\displaystyle\tilde{p}_{\mu}$ $\displaystyle=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}p_{\nu}\,,$ (55) which means that the frame $(\partial/\partial x^{\mu},\partial/\partial p_{\nu})$ transforms as $\displaystyle\frac{\partial}{\partial\tilde{x}^{\mu}}$ $\displaystyle=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}\frac{\partial}{\partial x_{\nu}}+\frac{\partial p_{\nu}}{\partial\tilde{x}^{\mu}}\frac{\partial}{\partial p_{\nu}}\,,$ (56) $\displaystyle\frac{\partial}{\partial\tilde{p}_{\mu}}$ $\displaystyle=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\nu}}\frac{\partial}{\partial p_{\nu}}\,.$ (57) On the other hand, the natural coframe $(dx^{\mu},dp_{\nu})$ changes as $\displaystyle d\tilde{x}^{\mu}$ $\displaystyle=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\nu}}dx^{\nu}\,,$ (58) $\displaystyle d\tilde{p}_{\mu}$ $\displaystyle=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}dp_{\nu}+\frac{\partial^{2}x^{\nu}}{\partial\tilde{x}^{\mu}\partial\tilde{x}^{\lambda}}p_{\nu}d\tilde{x}^{\lambda}\,.$ (59) Simlarly to that in Section III, the presence of a nonlinear connection, $O_{\mu\nu}$, allows one to split the cotangent bundle into a horizontal and a vertical subbundle. Inspired by the consideration of the Hamilton case considered in Ref. Barcaroli:2015xda (discussed below), we propose the following dual non-linear connection (constructed in Appendix A): $O_{\mu\nu}(x,p)=-m\left[N^{\alpha}{}_{\mu}\frac{(g_{\alpha\nu}-p_{\alpha}p_{\nu}/m^{2})}{F}-\partial_{\mu}\dot{\partial}_{\nu}F\right]\Bigg{|}_{(x,y(p))}\,,$ (60) where $p=p(y)$ is the kinematical map defined by Equation (53). By construction, these symbols have the transformation properties of a nonlinear connection, $\tilde{O}_{\mu\nu}=\frac{\partial x^{\lambda}}{\partial\tilde{x}^{\mu}}\frac{\partial x^{\epsilon}}{\partial\tilde{x}^{\nu}}O_{\lambda\epsilon}+\frac{\partial^{2}x^{\beta}}{\partial\tilde{x}^{\mu}\partial\tilde{x}^{\nu}}p_{\beta}\,.$ (61) Endowed with a nonlinear connection $O_{\mu\nu}$, one can decompose the tangent bundle of the cotangent bundle by the Whitney sum in each point $T_{u}\widetilde{T^{*}M}=O_{u}\oplus V_{u},\,\forall u\in\widetilde{T^{*}M}$. The subbundle $O_{u}$ is called horizontal space and is spanned by the frame, $\frac{\delta}{\delta x^{\mu}}=\delta_{\mu}=\frac{\partial}{\partial x^{\mu}}+O_{\mu\nu}\frac{\partial}{\partial p_{\nu}}\,,$ (62) and the subbundle $V_{u}$ is called vertical space and is spanned by the frame in each point of $\widetilde{T^{*}M}$: $\bar{\partial}^{\mu}=\frac{\partial}{\partial p_{\mu}}\,,$ (63) such that $T_{u}\widetilde{T^{*}M}=\text{span}\\{\delta_{\mu},\bar{\partial}^{\nu}\\}$. The transformation properties of the nonlinear connection are implied in the following rule for transforming this basis: $\displaystyle\frac{\delta}{\delta\tilde{x}^{\mu}}=\tilde{\delta}_{\mu}=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}\frac{\delta}{\delta x^{\nu}}=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}\delta_{\nu}\,,$ (64) $\displaystyle\frac{\partial}{\partial\tilde{p}_{\mu}}=\tilde{\bar{\partial}}^{\mu}=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\nu}}\frac{\partial}{\partial p_{\nu}}=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\nu}}\bar{\partial}^{\nu}\,.$ (65) Equivalently, with the nonlinear connection, we can decompose the cotangent space $T^{*}_{u}\widetilde{T^{*}M}=\text{span}\\{dx^{\mu},\delta p_{\nu}\\}$, where $\delta p_{\mu}=dp_{\mu}-O_{\nu\mu}dx^{\nu}\,.$ (66) Therefore, the dual basis transforms as $\displaystyle d\tilde{x}^{\mu}$ $\displaystyle=\frac{\partial\tilde{x}^{\mu}}{\partial x^{\nu}}dx^{\nu}\,,$ (67) $\displaystyle\delta\tilde{p}_{\mu}$ $\displaystyle=\frac{\partial x^{\nu}}{\partial\tilde{x}^{\mu}}\delta p_{\nu}\,.$ (68) Similarly to what has been done for the tangent bundle case, such a decomposition allows us to express a vector and a $1$-form via horizontal and vertical components, where now, the vertical component is considered along momenta instead of velocities, $\displaystyle X=X^{\mu}\delta_{\mu}+\bar{X}_{\mu}\bar{\partial}^{\mu}=X^{H}+X^{V}\,,$ (69) $\displaystyle\omega=\omega_{\mu}dx^{\mu}+\bar{\omega}^{\mu}\delta p_{\mu}=\omega^{H}+\omega^{V}\,.$ (70) Besides, the metric $\mathbb{H}(x,p)$ of the configuration space is defined as follows. Given a metric $h^{\mu\nu}(x,p)$, and the nonlinear connection $O_{\mu\nu}(x,p)$, the quantum phase space presents metrical properties given by the tensor, $\mathbb{H}(x,p)=h_{\mu\nu}(x,p)dx^{\mu}\otimes dx^{\nu}+h^{\mu\nu}(x,p)\delta p_{\mu}\otimes\delta p_{\nu}\,.$ (71) We refer to the tensor $\mathbb{H}$ as the $N$-lift to $\widetilde{T^{*}M}$ of the metric $h_{\mu\nu}$. The map between $y$ and $p$ cannot be done, in general, involving quantities that are parametrization-dependent because $p$ itself is parametrization-invariant, whereas $y$ is not. That is why one can only assume $y(p)$ for the definition of the metric $h^{F}_{\mu\nu}$. Endowed with these quantities, one can just extend the definition of $d$-tensors III.1 to the cotangent case, in which one only needs to consider the use of the nonlinear connection $O_{\mu\nu}$ and the adapted basis defined in this Section. The above implies that a $d$-tensor $T$ of type $(m+q,n+p)$ can be rewritten in the preferred basis as $\displaystyle T=T^{\mu_{1}...\mu_{m}}{}_{\nu_{1}...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}$ $\displaystyle\frac{\delta}{\delta x^{\mu_{1}}}\otimes...\otimes\frac{\delta}{\delta x^{\mu_{m}}}\otimes\frac{\partial}{\partial p_{\nu_{1}}}\otimes...\otimes\frac{\partial}{\partial p_{\nu_{n}}}$ $\displaystyle\otimes dx^{\alpha_{1}}\otimes...\otimes dx^{\alpha_{p}}\otimes\delta p_{\beta_{1}}\otimes...\otimes\delta p_{\beta_{q}}\,,$ (72) whose components transform according to usual linear transformation rules, as the one of Equation (34). ### IV.1 N-Linear Connection Equivalently, the notion of differentiation can be defined in the cotangent bundle through the $N$-linear connection $D$, which has the following coefficients in the frame $(\delta_{\mu},\bar{\partial}^{\nu})$ (see Theorem 4.9.1 in Ref. Miron:1994nvt ): $\displaystyle D_{\delta_{\nu}}\delta_{\mu}=H^{\alpha}_{\mu\nu}\delta_{\alpha}\,,\qquad D_{\delta_{\nu}}\bar{\partial}^{\mu}=-H^{\mu}_{\alpha\nu}\bar{\partial}^{\alpha}\,,$ (73) $\displaystyle D_{\bar{\partial}^{\nu}}\delta_{\mu}=C^{\alpha\nu}_{\mu}\delta_{\alpha}\,,\qquad D_{\bar{\partial}^{\nu}}\bar{\partial}^{\mu}=-C_{\alpha}^{\mu\nu}\bar{\partial}^{\alpha}\,.$ (74) Otherwise, in the frame $(dx^{\mu},\delta p_{\nu})$ one has (see Proposition 4.9.1 in Ref. Miron:1994nvt ) $\displaystyle D_{\delta_{\nu}}dx^{\mu}=-H^{\mu}_{\alpha\nu}dx^{\alpha}\,,\qquad D_{\delta_{\nu}}\delta p_{\mu}=H^{\alpha}_{\mu\nu}\delta p_{\alpha}\,,$ (75) $\displaystyle D_{\bar{\partial}^{\nu}}dx^{\mu}=-C^{\mu\nu}_{\alpha}dx^{\alpha}\,,\qquad D_{\bar{\partial}^{\nu}}\delta p_{\mu}=C_{\mu}^{\alpha\nu}\delta p_{\alpha}\,.$ (76) Considering a $N$-linear connection $D$ with set of coefficients, $D\Gamma(N)=(H^{\alpha}_{\mu\nu},C^{\alpha}_{\mu\nu})$, one can add to it a nonlinear connection, $N_{\mu\nu}$, that is in general independent of the coefficients of $D$, such that the new set is $D\Gamma=(N_{\mu\nu},H^{\alpha}_{\mu\nu},C^{\alpha}_{\mu\nu})$. For this reason, the derivative of a $d$-tensor in the cotangent bundle presents similar usual rules for dealing with up and down indices: $\displaystyle T^{\mu_{1}...\mu_{m}}{}_{\nu_{1}...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}{}_{|\epsilon}$ (77) $\displaystyle=\frac{\delta}{\delta x^{\epsilon}}T^{\mu_{1}...\mu_{m}}{}_{\nu_{1}...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}+H^{\mu_{1}}_{\gamma\epsilon}T^{\gamma...\mu_{m}}{}_{\nu_{1}...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}+...-H^{\gamma}_{\nu_{1}\epsilon}T^{\mu_{1}...\mu_{m}}{}_{\gamma...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}\,,$ $\displaystyle T^{\mu_{1}...\mu_{m}}{}_{\nu_{1}...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}{}^{||\epsilon}$ (78) $\displaystyle=\frac{\partial}{\partial p_{\epsilon}}T^{\mu_{1}...\mu_{m}}{}_{\nu_{1}...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}{}+C^{\mu_{1}\epsilon}_{\gamma}T^{\gamma...\mu_{m}}{}_{\nu_{1}...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}{}+...-C^{\gamma\epsilon}_{\nu_{1}}T^{\mu_{1}...\mu_{m}}{}_{\gamma...\nu_{n}\alpha_{1}...\alpha_{p}}{}^{\beta_{1}...\beta_{q}}{}\,.$ Let us note that from the kinematical map relating velocities and momenta, the coefficients $H^{\alpha}_{\mu\nu}(x,y(p))$ and $C^{\alpha}_{\mu\nu}(x,y(p))$ can be found as been parametrization-invariant. ### IV.2 Finsler–q-de Sitter (Cotangent Bundle Case) Here, we again consider the $q$-de Sitter-inspired case. Then, using the Finsler function (46), the momentum is given by Equation: (53) $\displaystyle p_{0}=\frac{m\dot{x}^{0}}{\sqrt{(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}}}-\ell\frac{m^{2}a^{2}(\dot{x}^{1})^{2}(a^{2}(\dot{x}^{1})^{2}+(\dot{x}^{0})^{2})}{2[(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2})]^{2}}\,,$ (79) $\displaystyle p_{1}=-\frac{ma^{2}\dot{x}^{1}}{\sqrt{(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}}}+\ell\frac{m^{2}a^{2}(\dot{x}^{0})^{3}\dot{x}^{1}}{((\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}))^{2}}\,,$ (80) which furnishes a helpful expression that is throughout this Section and is a common trick when trying to find momentum-dependent quantities from the Finsler approach: $\displaystyle\frac{m\dot{x}^{0}}{\sqrt{(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}}}=p_{0}+\ell\frac{a^{-2}(p_{1})^{2}(a^{-2}(p_{1})^{2}+(p_{0})^{2})}{2m^{2}}\,,$ (81) $\displaystyle\frac{ma\dot{x}^{1}}{\sqrt{(\dot{x}^{0})^{2}-a^{2}(\dot{x}^{1})^{2}}}=-a^{-1}p_{1}\left(1+\ell\frac{(p_{0})^{3}}{m^{2}}\right)\,.$ (82) The above expressions allow us to express the Finsler metric through its momentum dependence: $\displaystyle g_{\mu\nu}^{F}(x,\dot{x}(p))=h_{\mu\nu}^{F}(x,p)=\begin{pmatrix}1+\frac{3\ell p_{0}(p_{1})^{4}}{m^{4}}&-\frac{\ell a(p_{1})^{3}[(p_{1})^{2}-4(p_{0})^{2}]}{2m^{4}}\\\ -\frac{\ell a(p_{1})^{3}[(p_{1})^{2}-4(p_{0})^{2}]}{2m^{4}}&-a^{2}+\frac{\ell a^{2}(p_{0})^{3}[2(p_{0})^{2}+(p_{1})^{2}]}{m^{4}}\end{pmatrix}\,,$ (83) which can be called a ”Finsler-rainbow metric.” One can also find the induced non-linear connection in the cotangent bundle through the definition (60) to read as $\displaystyle O_{00}(x,p)=$ $\displaystyle-\frac{H\ell(p_{1})^{2}}{8m^{10}}\left[4(p_{0})^{10}+44(p_{0})^{8}(p_{1})^{2}+190(p_{0})^{6}(p_{1})^{4}-196(p_{0})^{4}(p_{1})^{6}\right.$ $\displaystyle\left.+31(p_{0})^{2}(p_{1})^{8}+32(p_{1})^{10}\right]\,,$ (84) $\displaystyle O_{01}(x,p)=$ $\displaystyle Hp_{1}-\frac{\ell Hp_{0}p_{1}}{8m^{10}}\left[-4m^{8}(p_{0})^{2}+8(p_{0})^{10}+32(p_{0})^{8}(p_{1})^{2}+206(p_{0})^{6}(p_{1})^{4}\right.$ $\displaystyle\left.-212(p_{0})^{4}(p_{1})^{6}+43(p_{0})^{2}(p_{1})^{8}+28(p_{1})^{10}\right]\,,$ (85) $\displaystyle O_{10}(x,p)=$ $\displaystyle Hp_{1}-\frac{H\ell p_{0}p_{1}}{8m^{10}}\left(-4m^{8}(p_{0})^{2}+4(p_{0})^{10}+140(p_{0})^{8}(p_{1})^{2}+2(p_{0})^{6}(p_{1})^{4}\right.$ $\displaystyle\left.-106(p_{0})^{4}(p_{1})^{6}+61(p_{0})^{2}(p_{1})^{8}+4(p_{1})^{10}\right)\,,$ (86) $\displaystyle O_{11}(x,p)=$ $\displaystyle Hp_{0}+\frac{H\ell}{8m^{10}}\left(4(p_{0})^{2}(p_{1})^{2}\left(m^{8}+3(p_{1})^{8}\right)+8m^{8}(p_{1})^{4}-8(p_{0})^{12}-124(p_{0})^{10}(p_{1})^{2}\right.$ $\displaystyle\left.-30(p_{0})^{8}(p_{1})^{4}+138(p_{0})^{6}(p_{1})^{6}-89(p_{0})^{4}(p_{1})^{8}-4(p_{1})^{12}\right)\,.$ (87) From these expressions, one can construct the decomposition of the tangent and cotangent spaces of the cotangent bundle into horizontal and vertical parts, accordingly. ## V Geometry of the Cotangent Bundle: Hamilton Geometry Besides the Finsler geometry, another interesting proposal for building a natural geometry for propagation of particles that probe a modified dispersion relation consists of the so-called Hamilton geometry. In this case, different from the Finsler geometry, we start with a geometric structure defined in the cotangent bundle (the definitions used in this metric follow that in the book Miron:1994nvt and in papers Barcaroli:2015xda ; Barcaroli:2016yrl ; Barcaroli:2017gvg ; Pfeifer:2018pty ). A Hamilton space is a pair, $(M,H(x,p))$, where $M$ is a smooth manifold and $H:T^{*}M\rightarrow{\mathbb{R}}$ is a continuous function on the cotangent bundle that satisfies the following properties: 1. 1. $H$ is smooth on the manifold $\widetilde{T^{*}M}$; 2. 2. the Hamilton metric, $h_{H}$, with components, $h_{H}^{\mu\nu}(x,p)=\frac{1}{2}\frac{\partial}{\partial p_{\mu}}\frac{\partial}{\partial p_{\nu}}H(x,p)\,,$ (88) is nondegenerate. Since one does not have an arc-length functional, worldlines as extremizing curves are an absent concept in this approach. Instead, the equations of motion of a particle that obeys a given Hamiltonian are given by the Hamilton equations of motion: $\displaystyle\dot{x}^{\mu}$ $\displaystyle=\frac{\partial H}{\partial p_{\mu}}\,,$ (89) $\displaystyle\dot{p}_{\mu}$ $\displaystyle=-\frac{\partial H}{\partial\dot{x}^{\mu}}\,.$ (90) Since this is just another metric structure defined in the cotangent bundle, the same results regarding the tools for coordinate transformations given by Equation (54) are applicable here. As the case of Hamiltonian mechanics, the definition of Poisson brackets is useful enough for our purposes. For two real valued functions $F(x,p)$ and $G(x,p)$, their Poisson brackets are given in Barcaroli:2015xda (the geometry of the cotangent bundle with deformed Hamiltonian can also be described with the language of symplectic geometry, which is reviewed in Ref. Nozari:2014qja ): $\\{F(x,p),G(x,p)\\}=\partial_{\mu}F\bar{\partial}^{\mu}G-\partial_{\mu}G\bar{\partial}^{\mu}F\,.$ (91) As above, in order to divide the tangent and cotangent spaces of the cotangent bundle into horizontal and vertical spaces, a non-linear connection is necessary, and the canonical choice is given in Theorem 5.5.1 of Ref. Miron:1994nvt and Definition 2 of Ref. Barcaroli:2015xda as $O_{\mu\nu}(x,p)=\frac{1}{4}(\\{h_{\mu\nu}^{H},H\\}+h^{H}_{\mu\alpha}\partial_{\nu}\bar{\partial}^{\alpha}H+h^{H}_{\nu\alpha}\partial_{\mu}\bar{\partial}^{\alpha}H)\,,$ (92) where $h_{\mu\nu}^{H}$ is the inverse of the metric $h^{\mu\nu}_{H}$. This non-linear connection allows us to use the basis $\delta_{\mu}=\partial_{\mu}+O_{\mu\nu}\bar{\partial}^{\nu}$ and $\bar{\partial}^{\mu}$ as a special basis of $T_{(x,p)}\widetilde{T^{*}M}$, and to use the basis $dx^{\mu}$ and $\delta p_{\mu}=dp_{\mu}-O_{\nu\mu}dx^{\nu}$ as a special basis of $T^{*}_{(x,p)}\widetilde{T^{*}M}$, which transforms according to Equations (64), (65) and (67), (68). Endowed with these coefficients, following Theorem 5.6.1 of Ref. Miron:1994nvt , there exists a unique $N$-linear connection $D\Gamma(O)=(H^{\alpha}_{\mu\nu},C_{\alpha}^{\mu\nu})$ such that: 1. 1. $O_{\mu\nu}$ is the canonical non-linear connection; 2. 2. the metric $h^{\mu\nu}_{H}$ is $h$-covariant constant (no horizontal non- metricity): $D_{\delta_{\alpha}}h_{H}^{\mu\nu}=0\,;$ (93) 3. 3. the metric $h^{\mu\nu}_{H}$ is $v$-covariant constant (no vertical non- metricity): $D{\bar{\partial}^{\alpha}}h_{H}^{\mu\nu}=0\,;$ (94) 4. 4. $D\Gamma(N)$ is horizontally torsion free: $T^{\alpha}_{\ \mu\nu}=H^{\alpha}_{\mu\nu}-H^{\alpha}_{\nu\mu}=0\,;$ (95) 5. 5. $D\Gamma(N)$ is vertically torsion free: $S_{\alpha}^{\ \mu\nu}=C_{\alpha}^{\mu\nu}-C_{\alpha}^{\nu\mu}=0\,;$ (96) 6. 6. the triple $(O_{\mu\nu},H^{\alpha}_{\mu\nu},C_{\alpha}^{\mu\nu})$ has coefficients given by $\displaystyle O_{\mu\nu}(x,p)=\frac{1}{4}(\\{h_{\mu\nu}^{H},H\\}+h^{H}_{\mu\alpha}\partial_{\nu}\bar{\partial}^{\alpha}H+h^{H}_{\nu\alpha}\partial_{\mu}\bar{\partial}^{\alpha}H)\,,$ (97) $\displaystyle H_{\alpha}^{\mu\nu}=\frac{1}{2}h_{H}^{\alpha\beta}(\delta_{\mu}h^{H}_{\beta\nu}+\delta_{\nu}h^{H}_{\beta\mu}-\delta_{\beta}h^{H}_{\mu\nu})\,,$ (98) $\displaystyle C_{\alpha}^{\mu\nu}=-\frac{1}{2}h_{\alpha\beta}^{H}\bar{\partial}^{\mu}h_{H}^{\beta\nu}\,.$ (99) This is called a Cartan $N$-linear covariant derivative. Equivalently, the notion of $d$-tensors and their derivatives discussed in Section IV.1 are applicable. ### V.1 Symmetries Hamilton geometry also allows one to encompass a DSR language, as was the case for Finsler geometry discussed in Section III.2. However, its realization does not come from the invariance of an interval $ds^{2}$, since one does not have it, but from the invariance of the Hamiltonian function $H(x,p)$. The approach, which we highlight here, was done starting from Definition 4 of Section II-D of Ref. Barcaroli:2015xda . In a Hamilton space $(M,H)$ with manifold $M$, and Hamiltonian $H$, let $X=\xi^{\mu}\partial_{\mu}$ be a vector field in the basis manifold $M$ and $X^{C}=\xi^{\mu}\partial_{\mu}-p_{\nu}\partial_{\mu}\xi^{\nu}\bar{\partial}^{\mu}$ be the so-called complete lift of $X$ to $\widetilde{T^{*}M}$. A symmetry of the Hamiltonian is a transformation generated by $X^{C}$, whose components satisfy $X^{C}(H)\xi^{\mu}\partial_{\mu}H-p_{\nu}\partial_{\mu}\xi^{\nu}\bar{\partial}^{\mu}H=0\,.$ (100) If one derivates this expression twice with respect to momenta, one gets the following result: $0=\frac{1}{2}\bar{\partial}^{\mu}\bar{\partial}^{\nu}X^{C}(H)=\xi^{\alpha}\partial_{\alpha}h_{H}^{\mu\nu}-h_{H}^{\mu\alpha}\partial_{\alpha}\xi^{\nu}-h_{H}^{\nu\alpha}\partial_{\alpha}\xi^{\mu}-p_{\beta}\partial_{\alpha}\xi^{\beta}\bar{\partial}^{\alpha}h_{H}^{\mu\nu}\,.$ (101) This is just the generalization of the killing equation to a general Hamilton space. In general, if $h_{H}$ does not depend on momenta, then it reduces to the standard Riemannian case. Besides, from the expression of the Poisson brackets (91), it can verified that such symmetries give rise to conserved charges $\xi^{\mu}p_{\mu}$; i.e., that Poisson commutes with the Hamiltonian: $\\{\xi^{\mu}p_{\mu},H\\}=0.$ (102) These are the charges that, at an algebraic level, can generate translations, boosts, and rotations, for instance. ### V.2 Hamilton–q-de Sitter (Cotangent Bundle Case) As an example, we rely on the results presented in Ref. Barcaroli:2015xda , which are as well inspired by the $q$-de Sitter Hamiltonian (45). In this case, the Hamilton metric, defined by Equation (88), reads: $\displaystyle h^{\mu\nu}_{H}(x,p)=\begin{pmatrix}1&-\ell p_{1}(1+2Hx^{0})\\\ -\ell p_{1}(1+2Hx^{0})&-(1+2Hx^{0})(1+\ell p_{0})\end{pmatrix}\,,$ (103) which, as can be seen, acquires a shape much simpler than the rainbow-Finsler one (83) due to the much direct way in which it is calculated. The non-linear connection can be read from Equation (92) and can be cast in a matrix form due to its simplicity: $\displaystyle O_{\mu\nu}(x,p)=\begin{pmatrix}H\ell p_{1}^{2}&Hp_{1}\\\ Hp_{1}&Hp_{0}(1-\ell p_{0})\end{pmatrix}\,.$ (104) As expected, it coincides with the case (IV.2) in the Riemannian case, i.e., when $\ell=0$. The Hamilton equations of motion can be found from Equation (89) and read: $\displaystyle\dot{x}^{0}-2p_{0}+\ell p_{1}^{2}(1+2Hx^{0})=0\,,$ (105) $\displaystyle\dot{x}^{1}+2p_{1}(1+Hx^{0})+2\ell p_{0}p_{1}(1+2Hx^{0})=0\,,$ (106) $\displaystyle\dot{p}_{0}-2Hp_{1}^{2}-2H\ell p_{0}p_{1}^{2}=0\,,$ (107) $\displaystyle\dot{p}_{1}=0\,.$ (108) The autoparallel (horizontal) curves of the non-linear connection satisfy (see Equation (8.2) in Ref. Miron:1994nvt ) $\dot{p}_{\mu}-O_{\nu\mu}\dot{x}^{\nu}=0\,,$ (109) and, as can be seen from Equation (104) for $O_{\mu\nu}$, the worldlines, defined from the Hamilton equations of motion, are not autoparallels of the non-linear connection. The symmetries have also been analyzed in Ref. Barcaroli:2015xda , where it has been noticed that the conserved charges that generate translations and the boost coincide with the results from Ref. Barcaroli:2015eqe that do not rely on the geometrical approach used in this paper. ## VI The Tangent-Bundle Version of Hamilton Geometry Endowed with Hamilton equations of motion (89), one has a map between the momenta and velocities from $\dot{x}^{\mu}=y^{\mu}=\partial H/\partial p_{\mu}$. When it is possible to invert this map to find $p_{\mu}=p_{\mu}(y)$ (as done in Appendix B of Ref. Barcaroli:2017gvg ), one derives an interesting map between the cotangent and tangent space version of Hamilton geometry. Indeed, using this map, a Hamilton metric defined in the tangent bundle reads: $\displaystyle g_{H}^{\mu\nu}(x,y)\doteq h_{H}^{\mu\nu}(x,p(y))\,.$ (110) The dual non-linear connection in this case has been discussed in Appendix C of Ref. Barcaroli:2015xda , and is given by $N(x,y)^{\mu}{}_{\nu}=2O(x,p(y))_{\nu\alpha}h_{H}^{\alpha\mu}(x,p(y))-(\partial_{\nu}\bar{\partial}^{\mu}H)|_{p=p(y)}\,.$ (111) Its main property is the preservation of the horizontal tangent spaces of the cotangent and tangent bundle connected through the kinematical map $y^{\mu}=\partial H/\partial p_{\mu}$. With this map, it is possible to define the dual non-linear and $N$-linear connections, now defined in the tangent bundle. It should be stressed that although this gives geometrical quantities defined in the tangent bundle, this does not represent a Finsler geometry, since there is no arc-length functional and the Hamilton metric is not, in general, 0-homogeneous to start with. ### Hamilton-$\kappa$-Poincaré (Tangent Bundle Case) The kinematical map that allows us to describe $y=y(p)$ is found by inverting the relation $y^{\mu}=\partial H/\partial p_{\mu}$ for the $q$-de Sitter Hamiltonian, given by $\displaystyle p_{0}=\frac{y^{0}}{2}+\ell\frac{(y^{1})^{2}}{8}\,,$ (112) $\displaystyle p_{1}=-\frac{y^{1}}{2}+H\frac{x^{0}y^{1}}{2}+\ell\frac{y^{0}y^{1}}{4}\,.$ (113) The metric in the tangent bundle reads: $\displaystyle g^{\mu\nu}_{H}(x,y)=\begin{pmatrix}1&\ell(Hx^{0}y^{1}+y^{1})/2\\\ \ell(Hx^{0}y^{1}+y^{1})/2&-(1+2Hx^{0})(1+\ell y^{0}/2)\end{pmatrix}\,.$ (114) The dual non-linear connection reads $\displaystyle N^{\mu}{}_{\nu}(x,y)=\begin{pmatrix}-H\ell(y^{1})^{2}/2&\ell hy^{0}y^{1}+hy^{1}\\\ Hy^{1}&-hy^{0}-3\ell h(y^{1})^{2}/4\end{pmatrix}\,.$ (115) In Section VII below, some key points of each approach are discussed while comparing the descriptions of configuration and phase spaces. ## VII Advantages and Difficulties of Each Formalism The approaches considered—Finsler and Hamilton spaces—present the points that can be considered positive or negative. In this Section, we highlight some of those points which look to be most important from theoretical and phenomenological points of view. ### VII.1 Finsler Geometry Let us emphasize that here not a complete list of positive or negative points are given, and, certainly, the points listed represent just our view on the subject under scrutiny and some points we are classifying in one way or another can be seen by others completely differently. #### VII.1.1 Advantages Preservation of the equivalence principle. Due to the presence of an arc- length functional, the extremizing geodesics of the Finsler function are the same worldlines of the Hamiltonian, from which the arc-length was derived. This means that, in the Finslerian language, the equivalence principle is satisfied, as soon as the worldlines are trajectories of free particles in this spacetime. There is a fundamental difference in comparison to the special or general relativity formulation, since these trajectories are now mass- dependent, since the Finsler function and the metric carry the mass of the particle due to Planck-scale effects. Intriguingly, although the metric does not present a massless limit (which is discussed below), it is possible to find trajectories of massless particles, which are compatible with the Hamiltonian formulation, by taking the limit $m\rightarrow 0$ in the geodesic Equation Amelino-Camelia:2014rga ; Lobo:2016xzq . This finding leads to some effects due to modifications of the trajectories of particles. For instance, one of the most explored avenues of quantum gravity phenomenology (maybe competing with threshold effects) is the time delay until particles with different energies might arrive at a detector after a (almost) simultaneous emission Jacob:2008bw ; Zhu:2022blp (for reviews, see Amelino-Camelia:2008aez ; Addazi:2021xuf ). This kind of experimental investigation is not exhausted, and novelties have arrived in the analysis of sets of gamma-ray bursts and candidate neutrinos emitted from them in the multimessenger astronomy approach Amelino-Camelia:2016ohi ; Amelino-Camelia:2022pja . Preservation of the relativity principle. This formalism allows one to derive and solve the killing equation, which furnishes infinitesimal symmetry transformations of the metric. It has been shown in Ref. Amelino- Camelia:2014rga that generators of these transformations can be constructed and identified with the transformations that are generally depicted in the doubly special relativity. The latter implies, in a preservation of the relativity principle that inertial frames should assign the same MDR to a given particle which, in its turn, implies that the deformation scale of quantum gravity is observer-independent, i.e., two observers would not assign different values, in the same system of units, to the quantum gravity scale. This preservation has important phenomenological consequences, such as the point that the threshold constraints on the quantum gravity parameter do not apply in the DSR scenario. The reason is that, accompanied by the deformation of the Lorentz (Poincaré) symmetries, comes a deformation of the composition law of momenta of particles (for instance $p$ and $q$), such that the nature of interaction vertices to not get modified when transforming from one frame to another: $\Lambda(p\oplus q)=\Lambda(p)\oplus\Lambda(q)\,,$ (116) where $\Lambda$ is a deformed Lorentz transformation induced by the killing vectors and $\oplus$ represents a modified composition of components of the involved momenta (this covariance condition usually needs a back-reaction on the boost parameter, but we do not dwell on that here; for more details, see Majid:2006xn ; Lobo:2021yem and references therein). Threshold constraints, such as the one placed in Ref. HAWC:2019gui , assumes that the composition of momenta is undeformed, although the dispersion relation is modified in a Lorentz invariance violation (LIV) scenario. When this is the case, processes that are forbidden in special relativity, such as the decay of the photon into an electron–positron pair, becomes kinematicaly allowed for a given threshold energy. The no observation of such decays allows one to place constraints on the quantum gravity parameter. When the dispersion relation is modified as well, what happens is that generally these kinds of processes remain forbidden or modifications in the threshold energies are so minute that they are unobservable for a quantum gravity parameter in the order of the Planck energy Lobo:2021yem . This is an important feature of "deforming" instead of "violating" the Lorentz symmetry. Preservation of the clock postulate. The availability of an arc-length functional leads to a possibility to analyze the consequences of having the proper time of a given particle given by it. If this is the case, then the worldlines or geodesics are just paths that extremize the proper time an observer measures in spacetime, similar to that in special relativity. One of the consequences of this feature consists of the possibility of connecting the time elapsed in the comoving frame of a particle during its lifetime (which is its lifetime at rest) and the coordinate time, which is the one that is assigned to this phenomenon in the laboratory coordinates. Using this expression, one can investigate the relativistic time dilation (responsible for the "twin paradox") or the so-called first clock effect (for further details on the first and also on the second clock effect, which can appear in theories with a non-metricity tensor; see Ref. Lobo:2018zrz ), in which, for instance, the lifetime of a particle is dilated in comparison to the one assigned in the laboratory. Due to Finslerian corrections, the lifetime of a particle in the laboratory would receive Planckian corrections, which, actually, is a novel avenue of phenomenological investigation that is being currently carried out Lobo:2020qoa ; Lobo:2021yem through the search for signatures in particle accelerators and cosmic rays. #### VII.1.2 Difficulties Absence of massless rainbow Finsler metric. The Finsler approach had emerged as an opportunity to describe in a consistent way the intuition that the quantum spacetime probed by a high-energy particle would present some energy- momentum (of the particle itself) corrections, which is justified by different approaches to quantum gravity Assanioussi:2014xmz ; Weinfurtner:2008if . Since then, proposals of rainbow metrics have considered a smooth transition from massive to massless cases, not only from the point of view of the trajectories, but from the metric itself. This is not the case for the Finsler approach presented here. Although the trajectories and symmetries are defined for both massive and massless cases by considering the $m\rightarrow 0$ limit, the rainbow metric of Finsler geometry, given by Equation (83), is certainly not defined for massless particles. The reason for this is the point that when passing from the Hamiltonian to the Lagrangian formalism, we defined an arc- length functional, which is not a legitimate action functional for massless particles. In other words, a crucial step for deriving the Finsler function is the handling of the Lagrange multiplier $\lambda$ of action (4), which can only be solved if the particle is massive, as can be found in Refs. Amelino- Camelia:2014rga ; Lobo:2016xzq ; Lobo:2016lxm ; Lobo:2020qoa . A possibility that has been explored consisted of not solving the equation for $\lambda$ and defining a metric that depends on $\lambda$ and on velocities from a Polyakov- like action for free particles (instead of the Nambu–Goto one given by the arc-length), which turned out to be out of the Finsler geometry scope Lobo:2016xzq ; Lobo:2016lxm . However, this possibility has not been further explored beyond preliminary investigations. The issue of the absence of a massless rainbow-Finsler metric could be circumvented by proposing a different kind of geometry, which from the very beginning started from the momenta formulation, like the other possibility described in this paper, namely the Hamilton geometry. Definition only through perturbations. the Finsler geometry has been considered in this paper in this context at most perturbatively around the quantum-gravity-length scale (or inverse of energy scale), which may be considered as a negative point if one aims to make it at a more fundamental or theoretical level. Nevertheless, from the pragmatic perspective of phenomenology, since such effects, if they exist, are minute, then the perturbative approach is enough for proposing new effects that could serve as avenues of experimental investigation. The handling of finite symmetries. Another issue that can be problematic is the handling of finite symmetries in the Finsler context. Up to today, the connection between Finsler geometry and quantum gravity phenomenology has not faced the issue of integrating the symmetries and finding finite versions of deformed Lorentz transformations. Some initial investigations were carried out in Ref. Lobo:2021yem from the momentum space perspective, but further issues are being currently faced by some authors of the present paper. ### VII.2 Hamilton Geometry The descriptions of the points in this section will be a bit shorter than the previous ones, because some universal points we already described above; therefore, we instruct the reader to check on them when that is the case. #### VII.2.1 Advantages Presence of a massless rainbow Hamilton metric. Differently from the Finsler case, the Hamilton geometry does not need an arc-length functional; instead, it only needs a given Hamiltonian, from which the metric, non-linear connection, and symmetries are derived. This means that from the very beginning, the massless limit of geometrical quantities exists. Does not require perturbative methods. Another positive point about the Hamilton geometry is the finding that one can handle with the exact form of the proposed Hamiltonian, and it does not need to consider perturbations around a certain scale. Instead, independently of the form of the (smooth) dispersion relation that arises from de facto approaches to quantum gravity, the geometry can be handled, as has been considered, e.g., in Refs. Barcaroli:2016yrl ; Barcaroli:2017gvg . Preservation of the relativity principle and the handling of symmetries. Due to the proximity of this approach to the way that the DSR formalism generally handles with Planck scale corrections, i.e., from the point of view of momentum space and Hamilton equations, the handling of symmetries is facilitated in this approach. For instance, it is straightforward to find the conserved charges from the killing vectors, which generate finite transformations that are momenta-dependent without tedious terms in the denominator of the equations when one is working in velocity space, as Finsler geometry is initially formulated (or without mass terms in the denominator in the Finsler version of the phase space). Generalization to curved spacetimes. This approach is considered in more curved space cases, beyond the $q$-de Sitter exemplified in this paper; for instance, its spherically symmetric and cosmological versions were explored giving rise to interesting phenomenological opportunities, from the point of view of time delays and gravitational redshift, among others (for some applications of Hamilton geometry in this context, see Pfeifer:2018pty and references therein). #### VII.2.2 Difficulties Non-geodesic trajectory. An issue that may be considered problematic is the point that the worldlines of particles, given by the Hamilton equations, are not geodesics of the non-linear connection that means that there exists a force term in the geodesic equation, which is in contrast with the Finsler case. This is a property of the Hamilton geometry, as has been shown in Ref. Barcaroli:2015xda , and is not specific to the $q$-de Sitter case analyzed here. Absence of the arc-length. The Hamilton geometry does not dwell with an arc- length functional that means that the only geodesics present are those of the non-linear or of the $N$-linear connections and there are no extremizing ones. The absence of a function that allows one to measure distances in spacetime can be seen as a difficulty of this geometry; if the distances cannot be calculated, one could wonder what such a metric means. Even if the norm of a tangent vector can be integrated, this integral would not be, in general, parametrization-independent, which is also a drawback of this tentative. Besides, the absence of an arc-length limits the phenomenology of the preservation of the clock postulate that was discussed in the Finsler case. ## VIII Final Remarks We revised two proposals that have been considered as candidates for describing the quantum configuration and phase spaces probed by particles whose kinematics are modified by a length scale identified as the quantum gravity scale. Finsler geometry starts from a configuration space framework that presents applications on its own in biology, thermodynamics, and modified gravity; and it finds a natural environment in quantum gravity phenomenology due to its power to describe a scenario in which important principles that guided physics in the XXth century, such as the relativity principle, are preserved even at a Planckian regime. Besides its traditional description in terms of the couple spacetime and velocity space (configuration space), we also explored its development in terms of the induced couple spacetime and momentum space (phase space), which is actually more appropriate for a pure quantum description than the configuration space. Some points that we consider positive and negative and which are consequences of the requirements for using the Finsler language, the derivation of an arc-length functional defined in the slit tangent bundle, are discussed in Section VII. The second case of the present study is the Hamilton geometry, whose properties are derived directly from the Hamiltonian itself, without the need to go through the definition of an arc-length. Actually, in general, the Hamilton metric does not even define a curve-parametrization-invariant length measure which brings some limitations to phenomenological investigations of this subject in quantum gravity. On the other hand, this issue circumvents some intrinsic difficulties of Finsler geometry, which were also discussed in Section VII. The goal of this paper was to review some topics of these two important geometries by using kinematical descriptions of particles whose behavior might present departures from special relativity results due to the effective quantum gravity influence. We also aimed to bring some points that we consider as under-explored perspectives on the subject by explicitly presenting some geometric quantities that are dual to those, in which those quantities were originally presented, such as the dual metrics and non-linear connections (whose Finslerian one was proposed in this paper, by inspiration of definitions in the Hamilton geometry literature) of Finsler and Hamilton geometries in the cotangent and tangent bundles, respectively. At least two global points could be considered insufficiently explored or unexplored in this subject. One is the geometry probed by an (non-)interacting multi-particle system. Some challenges of this problem can be found, for instance, in Ref. Hossenfelder:2014ifa , but the relations between the approaches there described and Finsler/Hamilton geometries remains unclear. Another point that remains unexplored consists of the dynamics of the configuration/phase space in a way that is compatible with quantum gravity phenomenology-inspired approaches. For instance, one could wonder if there exists a gravitational field theory defined in Finsler or Hamilton spaces that has $q$-de Sitter or other proposals as solutions, and how matter would interact in this scenario. The exploration of this topic might shed light on the one regarding a multi-particle system. These are more challenges that might be subjects of the future research in this area and which may help to build a bridge between quantum and modified gravities. ## Ackowledgements I.P.L. was partially supported by the National Council for Scientific and Technological Development—CNPq grant 306414/2020-1, and by the grant 3197/2021, Paraíba State Research Foundation (FAPESQ). I.P.L. would like to acknowledge the contribution of the COST Action CA18108. L.C.N.S. would like to thank Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) for partial financial support through the research project no. 164762/2020-5. V.B.B. is partially supported by CNPq through the research project no. 307211/2020-7. P.H.M. and S.A. thank Coordenação de Aperfeiçoamento de Pessoal de Nível Superior—Brazil (CAPES)—Finance Code 001, for financial support. G.V.S thank Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) for financial support. E.R. and G.M were supported by the PIBIC program of the Federal University of Paraíba. ## Appendix A Dual Finsler Nonlinear Connection The momenta of a particle in Finsler geometry, given by the following expression, $p_{\mu}=m\frac{\partial F}{\partial y^{\mu}}\equiv m\dot{\partial}_{\mu}F\,,$ (117) defines a kinematical map between velocity and momenta variables at each given point in the base manifold $M$. We refer to such a map as $\displaystyle\flat:\quad\widetilde{TM}\rightarrow\widetilde{T^{\ast}}M$ (118) $\displaystyle(x,y)\mapsto\flat(x,y)=(x,m\dot{\partial}F(x,y))=(x,p(x,y))\,.$ (119) Inspired by the construction of Ref. Barcaroli:2015xda , the condition that a nonlinear connection in the tangent bundle is dual to one in the cotangent bundle by a kinematical map, $\flat$, is that such an application maps the tangent space of the tangent bundle onto the tangent space of the cotangent bundle. This means that the differential of such a map maps the preferred basis of one tangent space, $\delta_{\mu}=\partial_{\mu}-N^{\nu}{}_{\mu}\dot{\partial}_{\nu}$, onto the other, $d\,\flat(\delta_{\mu})=\delta^{\prime}_{\mu}=\partial_{\mu}-O_{\mu\nu}\dot{\partial}^{\nu}$. This means that the action of this differential on a vector $X=X^{\mu}\partial_{\mu}+\dot{X}^{\mu}\dot{\partial}_{\mu}$ is given by $\displaystyle d\,\flat_{(x,y)}:\quad T_{(x,y)}\widetilde{TM}$ $\displaystyle\rightarrow T_{\flat(x,y)}\widetilde{T^{*}M}\,,$ (120) $\displaystyle X=X^{\mu}\partial_{\mu}+\dot{X}^{\mu}\dot{\partial}_{\mu}$ $\displaystyle\mapsto d\,\flat_{(x,y)}(X)=X^{\mu}d\,\flat_{(x,y)}(\partial_{\mu})+\dot{X}^{\mu}d\,\flat_{(x,y)}(\dot{\partial}_{\mu})$ (121) $\displaystyle=X^{\mu}(\partial_{\mu}+m\partial_{\mu}\dot{\partial}_{\nu}F\bar{\partial}^{\nu})+m\dot{X}^{\mu}\dot{\partial}_{\mu}\dot{\partial}_{\nu}F\bar{\partial}^{\nu}\,.$ (122) By acting on the basis vectors $\delta_{\mu}=\partial_{\mu}-N^{\nu}{}_{\mu}\dot{\partial}_{\nu}$, one finds: $\displaystyle d\,\flat_{(x,y)}(\delta_{\mu})=d\,\flat_{(x,y)}(\delta_{\mu})-N^{\nu}{}_{\mu}d\,\flat_{(x,y)}(\dot{\partial}_{\nu})=\partial_{\mu}+m\partial_{\mu}\dot{\partial}_{\nu}F\bar{\partial}^{\nu}-mN^{\nu}{}_{\mu}\dot{\partial}_{\nu}\dot{\partial}_{\alpha}F\bar{\partial}^{\alpha}\,.$ (123) In order to simplify this expression, the relation $2g_{\nu\alpha}=\dot{\partial}_{\nu}\dot{\partial}_{\alpha}F^{2}=\dot{\partial}_{\nu}(2F\dot{\partial}_{\alpha}F)$ is used that leads to $\dot{\partial}_{\nu}\dot{\partial}_{\alpha}F=\frac{g_{\nu\alpha}-p_{\nu}p_{\alpha}/m^{2}}{F}\,.$ (124) From this expression, one finds that $d\,\flat_{(x,y)}(\delta_{\mu})=\partial_{\mu}-m\left[N^{\alpha}{}_{\mu}\frac{(g_{\alpha\nu}-p_{\alpha}p_{\nu}/m^{2})}{F}-\partial_{\mu}\dot{\partial}_{\nu}F\right]\bar{\partial}^{\nu}=\partial_{\mu}+O_{\mu\nu}\bar{\partial}^{\nu}\,,$ (125) which leads to the dual nonlinear connection, $O_{\mu\nu}(x,p)=-m\left[N^{\alpha}{}_{\mu}\frac{(g_{\alpha\nu}-p_{\alpha}p_{\nu}/m^{2})}{F}-\partial_{\mu}\dot{\partial}_{\nu}F\right]\Bigg{|}_{(x,y(p))}\,.$ (126) ## References * (1) Bronstein, Matvei, “Quantum theory of weak gravitational fields,” Gen. Rel. Grav. 44 (2012) 267–283. * (2) Rovelli, Carlo, “Loop quantum gravity,” Living Rev. Rel. 11 (2008) 5. * (3) Ambjorn, J. and Jurkiewicz, J. and Loll, R., “Causal Dynamical Triangulations and the Quest for Quantum Gravity,” in Foundations of Space and Time: Reflections on Quantum Gravity, pp. 321–337. 4, 2010. arXiv:1004.0352 [hep-th]. * (4) Amelino-Camelia, Giovanni, “Quantum-Spacetime Phenomenology,” Living Rev. Rel. 16 (2013) 5, arXiv:0806.0339 [gr-qc]. * (5) Addazi, A. and others, “Quantum gravity phenomenology at the dawn of the multi-messenger era—A review,” Prog. Part. Nucl. Phys. 125 (2022) 103948, arXiv:2111.05659 [hep-ph]. * (6) Amelino-Camelia, Giovanni and da Silva, Malú Maira and Ronco, Michele and Cesarini, Lorenzo and Lecian, Orchidea Maria, “Spacetime-noncommutativity regime of Loop Quantum Gravity,” Phys. Rev. D 95 no. 2, (2017) 024028, arXiv:1605.00497 [gr-qc]. * (7) Brahma, Suddhasattwa and Ronco, Michele and Amelino-Camelia, Giovanni and Marciano, Antonino, “Linking loop quantum gravity quantization ambiguities with phenomenology,” Phys. Rev. D 95 no. 4, (2017) 044005, arXiv:1610.07865 [gr-qc]. * (8) Brahma, Suddhasattwa and Ronco, Michele, “Constraining the loop quantum gravity parameter space from phenomenology,” Phys. Lett. B 778 (2018) 184–189, arXiv:1801.09417 [hep-th]. * (9) Majid, S. and Ruegg, H., “Bicrossproduct structure of kappa Poincare group and noncommutative geometry,” Phys. Lett. B 334 (1994) 348–354, arXiv:hep-th/9405107. * (10) Lukierski, Jerzy and Ruegg, Henri and Nowicki, Anatol and Tolstoi, Valerii N., “Q deformation of Poincare algebra,” Phys. Lett. B 264 (1991) 331–338. * (11) Lukierski, Jerzy and Nowicki, Anatol and Ruegg, Henri, “New quantum Poincare algebra and k deformed field theory,” Phys. Lett. B 293 (1992) 344–352. * (12) Majid, S., Foundations of quantum group theory. Cambridge University Press, 2, 2011. * (13) Amelino-Camelia, Giovanni, “Doubly-Special Relativity: Facts, Myths and Some Key Open Issues,” Symmetry 2 (2010) 230–271, arXiv:1003.3942 [gr-qc]. * (14) Mattingly, David, “Modern tests of Lorentz invariance,” Living Rev. Rel. 8 (2005) 5, arXiv:gr-qc/0502097. * (15) Liberati, Stefano, “Tests of Lorentz invariance: a 2013 update,” Class. Quant. Grav. 30 (2013) 133001, arXiv:1304.5795 [gr-qc]. * (16) Amelino-Camelia, Giovanni, “Relativity in space-times with short distance structure governed by an observer independent (Planckian) length scale,” Int. J. Mod. Phys. D 11 (2002) 35–60, arXiv:gr-qc/0012051. * (17) Magueijo, Joao and Smolin, Lee, “Lorentz invariance with an invariant energy scale,” Phys. Rev. Lett. 88 (2002) 190403, arXiv:hep-th/0112090. * (18) Amelino-Camelia, Giovanni and Freidel, Laurent and Kowalski-Glikman, Jerzy and Smolin, Lee, “The principle of relative locality,” Phys. Rev. D 84 (2011) 084010, arXiv:1101.0931 [hep-th]. * (19) Proutorov, Evgenii and Matsuyama, Naoki and Koibuchi, Hiroshi, “Finsler geometry modeling and monte carlo study of liquid crystal elastomers under electric fields,” Journal of Physics: Condensed Matter 30 no. 40, (2018) 405101. * (20) Hehl, Friedrich W and Obukhov, Yuri N, Foundations of classical electrodynamics: Charge, flux, and metric, vol. 33. Springer Science & Business Media, 2003. * (21) Červenỳ, V, “Fermat’s variational principle for anisotropic inhomogeneous media,” Studia geophysica et geodaetica 46 no. 3, (2002) 567–588. * (22) Pfeifer, Christian, “Finsler spacetime geometry in Physics,” Int. J. Geom. Meth. Mod. Phys. 16 no. supp02, (2019) 1941004, arXiv:1903.10185 [gr-qc]. * (23) Magueijo, Joao and Smolin, Lee, “Gravity’s rainbow,” Class. Quant. Grav. 21 (2004) 1725–1736, arXiv:gr-qc/0305055. * (24) Weinfurtner, Silke and Jain, Piyush and Visser, Matt and Gardiner, C. W., “Cosmological particle production in emergent rainbow spacetimes,” Class. Quant. Grav. 26 (2009) 065012, arXiv:0801.2673 [gr-qc]. * (25) Assanioussi, Mehdi and Dapor, Andrea and Lewandowski, Jerzy, “Rainbow metric from quantum gravity,” Phys. Lett. B 751 (2015) 302–305, arXiv:1412.6000 [gr-qc]. * (26) Olmo, Gonzalo J., “Palatini Actions and Quantum Gravity Phenomenology,” JCAP 10 (2011) 018, arXiv:1101.2841 [gr-qc]. * (27) Ling, Yi and Li, Xiang and Zhang, Hong-bao, “Thermodynamics of modified black holes from gravity’s rainbow,” Mod. Phys. Lett. A 22 (2007) 2749–2756, arXiv:gr-qc/0512084. * (28) Lobo, Iarley P. and Santos, Luis C. N. and Bezerra, V. B. and Morais Graça, J. P. and Moradpour, H., “The extended phase space thermodynamics of Planck-scale-corrected Reissner-Nordström-anti-de Sitter black hole,” Nucl. Phys. B 972 (2021) 115568, arXiv:2110.05396 [gr-qc]. * (29) Gorji, M. A. and Nozari, K. and Vakili, B., “Gravity’s rainbow: a bridge between LQC and DSR,” Phys. Lett. B 765 (2017) 113–119, arXiv:1606.00910 [gr-qc]. * (30) Garattini, Remo and Lobo, Francisco S. N., “Gravity’s Rainbow and traversable wormholes,” in 14th Marcel Grossmann Meeting on Recent Developments in Theoretical and Experimental General Relativity, Astrophysics, and Relativistic Field Theories, vol. 2, pp. 1448–1453. 2017\. arXiv:1512.04470 [physics.gen-ph]. * (31) Amirabi, Z. and Halilsoy, M. and Mazharimousavi, S. Habib, “Thin-shell wormholes in rainbow gravity,” Mod. Phys. Lett. A 33 no. 09, (2018) 1850049. * (32) Bezerra, V. B. and Lobo, I. P. and Mota, H. F. and Muniz, C. R., “Landau Levels in the Presence of a Cosmic String in Rainbow Gravity,” Annals Phys. 401 (2019) 162–173, arXiv:1901.01596 [gr-qc]. * (33) Carvalho, Gabriel G. and Lobo, Iarley P. and Bittencourt, Eduardo, “Extended disformal approach in the scenario of Rainbow Gravity,” Phys. Rev. D 93 no. 4, (2016) 044005, arXiv:1511.00495 [gr-qc]. * (34) Lobo, Iarley P. and Carvalho, Gabriel G., “The geometry of null-like disformal transformations,” Int. J. Geom. Meth. Mod. Phys. 16 no. 11, (2019) 1950180, arXiv:1707.01784 [gr-qc]. * (35) Santos, L. C. N. and Bezerra, V. B., “Electrostatic self-interaction of charged particles in the space-time of a cosmic string in the context of gravity’s rainbow,” Gen. Rel. Grav. 51 no. 11, (2019) 145, arXiv:1911.01996 [gr-qc]. * (36) Bernhard Riemann (auth.), Jürgen Jost (eds.), On the Hypotheses Which Lie at the Bases of Geometry. Classic Texts in the Sciences. Birkhäuser Basel, 1 ed., 2016. * (37) Paul Finsler (auth.), Über Kurven und Flächen in allgemeinen Räumen. Lehrbücher und Monographien aus dem Gebiete der Exakten Wissenschaften 11. Springer Basel, 1 ed., 1951. * (38) Bao, David and Chern, S-S and Shen, Zhongmin, An introduction to Riemann-Finsler geometry, vol. 200. Springer Science & Business Media, 2000. * (39) Hohmann, Manuel and Pfeifer, Christian and Voicu, Nicoleta, “Mathematical foundations for field theories on Finsler spacetimes,” J. Math. Phys. 63 no. 3, (2022) 032503, arXiv:2106.14965 [math-ph]. * (40) Bernal, Antonio and Javaloyes, Miguel Ángel and Sánchez, Miguel, “Foundations of Finsler Spacetimes from the Observers’ Viewpoint,” Universe 6 no. 4, (2020) 55, arXiv:2003.00455 [gr-qc]. * (41) Minguzzi, E., “Light cones in Finsler spacetime,” Commun. Math. Phys. 334 no. 3, (2015) 1529–1551, arXiv:1403.7060 [math-ph]. * (42) Girelli, Florian and Liberati, Stefano and Sindoni, Lorenzo, “Planck-scale modified dispersion relations and Finsler geometry,” Phys. Rev. D 75 (2007) 064015, arXiv:gr-qc/0611024. * (43) Lobo, Iarley P. and Pfeifer, Christian, “Reaching the Planck scale with muon lifetime measurements,” Phys. Rev. D 103 no. 10, (2021) 106025, arXiv:2011.10069 [hep-ph]. * (44) Raetzel, Dennis and Rivera, Sergio and Schuller, Frederic P., “Geometry of physical dispersion relations,” Phys. Rev. D 83 (2011) 044047, arXiv:1010.1369 [hep-th]. * (45) Rodrigues, Ernesto and Lobo, Iarley P., “Revisiting Legendre transformations in Finsler geometry,” arXiv:2208.11406 [gr-qc]. * (46) Gubitosi, Giulia and Mercati, Flavio, “Relative Locality in $\kappa$-Poincaré,” Class. Quant. Grav. 30 (2013) 145002, arXiv:1106.5710 [gr-qc]. * (47) Miron, Radu and Anastasiei, Mihai, The Geometry of Lagrange Spaces: Theory and Applications. Fundam.Theor.Phys. Springer Netherlands, Dordrecht, 1994. * (48) Zhongmin Shen, Differential geometry of spray and Finsler spaces. Kluwer Academic Publishers, 1 ed., 2001. * (49) Amelino-Camelia, Giovanni and Barcaroli, Leonardo and Gubitosi, Giulia and Liberati, Stefano and Loret, Niccoló, “Realization of doubly special relativistic symmetries in Finsler geometries,” Phys. Rev. D 90 no. 12, (2014) 125030, arXiv:1407.8143 [gr-qc]. * (50) Lobo, Iarley P. and Loret, Niccoló and Nettel, Francisco, “Investigation of Finsler geometry as a generalization to curved spacetime of Planck-scale-deformed relativity in the de Sitter case,” Phys. Rev. D 95 no. 4, (2017) 046015, arXiv:1611.04995 [gr-qc]. * (51) Barcaroli, Leonardo and Gubitosi, Giulia, “Kinematics of particles with quantum-de Sitter-inspired symmetries,” Phys. Rev. D 93 no. 12, (2016) 124063, arXiv:1512.03462 [gr-qc]. * (52) Letizia, Marco and Liberati, Stefano, “Deformed relativity symmetries and the local structure of spacetime,” Phys. Rev. D 95 no. 4, (2017) 046007, arXiv:1612.03065 [gr-qc]. * (53) Lobo, Iarley P. and Loret, Niccoló and Nettel, Francisco, “Rainbows without unicorns: Metric structures in theories with Modified Dispersion Relations,” Eur. Phys. J. C 77 no. 7, (2017) 451, arXiv:1610.04277 [gr-qc]. * (54) Amelino-Camelia, Giovanni and Rosati, Giacomo and Bedić, Suzana, “Phenomenology of curvature-induced quantum-gravity effects,” Phys. Lett. B 820 (2021) 136595, arXiv:2012.07790 [gr-qc]. * (55) Lobo, Iarley P. and Pfeifer, Christian and Morais, Pedro H. and Batista, Rafael Alves and Bezerra, Valdir B., “Two-body decays in deformed relativity,” JHEP 09 (2022) 003, arXiv:2112.12172 [hep-ph]. * (56) Barcaroli, Leonardo and Brunkhorst, Lukas K. and Gubitosi, Giulia and Loret, Niccoló and Pfeifer, Christian, “Hamilton geometry: Phase space geometry from modified dispersion relations,” Phys. Rev. D92 no. 8, (2015) 084053, arXiv:1507.00922 [gr-qc]. * (57) Barcaroli, Leonardo and Brunkhorst, Lukas K. and Gubitosi, Giulia and Loret, Niccoló and Pfeifer, Christian, “Planck-scale-modified dispersion relations in homogeneous and isotropic spacetimes,” Phys. Rev. D95 no. 2, (2017) 024036, arXiv:1612.01390 [gr-qc]. * (58) Barcaroli, Leonardo and Brunkhorst, Lukas K. and Gubitosi, Giulia and Loret, Niccoló and Pfeifer, Christian, “Curved spacetimes with local $\kappa$-Poincaré dispersion relation,” Phys. Rev. D 96 no. 8, (2017) 084010, arXiv:1703.02058 [gr-qc]. * (59) Pfeifer, Christian, “Redshift and lateshift from homogeneous and isotropic modified dispersion relations,” Phys. Lett. B 780 (2018) 246–250, arXiv:1802.00058 [gr-qc]. * (60) Nozari, Kourosh and Gorji, M. A. and Hosseinzadeh, V. and Vakili, B., “Natural Cutoffs via Compact Symplectic Manifolds,” Class. Quant. Grav. 33 no. 2, (2016) 025009, arXiv:1405.4083 [gr-qc]. * (61) Jacob, Uri and Piran, Tsvi, “Lorentz-violation-induced arrival delays of cosmological particles,” JCAP 01 (2008) 031, arXiv:0712.2170 [astro-ph]. * (62) Zhu, Jie and Ma, Bo-Qiang, “Lorentz-violation-induced arrival time delay of astroparticles in Finsler spacetime,” Phys. Rev. D 105 no. 12, (2022) 124069, arXiv:2206.07616 [gr-qc]. * (63) Amelino-Camelia, Giovanni and D’Amico, Giacomo and Rosati, Giacomo and Loret, Niccoló, “In-vacuo-dispersion features for GRB neutrinos and photons,” Nature Astron. 1 (2017) 0139, arXiv:1612.02765 [astro-ph.HE]. * (64) Amelino-Camelia, Giovanni and Di Luca, Maria Grazia and Gubitosi, Giulia and Rosati, Giacomo and D’Amico, Giacomo, “Could quantum gravity slow down neutrinos?,” arXiv:2209.13726 [gr-qc]. * (65) Majid, Shahn, “Algebraic approach to quantum gravity. II. Noncommutative spacetime,” arXiv:hep-th/0604130. * (66) HAWC Collaboration, Albert, A. and others, “Constraints on Lorentz Invariance Violation from HAWC Observations of Gamma Rays above 100 TeV,” Phys. Rev. Lett. 124 no. 13, (2020) 131101, arXiv:1911.08070 [astro-ph.HE]. * (67) Lobo, I. P. and Romero, C., “Experimental constraints on the second clock effect,” Phys. Lett. B 783 (2018) 306–310, arXiv:1807.07188 [gr-qc]. * (68) Hossenfelder, Sabine, “The Soccer-Ball Problem,” SIGMA 10 (2014) 074, arXiv:1403.2080 [gr-qc].
# Identifying regime switches through Bayesian wavelet estimation: evidence from flood detection in the Taquari River Valley Flávia C. Motta Department of Statistics, Federal University of São Carlos, São Carlos, SP, Brazil Institute of Mathematics and Computer Sciences, University of São Paulo, São Carlos, SP, Brazil Michel H. Montoril Department of Statistics, Federal University of São Carlos, São Carlos, SP, Brazil (May 2023) ###### Abstract Two-component mixture models have proved to be a powerful tool for modeling heterogeneity in several cluster analysis contexts. However, most methods based on these models assume a constant behavior for the mixture weights, which can be restrictive and unsuitable for some applications. In this paper, we relax this assumption and allow the mixture weights to vary according to the index (e.g., time) to make the model more adaptive to a broader range of data sets. We propose an efficient MCMC algorithm to jointly estimate both component parameters and dynamic weights from their posterior samples. We evaluate the method’s performance by running Monte Carlo simulation studies under different scenarios for the dynamic weights. In addition, we apply the algorithm to a time series that records the level reached by a river in southern Brazil. The Taquari River is a water body whose frequent flood inundations have caused various damage to riverside communities. Implementing a dynamic mixture model allows us to properly describe the flood regimes for the areas most affected by these phenomena. ## 1 Introduction In several data analysis problems, we want to cluster observations between two groups. For instance, in many clinical studies, the goal is to classify patients according to disease absent or present (see Hall and Zhou,, 2003, Rindskopf and Rindskopf,, 1986, Hui and Zhou,, 1998). In contamination problems found in astronomy investigations, on the other hand, the aim is to separate the objects of interest, called members (e.g., stars), from foreground/background objects contaminating the sample, known as contaminants (see Walker et al.,, 2009). In genetics, studies based on microarray data are usually driven to detecting differentially expressed genes under two conditions, e.g., “healthy tissue versus diseased tissue” (see Bordes et al.,, 2006). To address these scenarios of bimodal data sets, two-component mixture models have shown to be excellent alternatives to cluster data observations within the group that better describes their features (Patra and Sen,, 2016). In this context, the mixture model with two components will assume that the sample of data observations $y_{1},\dots,y_{n}$ is, in fact, the realization of a random variable $Y$ that belongs to a population composed of two subpopulations, known as mixture components. Thus, at each point $t$, $t=1,\dots,n$, $Y$ is fitted according to some of the mixture components, dictated by a mixture weight $\alpha$. This setting may be very restrictive to some data sets. For instance, in epidemiological studies that evaluate the response to medications, the probability of classifying a patient in the group of “disease present” must be allowed to vary across time so that the longitudinal effect of the treatment can be properly measured. The same issue arises in quality control problems, where the probability of the supervised system operating in a failure-free regime is also not constant over time. In order to classify those features properly, under a mixture model assumption, the mixture weight should be allowed to vary according to the index (which could be time or location). In other words, it would be appropriate for the mixture weight to present a dynamic behavior. Assuming dynamic mixture weights for mixture models is an extension that has already been applied in different areas, from traffic flow applications (see Nagy et al.,, 2011) to investigations in genetics (see Montoril et al.,, 2019; 2021). As discussed in Montoril et al., (2021), this generalization is similar to the extension of Hidden Markov Models (HMM) into non-homogeneous Hidden Markov Models (NHMM), first described by Hughes and Guttorp, (1994). In both scenarios, one generalizes the model by considering unobserved varying probabilities. In the case of mixture models, those dynamic probabilities are the mixture weights, whereas, in HMM, they are the transition probabilities. It is important to emphasize that, although connected, dynamic mixture weights and transition probabilities are different things. Considering a “non-homogeneous” structure for the mixture model implies that, besides estimating the dynamic mixture weights, one also needs to estimate the component parameters, and that increases the challenge. For instance, in Montoril et al., (2019), from a frequentist approach, the authors rely on wavelets to perform the estimation of the dynamic weights, where they transform the data in order to deal with a nonparametric heteroscedastic regression. Nonetheless, their procedure depends on assuming known means and variances for the mixture components, which, in practice, may be unrealistic. In this work, unlike the aforementioned paper, the leading motivation is to provide a Bayesian approach that estimates not only the dynamic mixture weights but also the component parameters of a two-component mixture model. To accomplish this goal, we propose an efficient Gibbs sampling algorithm, which allows the distribution of the posterior draws to be used for inference purposes. Regarding the dynamic mixture weights, we use the data augmentation method by Albert and Chib, (1993) and incorporate Bayesian wavelet denoising techniques to estimate the dynamic behavior of the mixture weight. We do this to exploit the good properties of wavelets in curves’ estimation. Wavelets are families of basis functions that can be used to represent other functions, signals, and images as a series of successive approximations (Härdle et al.,, 2012, Abramovich et al.,, 2000). In statistical applications, these mathematical tools have been successfully used to solve problems in nonparametric regression (see Donoho and Johnstone,, 1994, Cai and Brown,, 1999); density estimation (see Donoho, 1993a, , Donoho et al.,, 1996, Hall and Patil,, 1995); time series analysis (see, e.g., Morettin,, 1996, Priestley,, 1996, Percival and Walden,, 1999); among many other areas. There is a vast literature that provides a review of wavelets in statistics (see, e.g., Vidakovic,, 1999, Ogden,, 1997). In this paper, wavelet bases are applied to enable the estimation of the dynamic mixture weights. To review the mathematical background and the terminology associated with the wavelet theory, in the following section, we provide a short introduction to the wavelet basis functions; the discrete wavelet transform (DWT); and, the Bayesian approach for denoising in a wavelet-based scenario. The remainder of the paper is organized as follows. In Section 3, we describe the dynamic mixture model considered in this paper and give details related to the MCMC sampling scheme constructed to perform the estimation. In Section 4, we present some numerical experiments. We first conduct Monte Carlo simulations to evaluate the method in a controlled setting. Then, we apply the MCMC algorithm to a river data set to identify periods when flood inundations occurred. ## 2 Wavelets In this work, we use the term wavelets to refer to a system of orthonormal basis functions for $L_{2}([0,1])$ or $L_{2}(\mathbb{R})$. The bases are generated by dyadic translations and dilations of the functions $\varphi(\cdot)$ and $\psi(\cdot)$, known, respectively, as the scaling and wavelet functions. These systems of integer-translates and dilates are given by $\displaystyle\varphi_{j_{0}k}(t)$ $\displaystyle=2^{j_{0}/2}\varphi(2^{j_{0}}t-k),\quad k\in\mathbb{Z},$ $\displaystyle\psi_{jk}(t)$ $\displaystyle=2^{j/2}\psi(2^{j}t-k),\quad j,k\in\mathbb{Z}.$ Thus, for any integer $j_{0}$ and $J$, a periodic function $f(t)\in L_{2}([0,1])$ can be approximated in $L_{2}$-sense as the projection onto a multiresolution space $V_{J}$: $f(t)=\sum\limits_{k=0}^{2^{j_{0}}-1}c_{j_{0}k}\varphi_{j_{0}k}(t)+\sum\limits_{j=j_{0}}^{J-1}\sum\limits_{k=0}^{2^{j}-1}d_{jk}\psi_{jk}(t),$ where $c_{j_{0}k}$’s are known as scaling coefficients and $d_{jk}$’s are called detail coefficients. The former are associated with the coarsest resolution level in witch $f(t)$ was decomposed, $j_{0}$. As a result, they capture the gross structure of $f(t)$. The detail coefficients, on the other hand, being linked to finer resolution levels, can capture local information about $f(t)$. Put simply, in moving from a coarser resolution level $j$ to a finer $j+1$, we are increasing the resolution at which a function is approximated, thus the expansion coefficients become more descriptive about the local features of $f(t)$. In practice, we access $f(t)\in L_{2}([0,1])$ through a grid of points in time or space in which $f$ is applied. Therefore, consider $\bm{f}=(f(1/n),f(2/n),\dots,f(n/n))^{T}$ to be a vector of samples of $f(t)$ on an equispaced grid of $n$ points, with $n=2^{J}$, for some positive integer $J$. To obtain the scaling and detail coefficients that approximate $\bm{f}$, we perform the discrete wavelet transform (DWT) of $\bm{f}$. In matrix notation, the DWT of $\bm{f}$ is $\bm{\theta}=\bm{Wf},$ (1) where $\bm{\theta}=(c_{00},d_{00},\bm{d}_{1}^{T},\dots,\bm{d}_{J-1}^{T})^{T}$ is a vector of size $n$, having both scaling and detail coefficients $\bm{d}_{j}=(d_{j0},d_{j1},\dots,d_{j2^{j}-1})^{T}$, and $\bm{W}$ is the DWT matrix with $(jk,i)$ entry given by $W_{jk,i}\sqrt{n}\approx\psi_{jk}(i/n)=2^{j/2}\psi(2^{j}i/n-k)$, $k=0,\dots,2^{j}-1$, $j=1,\dots,J-1$. (Abramovich et al.,, 1998). By orthogonality, the multiplication $\bm{W^{T}\theta}$ recovers the signal $\bm{f}$. This transformation from wavelet coefficients to fitted values is known as the inverse discrete wavelet transform (IDWT). One of the main advantages provided by the DWT is the sparse representation generally achieved. As shown by Donoho, 1993b , wavelets are unconditional bases for a range of function spaces, such as Hölder and Sobolev spaces, as well as spaces suitable for representing functions of ‘bounded variation’. As an aside, it is also worth mentioning that using Mallat’s pyramid algorithm (Mallat,, 1989), the DWT and IDWT are performed requiring only $\mathcal{O}(n)$ operations, which makes them very efficient in terms of computational speed and storage. These properties help to explain why wavelet bases are excellent tools to address problems of data analysis. In the following section, we present a brief review of handling the denoising problem within the wavelet domain, emphasizing the Bayesian framework due to its central role in the estimation process of this paper. ### 2.1 Bayesian wavelet denoising Consider the nonparametric regression model $\bm{y}=\bm{f}+\bm{e},$ (2) where $\bm{y}=(y_{1},\dots,y_{n})^{T}$ is the vector of observed values, $\bm{f}=(f(1/n),\dots,f(n/n))^{T}$ is the function of interest applied to a grid of $n$ equally spaced points, and $\bm{e}=(e_{1},\dots,e_{n})^{T}$ is a vector of zero-mean random variables. For most applications, $e_{t}$’s are independent and identically distributed normal random variables with zero mean and constant variance $\sigma^{2}$. The goal of nonparametric regression is to recover the unknown function $f$ from the noisy observations $\bm{y}$. With that in mind, Donoho and Johnstone, (1994) propose to transform the observations $\bm{y}$ to the wavelet domain, shrink the noisy wavelet coefficients or even equal them to zero, based on some threshold rule, and then estimate $\bm{f}$ by applying the IDWT to the regularized coefficients. This method is known in the literature as wavelet shinkage. Therefore, let $n$ be a power of two, $n=2^{J}$ for some positive integer $J$. Then, we can represent (2) in the wavelet domain as $\bm{d}^{*}=\bm{\theta}+\bm{\varepsilon},$ (3) where $\bm{d}^{*}=\bm{Wy}$, $\bm{\theta}=\bm{Wf}$, and $\bm{\varepsilon}=\bm{We}$, with $\bm{W}$ being the DWT matrix. From a Bayesian perspective, the wavelet shrinkage technique consists in assigning a prior distribution to each wavelet coefficient of the unknown function. The idea is that, by choosing a prior able to capture the sparseness associated with most wavelet decompositions, we can estimate $\bm{\theta}$, by imposing some Bayes rule on the resulting posterior distribution of the wavelet coefficients. Then, applying the IDWT to the estimated $\bm{\theta}$ gives us an estimation of $\bm{f}$. One of the most appropriate prior choices for modeling wavelet coefficients are the spike and slab priors. First consolidated within Bayesian variable selection methods (George and McCulloch,, 1993), these kinds of prior are a mixture between two components: one that concentrates its mass at values close to zero or even in zero (Dirac delta) and another whose mass is spread over a wide range of possible values for the unknown parameters. Choosing this mixture as prior to the distribution of wavelet coefficients allows the first component, known as spike, to capture the null wavelet coefficients, while the second component, called slab, describes the coefficients associated with the unknown function. A spike and slab prior frequently assigned to wavelet coefficients is the mixture between a point mass at zero and a Gaussian distribution (see, e.g., Abramovich et al.,, 1998). In this scenario, each detail wavelet coefficient is distributed following $\displaystyle\begin{split}&\pi_{j}\text{N}(0,\upsilon_{j}^{2})+(1-\pi_{j})\delta_{0}(\theta_{jk}),\\\ \end{split}$ (4) $k=0,1,\dots,2^{j}-1$, $j=0,1,\dots,J-1$, with $\delta_{0}$ being a point mass at zero. The prior specification is usually completed by assigning a diffuse prior to the scaling coefficient at the coarsest level $c_{00}$. Thus, the sample scaling coefficient obtained from the DWT of the data estimates $c_{00}$ (Abramovich et al.,, 1998). Under the prior (LABEL:eq:prior_normal), the posterior distribution for each detail coefficient is also a mixture between a Gaussian distribution and $\delta_{0}$, given by $\displaystyle\begin{split}\theta_{jk}|d_{jk}^{*}&\sim\pi_{\text{post}}\text{N}\left(\frac{\upsilon_{j}^{2}}{1+\upsilon_{j}^{2}}d_{jk}^{*},\frac{\upsilon_{j}^{2}}{1+\upsilon_{j}^{2}}\right)+(1-\pi_{\text{{post}}})\delta_{0}(\theta_{jk}),\\\ \pi_{\text{post}}&=\frac{\pi_{j}g_{\upsilon_{j}^{2}}(d_{jk}^{*})}{\pi_{j}g_{\upsilon_{j}^{2}}(d_{jk}^{*})+(1-\pi_{j})\phi(d_{jk}^{*})},\\\ \end{split}$ (5) $k=0,1,\dots,2^{j}-1$, $j=0,1,\dots,J-1$, where $\phi$ denotes the standard normal density and $g_{\upsilon_{j}^{2}}$ denotes the convolution between the slab component in (LABEL:eq:prior_normal) (in this case $\text{N}(0,\upsilon_{j}^{2})$) and $\phi$. Using $\gamma$ to denote the slab density and $\star$ to denote the convolution operator, we can write $g=\gamma\star\phi$. It should be stressed that, as shown by Abramovich et al., (1998), using the posterior medians as the pointwise estimates of $\bm{\theta}$ yields a thresholding rule. In other words, we are able to equal the estimated noisy coefficients to zero. In the Empirical Bayes thresholding method by Johnstone and Silverman, 2005a , Johnstone and Silverman, 2005b , the authors propose replacing the Gaussian component in (LABEL:eq:prior_normal) with heavy-tailed distributions, such as the Laplace density. This replacement intends to provide larger estimates for the non-null coefficients than those obtained from Gaussian distributions. In this scenario, considering the Laplace density as the slab component, the prior for each detail wavelet coefficient can be written as $\displaystyle\begin{split}&\pi_{j}\gamma_{a}(\theta_{jk})+(1-\pi_{j})\delta_{0}(\theta_{jk}),\\\ \end{split}$ (6) $k=0,1,\dots,2^{j}-1$, $j=0,1,\dots,J-1$, where $\gamma_{a}(x)$ denotes the Laplace density with scale parameter $a>0$, i.e., $\gamma_{a}(x)=\frac{a}{2}\exp(-a|x|),\quad x\in\mathbb{R}.$ (7) Johnstone and Silverman, 2005a , Johnstone and Silverman, 2005b thresholding method is called Empirical Bayes because the hyperparameters $\pi_{j}$ and $a$ are chosen empirically from the data, using a marginal maximum likelihood approach. Thus, for each resolution level $j$ of the wavelet transform, the arguments $\pi_{j}$ and $a$ that maximize the marginal log-likelihood are selected and plugged back into the prior. Then, the estimation of $\bm{\theta}$ is carried out with either posterior medians, posterior means, or other estimators. Under these circumstances, the posterior distribution is given by $\displaystyle\begin{split}\theta_{jk}|d_{jk}&\sim\pi_{\text{post}}f_{1}(\theta_{jk}|d_{jk})+(1-\pi_{\text{post}})\delta_{0}(\theta_{jk}),\\\ \pi_{\text{post}}&=\frac{\pi_{j}g_{a}(d_{jk}^{*})}{\pi_{j}g_{a}(d_{jk}^{*})+(1-\pi_{j})\phi(d_{jk}^{*})},\\\ \end{split}$ (8) $k=0,1,\dots,2^{j}-1$, $j=0,1,\dots,J-1$, with $f_{1}(\theta_{jk}|d_{jk})$ being the non-null mixture component and $g_{a}=\gamma_{a}\star\phi$. It can be shown that $f_{1}(\theta_{jk}|d_{jk})$ is a mixture of two truncated normal distributions. Define $f_{\text{TN}}(x|\mu,\sigma,\alpha,\beta)$ to be the density of a truncated normal distribution with location parameter $\mu$, scale parameter $\sigma$, minimum value $\alpha$ and maximum value $\beta$. Then, with a slight abuse of notation, we can write $f_{1}(\theta_{jk}|d_{jk})$ as $\displaystyle\begin{split}f_{1}(\theta_{jk}|d_{jk})&=\eta\times f_{\text{TN}}\left(\theta_{jk}\biggl{\rvert}\frac{d_{jk}}{\sigma_{j}}-a,1,0,+\infty\right)\\\ &\quad+(1-\eta)\times f_{\text{TN}}\left(\theta_{jk}\biggl{\rvert}\frac{d_{jk}}{\sigma_{j}}+a,1,-\infty,0\right),\end{split}$ (9) where $\displaystyle\eta=\frac{\exp{(-a\frac{d_{jk}}{\sigma_{j}})}\Phi(\frac{d_{jk}}{\sigma_{j}}-a)}{\exp{(a\frac{d_{jk}}{\sigma_{j}})}\tilde{\Phi}(\frac{d_{jk}}{\sigma_{j}}+a)+\exp{(-a\frac{d_{jk}}{\sigma_{j}})}\Phi(\frac{d_{jk}}{\sigma_{j}}-a)},$ with $\Phi$ denoting the standard normal cumulative function, and $\tilde{\Phi}=1-\Phi$. ## 3 The model Let $y_{1},\dots y_{n}$ be a random sample from the dynamic Gaussian mixture model $\displaystyle\begin{split}y_{t}&=(1-z_{t})x_{1t}+z_{t}x_{2t},\\\ x_{kt}|\mu_{k},\tau_{k}^{2}&\sim\text{N}(\mu_{k},\tau_{k}^{-2}),\quad k=1,2,\\\ z_{t}|\alpha_{t}&\sim\text{Bern}(\alpha_{t}),\quad t=1,\dots,n,\end{split}$ (10) where $z_{t}$’s are allocation variables that indicate to which mixture component the observations $y_{t}$’s belong to. The $z_{t}$ have a Bernoulli distribution with parameter $\alpha_{t}$, the mixture weight that has a dynamic behavior. In (10), the component parameters $\mu_{k}$ and $\tau_{k}^{2}$, $k=1,2$, and the dynamic mixture weights $\alpha_{t}$, $t=1,\dots,n$, are parameters to be estimated. Following Albert and Chib, (1993), we introduce a data augmentation approach by associating an auxiliary variable $l_{t}$ to each allocation variable $z_{t}$. In the original work, $l_{t}=\bm{x}_{t}^{T}\bm{\theta}+e_{t}$ and $e_{t}\sim\text{N}(0,1)$, where $\bm{{x}_{t}}$ is a vector of $p$ known covariates and $\bm{\theta}$ is a vector of $p$ unknown parameters. In greater detail, $z_{t}=1$, if $l_{t}>0$, and $z_{t}=0$, otherwise. However, unlike in Albert and Chib, (1993), where the design matrix $\bm{X}$ in the probit regression corresponds to the covariates related to $\alpha_{t}$, in this paper, $\bm{X}=\bm{W}^{T}$, where $\bm{W}$ is the DWT matrix. Thus, for every $t=1,\dots,n$, we have $\displaystyle\begin{split}l_{t}&=\bm{x}_{t}^{T}\bm{\theta}+e_{t},\\\ e_{t}&\sim\text{N}(0,1),\\\ \end{split}$ (11) where $\bm{x}_{t}$ corresponds to the $t$-th column of matrix $\bm{W}$ and $\bm{\theta}=(c_{00},d_{00},\bm{d}_{1}^{T},\dots,\bm{d}_{J-1}^{T})^{T}$ is the vector of wavelet coefficients, such that $n=p=2^{J}$. Therefore, the dynamic mixture weight $\alpha_{t}$, which is the probability of success of $z_{t}$, is given by the binary regression model, $\alpha_{t}=\Phi(\bm{x}_{t}^{T}\bm{\theta}),$ where $\Phi$ is the standard Gaussian cumulative function. ### 3.1 Bayesian estimation In this paper, the estimation of both component parameters and dynamic mixture weights is performed through a Gibbs sampling algorithm. By giving conjugate prior distributions to the parameters, we sample from their full conditional posterior distributions and make inferences about the parameter values (e.g., point and credible estimates). In this section, we first present the full conditional posterior distributions from which we draw the parameters of (10). Then, we detail the MCMC algorithm built to perform the sampling. In (10), since we are mostly interested in the estimation of the mixture weights, we assume that the sample $\bm{y}=(y_{1},\dots,y_{n})^{T}$ is a time series whose dependence structure is determined by the dynamic behavior of $\alpha_{t}$’s. In this setting, given the component parameters and the dynamic mixture weights, the observations $y_{t}$’s are conditionally independent, and we have $p(\bm{y}|\bm{\mu},\bm{\tau^{2}},\bm{z})=\prod_{t=1}^{n}p(y_{t}|z_{t},\bm{\mu},\bm{\tau^{2}})$. Thus, the complete-data likelihood function $p(\bm{y}|\bm{\mu},\bm{\tau^{2}},\bm{z})$ is given by $\displaystyle\prod\limits_{k=1}^{2}\left(\frac{\tau_{k}^{2}}{2\pi}\right)^{T_{k}/2}\exp{\left[-\frac{\tau_{k}^{2}}{2}\sum\limits_{t:z_{t}=k-1}(y_{t}-\mu_{k})^{2}\right]},$ where $T_{k}=\\#\\{t:z_{t}=k-1,\,t=1,2,...,n\\}$ and $s_{k}=\sum\limits_{t:z_{t}=k-1}y_{t}$ for $k=1,2$. For the complete-data Bayesian estimation of $\bm{\mu}=(\mu_{1},\mu_{2})^{T}$ and $\bm{\tau^{2}}=(\tau^{2}_{1},\tau^{2}_{2})^{T}$, $p(\bm{y}|\bm{\mu},\bm{\tau^{2}},\bm{z})$ is combined with prior distributions to obtain the posteriors. A common issue that arises in the Bayesian estimation of mixture models is the invariance of the mixture likelihood function under the relabelling of the mixture components, known as label switching. To address this problem in our approach, we adopt the simple constraint $\mu_{1}<\mu_{2}$ and reorder the pairs $(\mu_{k},\tau_{k}^{2})$ according to this restriction in the MCMC sampling scheme. Following the usual practice of assigning independent prior distributions to the component parameters (see Escobar and West,, 1995, Richardson and Green,, 2002), we assume $p(\bm{\mu},\bm{\tau_{k}^{2}})=p(\mu_{1})p(\tau_{1}^{2})p(\mu_{2})p(\tau_{2}^{2})$ and place the following priors on $\mu_{k}$ and $\tau^{2}_{k}$, $k=1,2$, $\displaystyle\mu_{k}\sim\text{N}(b_{0k},B_{0k}),$ (12) $\displaystyle\tau_{k}^{2}\sim\Gamma(c_{0k},C_{0k}).$ (13) For the sake of simplicity, hereafter we denote by $[\dots]$ the set of all remaining variables to be considered for the posterior in use. Hence, under the conjugate priors (12) and (13), one obtains the conditional posterior distributions for $\mu_{k}$ and $\tau_{k}^{2}$, $\displaystyle\mu_{k}|[\dots]\sim\text{N}(b_{k},B_{k}),$ (14) $\displaystyle\tau_{k}^{2}|[\dots]\sim\Gamma(c_{k},C_{k}),$ (15) where $\displaystyle\begin{split}B_{k}&=(B_{0k}^{-1}+\tau_{k}^{2}T_{k})^{-1},\\\ b_{k}&=B_{k}(\tau_{k}^{2}s_{k}+B_{0k}^{-1}b_{0k}),\\\ \end{split}\qquad\begin{split}C_{k}&=C_{0k}+\frac{\sum\limits_{t:z_{t}=k-1}(y_{t}-\mu_{k})^{2}}{2},\\\ c_{k}&=c_{0k}+\frac{T_{k}}{2}.\end{split}$ It is worth stressing that assuming the mixture weights to have a dynamic behavior does not interfere with the full conditional posteriors of the component parameters, because they are calculated as in the case of the ordinary (static) mixture model. Given the observations $\bm{y}$, the component parameters $\bm{\mu}$, $\bm{\tau^{2}}$ and $\bm{\alpha}=(\alpha_{1},\dots,\alpha_{n})^{T}$, the $z_{t}$’s are conditionally independent and $p(z_{t}=1|\bm{y},\bm{\mu},\bm{\tau^{2}},\bm{\alpha})\propto\alpha_{t}f_{N}(y_{t}|\mu_{2},\tau_{2}^{-2})$. Thus, one can easily show that, for each $t=1,\dots,n$, the full conditional posterior of $z_{t}$ is given by $\displaystyle\begin{split}z_{t}|[\dots]&\sim\text{Bern}(\beta_{t}),\\\ \beta_{t}&=\frac{\alpha_{t}f_{N}(y_{t}|\mu_{2},\tau_{2}^{-2})}{\alpha_{t}f_{N}(y_{t}|\mu_{2},\tau_{2}^{-2})+(1-\alpha_{t})f_{N}(y_{t}|\mu_{1},\tau_{1}^{-2})}.\end{split}$ (16) The latent variables introduced in (11) are unknown. However, given the vector of wavelet coefficients $\bm{\theta}$ and the allocation data $\bm{z}=(z_{1},\dots,z_{n})^{T}$, we can use the structure of the MCMC algorithm to draw $l_{1},\dots,l_{n}$ from their posterior distribution, which is $\displaystyle\begin{split}l_{t}|[\dots]&\sim\text{N}(\bm{x}_{t}^{T}\bm{\theta},1)\text{ truncated at left by 0 if }z_{t}=1,\\\ l_{t}|[\dots]&\sim\text{N}(\bm{x}_{t}^{T}\bm{\theta},1)\text{ truncated at right by 0 if }z_{t}=0.\end{split}$ (17) For the vector of parameters $\bm{\theta}$, Albert and Chib, (1993) derived the posterior distribution of $\bm{\theta}$ given $\bm{z}$ and $\bm{l}$ under diffuse and Gaussian priors. In this work, on the other hand, $\bm{\theta}$ is a vector of wavelet coefficients. As a result, we need a sparsity inducing prior able to address the noise $e_{t}$ in (11). Thus, following the discussion in Section 2.1, we suggest using spike and slab priors for the components of vector $\bm{\theta}$. In this scenario, we assume that the entries of $\bm{\theta}$ are mutually independent. For $t=2^{j}+k+1$, $k=0,\dots,2^{j}-1$ and $j=0,\dots,J-1$, this kind of prior can be specified as $\theta_{t}\sim(1-\pi_{j})\delta_{0}(\cdot)+\pi_{j}\gamma(\cdot),$ (18) where we consider $\gamma$ to be either the Gaussian distribution or the Laplace distribution as presented in (LABEL:eq:prior_normal) and in (LABEL:eq:prior_laplace), respectively. Following Abramovich et al., (1998), the prior specification is completed by assigning a diffuse prior on the scaling coefficient at the coarsest level $c_{00}$, in the first entry of vector $\bm{\theta}$. Under (18), the posterior distribution of $\theta_{t}$ is given by $\displaystyle\begin{split}\theta_{t}|[\dots]&\sim(1-\pi_{\text{{post}}})\delta_{0}(\theta_{t})+\pi_{\text{{post}}}f_{1}(\theta_{t}|\bm{w}_{t}^{T}\bm{l}),\\\ \pi_{\text{{post}}}&=\frac{\pi_{j}g(\bm{w}_{t}^{T}\bm{l})}{\pi_{j}g(\bm{w}_{t}^{T}\bm{l})+(1-\pi_{j})\phi(\bm{w}_{t}^{T}\bm{l})},\end{split}$ (19) where $\bm{w}_{t}$ is a column-vector corresponding to the $t$-th row of matrix $\bm{W}$, $f_{1}(\theta_{t}|\bm{w}_{t}^{T}\bm{l})$ is the posterior non-null mixture component and $g$ is the convolution between $\gamma$ and the standard normal distribution $\phi$, $g=\gamma\star\phi$. Regarding the hyperparameters of the spike and slab priors, that is, the sparsity parameter $\pi_{j}$ and the variance $\upsilon_{j}^{2}$ (Gaussian component) or the scale parameter $a$ (Laplace component), we follow the approach in Johnstone and Silverman, 2005a , Johnstone and Silverman, 2005b and estimate them jointly by maximizing the marginal log likelihood function, which is given by $\sum\limits_{i=1+2^{j}}^{2^{j+1}}\log\\{(1-\pi_{j})\phi(\bm{w}_{i}^{T}\bm{l})+\pi_{j}g(\bm{w}_{i}^{T}\bm{l})\\}.$ These values are then used in (19) to sample the vector $\bm{\theta}$ in the MCMC procedure, which is detailed in Algorithm 1. Algorithm 1 Gibbs sampling algorithm - Data augmentation 1:Choose number of iterations $N$. 2:Specify initial values for $\bm{\mu}^{(0)},\,{\bm{\tau^{2}}}^{(0)},\,\bm{z}^{(0)}=(z_{1}^{(0)},\dots,z_{n}^{(0)})^{T}$ and $\bm{\alpha}^{(0)}$. 3:for $i\leftarrow 1$ to $N$ do 4: Sample $\mu_{1}^{(i)}\sim p(\mu_{1}|[\dots])$. $\triangleright$ See (14) 5: Sample ${\tau_{1}^{2}}^{(i)}\sim p(\tau_{1}^{2}|[\dots])$. $\triangleright$ See (15) 6: Sample $\mu_{2}^{(i)}\sim p(\mu_{2}|[\dots])$. $\triangleright$ See (14) 7: Sample ${\tau_{2}^{2}}^{(i)}\sim p(\tau_{2}^{2}|[\dots])$. $\triangleright$ See (15) 8: if $\mu_{2}<\mu_{1}$ then 9: Permute the labeling of pairs $(\mu_{k}^{(i)},{\tau_{k}^{2}}^{(i)})$. 10: end if 11: Sample $z_{t}^{(i)}\sim p(z_{t}|[\dots])$, for $t=1,\dots,n$. $\triangleright$ See (16) 12: Sample $l_{t}^{(i)}\sim p(l_{t}|[\dots])$, for $t=1,\dots,n$. $\triangleright$ See (17) 13: Select $\upsilon_{j}^{2}\,/\,a$ and $\pi_{j}$ by marginal maximum likelihood. 14: Sample $\theta_{t}^{(i)}\sim p(\theta_{t}|[\dots])$, for $t=1,\dots,n$. $\triangleright$ See (19) 15: Calculate $\bm{\alpha}^{(i)}=\Phi(\bm{W^{T}\theta})$. $\triangleright$ $\bm{W}$ is the matrix form of the DWT. 16:end for As discussed in Section 2.1, using (18) as prior for $\theta_{t}$ allows the posterior medians to act like thresholding rules, equating to zero noisy coefficients. Because of this, we elect the absolute loss as the Bayes rule estimator for the numerical experiments performed using the MCMC method described in Algorithm 1. ## 4 Numerical Experiments In this section, we illustrate the estimation process discussed in the former sections by conducting Monte Carlo experiments and applying it to a river quota data set to identify flood regimes. In both studies, we implement Algorithm 1 running 6,000 iterations, discarding the first 1,000 as burn-in and performing thinning every 5 draws. We consider the following independent priors for the component parameters: $\mu_{1}\sim N(q_{1},s^{2})$, $\tau_{1}^{2}\sim\Gamma(0.01,0.01)$, $\mu_{2}\sim N(q_{3},s^{2})$, and $\tau_{2}^{2}\sim\Gamma(0.01,0.01)$, where $q_{1}$ and $q_{3}$ are the first and third quartiles, respectively, of the observed data and $s^{2}$ is the sample variance. The purpose of using the data statistics is to reduce subjectivity, and, by adopting the quartiles, to segregate the data into two groups. Concerning the wavelet bases used to perform the transforms, we use the coiflet basis with six vanishing moments. It is important to highlight that, according to other simulated studies, using other Daubechies wavelet bases provides similar results to those achieved by this specific coiflet basis. We do not present these supplementary analyses due to space limitations. ### 4.1 Monte Carlo simulations In our simulated investigations, we generate the artificial data sets by mixing two normally distributed samples of size 1,024, as defined in (10). In this case, we set the following values for the component parameters: $\mu_{1}=0$, $\mu_{2}=2$, $\tau_{1}^{2}=4$ and $\tau_{2}^{2}=4$. Concerning the dynamic mixture weights, we employ three different curves for $\alpha_{t}$: sinusoidal, blocks, and bumps, with the first being defined as $\alpha_{t}=0.4\,\cos(2\pi(t+\pi))+0.5$, and the last two being rescaled test functions introduced by Donoho and Johnstone, (1994). For all three behaviors of $\alpha_{t}$, we run 1,000 Monte Carlo replicates. Additionally, we regard both spike and slab priors, discussed in Section 2.1, for the distribution of the wavelet coefficients, namely: the spike and slab prior with Gaussian slab (SSG), and the spike and slab prior with Laplace slab (SSL). Hereafter, we use the acronyms, SSG and SSL, to refer to these priors. As mentioned in Section 3.1, the point estimates are the medians of the MCMC chains for each Monte Carlo replicate. To appraise the performance of the estimation as a whole, we calculate the average of these point estimates and their 95% HPD intervals. The results for the component parameters are presented in Table 1 and Table 2. It is worth noting that the method, under both priors, performs satisfactorily, with some estimates even coinciding with the parameter values, which, in turn, are encompassed by the HPD intervals in every $\alpha_{t}$’s scenario. Regarding the dynamic mixture weights, Figure 1 shows the results. For the sinusoidal scenario, we see that the method, considering both SSG and SSL priors, succeeds in mimicking the curve’s shapes. Although the bumps and blocks functions are less smooth than the sinusoidal, the method still can satisfactorily estimate their curves. In fact, for the bumps, the point estimates not only follow the sharp shape of the function but also captures the null values correctly. For the blocks scenario, the estimates properly mimic the discontinuity regions and the HPD intervals succeed at encompassing the entire curve. Table 1: Averages of the point estimates (95% HPD credible intervals) for the component parameters $\mu_{1},\tau_{1}^{2},\mu_{2}$ and $\tau_{2}^{2}$, based on 1,000 replications of data sets, considering the SSG prior to $\bm{\theta}$. $\alpha_{t}$’s curve | $\mu_{1}$ = 0 | $\tau_{1}^{2}$ = 4 | $\mu_{2}$ = 2 | $\tau_{2}^{2}$ = 4 ---|---|---|---|--- Sinusoidal | 0.00 (-0.04;0.06) | 4.00 (3.58;4.65) | 2.00 (1.95;2.04) | 4.00 (3.40;4.59) Bumps | 0.00 (-0.04;0.02) | 4.01 (3.59;4.38) | 1.90 (1.60;2.15) | 3.62 (1.06;6.45) Blocks | 0.00 (-0.04;0.06) | 4.06 (3.41;4.71) | 2.00 (1.95;2.06) | 4.00 (3.50;4.63) Table 2: Averages of the point estimates (95% HPD credible intervals) for the component parameters $\mu_{1},\tau_{1}^{2},\mu_{2}$ and $\tau_{2}^{2}$, based on 1,000 replications of data sets, considering the SSL prior to $\bm{\theta}$. $\alpha_{t}$’s curve | $\mu_{1}$ = 0 | $\tau_{1}^{2}$ = 4 | $\mu_{2}$ = 2 | $\tau_{2}^{2}$ = 4 ---|---|---|---|--- Sinusoidal | 0.00 (-0.05;0.05) | 4.05 (3.50;4.62) | 2.00 (1.95;2.04) | 3.99 (3.49;4.50) Bumps | 0.00 (-0.04;0.03) | 3.96 (3.34;4.53) | 1.89 (1.43;2.22) | 3.66 (0.71;6.56) Blocks | 0.02 (-0.15;0.05) | 3.91 (3.40;5.60) | 1.95 (1.28;2.07) | 3.85 (0.82;4.76) Figure 1: Estimates of the $\alpha_{t}$’s provided by SSG prior (left); and SSL prior (right). The curves assigned to $\alpha_{t}$ are, respectively: the sinusoidal (top), the bumps (middle), and the blocks (bottom). The full lines correspond to the $\alpha_{t}$’s curve, the dashed lines correspond to the average of the pointwise estimates and the shaded areas correspond to the 95% HPD intervals. ### 4.2 Taquari quota data set Part of the Taquari-Antas Hydrographic Basin (TAHB) in the state of Rio Grande do Sul (south of Brazil), the Taquari River is located in the upper domain of the Baixo Taquari-Antas Valley, a region that has been affected by an increasing number of extreme rainfall events in recent decades (Tognoli et al.,, 2021). As a result, on many occasions, the rain excess is not drained efficiently and floods riverside regions. This phenomenon is aggravated in urban areas, where the human occupation of floodplains and the soil impermeability contribute to reducing the infiltration capacity and overloading the drainage system, leading to flood inundations (Kurek,, 2016). As reported by Oliveira et al., (2018), Encantado is one of the cities adjacent to the course of the Taquari River most susceptible to fluvial inundations. The geomorphological and topographical characteristics of Encantado’s land favor the water accumulation and restrict its drainage (Oliveira et al.,, 2018). Furthermore, the urbanization of areas with high flood vulnerability in this municipality contributes to intensifying the occurrence of flood inundations (Kurek,, 2016). Because of these circumstances, we propose implementing Algorithm 1 to a time series of Taquari’s river quota to estimate the probability of an inundation regime in Encantado’s urban areas. A river quota is the height of the water body, conventionally measured in centimeters (cm), on a given region of the riverbank. The data set corresponds to the records of Encantado´s fluviometric station identified by the code 86720000. The monthly time series of this station comes from the Hidroweb system, an integrated platform of the National Water Resources Management System (SINGREH) available at https://www.snirh.gov.br/hidroweb/serieshistoricas. Figure 2 shows a map of Encantado, highlighting the station used in this study. Figure 2: Location map of the fluviometric station in the city of Encantado. In the upper-right corner, the Taquari Antas Hydrographic Basin in Rio Grande do Sul state, south of Brazil. To validate the estimated probabilities, we use a report from the Brazilian Geological Survey (CPRM) (Peixoto and Lamberty,, 2019) that records the months when floods occurred in Encantado. Therefore, we can see if the estimates of the mixture weight properly describe the flood regimes, no inundation and inundation, for each month. It is worth highlighting that since inundations can last for a couple of days or even more, there are no records of the specific days when these events took place, only the months. Because of that, and considering that the model is a mixture of two Gaussian distributions, we use the monthly average of the Taquari quota to estimate the probability associated with flood inundations. The period analyzed was from May 2004 to December 2014, consisting of 128 observations. Figure 3 presents this data set. Table 3 shows the point estimates for the component parameters that describe each flood regime. Note that the results provided by the method under the SSG prior are similar to those achieved by it assigning the SSL prior to the distribution of wavelet coefficients. Concerning the dynamic mixture weights, Figure 4 shows the estimates considering both priors for $\bm{\theta}$. By analyzing the results, we see that using the SSL prior allows estimating higher peaks for the probabilities related to inundation periods than using the SSG prior. In fact, under a Bayes classifier, if the method employs the SSG prior, it can detect neither the months when flood episodes were reported nor change points ($\\{t:\alpha_{t}=0.5\\}$). In summary, the method provides results consistent with the data on flood inundations in Encantado available in other works and reports (see Peixoto and Lamberty,, 2019, Tognoli et al.,, 2021). In addition, choosing the Laplace density in the spike and slab prior tends to provide dynamic weight estimates more capable of detecting floods. Figure 3: Monthly average of Taquari’s river quota (cm) from May 2004 to December 2014. Table 3: Medians (95% HPD credible intervals) for the component parameters $\mu_{1},\tau_{1}^{2},\mu_{2}$ and $\tau_{2}^{2}$ of the Taquari quota data set, based on the MCMC samples. Parameters | SSG prior | SSL prior ---|---|--- $\mu_{1}$ | 227.07 (210.09; 242.89) | 220.60 (206.25; 236.28) $\tau_{1}^{2}$ | 2.30e-4 (1.54e-4; 3.15e-4) | 2.58e-4 (1.77e-4; 3.45e-4) $\mu_{2}$ | 405.01 (316.38; 483.35) | 400.20 (355.72; 439.54) $\tau_{2}^{2}$ | 1.14e-4 (2.56e-5; 3.42e-4) | 1.04e-4 (3.65e-5; 1.85e-4) Figure 4: Estimates of the $\alpha_{t}$’s of the Taquari quota data provided by SSG prior (left); and SSL prior (right). The full (black) lines correspond to the point estimates (medians) and the dashed (blue) lines mark the months when flood inundations were reported by Peixoto and Lamberty, (2019). ## 5 Conclusion This paper presents an approach to identify regime switches in bimodal data sets. We use a two-component mixture model whose mixture weight varies according to some index, like time. This adaptation makes the model more flexible and adaptive to a broader range of clustering and classification problems. Furthermore, we use wavelet bases to estimate the dynamic behavior of the mixture weight due to their excellent properties when it comes to curves’ estimation. However, unlike other approaches in the literature that also rely on wavelets (see Montoril et al.,, 2019), here we consider a Bayesian framework and propose estimating the dynamic weights and the component parameters jointly through an efficient Gibbs sampling algorithm. We analyze the performance of this MCMC algorithm by conducting Monte Carlo experiments and illustrate the approach with an application to a river quota data set. Results from the simulations show that the method provides good estimates for the component parameters and the dynamic weights even when the function behind $\alpha_{t}$’s behavior is rougher. Additionally, the estimation performance using SSG prior is similar to the performance achieved when SSL prior is employed. The same does not apply to the results obtained in the river quota data set. For this application, we notice that implementing the method under the SSG prior to the wavelet coefficients yields smaller values for the probabilities associated with inundations occurrence than the estimates provided by using the SSL prior. This is likely because the Gaussian distribution does not have heavy tails, unlike the Laplace distribution. ## References * Abramovich et al., (2000) Abramovich, F., Bailey, T. C., and Sapatinas, T. (2000). Wavelet analysis and its statistical applications. Journal of the Royal Statistical Society: Series D (The Statistician), 49(1):1–29. * Abramovich et al., (1998) Abramovich, F., Sapatinas, T., and Silverman, B. W. (1998). Wavelet thresholding via a bayesian approach. Journal of the Royal Statistical Society. Series B (Statistical Methodology), 60(4):725–749. * Albert and Chib, (1993) Albert, J. H. and Chib, S. (1993). Bayesian analysis of binary and polychotomous response data. Journal of the American Statistical Association, 88(422):669–679. * Bordes et al., (2006) Bordes, L., Delmas, C., and Vandekerkhove, P. (2006). Semiparametric estimation of a two-component mixture model where one component is known. Scandinavian Journal of Statistics, 33(4):733–752. * Cai and Brown, (1999) Cai, T. and Brown, L. D. (1999). Wavelet estimation for samples with random uniform design. Statistics & Probability Letters, 42(3):313–321. * (6) Donoho, D. L. (1993a). Nonlinear wavelet methods for recovery of signals, densities, and spectra from indirect and noisy data. In In Proceedings of Symposia in Applied Mathematics, pages 173–205. American Mathematical Society. * (7) Donoho, D. L. (1993b). Unconditional bases are optimal bases for data compression and for statistical estimation. Applied and Computational Harmonic Analysis, 1(1):100–115. * Donoho et al., (1996) Donoho, D. L., Johnstone, I. M., Kerkyacharian, G., and Picard, D. (1996). Density estimation by wavelet thresholding. The Annals of Statistics, 24(2):508–539. * Donoho and Johnstone, (1994) Donoho, D. L. and Johnstone, J. M. (1994). Ideal spatial adaptation by wavelet shrinkage. Biometrika, 81(3):425–455. * Escobar and West, (1995) Escobar, M. D. and West, M. (1995). Bayesian density estimation and inference using mixtures. Journal of the American Statistical Association, 90(430):577–588. * George and McCulloch, (1993) George, E. I. and McCulloch, R. E. (1993). Variable selection via gibbs sampling. Journal of the American Statistical Association, 88(423):881–889. * Hall and Patil, (1995) Hall, P. and Patil, P. (1995). Formulae for mean integrated squared error of nonlinear wavelet-based density estimators. The Annals of Statistics, 23(3):905–928. * Hall and Zhou, (2003) Hall, P. and Zhou, X.-H. (2003). Nonparametric estimation of component distributions in a multivariate mixture. The Annals of Statistics, 31(1):201–224. * Härdle et al., (2012) Härdle, W., Kerkyacharian, G., Picard, D., and Tsybakov, A. (2012). Wavelets, approximation, and statistical applications, volume 129\. Springer Science & Business Media. * Hughes and Guttorp, (1994) Hughes, J. P. and Guttorp, P. (1994). A class of stochastic models for relating synoptic atmospheric patterns to regional hydrologic phenomena. Water Resources Research, 30(5):1535–1546. * Hui and Zhou, (1998) Hui, S. L. and Zhou, X.-H. (1998). Evaluation of diagnostic tests without gold standards. Statistical Methods in Medical Research, 7(4):354–370. * (17) Johnstone, I. and Silverman, B. W. (2005a). Ebayesthresh: R programs for empirical bayes thresholding. Journal of Statistical Software, 12:1–38. * (18) Johnstone, I. M. and Silverman, B. W. (2005b). Empirical bayes selection of wavelet thresholds. The Annals of Statistics, 33(4):1700–1752. * Kurek, (2016) Kurek, R. K. M. (2016). Analysis of floods in the vale do taquari/rs as subsidy the development of a forecast model. Master’s thesis, Federal University of Santa Maria. * Mallat, (1989) Mallat, S. G. (1989). A theory for multiresolution signal decomposition: the wavelet representation. IEEE transactions on pattern analysis and machine intelligence, 11(7):674–693. * Montoril et al., (2021) Montoril, M. H., Correia, L. T., and Migon, H. S. (2021). Bayesian estimation of dynamic weights in gaussian mixture models. * Montoril et al., (2019) Montoril, M. H., Pinheiro, A., and Vidakovic, B. (2019). Wavelet-based estimators for mixture regression. Scandinavian Journal of Statistics, 46(1):215–234. * Morettin, (1996) Morettin, P. A. (1996). From fourier to wavelet analysis of time series. In Prat, A., editor, COMPSTAT, pages 111–122, Heidelberg. Physica-Verlag HD. * Nagy et al., (2011) Nagy, I., Suzdaleva, E., Kárný, M., and Mlynářová, T. (2011). Bayesian estimation of dynamic finite mixtures. International Journal of Adaptive Control and Signal Processing, 25(9):765–787. * Ogden, (1997) Ogden, R. T. (1997). Essential wavelets for statistical applications and data analysis. Birkhäuser Boston Inc. * Oliveira et al., (2018) Oliveira, G. G. d., Eckhardt, R. R., Haetinger, C., and Alves, A. (2018). Caracterização espacial das áreas suscetíveis a inundações e enxurradas na bacia hidrográfica do rio taquari-antas. Geosciences= Geociências, 37(4):849–863. * Patra and Sen, (2016) Patra, R. K. and Sen, B. (2016). Estimation of a two-component mixture model with applications to multiple testing. Journal of the Royal Statistical Society. Series B (Statistical Methodology), 78(4):869–893. * Peixoto and Lamberty, (2019) Peixoto, C. A. B. and Lamberty, D. (2019). Setorização de áreas de alto e muito alto risco a movimentos de massa, enchentes e inundações: Encantado, rio grande do sul. Technical report, CPRM, Porto Alegre. * Percival and Walden, (1999) Percival, D. B. and Walden, A. T. (1999). Wavelet Methods for Time Series Analysis. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press. * Priestley, (1996) Priestley, M. B. (1996). Wavelets and time-dependent spectral analysis. Journal of Time Series Analysis, 17(1):85–103. * Richardson and Green, (2002) Richardson, S. and Green, P. J. (2002). On Bayesian Analysis of Mixtures with an Unknown Number of Components (with discussion). Journal of the Royal Statistical Society: Series B (Methodological), 59(4):731–792. * Rindskopf and Rindskopf, (1986) Rindskopf, D. and Rindskopf, W. (1986). The value of latent class analysis in medical diagnosis. Statistics in Medicine, 5(1):21–27. * Tognoli et al., (2021) Tognoli, F. M. W., Bruski, S. D., and Araujo, T. P. d. (2021). Data analysis reveals that extreme events have increased the flood inundations in the taquari river valley, southern brazil. Latin American Data in Science, 1(1):16–25. * Vidakovic, (1999) Vidakovic, B. (1999). Statistical modeling by wavelets, volume 503. John Wiley & Sons. * Walker et al., (2009) Walker, M. G., Mateo, M., Olszewski, E. W., Sen, B., and Woodroofe, M. (2009). Clean kinematic samples in dwarf spheroidals: An algorithm for evaluating membership and estimating distribution parameters when contamination is present. The Astronomical Journal, 137(2):3109.
# A Dynamic Model of Performative Human-ML Collaboration: Theory and Empirical Evidence Tom Sühr MPI for Intelligent Systems, Tübingen Tübingen AI Center <EMAIL_ADDRESS> &Samira Samadi MPI for Intelligent Systems, Tübingen Tübingen AI Center <EMAIL_ADDRESS> &Chiara Farronato Harvard Business School Harvard University, CEPR, NBER <EMAIL_ADDRESS> ###### Abstract Machine learning (ML) models are increasingly used in various applications, from recommendation systems in e-commerce to diagnosis prediction in healthcare. In this paper, we present a novel dynamic framework for thinking about the deployment of ML models in a performative, human-ML collaborative system. In our framework, the introduction of ML recommendations changes the data generating process of human decisions, which are only a proxy to the ground truth and which are then used to train future versions of the model. We show that this dynamic process in principle can converge to different stable points, i.e. where the ML model and the Human+ML system have the same performance. Some of these stable points are suboptimal with respect to the actual ground truth. We conduct an empirical user study with 1,408 participants to showcase this process. In the study, humans solve instances of the knapsack problem with the help of machine learning predictions. This is an ideal setting because we can see how ML models learn to imitate human decisions and how this learning process converges to a stable point. We find that for many levels of ML performance, humans can improve the ML predictions to dynamically reach an equilibrium performance that is around 92% of the maximum knapsack value. We also find that the equilibrium performance could be even higher if humans rationally followed the ML recommendations. Finally, we test whether monetary incentives can increase the quality of human decisions, but we fail to find any positive effect. Our results have practical implications for the deployment of ML models in contexts where human decisions may deviate from the indisputable ground truth. ## 1 Introduction Human-ML collaboration is increasingly used in various applications, from content moderation in social media [1] to predicting diagnoses in healthcare [2, 3] and making hiring decisions in human resources [4]. Companies that implement human-ML collaborative systems face three crucial challenges: 1) ML models learn from past human decisions, which are often only an approximation to the ground truth (noisy labels); 2) ML models are rolled out to help future human decisions, affecting the data-generating process of human-ML collaboration that then influences future updates to ML models (performative predictions [5]); and 3) the quality of the human-ML collaborative prediction of the ground truth may change as a function of incentives and other human factors. These challenges create a dynamic learning process. Without access to the ground truth, it is often difficult to know whether the learning process will reach an equilibrium state with a good approximation of the ground truth, if it is interrupted at a sub-optimal level, or if it does not reach a stable state at all. For intuition, we can focus on the decision of a healthcare company to develop and deploy an ML model to predict medical diagnoses from patient visits. The problem is made difficult by the fact that a doctor’s diagnoses can be wrong, and it is often too costly or time-consuming to identify the indisputable ground truth—i.e., the underlying true diagnosis of a patient—so the company typically uses all diagnoses to train their ML model, without distinction between good or bad diagnoses. In addition, the company typically evaluates the algorithm’s performance based on its ability to match those same doctor diagnoses, potentially replicating their mistakes. The dynamic deployment of updates to ML models that support doctor diagnoses could lead to a downward spiral of human+ML performance if the company deploys a bad model and the bad model adversely affects doctor decisions. Or, it can lead to continuous improvement until it reaches a stable point that is a good approximation to the indisputable ground truth. Without (potentially costly) efforts to measure the ground truth, the company has no way of distinguishing between downward spirals or continuous improvements. This raises a multitude of empirical questions regarding the governing mechanisms of this dynamic system. How do humans improve on ML predictions of different quality levels, and do financial incentives matter? Will the dynamic learning process converge to a good equilibrium even without the company knowing the actual ground truth labels? How do we design an experiment to test such a system in an empirical context? #### Contributions. In this paper, we present a novel framework for thinking about ML deployment strategies in a performative, human-AI collaborative system. We present a simple theoretical framework to identify conditions under which ML deployment strategies converge to stable points that are a good approximation to the ground truth, and conditions under which there are downward spirals away from the ground truth. To validate our theory, we provide an empirical study in which humans solve knapsack problems with the help of machine learning predictions. We conducted a user study with 1,408 participants, each of whom solved 10 knapsack problems. Knapsack problems are particularly suited to explore our questions because we can train machine learning models to replicate human decisions, while having access to the underlying ground truth (the optimal solution) to evaluate the learning process. Additionally, knapsack problems can be hard for humans, making the task not obvious; and problem instances can be generated and labeled perfectly at negligible cost. We highlight both theoretical and empirical contributions. Our theoretical framework introduces the _collaborative characteristic function_ as the function mapping the performance of ML models with respect to the ground truth to the performance of humans using those models when making decisions (human+ML). We define _utility_ the performance of decisions (made by ML alone, human alone, or human+ML) with respect to the ground truth. We note that utility cannot be easily quantified for many practical applications. Finally, we introduce the notion of _collaborative learning paths_ , each of which characterizes a possible dynamic deployment strategy. We show conditions under which this dynamic system theoretically reaches a stable point of utility for the firm. On the empirical side, we show how low-quality ML models can make humans perform worse than if they had no access to ML recommendations. Still, humans improve on ML models so that the deployment strategy of ML models with initial performance between 72% and 92% will converge to a performance that is around 92% of the value of the optimal knapsack solution. Two empirical findings constrain the system from converging to an even higher performance. First, humans do not respond to financial incentives for performance. Second, humans sometimes make decisions that are worse than the ML recommendation, despite the fact that it is fairly easy for them to compare their solution to the ML suggestion and pick the best of the two. Our results have practical implications for the deployment of ML models when humans are influenced by those models but their decisions deviate from an unknown ground truth. First, performance metrics of ML models can be misleading when the learning objective is based on comparisons against human decisions and those decisions can be wrong. Companies should thus exert efforts to assess the quality of human decisions and take that into account when training ML models. For example, in the medical setting, human diagnoses should be first verified or confirmed by external experts, or patients should be followed up to confirm the validity of initial diagnoses. At a minimum, ML models should be trained on subsets of data for which there is enough confidence that the decisions are correct. Second, our work highlights the strategic importance of deploying ML models that allow for convergence to a stable point with higher utility than humans alone. Such convergence is not guaranteed and, as argued above, difficult to assess. Third, our work calls for the need to adopt a dynamic approach when deploying algorithms that interact with human decisions, and those interactions are used for future model building. ## 2 Related Work There has been a growing body of work investigating various forms of human-AI collaboration. From learning-to-defer systems, where a model defers prediction tasks to humans if its own uncertainty is too high[6, 7, 8], to AI-assisted decision making where humans may or may not consult ML predictions to make a decision [9, 3, 2]. Several alternative decision mechanisms have also been explored [10, 11]. The application areas range from programming [12, 13], to healthcare [2, 3] and business consulting [14]. Related work also investigates factors influencing human-ML collaboration, such as explanations of ML predictions [15], monetary incentives [16], fairness constraints [17], and humans’ adaptability to model changes [18]. In this work, for the first time to the best of our knowledge, we examine the human+ML interaction from a dynamic perspective, where ML models learn from human decisions that are 1) the result of previous human+ML collaboration and 2) can arbitrarily deviate from the underlying ground truth. This paper is also inspired by an extensive line of work on performative prediction [5, 19, 20, 21], a theoretical framework in which predictions influence the outcome they intend to predict. We adapt the ideas of performative prediction to a context of human-ML collaboration and extend it in three major ways: 1) In our setting, the model predictions change the quality of the human-ML labels as a proxy for the ground truth (e.g., a doctor diagnosis), but the ground truth is held constant (e.g., the true patient diagnosis); 2) We introduce the concept of utility, to quantify the quality of a solution with respect to the ground truth. There can be several stable points with respect to model parameters in the performative prediction framework, but not all of them are equally good at approximating the indisputable ground truth; and 3) The ground truth is unknown, and the mapping between human or ML labels and the ground truth is not fixed. To the best of our knowledge, we are the first to explore performative predictions where the company deploying ML models is unaware of the model’s performance relative to the ground truth, and only knows its similarity to human labels. Our empirical application is also novel in that it investigates the implications of performative predictions for human-ML collaboration. ## 3 Problem Statement We consider a setting in which time is separable in discrete time epochs $t=1,...,T$. At each $t$, a firm deploys machine learning model $M_{t}\in\mathcal{M}$ with $M_{t}:\mathcal{X}\rightarrow\mathcal{Y}$. The model $M_{t}$ predicts a solution $Y\in\mathcal{Y}$ (e.g., a diagnosis) to a problem $X\in\mathcal{X}$ (e.g. the patient’s symptoms) as a function of past data. The firm employs expert humans $H\in\mathcal{H}$ with $H:\mathcal{X}\times\mathcal{Y}\rightarrow\mathcal{Y}$, who solve the problems with the help of the ML predictions. We will write $M_{t}(X)=Y_{M_{t}}$ and $H(X,Y_{M_{t}})=Y_{H_{t}}$. We assume that for all $X\in\mathcal{X}$, there exists an optimal solution $Y^{*}$, which is the undisputable ground truth. #### The Firm’s Learning Objective. In many real world applications, determining the ground truth label $Y^{*}$ can be extremely costly. For example, obtaining the correct medical diagnosis can often require the knowledge of various specialists (e.g., orthopedists, pediatricians, neurologists). Even when a single expert is enough, they can misdiagnose a patient’s symptoms. Yet, in many of these cases, using the human labels $Y_{H_{t}}$ as a proxy for $Y^{*}$ is the only feasible option to build ML models. We allow for the quality of $Y_{H_{t}}$ with respect to $Y^{*}$ to change. This means that two iterations of the ML model, $M_{t}$ and $M_{t+1}$, are trained on data from two different data generating processes, $(X,Y_{H_{t-1}})\sim~{}D_{t-1}$ and $(X,Y_{H_{t}})\sim D_{t}$, respectively. Without access to $Y^{*}$, the only feasible learning objective for a firm that wants to update its model parameters at time $t$ is the comparison between the latest human-ML collaborative labels with the new predictions.111We assume that models at time $t$ are trained exclusively on data from the previous period $t-1$, although we can generalize our setting to include any data points from 0 to $t-1$. For a given loss function $\mathit{l}:\mathcal{Y}\times\mathcal{Y}\rightarrow\mathbb{R_{+}}$ that is $L(Y_{M_{t}},Y_{H_{t-1}}):=\underset{H\in\mathcal{H}}{\mathbb{E}}[\underset{(X,Y_{H_{t-1}})\sim D_{t-1}}{\mathbb{E}}\mathit{l}(Y_{M_{t}},Y_{H_{t-1}})].$ (1) The firm wants to minimize the difference between the model predictions at time $t$ and the human labels at time $t-1$. We can write the firm’s problem as selecting a model $M_{t}$ to minimize the loss function in Equation 1: $\underset{M_{t}\in\mathcal{M}}{minimize}\text{ }L(Y_{M_{t}},Y_{H_{t-1}}).$ (2) For simplicity, we assume that at each time $t$, the firm collects enough data to perfectly learn the human-ML solution. In other words, with the optimal model, $L(Y_{M_{t}},Y_{H_{t-1}})=0$. We return to this assumption in Appendix A.7. #### Utility. In our scenario, the firm cannot quantify the true quality of a solution $Y$ with respect to $Y^{*}$. The loss in Equation 2 is just a surrogate for the loss $L(Y,Y^{*})$, which is impossible or too costly to obtain. The firm thus defines the human label as "ground truth," and maximizes the similarity between model and human solutions, without knowing how close the human or ML solutions are to the indisputable ground truth. In order to evaluate the firm’s progress in approximating $Y^{*}$, it is useful to define a measure of utility. ###### Definition 1. (Utility) Let $d_{X}$ be a distance measure on $\mathcal{Y}$ with respect to a given $X\in\mathcal{X}$. The function $\mathbb{U}:\mathcal{X}\times\mathcal{Y}\rightarrow\mathbb{R}$ is a utility function on $\mathcal{X}\times\mathcal{Y}$, if $\forall X\in\mathcal{X},Y_{min},Y,Y^{\prime},Y^{*}\in\mathcal{Y}$ 1. 1. $\exists Y_{min}\in\mathcal{Y}:\mathbb{U}(X,Y)\in[\mathbb{U}(X,Y_{min}),\mathbb{U}(X,Y^{*})]$ (bounded) 2. 2. $\exists\varepsilon>0:|d_{X}(Y,Y^{*})-d_{X}(Y^{\prime},Y^{*})|<\varepsilon\Rightarrow\mathbb{U}(X,Y)=\mathbb{U}(X,Y^{\prime})$ ($\varepsilon$-sensitive) 3. 3. $d_{X}(Y,Y^{*})+\varepsilon<d_{X}(Y^{\prime},Y^{*})\Rightarrow\mathbb{U}(X,Y)>\mathbb{U}(X,Y^{\prime})$ (proximity measure) The utility of a solution for the firm is maximal if $Y$ is $\varepsilon$-close to $Y^{*}$ with respect to the underlying problem $X$. The variable $\varepsilon$ should be interpreted as the threshold below which a firm perceives no difference between two outcomes, i.e., it does not care about infinitely small improvements. #### Collaborative Characteristic Function. As time $t$ increases, the firm hopes that the distributions $D_{t}$ shift closer to the optimal distribution $D^{*}$, where $(X,Y)=(X,Y^{*})$. In other words, for each model’s distance $d$, $d(D_{t},D^{*})>d(D_{t+1},D^{*})$. This could happen, for example, if humans were able to easily compare available solutions and pick the one that is closest to the indisputable ground truth. We can translate this continuous improvement into properties of the human decision function $H$ as follows: for all $t=1,...,T$ and $X\in\mathcal{X}$, $\underset{H\in\mathcal{H}}{\mathbb{E}}[\mathbb{U}(X,H(X,Y_{M_{t}}))]=\mathbb{U}(X,Y_{M_{t}})+\delta_{M_{t}}.$ (3) The firm’s hope is that $\delta_{M_{t}}\geq 0$ for $M_{t}$. Effectively, $\delta_{M_{t}}$ characterizes the human-ML collaboration for all utility levels of a model. If $\delta_{M_{t}}$ is positive, humans are able to improve on a ML prediction (and future model iterations will thus get better at approximating the ground truth). Instead, if $\delta_{M_{t}}$ is negative, humans will perform worse than the ML recommendations, and future model iterations will get progressively farther away from the ground truth. We define the function given by Equation 3 as the collaborative characteristic function: ###### Definition 2. (Collaborative Characteristic Function) For a utility function $\mathbb{U}$, humans $H\in\mathcal{H}$, we define the collaborative characteristic function $\Delta_{\mathbb{U}}:\mathbb{R}\rightarrow\mathbb{R}$ as follows: $\Delta_{\mathbb{U}}(\mathbb{U}(X,Y_{M}))=\underset{H\in\mathcal{H}}{\mathbb{E}}[\mathbb{U}(X,H(X,Y_{M}))]=\mathbb{U}(X,Y_{M})+\delta_{M}.$ The function $\Delta_{\mathbb{U}}$ can take any arbitrary form. Several factors can affect $\Delta_{\mathbb{U}}$, e.g., monetary incentives and ML explanations. In subsequent sections, we empirically approximate one such function. #### Collaborative Learning Path and Stable Points. Although $\Delta_{\mathbb{U}}$ has infinite support, a firm will only experience a discrete set of utility values achieved by humans with the help of ML recommendations. We call this the collaborative learning path. It is characterized by $\Delta_{\mathbb{U}}$, the utility of the first deployed model $s$, and the number of deployment cycles $T$: ###### Definition 3. (Collaborative Learning Path) Let $\Delta_{\mathbb{U}}$ be a collaborative characteristic function, $t=1,...,T\in\mathbb{N}_{\geq 1}$ the number of deployment cycles and $s=\mathbb{U}(X,Y_{M_{1}})$ the utility of the starting model. We define the collaborative learning path to be the function $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)=\underset{H\in\mathcal{H}}{\mathbb{E}}[\underset{X\in\mathcal{X}}{\mathbb{E}}(\mathbb{U}(H(X,Y_{M_{t}}))].$ ###### Definition 4. (Stable Point) A stable point $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$ occurs at $t$ if for all $t^{\prime}\geq t$, $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t^{\prime})=\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$. Stable points are states where the utility remains constant in all future model deployments. If $Y^{*}$ is unique for all $X$, then this is also a stable point for the distribution shifts. Whether a firm can reach a stable point on its collaborative learning function depends on the shape of $\Delta_{\mathbb{U}}$ and the initial model utility $s$. Figure 1 shows two examples of collaborative characteristic functions and collaborative learning paths. The 45-degree line includes the points where $\underset{X,H}{\mathbb{E}}[\mathbb{U}(X,H(X,Y))]=\underset{X}{\mathbb{E}}[\mathbb{U}(X,Y)]$. Stable points will always lie on this line, because a stable point requires $\delta_{t}\approx 0$ ($|\delta_{t}|\leq\epsilon$, where $\epsilon$ is defined in Appendix A.5 and denotes the smallest change in utility that is possible for a given $\varepsilon$ from definition 1). If $|\delta_{t}|>\epsilon$, it indicates that humans’ influence changes labels $Y$ relative to the most recent ML model, leading to a new data distribution. The model at $t+1$ will thus differ from $M_{t}$, preventing stability. When the model and human+ML labels differ, there are two possible cases. First, $\delta_{M_{t}}>\epsilon$, which implies that the collaborative characteristic function $\Delta_{\mathbb{U}}$ is above the 45-degree line on that portion of the domain (figure 1(a)). In this case, human+ML labels are closer to the indisputable ground truth than the model alone, which leads to improvements of subsequent model deployments. Second, if $\delta_{M_{t}}<-\epsilon$, the collaborative characteristic function is below the 45-degree line (figure 1(b)). In this case, human+ML labels are further away from the indisputable ground truth than the model alone, which leads to deterioration of subsequent model deployments. (a) Collaborative Improvement (b) Collaborative Harm Figure 1: Collaborative Improvement (left): The firm’s collaborative characteristic function and one collaborative learning path, if humans improve on the ML solution. The x-axis denotes the model expected utility, the y-axis denotes expected human+ML utility. The firm deploys a first model with utility (s). Then humans use the model and improve utility by $\delta_{1}$, leading to expected human+ML utility (1). The firm learns a new model with utility (b) on the new data distribution. This is viable under the assumption that the new model has the same utility as the previous period’s human+ML labels, i.e., we can move horizontally from (1) to the 45-degree line at (b). Humans can further improve utility by $\delta_{2}$, which leads to expected utility (2). The dynamic improvement process continues until it reaches stable point utility (6-d). Collaborative Harm (right): The firm deploys a model with expected utility (s) but the humans, when interacting with the model, decrease utility by $\delta_{1}$, with expected utility (1). The firm will thus learn a model of utility (b) on the new distribution. The downward spiral continues until stable point (d). We present the best-case and worst-case scenarios from Figure 1 as Propositions 1 and 2 below: ###### Proposition 1. (Collaborative Improvement) If $\Delta_{\mathbb{U}}(\mathbb{U}(X,Y_{M}))\geq\mathbb{U}(X,Y_{M})$ for all $M\in\mathcal{M},X\in\mathcal{X}$. Then $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$, is non-decreasing with $t=1,...,T$ and for sufficiently large $T$ it exists a $t^{\prime}\in[1,T]$ such that $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t^{\prime})$ is a stable point. ###### Proof. (sketch) Because $\mathbb{U}$ is bounded, $\delta_{M}$ must be 0 in the extreme points. Furthermore, because of the $\varepsilon$-sensitivity of $\mathbb{U}$, the steps $t$ until reaching the maximum utility are also bounded. It follows that there exists a $t\in\mathbb{N}$ such that $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)-\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t+1)=0$, which is a stable point. See Appendix A.6 for the complete proof. ∎ ###### Proposition 2. (Collaborative Harm) If $\Delta_{\mathbb{U}}(\mathbb{U}(X,Y_{M}))\leq\mathbb{U}(X,Y_{M})$ for all $M\in\mathcal{M},X\in\mathcal{X}$. Then $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$, is non-increasing with $t=1,...,T$ and for sufficiently large $T$ it exists a $t^{\prime}\in[1,T]$ such that $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t^{\prime})$ is a stable point. ###### Proof. Similar to the proof of Proposition 1. ∎ In practice, a firm’s collaborative characteristic function can take any arbitrary shape, with portions above and portions below the 45-degree line. As long as the function is continuous, at least one stable point exists, and possibly more. When more than one stable point exist, the firm would like to reach the stable point with the highest utility (i.e., the highest point of the characteristic function lying on the 45-degree line). However, since the firm does not have access to the indisputable ground truth, when it reaches a stable point it does not know where such point lies on the 45-degree line. In what follows, we empirically explore a context where 1) the ML model learns to imitate human decisions, and 2) it is easy for us to identify the indisputable ground truth. The setting allows us to draw at least a portion of the collaborative characteristic function, and explore the effects of human behavior on its shape, particularly the effect of monetary incentives and alternative selection criteria. We present study participants with instances of hard knapsack problems to answer the following research questions: RQ1: How do monetary incentives affect human performance? To keep our treatment condition manageable, we explore the effect of different levels of performance bonuses on $\mathbb{U}(H(X,.))$, i.e., the human performance without ML recommendations. RQ2: What is the shape of the human-ML collaborative characteristic function $\Delta_{\mathbb{U}}$? Here, we hold the performance bonus constant, and test humans’ effect $\delta_{M}$ on utility for different levels of ML performance. This will enable us to approximate $\Delta_{\mathbb{U}}$ for a specific task. ## 4 Experimental Setup In this section, we describe our user study. The goal of our experimental setup is to simulate an environment in which users work on difficult tasks with the help of ML. The company responsible for deploying ML models does not know the optimal solution $Y^{*}$ (e.g., the true patient’s diagnosis), and it trains ML models to replicate experts’ decisions (doctor diagnoses). To evaluate how the company’s models perform against $Y^{*}$, we need a setting in which we, as researchers, know the quality of any solution Y using a utility function $\mathbb{U}(X,Y)$. This allows us to make absolute quality assessments of solutions. Note that this is often unattainable in practice, as we argued in the introduction. The Knapsack Problem is particularly well suited for this context. #### The Knapsack Problem. In our experiment, users solve instances of the knapsack problem. An instance involves selecting which of $n=18$ items to pack into a knapsack, each with a weight $w$ and a value $v$. The objective is to maximize value without exceeding the weight limit $W$ of the knapsack (between 5 and 250). We focus on the one-dimensional 0-1 knapsack problem, in which participants choose which items to pack (see Appendix 6 for a formal definition). We constrain the weights, values, and capacity of our instances to integer values, to make them easier to interpret by humans. We describe the details of the knapsack problem generation in Appendix A.10. The knapsack problem has desirable properties for the empirical application of our dynamic framework. First, users do not require special training—beyond a short tutorial—to find a solution to the problem. Yet, the task is hard for humans, especially with a growing number of items [22]. Thus, the optimal solution $Y^{*}$ is not obvious. Second, we can generate solutions to the knapsack problem in two ways. The “optimal” solution can be found with dynamic programming. The “ML” solution can be found by imitating what humans select and computing the training loss as the difference between the items selected by participants versus items selected by a model. This setup allows us to quantify the utility of the proposed solution relative to the optimal solution. We define utility for the knapsack problem as follows: ###### Definition 5. (Economic Performance) For a knapsack instance $X=((w_{1},\cdots,w_{n}),(v_{1},\cdots,v_{n}),W)$ with optimal solution $\underset{x_{1},\cdot,x_{n};\sum_{i=1}^{n}x_{i}w_{i}\leq W}{\max}\sum_{i=1}^{n}x_{i}v_{i}=:Y^{*}$ and a valid solution $Y$ we call the function $\mathbb{U}_{\text{Econ}}(X,Y)=\frac{Y}{Y^{*}}$ the economic performance of $Y$ given $X$. Appendix A.4 contains details about $\mathbb{U}_{\text{Econ}}(X,Y)$ and discusses our results using an alternative utility function. Note that there can be multiple optimal combinations of items to pack, but the optimal value $Y^{*}$ is unique. #### Study Design. We recruited participants from Prolific222https://www.prolific.com/ exclusively from the UK to ensure familiarity with the currency and weight metrics used to describe the knapsack items and monetary incentives in the study. Appendix A.11 presents screenshots of the web interface for each step of the study. At the beginning of the study, participants received a tutorial on the knapsack problem, our web application’s interface, and the payment structure, described below. After the tutorial, the participants solved two practice problems and received feedback on their submission’s performance. For the main task, each participant received 10 knapsack problems generated by Algorithm 1. For each problem, they had 3 minutes to submit their solution. If the participant did not actively click on the submit button, the selected items were automatically submitted at the 3-minute mark. Participants could take unlimited breaks between problems. At the end of the study, we asked participants about their demographics, previous experience with the knapsack problem, and how much effort they put in solving the task. Every participant received a base payment of £2.00 (approx. $2.50) if they achieved at least 70% of the value of the optimal solution, averaged across the 10 knapsack instances they solved. We set the 70% threshold to discourage participants from randomly selecting items, as randomly-generated solutions that pick items until reaching the weight capacity have an average $\mathbb{U}_{\text{Econ}}$ around 60%. Model | None | q1 | q2 | q3 | q4 | q5 | q6 ---|---|---|---|---|---|---|--- Mean $\mathbb{U}_{\text{Econ}}(X,Y)$ | . | 0.717 | 0.800 | 0.844 | 0.884 | 0.899 | 0.920 SD | . | 0.083 | 0.105 | 0.098 | 0.105 | 0.088 | 0.085 No Bonus | N=102 | | | | | | 2-cent Bonus | N=98 | | | | | | 10-cent Bonus∗ | N=100+117 | N=64 | N=78 | N=194 | N=179 | N=70 | N=191 20-cent Bonus | N=96 | | | | | | Table 1: Matrix of treatment conditions. The columns denote information on the ML recommendation performance. The rows denote bonus payments for performance. The number of study participants are presented in the relevant cells. ∗We ran the 10-cent bonus treatment with no ML recommendation twice: once without a comprehension quiz for the bonus structure (100 participants) and once with the comprehension quiz (117). Participants were randomly allocated into four monetary treatments and seven algorithmic recommendations (see Table 1). All monetary conditions were tested while users had no access to algorithmic recommendations. Participants in the No Bonus condition did not receive any additional payments beyond the base payment. Participants in the 2-cent Bonus condition received an additional £0.02 for each percentage point of $\mathbb{U}_{\text{Econ}}$ above 70%. For example, if a participant achieved on average $\mathbb{U}_{\text{Econ}}$ = 85%, they would receive $\text{\textsterling}2.00+15\times\text{\textsterling}0.02=\text{\textsterling}2.30$. Participants in the 10-cent Bonus and 20-cent Bonus treatments had similar incentives for performance, but higher monetary rewards for each additional percentage point increase in performance (£0.10 and £0.20, respectively). We ran the 10-cent Bonus treatment twice. In the second round, we introduced a comprehension quiz to ensure that our participants understood the payment structure. Within the 10-cent Bonus with bonus comprehension quiz, we randomized access to ML recommendations. Users were randomly allocated to one of seven ML treatments. The control group had no ML recommendations. The other six groups had access to recommendations from a progressively better ML model, as Table 1 shows on each of the last six columns. We picked ML models with varying degrees of performance to approximate the collaborative characteristic function from Figure 1. For details on the model training for the ML recommendations, see Appendix A.8. ## 5 Results (a) Different bonus incentives. (b) Different ML recommendations. Figure 2: Human Performance Across Treatments. Error bars denote 95% confidence intervals based on standard errors clustered at the user level. Solid bars denote the average performance of the submitted solution, striped bars denote the performance if one picked the higher solution between the submitted solution and the provided ML recommendation. A total of 1,408 participants completed the study; we removed 119 participants due to forbidden browser reloads or uses of the browser’s back-button, which left 1,289 for the analyses below. See Appendix A.9 for an overview of participants’ demographics. On average, participants’ compensation implied an hourly wage of £12.17 ($15.22), which is above the UK minimum wage of £11.44. Please see Appendix A.3 for more payment details. We start by discussing the null results of monetary performance incentives (RQ1). Figure 2(a) shows the results of our monetary incentive experiment. On average, user economic performance without any bonus was 89.7% (light blue bar). None of the bonus alternatives are statistically distinguishable from the control group, nor from each other, and their point estimates are all between 88.6% (for the 20-cent bonus) and 90% (for the 10-cent bonus). The null effect of monetary incentives is not due to the fact that users did not understand the bonus structure. To test this hypothesis, we can compare the performance of users in the two 10-cent bonus treatments without algorithmic recommendations (third column in Figure 2(a) and first column in Figure 2(b), both yellow). These two treatments only differ by the fact that the one in Figure 2(b) had a comprehension quiz for the bonus structure. The difference in performance between the two treatments is a mere 0.9%, not statistically different from zero ($p=0.268$, based on standard errors clustered at the user level). RQ2: Because monetary incentives do not change human performance, we test the introduction of ML recommendations with a single bonus structure, the 10-cent bonus. Figure 2(b) presents the results. Focusing on the solid bars, three insights are noteworthy. First, comparing the first two columns (yellow and blue), models with low economic performance seem to lead humans to perform slightly worse than if they were not supported by ML recommendations (89.4% versus 90.9%). This comparison is not statistically significant ($p=0.147$), likely due to low statistical power, but the level difference is not trivial. Despite this, humans do improve performance relative to the algorithmic recommendations (89.4% versus 71.8%, $p=1.8$e-$28$), a result we come back to in Section 5.1. This suggests that people might reduce their effort in solving the problem when they have access to recommendations, but they do not completely eliminate effort (Appendix Figure 18 shows no clear patterns in time spent per problem across treatment conditions). Second, models with better performance lead to increases in human performance, as evidenced by the progressively increasing performance from q1 (72%) to q6 (92%). Third, even if human performance increases with the performance of the ML recommendation, the increments in performance are quantitatively fairly small and sometimes statistically indistinguishable from one another, going from 89.4% when the model’s performance is 72%, to 92.6% when the model’s performance is 92%.333Regression results, controlling for time taken to solve each problem, are presented in Appendix Table 3. ### 5.1 The Results Within Our Theoretical Framework Figure 3: Empirical Collaborative Characteristic Function. Confidence intervals are based on standard errors clustered at the participant level. Figure 3 embeds our empirical results in the theoretical framework presented in Section 3. On the x-axis, we plot the performance of the six ML models deployed in our study. On the y-axis, we plot the performance of the solutions submitted by humans who receive ML recommendations. Each of the points correspond to the six ML treatments of Figure 2(b). We linearly interpolate the estimated points to form an approximation of the collaborative characteristic function $\Delta_{\mathbb{U}}$ (solid blue line). The curve suggests that humans improve on the ML recommendations for ML performance levels between 70% and 92%. The estimated $\delta_{qi}$’s range from 17.5% ($p=1.8$e-$28$) for $q1$, to 0.5% ($p=0.46$) for $q6$. We denote $q6$ a stable point since the human improvement is estimated to be small and statistically indistinguishable from zero. The results imply that, for this portion of the domain, a firm can deploy a model with below-human performance and still converge to a stable point with 92% performance in subsequent deployments. The $\delta_{qi}$ improvements are always positive (or indistinguishable from zero for q6), but they could have been even larger. Indeed, in this specific setting, as participants add items to the knapsack, in principle, they can easily compare the value of their solution to the value of the ML recommendation (both of which appear at the top of the interface, see Appendix Figure 15). If humans had picked the highest between their solution and the ML recommendation, the collaborative characteristic function would have shifted upward to the dashed green line in Figure 3, and the stable point would have achieved a higher performance than 92%. The discrepancy between the solid and dashed lines increases as the ML model improves, suggesting that even in a straightforward comparison, humans do not follow ML recommendations when it is in their best interest to do so (the difference can also be seen by comparing the solid and striped bars in Figure 2(b)). Appendix Figure 19 decomposes the net effect into two parts. On one hand, as the model performance improves, humans are more likely to follow its recommendations. On the other, when they do not follow the ML recommendation, as the model performance improves, it is much more likely that the submitted solution is inferior compared to the recommendation. Under both solid and dashed collaborative characteristic functions, we can imagine possible collaborative learning paths, $\mathbb{L}_{\Delta_{\mathbb{U}}}$. With this shape of $\Delta_{\mathbb{U}}$, the deployment decision is simple: all collaborative learning paths will eventually reach a stable point at above human performance. ## 6 Conclusions We present a theoretical framework for human-ML collaboration in a dynamic setting where human labels can deviate from the indisputable ground truth. We introduce the collaborative characteristic function, which theoretically links the utility of ML models with respect to the indisputable ground truth, to the utility of humans using those same ML models to support their decisions. The collaborative characteristic function allows for multiple collaborative learning paths, depending on the utility of the initially deployed ML model. Each of the collaborative learning paths characterizes a possible ML deployment strategy and its ensuing dynamic learning process. We theoretically show conditions under which this dynamic system reaches a stable point through dynamic utility improvement or deterioration. We then present the empirical results of a large user study, which allows us to estimate points on the collaborative characteristic function of the knapsack problem. For ML models of performance between 72% and 92%, our empirical results suggest that the collaborative characteristic function lies above the 45-degree line. Any collaborative learning path starting at utility between 72% and 92% will thus converge to a stable point with utility around 92%. We explore two factors that can shift the collaborative characteristic function. We find that monetary incentives do not seem to affect human performance. However, we find that wherever applicable, a simple post-processing step that picks the best among available solutions (as is possible for the knapsack problem) can substantially shift the collaborative characteristic function upward, leading to stable equilibria of higher utility. Our work has a number of limitations. On the theoretical side, our collaborative learning paths assume that the firm is able to perfectly replicate human+ML performance in future ML models. Appendix A.7 discusses stability when learning does not exactly replicate previous human+ML performance. On the empirical side, to reduce costs while maintaining statistical power, we only randomized monetary incentives without ML recommendations, and we randomized the quality of ML recommendations while fixing monetary incentives. Studying the interaction of monetary incentives and ML performance is an important extension. The null result of monetary incentives should be interpreted within our context. First, the study participants received payments above minimum wage, and we only tested different levels of linear performance bonuses. It would be valuable to extend our work to evaluate the extent to which alternative base payments or non- linear bonuses may induce different levels of quality and effort by participants and thus collaborative characteristic functions of varying shapes. Our approximation of $\Delta_{\mathbb{U}}$ for the knapsack problem is naturally incomplete since we did not test every possible level of model performance. However, the six points of the curve that we empirically estimate make us fairly comfortable that a linear interpolation is reasonable, at least for model performances between our minimum and maximum. Future work could investigate the properties of $\Delta_{\mathbb{U}}$ that guarantee a unique optimal stable point, both theoretically and empirically. Provided that researchers have access to the indisputable ground truth, further empirical investigations of collaborative characteristic functions could also shed light on the shape of those functions for practically relevant tasks such as medical diagnoses or hiring decisions. Future work should also discuss fairness aspects of this framework, e.g., whether or not fair stable points exist and how a firm can reach them. More generally, we hope this work generates more interest in studying settings where ML deployments lead to changes in the data generating process, which have broad managerial and practical applications. ## References * [1] Vivian Lai, Samuel Carton, Rajat Bhatnagar, Q Vera Liao, Yunfeng Zhang, and Chenhao Tan. Human-ai collaboration via conditional delegation: A case study of content moderation. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1–18, 2022. * [2] Maia Jacobs, Melanie F Pradier, Thomas H McCoy Jr, Roy H Perlis, Finale Doshi-Velez, and Krzysztof Z Gajos. How machine-learning recommendations influence clinician treatment selections: the example of antidepressant selection. Translational psychiatry, 11(1):108, 2021. * [3] Krishnamurthy Dvijotham, Jim Winkens, Melih Barsbey, Sumedh Ghaisas, Robert Stanforth, Nick Pawlowski, Patricia Strachan, Zahra Ahmed, Shekoofeh Azizi, Yoram Bachrach, et al. Enhancing the reliability and accuracy of ai-enabled diagnosis via complementarity-driven deferral to clinicians. Nature Medicine, 29(7):1814–1820, 2023. * [4] Andi Peng, Besmira Nushi, Emre Kiciman, Kori Inkpen, and Ece Kamar. Investigations of performance and bias in human-ai teamwork in hiring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 12089–12097, 2022. * [5] Juan Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, and Moritz Hardt. Performative prediction. In International Conference on Machine Learning, pages 7599–7609. PMLR, 2020. * [6] Corinna Cortes, Giulia DeSalvo, and Mehryar Mohri. Learning with rejection. In Algorithmic Learning Theory: 27th International Conference, ALT 2016, Bari, Italy, October 19-21, 2016, Proceedings 27, pages 67–82. Springer, 2016. * [7] Mohammad-Amin Charusaie, Hussein Mozannar, David Sontag, and Samira Samadi. Sample efficient learning of predictors that complement humans. In International Conference on Machine Learning, pages 2972–3005. PMLR, 2022. * [8] Hussein Mozannar, Hunter Lang, Dennis Wei, Prasanna Sattigeri, Subhro Das, and David Sontag. Who should predict? exact algorithms for learning to defer to humans. In International conference on artificial intelligence and statistics, pages 10520–10545. PMLR, 2023. * [9] Hussein Mozannar, Jimin Lee, Dennis Wei, Prasanna Sattigeri, Subhro Das, and David Sontag. Effective human-ai teams via learned natural language rules and onboarding. Advances in Neural Information Processing Systems, 36, 2024. * [10] Mark Steyvers, Heliodoro Tejeda, Gavin Kerrigan, and Padhraic Smyth. Bayesian modeling of human–ai complementarity. Proceedings of the National Academy of Sciences, 119(11):e2111547119, 2022. * [11] Hussein Mozannar, Gagan Bansal, Adam Fourney, and Eric Horvitz. Reading between the lines: Modeling user behavior and costs in ai-assisted programming. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1–16, 2024. * [12] Arghavan Moradi Dakhel, Vahid Majdinasab, Amin Nikanjam, Foutse Khomh, Michel C Desmarais, and Zhen Ming Jack Jiang. Github copilot ai pair programmer: Asset or liability? Journal of Systems and Software, 203:111734, 2023. * [13] Hussein Mozannar, Gagan Bansal, Adam Fourney, and Eric Horvitz. When to show a suggestion? integrating human feedback in ai-assisted programming. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 10137–10144, 2024. * [14] Fabrizio Dell’Acqua, Edward McFowland, Ethan R Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R Lakhani. Navigating the jagged technological frontier: Field experimental evidence of the effects of ai on knowledge worker productivity and quality. Harvard Business School Technology & Operations Mgt. Unit Working Paper, 24(24-013), 2023. * [15] Helena Vasconcelos, Matthew Jörke, Madeleine Grunde-McLaughlin, Tobias Gerstenberg, Michael S Bernstein, and Ranjay Krishna. Explanations can reduce overreliance on ai systems during decision-making. Proceedings of the ACM on Human-Computer Interaction, 7(CSCW1):1–38, 2023. * [16] Nikhil Agarwal, Alex Moehring, Pranav Rajpurkar, and Tobias Salz. Combining human expertise with artificial intelligence: Experimental evidence from radiology. NBER Working Paper No. 31422, 2023. * [17] Tom Sühr, Sophie Hilgard, and Himabindu Lakkaraju. Does fair ranking improve minority outcomes? understanding the interplay of human and algorithmic biases in online hiring. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, pages 989–999, 2021. * [18] Gagan Bansal, Besmira Nushi, Ece Kamar, Daniel S Weld, Walter S Lasecki, and Eric Horvitz. Updates in human-ai teams: Understanding and addressing the performance/compatibility tradeoff. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 2429–2437, 2019. * [19] Celestine Mendler-Dünner, Juan Perdomo, Tijana Zrnic, and Moritz Hardt. Stochastic optimization for performative prediction. Advances in Neural Information Processing Systems, 33:4929–4939, 2020. * [20] Moritz Hardt, Meena Jagadeesan, and Celestine Mendler-Dünner. Performative power. Advances in Neural Information Processing Systems, 35:22969–22981, 2022. * [21] Celestine Mendler-Dünner, Frances Ding, and Yixin Wang. Anticipating performativity by predicting from predictions. Advances in neural information processing systems, 35:31171–31185, 2022. * [22] Carsten Murawski and Peter Bossaerts. How humans solve complex problems: The case of the knapsack problem. Scientific reports, 6(1):34851, 2016. * [23] David Pisinger. Where are the hard knapsack problems? Computers & Operations Research, 32(9):2271–2284, 2005. ## Appendix A Appendix ### A.1 Data and Code The code for model training, data generation, the web application for user study and our data analysis and plotting can be found in retracted for anonymity; all files are part of the submission as .zip file ### A.2 The Knapsack Problem ###### Definition 6. (0-1 Knapsack Problem) We call $maximize\sum_{i=1}^{n}v_{i}x_{i}\text{ s.t. }\sum_{i=1}^{n}w_{i}x_{i}\leq W$ $\text{ with }x_{i}\in\\{0,1\\},v_{i},w_{i},W\in\mathbb{R_{+}}$ the 0-1 Knapsack Problem. ### A.3 Payment Details We calculated the base payment assuming an average time of 19 minutes to complete the study. The base payment was adjusted upward if the median time to completion was longer than 19 minutes. We adjusted the payment despite the fact that many participants finished our survey but did not enter the completion code directly afterwards. This sometimes increased the median time to completion. ### A.4 Analysis with Optimality ###### Definition 7. (Optimality) For a knapsack instance $X$ with optimal solution $Y^{*}$ and a valid solution $Y$ we call the function $\mathbb{U}_{\text{Opt}}(X,Y)=\begin{cases}1&\text{if, }Y=Y^{*},\\\ 0&\text{else }\end{cases}$ the optimality of $Y$ given $X$. Furthermore, we call $\underset{X}{\mathbb{E}}[\mathbb{U}_{\text{Opt}}(X,Y)]$, the optimal solution rate over all $X$. ###### Observation 1. Economic performance and Optimality are utility functions (1). ###### Proof. We start with the proof that Economic Performance is a utility function. 1) Economic performance is bounded between 0 (for an empty knapsack) and 1, for the optimal value of the knapsack. 2) There exists an $\varepsilon>0$, which is the minimum value of an item for the knapsack problem. The value of that item is the smallest possible distance between two solutions which are not equally good. 3) Because the $Y$ in our case is the sum of the values of the items in the knapsack and $Y*$ is the maximum possible value of the knapsack, any value that is closer to the optimal solution has also higher economic performance because the numerator grows. We chose $\varepsilon$ to be the minimum item value, thus this minimum increase in value between solutions is fulfilled. In summary, Economic Performance satisfies all three criteria of a utiltiy function. We continue with the proof that optimality is a utility function. 1) it is 0 or 1 and thus bounded. 2) If we choose $0<\varepsilon<1$, then $\varepsilon$-sensitivity is satisfied. 3) Is always true for the choice of our $\varepsilon$. Assume for example $\varepsilon=0.5$, then it is that $d(1,1)+0.5<d(0,1)$ and $\mathbb{U}(1)>\mathbb{U}(0)$. This statement is true for all $0<\varepsilon<1$ which is what we specified for $\varepsilon$. ∎ Optimality is the function that indicates whether a solution to a knapsack problem has the optimal value or not. Figure 4 shows the empirical collaborative characteristic function for optimality as utility function. The humans achieve approximately $20\%$ optimalty without ML advice. The effect of human on human-ML performance is significant for all models ($p<0.001$). Interestingly, the effect is large even beyond human performance. Furthermore, for models q1,q2,3 with extremely low utility (average optimality of almost $0\%$), human effects on the overall outcome is large and close to human performance. As in Figure 3, the utility gain of rationally acting humans would have been larger for most models. Our observations suggest that stable points of optimality would lie above human performance without ML adivce. Figure 4: Empirical Collaborative Characteristic Function for the "Optimality" utility function. Confidence intervals are based on standard errors clustered at the participant level. ### A.5 Comments on the Definition of Utility We want to denote that $\varepsilon$-sensitivity implies the following: ###### Observation 2. $\exists\epsilon,\varepsilon>0:|d_{X}(Y,Y^{*})-d_{X}(Y^{\prime},Y^{*})|=\varepsilon\Rightarrow|\mathbb{U}(X,Y)-\mathbb{U}(X,Y^{\prime})|=\epsilon$ This means that there is a minimum utility change that we call $\epsilon$. ### A.6 Proof of Propositon 1&2 #### Proposition 1 (Collaborative Improvement) If $\Delta_{\mathbb{U}}(\mathbb{U}(X,Y_{M}))\geq\mathbb{U}(X,Y_{M})$ for all $M\in\mathcal{M},X\in\mathcal{X}$. Then $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$, is non-decreasing with $t=1,...,T$ and for sufficiently large $T$ it exists a $t^{\prime}\in[1,T]$ such that $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t^{\prime})$ is a stable point. ###### Proof. Let $t\in 1,...,T$ be the number of deployment (epochs) that a firm will make. The firm perfectly learns the data distribution in every epoch, in other words, we assume that $L(Y_{M_{t}},Y_{H_{t-1}}=0,\forall t$. Furthermore, it is $\Delta_{\mathbb{U}}(\mathbb{U}(X,Y_{M}))\geq\mathbb{U}(X,Y_{M})$ for all $M\in\mathcal{M},X\in\mathcal{X}$. We first show that $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$ is non-decreasing with $t$. For that, assume that there exists $t$ for which $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)>\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t+1)$. But $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t+1)=\underset{X\in\mathcal{X}}{\mathbb{E}}(\mathbb{U}(H(X,Y_{M_{t+1}})))\geq^{\delta_{i}\geq 0}\underset{X\in\mathcal{X}}{\mathbb{E}}(\mathbb{U}(Y_{M_{t+1}}))=^{L(Y_{M_{t+1}},Y_{H_{t}})=0}\underset{X\in\mathcal{X}}{\mathbb{E}}(\mathbb{U}(H(X,Y_{M_{t}})))=\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$. It follows that $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$ must be non- decreasing. Now we show that there exists a $t^{\prime}\in[1,T]$ such that $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t^{\prime})$ is a stable point for sufficiently large $T$. For this, consider that $\mathbb{U}$ has a maximum $\mathbb{U}(Y^{*})$ (Property 1 (bounded) of definition 1) and there exists a minimum increment of utility $\epsilon$ (see A.5) in each deployment. If we do not achieve at least $\epsilon$ increment in utility, we have reached a stable point. Thus, we can write the maximum utility as $\mathbb{U}(Y^{*})=\mathbb{U}(Y_{M_{t}})+N\epsilon$. For sufficiently large ($T\geq N+1)$, this implies that we reached maximum utility with $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,T)$, and every deployment beyond that must have equal utility. ∎ #### Proposition 2 (Collaborative Harm) If $\Delta_{\mathbb{U}}(\mathbb{U}(X,Y_{M}))\leq\mathbb{U}(X,Y_{M})$ for all $M\in\mathcal{M},X\in\mathcal{X}$. Then $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t)$, is non-increasing with $t=1,...,T$ and for sufficiently large $T$ it exists a $t^{\prime}\in[1,T]$ such that $\mathbb{L}_{\Delta_{\mathbb{U}}}(s,t^{\prime})$ is a stable point. ###### Proof. Analogous to the proof of Proposition 1. ∎ ### A.7 Perfect vs Imperfect learning In this section we discuss what changes if we loosen the assumption $L(Y_{M_{t}},Y_{H_{t-1}})=0$. We call this assumption the "perfect learner" assumption because the firm perfectly learns the human labels from epoch $t-1$ with a model in epoch $t$. In the following, we consider an imperfect learner such that $L(Y_{M_{t}},Y_{H_{t-1}})=\sigma$. For now, we limit our discussion to the cases of Proposition 1 1 and Proposition 2 2. In the case of collaborative improvement ($\Delta_{\mathbb{U}}(\mathbb{U}(X,Y_{M}))\geq\mathbb{U}(X,Y_{M})$) the human will improve on any model that the firm can deploy. However, if $L(Y_{M_{t}},Y_{H_{t-1}})=\sigma\Rightarrow(\mathbb{U}(Y_{M_{t}})-\mathbb{U}(Y_{M_{t-1}}))<0$ then our error $\sigma$ is larger than what we gain by putting a human in the loop. If the above statement is true for all $M_{t}$, then the imperfection has created collaborative harm and we arrived in the case of Proposition 2. However, this means that this would lead to a stable point. If we consider a second scenario where $L(Y_{M_{t}},Y_{H_{t-1}})=\sigma\Rightarrow(\mathbb{U}(Y_{M_{t}})-\mathbb{U}(Y_{M_{t-1}}))>0$ for all $M_{t}$. Then we are still in the scenario of collaborative improvement which means that we will reach a stable point. We can think about this when looking at Figure 1. An imperfect learner which is still a collaborative improvement setting is effectively like tilting the green dashed lines with a negative slope. We would still reach point 6-d but would require more iterations than with perfect learning. In summary, this is an equally empirical question as it is for a perfect learner. The question is much humans improve the system’s performance and in the case of an imperfect learner how much of that improvement gets "eaten up" by learning errors. ### A.8 Model Training We release the code required for training our models, our model parameters and all predictions for the instances together with the instances that participants saw. Learning to solve the knapsack problem is a research area for itself, however for the small, one-dimensional case of our experiment, it is possible on consumer hardware. We only train models for knapsack instances with 18 items. As input features we concatenate weights $w_{1},...,w_{18}$, values $v_{1},...,v_{18}$, the weight constraint $W$, the sum of the weights and the sum of the values. Thus, our input dimension is 39. Our goal was to train models with a broad spectrum of economic performances, not to solve the knapsack problem perfectly. We added 5 fully connected layers, 4 of them with ReLU activation functions. We use torch.Sigmoid() for our outputs. The output dimension was 18 and the output values in each index can be interpreted as the likelihood that the item belongs to a solution or not. For more details on the architecture, see our code. We want to highlight two important aspects of how we thought about the model training. First, did not want to use any prior knowledge that a firm in our setting could not have either. For example, if we could have known the utility of a knapsack solution (economic performance or optimality) we could have just directly maximized it, or if we could know the optimal solution, we could have just used the distance to the optimal solution as our loss. Instead, we used the binary cross-entropy between the label and prediction as our loss. The label was a 18-dimensional 0-1 vector. If the i-th entry of this output vector is 1, it means that the i-th item is in the knapsack and otherwise not. Thus we simply minimized the differences between chosen items in our training data and those of our model. For us, this was a reasonable analogy for the application context of healthcare in which every "item" is a diagnosis or a symptom (e.g. an ICD10 code). Because our financial budget was limited and we wanted to test multiple models, we trained all models on optimally solved knapsack instances. It would have also created a lot of overhead and space for errors if we would have collected the data of model q1 then trained q2 and rerun the user study. Training them all on generated labels made it possible to run more treatments at once. We still wanted to use ML models instead of solutions produced with dynamic programming, because we wanted to incorporate the distributional character of ML predictions (see Figure 8) and study the reaction to different quantiles of solution quality in greater detail in future work. However, we had to include two pieces of prior knowledge in order to achieve better model performance (especially for q5 and q6). First, we sorted the items by density (value/weight). This is a big advantage in general, but only a small one for our knapsack instances because weights and values are strongly correlated. Second, we normalized weights and values in a pre-processing step. In our setting, both operations could not have been done by the firm (what is a normalized symptom)? However, with those minor modifications we were able to create a larger range of models without massive resources and still just immitate the "human" label without incorporating anything in the loss. In a post-processing step, we sorted the items by sigmoid outputs. We then added items to the knapsack until the weight constraint was reached. From that item selection, we calculated the actual knapsack values. For more details, please visit our github repository To be added after acceptance. ### A.9 Overview statistics Figure 5 shows the overview of the answer to the demographic questions in the end of our study. Most participants held an Undergraduate degree, were between 25 and 44 years old and have not heard about the knapsack problem before completing the study. $50.1\%$ of the participants identified as female $48.6\%$ as male and $0.8\%$ as non-binary or non-gender conforming. $96.8\%$ of the participants have not heard about the knapsack problem before this study. Figure 6 shows the perceived difficulty of the task for the participants, as well as the reported effort the participants put to complete the task. Most participants perceived the task as neutral to hard and put in large to very large effort (self-reportedly). Figure 7 shows how much effort people think they would have spent with or without the help of ML. It seems like participants who had no ML help think they would put less effort in the task. People who had the help of ML reported to put about as much effort as all participants reported to put in right now. Future work should investigate these perceptions in detail. Figure 5: Highest level of education completed, age group, gender and whether participants have heard of the knapsack problem before this study. Figure 6: Perceived difficulty of the task versus the reported level of effort participants reported in our study Figure 7: How much effort participants thought that they would have spent with/without the help of ML ### A.10 Generating Hard Knapsack Problems Knapsack problems where the weights $w_{i}$ and values $v_{i}$ are strongly, yet imperfectly, correlated [23, 22] tend to be hard to solve. We generate knapsack instances with strong correlations ($r\in[0.89,1.00]$, mean $r=.9814$) using Algorithm 1, following the criteria for difficult problems outlined by [23]. In our experiment, users solve knapsack instances with $n=18$ items, $W_{min}=5$, $W_{max}=250$. We constrain the weights, values, and capacity of our instances to integer values, to make them easier to interpret by humans. Algorithm 1 Generate hard knapsack instance number of items $n\geq 0$, knapsack capacity range $W_{min},W_{max}>0$ $W\leftarrow random.uniform.integer(W_{min},W_{max})$ $w\leftarrow random.uniform.integer(1,W,n)$ $\triangleright$ n-dimensional vector of weights $i\leftarrow 1$ while $i\leq n$ do $v_{i}\leftarrow max(1,random.uniform.integer(w_{i}-\lfloor\frac{W}{10}\rfloor,w_{i}+\lfloor\frac{W}{10}\rfloor)$ end while ### A.11 Survey Design Figure 8: Distribution of economic performances of solutions by the six models we deployed in our experiment. Figure 9: Tutorial 1/5 Figure 10: Tutorial 2/5 Figure 11: Tutorial 3/5 (with ML treatment) Figure 12: Tutorial 4/5 (with 10 cents/ppt monetary incentive) Figure 13: Tutorial 5/5 (with comprehension quiz) Figure 14: Feedback to a practice problem Figure 15: Interface for the main task: 1) the knapsack capacity, 2) sum of weights of selected items, 3) sum of values of selected items, 4) remaining time, 5) items with weights and values, 6) machine learning solution (only visible if user receives corresponding treatment). Clicking on gray items adds them to the knapsack if the weight allows it, and clicking on green items removes them from the knapsack. The total weight and value of selected items is shown at the top and automatically updated. Figure 16: Demographic questions after tasks Figure 17: Score screen for performance feedback in the end ### A.12 Futher Statistical insights | $\mathbb{U}_{\text{Econ}}(X,H(X))$ | $\mathbb{U}_{\text{Econ}}(X,H(X))$ ---|---|--- Intercept | $0.7620^{***}$ | $0.6957^{***}$ | $(0.0184)$ | $(0.0408)$ 02-cent bonus | $0.0003$ | $0.1087^{*}$ | $(0.0077)$ | $(0.0484)$ 10-Cent bonus | $0.0011$ | $0.0683$ | $(0.0076)$ | $(0.0499)$ 20-Cent bonus | $-0.0087$ | $0.0787$ | $(0.0079)$ | $(0.0478)$ log(seconds spent) | $0.0322^{***}$ | $0.0481^{***}$ | $(0.0039)$ | $(0.0093)$ 02-cent bonus $\cdot$ log(seconds spent) | — | $-0.0260^{*}$ | | $(0.0112)$ 10-cent bonus $\cdot$ log(seconds spent) | — | $-0.0161$ | | $(0.0115)$ 20-cent bonus $\cdot$ log(seconds spent) | — | $-0.0210$ | | $(0.0113)$ N | 3,960 | 3,960 Adj.R2 | 0.0613 | 0.0661 | $\mathbb{U}_{\text{Econ}}(X,H(X,Y))$ | $\mathbb{U}_{\text{Opt}}(X,H(X,Y))$ ---|---|--- Intercept | $0.8082^{***}$ | $-0.0297$ | $(0.0049)$ | $(0.0245)$ q1 ($72\%$) | $-0.0131$ | $-0.0408$ | $(0.0044)^{**}$ | $(0.0260)$ q2 ($80\%$) | $0.0011$ | $-0.0161$ | $(0.0043)$ | $(0.0283)$ q3 ($84\%$) | $0.0048$ | $-0.0324$ | $(0.0049)$ | $(0.0198)$ q4 ($88\%$) | $0.0151^{***}$ | $0.0333$ | $(0.0033)$ | $(0.0214)$ q5 ($90\%$) | $0.0247^{***}$ | $0.0518^{*}$ | $(0.0043)$ | $(0.0263)$ q6 ($92\%$) | $0.0211^{***}$ | $0.0880^{***}$ | $(0.0033)$ | $(0.0213)$ log(seconds spent) | $0.0240^{***}$ | $0.0533^{***}$ | $(0.0010)$ | $(0.0045)$ N | 8,930 | 8,930 Adj.R2 | 0.0733 | 0.0281 Table 2: Linear regressions with clustered standard errors on participant id. Effect of monetary incentive on $\mathbb{U}_{\text{Econ}}$ of human solutions (left). Effect of ML recommendation on different levels of economic performance $\mathbb{U}_{\text{Econ}}$ (right). Standard errors in parentheses. * $p<0.05$, ** $p<0.01$, *** $p<0.001$. (a) Different bonus incentives. (b) Different ML recommendations. Figure 18: Time Spent Across Treatment Conditions | $\mathbb{U}_{\text{Econ}}(X,H(X))$ ---|--- | (1) | (2) | (3) Intercept | $0.7620^{***}$ | $0.7676^{***}$ | $0.8082^{***}$ | $(0.0184)$ | $(0.0276)$ | $(0.0049)$ 2-cent Bonus | $0.0003$ | | | $(0.0077)$ | | 10-cent Bonus | $0.0011$ | | | $(0.0076)$ | | 20-cent Bonus | $-0.0087$ | | | $(0.0079)$ | | Comprehension Quiz | | $0.0104$ | | | $(0.0076)$ | q1 ($72\%$) | | | $-0.0131$ | | | $(0.0044)^{**}$ q2 ($80\%$) | | | $0.0011$ | | | $(0.0043)$ q3 ($84\%$) | | | $0.0048$ | | | $(0.0049)$ q4 ($88\%$) | | | $0.0151^{***}$ | | | $(0.0033)$ q5 ($90\%$) | | | $0.0247^{***}$ | | | $(0.0043)$ q6 ($92\%$) | | | $0.0211^{***}$ | | | $(0.0033)$ log(seconds spent) | $0.0322^{***}$ | $0.0317^{***}$ | $0.0240^{***}$ | $(0.0039)$ | $0.0064$ | $(0.0010)$ N | 3,960 | 2170 | 8,930 Adj.R2 | 0.0613 | 0.0506 | 0.00733 Included Bonus Treatments | All | 10-cent | 10-cent Included ML Treatments | No ML | No ML | All ML Comprehension Quiz | No | Both | Yes Table 3: Linear regressions of economic performance $\mathbb{U}_{\text{Econ}}$ on dummies for the various treatment conditions. Column 1 includes all treatment conditions without ML recommendations and without comprehension quiz. It tests the difference in performance across different bonus levels. Column 2 includes the two treatment conditions without ML recommendation and with 10-cent bonus. The difference between the two treatment conditions is the presence of a comprehension quiz for the bonus structure. Column 3 includes all treatments with comprehension quiz and 10-cent bonus. It tests the difference in performance across ML recommendations with different performance. Standard errors, in parentheses, are clustered at the participant level. * $p<0.05$, ** $p<0.01$, *** $p<0.001$. (a) Rate of ML advice usage increased with better performance. (b) Participants ignore the ML recommendation with better performance. Figure 19: ML-usage increased with better ML performance. Share of ignored ML solutions did also increase with better performance.
# Stochastic Dynamics of Noisy Average Consensus: Analysis and Optimization Tadashi Wadayama and Ayano Nakai-Kasai Part of this research was presented at IEEE International Symposium on Information Theory 2022 (ISIT2022) [1]. 1Nagoya Institute of Technology, Gokiso, Nagoya, Aichi 466-8555, Japan, <EMAIL_ADDRESS><EMAIL_ADDRESS> ###### Abstract A continuous-time average consensus system is a linear dynamical system defined over a graph, where each node has its own state value that evolves according to a simultaneous linear differential equation. A node is allowed to interact with neighboring nodes. Average consensus is a phenomenon that the all the state values converge to the average of the initial state values. In this paper, we assume that a node can communicate with neighboring nodes through an additive white Gaussian noise channel. We first formulate the noisy average consensus system by using a stochastic differential equation (SDE), which allows us to use the Euler-Maruyama method, a numerical technique for solving SDEs. By studying the stochastic behavior of the residual error of the Euler-Maruyama method, we arrive at the covariance evolution equation. The analysis of the residual error leads to a compact formula for mean squared error (MSE), which shows that the sum of the inverse eigenvalues of the Laplacian matrix is the most dominant factor influencing the MSE. Furthermore, we propose optimization problems aimed at minimizing the MSE at a given target time, and introduce a deep unfolding-based optimization method to solve these problems. The quality of the solution is validated by numerical experiments. ## I Introduction Continuous-time average consensus system is a linear dynamical system defined over a graph [3]. Each node has its own state value, and it evolves according to a simultaneous linear differential equation where a node is only allowed to interact with neighboring nodes. The ordinary differential equation (ODE) at the node $i(1\leq i\leq n)$ governing the evolution of the state value $x_{i}(t)$ of the node $i$ is given by $\displaystyle\frac{dx_{i}(t)}{dt}=-\sum_{j\in{\cal N}_{i}}\mu_{ij}(x_{i}(t)-x_{j}(t)).$ (1) The set ${\cal N}_{i}$ denote the neighboring nodes of node $i$, while the positive scalar $\mu_{ij}$ denotes the edge weight associated with the edge $(i,j)$. The same ODE applies to all other nodes as well. These dynamics gradually decrease the differences between the state values of neighboring nodes, leading to a phenomenon called average consensus that the all the state values converge to the average of the initial state values [2]. The average consensus system has been studied in numerous fields such as multi-agent control [4], distributed algorithm [5], formation control [6]. An excellent survey on average consensus systems can be found in [3]. In this paper, we will examine average consensus systems within the context of communications across noisy channels, such as wireless networks. Specifically, we consider the scenario in which nodes engage in local wireless communication, such as drones flying in the air or sensors dispersed across a designated area. It is assumed that each node can only communicate with neighboring nodes via an additive white Gaussian noise (AWGN) channel. The objective of the communication is to aggregate the information held by all nodes through the application of average consensus systems. As previously stated, the consensus value is the average of the initial state values. In this setting, we must account for the impact of Gaussian noise on the differential equations. The differential equation for a noisy average consensus system takes the form: $\displaystyle\frac{dx_{i}(t)}{dt}=-\sum_{j\in{\cal N}_{i}}\mu_{ij}(x_{i}(t)-x_{j}(t))+\alpha W_{i}(t),$ (2) where $W_{i}(t)$ represents an additive white Gaussian process, and $\alpha$ is a positive constant. The noise $W_{i}(t)$ can be considered as the sum of the noises occurring on the edges adjacent to the node $i$. In a noiseless average consensus system, it is well-established that the second smallest eigenvalue of the Laplacian matrix of the graph determines the convergence speed to the average [5]. The convergence behavior of a noisy system may be quite different from that of the noiseless system due to the presence of edge noise. However, the stochastic dynamics of such a system has not yet been studied. Studies on discrete-time consensus protocols subject to additive noise can be found in [11][12], but to the best of our knowledge, there are no prior studies on continuous-time noisy consensus systems. The main goal of this paper is to study the stochastic dynamics of continuous- time noisy average consensus system. The theoretical understanding of the stochastic behavior of such systems will be valuable for various areas such as multi-agent control and the design of consensus-based distributed algorithms for noisy environments. The primary contributions of this paper are as follows. We first formulate the noisy average consensus systems using stochastic differential equations (SDE) [7][8]. This SDE formulation facilitates mathematically rigorous treatment of noisy average consensus. We use the Euler-Maruyama method [7] for solving the SDE, which is a numerical method for solving SDEs. We derive a closed-form mean squared error (MSE) formula by analyzing the stochastic behavior of the residual errors in the Euler-Maruyama method. We show that the MSE is dominated by the sum of the inverse eigenvalues of the Laplacian matrix. However, minimizing the MSE at a specific target time is a non-trivial task because the objective function involves the sum of the inverse eigenvalues. To solve this optimization problem, we will propose a deep unfolding-based optimization method. The outline of the paper is as follows. In Section 2, we introduce the mathematical notation used throughout the paper, and then provide the definition and fundamental properties of average consensus systems. In Section 3, we define a noisy average consensus system as a SDE. In Section 4, we present an analysis of the stochastic behavior of the consensus error and derive a concise MSE formula. In Section 5, we propose a deep unfolding-based optimization method for minimizing the MSE at a specified target time. Finally, in Section 6, we conclude the discussion. ## II Preliminaries ### II-A Notation The following notation will be used throughout this paper. The symbols $\mathbb{R}$ and $\mathbb{R}_{+}$ represent the set of real numbers and the set of positive real numbers, respectively. The one dimensional Gaussian distribution with mean $\mu$ and variance $\sigma^{2}$ is denoted by ${\cal N}(\mu,\sigma^{2})$. The multivariate Gaussian distribution with mean vector $\bm{\mu}$ and covariance matrix $\bm{\Sigma}$ is represented by ${\cal N}(\bm{\mu},\bm{\Sigma})$. The expectation operator is denoted by ${\sf E}[\cdot]$. The notation $\mbox{diag}(\bm{x})$ is the diagonal matrix whose diagonal elements are given by $\bm{x}\in\mathbb{R}^{n}$. The matrix exponential $\exp(\bm{X})(\bm{X}\in\mathbb{R}^{n\times n})$ is defined by $\displaystyle\exp(\bm{X})\equiv\sum_{k=0}^{\infty}\frac{1}{k!}\bm{X}^{k}.$ (3) The Frobenius norm of $\bm{X}\in\mathbb{R}^{n\times n}$ is denoted by $\|\bm{X}\|_{F}$. The notation $[n]$ denotes the set of consecutive integers from $1$ to $n$. ### II-B Average Consensus Let $G\equiv(V,E)$ be a connected undirected graph where $V=[n]$. Suppose that a node $i\in V$ can be regarded as an agent communicating over the graph $G$. Namely, a node $i$ and a node $j$ can communicate with each other if $(i,j)\in E$. We will not distinguish $(i,j)$ and $(j,i)$ because the graph $G$ is undirected. Each node $i$ has a state value $x_{i}(t)\in\mathbb{R}$ where $t\in\mathbb{R}$ represents continuous-time variable. The neighborhood of a node $i\in V$ is represented by $\displaystyle{\cal N}_{i}\equiv\\{j\in V:(j,i)\in E,i\neq j\\}.$ (4) Note that the node $i$ is excluded from ${\cal N}_{i}$. For any time $t$, a node $i\in V$ can access the self-state $x_{i}(t)$ and the state values of its neighborhood, i.e., $x_{j}(t),j\in{\cal N}_{i}$ but cannot access to the other state values. In this section, we briefly review the basic properties of the average consensus processes [3]. We now assume that the set of state values $\bm{x}(t)\equiv(x_{1}(t),x_{2}(t),\ldots,x_{n}(t))^{T}$ are evolved according to the simultaneous differential equations $\displaystyle\frac{dx_{i}(t)}{dt}=-\sum_{j\in{\cal N}_{i}}\mu_{ij}(x_{i}(t)-x_{j}(t)),\quad i\in[n],$ (5) where the initial condition is $\displaystyle\bm{x}(0)=\bm{c}\equiv(c_{1},c_{2},\ldots,c_{n})^{T}\in\mathbb{R}^{n}.$ (6) The edge weight $\mu_{ij}$ follows the symmetric condition $\displaystyle\mu_{ij}=\mu_{ji},\quad i\in[n],j\in[n].$ (7) Let $\bm{\Delta}\equiv(\Delta_{1},\Delta_{2},\ldots,\Delta_{n})^{T}\in\mathbb{R}_{+}^{n}$ be a degree sequence where $\Delta_{i}$ is defined by $\displaystyle\Delta_{i}\equiv\sum_{j\in{\cal N}_{i}}\mu_{ij},\quad i\in[n].$ (8) The continuous-time dynamical system (5) is called an average consensus system because a state value converges to the average of the initial state values at the limit of $t\rightarrow\infty$, i.e, $\displaystyle\lim_{t\rightarrow\infty}\bm{x}(t)=\frac{1}{n}\left(\sum_{i=1}^{n}c_{i}\right)\bm{1}{=\gamma\bm{1}},$ (9) where the vector $\bm{1}$ represents $(1,1,\ldots,1)^{T}$ and $\gamma$ is defined by $\displaystyle\gamma\equiv\frac{1}{n}\sum_{i=1}^{n}c_{i}.$ (10) We define the Laplacian matrix $\bm{L}\equiv\\{L_{ij}\\}\in\mathbb{R}^{n\times n}$ of this consensus system as follows: $\displaystyle L_{ij}$ $\displaystyle=\Delta_{i},\quad i=j,\ i\in[n],$ (11) $\displaystyle L_{ij}$ $\displaystyle=-\mu_{ij},\quad i\neq j\mbox{ and }(i,j)\in E,$ (12) $\displaystyle L_{ij}$ $\displaystyle=0,\quad i\neq j\mbox{ and }(i,j)\notin E.$ (13) From this definition, a Laplacian matrix satisfies $\displaystyle\bm{L}\bm{1}$ $\displaystyle=\bm{0},$ (14) $\displaystyle\mbox{diag}(\bm{L})$ $\displaystyle=\bm{\Delta},$ (15) $\displaystyle\bm{L}$ $\displaystyle=\bm{L}^{T}.$ (16) Note that the eigenvalues of the Laplacian matrix $\bm{L}$ are nonnegative real because $\bm{L}$ is a positive semi-definite symmetric matrix. Let $\lambda_{1}=0<\lambda_{2}\leq\ldots\leq\lambda_{n}$ be the eigenvalues of $\bm{L}$ and $\bm{\xi}_{1},\bm{\xi}_{2},\ldots,\bm{\xi}_{n}$ be the corresponding orthonormal eigenvectors. The first eigenvector $\bm{\xi}_{1}\equiv(1/\sqrt{n})\bm{1}$ is corresponding to the eigenvalue $\lambda_{1}=0$, which results in $\bm{L}\bm{\xi}_{1}=0$. By using the notion of the Laplacian matrix, the dynamical system (5) can be compactly rewritten as $\displaystyle\frac{d\bm{x}(t)}{dt}=-\bm{L}\bm{x}(t),$ (17) where the initial condition is $\bm{x}(0)=\bm{c}$. The dynamical behaviors of the average consensus system (17) are thus characterized by the Laplacian matrix $\bm{L}$. Since the ODE (17) is a linear ODE, it can be easily solved. The solution of the ODE (17) is given by $\displaystyle\bm{x}(t)=\exp(-\bm{L}t)\bm{x}(0),\quad t\geq 0.$ (18) Let $\bm{U}\equiv(\bm{\xi}_{1},\bm{\xi}_{2},\ldots,\bm{\xi}_{n})\in\mathbb{R}^{n\times n}$ where $\bm{U}$ is an orthogonal matrix. The Laplacian matrix $\bm{L}$ can be diagonalized by using $\bm{U}$, i.e., $\displaystyle\bm{L}=\bm{U}\mbox{diag}(\lambda_{1},\ldots,\lambda_{n})\bm{U}^{T}.$ (19) On the basis of the diagonalization, we have the spectral expansion of the matrix exponential: $\displaystyle\exp(-\bm{L}t)$ $\displaystyle=\exp(-\bm{U}\mbox{diag}(\lambda_{1},\ldots,\lambda_{n})\bm{U}^{T}t)$ $\displaystyle=\bm{U}\exp(-\mbox{diag}(\lambda_{1},\ldots,\lambda_{n})t)\bm{U}^{T}$ $\displaystyle=\sum_{i=1}^{n}\exp(-\lambda_{i}t)\bm{\xi}_{i}\bm{\xi}_{i}^{T}.$ (20) Substituting this to $\bm{x}(t)=\exp(-\bm{L}t)\bm{x}(0)$, we immediately have $\displaystyle\bm{x}(t)=\frac{1}{n}\bm{1}(\bm{1}^{T})\bm{c}+\sum_{i=2}^{n}\exp(-\lambda_{i}t)\bm{\xi}_{i}\bm{\xi}_{i}^{T}\bm{c}.$ (21) The second term of the right-hand side converges to zero since $\lambda_{k}>0$ for $k=2,3,\ldots,n$. This explains why average consensus happens, i.e., the convergence to the average of the initial state values (9). The second smallest eigenvalue $\lambda_{2}$, called algebraic connectivity [10], determines the convergence speed because $\exp(-\lambda_{2}t)\bm{\xi}_{2}\bm{\xi}_{2}^{\sf T}$ shows the slowest convergence in the second term. ## III Noisy average consensus system ### III-A SDE formulation The dynamical model (2) containing a white Gaussian noise process is mathematically challenging to handle. We will use a common approach of approximating the white Gaussian process by using the standard Wiener process. Instead of model (2), we will focus on the following stochastic differential equation (SDE) [8] $d\bm{x}(t)=-\bm{L}\bm{x}(t)dt+\alpha d\bm{b}(t)$ (22) to study the noisy average consensus system. The parameter $\alpha$ is a positive real number, and it represents the intensity of the noises. The stochastic term $\bm{b}(t)$ represents the $n$-dimensional standard Wiener process. The elements of $\bm{b}(t)=(b_{1}(t),b_{2}(t),\ldots,b_{n}(t))^{T}$ are independent one dimensional-standard Wiener processes. For the Wiener process $b(t)$, we have $b(0)=0$, $E[b(t)]=0$, and it satisfies $\displaystyle b(t)-b(s)\sim{\cal N}(0,t-s),\ 0\leq s\leq t.$ (23) ### III-B Approaches for studying stochastic dynamics Our primary objective in the following analysis is to investigate the stochastic dynamics of the noisy average consensus system, focusing on deriving the mean and covariance of the solution $\bm{x}(t)$ for the SDE (22). There are two approaches to analyze the system. The first approach relies on the established theory of Ito calculus [8], which is used to handle stochastic integrals directly (see Fig. 1). Ito calculus can be applied to derive the first and second moments of the solution of (22). Alternatively, the second approach employs the Euler-Maruyama (EM) method [7] and utilizes the weak convergence property [7] of the EM method. We will adopt the latter approach in our analysis, as it does not require knowledge of advanced stochastic calculus if we accept the weak convergence property. Additionally, this approach can be naturally extended to the analysis on the discrete-time noisy average consensus system. Furthermore, the EM method plays a key role in the optimization method to be presented in Section V. Our analysis motivates the use EM method for optimizing the covariance. Figure 1: Two approaches for deriving the mean and covariance of $\bm{x}(t)$. This paper follows the lower path using the EM method. ### III-C Euler-Maruyama method We use the Euler-Maruyama method corresponding to this SDE so as to study the stochastic behavior of the solution of the SDE (22) defined above. The EM method is well-known numerical method for solving SDEs [7]. Assume that we need numerical solutions of a SDE in the time interval $0\leq t\leq T$. We divide this interval into $N$ bins and let $t_{k}\equiv k\eta,\ k=0,1,\ldots,N$ where the interval $\eta$ is given by $\eta\equiv{T}/{N}.$ Let us define a discretized sample $\bm{x}^{(k)}$ be $\bm{x}^{(k)}\equiv\bm{x}(t_{k}).$ It should be noted that, the choice of the width $\eta$ is crucial in order to ensure the stability and the accuracy of the EM method. A small width leads to a more accurate solution, but requires more computational time. A large width may be computationally efficient but may lead to instability in the solution. The recursive equation of the EM method corresponding to SDE (22) is given by $\displaystyle\bm{x}^{(k+1)}=\bm{x}^{(k)}-\eta\bm{L}\bm{x}^{(k)}+\alpha\bm{w}^{(k)},\ k=0,1,2,\ldots,N,$ (24) where each element of $\bm{w}^{(k)}\equiv(w_{1}^{(k)},w_{2}^{(k)},\ldots,w_{n}^{(k)})^{T}$ follows $w_{i}^{(k)}\sim{\cal N}(0,\eta).$ In the following discussion, we will use the equivalent expression [7]: $\displaystyle\bm{x}^{(k+1)}=\bm{x}^{(k)}-\eta\bm{L}\bm{x}^{(k)}+\alpha\sqrt{\eta}\bm{z}^{(k)},\ k=0,1,2,\ldots,N,$ (25) where $\bm{z}^{(k)}$ is a random vector following the multivariate Gaussian distribution ${\cal N}(\bm{0},\bm{I})$. The initial vector $\bm{x}^{(0)}$ is set to be $\bm{c}$. This recursive equation will be referred to as the Euler- Maruyama recursive equation. Figure 2 presents a solution evaluated with the EM method. The cycle graph with 10 nodes with the degree sequence $\bm{d}=(2,2,\ldots,2)$ is assumed. The initial value is randomly initialized as $\bm{x}(0)\sim{\cal N}(0,\bm{I})$. We can confirm that the state values are certainly converging to the average value $\gamma$ in the case of noiseless case (left). On the other hand, the state vector fluctuates around the average in the noisy case (right). Figure 2: Trajectories of $\bm{x}(t_{k})=(x_{1}(t_{k}),\ldots,x_{n}(t_{k}))$ estimated by using the EM method. Cycle graph with 10 nodes were used. The range $[0,10.0]$ are discretized with $N=100$ points. The consensus average value is $\gamma=-0.3267$. Left panel: noiseless case $(\alpha=0.0)$, Right panel: noisy case $(\alpha=0.1)$. ## IV Analysis for Noisy average consensus ### IV-A Recursive equation for residual error In the following, we will analyze the stochastic behavior of the residual error. This will be the basis for the MSE formula to be presented. Recall that the initial state vector is $\bm{c}=(c_{1},c_{2},\ldots,c_{n})^{T}$ and that the average of the initial values is denoted by $\gamma$. Since the set of eigenvectors $\\{\bm{\xi}_{1},\ldots,\bm{\xi}_{n}\\}$ of $\bm{L}$ is an orthonormal base, we can expand the initial state vector $\bm{c}$ as $\displaystyle\bm{c}=\zeta_{1}\bm{\xi}_{1}+\zeta_{2}\bm{\xi}_{2}+\cdots+\zeta_{n}\bm{\xi}_{n},$ (26) where the coefficient is obtained by $\zeta_{i}=\bm{c}^{T}\bm{\xi}_{i}(i\in[n])$. Note that $\zeta_{1}\bm{\xi}_{1}=\gamma\bm{1}$ holds. At the initial index $k=0$, the Euler-Maruyama recursive equation becomes $\displaystyle\bm{x}^{(1)}=\bm{x}^{(0)}-\eta\bm{L}\bm{x}^{(0)}+\alpha\sqrt{\eta}\bm{z}^{(0)}.$ (27) Substituting (26) into the above equation, we have $\displaystyle\bm{x}^{(1)}$ $\displaystyle=\bm{x}^{(0)}-\eta\bm{L}(\zeta_{1}\bm{\xi}_{1}+\zeta_{2}\bm{\xi}_{2}+\cdots+\zeta_{n}\bm{\xi}_{n})+\alpha\sqrt{\eta}\bm{z}^{(0)}$ $\displaystyle=\bm{x}^{(0)}-\eta\zeta_{1}\bm{L}\bm{\xi}_{1}-\eta L(\bm{x}^{(0)}-\zeta_{1}\bm{\xi}_{1})+\alpha\sqrt{\eta}\bm{z}^{(0)}$ $\displaystyle={\bm{x}^{(0)}-\eta\bm{L}(\bm{x}^{(0)}-\gamma\bm{1})+\alpha\sqrt{\eta}\bm{z}^{(0)}},$ (28) where the equations $L\bm{\xi}_{1}=\bm{0}$ and $\zeta_{1}\bm{\xi}_{1}=\gamma\bm{1}$ are used in the last equality. Subtracting $\gamma\bm{1}$ from the both sides, we get $\displaystyle\bm{x}^{(1)}-\gamma\bm{1}=(\bm{I}-\eta\bm{L})(\bm{x}^{(0)}-\gamma\bm{1})+\alpha\sqrt{\eta}\bm{z}^{(0)}.$ (29) For the index $k\geq 1$, the Euler-Maruyama recursive equation can be written as $\displaystyle\bm{x}^{(k+1)}=(\bm{I}-\eta\bm{L})\bm{x}^{(k)}+\alpha\sqrt{\eta}\bm{z}^{(k)}.$ (30) Subtracting $\gamma\bm{1}$ from the both sides, we have $\displaystyle\bm{x}^{(k+1)}-\gamma\bm{1}=(\bm{I}-\eta\bm{L})\bm{x}^{(k)}-\gamma\bm{1}+\alpha\sqrt{\eta}\bm{z}^{(k)}.$ (31) By using the relation $(\bm{I}-\eta\bm{L})\gamma\bm{1}=\gamma\bm{1},$ we can rewrite the above equation as $\displaystyle\bm{x}^{(k+1)}-\gamma\bm{1}$ $\displaystyle=(\bm{I}-\eta\bm{L})\bm{x}^{(k)}-(\bm{I}-\eta\bm{L})\gamma\bm{1}+\alpha\sqrt{\eta}\bm{z}^{(k)}$ $\displaystyle=(\bm{I}-\eta\bm{L})(\bm{x}^{(k)}-\gamma\bm{1})+\alpha\sqrt{\eta}\bm{z}^{(k)}.$ (32) It can be confirmed the above recursion (32) is consistent with the initial equation (29). We here summarize the above argument as the following lemma. ###### Lemma 1 Let $\bm{e}^{(k)}\equiv\bm{x}^{(k)}-\gamma\bm{1}$ be the residual error at index $k$. The evolution of the residual error of the EM method is described by $\displaystyle\bm{e}^{(k+1)}=(\bm{I}-\eta\bm{L})\bm{e}^{(k)}+\alpha\sqrt{\eta}\bm{z}^{(k)}$ (33) for $k\geq 0$. The residual error $\bm{e}^{(k)}$ denotes the error between the average vector $\gamma\bm{1}$ and the state vector $\bm{x}^{(k)}$ at time index $k$. By analyzing the statistical behavior of $\bm{e}^{(k)}$, we can gain insight into the stochastic properties of the dynamics of the noisy consensus system. ### IV-B Asymptotic mean of residual error Let a vector $\bm{x}\sim{\cal N}(\bm{\mu},\bm{\Sigma})$. Recall that the vector obtained by a linear map $\bm{y}=\bm{A}\bm{x}$ also follows the Gaussian distribution, i.e., $\displaystyle\bm{y}\sim{\cal N}(\bm{A}\bm{\mu},\bm{A}\bm{\Sigma}\bm{A}^{T}),$ (34) where $\bm{A}\in\mathbb{R}^{n\times n}$. If two Gaussian vectors $\bm{a}\sim{\cal N}(\bm{\mu}_{a},\bm{\Sigma}_{a})$ and $\bm{b}\sim{\cal N}(\bm{\mu}_{b},\bm{\Sigma}_{b})$ are independent, the sum $\bm{z}=\bm{a}+\bm{b}$ becomes also Gaussian, i.e, $\displaystyle\bm{z}\sim{\cal N}(\bm{\mu}_{a}+\bm{\mu}_{b},\bm{\Sigma}_{a}+\bm{\Sigma}_{b}).$ (35) In the recursive equation (33), it is evident that $\bm{e}^{(1)}$ follows a multivariate Gaussian distribution because $\displaystyle\bm{e}^{(1)}=(\bm{I}-\eta\bm{L})(\bm{c}-\gamma\bm{1})+\alpha\sqrt{\eta}\bm{z}^{(0)}$ (36) is the sum of a constant vector and a Gaussian random vector. From the above properties of Gaussian random vectors, the residual error vector $\bm{e}^{(k)}$ follows the multivariate Gaussian distribution ${\cal N}(\bm{\mu}^{(k)},\bm{\Sigma}^{(k)})$ where the mean vector $\bm{\mu}^{(k)}$ and the covariance matrix $\bm{\Sigma}^{(k)}$ are recursively determined by $\displaystyle\bm{\mu}^{(k+1)}$ $\displaystyle=(\bm{I}-\eta\bm{L})\bm{\mu}^{(k)},$ (37) $\displaystyle\bm{\Sigma}^{(k+1)}$ $\displaystyle=(\bm{I}-\eta\bm{L})\bm{\Sigma}^{(k)}(\bm{I}-\eta\bm{L})^{T}+\alpha^{2}\eta\bm{I}$ (38) for $k\geq 0$ where the initial values are formally given by $\displaystyle\bm{\mu}^{(0)}$ $\displaystyle=\bm{c}-\gamma\bm{1},$ (39) $\displaystyle\bm{\Sigma}^{(0)}$ $\displaystyle=\bm{O}.$ (40) Solving the recursive equation, we can get the asymptotic mean formula as follows. ###### Lemma 2 Suppose that $T>0$ is given. The asymptotic mean at $N\rightarrow\infty$ is given by $\displaystyle\lim_{N\rightarrow\infty}\bm{\mu}^{(N)}=\exp(-\bm{L}T)(\bm{c}-\gamma\bm{1}).$ (41) (Proof) The mean recursion is given as $\bm{\mu}^{(k)}=(\bm{I}-\eta\bm{L})^{k}(\bm{c}-\gamma\bm{1})$ for $k\geq 1$. Recall that the eigenvalue decomposition of $\bm{L}$ is given by $\bm{L}=\bm{U}\mbox{diag}(\lambda_{1},\ldots,\lambda_{n})\bm{U}^{T}$. From $\displaystyle\bm{I}-\eta\bm{L}=\bm{U}(\bm{I}-\eta\mbox{diag}(\lambda_{1},\ldots,\lambda_{n}))\bm{U}^{T},$ (42) we have $\displaystyle(\bm{I}-\eta\bm{L})^{k}=\bm{U}\mbox{diag}((1-\eta\lambda_{1})^{k},\ldots,(1-\eta\lambda_{n})^{k})\bm{U}^{T}.$ (43) This implies, from the definition of exponential function, $\displaystyle\lim_{N\rightarrow\infty}\left(\bm{I}-\frac{T}{N}\bm{L}\right)^{N}=\exp(-\bm{L}T),$ (44) where $\eta=T/N$. It is easy to confirm that the claim of this lemma is consistent with the continuous solution of noiseless case (18). Namely, at the limit of $\alpha\rightarrow 0$, the state evolution of the noisy system converges to that of the noiseless system. ### IV-C Asymptotic covariance of residual error We here discuss the asymptotic behavior of the covariance matrix $\bm{\Sigma}^{(N)}$ at the limit of $N\rightarrow\infty$. ###### Lemma 3 Suppose that $T>0$ is given. The asymptotic covariance matrix at $N\rightarrow\infty$ is given by $\displaystyle\lim_{N\rightarrow\infty}\bm{\Sigma}^{(N)}=\bm{U}\mbox{diag}\left(\alpha^{2}T,\theta_{2},\theta_{3},\ldots,\theta_{n}\right)\bm{U}^{T},$ (45) where $\theta_{i}$ is defined by $\displaystyle\theta_{i}\equiv\frac{\alpha^{2}}{2\lambda_{i}}\left(1-e^{-2\lambda_{i}T}\right).$ (46) (Proof) Recall that $\displaystyle\bm{I}-\eta\bm{L}=\bm{U}\mbox{diag}(1,1-\eta\lambda_{2}\ldots,1-\eta\lambda_{n})\bm{U}^{T}.$ (47) Let $\bm{\Sigma}^{(k)}=\bm{U}\mbox{diag}(s_{1}^{(k)},\ldots,s_{n}^{(k)})\bm{U}^{T}$. A spectral representation of the covariance evolution (38) is thus given by $\displaystyle\mbox{diag}(s_{1}^{(k+1)},\ldots,s_{n}^{(k+1)})$ $\displaystyle=\mbox{diag}(s_{1}^{(k)},s_{2}^{(k)}(1-\eta\lambda_{2})^{2}\ldots,s_{n}^{(k)}(1-\eta\lambda_{n})^{2})+\alpha^{2}\eta\bm{I},$ (48) where $s_{i}^{(0)}=0$. The first component follows a recursion $s_{1}^{(k+1)}=s_{1}^{(k)}+\alpha^{2}\eta$ and thus we have $s_{1}^{(N)}=\alpha^{2}\eta N=\alpha^{2}T.$ Another component follows $\displaystyle s_{i}^{(k+1)}=s_{i}^{(k)}(1-\eta\lambda_{i})^{2}+\alpha^{2}\eta.$ (49) Let us consider the characteristic equation of (49) which is given by $\displaystyle s=s(1-\eta\lambda_{i})^{2}+\alpha^{2}\eta.$ (50) The solution of the equation is given by $\displaystyle s=\frac{\alpha^{2}\eta}{1-(1-\eta\lambda_{i})^{2}}.$ (51) The above recursive equation (49) thus can be transformed as $\displaystyle s_{i}^{(k+1)}-s=(s_{i}^{(k)}-s)(1-\eta\lambda_{i})^{2}.$ (52) From the above equation, $s_{i}^{(N)}$ can be solved as $\displaystyle s_{i}^{(N)}=s+(s_{i}^{(0)}-s)(1-\eta\lambda_{i})^{2N}.$ (53) Taking the limit $N\rightarrow\infty$, we have $\displaystyle\lim_{N\rightarrow\infty}s_{i}^{(N)}=\frac{\alpha^{2}}{2\lambda_{i}}\left(1-e^{-2\lambda_{i}T}\right).$ (54) We thus have the claim of this lemma. ### IV-D Weak convergence of Euler-Maruyama method As previously noted, the asymptotic mean (41) is consistent with the continuous solution. The weak convergence property of the EM method [7] allows us to obtain the moments of the error at time $t$. We will briefly explain the weak convergence property. Suppose a SDE with the form: $\displaystyle d\bm{x}(t)=\phi(\bm{x}(t))dt+\psi(\bm{x}(t))d\bm{b}(t).$ (55) If $\phi$ and $\psi$ are bounded and Lipschitz continuous, then the finite order moment estimated by the EM method converges to the exact moment of the solution $\bm{x}(t)$ at the limit $N\rightarrow\infty$ [7]. This property is called the weak convergence property. In our case, the SDE (22) has bounded and Lipschitz continuous coefficient functions, i.e, $\phi(\bm{x})=-\bm{L}\bm{x}$ and $\psi(\bm{x})=\alpha$. Hence, we can employ the weak convergence property in our analysis. Suppose $\bm{x}(t)$ is a solution of SDE (22) with the initial condition $\bm{x}(0)=\bm{c}$. Let $\bm{\mu}(t)$ be the mean vector of the residual error $\bm{e}(t)=\bm{x}(t)-\gamma\bm{1}$ and $\bm{\Sigma}(t)$ is the covariance matrix of the residual error $\bm{e}(t)$. ###### Theorem 1 For a positive real number $t>0$, the mean and the covariance matrix of the residual error $\bm{e}(t)$ are given by $\displaystyle\bm{\mu}(t)$ $\displaystyle=\exp(-\bm{L}t)(\bm{c}-\gamma\bm{1})$ (56) $\displaystyle\bm{\Sigma}(t)$ $\displaystyle=\bm{U}\mbox{diag}\left(\alpha^{2}t,\theta_{2},\theta_{3},\ldots,\theta_{n}\right)\bm{U}^{T}.$ (57) (Proof) Due to the weak convergence property of the EM method, the first and second moments of the error are converged to the asymptotic mean and covariance of the EM method [7], i.e., $\displaystyle\bm{\mu}(T)$ $\displaystyle=\lim_{N\rightarrow\infty}\bm{\mu}^{(N)}$ (58) $\displaystyle\bm{\Sigma}(T)$ $\displaystyle=\lim_{N\rightarrow\infty}\bm{\Sigma}^{(N)},$ (59) where $N$ and $T$ are related by $T=\eta N$. Applying Lemmas 2 and 3 and replacing the variable $T$ by $t$ provide the claim of the theorem. ### IV-E Mean squared error In the following, we assume that the initial state vector $\bm{c}$ follows Gaussian distribution ${\cal N}(\bm{0},\bm{I})$. In this setting, $\bm{\mu}(t)$ also follows multivariate Gaussian distribution with the mean vector $\bm{0}$ and the covariance matrix $\bm{Q}(t)\bm{Q}(t)^{T}$ where $\displaystyle\bm{Q}(t)\equiv\exp(-\bm{L}t)\left(\bm{I}-\frac{1}{n}\bm{1}(\bm{1}^{T})\right)$ (60) because $\bm{\mu}(t)$ can be rewritten as $\displaystyle\bm{\mu}(t)$ $\displaystyle=\exp(-\bm{L}t)(\bm{c}-\gamma\bm{1})=\exp(-\bm{L}t)\left(\bm{I}-\frac{1}{n}\bm{1}(\bm{1}^{T})\right)\bm{c}.$ (61) By using the result of Theorem 1, we immediately have the following corollary indicating the MSE formula. ###### Corollary 1 The mean squared error (MSE) $\displaystyle{\sf MSE}(t)\equiv{\sf E}[\|\bm{x}(t)-\gamma\bm{1}\|_{2}^{2}]$ (62) is given by $\displaystyle{\sf MSE}(t)$ $\displaystyle=\alpha^{2}t+\frac{\alpha^{2}}{2}\sum_{i=2}^{n}\frac{1-e^{-2\lambda_{i}t}}{\lambda_{i}}+\mbox{tr}(\bm{Q}(t)\bm{Q}(t)^{T}).$ (Proof) We can rewrite $\bm{x}(t)$ as: $\displaystyle\bm{x}(t)=\gamma\bm{1}+\bm{Q}(t)\bm{c}+\bm{w},$ (63) where $\bm{w}\sim{\cal N}(\bm{0},\bm{\Sigma}(t))$, and $\bm{w}$ and $\bm{c}$ are independent. We thus have $\displaystyle{\sf MSE}(t)$ $\displaystyle=\mbox{tr}(\bm{\Sigma}(t))+\mbox{tr}(\bm{Q}(t)\bm{Q}(t)^{T})$ $\displaystyle=\alpha^{2}t+\frac{\alpha^{2}}{2}\sum_{i=2}^{n}\frac{1-e^{-2\lambda_{i}t}}{\lambda_{i}}+\mbox{tr}(\bm{Q}(t)\bm{Q}(t)^{T})$ (64) due to Theorem 1. Since the value of the term $\mbox{tr}(\bm{Q}(t)\bm{Q}(t)^{T})$ is exponentially decreasing with $t$, $\mbox{tr}(\bm{\Sigma}(t))$ is dominant in ${\sf MSE}(t)$ for sufficiently large $t$. For sufficiently large $t$, the MSE is well approximated by the asymptotic MSE (AMSE) as $\displaystyle{\sf MSE}(t)\simeq{\sf AMSE}(t)\equiv\alpha^{2}t+\frac{\alpha^{2}}{2}\sum_{i=2}^{n}\frac{1}{\lambda_{i}}$ (65) because $\mbox{tr}(\bm{Q}(t)\bm{Q}(t)^{T})$ is negligible, and $1-e^{-2\lambda_{i}t}$ can be well approximated to $1$. We can observe that the sum of inverse eigenvalue $\sum_{i=2}^{n}({1}/{\lambda_{i}})$ of the Laplacian matrix determines the intercept of the ${\sf AMSE}(t)$. In other words, the graph topology influences the stochastic error behavior through the sum of inverse eigenvalues of the Laplacian matrix. Figure 3 presents a comparison of ${\sf MSE}(t)$ evaluated by the EM method (25) and the formula in (64). In this experiment, the cycle graph with 10 nodes is used. The values of ${\sf AMSE}(t)$ are also included in Fig. 3. We can see that the theoretical values of ${\sf MSE}(t)$ and estimated values by the EM method are quite close. Figure 3: Comparison of MSE: The label Euler-Maruyama represents ${\sf MSE}(t)$ estimated by using samples generated by the EM method. Theoretical ${\sf MSE}(t)$ represents the values evaluated by (64).Theoretical ${\sf AMSE}(t)$ represents the values of ${\sf AMSE}(t)$. Cycle graph with 10 nodes with $\bm{d}=(2,2,\ldots,2)$ are used. The parameter setting is as follows: $N=250$, $T=10$, $\alpha=0.2$. $5000$ samples are generated by the EM method for estimating ${\sf MSE}(t)$. ## V Minimization of mean squared error ### V-A Optimization Problems A and B In the previous section, we demonstrated that the MSE can be expressed in closed-form. It is natural to optimize the edge weights $\\{\mu_{ij}\\}$ in order to decrease the value of the MSE. The optimization of the edge weights is equivalent to the optimization of the Laplacian matrix $\bm{L}$. There exist several related works that aim to achieve a similar goal for noise-free systems. For example, Xiao and Boyd [5] proposed a method to minimize the second eigenvalue to achieve the fastest convergence to the average. They formulated the optimization problem as a convex optimization problem, which can be solved efficiently. Kishida et al. [13] presented a deep unfolding- based method for optimizing time-dependent edge weights, yet these methods are not applicable to systems with noise. Optimizing the MSE may be a non-trivial task as it involves the sum of the inverse eigenvalues of the Laplacian matrix. In this subsection, we will present two optimization problems of edge weights. #### V-A1 Optimization problem A Assume that a degree sequence $\bm{d}\in\mathbb{R}^{+}$ is given in advance. The optimization problem A is the minimization problem of ${\sf MSE}(t^{*})$ under the given degree sequence where $t^{*}$ is the predetermined target time given in advance. The precise formulation of the problem is given as follows: $\displaystyle\mbox{minimize }{\sf MSE}(t^{*})$ subject to: $\displaystyle\bm{L}$ $\displaystyle=\\{L_{ij}\\}\in\mathbb{R}^{n\times n}$ (66) $\displaystyle\bm{L}$ $\displaystyle=\bm{L}^{T}$ (67) $\displaystyle\bm{L}$ $\displaystyle\bm{1}=\bm{0}$ (68) $\displaystyle\|\mbox{diag}(\bm{L})-\bm{d}\|_{2}<\theta$ (69) $\displaystyle L_{ij}$ $\displaystyle=0,\ (i,j)\notin E.$ (70) The constraint (67) is imposed for the symmetry of the edge weight $\mu_{ij}=\mu_{ji}$ for $(i,j)\in E$. The row sum constraint (68) is needed for satisfying (8). The constraint (69) means that $\bm{L}$ should be close enough to the given degree sequence. The positive constant $\theta$ can be seen as a tolerance parameter. One way to interpret the optimization problem A is to consider the graph $G$ representing the wireless connection between terminals $i\in[n]$. The degree sequence $\bm{d}=(d_{1},d_{2},\ldots,d_{n})$ can be seen as an allocated receive total wireless power, i.e., the terminal $i$ can receive the neighbouring signals up to the total power $d_{i}$. If an average consensus protocol is used in such a wireless network for specific applications, it is desirable to optimize the ${\sf MSE}(t^{*})$ while satisfying the power constraint. #### V-A2 Optimization problem B Assume that a real constant $D\in\mathbb{R}^{+}$ is given in advance. The optimization problem B is the minimization problem of ${\sf MSE}(t^{*})$ under the situation that the diagonal sum of the Laplacian matrix $\bm{L}$ is equal to $D$. The formulation is given as follows: $\displaystyle\mbox{minimize }{\sf MSE}(t^{*})$ subject to: $\displaystyle\bm{L}$ $\displaystyle=\\{L_{ij}\\}\in\mathbb{R}^{n\times n}$ (71) $\displaystyle\bm{L}$ $\displaystyle=\bm{L}^{T}$ (72) $\displaystyle\bm{L}$ $\displaystyle\bm{1}=\bm{0}$ (73) $\displaystyle\left|\sum_{i=1}^{n}L_{ii}-D\right|<\theta$ (74) $\displaystyle L_{ij}$ $\displaystyle=0,\ (i,j)\notin E.$ (75) Following the interpretation above, the power allocation is also optimized in this problem. ### V-B Minimization based on deep-unfolded EM method Advances in deep neural networks have had a strong impact on the design of algorithms for communications and signal processing [14, 15, 16]. Deep unfolding can be seen as a very effective way to improve the convergence of iterative algorithms. Gregor and LeCun introduced the Learned ISTA (LISTA) [21]. Borgerding et al. also proposed variants of AMP and VAMP with trainable capability [19][20]. Trainable ISTA(TISTA) [23] is another trainable sparse signal recovery algorithm with fast convergence. TISTA requires only a small number of trainable parameters, which provides a fast and stable training process. Another advantage of deep unfolding is that it has a relatively high interpretability of learning results. The concept behind deep unfolding is rather simple. We can embed trainable parameters into the original iterative algorithm and then unfold the signal- flow graph of the original algorithm. The standard supervised training techniques used in deep learning, such as Stochastic Gradient Descent (SGD) and back propagation, can then be applied to the unfolded signal-flow graph to optimize the trainable parameters. The combination of deep unfolding and the differential equation solvers [24] is a current area of active research in scientific machine learning. It should be noted, however, that the technique is not limited to applications within scientific machine learning. In this subsection, we introduce an optimization algorithm that is based on the deep-unfolded EM method. The central idea is to use a loss function that approximates ${\sf MSE}(t^{*})$. By using a stochastic gradient descent approach with this loss function, we can obtain a near-optimal solution for both optimization problems A and B. The proposed method can be easily implemented using any modern neural network framework that includes an automatic differentiation mechanism. The following subsections will provide a more detailed explanation of the proposed method. #### V-B1 Mini-batch for optimization In an optimization process described below, a number of mini-batches are randomly generated. A mini-batch consists of $\displaystyle{\cal M}\equiv\\{(\bm{c}_{1},\gamma_{1}),(\bm{c}_{2},\gamma_{2}),\ldots(\bm{c}_{K},\gamma_{K})\\}.$ (76) The size parameter $K$ is called the mini-batch size. The initial value vector $\bm{c}_{i}$ follows Gaussian distribution, i.e., $\bm{c}_{i}\sim{\cal N}(\bm{0},\bm{I})(i\in[n])$. The corresponding average value are obtained by $\gamma_{i}\equiv(1/n)\bm{c}_{i}^{T}\bm{I}$. #### V-B2 Loss function for Optimization problem A The loss function corresponding to a mini-batch ${\cal M}$ is given by $\displaystyle E_{\cal M}(\bm{L})\equiv\frac{1}{K}\sum_{i=1}^{K}\|\bm{\chi}(\bm{c}_{i})-\gamma_{i}\bm{1}\|^{2}_{2}+P_{A}(\bm{L}),$ (77) where $\bm{\chi}(\bm{c}_{i})\equiv\bm{x}^{(N)}$ is the random variable given by the Euler-Maruyama recursion: $\displaystyle\bm{x}^{(k+1)}=\bm{x}^{(k)}-\eta\bm{L}\bm{x}^{(k)}+\alpha\sqrt{\eta}\bm{z}^{(k)},\ k=0,1,2,\ldots,N,$ (78) with $\bm{x}^{(0)}=\bm{c}_{i}$. The first term of the loss function can be regarded as an approximation of ${\sf MSE}(t^{*})$: $\displaystyle\frac{1}{K}\sum_{i=1}^{K}\|\bm{\chi}(\bm{c}_{i})-\gamma_{i}\bm{1}\|^{2}_{2}\simeq{\sf MSE}(t^{*})$ (79) for sufficiently large $K$ and $T=t^{*}$. The function $P_{A}(\bm{L})$ is a penalty function corresponding to the constraints (67)–(70) defined by $\displaystyle P_{A}(\bm{L})$ $\displaystyle\equiv\rho_{1}\|\bm{L}-\bm{L}^{T}\|_{F}^{2}+\rho_{2}\|\bm{L}\bm{1}\|_{2}^{2}+\rho_{3}\|\mbox{diag}(\bm{L})-\bm{d}\|_{2}^{2}$ $\displaystyle+\rho_{4}\|\bm{L}$ $\displaystyle\odot\bm{M}\|_{F}^{2},$ (80) where $\bm{M}=\\{M_{ij}\\}$ is the mask matrix defined by $\displaystyle M_{ij}\equiv\left\\{\begin{array}[]{cc}1,&(i,j)\notin E\\\ 0,&\mbox{otherwise}.\end{array}\right.$ (83) The operator $\odot$ represents the Hadamard matrix product. The positive constants $\rho_{i}(i\in[4])$ controls relative strength of each penalty term. The first term of the penalty function corresponds to the symmetric constraint (67). The term $\|\bm{L}\bm{1}\|_{2}^{2}$ is the penalty term for the row sum constraint (68). The third term $\|\mbox{diag}(\bm{L})-\bm{d}\|_{2}^{2}$ is included for the degree constraint. The last term $\|\bm{L}\odot\bm{M}\|_{F}^{2}$ enforces $L_{ij}$ to be very small if $(i,j)\notin E$. Due to these penalty terms in $P_{A}(\bm{L})$, the violations on the constraints (67)–(70) are suppressed in an optimization process. #### V-B3 Loss function for Optimization problem B For Optimization problem B, we use almost the same same loss function: $\displaystyle E_{\cal M}(\bm{L})\equiv\frac{1}{K}\sum_{i=1}^{K}\|\bm{\chi}(\bm{c}_{i})-\gamma_{i}\bm{1}\|^{2}_{2}+P_{B}(\bm{L}).$ (84) In this case, we use the penalty function matched to the feasible conditions of Optimization problem B: $\displaystyle P_{B}(\bm{L})$ $\displaystyle\equiv\rho_{1}\|\bm{L}-\bm{L}^{T}\|_{F}^{2}+\rho_{2}\|\bm{L}\bm{1}\|_{2}^{2}+\rho_{3}\left(\sum_{i=1}^{n}L_{ii}-D\right)^{2}$ $\displaystyle+\rho_{4}\|\bm{L}$ $\displaystyle\odot\bm{M}\|_{F}^{2}.$ (85) The third term of $P_{B}(\bm{L})$ corresponds to the diagonal sum condition (74). #### V-B4 Optimization process The optimization process is summarized in Algorithm 1. This optimization algorithm is mainly based on the Deep-unfolded Euler-Maruyama (DU-EM) method for approximating ${\sf MSE}(t^{*})$. The initial value of the matrix $\bm{L}$ is assumed to be the $n\times n$ zero matrix $\bm{O}^{n\times n}$. The main loop can be regarded as a stochastic gradient descent method minimizing the loss values. The update of $\bm{L}$ (line 5) can be done by any optimizer such as the Adam optimizer. The gradient of the loss function (line 4) can be easily evaluated by using an automatic differentiation mechanism included in recent neural network frameworks such as TensorFlow, PyTorch, Jax, and Flux.jl with Julia. The block diagram of the Algorithm 1 is shown in Fig. 4. Algorithm 1 Optimization process using DU-EM method 0: graph $G$, tolerance $\theta$, degree sequence $\bm{d}$ or degree sum $D$ 0: Laplacian matrix $\bm{L}_{out}$ 1: Let $\bm{L}\equiv\bm{O}^{n\times n}$ 2: for $i=1$ to $I$ do 3: Generate a mini-batch ${\cal M}$ randomly. 4: Compute the gradient of the loss function $\displaystyle\bm{g}\equiv\nabla E_{\cal M}(\bm{L})$ 5: The matrix $\bm{L}$ is updated by using $\bm{g}$. 6: end for 7: $\bm{L}_{out}\equiv\mbox{round}_{\theta,*}(\bm{L})$ Figure 4: Block diagram of optimization process in Algorithm 1. The core of the algorithm is the DU-EM method for approximating ${\sf MSE}(t^{*})$. Several standard deep learning techniques such as back propagation and stochastic gradient descent can be applied to update the trainable matrix $\bm{L}$. The stochastic optimization process outlined in Algorithm 1 is unable to guarantee that the obtained solution will be strictly feasible. To ensure feasibility, it is necessary to search for a feasible solution that is near the result obtained by optimization. This is accomplished by using the round function $\mbox{round}_{\theta,*}(\cdot)$ at line 7 of Algorithm 1. The specific details for the round function used for optimization problem A are outlined in Algorithm 2. The first step in the algorithm, $\bm{L}\equiv(\bm{L}_{in}+\bm{L}_{in}^{T})/2$, ensures that the resulting matrix is symmetric. The nested loop from line 2 to line 7 is used to enforce the degree constraint and the constraint $L_{ij}=0\ (i,j)\notin E$. The single loop from line 9 to line 11 is implemented to satisfy the constraint $\bm{L}\bm{1}=\bm{0}$. The output of the round function $\mbox{round}_{\theta,\bm{d}}(\cdot)$ guarantees that the constraints (67)-(70) of optimization problem A are strictly satisfied. A similar round function can be constructed for optimization problem B, which is presented in Algorithm 3. Algorithm 2 Round function $\mbox{round}_{\theta,\bm{d}}(\cdot)$ for Opt. prob. A 0: Matrix $\bm{L}_{in}$, degree sequence $\bm{d}$, threshold value $\theta$ 0: Laplacian matrix $\bm{L}_{out}$ satisfying (67)–(70) 1: Let $\bm{L}\equiv(\bm{L}_{in}+\bm{L}_{in}^{T})/2$ 2: for $i=1$ to $n$ do 3: $L_{ii}\equiv d_{i}$ 4: for $j=1$ to $n$ do 5: If $(i,j)\notin E$, then let $L_{ij}\equiv 0$ 6: end for 7: end for 8: $\bm{\epsilon}=(\epsilon_{1},\ldots,\epsilon_{n})^{T}\equiv\bm{L}\bm{1}$ 9: for $i=1$ to $n$ do 10: Let $L_{ii}\equiv L_{ii}-\epsilon_{i}$ 11: end for 12: if $\|\mbox{diag}(\bm{L})-\bm{d}\|_{2}\geq\theta$ then 13: Quit with declaration “optimization failed” 14: end if 15: Output $\bm{L}_{out}\equiv\bm{L}$ Algorithm 3 Round function $\mbox{round}_{\theta,D}(\cdot)$ for Opt. prob. B 0: Matrix $\bm{L}_{in}$, degree sum $D$, threshold value $\theta$ 0: Laplacian matrix $\bm{L}_{out}$ 1: Let $\bm{L}\equiv(\bm{L}_{in}+\bm{L}_{in}^{T})/2$ 2: for $i=1$ to $n$ do 3: for $j=1$ to $n$ do 4: If $(i,j)\notin E$, then let $L_{ij}\equiv 0$ 5: end for 6: end for 7: $\bm{\epsilon}=(\epsilon_{1},\ldots,\epsilon_{n})^{T}\equiv\bm{L}\bm{1}$ 8: for $i=1$ to $n$ do 9: Let $L_{ii}\equiv L_{ii}-\epsilon_{i}$ 10: end for 11: if $\left|\sum_{i=1}^{n}L_{ii}-D\right|\geq\theta$ then 12: Quit with declaration “optimization failed” 13: end if 14: Output $\bm{L}_{out}\equiv\bm{L}$ ## VI Numerical results ### VI-A Choice of Number of bins for EM-method In the previous sections, we proposed a DU-based optimization method. This section presents results of numerical experiments. For these experiments, we used the automatic differentiation mechanism provided by Flux.jl [25] on Julia Language [26]. Before discussing the optimization of MSE, we first examine the choice number of bins, $N$. Small $N$ is beneficial for computational efficiency but it may lead to inaccurate estimation of MSE. In this subsection, we will compare the Monte carlo estimates of MSE estimated by the EM-method. The Karate graph is a well-known graph of a small social network. It represents the relationships between 34 members of a karate club at a university. The graph consists of 34 nodes, which represent the members of the club, and 78 edges, which represent the relationships between the members. Figure 5 compares three cases, i.e., $N=100,250,1000$. No visible difference can be observed in the range from $T=0$ to $T=5$. In the following experiments, we will use $N=250$ for EM-method. Figure 5: MSE values estimated by Monte Carlo method based on the EM method. Karate graph ($n=34$) and its unweighted Laplacian is assumed. ### VI-B Petersen graph (Optimization problem A) Petersen graph is a 3-regular graph with $n=10$ nodes (Fig.6(a)). In this subsection, we will examine the behavior of our optimization algorithm of ${\sf MSE}(t)$ for Petersen graph. Figure 6: Small graphs: (a) Cycle graph, (b) Petersen graph, (c) House graph. An adjacency matrix $\bm{A}\equiv\\{A_{ij}\\}\in\mathbb{R}^{n\times n}$ of a graph $G\equiv(V,E)$ is defined by $\displaystyle A_{ij}\equiv\left\\{\begin{array}[]{cc}1,&(i,j)\in E\\\ 0,&\mbox{otherwise}.\end{array}\right.$ (88) An unweighted Laplacian matrix $\bm{L}$ is defined by $\displaystyle\bm{L}\equiv\bm{D}-\bm{A},$ (89) The degree matrix $\bm{D}=\\{D_{ij}\\}$ is a diagonal matrix where $D_{ii}$ is the degree of the node $i$. Namely, an unweighted Laplacian corresponds the case where $\mu_{ij}=\mu_{ji}=1$ for any $(i,j)\in E$. In the following discussion, let $\bm{L}_{P}$ be the unweighted Laplacian matrix of Petersen graph. We assume Optimization problem A with the degree sequence $\bm{d}\equiv\mbox{diag}(\bm{L}_{P})=(3,3,\ldots,3)$. The parameter setting is as follows. The mini-batch size is set to $K=25$. The noise intensity is $\alpha=0.3$. The penalty coefficients are $\rho_{1}=\rho_{2}=\rho_{3}=\rho_{4}=10$. For time discretization, we use $T=4,N=250$. The number of iterations for an optimization process is set to 3000. The tolerance parameter is set to $\theta=0.1$. In the optimization process, we used the Adam optimizer with a learning rate of 0.01. The loss values of an optimization process of Algorithm 1 are presented in Fig.7. In the initial stages of the optimization process, the loss value is relatively high since the initial $\bm{L}$ is set to the zero matrix, which means that the system cannot achieve average consensus. The loss value decreases monotonically until around iteration 700, after which it fluctuates within a range of $700\leq k\leq 3000$. The graph shows that the matrix $\bm{L}$ in Algorithm 1 is being updated appropriately and that the loss value, which approximates ${\sf MSE}(t)$, is decreasing. Figure 7: Loss values in an optimization process: Optimization prob. A for Petersen graph Let us denote the Laplacian matrix obtained by the optimization process as $\bm{L}^{*}$. Table I summarizes several important quantities regarding $\bm{L}^{*}$. The top 4 rows of Table I indicate that $\bm{L}^{*}$ is certainly a feasible solution satisfying (67)–(70) because we set $\theta=0.1$. This numerical results confirms that the round function $\mbox{round}_{\theta,\bm{d}}(\cdot)$ works appropriately. The last row of Table I shows that $\bm{L}^{*}$ is very close to the unweighted Laplacian matrix $\bm{L}_{P}$. Since Petersen graph is regular and has high symmetry, it is conjectured that $\bm{L}_{P}$ is the optimal solution for Optimization problem A. Thus, the closeness between $\bm{L}_{P}$ and $\bm{L}^{*}$ can be seen as a convincing result. TABLE I: Several quantities on optimization result $\bm{L}^{*}$ $\|\bm{L}^{*}-\bm{L}^{*T}\|_{F}$ | 0 ---|--- $\|\bm{L}^{*}\bm{1}\|_{2}$ | 0 $\|\bm{L}^{*}\odot M\|_{F}$ | 0 $\|\mbox{diag}(\bm{L}^{*})-\bm{d}\|_{2}$ | $5.74\times 10^{-3}$ $\|\bm{L}_{P}-\bm{L}^{*}\|_{F}$ | 0.188 The MSE values of the optimization result $\bm{L}^{*}$ and the unweighted Laplacian matrix $\bm{L}_{P}$ are presented in Fig.8. These values are evaluated by the MSE formula (64). No visible difference can be seen between two curves. This means that Algorithm 1 successfully found a good solution for Optimization problem A in this case. Figure 8: Petersen graph: MSE values of the optimization result $\bm{L}^{*}$ and the unweighted Laplacian matrix $\bm{L}_{P}$. ### VI-C Karate graph (Optimization problem A) We here consider Optimization problem A on the Karate graph. Let $\bm{L}_{K}$ be the unweighted Laplacian matrix of the Karate graph. The target degree sequence is set to $\bm{d}\equiv\mbox{diag}(\bm{L})=(16,9,10,\ldots,12,17).$ The parameter setting for an optimization process is given as follows. The mini-batch size is set to $K=50$. The noise intensity is set to $\alpha=0.3$. The penalty coefficients are $\rho_{1}=\rho_{2}=\rho_{3}=\rho_{4}=10$. We use $T=2,N=250$ for DU-EM method. The number of iterations for an optimization process is set to 5000. The tolerance is set to $\theta=0.1$. In the optimization process, we used the Adam optimizer with learning rate 0.01. Assume that $\bm{L}^{*}$ is the Laplacian matrix obtained by an optimization process. The matrix $\bm{L}^{*}$ is a feasible solution satisfying all the constraints (67)–(70). For example, we have $\|\mbox{diag}(\bm{L}^{*})-\bm{d}\|_{2}=0.0894<0.1$. Figure 9 presents the absolute values of non-diagonal elements in $\bm{L}_{K}$ and $\bm{L}^{*}$. According to its definition, the absolute value of a non-diagonal element of $\bm{L}_{K}$ take the value one (left panel). On the other hand, we can observe that non-diagonal elements of $\bm{L}^{*}$ takes the absolute values in the range $0$ to $1.5$. Figure 9: Absolute values of non-diagonal elements in $\bm{L}_{K}$ and $\bm{L}^{*}$: (left panel) Laplacian matrix ${\bm{L}}_{K}$ of the Karate graph, (right panel) The Laplacian matrix $\bm{L}^{*}$ obtained by an optimization process. We present the MSE values of the optimization result $\bm{L}^{*}$ and the unweighted Laplacian matrix $\bm{L}_{K}$ in Fig.10. These values are evaluated by the MSE formula (64). It can be seen that the optimized Laplacian $\bm{L}^{*}$ provides smaller MSE values. In this case, appropriate assignment of weights $\mu_{ij}$ improves the noise immunity of the system. The inverse eigenvalue sums of the Laplacian matrices $\bm{L}_{K}$ and $\bm{L}^{*}$ are 13.83 and 13.41, respectively. In this case, the optimization process of Algorithm 1 can successfully provide a feasible Laplacian matrix with smaller inverse eigenvalue sum. As shown in (65), the inverse eigenvalue sum determines the behavior of ${\sf MSE}(t)$. Figure 10: Karate graph: MSE values of $\bm{L}^{*}$ ($\bm{d}=\mbox{diag}(\bm{L}_{K})$) and the unweighted Laplacian matrix $\bm{L}_{K}$. ### VI-D House graph (Optimization problem B) The house graph (Fig.6(c)) is a small irregular graph with 5 nodes defined by the adjacency matrix: $\displaystyle\bm{A}=\begin{pmatrix}0&1&1&0&0\\\ 1&0&0&1&0\\\ 1&0&0&1&1\\\ 0&1&1&0&1\\\ 0&0&1&1&0\\\ \end{pmatrix}.$ (90) We thus have the unweighted Laplacian $\bm{L}_{H}$ of the house graph as $\displaystyle\bm{L}_{H}=\begin{pmatrix}2&-1&-1&0&0\\\ -1&2&0&-1&0\\\ -1&0&3&-1&-1\\\ 0&-1&-1&3&-1\\\ 0&0&-1&-1&2\\\ \end{pmatrix},$ (91) where the diagonal sum of $\bm{L}_{H}$ is 12. We made two optimizations for $D=12$ and $D=20$. The parameter setting is almost the same as the one used in the previous subsection. Only the difference is to use $\rho_{3}=0.1$ as the diagonal sum penalty constant. As results of the optimization processes, we have two Laplacian matrices $\bm{L}^{*}_{12}$ $(D=12)$ and $\bm{L}^{*}_{24}$ $(D=24)$ as follows: $\displaystyle\bm{L}^{*}_{12}=\begin{pmatrix}2.29&-1.05&-1.23&0&0\\\ -1.05&2.29&0&-1.24&0\\\ -1.23&0&2.70&-0.44&-1.03\\\ 0&-1.24&-0.44&2.71&-1.03\\\ 0&0&-1.03&-1.03&2.06\\\ \end{pmatrix}$ (92) $\displaystyle\bm{L}^{*}_{24}=\begin{pmatrix}4.80&-2.70&-2.09&0&0\\\ -2.70&4.79&0&-2.08&0\\\ -2.09&0&4.81&-0.37&-2.35\\\ 0&-2.08&-0.37&4.85&-2.40\\\ 0&0&-2.35&-2.40&4.74\\\ \end{pmatrix}$ (93) The diagonal sums of $\bm{L}^{*}_{12}$ and $\bm{L}^{*}_{24}$ are 12.04 and 23.99, respectively. Compared with $\bm{L}^{*}_{12}$ with $\bm{L}_{H}$, the diagonal elements of $\bm{L}^{*}_{12}$ are more flat: $\displaystyle\mbox{diag}(\bm{L}^{*}_{12})$ $\displaystyle=(2.29,2.29,2.70,2.71,2.06)^{T},$ (94) $\displaystyle\mbox{diag}(\bm{L}_{H})$ $\displaystyle=(2,2,3,3,2)^{T}.$ (95) The MSE values of the optimization result $\bm{L}^{*}_{12},\bm{L}^{*}_{24}$ and the unweighted Laplacian matrix $\bm{L}_{H}$ are shown in Fig.11. We can observe that $\bm{L}_{12}^{*}$ achieves slightly smaller MSE values compared with the unweighted Laplacian matrix $\bm{L}_{H}$. The Laplacian matrix $\bm{L}^{*}_{24}$ provides much smaller MSE values than those of $\bm{L}_{H}$. The sums of inverse eigenvalues are $1.64,1.59,0.82$ for $\bm{L}_{H}$, $\bm{L}_{12}^{*}$, and $\bm{L}_{24}^{*}$, respectively. Figure 11: House graph: MSE values of $\bm{L}^{*}_{12},\bm{L}^{*}_{20}$ and the unweighted Laplacian matrix $\bm{L}_{H}$. ### VI-E Barabási-Albert (BA) random graphs (Optimization problem B) As an example of random scale-free networks, we here handle Barabási-Albert random graph which use a preferential attachment mechanism. The number of edges between a new node to existing nodes is assumed to be 5. In this experiment, we generated an instance of Barabási-Albert random graph with 50 nodes. The unweighted Laplacian of the instance is denoted by $\bm{L}_{B}$. The sum of the diagonal elements of $\bm{L}_{B}$ is $450$. The parameter setting for optimization is the same as the one used in the previous subsection except for $D=450$. The output of the optimization algorithm is referred to as $\bm{L}^{*}$. Figure 12 presents the MSE values of the original unweighted Laplacian $\bm{L}_{B}$ and the optimization output $\bm{L}^{*}$. We can observe that the optimized MSE values are substantially smaller than those of the unweighted Laplacian $\bm{L}_{B}$. The sums of inverse eigenvalues for $\bm{L}^{*}$ and $\bm{L}_{B}$ are $6.44$ and $7.16$, respectively. Figure 13 illustrates the values of diagonal elements of $\bm{L}^{*}$ and $\bm{L}_{B}$. It can be observed that the values distribution of $\bm{L}^{*}$ is almost flat although the values of $\bm{L}_{B}$ varies from 5 to 21. This observation is consistent with the tendency observed in the previous subsection regarding the house graph. Figure 12: Barabási-Albert random graph: MSE values of $\bm{L}^{*}$ and the unweighted Laplacian matrix $\bm{L}_{B}$. Figure 13: Comparison of diagonal elements of $\bm{L}^{*}$ and $\bm{L}_{B}$. ## VII Conclusion In this paper, we have formulated a noisy average consensus system through a SDE. This formulation allows for an analytical study of the stochastic dynamics of the system. We derived a formula for the evolution of covariance for the EM method. Through the weak convergence property, we have established Theorem 1 and derived a MSE formula that provides the MSE at time $t$. Analysis of the MSE formula reveals that the sum of inverse eigenvalues for the Laplacian matrix is the most significant factor impacting the MSE dynamics. To optimize the edge weights, a deep unfolding-based technique is presented. The quality of the solution has been validated by numerical experiments. It is important to note that the theoretical understanding gained in this study will also provide valuable perspective on consensus-based distributed algorithms in noisy environments. In addition, the methodology for optimization proposed in this paper is versatile and can be adapted for various algorithms operating on graphs. The exploration of potential applications will be an open area for further studies. ## Acknowledgement This study was supported by JSPS Grant-in-Aid for Scientific Research (A) Grant Number 22H00514. The authors thank Prof. Masaki Ogura for letting us know the related work [11] on discrete-time average consensus systems. ## References * [1] T. Wadayama and A. Nakai-Kasai, “Continuous-time noisy average consensus system as Gaussian multiple access channel,” IEEE International Symposium on Information Theory, (ISIT) 2022. * [2] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of agents with switching topology and time-delays,” IEEE Trans. Automat. Contr., vol. 49, no. 9, pp. 1520–1533, Sept. 2004. * [3] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proceedings of the IEEE, vol. 95, no. 1, pp. 215–233, 2007\. * [4] W. Reb and R. W. Beard, “Consensus seeking in multi-agent systems under dynamically changing interaction topologies,” IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 655–661, 2005. * [5] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,” Systems and Control Letters, vol. 53, pp. 65–78, 2004. * [6] W. Ren, “Consensus strategies for cooperative control of vehicle formations,” IET Control Theory and Applications, vol.2, pp. 505–512, 2007. * [7] P. E. Kloeden and E. Platen, “Numerical solution of stochastic differential equations,” Springer-Verlag, 1991. * [8] B. Oksendal, “Stochastic differential equations: an introduction with applications,” Springer, 2010. * [9] C. Godsil and G. F. Royle, “Algebraic graph theory,” Springer, 2001. * [10] F. Chung, “Spectral graph theory,” American Mathematical Society, 1997. * [11] A. Jadbabaie and A. Olshevsky, “On performance of consensus protocols subject to noise: Role of hitting times and network structure,” IEEE 55th Conference on Decision and Control (CDC), pp. 179-184, 2016. * [12] R. Rajagopal and M. J. Wainwright, “Network-based consensus averaging with general noisy channels,” IEEE Trans. Signal Process., vol. 59, no. 1, pp. 373–385, Jan. 2011. * [13] M. Kishida, M. Ogura, Y. Yoshida, and T. Wadayama, “Deep learning-based average consensus,” IEEE Access, vol. 8, pp. 142404 - 142412, 2020. * [14] B. Aazhang, B. P. Paris and G. C. Orsak, “Neural networks for multiuser detection in code-division multiple-access communications,” IEEE Trans. Comm., vol. 40, no. 7, pp. 1212-1222, Jul. 1992. * [15] E. Nachmani, Y. Beéry and D. Burshtein, “Learning to decode linear codes using deep learning,” 2016 54th Annual Allerton Conf. Comm., Control, and Computing, 2016, pp. 341-346. * [16] T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,” IEEE Trans. Cog. Comm. Net., vol. 3, no. 4, pp. 563-575, Dec. 2017. * [17] Y. A. LeCun, L. Bottou, G. B. Orr, and K. R. Müller, “Efficient backprop,” in Neural networks: Tricks of the trade, G. B. Orr and K. R. Müller, Eds. Springer-Verlag, London, UK, 1998, pp. 9-50. * [18] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Nature, vol. 323, no. 6088, pp. 533-536, Oct. 1986. * [19] M. Borgerding and P. Schniter, “Onsager-corrected deep learning for sparse linear inverse problems,” 2016 IEEE Global Conf. Signal and Inf. Process. (GlobalSIP), Washington, DC, Dec. 2016, pp. 227-231. * [20] M. Borgerding, P. Schniter, and S. Rangan, “AMP-inspired deep networks for sparse linear inverse problems, ” IEEE Trans, Sig. Process. vol 65, no. 16, pp. 4293-4308 Aug. 2017. * [21] K. Gregor, and Y. LeCun, “Learning fast approximations of sparse coding,” Proc. 27th Int. Conf. Machine Learning, pp. 399-406, 2010. * [22] A. Balatsoukas-Stimming and C. Studer, “Deep Unfolding for Communications Systems: A Survey and Some New Directions,” 2019 IEEE International Workshop on Signal Processing Systems (SiPS), pp. 266-271, 2019. * [23] D. Ito, S. Takabe, and T. Wadayama, ”Trainable ISTA for sparse signal recovery,” IEEE Transactions on Signal Processing, vol. 67, no. 12, pp. 3113-3125, 2019. * [24] C. Rackauckas, Y. Ma, J. Martensen, C. Warner, K. Zubov, R. Supekar, D. Skinner, A. Ramadhan, and A. Edelman, “Universal differential equations for scientific machine learning,” arXiv:2001.04385, 2020. * [25] M. Innes, “Flux: Elegant machine learning with Julia,” Journal of Open Source Software, 2018. * [26] J. Bezanson, S. Karpinski, B. Viral, and A. Edelman, “Julia: A fast dynamic language for technical computing,” arXiv preprint arXiv:1209.5145, 2012. * [27] R. Albert, A.-L. Barabasi, “Statistical mechanics of complex networks,” American Physical Society, Rev. Mod. Phys., vol.74 pp. 47–97, 2002.
# Discontinuous phase transition from ferromagnetic to oscillating states in a nonequilibrium mean-field spin model Laura Guislain Univ. Grenoble Alpes, CNRS, LIPhy, 38000 Grenoble, France Eric Bertin Univ. Grenoble Alpes, CNRS, LIPhy, 38000 Grenoble, France ###### Abstract We study a nonequilibrium ferromagnetic mean-field spin model exhibiting a phase with spontaneous temporal oscillations of the magnetization, on top of the usual paramagnetic and ferromagnetic phases. This behavior is obtained by introducing dynamic field variables coupled to the spins through non- reciprocal couplings. We determine a nonequilibrium generalization of the Landau free energy in terms of the large deviation function of the magnetization and of an appropriately defined smoothed stochastic time derivative of the magnetization. While the transition between paramagnetic and oscillating phase is continuous, the transition between ferromagnetic and oscillating phases is found to be discontinuous, with a coexistence of both phases, one being stable and the other one metastable. Depending on parameter values, the ferromagnetic points may either be inside or outside the limit cycle, leading to different transition scenarios. The stability of these steady states is determined from the large deviation function. We also show that in the coexistence region, the entropy production has a pronounced maximum as a function of system size. ## I Introduction A number of systems driven far from equilibrium are known to exhibit spontaneous collective oscillations. This is the case for instance for coupled oscillators, like the Kuramoto model with distributed frequencies [1, 2], or in models of identical coupled noisy oscillators [3, 4]. Interestingly, spontaneous oscillations have also been reported in systems composed of a large number of coupled units which individually do not oscillate in the absence of interaction. Standard examples include different types of chemical oscillators [5], and recent experimental and theoretical studies have also reported spontaneous oscillations in populations of biological cells [6, 7], assemblies of active particles with non-reciprocal interactions [8, 9], biochemical clocks [10, 11, 12], droplets in a fluid binary mixture [13], models of population dynamics [14, 15], socio-economic models [16, 17] or nonequilibrium spin systems [18, 19, 20]. At the deterministic level of description, valid in the infinite system size limit, the spontaneous emergence of oscillations can be described as a Hopf bifurcation [21] in the framework of dynamical system theory. However, many situations of experimental relevance involve mesoscopic systems for which fluctuations cannot be neglected, as in the case of biochemical clocks for instance [22]. An important consequence of the presence of fluctuations is that the coherence time of oscillations becomes finite [23, 24, 25, 26, 27]. At a phenomenological level, the onset of oscillations in a fluctuating system may be described as a stochastic Hopf bifurcation [28, 29]. Yet, a deeper understanding would require to cast this phenomenon in the general framework of nonequilibrium phase transitions, by explicitly connecting the collective level of description to the microscopic dynamics. One may in particular interpret the onset of oscillations in a large system of interacting degrees of freedom by extending to far-from-equilibrium systems the thermodynamic framework of phase transition introduced at equilibrium. Along this line, the entropy production density has been shown to play the role of a generalized thermodynamic potential, with a discontinuous derivative at the onset of spontaneous oscillations [30, 31, 32, 14, 33, 34, 11, 35, 36, 37]. Another approach to phase transitions consists in introducing order parameters associated with spontaneously broken symmetries [38]. At a mean-field level of description, one may then introduce a Landau free energy and determine its expansion close to the phase transition. While this approach has been originally designed for equilibrium systems, several recent works have extended it to different types of nonequilibrium situations in the context of spin models, to describe relaxation effects [39, 40], or the driving by an oscillatory field or by multiple heat baths [41]. Based on a large deviation theory approach, the spontaneous transition from a paramagnetic to an oscillating state has been recently described in a nonequilibrium Landau framework [42]. In this paper, we extend the results of Ref. [42] by considering within the same nonequilibrium Landau framework the transition from a ferromagnetic state to a state of spontaneous collective oscillations. We study a mean-field spin model where spins are coupled to dynamic fields in a non-reciprocal way, resulting in a breaking of detailed balance which allows for the onset of oscillations in some parameter range. The presence of spontaneous oscillations may be interpreted as an instance of a non-reciprocal phase transition [43, 44]. Both spin and field variables are also subjected to ferromagnetic couplings, with different values. The phase transition is characterized by determining a large deviation function of the magnetization and of a stochastic variable playing the role of a smoothed time derivative of the magnetization. The transition from the ferromagnetic state to the oscillating state is found to be discontinuous, with a coexistence of the two phases in the transition region. The large deviation function allows us to determine which phase is stable or metastable. We also characterize finite size effects in terms of entropy production. The paper is organized as follows. The model is defined in Sec. II, the method is presented in Sec. III and the main results of Ref. [42] on the continuous phase transition from paramagnetic to oscillating states are summarized and extended in Sec. IV. Then, Sec. V characterizes a first scenario of discontinuous phase transition from ferromagnetic to oscillating states, whereby the limit cycle surrounds the ferromagnetic points. A second scenario, in which ferromagnetic points stand outside the limit cycle, is studied in Sec. VI. Conclusions and perspectives are summarized in Sec. VII. ## II Model description ### II.1 Definition of the model We consider a generalization of the kinetic mean-field Ising model with ferromagnetic interactions introduced in [42], and sharing some similarities with related models having two spin populations [45, 18] or subjected to a feedback control [19, 46]. The model involves $2N$ microscopic variables: $N$ spins $s_{i}=\pm 1$ and $N$ fields $h_{i}=\pm 1$. We define the magnetization $m$ and average field $h$ as $m=\frac{1}{N}\sum_{i=1}^{N}s_{i}\,,\qquad h=\frac{1}{N}\sum_{i=1}^{N}h_{i}\,.$ (1) The stochastic dynamics consists in randomly flipping a single spin $s_{i}=\pm 1$ with rate $w_{s}^{\pm}$, or a single field $h_{i}=\pm 1$ with rate $w_{h}^{\pm}$. In a mean-field spirit, the flipping rates $w_{s}^{\pm}$ and $w_{h}^{\pm}$ are independent of $i$, and depend only on $m$ and $h$, $w_{s}^{\pm}(m,h)=\frac{1}{1+e^{\beta\Delta E_{s}^{\pm}(m,h)}},\quad w_{h}^{\pm}=\frac{1}{1+e^{\beta\Delta E_{h}^{\pm}(m,h)}},$ (2) with $\beta=T^{-1}$ the inverse temperature and $\Delta E_{s,h}^{\pm}(m,h)$ the variation of $E_{s,h}(m,h)$ when flipping a spin $s_{i}=\pm 1$ or a field $h_{i}=\pm 1$, where $\displaystyle E_{s}(m,h)$ $\displaystyle=-N\left(\frac{J_{1}}{2}m^{2}+\frac{J_{2}}{2}h^{2}+mh\right),$ (3) $\displaystyle E_{h}(m,h)$ $\displaystyle=E_{s}(m,h)+\mu Nhm\,.$ (4) When $\mu=0$, $E_{h}=E_{s}$ and the transition rates satisfy detailed balance with respect to the equilibrium probability distribution $P_{\rm eq}\propto e^{-\beta E_{s}}$. Detailed balance is broken as soon as $\mu\neq 0$, and $\mu$ can thus be interpreted as a parameter controling the distance to equilibrium. For fixed values of the interactions $J_{1}$ and $J_{2}$, the temperature $T$ and the distance to equilibrium $\mu$ are the two control parameters of the model. We denote as $\mathcal{C}=(s_{1},\dots,s_{N},h_{1},\dots,h_{N})$ the microscopic configuration of the system. Flips of the variables $s_{i}$ and $h_{i}$ are encoded into formal transition rates $W(\mathcal{C}^{\prime}|\mathcal{C})$ from a configuration $\mathcal{C}$ to a configuration $\mathcal{C}^{\prime}$. The probability $P(\mathcal{C},t)$ of a configuration $\mathcal{C}$ at time $t$ evolves according to the master equation $\partial_{t}P(\mathcal{C},t)=\sum_{\mathcal{C}^{\prime}(\neq\mathcal{C})}\big{[}W(\mathcal{C}|\mathcal{C}^{\prime})P(\mathcal{C}^{\prime},t)-W(\mathcal{C}^{\prime}|\mathcal{C})P(\mathcal{C},t)\big{]}.$ (5) ### II.2 Phase diagram in the deterministic limit We first study the bifurcation diagram of the system obtained in the deterministic limit $N\to\infty$. We compute the time derivatives $d_{t}\langle m\rangle$ and $d_{t}\langle h\rangle$ using the master equation Eq. (5), where $\langle x\rangle=\sum_{\mathcal{C}}m(\mathcal{C})P(\mathcal{C})$ for any observable $x$. We assume that the law of large numbers holds in the limit $N\to\infty$ so that $\langle f(m,h)\rangle\to f(m,h)$ for any regular function $f$. Finally we obtain a set of deterministic equations on $m(t)$ and $h(t)$ (see Appendix A): $\displaystyle\frac{dm}{dt}$ $\displaystyle=-m+\tanh[\beta(J_{1}m+h)],$ (6) $\displaystyle\frac{dh}{dt}$ $\displaystyle=-h+\tanh[\beta(J_{2}h+(1-\mu)m)].$ (7) We explore regimes where the magnetization $m(t)$ may exhibit oscillations. In dynamical systems theory, a limit cycle may generally be described in the plane defined by a variable and its time derivative, thus we introduce a new variable $\dot{m}=dm/dt$. The set of deterministic equations become $\displaystyle\frac{dm}{dt}=\dot{m}\,,\qquad\frac{d\dot{m}}{dt}=Y(m,\dot{m})$ (8) where $Y(m,\dot{m})$ has a lengthy expression, given in Appendix B [see Eq. (104)]. $Y(m,\dot{m})$ satisfies the symmetry $Y(-m,-\dot{m})=-Y(m,\dot{m})$. To study the fixed points of Eq. (8) and their stability, we decompose $Y(m,\dot{m})$ into a $\dot{m}$-independent contribution $Y(m,0)=-V^{\prime}(m)$ (9) [see Appendix B, Eq. (105) for its explicit expression] and a $\dot{m}$-dependent contribution $\dot{m}g(m,\dot{m})=Y(m,\dot{m})-Y(m,0)\,,$ (10) which defines the function $g(m,\dot{m})$. From Eq. (8), the fixed points $(m,\dot{m})=(m_{0},0)$ satisfy $Y(m_{0},0)=0$, and thus $V^{\prime}(m_{0})=0$ according to Eq. (9). One finds in particular that $(m,\dot{m})=(0,0)$ is always a fixed point of the system, because $V^{\prime}(0)=0$ by symmetry. Linearizing the dynamics around a fixed point $(m_{0},0)$, $m=m_{0}+\delta m$, $\dot{m}=\delta\dot{m}$, one has from Eq. (8) $\frac{d}{dt}\begin{pmatrix}\delta m\\\ \delta\dot{m}\end{pmatrix}=\mathbf{M}\begin{pmatrix}\delta m\\\ \delta\dot{m}\end{pmatrix}$ (11) with $\mathbf{M}=\begin{pmatrix}0&1\\\ -V^{\prime\prime}(m_{0})&g(m_{0},0)\end{pmatrix}.$ (12) The linear stability of the fixed point $(m_{0},0)$ is determined by the sign of the eigenvalues of the matrix $\mathbf{M}$, $\lambda_{\pm}=\frac{1}{2}g(m_{0},0)\left(1\pm\sqrt{1-\frac{4V^{\prime\prime}(m_{0})}{g(m_{0},0)^{2}}}\right).$ (13) The fixed point $(m_{0},0)$ is stable if both $\lambda_{+}$ and $\lambda_{-}$ are negative (or have a negative real part), implying that $V^{\prime\prime}(m_{0})>0$ and $g(m_{0},0)<0$. We see in particular from Eq. (13) that the fixed point $(m_{0},0)$ becomes unstable when $g(m_{0},0)$ is positive. We define the critical temperature $T_{c}=(J_{1}+J_{2})/2$ and the dimensionless deviation from $T_{c}$, $\varepsilon=\frac{T_{c}-T}{T_{c}}\,.$ (14) Using expression (104) of $Y(m,\dot{m})$, we get for $m_{0}=0$ and small $\varepsilon$ that $g(0,0)=a_{0}\varepsilon$, with $a_{0}=2T/T_{c}$. We also have $V^{\prime\prime}(0)=(\mu-\mu_{l}(T))/T^{2}$, where we define $\mu_{l}(T)$ as $\mu_{l}(T)=1-(J_{1}-T)(J_{2}-T).$ (15) Hence, the fixed point $(m,\dot{m})=(0,0)$ is linearly stable for $T>T_{c}$ [$\varepsilon<0$, implying $g(0,0)<0$] provided that $\mu>\mu_{l}(T)$ [implying $V^{\prime\prime}(0)>0$], and unstable otherwise. We define $\mu_{c}=\mu_{l}(T_{c})$. Two examples of stability diagrams, obtained from the numerical evaluation of the fixed points and their stability [given by the sign of the eigenvalues of Eq. (13)], are shown in Fig. 1 for different values of $J_{1}$ and $J_{2}$. Trajectories and existence of limit cycles are obtained from the numerical integration of Eqs. (6) and (7). A stable paramagnetic fixed point [denoted as P in Fig. 1(a,b)] is found at high enough temperature, while this point becomes unstable at low temperature. For low values of $(T,\mu)$, two symmetric ferromagnetic stable fixed points (F) are observed. At low $T$ and high $\mu$, an oscillating behavior (O) is observed. The transition lines between the three different behaviors meet at a tricritical point ($T_{c}$, $\mu_{c}$), see Fig. 1(a,b). Figure 1: Phase diagram of the deterministic dynamics Eq. (8) obtained numerically for (a) $J_{1}=0.6$, $J_{2}=0.4$ and (b) $J_{1}=J_{2}=0.5$. Three distinct behaviors are observed: a stable paramagnetic point (P), two stable ferromagnetic points (F), and an oscillating phase (O). On the thick dashed and dotted lines, the limit cycle and the ferromagnetic points coexist. On the blue dotted line, the ferromagnetic points are inside the limit cycle (Type-I coexistence) and on the red dashed line the ferromagnetic points are outside the limit cycle (Type-II coexistence). When the two lines meet, the bifurcation is heteroclinic. The orange line corresponds to $\mu_{l}(T)$ [Eq. (15)]: in plain for $T>T_{c}$ where it represents the limit of stability of the paramagnetic points and dotted for $T<T_{c}$ as indication. (c)-(f) Examples of trajectories $m(t)$ and phase space ($m(t),\dot{m}(t))$: (c),(d) Type-I coexistence for $J_{1}=0.6$, $J_{2}=0.4$, $T/T_{c}=0.9$ and $(\mu-\mu_{c})/\mu_{c}=-3.5\times 10^{-3}$ (blue dot in (a)); (e),(f) Type-II coexistence for $J_{1}=J_{2}=0.5$, $T/T_{c}=0.9$ and $(\mu-\mu_{c})/\mu_{c}=8\times 10^{-3}$ (red dot in (b)). Depending on the value of $\mu$, the bifurcation from the paramagnetic point to a limit cycle at $T_{c}$ which occurs for $\mu>\mu_{c}$ can either be continuous (supercritical Hopf bifurcation) or discontinuous (subcritical Hopf bifurcation) [42]. It is generically continuous when the couplings $J_{1}$ and $J_{2}$ are positive. The transition from the ferromagnetic points to a limit cycle is discontinuous [except for the particular values $J_{2}=\pm 2+J_{1}$] and we observe small regions of the parameter space where the ferromagnetic points and the limit cycle coexist. In Fig. 1, they are represented by thick dotted and dashed lines. Depending on the values of the parameters ($T$, $J_{1}$ and $J_{2}$), the ferromagnetic points can be either inside or outside the limit cycle, which leads to a topological classification of the transition into two differents types. In the following, we call discontinuous transition of Type I the case when the ferromagnetic points are inside the limit cycle, and discontinuous transition of Type II the case when the ferromagnetic points are outside the limit cycle. A discontinuous transition of Type I is typically found close to $T_{c}$ for $J_{1}>J_{2}$ (under additional assumptions that are specified below), as illustrated for $J_{1}=0.6$ and $J_{2}=0.4$ by the dotted blue line in Fig. 1(a). An example of trajectory $m(t)$ and phase space plot $(m(t),\dot{m}(t))$ is represented in Fig. 1(c),(d). A discontinuous transition of Type II is found for $J_{1}\leq J_{2}$ (under additional assumptions that are specified below) and $T<T_{c}$, see Fig. 1(b) for $J_{1}=J_{2}=0.5$. The corresponding trajectory $m(t)$ and its phase space representation $(m(t),\dot{m}(t))$ is plotted in Fig. 1(e),(f). The farther from $T_{c}$, the closer the ferromagnetic points and the limit cycle are. Note that in the case $J_{1}=0.6$ and $J_{2}=0.4$, one observes that for lower temperatures, the ferromagnetic points sit outside the limit cycle [dashed red line in Fig. 1(a)], similarly to the behavior displayed in Fig. 1(e), (f). At the point where the dotted blue line meets the dashed red line, a limit cycle joining the two ferromagnetic points and with an infinite period appears when the ferromagnetic points loose stability, corresponding to a heteroclinic bifurcation. Note also that for $J_{2}=\pm 2+J_{1}$, the transition is neither of type I or II, but is a continuous transition from the ferromagnetic points to the limit cycle corresponding to a heteroclinic bifurcation. We do not study this particular case in this paper, but a comment on the specificity of this case is made in Sec. VI.5. ### II.3 Close to the tricritical point Figure 2: Sign of $v_{1}(T_{c},\mu_{c})$ from Eq. (18) in the plane ($J_{1}$, $J_{2}$). In regions with $v_{1}>0$, the discontinuous transition between ferromagnetic and oscillating phases in the $(T,\mu)$ phase diagram is of Type I: a limit cycle coexists with ferromagnetic points located inside the cycle (see Fig. 1(a) and Sec. V). In the case $v_{1}=0$ with $J_{1}=J_{2}$ (red line), the discontinuous transition between ferromagnetic and oscillating phases in the $(T,\mu)$ phase diagram is of Type II: a limit cycle coexists with ferromagnetic points located outside the cycle (see Fig. 1(b) and Sec. VI). The case $v_{1}=0$ with $J_{1}\neq J_{2}$ is not discussed in details in this work. In the following, we focus on the transition close to $T_{c}$, i.e., for small $\varepsilon$ in order to use a perturbative theory. We consider that $m$ is small such that only the first orders of the series expansion of $V(m)$ are necessary. One finds at order $m^{4}$ for $V(m)$ and at order $m^{2}$ for $g(m,0)$ that $\displaystyle V(m)=\frac{\mu-\mu_{l}(T)}{2T^{2}}m^{2}+\frac{v_{1}(T,\mu)}{4}m^{4}+V_{0},$ (16) $\displaystyle g(m,0)=a_{0}\varepsilon-a_{1}m^{2},$ (17) where $V_{0}$ is at this stage an arbitrary constant, and where $v_{1}(T,\mu)$, $a_{0}$ and $a_{1}$ are given in Appendix B. In particular one has $a_{0},a_{1}>0$ and $v_{1}(T_{c},\mu_{c})=\frac{(J_{1}-J_{2})[4-(J_{1}-J_{2})^{2}]}{12(J_{1}+J_{2})}\,.$ (18) The sign of $v_{1}(T_{c},\mu_{c})$, which plays a key role in the behavior of the model, thus depends on the relative values of $J_{1}$ and $J_{2}$ (see Fig. 2). When $v_{1}>0$, ferromagnetic points exist for $\mu<\mu_{l}(T)$, and are given by $m_{0}^{2}=\frac{\mu_{l}(T)-\mu}{T^{2}v_{1}}\,,$ (19) i.e., nonzero solutions of the equation $V^{\prime}(m_{0})=0$. According to Eq. (13), and given that $V^{\prime\prime}(m_{0})>0$, ferromagnetic points are linearly stable when $g(m_{0},0)<0$, which corresponds to $\mu<\mu_{F}(T)\equiv\mu_{l}(T)-\frac{\varepsilon a_{0}T^{2}v_{1}}{a_{1}}\,.$ (20) Numerically, one observes that before ferromagnetic points become linearly unstable upon increasing $\mu$, they coexist over a tiny parameter range with a limit cycle that surrounds them. An example is given for $J_{1}=0.6$ and $J_{1}=0.4$ in Fig. 3, which displays the ferromagnetic points and the extension of the limit cycle as a function of $\mu$ at fixed $\varepsilon$ [Fig. 3(a)], as well as the corresponding trajectories $m(t)$ [Fig. 3(b)] and the coexisting trajectories in the phase space $(m,\dot{m})$ [Fig. 3(c)]. Unlike for smaller values of $T$, we observe that close to the tricritical point, the ferromagnetic points and the limit cycle are well separated. We also observe in this regime that the two symmetries $m\,\mapsto\,-m$ and $\dot{m}\,\mapsto\,-\dot{m}$ are separately valid to a good approximation, while they were previously valid only under the simultaneous transformation $(m,\dot{m})\,\mapsto\,(-m,-\dot{m})$. When $v_{1}\leq 0$, higher order terms in the expansion of $V(m)$ are necessary to obtain the ferromagnetic points and their stability. Numerically, we observe two scenarios: the first one is a heteroclinic bifurcation, the ferromagnetic points loose stability and a limit cycle with infinite period arises. This is the case in particular for $J_{2}=\pm 2+J_{1}$, for which $v_{1}(T_{c},\mu_{c})=0$. The second scenario is that before disappearing, the ferromagnetic points coexist with a small elliptic limit cycle. This is the case for $J_{1}=J_{2}=0.5$, see Fig. 4 where an example of the evolution with $\mu$ (at fixed $\varepsilon$) of the ferromagnetic fixed points and of the limit cycle are displayed, together with examples of trajectories $m(t)$. When $v_{1}(T_{c},\mu_{c})<0$, one finds numerically that the ferromagnetic phase and the paramagnetic phase coexist for $T\gtrsim T_{c},\mu\gtrsim\mu_{l}(T)$, so that $(T_{c},\mu_{c})$ is no longer a tricritical point, whereas for $v_{1}(T_{c},\mu_{c})=0$ which is verified for $J_{1}=J_{2}$ and for $J_{2}=\pm 2+J_{1}$, one finds that the three transition lines meet at $(T_{c},\mu_{c})$ (see Fig. 1 for $J_{1}=J_{2}=0.5$ for an example of bifurcation diagram). In the following, we focus on the case where $v_{1}(T_{c},\mu_{c})\geq 0$ where the three phases meet at the critical point $(T_{c},\mu_{c})$, in order to perform a perturbative analysis close to the tricritical point. Figure 3: Type-I coexistence close to the tricritical point, corresponding to ferromagnetic points inside the limit cycle, for $\varepsilon=(T_{c}-T)/T_{c}=10^{-4}$ and $J_{1}=0.6$, $J_{2}=0.4$. (a) Values of $m$ for the ferromagnetic points (orange lines) and for the limit cycle (blue shaded area) along the transition. At $\mu=\mu_{F}(T)$ [Eq. (20)] ferromagnetic points become linearly unstable. (b) Examples of trajectories for $(\mu-\mu_{F})/\mu_{F}=-5\times 10^{-7}$. (c) Corresponding phase space representation in the plane ($m$, $\dot{m}$). Figure 4: Type-II coexistence close to the tricritical point, corresponding to ferromagnetic points outside the limit cycle, for $\varepsilon=10^{-3}$ and $J_{1}=J_{2}=0.5$. (a) Values of $m$ for the ferromagnetic points (orange lines) and for the limit cycle (blue shaded area) along the transition. At $\mu=\mu_{F}(T)$ [here, $(\mu_{F}-\mu_{c})/\mu_{c}\approx 1.5\times 10^{-5}$] the ferromagnetic points become unstable. (b) Examples of trajectories for $(\mu-\mu_{F})/\mu_{F}=-5\times 10^{-6}$. (c) Corresponding phase space representation in the plane ($m$, $\dot{m}$). ## III Generalized Landau theory Figure 5: Trajectories $m(t)$: the dashed blue and orange lines correspond to deterministic trajectories and the green line to a trajectory for a finite system size $N=5000$ obtained from stochastic numerical simulations. Parameters: $J_{1}=J_{2}=0.5$, $\varepsilon=5\times 10^{-2}$ and $(\mu-\mu_{c})/\mu_{c}=3.88\times 10^{-3}$. Jumps between noisy oscillatory states and ferromagnetic states are observed. The deterministic limit provides knowledge on the different stable fixed points or limit cycles that are present in the system. However it lacks information on the behavior of the system at finite size $N$, such as knowledge on the macroscopic fluctuations around the stable points or cycles. But most importantly, in case of coexistence of solutions in the limit $N\to\infty$, the deterministic approach fails to predict which solution is the most stable one at finite but large size $N$. In addition, for moderate size $N$, one observes jumps between noisy oscillatory states and ferromagnetic states, as illustrated in Fig. 5. A statistical description of such a situation where the ferromagnetic points and the limit cycle are both linearly stable in the deterministic limit would thus be useful. We briefly recall in this section the nonequilibrium generalization of the Landau theory developed in [42], which allows for a description of phase transitions to oscillating states. ### III.1 Stochastic time derivative $\dot{m}$ We first introduce a new variable $\dot{m}$ that plays the role of a smoothed time derivative of the magnetization for finite-size systems. Following [42], we formally define the stochastic derivative $\dot{m}(\mathcal{C})$ of the magnetization $m(\mathcal{C})$ as $\dot{m}(\mathcal{C})=\sum_{\mathcal{C}^{\prime}(\neq\mathcal{C})}\left[m\left(\mathcal{C}^{\prime}\right)-m\left(\mathcal{C}\right)\right]W(\mathcal{C}^{\prime}|\mathcal{C}),$ (21) such that on average $d\langle m\rangle/dt=\langle\dot{m}\rangle$. Eq. (21) thus associates with each microscopic configuration $\mathcal{C}$ an observable $\dot{m}(\mathcal{C})$, which is a smoothed time derivative of $m$ because it is averaged over all possible transitions $\mathcal{C}\to\mathcal{C}^{\prime}$, for a fixed configuration $\mathcal{C}$. The advantage of this definition is that fluctuations of $\dot{m}$ are typically on the same scale as that of $m$, which is a key property for the large deviation approach described below. Taking instead the time derivative of $m\big{(}\mathcal{C}(t)\big{)}$ would lead to diverging, white-noise-like fluctuations which are not appropriate to develop a generalized Landau theory. Under the mean-field assumption, the formal transition rate $W(\mathcal{C}^{\prime}|\mathcal{C})$ can be reexpressed in terms of the flipping rates $w_{s}^{\pm}(m,h)$ to flip a spin $s_{i}=\pm 1$ defined in Eq. (2). When flipping a spin $s_{i}=\pm 1$, the magnetization change is given by $m(\mathcal{C}^{\prime})-m(\mathcal{C})=\mp 2/N$. Since there are $N(1\pm m)/2$ possibilities to choose a spin $s_{i}=\pm 1$, one finds: $\dot{m}=(1-m)w_{s}^{-}(m,h)-(1+m)w_{s}^{+}(m,h).$ (22) Using the expression (2) of the flipping rates $w_{s}^{\pm}(m,h)$, Eq. (22) becomes: $\dot{m}=-m+\tanh[\beta(J_{1}m+h)].$ (23) Note that the functional relation $\dot{m}(m,h)$ turns out to be identical to the functional relation (6) obtained in the deterministic limit $N\to\infty$. However, Eq. (23) is valid for any finite $N$, and the variables $m$ and $h$ are here stochastic variables. ### III.2 Large deviation function At finite size $N$, the dynamics of the system is determined by the master equation (5). Instead of considering $P(\mathcal{C})$ which involves $2^{2N}$ configurations $\mathcal{C}=\\{s_{1},\dots,s_{N},h_{1},\dots,h_{N}\\}$, we consider $P_{N}(m,\dot{m})$ the joint stationary probability density of the global observables $m$ and $\dot{m}$. The variations of $m$ and $\dot{m}$ during a transition scale as $1/N$. We introduce $\mathbf{d}_{k}$ such that $(\Delta m,\Delta\dot{m})=\pm\mathbf{d}_{k}/N$ with $k=1$ when flipping a spin $s_{i}=\pm 1$ and $k=2$ when flipping a field $h_{i}=\pm 1$, so that we have: $\displaystyle\textbf{d}_{1}$ $\displaystyle=\left(-2,2-2\beta J_{1}+2\beta J_{1}(m+\dot{m})^{2}\right),$ (24) $\displaystyle\textbf{d}_{2}$ $\displaystyle=-\left(0,-2\beta+2\beta(m+\dot{m})^{2}\right).$ (25) We note $NW_{k}^{\pm}$ the coarse-grained transition rates from a configuration $(m,h)$ to $(m^{\prime},h^{\prime})$, with $m^{\prime}=m\mp 2/N$ and $h^{\prime}=h$ if $k=1$, and $m^{\prime}=m$ and $h^{\prime}=h\mp 2/N$ if $k=2$. One has: $\displaystyle W_{1}^{\pm}$ $\displaystyle=\frac{(1\pm m)/2}{1+\exp[\pm 2\beta(J_{1}m+h)]},$ (26) $\displaystyle W_{2}^{\pm}$ $\displaystyle=\frac{(1\pm h)/2}{1+\exp[\pm 2\beta(J_{2}h+(1-\mu)m)]}.$ The coarse-grained master equation governing the evolution of $P_{N}(m,\dot{m})$ reads [42]: $\displaystyle\partial_{t}$ $\displaystyle P_{N}(m,\dot{m})=N\sum_{k,\sigma}\bigg{[}-W_{k}^{\sigma}(m,\dot{m})P_{N}(m,\dot{m})$ (27) $\displaystyle+W_{k}^{\sigma}\left((m,\dot{m})-\frac{\sigma\textbf{d}_{k}}{N}\right)P_{N}\left((m,\dot{m})-\frac{\sigma\textbf{d}_{k}}{N}\right)\bigg{]}$ where $k=1,2$ and $\sigma=\pm$. For large $N$, the stationary joint distribution $P_{N}(m,\dot{m})$ takes a large deviation form [47] $P_{N}(m,\dot{m})\underset{N\to\infty}{\sim}\exp[-N\phi(m,\dot{m})]\,,$ (28) a property justified by the theory of Markov jump processes with vanishing jump size [48]. Beside providing information on the fluctuations at finite system size, the large deviation function (or rate function) $\phi(m,\dot{m})$ determines the macroscopic phase of the system. Linearly stable solutions of the deterministic equations correspond to local minima of the large deviation function. When two or more linearly stable solutions are present in the deterministic equations, the global minima of $\phi$ gives the macroscopic phase of the system (i.e., the most stable one). Injecting the large deviation form (28) into Eq. (1) gives to order $(\nabla\phi)^{2}$, $\displaystyle\dot{m}\partial_{m}\phi+Y(m,\dot{m})\partial_{\dot{m}}\phi+D_{11}(\partial_{m}\phi)^{2}$ (29) $\displaystyle+2D_{12}\partial_{m}\phi\partial_{\dot{m}}\phi+D_{22}(\partial_{\dot{m}}\phi)^{2}=0\,,$ with $(\dot{m},Y(m,\dot{m}))=\sum_{k,\sigma}-\sigma\mathbf{d}_{k}W_{k}^{\sigma}\,.$ (30) The function $Y(m,\dot{m})$ is found to be the same function as the one introduced in the deterministic limit in Eq. (8). We introduce $\mathbf{D}=\\{D_{ij}(m,\dot{m})\\}$ as $\mathbf{D}\equiv\sum_{k}\mathbf{d}_{k}\cdot\mathbf{d}_{k}^{T}W_{k}^{\sigma},$ (31) whose explicit expression is given in Appendix B. We use the decomposition introduced in Eq. (10), $Y(m,\dot{m})=-V^{\prime}(m)+\dot{m}g(m,\dot{m})\,$ (32) and we focus, in this paper, on obtaining the large deviation function in regions where a fixed point ($m_{0},0$) looses stability in the deterministic limit, i.e., where $g(m_{0},0)$ changes sign, in order to use a perturbative framework in terms of the small parameter $u_{0}\equiv g(m_{0},0).$ (33) We assume $\nabla\phi=O(u_{0})$ since quadratic terms in $\nabla\phi$ have to balance the contribution in $u_{0}\dot{m}\partial_{\dot{m}}\phi$. At order $u_{0}$, Eq. (LABEL:eq:phi:quadrat) reduces to $\dot{m}\partial_{m}\phi-V^{\prime}(m)\partial_{\dot{m}}\phi=0.$ (34) The general solution of Eq. (34) reads [42] $\phi(m,\dot{m})=f\big{(}H(m,\dot{m})\big{)}+f_{0}$ (35) where the function $H(m,\dot{m})$ takes a form similar to a Hamiltonian, $H(m,\dot{m})=\frac{\dot{m}^{2}}{2}+V(m)\,.$ (36) The minimum value of $V(m)$ is set to $V=0$, and $f$ is at this stage an arbitrary function, satisfying for convenience $f(0)=0$. The constant $f_{0}$ in Eq. (35) ensures that the minimal value of $\phi(m,\dot{m})$ is zero. Contributions of order $u_{0}^{2}$ to Eq. (LABEL:eq:phi:quadrat) yield a condition determining the derivative $f^{\prime}(H)$ (see [42] for a detailed derivation) $f^{\prime}(H)=-\frac{\int_{m_{1}}^{m_{2}}dm\,\sqrt{2[H-V(m)]}\,g\left(m,\sqrt{2[H-V(m)]}\right)}{\int_{m_{1}}^{m_{2}}dm\,\bigg{[}D_{11}\frac{V^{\prime}(m)^{2}}{\sqrt{2[H-V(m)]}}+2D_{12}V^{\prime}(m)+D_{22}\sqrt{2[H-V(m)]}\bigg{]}}\,,$ (37) where $m_{1}$ and $m_{2}$ are such that $V(m_{1})=V(m_{2})=H$ and $V(m)\leq H$ for $m_{1}\leq m\leq m_{2}$. The form Eq. (35) of the large deviation function $\phi(m,\dot{m})$ can be interpreted as giving a statistical weight to deterministic trajectories determined by the Hamiltonian dynamics $\frac{dm}{dt}=\frac{\partial H}{\partial\dot{m}},\qquad\frac{d\dot{m}}{dt}=-\frac{\partial H}{\partial m},$ (38) valid at order $\varepsilon$, where the Hamiltonian $H(m,\dot{m})$ is defined in Eq. (36). Denoting $m_{0}$ a minimum of $V(m)$ (we assume here for simplicity that $V(m)$ has a single minimum or two symmetric minima) the case $H=V(m_{0})$ corresponds to a fixed point ($m_{0},0$) of the deterministic dynamics, whereas values $H>V(m_{0})$ correspond to closed orbits, and thus to oscillations. The most probable value of $H$, and thus the macroscopically observed behavior, is determined by the global minimum $H^{*}$ of $f(H)$. Note that the method used here follows similar lines as the determination of nonequilibrium potentials in dissipative dynamical systems [49, 50, 51]. When the three phases meet at the critical point $(T_{c},\mu_{c})$, the ferromagnetic points, noted $\pm m_{0}$, have a small amplitude ($m_{0}\leq\varepsilon$), so that $g(m_{0},0)\sim\varepsilon$ with $\varepsilon=(T_{c}-T)/T_{c}$. Therefore, close to the critical point, i.e., for small $\varepsilon$, the framework described above can be used to obtain the large deviation function $\phi(m,\dot{m})$ and thus the probability density $P_{N}(m,\dot{m})$ for large $N$. ## IV Continuous transition from a paramagnetic phase to an oscillating phase In this section, we briefly recall and extend results presented in [42] for the continuous transition observed when $J_{1}$ and $J_{2}$ are positive, at $T=T_{c}$, for $\mu>\mu_{c}$, from a high-$T$ paramagnetic phase to a low-$T$ oscillating phase [vertical green line in Fig. 1(a), (b)], corresponding to a Hopf bifurcation at the deterministic limit. Above $T_{c}$ ($\varepsilon<0$), the system is in a paramagnetic phase, whereas below $T_{c}$ ($\varepsilon>0$) it is in an oscillating phase. ### IV.1 Transition from a paramagnetic phase to an elliptic limit cycle #### IV.1.1 Large deviation function Figure 6: Large deviation function around the paramagnetic-oscillating transition. (a) Examples of $f(H)$ for $\varepsilon=10^{-3}$ (blue curve) and for $\varepsilon=-10^{-3}$ (orange curve). The inset represents the shape of $V(m)$. (b) Colormap of $\phi(m,\dot{m})$ in the space $(m,\dot{m})$ for $\varepsilon=10^{-3}$. Scalings with $\varepsilon$: $m\sim\varepsilon^{1/2}$, $\dot{m}\sim\varepsilon^{1/2}$, $H\sim\varepsilon$ and $f=\phi\sim\varepsilon^{2}$. Parameters: $J_{1}=0.6$, $J_{2}=0.4$, $(\mu-\mu_{c})/\mu_{c}=1$. As in the paramagnetic phase $m$ and $\dot{m}$ are small, we use a power- series expansion of $V(m)$ and $g(m,\dot{m})$ in $m$ and $\dot{m}$. At the lowest order required to describe the transition, one has $\displaystyle V(m)=\frac{\mu-\mu_{l}(T)}{2T^{2}}m^{2},$ (39) $\displaystyle g(m,\dot{m})=a_{0}\varepsilon-a_{1}m^{2}-a_{2}m\dot{m}-a_{3}\dot{m}^{2},$ (40) where $\varepsilon=(T_{c}-T)/T_{c}$; $\mu_{l}(T)$ and $a_{0}$ were introduced previously in Eq. (17) and their expressions are recalled in Appendix B along with the expressions of $a_{1}$ and $a_{2}$, which are all positive quantities. Compared to Eq. (16), we only keep the quadratic term in $V(m)$, which is positive around $T_{c}$ for $\mu>\mu_{c}=\mu_{l}(T_{c})$. Higher order terms are necessary only when the quadratic term is negative, in order to describe ferromagnetic order, or when the quadratic term is small, which is discussed in the next section. An illustration of the quadratic potential $V(m)$ is plotted in the inset of Fig. 6(a). In the deterministic limit, the oscillating phase appears for $\varepsilon>0$ when the paramagnetic point $(0,0)$ looses stability. The small perturbative parameter $u_{0}$ introduced in Sec. III [see Eq. (33)] is proportional to $\varepsilon$, since here $m_{0}=0$ and $u_{0}=g(0,0)=a_{0}\varepsilon$ from Eq. (40). Hence the formalism introduced in the previous section to obtain the large deviation function $\phi(m,\dot{m})=f(H)$ can be used to describe the phase transition for small $\varepsilon$, when the perturbative approach is valid. We find from Eq. (37), after integration, $f(H)=-\varepsilon aH+bH^{2},$ (41) with $a=\frac{T^{2}a_{0}}{T^{2}D_{22}+D_{11}[\mu-\mu_{l}(T)]}$ (42) and $b=\frac{a_{1}T^{4}+3T^{2}(\mu-\mu_{l})a_{3}}{4(\mu-\mu_{l})[T^{2}D_{22}+D_{11}(\mu-\mu_{l})]}.$ (43) When $\varepsilon<0$, $f(H)$ is minimal for $H=0$, which corresponds to the paramagnetic phase. When $\varepsilon>0$, $f(H)$ has a minimum in $H^{*}=\varepsilon a/2b$, see Fig. 6(a) for examples of $f(H)$ around the paramagnetic-oscillating transition. The equation $H(m,\dot{m})=H^{*}$ describes an ellipse in the phase space $(m,\dot{m})$ as depicted in Fig. 6(b). The period $\tau$ of a limit cycle described by $V(m)+\dot{m}^{2}/2=H^{*}$ is obtained as $\tau=2\int_{-m^{*}}^{m^{*}}\frac{dm}{\dot{m}}=\sqrt{2}\int_{-m^{*}}^{m^{*}}\frac{dm}{\sqrt{H^{*}-V(m)}}\,,$ (44) where $m^{*}$ is such that $H^{*}=V(m^{*})$. Using expression (39) of $V(m)$, we find $\tau=\frac{2\sqrt{2}\pi T}{\sqrt{\mu-\mu_{l}(T)}}.$ (45) #### IV.1.2 Order parameters The paramagnetic to oscillating phase transition is characterized by two order parameters, $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$, where $\langle x\rangle=\int dmd\dot{m}P_{N}(m,\dot{m})\,x(m,\dot{m})$, and where the observable $x$ stands for $m^{2}$ or $\dot{m}^{2}$ [or, more generally, any even function $x(m,\dot{m})$]. Using Eqs. (28) and (35), $P_{N}(m,\dot{m})$ can be approximated by its properly normalized large deviation form, $P_{N}(m,\dot{m})\approx\frac{\exp\big{[}-Nf\big{(}H(m,\dot{m})\big{)}\big{]}}{\int dm^{\prime}d\dot{m}^{\prime}\,\exp\big{[}-Nf\big{(}H(m^{\prime},\dot{m}^{\prime})\big{)}\big{]}}.$ (46) Then from Eq. (36), $\int d\dot{m}$ can be replaced by $\int dH/\sqrt{2[H-V(m)]}$, so that $\langle x\rangle$ becomes, making the integration intervals explicit: $\langle x\rangle=\frac{\int_{-1}^{1}dm\int_{V(m)}^{\infty}\frac{dH}{\sqrt{H-V(m)}}\,x\,e^{-Nf(H)}}{\int_{-1}^{1}dm\int_{V(m)}^{\infty}\frac{dH}{\sqrt{H-V(m)}}e^{-Nf(H)}}\,.$ (47) From Eqs. (39) and (41) one can then compute the values of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ using Eq. (47). In the limit of large system sizes, in the paramagnetic phase ($\varepsilon<0$), $\langle\dot{m}^{2}\rangle=1/(|\varepsilon|aN)$ which vanishes in the limit $N\to\infty$. By contrast, in the oscillating phase ($\varepsilon<0)$, $\langle\dot{m}^{2}\rangle=H^{*}\sim\varepsilon$ is constant in the limit $N\to\infty$. At the transition, $\varepsilon=0$, $\langle\dot{m}^{2}\rangle=1/\sqrt{\pi bN}$. For smaller system sizes $N\ll|\varepsilon|^{-2}$, $f(H)$ can be approximated as $f(H)=bH^{2}$ so that one finds $\langle\dot{m}^{2}\rangle\sim N^{-1/2},$ (48) which is independent of $\varepsilon$. ### IV.2 Non-elliptic limit cycle near the tricritical point #### IV.2.1 Large deviation function Figure 7: Large deviation function close to the tricritical point. (a) Example of $f(H)$ from Eq. (52); the inset represents the shape of $V(m)$. (b) Colormap of $\phi(m,\dot{m})$ in the space $(m,\dot{m})$. Note the different scalings $m\sim\varepsilon^{1/2}$ and $\dot{m}\sim\varepsilon$. Parameters: $J_{1}=0.6$, $J_{2}=0.4$, $\mu=\mu_{c}$ and $\varepsilon=10^{-3}$. The three phases (paramagnetic, oscillating and ferromagnetic) meet at the tricritical point $T=T_{c}$ and $\mu=\mu_{c}$. We now look at how the limit cycle changes when approaching the tricritical point at $\mu=\mu_{c}$ for $T<T_{c}$. We have $\mu_{c}-\mu_{l}(T)=T_{c}^{2}\varepsilon^{2}\,,$ (49) such that at $\mu=\mu_{c}$, the quadratic term $\propto(\mu_{c}-\mu_{l})m^{2}$ in the expression of $V(m)$ scales as $\varepsilon^{2}m^{2}$. Keeping only the quadratic term in $V(m)$, we would find $m\sim\varepsilon^{1/2}$ by following the same reasoning as in Sec. IV.1.1. Hence, the quadratic term becomes of order $\varepsilon^{3}$ whereas the next-order term $\propto m^{4}$ scales as $\varepsilon^{2}$, so that the assumption of neglecting the $m^{4}$-term in the expansion of $V(m)$ is inconsistent. To obtain the right behavior at $\mu=\mu_{c}$, it is thus necessary to include the contribution of $m^{4}$ in $V(m)$. At $\mu=\mu_{c}$, we have $V(m)=\frac{\varepsilon^{2}T_{c}^{2}}{2T^{2}}m^{2}+\frac{v_{1}(T,\mu)}{4}m^{4},$ (50) with $v_{1}(T,\mu)$ given in Appendix B. In the following, we assume that the scaling $m\sim\varepsilon^{1/2}$ remains valid close to $\mu_{c}$, and we check below that the assumption is consistent. [Note that the reason why the scaling $m\sim\varepsilon^{1/2}$ will eventually prove valid is different from the one mentioned in the previous paragraph, which relies only on the quadratic term in $V(m)$]. Under this assumption, the term in $\varepsilon^{2}m^{2}\sim\varepsilon^{3}$ in Eq. (50) is negligible compared to the term of order $m^{4}$. Hence, to leading order, we can use the simple form $V(m)=\frac{v_{1}(T,\mu)}{4}m^{4}.$ (51) We also consider that $v_{1}(T_{c},\mu_{c})>0$ as discussed in Sec. II.2 (see Fig. 2 for possible values of $J_{1}$, $J_{2}$) and that $v_{1}(T,\mu_{c})$ remains positive for small $\varepsilon$. An example of $V(m)$ is plotted in the inset of Fig. 7(a). Using the same expression of $g(m,\dot{m})$ as before, Eq. (40), we obtain from Eq. (37) after integration, $f(H)=-\frac{\varepsilon a_{0}}{D_{22}}H+cH^{3/2}$ (52) with $c=\frac{8\,\Gamma\left(\frac{3}{4}\right)^{4}a_{1}}{5\pi^{2}D_{22}\sqrt{v_{1}}}\,$ (53) where $\Gamma(x)=\int_{0}^{\infty}t^{x-1}e^{-t}dt$. One finds that $f(H)$ has a minimum for $H^{*}=\left(\frac{2\varepsilon a_{0}}{3cD_{22}}\right)^{2}$ (54) [see Fig. 7(a)] corresponding to a limit cycle in the phase space $(m,\dot{m})$. However, $H$ is no longer quadratic in $m$, because of the quartic form (51) of $V(m)$. Hence the limit cycle is no longer elliptic, see Fig. 7(b) for a colormap of $\phi(m,\dot{m})$ in the phase space $(m,\dot{m})$. We recall that for $\mu\gg\mu_{c}$, one had $m\sim\dot{m}\sim\varepsilon^{1/2}$, whereas now one finds distinct scalings $m\sim\varepsilon^{1/2}$ and $\dot{m}\sim\varepsilon$. These scalings are obtained by using $H^{*}\sim\varepsilon^{2}$ from Eq. (54), and the expression (36) of $H(m,\dot{m})$ together with the quartic form (51) of $V(m)$. Finally, evaluating the oscillation period using Eq. (44), we find $\tau=\frac{4\sqrt{\pi}\Gamma\\!\left(\frac{5}{4}\right)}{\Gamma\left(\frac{3}{4}\right)(v_{1}H^{*})^{1/4}}\,.$ (55) As $H^{*}\sim\varepsilon^{2}$, the period diverges as $\varepsilon^{-1/2}$ when $\varepsilon\to 0$. Figure 8: (a) $P_{N}(m,\dot{m})$ obtained from stochastic numerical simulations. (b) Theoretical $P_{N}(m,\dot{m})$ evaluated by including leading-order corrections, as given in Eq. (60). Parameters: $\varepsilon=10^{-2}$, $\mu=\mu_{l}(T)$, $N=10^{7}$. ### IV.3 Comparison with stochastic simulations and need for higher order corrections The method used to obtain analytically the large deviation function, developed in Sec. III, relies on two main assumptions: $N$ is large and $\varepsilon$ is small. We now compare the analytical results with numerical simulations of the stochastic spin model. We use the Gillespie algorithm [52] to simulate the stochastic dynamics with the rates given by Eq. (26) for a time-interval $\tau$. The initial condition for the simulations are $m=0$ and $h=0$. In this algorithm, time-steps are of $O(N^{-1})$ such that the number of steps required to have $t=1$ is of $O(N)$. To observe an non-elliptic limit cycle close to $T_{c}$, one needs to have $N|f(H^{*})|\gg 1$ which corresponds to $N\varepsilon^{3}\gg 1$ [see Eqs. (52) and Eq. (54)]. For example, a value $\varepsilon=10^{-3}$ would require simulations with at least $N\sim 10^{11}$. To obtain data with converged statistics depicting the transition, we make simulations for larger $\varepsilon$ where the approximations made in Sec III are no longer expected to be quantitatively valid. We now discuss the notable differences observed in numerical simulations due to larger $\varepsilon$ values and smaller system sizes $N$. In Fig. 8, we plot $P_{N}(m,\dot{m})$ obtained from stochastic simulations for $(T_{c}-T)/T_{c}=\varepsilon=10^{-2}$ and for $\mu=\mu_{l}(T)$. We take $\mu=\mu_{l}(T)$ (instead of $\mu=\mu_{c}$) so that the term in front of $m^{2}$ in $V(m)$ is exactly zero. Significant discrepancies are observed between the simulation results and the theoretical predictions of the perturbative approach described in Sec. III. From leading order calculations in $\varepsilon$, we obtain that $\phi(m,\dot{m})=f(H)$ with $H=V(m)+\dot{m}^{2}/2$ which has, in particular, two consequences. First, $m$ and $\dot{m}$ are decoupled, and the symmetries $m\,\mapsto\,-m$ and $\dot{m}\,\mapsto\,-\dot{m}$ hold independently. Second, the probability $P_{N}(m,\dot{m})$ is uniform along the limit cycle, corresponding in a dynamical view (and in the deterministic limit) to a constant ‘speed’ along the limit cycle. Both of these characteristics are not observed in the stochastic simulations, see Fig. 8(a). These differences come from higher order corrections, in $\varepsilon$ and in $N$ in the probability density $P_{N}(m,\dot{m})$. Similar corrections were studied in [49] in the context of noisy dynamical systems. We now give an example of the first corrections for $\mu=\mu_{l}(T)$. The detailed steps of the derivation are given in [53]. To perform a systematic $\varepsilon$-expansion, we introduce rescaled variables $\tilde{H}=H/H^{*}$ and $x=m/m_{0}$ with $H^{*}$ given in Eq. (54) and $m_{0}=(4H^{*}/v_{1})^{1/4}$ consistently with Eq. (19). At lowest order in $\varepsilon$, from Eq. (52), one finds $\phi(m,\dot{m})\sim\varepsilon^{3}\tilde{\phi}_{1}(x,\tilde{H})$, where $\tilde{\phi}_{1}$ is a rescaled function independent of $\varepsilon$. We introduce $C(m,\dot{m})$ the correction of $O(N^{0})$ to $\ln P_{N}(m,\dot{m})$, $P_{N}(m,\dot{m})\propto\exp[-N\phi(m,\dot{m})+C(m,\dot{m})],$ (56) and we expand $\phi(m,\dot{m})$ and $C(m,\dot{m})$ in power series of $\varepsilon^{1/2}$ [53], $\phi(m,\dot{m})=\varepsilon^{3}\sum_{i=1}^{\infty}\varepsilon^{(i-1)/2}\tilde{\phi}_{i}(x,\tilde{H})$ (57) and $C(m,\dot{m})=\sum_{i=0}^{\infty}\varepsilon^{i/2}\tilde{C}_{i}(x,\tilde{H}).$ (58) Injecting these expressions into the master equation on $P_{N}(m,\dot{m})$, one finds equations on $\tilde{\phi}_{i}$ and $\tilde{C}_{i}$ at each order $i$ [53]. For the lowest order, we find $\phi_{1}(m,\dot{m})=\varepsilon^{3}\tilde{\phi}_{1}(x,\tilde{H})=f(H)$ [Eq. (52)] and $C_{0}(m,\dot{m})=\varepsilon^{1/2}\tilde{C}_{0}(x,\tilde{H})=\varepsilon^{1/2}c_{0}$ a constant given by the normalization of $P_{N}(m,\dot{m})$. For $\tilde{\phi}_{2}(x,\tilde{H})$ and $\tilde{C}_{1}(x,\tilde{H})$ one finds: $\displaystyle\tilde{\phi}_{2}(x,\tilde{H})$ $\displaystyle=A\left(1-\sqrt{\tilde{H}}\right)\tilde{H}x\left[a_{0}\,{}_{2}F_{1}\left(-\frac{1}{2},\frac{1}{4},\frac{5}{4},x^{4}\right)-\frac{2}{3}\frac{D_{22}}{\alpha}x^{2}\,{}_{2}F_{1}\left(-\frac{1}{2},\frac{3}{4},\frac{7}{4},x^{4}\right)\right]\,,$ (59) $\displaystyle\tilde{C}_{1}(x,\tilde{H})$ $\displaystyle=B\tilde{H}^{1/4}\left[-6a_{0}x\sqrt{1-x^{4}}+30a_{0}\sqrt{v_{1}}x\,{}_{2}F_{1}\left(-\frac{1}{2},\frac{1}{4},\frac{5}{4},x^{4}\right)-8\frac{D_{22}}{\alpha}x^{3}{}_{2}F_{1}\left(\frac{1}{2},\frac{5}{4},\frac{7}{4},x^{4}\right)\right]\,,$ where $\alpha$, $A$ and $B$ depend on $v_{1}$, $a_{0}$ and $a_{1}$ and are given in Appendix B, and ${}_{2}F_{1}(a,b,c,x)$ denotes the hypergeometric function. The correction $\tilde{\phi}_{2}$ changes the orientation and the shape of the limit cycle, whereas the correction $\tilde{C}_{1}$ breaks the uniformity of the probability $P_{N}(m,\dot{m})$ along the limit cycle. We plot in Fig. 8 the following expression of $P_{N}(m,\dot{m})$ that includes leading corrections, $P_{N}(m,\dot{m})=\exp[-N(\phi_{1}+\phi_{2})+C_{0}+C_{1}]$ (60) with the definitions $\phi_{i}(m,\dot{m})=\varepsilon^{3+(i-1)/2}\tilde{\phi_{i}}(x,\tilde{H})$ and $C_{i}(m,\dot{m})=\varepsilon^{i/2}\tilde{C}_{i}(x,\tilde{H})$. The main features of the probability density obtained from the simulations are captured by these leading corrections. ## V Type-I discontinuous transition between ferromagnetic and oscillating phases In this section, we investigate the properties, near the tricritical point $(T_{c},\mu_{c})$, of the ferromagnetic to oscillating phase transition where a limit cycle appears around the ferromagnetic points, called coexistence of Type I. This case corresponds to $v_{1}(T_{c},\mu_{c})>0$, see Fig. 2 as well as the phase diagram of Fig. 1(a) and the trajectories displayed in Figs. 1(c) and 1(e). ### V.1 Large deviation function and phase diagram #### V.1.1 Validity of the perturbative approach We start from the generic expansion of the potential $V(m)$ given in Eq. (16), recalled here for clarity, $V(m)=\frac{\mu-\mu_{l}(T)}{2T^{2}}m^{2}+\frac{v_{1}(T,\mu)}{4}m^{4}+V_{0},$ (61) where $V_{0}$ is chosen such that $V(m)\geq 0$ and its minimal value is zero. An example of $V(m)$ for $\mu<\mu_{l}(T)$ is given in Fig. 9. We recall that $g(m,\dot{m})$ is given in Eq. (40). As discussed in Sec. II.3, ferromagnetic points $m_{0}^{2}=(\mu_{l}(T)-\mu)/T^{2}v_{1}$ exist for $\mu<\mu_{l}(T)$, and are locally stable for $\mu\leq\mu_{F}$ with $\mu_{F}(T)=\mu_{l}(T)-\varepsilon a_{0}T^{2}v_{1}/a_{1}$. In this section, we focus on the region where the ferromagnetic points loose stability ($\mu\approx\mu_{F}$), thus: $\mu_{l}(T)-\mu\sim\varepsilon\,$ (62) and one has $m_{0}^{2}\sim\varepsilon$ and thus $u_{0}=g(m_{0},0)\sim\varepsilon$. For small $\varepsilon=(T_{c}-T)/T_{c}$, the main assumption made in Sec. III, i.e., that $u_{0}$ is small, is verified, and we can use the method developed in this section to obtain the large deviation function and study the phase transition from a ferromagnetic phase to an oscillating phase. #### V.1.2 Typical $f(H)$ and phase diagram Figure 9: Large deviation function for Type-I discontinuous transition between ferromagnetic and oscillating phases. (a) $f(H)$ determined numerically from Eq. (37) (full line); the dashed line corresponds to local approximations described in Sec. V.1.3. Inset: shape of $V(m)$ with two minima. (b) Colormap of $\phi(m,\dot{m})$ in the plane $(m,\dot{m})$. Note the different scalings $m\sim\varepsilon^{1/2}$ and $\dot{m}\sim\varepsilon$. Parameters: $J_{1}=0.6$, $J_{2}=0.4$, $(\mu-\mu_{c})/\mu_{c}=-3.18\times 10^{-5}$ and $\varepsilon=10^{-3}$. In general, except for particular cases as the one described in the previous section, one cannot obtain explicit analytical expressions of $f^{\prime}(H)$ from Eq. (37), and one needs to perform a numerical integration of the integrals in Eq. (37) to determine $f(H)$. An example of $f(H)$, for $\mu\leq\mu_{l}(T)$, numerically obtained from Eq. (37), is plotted in Fig. 9(a). We observe that $f(H)$ has two local minima: one in $H=0$ corresponding to the ferromagnetic points $m=m_{0}$ and $\dot{m}=0$ [since $V(m_{0})=0$], and one for $H=H^{*}>0$ corresponding to a limit cycle in the phase space $(m,\dot{m})$. We numerically obtain (not shown) that $H^{*}\sim\varepsilon^{2}.$ (63) An example of colormap of $\phi(m,\dot{m})=f\big{(}H(m,\dot{m})\big{)}$ in the phase space $(m,\dot{m})$ is displayed in Fig. 9(b). Here, the most stable phase is the oscillating phase as $f(0)>f(H^{*})$. Contrary to Sec. IV, no analytical expression of $H^{*}$ is available in the present case. The transition from the ferromagnetic phase to the oscillating phase takes place when $f(0)=f(H^{*})$; we note $\mu_{t}(T)$ the value of $\mu$ at the transition. The value of $H$ jumps from $H=0$ to the nonzero value $H^{*}$ at the transition, meaning that the latter is discontinuous. We obtain numerically that $(\mu_{c}-\mu_{t})/\mu_{c}\sim\varepsilon$ where $\varepsilon=(T_{c}-T)/T_{c}$ [see Fig. 10 with $(\mu_{c}-\mu_{F})/\mu_{c}\sim\varepsilon$]. From the numerical determination of $f(H)$, one obtains a phase diagram in the space $(\varepsilon,\mu)$ with the determination of the different phases: ferromagnetic phase (F), oscillating phase (O) or the phase where both coexist, with one being more stable than the other. A close up on the phase diagram near the tricritical point for $J_{1}=0.6$ and $J_{2}=0.4$ is plotted in Fig. 10, where we represent $(\mu-\mu_{F})/\mu_{F}$, with $\mu_{F}$ given in Eq. (20), in order to visualize the different phases. The limits of existence of the ferromagnetic and oscillating states can also be obtained in the deterministic limit, but for the determination of $\mu_{t}$ (which characterizes the most stable phase) it is necessary to consider finite system sizes using the large deviation approach. Figure 10: Close up of the phase diagram of Fig. 1(a) near the tricritical point, for $J_{1}=0.6$ and $J_{2}=0.4$, in the reduced parameters $\varepsilon=(T_{c}-T)/T_{c}$ and $(\mu-\mu_{F}(T))/\mu_{F}(T)$, where $\mu_{F}$ is defined in Eq. (20) and $(\mu_{c}-\mu_{F}(T))/\mu_{c}\sim\varepsilon$. O corresponds to the oscillating phase, F to the ferromagnetic phase. In the hatched area, both phases are locally stable. The most stable phase, given by the global minima of $f(H)$, is either the ferromagnetic phase [(O)+F] or the oscillating phase [O+(F)], separated by the transition line $\mu_{t}$. $\mu_{l}(T)$ corresponds to the limit of existence of the ferromagnetic points in the deterministic limit, $\mu_{F}$ to their linear stability limit, and $\mu_{O}$ to the limit of existence of the oscillating state at deterministic level. #### V.1.3 Local analytical expressions of $f(H)$ In most cases, keeping only the first orders of the series expansions of $V(m)$ and $g(m,\dot{m})$ is not enough to obtain an analytical expression of $f^{\prime}(H)$. Still, local approximations can be obtained. Near a minimum $m_{0}$ of $V(m)$, a quadratic expansion of $V$ gives $f(H)=f_{F}(H)\equiv-\frac{g(m_{0},0)}{D_{22}+2v_{1}m_{0}^{2}D_{11}}H,$ (64) where $f_{F}(H)$ stands for the local approximate expression of $f(H)$ in the ferromagnetic state. We recover that the point $(m,\dot{m})=(m_{0},0)$ is stable when $g(m_{0},0)<0$. This expression of $f(H)$ is valid for small $H$ only. For $H\gg H^{*}$, we recover that $f(H)\approx cH^{3/2}$, with $c$ defined in Eq. (53). For intermediate values of $H$ ($H\sim H^{*}$), we do not have an analytical expression of $f(H)$. However the regime $H\gg H^{*}$ is similar to the one obtained for $\mu=\mu_{c}$, which suggests that the form of $f(H)$ obtained for $\mu\approx\mu_{c}$ in Eq. (52) remains approximately valid up to a redefinition of coefficient values. One can perform a local fit of the form $f(H)=f_{O}(H)+f(H^{*})$ with $f_{O}(H)=\tilde{c}\left(H^{3/2}-\frac{3}{2}\sqrt{H^{*}}H+\frac{1}{2}H^{*3/2}\right),$ (65) where the parameters $\tilde{c}$ and $H^{*}$ are fitted on the numerically evaluated $f(H)$ to get a local approximation of $f(H)$ near $H^{*}$ (see Fig. 9 for an example of a fit of $f(H)$ close to its minimum). The functional form (65) provides a reasonable description of the large $H$ behavior of $f(H)$, and is more accurate than a simple parabolic fit around the minimum $H=H^{*}$. ### V.2 Scalings of order parameters with system size at the transition #### V.2.1 Large-$N$ scaling at $\mu=\mu_{t}$ Using the two local approximations of $f(H)$ given in Eqs. (64) and (65), we study the behaviors of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ in the large-$N$ limit when approaching the critical point $\varepsilon=0$ where the three phases meet. In the ferromagnetic phase ($\mu<\mu_{t}$), using Eq. (64) and $g(m_{0},0)=a_{0}\varepsilon-a_{1}m_{0}^{2}$, one finds in the large-$N$ limit, $\displaystyle\langle m^{2}\rangle$ $\displaystyle=m_{0}^{2}=\frac{\mu_{l}(T)-\mu}{T^{2}v_{1}},$ (66) $\displaystyle\langle\dot{m}^{2}\rangle$ $\displaystyle=\frac{D_{22}+2v_{1}m_{0}^{2}D_{11}}{(a_{1}m_{0}^{2}-a_{0}\varepsilon)N}.$ (67) We recover the results of the deterministic limit for $N\to\infty$, $\langle m^{2}\rangle=m_{0}^{2}\sim\varepsilon$ and $\langle\dot{m}^{2}\rangle=0$. For large but finite $N$, we obtain that $\langle\dot{m}^{2}\rangle\sim\varepsilon^{-1}N^{-1}$. In the oscillating phase ($\mu>\mu_{t}$), $f(H)$ is minimal in $H^{*}$, so that for large enough $N$ we can replace $e^{-Nf(H)}$ by $\delta(H-H^{*})$ in Eq. (47), yielding $\displaystyle\langle m^{2}\rangle$ $\displaystyle=\frac{\int_{-m^{*}}^{m^{*}}dm\frac{m^{2}}{\sqrt{H^{*}-V(m)}}}{\int_{-m^{*}}^{m^{*}}dm\frac{1}{\sqrt{H^{*}-V(m)}}},$ (68) $\displaystyle\langle\dot{m}^{2}\rangle$ $\displaystyle=\frac{\int_{-m^{*}}^{m^{*}}dm\sqrt{2(H^{*}-V(m))}}{\int_{-m^{*}}^{m^{*}}dm\frac{1}{\sqrt{2(H^{*}-V(m))}}},$ (69) where $m^{*}$ is such that $H^{*}=V(m^{*})$. Both $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ reach constant values at large $N$. We obtain numerically that $\langle m^{2}\rangle\sim\varepsilon$ and $\langle\dot{m}^{2}\rangle\sim\varepsilon^{2}$, which are the same scalings as the one observed for the nonelliptic limit cycle for $\mu=\mu_{c}$ (see Sec. IV.2). #### V.2.2 Moderate-$N$ scaling at $\mu=\mu_{t}$ Figure 11: Order parameters $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ as a function of system size $N$, for $\varepsilon\in[10^{-3},10^{-4},7\times 10^{-6},6\times 10^{-7},4.6\times 10^{-8},3.6\times 10^{-9}$] at $\mu=\mu_{t}(\varepsilon)$. Order parameters are evaluated numerically from Eqs. (37) and (47). (a) $\langle m^{2}\rangle$ vs $N$. (b) $\langle\dot{m}^{2}\rangle$ vs $N$. The red dashed line corresponds to the scaling predictions for moderate values of $N$ from Eqs. (70) and (71), with $\langle m^{2}\rangle\sim N^{-1/3}$ and $\langle\dot{m}^{2}\rangle\sim N^{-2/3}$. (c) $\langle m^{2}\rangle/\varepsilon$ and (d) $\langle\dot{m}^{2}\rangle\varepsilon^{2}$ vs the rescaled system size $N/\varepsilon^{-3}$, showing data collapse. Parameters: $J_{1}=0.6$, $J_{2}=0.3$. Unlike for large values of $N$, for intermediate values of $N$, $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ are found not to depend much on $\varepsilon$ and on $\mu$. In Figs. 11(a) and 11(b), we plot $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ for different $\varepsilon$ at the transition, $\mu=\mu_{t}(\varepsilon)$. We observe that $\langle m^{2}\rangle\sim N^{-1/3}$ and $\langle\dot{m}^{2}\rangle\sim N^{-2/3}$ for moderate $N$ values. These scaling behaviors can be understood as follows. The expression of the average $\langle x\rangle$ contains an integral over $H$ with the factor $\exp[-Nf(H)]$, see Eq. (47). For moderate $N$, the integral is dominated by the $H^{3/2}$-term in $f(H)$. This can be justified by performing the change of variable $z=NH^{3/2}$ in the integral. One then finds that higher order powers of $H$ in $f(H)$ are negligible when $N\gg 1$, while linear contributions in $H$ are also negligible as long as $\sqrt{H^{*}}N^{1/3}\ll 1$, i.e., $N\ll(H^{*})^{-3/2}$, where $H^{*}$ is small for small $\varepsilon$ [see Eq. (63)]. The integral is then dominated by the contribution of the region $z\sim 1$, i.e., which corresponds to $H\gg H^{*}$. In addition, Eq. (47) also contains an integral over $m$. Due to the factor $[H-V(m)]^{-1/2}$ in the integrals, values of $m$ which contribute the most are where $V(m)\approx v_{1}m^{4}/4$. In a similar way as above, this can be justified by performing the change of variable $z^{\prime}=m/H^{1/4}$ in the integral with the expression of $V(m)$ given in Eq. (61). One then finds that higher order powers of $m$ are negligible when $H\ll 1$ and the quadratic order is negligible when $v_{0}(\mu-\mu_{l})\ll H^{1/4}$ which is verified for $H\gg H^{*}$ as $\mu-\mu_{l}\sim\varepsilon$ [Eq. (62)] and $H^{*}\sim\varepsilon^{2}$ [Eq. (63)]. Using these two approximations on $f(H)$ and $V(m)$, we find: $\langle m^{2}\rangle=\frac{4\Gamma\left(\frac{5}{6}\right)\Gamma\left(\frac{3}{4}\right)^{4}}{\pi^{5/2}\sqrt{v_{1}}c^{1/3}}N^{-1/3},$ (70) and $\langle\dot{m}^{2}\rangle=\frac{4\Gamma\left(\frac{7}{6}\right)}{3\sqrt{\pi}c^{2/3}}N^{-2/3},$ (71) where $c$ is given in Eq. (53). We plot these quantities in red in Figs. 11(a) and 11(b) alongside the numerical values obtained from Eq. (47) using the numerical evaluation of $f(H)$ from Eq. (37). We note $N^{*}$ the crossover value of $N$ between the moderate- and large-$N$ regimes. The crossover takes place when the value of $\langle m^{2}\rangle\sim N^{-1/3}$ in the moderate-$N$ approximation is comparable to the one in the large-$N$ approximation, $\langle m^{2}\rangle\sim\varepsilon$. One thus finds that $N^{*}$ behaves as $N^{*}\sim\varepsilon^{-3}.$ (72) [Note that, according to the integration argument above, $N^{*}\sim(H^{*})^{-3/2}$, implying $H^{*}\sim\varepsilon^{2}$, consistently with Eq. (63)]. A similar argument for $\dot{m}$ yields the same scaling for $N^{*}$: in the moderate-$N$ regime ($N\ll N^{*}$), $\langle\dot{m}^{2}\rangle\sim N^{-2/3}$, while in the large-$N$ approximation ($N\gg N^{*}$), one has $\langle\dot{m}^{2}\rangle\sim\varepsilon^{2}$ in the oscillating phase and $\langle\dot{m}^{2}\rangle\sim\varepsilon^{-1}N^{-1}$ in the ferromagnetic phase, which both give a crossover at $N^{*}\sim\varepsilon^{-3}$. Focusing on the oscillating phase, these scaling behaviors of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ can be encompassed into two scaling functions $\langle m^{2}\rangle=\varepsilon\,\mathcal{F}_{m}(\varepsilon^{3}N),\quad\langle\dot{m}^{2}\rangle=\varepsilon^{2}\,\mathcal{F}_{\dot{m}}(\varepsilon^{3}N),$ (73) with asymptotic behaviors $\mathcal{F}_{m}(x)\sim x^{-1/3}$ and $\mathcal{F}_{\dot{m}}(x)\sim x^{-2/3}$ for $x\to 0$, while both functions go to constant values for $x\to\infty$. Figs. 11(c) and 11(d) show the data collapse obtained by plotting the rescaled variables $\langle m^{2}\rangle/\varepsilon$ and $\langle\dot{m}^{2}\rangle/\varepsilon^{2}$ versus the rescaled system size $N/\varepsilon^{-3}$, for different values of $\varepsilon$ at $\mu=\mu_{t}(\varepsilon)$. ### V.3 Detailed study of the crossover regime We reported above two distinct scaling regimes $N\ll N^{*}$ and $N\gg N^{*}$ of the observables $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ as a function of system size $N$ for $\mu=\mu_{t}(\varepsilon)$, and we identified the scaling with $\varepsilon$ of the crossover size $N^{*}$. We now investigate in more details the behavior of these observables in the crossover regime $N\sim N^{*}$, now focusing on the effect of the variations of $\mu$ close to the transition value $\mu_{t}$, for a fixed $\varepsilon$. We find in particular that in the crossover regime, $\langle m^{2}\rangle$ has a non- monotonic behavior as a function of $N$, whose details significantly depend on $\mu$. The behavior of $\langle\dot{m}^{2}\rangle$, while monotonic as a function of $N$, is found to strongly depend on $\mu$. #### V.3.1 Influence of $\mu$ on the crossover regime Figure 12: (a) $\langle m^{2}\rangle$ and (b) $\langle\dot{m}^{2}\rangle$ as a function of system size $N$, for different values of $\mu$ close to $\mu_{c}$, corresponding to $(\mu_{c}-\mu)/\mu_{c}\in[4.0,3.6,3.3,3.1,2.5]\times 10^{-5}$ from darker to lighter colors. The transition between ferromagnetic and oscillating states takes place at $\mu_{t}$ given by $(\mu_{c}-\mu_{t})/\mu_{c}\approx 3.2\times 10^{-5}$. Parameters: $J_{1}=0.6$, $J_{2}=0.4$, and $\varepsilon=10^{-3}$. Order parameters are evaluated numerically in the same way as in Fig. 11. As mentioned above, the observables $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ are seen to have a very weak dependence on $\varepsilon$ and $\mu$ in the moderate-$N$ regime. For large $N$, the value of $\langle\dot{m}^{2}\rangle$ is found to be significantly different in the oscillating phase $(\mu>\mu_{t})$ where $\langle\dot{m}^{2}\rangle\sim\varepsilon^{2}$ and in the ferromagnetic phase $(\mu<\mu_{t})$ where $\langle\dot{m}^{2}\rangle\sim\varepsilon^{-1}N^{-1}$. In contrast, the value of $\langle m^{2}\rangle$ is similar in both phases, with $\langle m^{2}\rangle\sim\varepsilon$. In Fig. 12, $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ are plotted as a function of system size $N$ in the crossover regime $N\sim N^{*}$, for different values of $\mu$ across the transition, keeping $\varepsilon$ fixed. After an initial decay for $N\ll N^{*}$, we observe that $\langle m^{2}\rangle$ slightly increases before reaching a constant value, as expected in the limit $N\to\infty$. For some values of $\mu$, like for $(\mu-\mu_{c})/\mu_{c}=-3.3\times 10^{-5}$, a second decay is observed before reaching the asymptotic constant value. The behavior of $\langle\dot{m}^{2}\rangle$ is significantly different from that of $\langle m^{2}\rangle$ as the large-$N$ limit yields two different behaviors in the ferromagnetic or in the oscillating phase. We observe that for $\mu<\mu_{t}$, $\langle\dot{m}^{2}\rangle$ first reaches a plateau for a significant range of $N$, before steeply decreasing to eventually reach the large-$N$ scaling $\langle\dot{m}^{2}\rangle\sim N^{-1}$. #### V.3.2 Interpretation as a finite-size phase coexistence The observed non-trivial behaviors can be given a simple interpretation in terms of finite-size phase coexistence and metastability. For a finite-size system, a metastable state has a finite probability to be visited, and this probability decreases exponentially with system size. Based on this idea, we introduce a simple decomposition of average values into contributions of each phase, and show that such a decomposition is sufficient to account for most of the observed behaviors. Figure 13: (a) $\langle m^{2}\rangle$ and (b) $\langle\dot{m}^{2}\rangle$ vs system size $N$. In black, data from Fig. 12 for $(\mu_{c}-\mu)/\mu_{c}=3.3\times 10^{-5}$. The dashed lines correspond to $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ computed for $f=f_{H}$ [Eq. (64)] (orange) and for $f=f_{O}$ [Eq. (65)] (blue). The plain red lines correspond to $\langle x(m,H)\rangle_{\mathrm{approx}}$ defined in Eq. (74). Here, $f(H^{*})=1.1\times 10^{-12}$. Starting from the expression of an average observable given in Eq. (47), we split the semi-axis $H\geq 0$ into two regions, separated by the value $H_{0}$ corresponding to the local maximum of $f(H)$. For small $H$, i.e., near the ferromagnetic points, $f(H)$ is linear [see Eq. (64)]. For $H$ around $H^{*}$, we write $f(H)=f_{O}(H)+f(H^{*})$ with $f_{O}$ from Eq. (65) where $\tilde{c}$, $H^{*}$ and $f(H^{*})$ are parameters fitted on the numerically evaluated $f(H)$. We consider $N$ large enough so that the integration interval can be extended to the entire real axis due to the rapidly decaying factor $\exp[-Nf(H)]$. Then, for any quantity $x(m,H)$, we use the approximate expression of the average value $\langle x(m,H)\rangle$, $\langle x(m,H)\rangle_{\mathrm{approx}}=\frac{\langle x\rangle_{F}+\langle x\rangle_{O}C_{1}\sqrt{N}e^{-Nf(H^{*})}}{1+C_{1}\sqrt{N}e^{-Nf(H^{*})}}$ (74) with $C_{1}=\frac{1}{\sqrt{N}}\frac{\int_{-1}^{1}dm\int_{V(m)}^{\infty}dH\frac{1}{\sqrt{H-V(m)}}e^{-Nf_{O}(H)}}{\int_{-1}^{1}dm\int_{V(m)}^{\infty}dH\frac{1}{\sqrt{H-V(m)}}e^{-Nf_{F}(H)}}.$ (75) In Eq. (74), $\langle x\rangle_{F}$ (resp. $\langle x\rangle_{O}$) corresponds to the ‘pure-state’ average computed in the ferromagnetic state with $f(H)=f_{F}(H)$ defined in Eq. (64) [resp. $f(H)=f_{O}(H)$ in the oscillating state, see Eq. (65)]. We obtain $C_{1}=-\frac{4g(m_{0},0)m_{0}\sqrt{v_{1}}H^{*1/4}}{\sqrt{3\pi\tilde{c}}(D_{22}+2v_{1}m_{0}^{2}D_{11})}\int_{m*}^{1}\frac{dm}{\sqrt{H^{*}-V(m)}},$ (76) with $m^{*}$ such that $V(m^{*})=H^{*}$. The oscillating phase has a contribution weighted with the factor $C_{1}\sqrt{N}e^{-Nf(H^{*})}/(1+C_{1}\sqrt{N}e^{-Nf(H^{*})})$. When $f(H^{*})>0$ (i.e., the ferromagnetic phase is the most stable one), the contribution of the oscillating phase disappears at large $N$ but is non- negligible for $N\sim f(H^{*})$. In Fig. 13, the ‘pure-state’ averages $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ evaluated using either $f=f_{H}$ for the ferromagnetic state, or $f=f_{O}$ for the oscillating state, as well as the ‘mixed-state’ approximation Eq. (74) are compared to the values of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ obtained from the numerically evaluated $f(H)$ (same data as on Figs. 11 and 12). $\langle m^{2}\rangle$ increases due to the influence of the oscillating phase as $\langle m^{2}\rangle$ is higher for $f=f_{O}$ than for $f=f_{H}$; then for larger $N$, it decreases to its expected value in the ferromagnetic state. When $f(H^{*})<0$ (i.e., the oscillating phase is the most stable one), we observe a monotonous increase between the moderate-$N$ decay the asymptotic large-$N$ value (see Fig. 12). The influence of the oscillating phase is even more pronounced for $\langle\dot{m}^{2}\rangle$, as we observe that for moderate $N$, $\langle\dot{m}^{2}\rangle$ is almost constant and equal to the value expected in the oscillating state (obtained using $f=f_{O}$), before eventually steeply decreasing when $Nf(H^{*})\sim 1$. #### V.3.3 Comparison with stochastic simulations Figure 14: Comparison with stochastic simulations of the spin model. (a) $\langle m^{2}\rangle$ and (b) $\langle\dot{m}^{2}\rangle$ obtained from stochastic numerical simulations, as a function of $N$, for $(\mu_{c}-\mu)/\mu_{c}\in[4.44,4.04,3.64,3.24]\times 10^{-3}$ (from darker to lighter colors). Parameters: $J_{1}=0.6$, $J_{2}=0.4$, and $\varepsilon=10^{-1}$. In Fig. 14, $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ obtained from stochastic simulations of the spin model are plotted for different system sizes $N$, for $\varepsilon=(T_{c}-T)/T_{c}=10^{-1}$. Qualitatively, the moderate- and large-$N$ regimes are visible on the data. However, the decay of $\langle m^{2}\rangle$ in the moderate-$N$ regime is significantly slower than the theoretically predicted behavior $\langle m^{2}\rangle\sim N^{-1/3}$. This is most likely due to the fact that $\varepsilon$ is not small enough to enter the asymptotic low-$\varepsilon$ regime, as discussed below. The moderate-$N$ decay of $\langle\dot{m}^{2}\rangle$ seems better described by the theoretical prediction $\langle\dot{m}^{2}\rangle\sim N^{-2/3}$, although significant deviations are also visible. In the large-$N$ regime, $\langle\dot{m}^{2}\rangle$ reaches a constant value $\sim\varepsilon^{2}$ in the oscillating phase ($\mu>\mu_{t}$), or decreases as $\varepsilon^{-1}N^{-1}$ in the ferromagnetic phase ($\mu<\mu_{t}$). The transition between the moderate- and large-$N$ regimes takes place around $N^{*}\approx 10^{3}\approx\varepsilon^{-3}$, as expected. To understand the discrepancies found between stochastic simulations data and theoretical predictions, we note that the main approximation made to obtain the power laws $\langle m^{2}\rangle\sim N^{-1/3}$ and $\langle\dot{m}^{2}\rangle\sim N^{-2/3}$ is the assumption $\sqrt{H^{*}}N^{1/3}\ll 1$ (see Sec. V.2.2), which is not valid here for moderate $N$ values. Moreover, we showed in Sec. IV.3 that there are discrepancies between the theory and the simulations for low values of $\varepsilon$ when $N$ is not large enough. We discuss this issue in more details in the next subsection. #### V.3.4 Discussion on the low-$\varepsilon$ and large-$N$ approximations Figure 15: (a) $P_{N}(m,\dot{m})$ obtained from stochastic simulations of the spin model for $N=5\times 10^{5}$. (b) Trajectories ($m(t),\dot{m}(t))$ in the phase space $(m,\dot{m})$ obtained in the deterministic limit. The color corresponds to $v(m,\dot{m})^{-1}$, where $v(m,\dot{m})$ is the local speed along the limit cycle, and can be interpreted as the local density $p(m,\dot{m})\propto v(m,\dot{m})^{-1}$. The two black dots are added for visual purpose and correspond to the ferromagnetic points. Parameters: $J_{1}=0.6$, $J_{2}=0.4$, $\varepsilon=10^{-1}$ and $(\mu_{c}-\mu)/\mu_{c}=3.5\times 10^{-3}$. At the deterministic limit, both the limit cycle and the ferromagnetic points are linearly stable. In Fig. 14, we compared $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ obtained from the theoretical results of Sec. III with stochastic simulations of the spin model. We observed that the behavior is qualitatively the same, but we did not obtain quantitative results. We now discuss the low-$\varepsilon$ approximation and its consequences. We plot in Fig. 15(a) an example of $P_{N}(m,\dot{m})$ obtained from numerical simulations for $N=5\times 10^{5}$, $\varepsilon=10^{-1}$ and $(\mu_{c}-\mu)/\mu_{c}=3.5\times 10^{-3}$. In the deterministic limit, both the limit cycle and the ferromagnetic points are linearly stable for these parameter values. We observe a significant difference with the results obtained in this section, similar to what was observed in Sec. IV.3: there is no individual symmetry $m\,\mapsto\,-m$ or $\dot{m}\,\mapsto\,-\dot{m}$, and the probability density is not constant along the limit cycle. In Sec. IV, we showed that higher order corrections [both in $\phi(m,\dot{m})$ and $C(m,\dot{m})$] to the large deviation form of $P_{N}(m,\dot{m})$ may account for discrepancies between analytical predictions and numerical results of stochastic simulations. We recall that corrections in $\varepsilon$ to the large deviation function $\phi$ lead to changes in the shape of the limit cycle, and are responsible for the breaking of the individual reversal symmetry in $m$ and $\dot{m}$. In contrast, corrections in $N^{-1}$ and $\varepsilon$ to the large deviation function, given by the function $C(m,\dot{m})$ [Eq. (56)], break the uniformity of the probability density along the limit cycle. In Sec. IV, we computed the first corrections analytically for $\mu=\mu_{l}(T)$. However, these corrections are more complicated to compute for any $\mu$, and we thus propose here a different way to determine corrections to the large deviation function. In the deterministic limit $N\to\infty$, Eq. (56) can be rewritten in the form $P_{N}(m,\dot{m})\underset{N\to\infty}{\to}\exp[C(m,\dot{m})]\,\delta(\phi(m,\dot{m})),$ (77) and the probability density on the limit cycle can be obtained from the local speed $v(m,\dot{m})=[\left(dm/dt\right)^{2}+\left(d\dot{m}/dt\right)^{2}]^{1/2},$ (78) since $P(m,\dot{m})\propto v(m,\dot{m})^{-1}$. The deterministic limit provides information on the location of the minima of the function $\phi(m,\dot{m})$, which corresponds to the limit cycle, as well as the value of $C(m,\dot{m})$ on the limit cycle. To obtain these corrections, we compute the trajectory $(m(t),\dot{m}(t))$ in the deterministic limit [using Eq. (8)] and we plot, in Fig. 15(b) the trajectories in the phase space $(m,\dot{m})$ where the color can be interpreted as the local density $P(m,\dot{m})\propto v(m,\dot{m})^{-1}$ along the limit cycle. We observe the same shape of limit cycle as in the stochastic simulations, and we recover a higher probability density near the axis $\dot{m}=0$ (close to the ferromagnetic points), in qualitative agreement with numerical results. ### V.4 Entropy production Figure 16: Entropy production $\Sigma$ vs. $N$, for different values of $\mu$ such that $(\mu_{c}-\mu)/\mu_{c}\in[4.0,3.6,3.4,3.33,3.29,3.28,3.275,3.272,3.271]\times 10^{-5}$ from darker to lighter colors. The inset represents $N_{\Sigma}$, the value of $N$ for which $\Sigma$ is maximal, as a function of $\mu-\mu_{t}$. The dashed line indicates an exponent $-1$. The entropy production is evaluated numerically from Eq. (80) where the averages are computed as in Fig. 11. Parameters: $J_{1}=0.6$, $J_{2}=0.4$, and $\varepsilon=10^{-3}$; $(\mu_{c}-\mu_{t})/\mu_{c}=3.271\times 10^{-5}$. Beyond the order parameter $\langle\dot{m}^{2}\rangle$, the transition to an oscillating phase may also be characterized thermodynamically as a transition from microscopic to macroscopic irreversibility [31, 37], by introducing the entropy production density $\sigma=\Sigma/N$ in the limit $N\to\infty$, where the steady-state entropy production $\Sigma$ identifies with the entropy flux [54, 55], $\Sigma=\frac{1}{2}\sum_{\mathcal{C},\mathcal{C}^{\prime}}\big{[}W(\mathcal{C}^{\prime}|\mathcal{C})P(\mathcal{C})-W(\mathcal{C}|\mathcal{C}^{\prime})P(\mathcal{C}^{\prime})\big{]}\,\ln\frac{W(\mathcal{C}^{\prime}|\mathcal{C})}{W(\mathcal{C}|\mathcal{C}^{\prime})}\,.$ (79) We briefly investigate the influence of the bistability of the system on the entropy production. In the large-$N$ and small-$\varepsilon$ limits, one has (see Appendix C and [42]) $\frac{\Sigma}{N}=\left[1+(T-J_{1})^{2}\right]\left\langle\dot{m}^{2}\right\rangle+T^{2}\left\langle V^{\prime}(m)^{2}\right\rangle.$ (80) In the paramagnetic phase or ferromagnetic phase, one finds $\Sigma=O(N^{0})$, and in an oscillating phase $\Sigma=O(N)$. Using Eq. (80), we compute the entropy production numerically for different system sizes; the results are plotted in Fig. 16. For large $N$, we recover that $\Sigma$ is independent of $N$ in the ferromagnetic phase, whereas $\Sigma\sim N$ in the oscillating phase. However, for moderate values of $N$, one has $\Sigma\sim N^{1/3}$, due to the scaling $\langle\dot{m}^{2}\rangle\sim N^{-2/3}$ obtained in Sec. V.3. Note that this scaling is different from the scaling of the usual transition to an oscillating phase with an elliptic limit cycle, where one finds for moderate $N$, $\Sigma\sim N^{1/2}$, as a consequence of the scaling $\langle\dot{m}^{2}\rangle\sim N^{-1/2}$ [see Eq. (48)]. In the ferromagnetic phase, due to the influence of the oscillating phase, the entropy production increases before having a steep decrease to its constant value. Interestingly, this ‘overshoot’ effect is still present for $\mu$ slightly below $\mu_{O}$ (see Fig. 10), that is when the limit cycle no longer exists at the deterministic level. In this situation, the fluctuations described by the large deviation function keep track of the nearby existence of the limit cycle in parameter space, and are still able to generate a non-monotonous behavior. We introduce $N_{\Sigma}$ the value of $N$ where the entropy production is maximal. The evolution of $N_{\Sigma}$ with $\mu-\mu_{t}$, where $\mu_{t}$ is the value of $\mu$ at the transition between the oscillating and the ferromagnetic phases, is plotted in the inset of Fig. 16 over a range of small values of $\mu-\mu_{t}$. Numerical data can be approximately described by a power-law decay $N_{\Sigma}\sim(\mu-\mu_{t})^{-1}$, although no theoretical prediction is available to support this scaling relation. Accordingly, for $N$ larger than $N_{\Sigma}$, the entropy production drops by an amount $\Delta\Sigma\sim N_{\Sigma}\sim(\mu-\mu_{t})^{-1}$, before reaching its asymptotic constant value. ## VI Type-II discontinuous transition between ferromagnetic and oscillating phases In this section, we investigate the properties of the transition of Type II between the ferromagnetic and oscillating phases, near the tricritical point $(T_{c},\mu_{c})$. In this case, obtained for $J_{1}=J_{2}$, a small, almost elliptic limit cycle around the center is observed. The Type II scenario is illustrated in the phase diagram of Fig. 1(b) and on the trajectories of Figs. 1(d) and 1(f). All figures in this section are obtained with $J_{1}=J_{2}=0.5$. ### VI.1 Large deviation function and phase diagram #### VI.1.1 Validity of the perturbative approach For $J_{1}=J_{2}$, one has $\mu_{c}=1$ and $T_{c}=J_{1}$. The main difference with the previous case is that $v_{1}(T,\mu)$, the factor in front of $m^{4}$ in $V(m)$ [Eq. (61)] vanishes at $(T_{c},\mu_{c})$. We have to leading order in an expansion in $\varepsilon$ and $\mu-\mu_{c}$, $v_{1}(T,\mu)=-\frac{J^{2}}{12}\varepsilon-\frac{1}{4J^{2}}(\mu-\mu_{c})$ (81) with $\varepsilon=(T_{c}-T)/T_{c}$. Corrections to Eq. (81) include terms proportional to $(\mu-\mu_{c})^{2}$, $\varepsilon(\mu-\mu_{c})$ and $\varepsilon^{2}$. Numerically, we observe that the transition between the oscillating and ferromagnetic phases takes place for $(\mu-\mu_{c})\ll\varepsilon$ so that the term in $\mu-\mu_{c}$ can be neglected and we write $v_{1}=-\alpha\varepsilon$ with $\alpha=J^{2}/12$. As $v_{1}<0$, higher order terms in the expansion of $V(m)$ are necessary to compensate for the term $-\varepsilon m^{4}$ and thus to describe the ferromagnetic points. Numerically, we observe ferromagnetic points whose amplitude goes to zero with $\varepsilon$. The coefficients of the terms proportional to $m^{6}$ and $m^{8}$ in the expansion of $V(m)$ obtained from Eq. (105) scale as $\varepsilon$ for small $\varepsilon$, which would give ferromagnetic points independent of $\varepsilon$ in this limit, if only these terms were retained. Expanding $V(m)$ further, we find that the coefficient of the term proportional to $m^{10}$ is independent of $\varepsilon$. Assuming that the ferromagnetic point $m_{0}$ results from the balance of the terms in $m^{4}$ and in $m^{10}$, i.e., $\varepsilon m_{0}^{4}\sim m_{0}^{10}$, yields $m_{0}\sim\varepsilon^{1/6}$. We now check a posteriori that the assumption to neglect the terms in $m^{6}$ and $m^{8}$ was valid. For $m\sim\varepsilon^{1/6}$, one has $\varepsilon m^{6}\sim\varepsilon^{2}$ and $\varepsilon m^{8}\sim\varepsilon^{7/3}$ which are both much smaller than the term $m^{10}\sim\varepsilon^{5/3}$ for $\varepsilon\to 0$, so that neglecting the terms in $m^{6}$ and $m^{8}$ was justified for $m\sim m_{0}$. We thus write the following minimal form for $V(m)$, $V(m)=\frac{\mu-\mu_{l}(T)}{2T^{2}}m^{2}-\frac{\alpha\varepsilon}{4}m^{4}+\frac{v_{4}}{10}m^{10}+V_{0}$ (82) where $v_{4}=J^{2}/81$. $V_{0}$ is such that $V(m)\geq 0$ and the minimal value of $V(m)$ is zero. An example of the shape of $V(m)$ is given in Fig. 17(a). Until now, we have considered only $V(m)$ with one or two minima, whereas now it can have three of them. As we now show, this has important consequences which makes this case of interest, and quite different from the previous ones. We found that the ferromagnetic points are $m_{0}\sim\varepsilon^{1/6}$ when $\mu-\mu_{c}\ll\varepsilon$, so that $u_{0}=g(m_{0},0)\sim\varepsilon^{1/3}$ is small. Thus, for small $\varepsilon$ and close to $\mu_{c}$, one can use the perturbative method described in Sec. III to obtain the large deviation function. #### VI.1.2 Typical $f(H)$ and phase diagram Figure 17: Type-II transition between ferromagnetic and oscillating states. (a) Potential $V(m)$; the dashed lines correspond to $m=\pm m_{l}$. (b) Representation of the different areas in the plane $(m,\dot{m})$. (c) $f(H)$ in the different areas defined in Eq. (83). (d) Colormap of $\phi(m,\dot{m})$ in the plane $(m,\dot{m})$. Note the different scalings $m\sim\varepsilon^{1/2}$ and $\dot{m}\sim\varepsilon$. Parameters: $J_{1}=J_{2}=0.5$, $(\mu-\mu_{l})/\mu_{l}=4.55\times 10^{-4}$ and $\varepsilon=10^{-2}$. The main difference with the previous case is as follows. As illustrated in Fig. 17(a), the condition $V(m)=H$ may correspond to six values of $m$ instead of only two or four previously, when considering values of $H$ close to the local minima of $V(m)$ with $m\neq 0$ (ferromagnetic points). In the definition of $f^{\prime}(H)$ given in Eq. (37), we integrate over $m_{1}(H)$ and $m_{2}(H)$ such that $V(m_{1})=V(m_{2})=H$ and $V(m)\leq H$ for $m\in[m_{1},m_{2}]$. Thus, for a given value of $H$, $f$ can have different values depending on the range of values of $m$ over which the integral is computed. We note $m_{l}$ the positive value of $m$ where $V(m)$ has a local maximum. In the phase space $(m,\dot{m})$ there are four different areas, which are represented in Fig. 17(b). A first area around the center corresponds to small $m$ and $\dot{m}$, where $|m|<|m_{l}|$ and $V(m)<V(m_{l})$, which is denoted area 1 and is represented in green. Two symmetric domains situated around the ferromagnetic points, where $|m|>m_{l}$ and $V(m)<V(m_{l})$, correspond to area 2 and are represented in red. A last area for higher values of $H$, is denoted area 3 and is represented in blue. We define three different functions $f$, one for each area: $\phi(m,\dot{m})\\!=\\!\left\\{\\!\begin{array}[]{ll}f_{1}(H(m,\dot{m}))\\!&\\!\text{if }|m|\\!<\\!m_{l}\text{ and }V(m)\\!<\\!V(m_{l})\\\ f_{2}(H(m,\dot{m}))\\!&\\!\text{if }|m|\\!>\\!m_{l}\text{ and }V(m)\\!<\\!V(m_{l})\\\ f_{3}(H(m,\dot{m}))\\!&\\!\text{otherwise.}\end{array}\right.$ (83) As $f$ is defined up to a constant in every area, we impose that $f_{1}(0)=0$ and we assume $f(H)$ to be continuous at the border between two different areas. In Fig. 17(c), (d), examples of $f(H)$ and $\phi(m,\dot{m})$ are plotted. A limit cycle around the center, and two ferromagnetic points are locally stable. Once again, the most stable phase is given by the global minima of $f$, here the oscillating phase. From the numerical determination of $f(H)$ and its minima, one obtains the phase diagram in the parameter space ($T,\mu$). We plot in Fig. 18 the phase diagram for $J_{1}=J_{2}$ close to the tricritical point $(T_{c},\mu_{c})$. We introduce the line $\mu_{F}$ indicating the existence of the ferromagnetic points and the line $\mu_{O}$ indicating the existence of the oscillating state. We also introduce $\mu_{t}(T)$ the value of $\mu$ at the transition, such that $f_{2}(H(m_{0},0))=f_{1}(H^{*})$ (where $m_{0}$ corresponds to the ferromagnetic point and $H^{*}$ is where $f_{1}(H)$ is minimal). Numerically, we obtain (see Fig. 18) that $(\mu_{t}-\mu_{c})/\mu_{c}\sim\varepsilon^{4/3}.$ (84) Indeed, the transition almost takes place when the ferromagnetic points disappear, meaning that $\mu_{t}\approx\mu_{F}$. The ferromagnetic points disappear when the term in $m^{2}$ balances the $m^{4}$ term in $V(m)$ at $m_{0}\sim\varepsilon^{1/6}$, so that $(\mu_{F}-\mu_{l})m_{0}^{2}\sim\varepsilon m_{0}^{4}$. As $\mu_{l}\sim\mu_{c}$ [Eq. (49)], this gives $(\mu_{F}-\mu_{c})/\mu_{c}\sim\varepsilon^{4/3}$. We observe that when the ferromagnetic phase and the oscillating phase coexist, the ferromagnetic phase is almost always the most stable one. #### VI.1.3 Approximate local analytical expressions of $f(H)$ To go beyond the numerical evaluation of $f(H)$, we now try to obtain an approximate analytical expression of $f(H)$, which will be helpful in particular to determine the scaling regimes of the order parameters $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$. As in Sec. V, one cannot obtain a full analytic expression of $f(H)$, and we thus focus on local approximations. For an expression of $f(H)$ around the ferromagnetic points, a quadratic expansion of $V$ around one of its local minima $m_{0}$ gives $f_{2}(H)=f_{F}(H)+\bar{f}_{2}$ where $f_{F}(H)=-\frac{g(m_{0},0)}{D_{22}+6v_{4}m_{0}^{8}D_{11}}\left[H-V(m_{0})\right],$ (85) and $\bar{f}_{2}$ is a constant such that $f$ is continuous in $H=V(m_{l})$. We note here that $H(m_{0},0)\neq 0$ for the ferromagnetic points unlike in the previous section as now $V(m_{0})$ can be nonzero if the global minimum of $V$ is for $m=0$. For the area around the center, the leading term of $V(m)$ is the $m^{2}$ term, leading to the same $f(H)$ as for the Hopf bifurcation [see Eq. (41)], $f_{1}(H)=f_{O}(H)+\bar{f}_{1}$ where $f_{O}(H)=-\varepsilon aH+bH^{2}+\frac{(\varepsilon a)^{2}}{b}$ (86) with $a$ and $b$ given in Eqs. (42) and (43), and $\bar{f}_{1}$ a constant chosen such that $f(H)$ is continuous in $H=V(m_{l})$. The minimum of $f_{O}(H)$ corresponds to an elliptic limit cycle around the center, as depicted in Fig. 17(d), with $H^{*}=\frac{a\varepsilon}{2b}\sim\varepsilon[\mu-\mu_{l}(T)]\,,$ (87) as $b\sim[\mu-\mu_{l}(T)]^{-1}$ from Eq. (43). At the transition ($\mu=\mu_{t}$), we find [Eqs. (49) and (84)], $\mu_{t}-\mu_{l}(T)\sim\varepsilon^{4/3}$ (88) and thus $H^{*}\sim\varepsilon^{7/3}$. Figure 18: Close up of the phase diagram of Fig. 1(b) for $J_{1}=J_{2}=0.5$ for $T<T_{c}$. $\mu_{F}$ corresponds to the limit of existence of the ferromagnetic points (obtained from $V(m)$), and $\mu_{O}$ of the limit cycle (obtained in the deterministic limit). The transition between the F and O phases takes place for $\mu=\mu_{t}$. In the hatched area both the ferromagnetic points and the limit cycle are local minima of $f(H)$. As $\mu_{F}$ and $\mu_{t}$ are very close to each other, in most of the coexistence region, the ferromagnetic phase is the most stable one. Scalings of transition lines: $(\mu_{t}-\mu_{c})/\mu_{c}\sim\varepsilon^{4/3}$, $(\mu_{F}-\mu_{c})/\mu_{c}\sim\varepsilon^{4/3}$ and $(\mu_{O}-\mu_{c})/\mu_{c}\sim\varepsilon^{2}$, with $\varepsilon=(T_{c}-T)/T_{c}$. In the phase diagram of Fig. 18, we observe that the transition line $\mu_{t}$ is very close to the line $\mu_{F}$ indicating the limit of existence of the ferromagnetic points. Hence in most of the coexistence region, the ferromagnetic phase is the most stable phase. This can be explained with the following argument. Around the ferromagnetic points, $f^{\prime}(H)\sim\varepsilon^{1/3}$ whereas near the limit cycle, $f^{\prime}(H)\sim\varepsilon$. The slope of $f$ near the ferromagnetic points is much steeper that around the limit cycle (see Fig. 17 for an example of $f$). When the area around the ferromagnetic points exists, it rapidly becomes the global minimum of $f$ when varying $\mu$ at fixed $\varepsilon$. ### VI.2 Multiple scalings of order parameters with $N$ at the transition In Figs. 19(a) and 19(b) we plot $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ as a function of $N$ at the transition, for $\mu=\mu_{t}(\varepsilon)$. We observe three different regimes depending on the value of $N$ and $\varepsilon$, that are described below. #### VI.2.1 Large-$N$ scaling at $\mu=\mu_{t}$ Using the local approximations, we obtain the behaviors of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ in the large-$N$ limit when approaching the critical point $\varepsilon=0$ where the three phases meet. In the ferromagnetic phase, using Eq. (64), we find that $\langle m^{2}\rangle=m_{0}^{2}\sim\varepsilon^{1/3}$ and $\langle\dot{m}^{2}\rangle=\frac{D_{22}+6v_{4}m_{0}^{8}D_{11}}{-g(m_{0},0)N},$ (89) so that $\langle\dot{m}^{2}\rangle\sim\varepsilon^{-1/3}N^{-1}$ as $|g(m_{0},0)|\sim m_{0}^{2}\sim\varepsilon^{-1/3}$. Due to the $m^{10}$ term in $V(m)$, we obtain power laws in $\varepsilon$ with critical exponents quite different from the corresponding values previously obtained. In the oscillating phase, one finds $\langle m^{2}\rangle\sim\varepsilon$ similarly to the elliptic limit cycle obtained in Sec. IV, and $\langle\dot{m}^{2}\rangle\sim\varepsilon[\mu_{t}-\mu_{l}(T)]$. One has $\mu_{t}-\mu_{l}\sim\varepsilon^{4/3}$ [Eq. (88)] such that one finds $\langle\dot{m}^{2}\rangle\sim\varepsilon^{7/3}$ in the oscillating phase. #### VI.2.2 Moderate-$N$ scaling at $\mu=\mu_{t}$ Figure 19: Order parameters $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ as a function of the system size $N$, for $\varepsilon\in[10^{-4},10^{-6},10^{-8},10^{-10}]$ from lighter to darker colors at $\mu=\mu_{t}(\varepsilon)$. Order parameters are evaluated numerically from Eqs. (37) and (47). (a) $\langle m^{2}\rangle$ vs $N$. (b) $\langle\dot{m}^{2}\rangle$ vs $N$. The red dashed line corresponds to the theoretical prediction for low $N$ [Eqs. (90) and (91)], with a scaling of $N^{-1/6}$ for $\langle m^{2}\rangle$ and $N^{-5/6}$ for $\langle\dot{m}^{2}\rangle$. (c) $\langle m^{2}\rangle$ rescaled by $\varepsilon^{1/3}$ and (d) $\langle\dot{m}^{2}\rangle$ rescaled by $\varepsilon^{5/3}$ versus rescaled system size $N/\varepsilon^{-2}$, highlighting the crossover between the moderate and intermediate-$N$ regimes. (e) $\langle m^{2}\rangle$ rescaled by $\varepsilon$ and (f) $\langle\dot{m}\rangle$ rescaled by $\varepsilon^{7/3}$ versus rescaled system size $N/\varepsilon^{-10/3}$, highlighting the crossover to the large-$N$ regime. Parameters: $J_{1}=J_{2}=0.5$. For moderate values of $N$, we observe that $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ decrease with $N$ and are independent of $\varepsilon$. Similarly to Sec. V.2.2, for low values of $N$ we can keep only the first $\varepsilon$-independent term in $V(m)$, namely here $V(m)=v_{4}m^{10}/10$. Using Eq. (37), this approximation gives $f(H)=AH^{5/6}$, leading to $\displaystyle\langle m^{2}\rangle\sim N^{-1/6},$ (90) $\displaystyle\langle\dot{m}^{2}\rangle\sim N^{-5/6},$ (91) where the exact asymptotic relations including prefactors are given in Appendix D, and are plotted in dashed red lines in Fig. 19(a,b). #### VI.2.3 Intermediate-$N$ scaling at $\mu=\mu_{t}$ For intermediate values of $N$, we observe that for all $\varepsilon$, $\langle m^{2}\rangle\sim N^{-1/2}$ and $\langle\dot{m}^{2}\rangle\sim N^{-1/2}$. Indeed, for values of $N$ such that $f(H(m_{l}))\sim N^{-1}$ where $m_{l}$ is the positive junction point between the different areas (see Fig. 17), because the ferromagnetic areas are small, the main contribution to the integrals corresponds to $m$ in area 1 (in the center). In this area, one has $f(H)\sim f_{O}(H)$ where $f_{O}(H)$ is given in Eq. (86). The leading correction in $N$ of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ is given by $\langle\dot{m}^{2}\rangle=\frac{\mu-\mu_{l}(T)}{T^{2}}\langle m^{2}\rangle=H^{*}+\frac{e^{-NbH^{*2}}}{2\sqrt{\pi bN}},$ (92) with $H^{*}=a\varepsilon/b$. We recall that $b\sim(\mu-\mu_{l})^{-1}\sim\varepsilon^{-4/3}$ [Eq. (88)] at the transition, thus we find $\langle m^{2}\rangle\sim\varepsilon^{-2/3}N^{-1/2}$ and $\langle\dot{m}^{2}\rangle\sim\varepsilon^{2/3}N^{-1/2}$ for $N\varepsilon^{2}\ll 1$. We note $N_{1}$ the crossover value of $N$ between those the moderate- and intermediate-$N$ regimes, and $N_{2}$ the crossover value between the intermediate- and large-$N$ regimes. For $N\ll N_{1}$, $\langle m^{2}\rangle\sim N^{-1/6}$ while for $N_{1}\ll N\ll N_{2}$, $\langle m^{2}\rangle\sim\varepsilon^{-2/3}N^{-1/2}$, so that $N_{1}\sim\varepsilon^{-2}$. The same argument holds for $\langle\dot{m}^{2}\rangle$: for $N\ll N_{1}$, $\langle\dot{m}^{2}\rangle\sim N^{-5/6}$ and for $N_{1}\ll N\ll N_{2}$, $\langle\dot{m}^{2}\rangle\sim\varepsilon^{2/3}N^{-1/2}$, also implying $N_{1}\sim\varepsilon^{-2}$. These different scaling behaviors for $\langle m^{2}\rangle$ around the first crossover regime $N\sim N_{1}$ can be encompassed into a single scaling function $\langle m^{2}\rangle=\varepsilon^{1/3}\,\mathcal{F}_{m,1}(\varepsilon^{2}N),$ (93) where $\mathcal{F}_{m,1}(x)$ asymptotically behaves as $\mathcal{F}_{m,1}(x)\sim x^{-1/6}$ for $x\to 0$ and $\mathcal{F}_{m,1}(x)\sim x^{-1/2}$ for $x\to\infty$. In a similar way, $\langle\dot{m}^{2}\rangle$ can be expressed in terms of a scaling function, $\langle\dot{m}^{2}\rangle=\varepsilon^{5/3}\,\mathcal{F}_{\dot{m},1}(\varepsilon^{2}N),$ (94) with asymptotic behaviors $\mathcal{F}_{\dot{m},1}(x)\sim x^{-5/6}$ for $x\to 0$ and $\mathcal{F}_{\dot{m},1}(x)\sim x^{-1/2}$ for $x\to\infty$. We plot in Fig. 19(c,d) $\langle m^{2}\rangle/\varepsilon^{1/3}$ and $\langle\dot{m}^{2}\rangle/\varepsilon^{5/3}$ as a function of $N/\varepsilon^{-2}$, which is proportional to the rescaled system size $N/N_{1}$. As expected, the different curves corresponding to different values of $\varepsilon$ collapse for moderate up to intermediate values of $N$. We now turn to the second crossover $N\sim N_{2}$ between the intermediate- and large-$N$ regimes. For $N_{1}\ll N\ll N_{2}$, one has $\langle m^{2}\rangle\sim\varepsilon^{-2/3}N^{-1/2}$ while for $N\gg N_{2}$, $\langle m^{2}\rangle\sim\varepsilon$. Balancing the two contributions thus gives $N_{2}\sim\varepsilon^{-10/3}$ (note that $N_{2}\gg N_{1}$). The same argument holds for $\langle\dot{m}^{2}\rangle$: for $N_{1}\ll N\ll N_{2}$, one finds $\langle\dot{m}^{2}\rangle\sim\varepsilon^{2/3}N^{-1/2}$ and for $N\gg N_{2}$, $\langle\dot{m}^{2}\rangle\sim\varepsilon^{7/3}$, which also gives $N_{2}\sim\varepsilon^{-10/3}$. These scaling behaviors of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ can be encompassed into two scaling functions $\langle m^{2}\rangle=\varepsilon\,\mathcal{F}_{m,2}(\varepsilon^{10/3}N),\quad\langle\dot{m}^{2}\rangle=\varepsilon^{7/3}\,\mathcal{F}_{\dot{m},2}(\varepsilon^{10/3}N),$ (95) with asymptotic behaviors $\mathcal{F}_{m,2}(x)\sim\mathcal{F}_{\dot{m},2}(x)\sim x^{-1/2}$ for $x\to 0$, while both functions go to constant values for $x\to\infty$. In Fig. 19(e,f), we plot $\langle m^{2}\rangle/\varepsilon$ and $\langle\dot{m}^{2}\rangle/\varepsilon^{7/3}$ as a function of the rescaled system size $N/\varepsilon^{-10/3}\sim N/N_{2}$. As expected, for different values of $\varepsilon$, the different curves collapse for intermediate up to large values of $N$. ### VI.3 Crossover between intermediate and large $N$ regimes We reported above three distinct scaling regimes, separated by $N_{1}$ and $N_{2}$, for the observables $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ as a function of the system size $N$ for $\mu=\mu_{t}(\varepsilon)$, and we identified the scalings with $\varepsilon$ of the crossover sizes $N_{1}$ and $N_{2}$. We now investigate in more details the behavior of the observables in the crossover regimes, focusing on the effect of the variable $\mu$ close to the transition value $\mu_{t}$ for a fixed $\varepsilon$. We numerically find that in the moderate- and intermediate-N regimes, the two observables only weakly depend on $\mu$, unlike for large-$N$ values. Thus, we now focus on the dependence of the observables $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ on $\mu$ for a fixed $\varepsilon$, in the crossover regime between intermediate- and large-$N$ values ($N\sim N_{2}$). #### VI.3.1 Influence of $\mu$ on the second crossover regime Figure 20: (a) $\langle m^{2}\rangle$ and (b) $\langle\dot{m}^{2}\rangle$ as a function of the system size $N$, for $(\mu-\mu_{c})/\mu_{c}\in[8.310,8.340,8.348,8.350]\times 10^{-7}$ from darker to lighter colors. Parameters: $J_{1}=J_{2}=0.5$ and $\varepsilon=10^{-4}$. The transition takes place at $(\mu_{t}-\mu_{c})/\mu_{c}\approx 8.349\times 10^{-7}$. Order parameters are evaluated numerically in the same way as in Fig. 19. The black dashed line correspond to $\langle x(m,H)\rangle_{\mathrm{approx}}$ defined in Eq. (96). In the large-$N$ limit, the observables $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ are discontinuous at the transition, whereas for moderate and intermediate-$N$ values, they do not depend much on the value of $\mu$. In Fig. 20, $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ are plotted as a function of system size $N$ for different values of $\mu$ along the transition, at fixed $\varepsilon$. Like in Sec. V, we observe jumps in $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ in the ferromagnetic phase ($\mu<\mu_{t}$) at a finite system size, which takes place for higher values of $N$ when approaching the transition. The two values of $\langle m^{2}\rangle$ in the different phases are nonzero whereas $\langle\dot{m}^{2}\rangle$ goes from a nonzero value (in the oscillating phase) to a value decreasing as $N^{-1}$. The main difference with Sec. V is that the jump in $\langle m^{2}\rangle$ is now much more pronounced because of its different dependence on $\varepsilon$. Indeed, we had in Sec. V that $\langle m^{2}\rangle\sim\varepsilon$ in both phases, whereas now $\langle m^{2}\rangle\sim\varepsilon$ in the oscillating phase and $\langle m^{2}\rangle\sim\varepsilon^{1/3}$ in the ferromagnetic phase, leading for small $\varepsilon$ to a strong mismatch of $\langle m^{2}\rangle$ between the two phases. #### VI.3.2 Approximation in terms of phase coexistence Here again, the observed behaviors can be given a simple interpretation in terms of finite-size phase coexistence and metastability. Hence, as in Sec. V, we introduce a simple decomposition of average values into contributions of each phases. For any quantity $x(m,H)$, we introduce the approximate average value obtained by taking into account the statistical weight of each phase, $\langle x(m,H)\rangle_{\rm approx}=\frac{\langle x\rangle_{F}+\langle x\rangle_{O}C_{2}\sqrt{N}e^{-N(\bar{f}_{2}-\bar{f}_{1})}}{1+C_{2}\sqrt{N}e^{-N(\bar{f}_{2}-\bar{f}_{1})}}$ (96) where the ‘pure-state’ averages $\langle x\rangle_{F}$ and $\langle x\rangle_{O}$ are respectively obtained from the ferromagnetic state large deviation function $f_{F}(H)$ given in Eq. (85), and from the oscillating state large deviation function $f_{O}(H)$ given in Eq. (86); $\bar{f}_{i}$ is the minimum of $f_{i}(H)$ ($i=1$, $2$). The constant $C_{2}$, whose expression is similar to that of the constant $C_{1}$ given in Eq. (75), but with different expressions for $f_{F}$ and $f_{O}$ from the ones found in Sec. V, now becomes $C_{2}=\frac{-g(m_{0},0)}{(D_{22}+6v_{4}m_{0}^{8}D_{11})}\sqrt{\frac{6\pi v_{4}}{bv_{0}}}m_{0}^{4}\,.$ (97) In Sec. V, we could not have a local expression of $f_{O}(H)$ in the oscillating phase, so we used fitting parameters for the coefficients $\tilde{c}$ and $H^{*}$. Here, both $f_{F}$ and $f_{O}$ are known analytically. The only quantity which is not known analytically and is a fitted parameter, obtained from the numerical evaluation of $f(H)$, is $\bar{f}_{2}-\bar{f}_{1}$. This decomposition is plotted for different values of $\mu$ in Fig. 20 in black dashed lines. The jump in $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ is well described by this simple decomposition. When increasing $N$, the influence of the oscillating phase dominates until $f(H^{*})\sim N^{-1}$. #### VI.3.3 Comparison with stochastic simulations Figure 21: Comparison with stochastic simulations of the spin model. (a) $\langle m^{2}\rangle$ and (b) $\langle\dot{m}^{2}\rangle$ obtained from numerical simulations as a function of the system size $N$, for $(\mu-\mu_{c})/\mu_{c}\in[2,3,4]\times 10^{-3}$ (from darker to lighter colors). Parameters: $J_{1}=J_{2}=0.5$ and $\varepsilon=5\times 10^{-2}$. We plot in Fig. 21 $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ computed from stochastic simulations for different system sizes and different $\mu$, with $\varepsilon=(T_{c}-T)/T_{c}=5\times 10^{-2}$. We observe the different expected behaviors described above. At large $N$, a jump of $\langle m^{2}\rangle$ is observed between the ferromagnetic phase (high values of $\langle m^{2}\rangle\sim\varepsilon^{1/3}$ for low values of $\mu-\mu_{c}$) and the oscillating phase (low values of $\langle m^{2}\rangle\sim\varepsilon$ for higher values of $\mu-\mu_{c}$). A jump of $\langle\dot{m}^{2}\rangle$ is also observed: $\langle\dot{m}^{2}\rangle$ is constant and of order $\varepsilon^{7/3}\sim 10^{-3}$ for sufficiently high values of $\mu-\mu_{c}$, while it decreases as $N^{-1}$ for lower values of $\mu-\mu_{c}$. Similarly to Sec. V, we are able to describe the qualitative behavior of the observables $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$. However, for the restricted range of values of $\varepsilon$ and $N$ accessible in stochastic simulations, we are not able to reach a quantitative agreement with analytical predictions obtained in the small-$\varepsilon$, large-$N$ limit. We now briefly discuss the effect of considering finite values of $\varepsilon$ and $N$. Figure 22: (a) $P_{N}(m,\dot{m})$ obtained from stochastic simulations for $N=2\times 10^{4}$. (b) Trajectories ($m(t),\dot{m}(t))$ in the phase space $(m,\dot{m})$ obtained in the deterministic limit. The color corresponds to $v(m,\dot{m})^{-1}$ where $v(m,\dot{m})$ is the local velocity, as the local density $p(m,\dot{m})\sim v(m,\dot{m})^{-1}$. The black red dots are added for visual purpose and correspond to the ferromagnetic points. Parameters: $J_{1}=J_{2}=0.5$, $\varepsilon=5\times 10^{-2}$ and $(\mu-\mu_{c})/\mu_{c}=3.981\times 10^{-3}$. In Fig. 22(a), we plot an example of $P_{N}(m,\dot{m})$ obtained from stochastic simulations for $\varepsilon=10^{-1}$ and $N=5\times 10^{5}$, for a value of $\mu$ where both the limit cycle and the ferromagnetic points are linearly stable in the deterministic limit. We observe important discrepancies with the theory described in Sec. VI.1, which assumed a small elliptic limit cycle in the center with uniform probability along the cycle, and ferromagnetic points outside the cycle. From a dynamical viewpoint, a non- uniform probability along the cycle in an ensemble approach means that an individual system goes along the cycle at a non-uniform speed. Here, as shown in Fig. 22(a), the limit cycle is hardly visible, meaning the probability is strongly non-uniform along the cycle. To understand this result, we plot in Fig. 22(b) trajectories, obtained in the deterministic limit, in the phase space $(m,\dot{m})$ where the color codes for $v(m,\dot{m})^{-1}$, with $v(m,\dot{m})=[\left(dm/dt\right)^{2}+\left(d\dot{m}/dt\right)^{2}]^{1/2}$ the local speed on the cycle. The quantity $v(m,\dot{m})^{-1}$ is proportional the local probability density $p(m,\dot{m})$ along the limit cycle. We observe that the limit cycle is not elliptic, and that the speed $v(m,\dot{m})$ is far from uniform along the cycle. The probability density $p(m,\dot{m})$ is much higher close to the ferromagnetic points, in qualitative agreement with stochastic simulations. This discrepancy with theoretical predictions derived in Sec. VI.1 comes from both the small-$\varepsilon$ and large-$N$ approximations made to obtain analytical results. The small-$\varepsilon$ approximation gives an elliptic limit cycle, and the large-$N$ one gives a constant speed along the limit cycle [see Sec. IV.3 and Sec. V]. ### VI.4 Entropy production Figure 23: Entropy production $\Sigma$ vs. $N$, for $(\mu-\mu_{c})/\mu_{c}\in[6,7.2,8.0,8.2,8.3,8.330,8.343,8.348,8.352]\times 10^{-7}$ from darker to lighter colors. The entropy production is evaluated numerically from Eq. (80). Inset: $N_{\Sigma}$, the value of $N$ where $\Sigma$ is maximal, as a function of $\mu-\mu_{t}$, with $(\mu_{t}-\mu_{c})/\mu_{c}=8.352\times 10^{-7}$. A slope $-1.3$ is indicated by a dashed line, as a guide to the eye. Parameters: $J_{1}=J_{2}=0.5$ and $\varepsilon=10^{-4}$. Similarly to Sec. V, we discuss the behavior of the entropy production Eq. (80) with system size at the transition, which is plotted in Fig. 23 for different values of $\mu$ across the transition ($\mu\approx\mu_{t}$). For low values of $N$, the entropy production increases as $N^{1/6}$ as $\langle\dot{m}^{2}\rangle\sim N^{-5/6}$. For larger values of $N$ and for $\mu>\mu_{t}$, $\Sigma\sim N$ as expected in an oscillating phase. For $\mu<\mu_{t}$, as in Sec. V, we observe that $\Sigma$ increases like in the oscillating phase before steeply decreasing to a constant value. This behavior is similar to the one observed in Sec. V and is a consequence of the proximity of the oscillating phase. We again denote as $N_{\Sigma}$ the value of $N$ when $\Sigma$ is maximal. In the inset of Fig. 23, we plot $N_{\Sigma}$ as a function of the distance to the transition $\mu-\mu_{t}$, showing an approximate power-law divergence of $N_{\Sigma}$ when $\mu-\mu_{t}\to 0$. As a result, getting closer to the transition, the drop of $\Sigma$ takes place for larger $N$ and thus becomes bigger, since it eventually decays to approximately the same asymptotic large-$N$ value for all $\mu-\mu_{t}$. ### VI.5 Comment on the continuous transition for $J_{2}=\pm 2+J_{1}$ In this section, we investigated the properties of the discontinuous transition of Type II between the ferromagnetic phase and the oscillating phase taking place for $J_{1}=J_{2}$ where $v_{1}(T_{c},\mu_{c})=0$. As seen in Fig. 2, the condition $v_{1}(T_{c},\mu_{c})=0$ is also satisfied for $J_{2}=\pm 2+J_{1}$, and it would thus be natural to also study the transition in this case. However, in the deterministic limit we find for $J_{2}=\pm 2+J_{1}$ that the transition between the ferromagnetic phase and the oscillating phase is of a different type: the transition is continuous, and the ferromagnetic points turn into a limit cycle with a infinite period at the transition. We now comment further on the difference between the two cases $J_{1}=J_{2}$ and $J_{2}=\pm 2+J_{1}$ and why the method presented in this paper does not allow for a characterization of continuous transitions between ferromagnetic points and a limit cycle with infinite period. In Sec. VI.1, we computed the large deviation locally using Eq. (37) and assumed that the large deviation function is continuous in $m=m_{l}$ in order to obtain the large deviation function in all three areas, and thus in the whole plane ($m,\dot{m}$). However, an important issue is that at the point where the different areas meet, $m=m_{l}$, we have $V^{\prime}(m_{l})=0$ and $g(m_{l},0)\neq 0$. Thus, the assumption $V^{\prime}(m)\gg\dot{m}g(m,\dot{m})$ made to split the different orders of Eq. (LABEL:eq:phi:quadrat) is not valid close to $m=m_{l}$. Therefore, we expect to get corrections close to $m_{l}$ which are not taken into account here. When the ferromagnetic points and the limit cycle are well separated and far from $m_{l}$, which is true for $J_{1}=J_{2}$ as we have $m\sim\varepsilon^{1/6}$ for the ferromagnetic phase and $m\sim\varepsilon^{1/2}$ for the oscillating phase, the corrections close to $m_{l}$ do not affect the qualitative behavior at the transition, and thus the method presented in this section describes well the phase transition. However, if one or both of them are close to $m_{l}$, corrections, that are not taken into account in this paper and which would require a different approach, are necessary. For $J_{2}=\pm 2+J_{1}$, we find that the $m^{6}$-term in $V(m)$ is independent of $\varepsilon$ (whereas for $J_{1}=J_{2}$ it scales as $\varepsilon$) and thus is enough to describe the ferromagnetic points, which turn out to scale as $m\sim\varepsilon^{1/2}$ (as seen by balancing the terms in $\varepsilon m^{4}$ and $m^{6}$). If we blindly apply the method described in this section, we find a limit cycle with $m\sim\varepsilon^{1/2}$ in between the ferromagnetic points, which have the same scaling. The limit cycle is very close to the point $m=m_{l}$ where $V^{\prime}(m_{l})=0$ and where the theory breaks downs. Furthermore, from the definition of the period of the limit cycle Eq. (44), we find that when $H^{*}\approx V(m_{l})$, the period is very large, and diverges when $H^{*}=V(m_{l})$. Thus, by applying the method without enough care, we would still recover some qualitative properties of the transition: the ferromagnetic points and the limit cycle are very close to each other and the period of the limit cycle is very large. Yet, this would not be a correct description of the transition as we would find a discontinuous transition instead of a continuous one as observed numerically. ## VII Conclusion We have considered a mean-field spin model with a dynamics breaking detailed balance due to the non-reciprocal couplings between spins and auxiliary dynamic fields. The presence of ferromagnetic interactions between spins on one side, and between dynamic fields on the other side, allows for the presence of both ferromagnetic and spontaneously oscillating phases. We have characterized in details the transition between these two phases, showing that it is discontinuous with the coexistence of both ferromagnetic and oscillating states, one state being stable and the other one metastable. The relative stability of both states is determined by a large deviation function, generalizing the Landau free energy, that we evaluated explicitly in different cases thanks to a perturbative framework. Two main scenarios are discussed, whether the ferromagnetic points turn out to be inside or outside the limit cycle. In addition, we found that the entropy production is peaked as a function of system size, leading to a maximally dissipative system for an optimal finite system size. A natural generalization of this work may be to try to extend these results beyond mean-field, by considering finite-dimensional systems, with the goal to formulate a Ginzburg-Landau theory extending the present Landau framework based on a large deviation principle. Such a theory might then be amenable to a renormalization group treatment, extending the results of [3, 4] which considered the synchronization of coupled oscillators. Here, we have started from more basic ingredients, in the sense that the microscopic degrees of freedom of the model (i.e., the spins and dynamic fields) do not oscillate in the absence of interaction. Connecting these types of models to previous results obtained on the synchronization transition is thus an interesting challenge for future work. ###### Acknowledgements. L. G. acknowledges funding from the French Ministry of Higher Education and Research. ## Appendix A Derivation of the deterministic evolution equations In this appendix, we derive the deterministic evolution equations Eqs. (6) and (7) from the microscopic spin and field dynamics. The dynamics of the system is determined from the master equation [see Eq. (5)]. As the average is defined as $\langle x\rangle=\sum_{\mathcal{C}}x(\mathcal{C})P(\mathcal{C})$, we find after rearranging terms, $\displaystyle d_{t}\langle m\rangle=\sum_{\mathcal{C}}P(\mathcal{C})\sum_{\mathcal{C}^{\prime}\neq\mathcal{C}}\left[m(\mathcal{C}^{\prime})-m(\mathcal{C})\right]W(\mathcal{C}^{\prime}|\mathcal{C}),$ (98) $\displaystyle d_{t}\langle h\rangle=\sum_{\mathcal{C}}P(\mathcal{C})\sum_{\mathcal{C}^{\prime}\neq\mathcal{C}}\left[h(\mathcal{C}^{\prime})-h(\mathcal{C})\right]W(\mathcal{C}^{\prime}|\mathcal{C}),$ (99) with the shorthand notation $d_{t}\equiv d/dt$. From a configuration $\mathcal{C}=\\{s_{1},...,s_{N},h_{1},...,h_{N}\\}$ with magnetization $m=N^{-1}\sum_{i=1}^{N}s_{i}$ and average field $h=N^{-1}\sum_{i=1}^{N}h_{i}$, there are $(1\pm m)N/2$ possibilities to flip a spin $s_{i}=\pm 1$ and $(1\pm h)N/2$ possibilities to flip a field $h_{i}=\pm 1$. For a flip of a spin $s_{i}=\pm 1$, $m(\mathcal{C}^{\prime})-m(\mathcal{C})=\mp 2/N$ and for a flip of a field $h_{i}=\pm 1$, $h(\mathcal{C}^{\prime})-h(\mathcal{C})=\mp 2/N$. Thus, using the definition of the transition rates given in the main text [Eq. (2)], we find: $\displaystyle d_{t}\langle m\rangle$ $\displaystyle=\big{\langle}-m+\tanh[\beta(J_{1}m+h)]\,\big{\rangle},$ (100) $\displaystyle d_{t}\langle h\rangle$ $\displaystyle=\big{\langle}-h+\tanh[\beta(J_{2}h+(1-\mu)m)]\,\big{\rangle}.$ (101) Assuming that the law of large numbers applies in the limit $N\to\infty$, $m$ and $h$ obey the following deterministic equations: $\displaystyle d_{t}m$ $\displaystyle=-m+\tanh[\beta(J_{1}m+h)],$ (102) $\displaystyle d_{t}h$ $\displaystyle=-h+\tanh[\beta(J_{2}h+(1-\mu)m)].$ (103) These deterministic equations can be used to determine the macroscopic phase when a single solution exists for given values of the control parameters $\beta=T^{-1}$ and $\mu$. When two solutions exist, the most stable one has to be determined from the large deviation function approach, as explained in the main text. ## Appendix B Values of the different functions and coefficients of the model In this appendix, we give the values of the different functions and coefficients introduced in the main text. The function $Y(m,\dot{m})$ introduced in Eq. (8) as $d_{t}\dot{m}=Y(m,\dot{m})$ has been split into a $\dot{m}$-independent part, $V^{\prime}(m)=Y(m,0)$ and a $\dot{m}$-dependent part $\dot{m}g(m,\dot{m})=Y(m,\dot{m})-Y(m,0)$. From Eqs. (6) and (7), we have: $\displaystyle Y(m,\dot{m})=$ $\displaystyle\beta J_{1}m+(-1+\beta J_{1})\dot{m}-\tanh^{-1}(m+\dot{m})+\beta\tanh[J_{2}\tanh^{-1}(m+\dot{m})+\beta(1-\mu- J_{1}J_{2})m]$ (104) $\displaystyle+(m+\dot{m})^{2}\left[\tanh^{-1}(m+\dot{m})-\beta\tanh[J_{2}\tanh^{-1}(m+\dot{m})+\beta(1-\mu- J_{1}J_{2})m]-\beta J_{1}(m+\dot{m})\right]\,,$ $V^{\prime}(m)=-\beta J_{1}m+\beta J_{1}m^{3}+(1-m^{2})\tanh^{-1}(m)-\beta(1-m^{2})\tanh[J_{2}\tanh^{-1}(m)+\beta(1-\mu- J_{1}J_{2})m]\,.$ (105) In the main text, we introduced the following expansions of $V(m)$ and $g(m,\dot{m})$ [see Eqs. (16) and (17)], $V(m)=\frac{v_{0}}{2}m^{2}+\frac{v_{1}}{4}m^{4},$ (106) and $g(m,\dot{m})=a_{0}\varepsilon-a_{1}m^{2}-a_{2}m\dot{m}-a_{3}\dot{m}^{2}.$ (107) The coefficients appearing in these expansions are given by $\displaystyle v_{0}=(\mu-1)/T^{2}+(1-J_{1}/T)(1-J_{2}/T),$ (108) $\displaystyle\begin{aligned} v_{1}=&-2/3+(2J_{2}+3J_{1})/3T-(\mu-1+J_{1}J_{2})/T^{2}\\\ &-(\mu-1-J_{2}T+J_{1}J_{2})^{3}/3T^{4},\end{aligned}$ (109) $\displaystyle\varepsilon=(T_{c}-T)/T_{c},$ (110) $\displaystyle a_{0}=2T_{c}/T,$ (111) $\displaystyle\begin{aligned} a_{1}&=-2+(2J_{2}+J_{2}^{3}+3J_{1})/T\\\ &+J_{2}(-1+J_{1}J_{2}+\mu)^{2}/T^{3}\\\ &-2(1+J_{2}^{2})(-1+J_{1}J_{2}+\mu)/T^{2},\end{aligned}$ (112) $\displaystyle\begin{aligned} a_{2}&=-2+2\beta J_{2}+\beta J_{2}^{3}+3\beta J_{1}\\\ &+\beta^{2}(1+J_{2}^{2})(-1+J_{1}J_{2}+\mu),\end{aligned}$ (113) $\displaystyle a_{3}=-2/3+(2J_{2}+J_{2}^{3}+3J_{1})/3T.$ (114) In addition, we introduced in Eq. (31) the coefficients $D_{11}(m,\dot{m})$, $D_{12}(m,\dot{m})=D_{21}(m,\dot{m})$ and $D_{22}(m,\dot{m})$, which read: $\displaystyle D_{11}=1-m(m+\dot{m}),$ (115) $\displaystyle D_{12}=[1-m(m+\dot{m})][-1+\beta J_{1}(1-(m+\dot{m})^{2})]$ (116) $\displaystyle\begin{aligned} D_{22}&=[1-m(m+\dot{m})][-1+\beta J_{1}(1-(m+\dot{m})^{2})]^{2}\\\ &+[1-h(h+\dot{h})]\beta^{2}[1-(m+\dot{m})^{2}]^{2}\end{aligned}$ (117) with $h=T\tanh^{-1}(m+\dot{m})-J_{1}m$ (118) and $\dot{h}(m,\dot{m})=T\frac{Y(m,\dot{m})+\dot{m}}{1-(m+\dot{m})^{2}}-J_{1}\dot{m}.$ (119) In the main text, we use the coefficients evaluated at $m=0$ and $\dot{m}=0$, which simplify to: $\displaystyle D_{11}(0,0)=1,$ (120) $\displaystyle D_{12}(0,0)=-1+J_{1}/T,$ (121) $\displaystyle D_{22}(0,0)=1/T^{2}+(J_{1}/T-1)^{2}.$ (122) ## Appendix C Entropy production In steady-state, the entropy production can be identified with the entropy flux, which is defined from the microscopic configurations $\mathcal{C}=\\{s_{1},...,s_{N},h_{1},...,h_{N}\\}$ as $\Sigma=\sum_{\mathcal{C},\mathcal{C}^{\prime}}W(\mathcal{C}^{\prime}|\mathcal{C})P(\mathcal{C})\,\ln\frac{W(\mathcal{C}^{\prime}|\mathcal{C})}{W(\mathcal{C}|\mathcal{C}^{\prime})}\,.$ (123) Note that Eq. (123) is equivalent to the definition (80) given in the main text, up to a symmetrization of expression (123). To compute the entropy production, we aim at changing the sum over the configurations into integrals over $m$ and $\dot{m}$. We thus replace $\sum_{\mathcal{C}}$ by $\int dmd\dot{m}\sum_{\mathcal{C}\in S(m,\dot{m})}$, where $S(m,\dot{m})$ denotes the ensemble of configurations $\mathcal{C}$ with $m(\mathcal{C})=m$ and $\dot{m}(\mathcal{C})=\dot{m}$. We transform the integral over $\mathcal{C}^{\prime}$ into a sum over all possible transitions. We recall that spin reversals are labelled with $k=1$ and field reversals with $k=2$, keeping track of the sign $\sigma=\pm$ of the variable prior to reversal. We denote as $W_{k}^{\sigma}$ the coarse-grained transition rates given in Eq. (26) and $n_{k}^{\sigma}$ the fraction of possible transitions, $n_{1}^{\pm}=\frac{1}{2}(1\pm m),\quad n_{2}^{\pm}=\frac{1}{2}(1\pm h).$ (124) We find $\Sigma=N\sum_{k,\sigma}\left\langle W_{k}^{\sigma}(\mathbf{x})\ln\left[\\!\frac{W_{k}^{\sigma}(\mathbf{x})n_{k}^{-\sigma}(\mathbf{x}+\frac{\sigma\mathbf{d}_{k}}{N})}{W_{k}^{-\sigma}(\mathbf{x}+\frac{\sigma\mathbf{d}_{k}}{N})n_{k}^{\sigma}(\mathbf{x})}\right]\\!\right\rangle$ (125) with the shorthand notation $\mathbf{x}=(m,\dot{m})$. To further simplify notations, we make the dependence on $\mathbf{x}$ implicit in what follows. At leading order in $N$, we find $\Sigma=N\sum_{k}\left\langle\left(W_{k}^{+}-W_{k}^{-}\right)\ln\left[\frac{W_{k}^{+}n_{k}^{-}}{W_{k}^{-}n_{k}^{+}}\right]\right\rangle.$ (126) We define $\Sigma_{k}=\left\langle\left(W_{k}^{+}-W_{k}^{-}\right)\ln\left[\frac{W_{k}^{+}n_{k}^{-}}{W_{k}^{-}n_{k}^{+}}\right]\right\rangle.$ (127) For $k=1$, we have $W_{1}^{+}-W_{1}^{-}=\dot{m}/2$ and $\ln\left[\frac{W_{k}^{+}n_{k}^{-}}{W_{k}^{-}n_{k}^{+}}\right]=2\tanh^{-1}(m+\dot{m}),$ (128) thus we find that $\Sigma_{1}=N\left\langle\dot{m}\tanh^{-1}(m+\dot{m})\right\rangle.$ (129) In the main text, we showed that close to a transition, one generically has a scaling behavior $\dot{m}\sim\varepsilon^{\alpha}$ (with $\alpha>0$ an exponent depending on the specific transition considered) and $\langle m\dot{m}\rangle=0$. Thus, keeping the lowest order in $\varepsilon$, we find $\Sigma_{1}/N=\langle\dot{m}^{2}\rangle.$ (130) For $k=2$, we recall that we have $h(m,\dot{m})=-J_{1}m+T\tanh^{-1}[m+\dot{m}]$ and we introduce $\dot{h}=-h+\tanh[\beta(J_{2}h+(1-\mu)m)]$, the equivalent of $\dot{m}$ for the fields variables $h_{i}$, given in Eq. (119) as a function of $m$ and $\dot{m}$. We find $\Sigma_{2}=N\langle\dot{h}\tanh^{-1}(h+\dot{h})\rangle$. As spins and fields play symmetric roles, one also has $\langle h\dot{h}\rangle=0$ close to a transition. Using that $Y(m,\dot{m})=-V^{\prime}(m)$ and $\langle V^{\prime}(m)\dot{m}\rangle=0$ at the lowest order in $\varepsilon$, we find $\frac{\Sigma_{2}}{N}=\langle T^{2}V^{\prime}(m,\dot{m})^{2}+(T-J_{1})^{2}\dot{m}^{2}\rangle.$ (131) Gathering contributions, one finds $\frac{\Sigma}{N}=\left[1+(T-J_{1})^{2}\right]\left\langle\dot{m}^{2}\right\rangle+T^{2}\left\langle V^{\prime}(m)^{2}\right\rangle$ (132) at leading order in $\varepsilon$ and $N$. ## Appendix D Moderate-$N$ approximation of Sec. VI In this appendix, we give the exact expressions of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ for moderate values of $N$ for the Type-II discontinuous transition of Sec. VI. For intermediate values of $N$, only high values of $V(m)$ contribute, thus we consider that $V(m)=v_{4}m^{10}/10$. Using Eq. (37), we find $f(H)=AH^{6/5}\,,$ (133) with $A=\frac{9\Gamma\left(\frac{3}{10}\right)^{2}a_{1}}{(2^{6}5^{4}v_{4})^{1/5}\sqrt{\pi}\Gamma\left(\frac{1}{10}\right)}\,.$ (134) Thus, using the definition of $\langle m^{2}\rangle$ and $\langle\dot{m}^{2}\rangle$ from Eq. (47) we find: $\displaystyle\langle m^{2}\rangle=\frac{c_{m}}{(a_{1}v_{4})^{1/6}}N^{-1/6},$ (135) $\displaystyle\langle\dot{m}^{2}\rangle=c_{\dot{m}}\frac{v_{4}^{1/6}}{a_{1}^{5/6}}N^{-5/6},$ (136) with $\displaystyle c_{m}=\frac{5^{1/3}\Gamma\left(\frac{3}{10}\right)\Gamma\left(\frac{2}{3}\right)}{3\times 2^{2/5}\pi}\left(\frac{\Gamma\left(\frac{13}{10}\right)}{\Gamma\left(\frac{11}{10}\right)}\right)^{5/6}\\!\left(\frac{\Gamma\left(\frac{9}{5}\right)}{\Gamma\left(\frac{8}{5}\right)}\right)^{1/6}\\!\approx 0.53,$ (137) $\displaystyle c_{\dot{m}}=\frac{5^{2/3}\Gamma\left(\frac{3}{5}\right)\Gamma\left(\frac{4}{3}\right)\left(\Gamma\left(\frac{11}{10}\right)\Gamma\left(\frac{9}{5}\right)\right)^{5/6}}{2\sqrt{\pi}\Gamma\left(\frac{8}{5}\right)^{11/6}\Gamma\left(\frac{13}{10}\right)^{5/6}}\approx 1.33.$ (138) ## References * Acebrón _et al._ [2005] J. A. Acebrón, L. L. Bonilla, C. J. Pérez Vicente, F. Ritort, and R. Spigler, Rev. Mod. Phys. 77, 137 (2005). * Gupta _et al._ [2014] S. Gupta, A. Campa, and S. Ruffo, J. Stat. Mech.: Theor. Exp. , R08001 (2014). * Risler _et al._ [2004] T. Risler, J. Prost, and F. Jülicher, Phys. Rev. Lett. 93, 175702 (2004). * Risler _et al._ [2005] T. Risler, J. Prost, and F. Jülicher, Phys. Rev. E 72, 016130 (2005). * Nicolis [1986] G. Nicolis, Rep. Prog. Phys. 49, 873 (1986). * Kamino _et al._ [2017] K. Kamino, Y. Kondo, A. Nakajima, M. Honda-Kitahara, K. Kaneko, and S. Sawai, Proc. Natl. Acad. Sci. USA 114, E4149 (2017). * Wang and Tang [2019] S.-W. Wang and L.-H. Tang, Nat. Commun. 10, 5613 (2019). * Saha _et al._ [2020] S. Saha, J. Agudo-Canalejo, and R. Golestanian, Phys. Rev. X 10, 041009 (2020). * You _et al._ [2020] Z. You, A. Baskaran, and M. C. Marchetti, Proc. Natl. Acad. Sci. USA 117, 19767 (2020). * Cao _et al._ [2015] Y. Cao, H. Wang, Q. Ouyang, and Y. Tu, Nat. Phys. 11, 772 (2015). * Nguyen _et al._ [2018] B. Nguyen, U. Seifert, and A. C. Barato, J. Chem. Phys. 149, 045101 (2018). * Aufinger _et al._ [2022] L. Aufinger, J. Brenner, and F. C. Simmel, Nat. Commun. 13, 2852 (2022). * Devailly _et al._ [2015] C. Devailly, C. Crauste-Thibierge, A. Petrosyan, and S. Ciliberto, Phys. Rev. E 92, 052312 (2015). * Andrae _et al._ [2010] B. Andrae, J. Cremer, T. Reichenbach, and E. Frey, Phys. Rev. Lett. 104, 218102 (2010). * Duan _et al._ [2019] D. Duan, B. Niu, and J. Wei, Chaos, Solitons and Fractals 123, 206 (2019). * Gualdi _et al._ [2015] S. Gualdi, J.-P. Bouchaud, G. Cencetti, M. Tarzia, and F. Zamponi, Phys. Rev. Lett. 114, 088701 (2015). * Yi _et al._ [2015] S. D. Yi, S. K. Baek, G. Chevereau, and E. Bertin, J. Stat. Mech.: Theor. Exp. , P11001 (2015). * Collet _et al._ [2016] F. Collet, M. Formentin, and D. Tovazzi, Phys. Rev. E. 94, 042139 (2016). * De Martino and Barato [2019] D. De Martino and A. C. Barato, Phys. Rev. E 100, 062123 (2019). * Dai Pra _et al._ [2020] P. Dai Pra, M. Formentin, and P. Guglielmo, J. Stat. Phys. 179, 690 (2020). * Crawford [1991] J. D. Crawford, Rev. Mod. Phys. 63, 991 (1991). * Fei _et al._ [2018] C. Fei, Y. Cao, Q. Ouyang, and Y. Tu, Nat. Commun. 9, 1434 (2018). * Gaspard [2002] P. Gaspard, J. Chem. Phys. 117, 8905 (2002). * Barato and Seifert [2016] A. C. Barato and U. Seifert, Phys. Rev. X 6, 041053 (2016). * Barato and Seifert [2017] A. C. Barato and U. Seifert, Phys. Rev. E 95, 062409 (2017). * Oberreiter _et al._ [2022] L. Oberreiter, U. Seifert, and A. C. Barato, Phys. Rev. E 106, 014106 (2022). * Remlein _et al._ [2022] B. Remlein, V. Weissmann, and U. Seifert, Phys. Rev. E 105, 064101 (2022). * Sagués _et al._ [2007] F. Sagués, J. M. Sancho, and J. García-Ojalvo, Rev. Mod. Phys. 79, 829 (2007). * Xu _et al._ [2020] H.-Y. Xu, Y.-P. Luo, J.-W. Wu, and M.-C. Huang, Physica D 411, 132612 (2020). * Crochik and Tomé [2005] L. Crochik and T. Tomé, Phys. Rev. E 72, 057103 (2005). * Xiao _et al._ [2008] T. J. Xiao, Z. Hou, and H. Xin, J. Chem. Phys. 129, 114506 (2008). * Xiao _et al._ [2009] T. Xiao, Z. Hou, and H. Xin, J. Phys. Chem. B 113, 9316 (2009). * Barato and Hinrichsen [2012] A. C. Barato and H. Hinrichsen, J. Phys. A: Math. Theor. 45, 115005 (2012). * Tomé and de Oliveira [2012] T. Tomé and M. J. de Oliveira, Phys. Rev. Lett. 108, 020601 (2012). * Noa _et al._ [2019] C. E. F. Noa, P. E. Harunari, M. J. de Oliveira, and C. E. Fiore, Phys. Rev. E 100, 012104 (2019). * Martynec _et al._ [2020] T. Martynec, S. H. L. Klapp, and S. A. M. Loos, New J. Phys. 22, 093069 (2020). * Seara _et al._ [2021] D. S. Seara, B. B. Machta, and M. P. Murrell, Nat. Commun. 12, 392 (2021). * Le Bellac [1992] M. Le Bellac, _Quantum and Statistical Field Theory_ (Oxford Science Publications, Oxford, 1992). * Meibohm and Esposito [2022] J. Meibohm and M. Esposito, Phys. Rev. Lett. 128, 110603 (2022). * Holtzman and Raz [2022] R. Holtzman and O. Raz, Commun. Phys. 5, 280 (2022). * Aron and Chamon [2020] C. Aron and C. Chamon, SciPost Phys. 8, 074 (2020). * Guislain and Bertin [2023] L. Guislain and E. Bertin, Phys. Rev. Lett. 130, 207102 (2023). * Fruchart _et al._ [2021] M. Fruchart, R. Hanai, P. Littlewood, and V. Vitelli, Nature 592, 363 (2021). * Martin _et al._ [2023] D. Martin, D. Daniel Seara, Y. Avni, M. Fruchart, and V. Vitelli, (2023), arXiv:2307.08251 . * Collet [2014] F. Collet, J. Stat. Phys. 157, 1301 (2014). * Sinelschikov _et al._ [2023] D. Sinelschikov, A. Poggialini, M. F. Abbate, and D. De Martino, Emergence of collective self-oscillations in minimal lattice models with feedback (2023), arXiv:2306.01823 . * Touchette [2009] H. Touchette, Physics Reports 478 (2009). * Knessl _et al._ [1985] C. Knessl, B. J. Matkowsky, Z. Schuss, and C. Tier, SIAM J. Appl. Math. 46, 1006 (1985). * Graham and Tél [1987] R. Graham and T. Tél, Phys. Rev. A 35, 1328 (1987). * Graham and Tél [1984] R. Graham and T. Tél, Journal of Statistical Physics 35, 729 (1984). * Graham [1989] R. Graham, Journal of Statistical Physics 54, 1207 (1989). * Gillespie [2007] D. T. Gillespie, Annu. Rev. Phys. Chem. 58, 35 (2007). * [53] See Supplemental Material at XXX. * Schnackenberg [1976] J. Schnackenberg, Rev. Mod. Phys. 48, 571 (1976). * Gaspard [2004] P. Gaspard, J. Stat. Phys. 117, 599 (2004). Supplementary Information: Discontinuous phase transition from ferromagnetic to oscillating states in a nonequilibrium mean-field spin model ## Corrections to the large deviation function for $\mu=\mu_{l}$ Stochastic simulations are made for values of $\varepsilon$ and $N$ for which the assumptions of small $\varepsilon$ and large $N$ are not fully satisfied.
Statistical estimation of full-sky radio maps from 21cm array visibility data using Gaussian Constrained Realisations–A # Statistical estimation of full-sky radio maps from 21cm array visibility data using Gaussian Constrained Realisations Katrine A. Glasscock1 E-mail<EMAIL_ADDRESS>(KAG) 0000-0001-6894-0902 Philip Bull1,2 0000-0001-5668-3101 Jacob Burba1 0000-0002-8465-9341 Hugh Garsden1 0009-0001-3949-9342 Michael J. Wilensky1 1Jodrell Bank Centre for Astrophysics 0000-0001-7716-9312 University of Manchester Manchester M13 9PL UK 2Department of Physics and Astronomy University of Western Cape Cape Town 7535 South Africa (Accepted XXX. Received YYY; in original form ZZZ; 2024) ###### Abstract An important application of next-generation wide-field radio interferometers is making high dynamic range maps of radio emission. Traditional deconvolution methods like CLEAN can give poor recovery of diffuse structure, prompting the development of wide-field alternatives like Direct Optimal Mapping and $m$-mode analysis. In this paper, we propose an alternative Bayesian method to infer the coefficients of a full-sky spherical harmonic basis for a drift-scan telescope with potentially thousands of baselines. The can precisely encode the uncertainties and correlations between the parameters used to build the recovered image. We use Gaussian Constrained Realisations (GCR) to efficiently draw samples of the spherical harmonic coefficients, despite the very large parameter space and extensive sky-regions of missing data. Each GCR solution provides a complete, statistically-consistent gap-free realisation of a full- sky map conditioned on the available data, even when the interferometer’s field of view is small. Many realisations can be generated and used for further analysis and robust propagation of statistical uncertainties. In this paper, we present the mathematical formalism of the spherical harmonic GCR- method for radio interferometers. We focus on the recovery of diffuse emission as a use case, along with validation of the method against simulations with a known diffuse emission component. ###### keywords: reionisation, observations, large scale structure of universe – methods: statistical – techniques: interferometric ## 1 Introduction Through measurements of the Cosmic Microwave Background (CMB), significant progress has been made in studying the very early Universe (Planck Collaboration et al., 2020), but the subsequent epochs up to and including the Epoch of Reionisation still leaves much to be discovered. A promising method to explore these intermediate epochs of cosmic history is the redshifted 21 cm line from neutral hydrogen. The signal can be used as a cosmological probe of structure formation, as it traces the distribution and evolution of the neutral hydrogen that fills the early Universe and forms the first stars and galaxies during Cosmic Dawn and also provides an inverse tracer of the reionisation of the intergalactic medium during the later Epoch of Reionisation (EoR) (Pritchard & Loeb, 2012; Liu & Shaw, 2020). Importantly, this tracer is redshift dependent, providing line-of-sight distance information that adds an extra dimension compared to the CMB, (Furlanetto et al., 2006; Morales & Wyithe, 2010; Liu & Shaw, 2020). Studies of the 21 cm signal are often divided into either measuring the spatially averaged _global signal_ , or measuring the statistical properties of the spatial fluctuations in the brightness temperature field. The 21 cm signal from the above-mentioned epochs is redshifted to the range $z\sim 27-6$, corresponding to frequencies of $50-$200\text{\,}\mathrm{MHz}$$. A number of experiments have been built with the purpose of searching for this signal, and many have already reported upper limits. A non-exhaustive list includes the Murchison Widefield Array; MWA (Bowman et al., 2013; Tingay et al., 2013; Wayth et al., 2018), the Donald C. Backer Precision Array to Probe the Epoch of Reionisation; PAPER (Parsons et al., 2012; Ali et al., 2015), The LOw Frequency Array; LOFAR (van Haarlem et al., 2013; Patil et al., 2017), and the Hydrogen Epoch of Reionization Array; HERA (DeBoer et al., 2017; The HERA Collaboration et al., 2022). The biggest technical challenge to date is the proper handling of foreground contaminants and how they are distorted by the receiving instrumentation. The dominant contaminant is typically the diffuse emission that is present at all parts of the sky, and is roughly a factor of around $10^{5}$ times brighter than the underlying 21 cm signal (Liu & Shaw, 2020). The main component of diffuse emission is the Galactic synchrotron emission, but there are also contributions from extragalactic free-free emission, bright Galactic radio point sources, and unresolved extragalactic point sources, (Santos et al., 2005). In addition there may be as-yet unknown diffuse components, e.g. associated with apparent excess background emission as claimed by ARCADE 2 (Fixsen et al., 2011; Singal et al., 2010) and OVRO-LWA (Dowell & Taylor, 2018), or potential dark matter annihilation signals (Evoli et al., 2014; Lopez-Honorez et al., 2016). Being able to robustly separate known but uncertain Galactic and extra-galactic components from novel sources of emission would greatly aid the physical interpretation of all of these effects, particularly as highly sensitive but complex next-generation arrays such as SKAO come online. The issue of modelling diffuse emission is especially important for close- packed radio arrays. Sparse interferometers such as the VLA or SKA-MID are essentially blind to large-scale emission, as the antenna configuration resolves out large angular scales. Compact arrays with smaller antennas are designed to make high dynamic range measurements on large angular scales however, which means the bright diffuse emission becomes an unavoidable issue. Imaging with interferometers is complicated, particularly for diffuse emission, as the effect of the interferometer response (its point spread function; PSF) must be deconvolved. For regular close-packed arrays, there is often highly incomplete coverage of the $(u,v)$-plane, since there are many baselines of the same length. This ‘redundancy’ of the baselines results in artefacts like grating lobes in the PSF (Dillon & Parsons, 2016). The typical lack of uniform $(u,v)$-coverage has given rise to the use of specialised deconvolution algorithms, where the so-called _dirty image_ is deconvolved to result in a more uniform image with imaging artifacts removed or suppressed. Many traditional pipelines make use of the CLEAN algorithm, which works by iteratively removing and replacing bright sources in the dirty image with their calculated point spread functions (PSFs) to get rid of the point source side lobes, (Högbom, 1974; Cornwell, T. J., 2009). CLEAN, however, does not correct for diffuse emission and additionally the statistics of the resulting image are not well known, particularly as CLEAN acts as a non-linear transformation of the data. Some improvements have been made to the CLEAN algorithm through Multi-Scale CLEAN (Cornwell, 2008) and WSCLEAN (Offringa & Smirnov, 2017), which can be used in either (or both) multi-scale and multi-frequency mode. Frequency dependence is introduced by dividing the full bandwidth into several output channels and then treating them individually. In multi-scale mode, further iterations are introduced using sub-loops and going through the scales one-by-one by convolving the dirty image with the corresponding space-kernel. Even so, both methods still build upon the same CLEAN principle of using maxima to replace sources with an ideal PSF response. An alternative approach is Direct Optimal Mapping (DOM), which uses a maximum likelihood method to estimate the beam-weighted map of the sky using a pixel- basis, (Dillon et al., 2015; Xu et al., 2022). DOM is thus a one-shot process that estimates both the mean and the covariance in a lossless way, i.e. without losing information on model parameters. The model is built around a measurement matrix $\boldsymbol{\mathbf{A}}$ that maps a pixelated sky map to the visibilities. The choice of $\boldsymbol{\mathbf{A}}$ can be problematic however, and regularisation is needed in order to overcome degeneracies and ensure uniqueness of the map solutions that are obtained by ‘inverting’ the measuring matrix. Furthermore, the noise covariance $\boldsymbol{\mathbf{N}}$ is also affected by the measurement matrix, requiring a choice for how to beam-weight the recovered map. While in principle the DOM formulation accounts for the whole sky, much smaller faceted maps are typically used to reduce the computational costs. Like DOM, $m$-mode analysis is also typically implemented as a maximum likelihood estimator that will produce an estimate of the mean and covariance of the sky, although now in a spherical harmonic basis instead (Shaw et al., 2014, 2015; Eastwood et al., 2018). This method is developed specifically for drift-scan telescopes, taking advantage of simplifications that arise after applying a spherical transform to both the sky intensity and beam function, as well as performing a Fourier transform of the visibilities along the time direction, into $m$-mode space. As with DOM, proper care must be taken to ensure uniqueness of the solutions, so the inverse of the beam transfer function is replaced with the Moore-Penrose pseudo-inverse. Both DOM and $m$-mode analysis are maximum likelihood estimators, and therefore only provide moments of the underlying posterior distribution of the sky model parameters (although see Chapter 6 of Kriele (2022), which recasts $m$-mode analysis as a Bayesian linear model). In this paper, we introduce a statistical method that, like $m$-mode analysis, is based on a spherical harmonic description of the sky (although we do not use an $m$-mode transform here). We use an explicitly Bayesian treatment of the map reconstruction problem, introducing a prior to the system which provides a regularisation, particularly in regions of missing data, and using the Gaussian Constrained Realisation (GCR) method to make sampling of the large number of spherical harmonic coefficients tractable. Each sample vector drawn from the posterior is a valid realisation of the whole sky, without any missing data regions, and we can treat it like we would treat ‘ideal’ data without noise etc. Uncertainties can be propagated by passing multiple realisations through subsequent processing steps and then inspecting their distribution. By using an explicit forward model, we can also avoid ad-hoc beam weighting steps, which makes interpretation of the results more straightforward. Furthermore, this sampler can be incorporated in a larger ‘Gibbs sampling’ framework that also samples other sky and instrumental parameters (e.g. Eriksen et al., 2008; Kennedy et al., 2023; CHIME Collaboration, 2023). Naturally, calculating multiple samples will require more computational power than solving once for the maximum likelihood (or for the maximum _a posteriori_) solution. We propose some approaches to reduce the computational burden, such as specialising to the drift-scan telescope case and using Wigner $\mathfrak{D}$-matrices to account for the sky rotating above the instrument. This removes the need to explicitly calculate the visibility response function (which maps the spherical harmonic coefficients to the visibilities) at multiple times. The paper is structured as follows: In Section 2 we build and present the visibility response model and cover the Bayesian methods of Wiener filtering and Gaussian Constrained realisations. Both methods are related to Gibbs sampling, a future prospect of this work. Section 3 describes the diffuse emission foreground model as well as the visibility simulation- and sampling parameters. In Section 4 we present the performance for our reference (standard) case as well as a comparative analysis of various noise- and prior levels and ten other simulation scenarios including varying the field of view of the array. Finally, in Section 5 we conclude. ## 2 Bayesian sampling of diffuse foregrounds In this section we develop the mathematical formalism to statistically sample spherical harmonic modes on the whole sky given radio interferometer visibility data. ### 2.1 Data model The data model in this work is based on a single diffuse component modelled as a sum of spherical harmonic modes at each frequency. Ultimately, this will only be one component of a more comprehensive sky model involving contributions from point sources, the EoR signal, etc. (which do not need to be modelled using a spherical harmonic basis). It would also be possible to define a particular functional form for the frequency dependence and interpret the spherical harmonic coefficients as amplitudes of this spectral template, e.g. by defining the specific intensity of the sky as $\displaystyle I(\nu,\theta,\phi)$ $\displaystyle=\sum_{\ell,m}a_{\ell m}Y_{\ell m}(\theta,\phi)f(\nu)$ (1) with frequency dependence $f(\nu)$. A common choice of frequency dependence is a power law with spectral index $\beta$. These scenarios involve only simple modifications to the single component (and per-frequency) model that we consider here however, and so for the sake of simplicity we will not consider them further here. We begin by writing out an expression for the complex visibilities observed by an interferometer. A baseline separated by vector $\boldsymbol{\mathbf{b}}$ given by antennas $i,j$ roughly probes the sky brightness temperature at Fourier mode $\boldsymbol{\mathbf{u}}=\boldsymbol{\mathbf{b}}/\lambda$, (Liu & Shaw, 2020). The visibility that the interferometer measures by that baseline $\boldsymbol{\mathbf{b}}$ is then given by, $\displaystyle V_{ij}(\nu,t)=\int\text{d}^{2}\Omega\,A_{i}(\nu,\boldsymbol{\mathbf{\theta}})A_{j}^{*}(\nu,\boldsymbol{\mathbf{\theta}})\,I(\nu,\boldsymbol{\mathbf{\theta}})\,e^{-2\pi\boldsymbol{\mathbf{u}}(\nu)\cdot\boldsymbol{\mathbf{\theta}}}+n_{ij},$ (2) where $A_{i}$ and $A_{j}^{*}$ are the E-field beams for each antenna, $I$ is the specific intensity of the sky, the exponential term describes the fringes where $\boldsymbol{\mathbf{\theta}}$ is the topocentric coordinates of the sources, and $n_{ij}$ is the baseline dependent noise. The per-frequency spherical harmonic expansion of the sky model is given as, $\displaystyle I(\nu,\theta,\phi)=\sum_{\ell,m}a_{\ell m}(\nu)Y_{\ell m}(\theta,\phi),$ (3) where $\theta$ and $\phi$ are the declination and right ascension in equatorial coordinates respectively. For a typical drift scan array the motion of the sky is especially simple in an equatorial coordinate system, as it is a simple right ascension (or LST) rotation. There are many different conventions on which $a_{\ell m}$ modes to include, when working with spherical harmonics. For the sake of reducing computational resources, it makes sense to manipulate the $a_{\ell m}$-vector to contain only the minimal required modes to solve the system. Firstly, it should be noted that even though the visibilities are complex entities, the actual sky needs to be a real field. The following symmetry relation for the $a_{\ell m}$ modes must be satisfied, $\displaystyle{a_{\ell m}}=(-1)^{m}a^{*}_{l,-m},$ (4) or, split into real and imaginary parts, $\displaystyle a_{\ell,+m}^{\rm re}=(-1)^{m}a_{\ell,-m}^{\rm re}\quad{\rm and}\quad a_{\ell,+m}^{\rm im}=(-1)^{m+1}a_{\ell,-m}^{\rm im}\,.$ (5) This means that only $m\geq 0$ modes are necessary and all negative $m$ modes are naturally excluded from the $a_{\ell m}$-vector. Generally, the spherical harmonic coefficients are complex-valued. This can present complications when dealing with vectors of $a_{\ell m}$-values numerically. We therefore split the $a_{\ell m}$-values into their real- and imaginary parts. The $a_{\ell m}$-vector now consists first of all the real- and then the imaginary modes. Moreover, since the $m=0$ imaginary parts will always be zero in order to satisfy the (anti-)symmetry condition of Eq. 5, these spherical harmonic modes are removed while sampling and injected back in afterwards. For a given $\ell_{\text{max}}$ this leads to a total number of $a_{\ell m}$ modes of $N_{\text{modes}}=({\ell_{\text{max}}}+1)^{2}$. We next define the visibility response operator, $\delta V_{ij}^{\ell m}(\nu,t)$, which gives the projection from a spherical harmonic vector to a set of radio interferometer visibilities. The visibility response is dependent on the antenna array configuration and location, which means the primary beam function is also absorbed into this operator. The full visibility will then be the visibility response function applied to the $a_{\ell m}$ modes summed over all $(\ell,m)$-values, $\displaystyle V_{ij}=\sum_{\ell m}\delta V_{ij}^{\ell m}(\nu,t)\,a_{\ell m}.$ (6) where the visibility response is defined as, $\displaystyle\delta V_{ij}^{\ell m}(\nu,t)=\int\text{d}^{2}\Omega\,A_{i}(\nu,\boldsymbol{\mathbf{\theta}})A_{j}^{*}(\nu,\boldsymbol{\mathbf{\theta}})\,Y_{\ell m}(\alpha,\delta)\,e^{-2\pi\boldsymbol{\mathbf{u}}_{ij}(\nu)\cdot\boldsymbol{\mathbf{\theta}}}.$ (7) In this expression, $\boldsymbol{\mathbf{\theta}}$ are in topocentric coordinates (e.g. local altitude/azimuth), and $\alpha=\alpha(\boldsymbol{\mathbf{\theta}},t)$ and $\delta=\delta(\boldsymbol{\mathbf{\theta}},t)$ are the RA and Dec coordinates corresponding to a given topocentric pointing $\boldsymbol{\mathbf{\theta}}$ at local sidereal time $t$. This operator can be computed ahead of time for a given array configuration, which determines the available baseline vectors $\mathbf{u}$ and set of antenna E-field beams $A_{i}$. ### 2.2 Wigner $\mathfrak{D}$-matrix formalism In the above section, the visibility response is defined as a function of both time and frequency. Instead of simulating the visibility response operator for each time separately, we can choose a single reference time and apply a rotation to get the response at any other desired LST. This is made simpler by choosing the spherical harmonic basis to align with the rotation axis of the sky as seen by a drift-scanning telescope, i.e. by defining the spherical harmonic basis in equatorial coordinates. In this case, a simple azimuthal rotation around the celestial axis implements the mapping between RA and LST, while the declination stays constant. We begin by defining the rotational matrix $\mathcal{R}_{\ell m\ell^{\prime}m}$ that transforms the spherical harmonic modes at a reference sidereal time $t_{\text{ref}}$ to the updated sidereal time, $\displaystyle a_{\ell m}(t)$ $\displaystyle=\mathcal{R}_{\ell m\ell^{\prime}m^{\prime}}(t)\,a_{\ell^{\prime}m^{\prime}}(t_{\text{ref}}).$ (8) For spherical harmonics, this is given by the Wigner $\mathfrak{D}$-matrix, which is a unitary matrix in an irreducible representation of the SO(3) group (Tung, 1985). The spherical harmonics transform as $\displaystyle Y_{\ell}^{m}(\theta^{\prime},\phi^{\prime})$ $\displaystyle=\sum_{m^{\prime}=-\ell}^{\ell}Y_{\ell}^{m^{\prime}}(\theta,\phi)\,{\mathfrak{D}}_{m^{\prime}m}^{\ell}(\alpha,\beta,\gamma),$ (9) where ${\mathfrak{D}}_{mm^{\prime}}^{\ell}(\alpha,\beta,\gamma)$ is the $\mathfrak{D}$-matrix given by the three Euler angles (that describe sequential rotations around three axes). Note, that for any rotation with $\mathfrak{D}$-matrices, the spherical harmonic of degree $\ell$ and order $m$ transforms into a linear combination of spherical harmonics to the same degree $\ell$. The $\mathfrak{D}$-matrix itself is given as $\displaystyle{\mathfrak{D}}^{\ell}_{m^{\prime}m}(\alpha,\beta,\gamma)$ $\displaystyle=e^{-im^{\prime}\alpha}\,d^{\ell}_{m^{\prime}m}(\beta)\,e^{-im\gamma},$ (10) where $d^{\ell}_{m^{\prime}m}(\beta)$ are the corresponding reduced Wigner matrices. A table of the (small) $d$-matrices can be found in Dong (1994). Note the lack of mixing between $\ell$ modes. For a zenith-pointing drift-scan telescope, the sky rotation in a time interval $t-t_{\rm ref}$ can be mapped directly to an azimuthal rotation angle. The full visibility given in Eq. 6 can then be written as, $\displaystyle V_{ij}(\nu,t)=\sum_{\ell m}\delta V_{ij}^{\ell m}(\nu,t_{\text{ref}})\,\sum_{m^{\prime}}{\mathfrak{D}}^{\ell}_{m\,m^{\prime}}(t)\,a_{\ell m^{\prime}}(t_{\rm ref}),$ (11) where $\delta V_{ij}^{\ell m}(\nu,t_{\text{ref}})$ is the pre-computed visibility response at a reference time $t_{\text{ref}}$, ${\mathfrak{D}}^{\ell}_{m\,m^{\prime}}(t)$ is the appropriate $\mathfrak{D}$-matrix for a given LST, and $a_{\ell m^{\prime}}$ are the spherical harmonic modes of the sky at $t_{\text{ref}}$. If we let $p$ and $q$ label the frequency and time samples, $\nu_{p}$ and $t_{q}$, we can rewrite the above in index notation as $\displaystyle V_{ijpq}={X}_{ijpq\ell m}\,\,a_{\ell m},$ (12) where repeated indices denote summation, and ${X}_{ijpq\ell m}$ is the combined operator for the two sums in Eq. 11. The combined visibility response and $\mathfrak{D}$-matrices now form the full $\boldsymbol{\mathbf{X}}$-operator, which is linear and equivalent to that of Eq. 6. ### 2.3 Wiener filter – the maximum a posteriori solution Now, having established our visibility model based on the linear operator $\boldsymbol{\mathbf{X}}$ for the visibility response using a spherical harmonics basis — we want to obtain the posterior distribution for the spherical harmonic modes given the data, noise, and priors, in order to generate samples of the diffuse emission sky. The steps outlined in this section and the next are steps of increasing complexity in a Bayesian hierarchy, ultimately preparing for implementing this model into a full Gibbs sampling scheme. Here, we go from the Wiener filter solution; a simple maximum a posteriori (MAP) solution, to drawing independent samples from the Gaussian Constrained Realisation (GCR) equation. Thereby, as a first step in the Bayesian hierarchy we find the Wiener filter solution, which also acts as the central value for drawing the realisations. From Bayes’s theorem we can get the conditional distribution on the $a_{\ell m}$ modes, $\displaystyle P(\boldsymbol{\mathbf{a}}\,|\,\boldsymbol{\mathbf{d}},\boldsymbol{\mathbf{N}},\boldsymbol{\mathbf{a}}_{0},\boldsymbol{\mathbf{S}})\propto\,$ $\displaystyle P(\boldsymbol{\mathbf{d}}|\,\boldsymbol{\mathbf{a}},\boldsymbol{\mathbf{N}})\,P(\boldsymbol{\mathbf{a}}|\,\boldsymbol{\mathbf{a}}_{0},\boldsymbol{\mathbf{S}}),$ (13) when conditioning on the known covariances of the signal $\boldsymbol{\mathbf{S}}$ and noise $\boldsymbol{\mathbf{N}}$, the data-vector $\boldsymbol{\mathbf{d}}$, and the mean of the prior on the $a_{\ell m}$ modes $\boldsymbol{\mathbf{a}}_{0}$. For simplification of the notation we have dropped the $(\ell,m)$-indices and simply denote the $a_{\ell m}$-vector as $\boldsymbol{\mathbf{a}}$ as well as dropping the baseline indices ($i,j$). It is assumed that generally the signal prior and covariance will be independent of the data and noise covariance. For simplicity we are keeping $\boldsymbol{\mathbf{S}}$ fixed, however, in a full Gibbs sampling scheme it would be possible to sample $\boldsymbol{\mathbf{S}}$ as well. Furthermore, we assume that the noise is Gaussian thus rewriting Eq. 13 as: $\displaystyle P(\boldsymbol{\mathbf{a}}\,|\,\boldsymbol{\mathbf{d}},\boldsymbol{\mathbf{N}},\boldsymbol{\mathbf{a}}_{0},\boldsymbol{\mathbf{S}})\propto e^{-(\boldsymbol{\mathbf{d}}-\boldsymbol{\mathbf{Xa}})^{\dagger}\boldsymbol{\mathbf{N}}^{-1}(\boldsymbol{\mathbf{d}}-\boldsymbol{\mathbf{Xa}})}e^{-(\boldsymbol{\mathbf{a}}-\boldsymbol{\mathbf{a}}_{0})^{\dagger}\boldsymbol{\mathbf{S}}^{-1}(\boldsymbol{\mathbf{a}}-\boldsymbol{\mathbf{a}}_{0})}.$ (14) For a Gaussian distribution the maximum of the posterior distribution is the same as the maximum of the log-posterior, hence we can find the Wiener filter solution by maximising the partial derivative of the exponent of Eq. 14 with respect to the signal, $\boldsymbol{\mathbf{a}}$: $\displaystyle\left.\frac{\partial}{\partial\boldsymbol{\mathbf{a}}}\right|_{\boldsymbol{\mathbf{a}}=\boldsymbol{\mathbf{\hat{a}}}}\left(-(\boldsymbol{\mathbf{d}}-\boldsymbol{\mathbf{Xa}})^{\dagger}\boldsymbol{\mathbf{N}}^{-1}(\boldsymbol{\mathbf{d}}-\boldsymbol{\mathbf{Xa}})-(\boldsymbol{\mathbf{a}}-\boldsymbol{\mathbf{a}}_{0})^{\dagger}\boldsymbol{\mathbf{S}}^{-1}(\boldsymbol{\mathbf{a}}-\boldsymbol{\mathbf{a}}_{0})\right)=0,$ (15) resulting in, $\displaystyle\boldsymbol{\mathbf{d}}^{\dagger}\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{X}}+\boldsymbol{\mathbf{a_{0}}}^{\dagger}\boldsymbol{\mathbf{S}}^{-1}=\boldsymbol{\mathbf{a}}^{\dagger}\boldsymbol{\mathbf{X}}^{\dagger}\boldsymbol{\mathbf{N}}^{-1}+\boldsymbol{\mathbf{a}}^{\dagger}\boldsymbol{\mathbf{S}}^{-1}.$ (16) Now, taking advantage of the fact that covariance matrices are Hermitian, we can complex conjugate the entire expression and rearrange to obtain the Wiener filter solution, $\displaystyle\left[\boldsymbol{\mathbf{X}}^{\dagger}\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{X}}+\boldsymbol{\mathbf{S}}^{-1}\right]\boldsymbol{\mathbf{a}}_{\textup{wf}}=\left(\boldsymbol{\mathbf{X}}^{\dagger}\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{d}}+\boldsymbol{\mathbf{S}}^{-1}\boldsymbol{\mathbf{a}}_{0}\right).$ (17) Returning to Eq.14; multiplying two multivariate Gaussians will result in a new multivariate Gaussian. By completing the square it can be shown that this can be written as proportional to a multivariate Gaussian with inverse covariance matrix given as $\displaystyle\boldsymbol{\mathbf{\Sigma}}^{-1}=\boldsymbol{\mathbf{X}}^{\dagger}\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{X}}+\boldsymbol{\mathbf{S}}^{-1},$ (18) and where the Wiener filter solution acts as the mean, $\displaystyle\boldsymbol{\mathbf{\hat{a}}}=\boldsymbol{\mathbf{\Sigma}}\left(\boldsymbol{\mathbf{X}}^{\dagger}\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{d}}+\boldsymbol{\mathbf{S}}^{-1}\boldsymbol{\mathbf{a}}_{0}\right).$ (19) Even though the Wiener filter is the maximum a posteriori solution, it is generally a biased estimator as described in more detail in Kennedy et al. (2023). Furthermore, the Wiener filter is a summary statistic but we are interested in generating actual samples that are complete and statistically- consistent realisations of the full-sky map. For this, the Wiener filter can instead act as the central value for drawing the samples. Figure 1: The absolute values of the visibility response operator $X_{\text{re}}$ on a log-scale and with dimensions (Nmodes) $\times$ (NLSTs $\times$ Nfreq $\times$ Nbl). Here it is shown for $\ell_{\text{max}}=20$, LST $=$ 0–8 $\mathrm{h}$, $\nu=$ 100–101 $\mathrm{MHz}$ and 10 close packed antennas that form 45 baselines as shown in Fig. 2. The darker is is, the less that frequency/LST/baseline contributes to measuring that $(m,\ell)$-value. Note that this is the _real-part only_ of $X$ (the imaginary part showing similar structure) but it covers both the real- and imaginary parts of the $a_{\ell m}$ modes. ### 2.4 Gaussian Constrained Realisations In order to draw samples from the full conditional distribution $P(\boldsymbol{\mathbf{a}}\,|\,\boldsymbol{\mathbf{d}},\boldsymbol{\mathbf{N}},\boldsymbol{\mathbf{a}}_{0},\boldsymbol{\mathbf{S}})$, we can take advantage of the fact that we can describe our model as a multivariate Gaussian distribution. The Wiener filter solution described by its mean and covariance given in eqs. 18 and 19 acts as a first step in the hierarchy to which we can add random normal realisation terms of the signal $\boldsymbol{\mathbf{\omega}}_{a}$ and noise terms $\boldsymbol{\mathbf{\omega}}_{d}$ correctly scaled by their respective covariances to draw the constrained realistations $\boldsymbol{\mathbf{a}}_{\textup{cr}}$, (Eriksen et al., 2008). This leads to the Gaussian Constrained Realisation (GCR) equation, $\displaystyle\left[\boldsymbol{\mathbf{X}}^{\dagger}\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{X}}+\boldsymbol{\mathbf{S}}^{-1}\right]\boldsymbol{\mathbf{a}}_{\textup{cr}}=\left(\boldsymbol{\mathbf{X}}^{\dagger}\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{d}}+\boldsymbol{\mathbf{S}}^{-1}\boldsymbol{\mathbf{a}}_{0}+\boldsymbol{\mathbf{S}}^{-\frac{1}{2}}\boldsymbol{\omega}_{a}+\boldsymbol{\mathbf{X}}^{\dagger}\boldsymbol{\mathbf{N}}^{-\frac{1}{2}}\boldsymbol{\mathbf{\omega}}_{d}\right).$ (20) In turn Eq. 20 reduces back to the mean $\langle\boldsymbol{\mathbf{a}}_{\textup{cr}}\rangle=\boldsymbol{\mathbf{\hat{a}}}$, since the unit variance Gaussian vectors $\boldsymbol{\mathbf{\omega}}_{a}$ and $\boldsymbol{\mathbf{\omega}}_{d}$ have zero mean. By solving Eq. 20 repeatedly we can draw independent samples of the $a_{\ell m}$ modes on the sky consistent with the given data vector $\boldsymbol{\mathbf{d}}$, chosen signal prior $\boldsymbol{\mathbf{a}}_{0}$, the visibility response operator $\boldsymbol{\mathbf{X}}$, and the covariances $\boldsymbol{\mathbf{S}}$ and $\boldsymbol{\mathbf{N}}$. For the sake of keeping computational costs low, we have kept to 100 samples per set of parameters for this paper. In the case studied in this paper, there is only missing data outside of the FOV defined by the primary beam and the horizon; both of which is dealt with through the visibility response function. In case of actual gaps in the data, for instance due to RFI flagging, one can define a set of weights $\boldsymbol{\mathbf{w}}$ to redefine the inverse noise covariance, $\displaystyle\boldsymbol{\mathbf{N}}_{\boldsymbol{\mathbf{w}}}^{-1}=\boldsymbol{\mathbf{ww}}^{T}\circ\boldsymbol{\mathbf{N}}^{-1},$ (21) where $\circ$ denotes element-wise multiplication. Using this weighted noise covariance in Eq. 20 means we automatically zero the contribution from the data inside the mask. The signal covariance, however, does not necessarily go to zero inside the masked regions and thereby _takes over_ the signal estimation in the absence of information from the likelihood function. Note that this is not equivalent to drawing samples from the prior distribution and simply filling-in the masked regions. The samples generated within the flagged regions will be informed by both the prior and data from the un-flagged regions. In the future this process will be done within the context of Gibbs sampling (Geman & Geman, 1984), which is a Bayesian method to sample directly from the joint posterior of more complicated high-dimensional distributions. In this instance, we focus only on the distribution for the spherical harmonic coefficients, and use a conjugate gradient solver to obtain $\boldsymbol{\mathbf{a}}_{\textup{cr}}$ from the GCR equation. Most solvers do not handle complex quantities well so to overcome this, we saw it necessary to ‘realify’ the full system as described in App. A. When implemented into a Gibbs sampling scheme, it will enable us to also sample the signal covariance $\boldsymbol{\mathbf{S}}$. Gibbs sampling is an iterative method, that samples from each conditional distribution in turn and thereby effectively samples from the joint posterior. For each iteration it updates the conditional variables with the samples obtained from the previous step, $\displaystyle\boldsymbol{\mathbf{a}}_{i+1}$ $\displaystyle\leftarrow P\\!\left(\boldsymbol{\mathbf{a}}_{i}\,|\,\boldsymbol{\mathbf{d}},\boldsymbol{\mathbf{N}},\boldsymbol{\mathbf{a_{0}}},\boldsymbol{\mathbf{S}}_{i}\right),$ (22a) $\displaystyle\boldsymbol{\mathbf{S}}_{i+1}$ $\displaystyle\leftarrow P\\!\left(\boldsymbol{\mathbf{S}}_{i}\,|\,\boldsymbol{\mathbf{a}}_{i+1}\right),$ (22b) where $\leftarrow$ represents generating samples through evaluating the conditional distribution and $i$ indexes the Gibbs iteration number. First, the ${a_{\ell m}}$ modes are sampled and it is noticed that the conditional distribution takes the same form as in Eq. 13. We can therefore use the same arguments as before; we are still dealing with a multivariate Gaussian distribution and therefore end up at the GCR equation of Eq. 20. The sample obtained from Eq. 22a already contains the conditioning on the noise covariance $\boldsymbol{\mathbf{N}}$, data $\boldsymbol{\mathbf{d}}$, and prior mean $\boldsymbol{\mathbf{a_{0}}}$, so there is no need to condition again on these parameters when sampling $\boldsymbol{\mathbf{S}}$ in Eq. 22b. We leave a full exploration of sampling the signal covariance (which encodes the spherical harmonic angular power spectra) to later work. Figure 2: The standard array layout used in the majority of the simulations, consisting of 10 closed packed antennas with a diameter of $14\text{\,}\mathrm{m}$ and a separation of $14.6\text{\,}\mathrm{m}$. ## 3 Simulations and sky model The structure of the simulated visibility response function has already been laid out in Sec. 2.1 where it is also made clear that the simulation is dependent on which LSTs are chosen, $\ell_{\text{max}}$, frequency $\nu$, baseline orientation and length $b$, and dish size $\theta_{D}$. This section will describe how the simulations were obtained as well as the standard parameters used for the reference results, which from here on will be referred to as the _standard_ case. More realisations have been drawn using a variation of both the simulation/observation parameters as well as samples from using a variation of noise and prior levels. The details of these parameter variations are explained in Sections 4.2–4.4. ### 3.1 Precomputing the visibility response operator To simulate the visibility response function we used the hydra111https://github.com/HydraRadio/ code, which builds upon matvis (Kittiwisit et al., 2023) to simulate the visibility per $a_{\ell m}$-mode as well as per baseline, frequency, and LST. The absolute values of the real-part of the visibility response operator, $\boldsymbol{\mathbf{X}}_{\rm re}$, is shown in Fig. 1 also showing how the operator’s dimensions are packed. The visibility response is only simulated for Stokes-I. We have assumed the most general case where we want to solve for a different set of ${a_{\ell m}}$ modes for each frequency channel. The simulations cover the narrow bandwidth of $100$–$102\text{\,}\mathrm{MHz}$ in just two channels, in order to keep the problem small for the time being. The LST range spans $0$–$8\text{\,}\mathrm{h}$ in 10 steps. The primary beam model is calculated using pyuvsim222https://github.com/RadioAstronomySoftwareGroup/pyuvsim which provides an analytic Gaussian beam model assumed to be identical for each antenna, with the width $\theta\simeq\lambda/b$ where $b$ is length of the baseline $\boldsymbol{\mathbf{b}}$. As standard the distance between each antenna is $14.6\text{\,}\mathrm{m}$ (HERA-like) and the dishes themselves are modelled as hyperbolic dishes of diameter $\theta_{\text{D}}=$14\text{\,}\mathrm{m}$$ with no side-lobes. The model is simple, but has been specifically chosen so that we can clearly understand and validate the results. The impact of different beams is an interesting study in its own right, which we defer to future work. The standard array layout used in the simulations is a small closed-packed redundant array to represent a subsection of the full HERA array. A redundant layout benefits from sampling the same modes many times, thus increasing sensitivity especially to the relatively large angular scales relevant for EoR experiments. In this configuration we consider only 10 receivers as shown in Fig. 2. The subset is sufficiently small to both keep computational costs low and with its resulting 45 baselines it still offers multiple redundant baseline groups and variation in both direction and baseline length. As with HERA the array used in the simulations is a drift-scan instrument pointing towards zenith placed at latitude $\sim$-31\text{\,}\mathrm{\SIUnitSymbolDegree}$$. From Fig. 1 it is already clear that some baselines contribute more than others. For instance the baseline of antennas $(3,6)$ is not very responsive, which is the long E-W $43.8\text{\,}\mathrm{m}$ baseline (Fig. 2) whereas all the short $14.6\text{\,}\mathrm{m}$ baselines contribute significantly more. ### 3.2 Diffuse emission sky model As we aim to sample the diffuse emission foregrounds specifically, no other foregrounds are included in the foreground model. For the purely diffuse emission foreground sky model we use the Global Sky Model (GSM2016; Zheng et al., 2017) as implemented by pyGDSM (Price, 2016). GSM2016 is based on principal component analysis of a large set of multi-frequency datasets (spanning $10\text{\,}\mathrm{MHz}$ to $5\text{\,}\mathrm{THz}$) as well as performing a blind separation into physical components revealing five components identified as synchrotron emission, free-free emission, cold dust, warm dust, and the CMB anisotropy. The sky model is provided as frequency dependent HEALpix maps where we use only the GSM2016 map at reference frequency $100\text{\,}\mathrm{MHz}$. The foreground map is then passed to HEALpy to get the true $a_{\ell m}$ modes, $\boldsymbol{\mathbf{a}}_{\text{true}}$. As standard, the resolution is set to nside $=128$, which corresponds to a HEALPix angular pixel size of $0.46\text{\,}\mathrm{\SIUnitSymbolDegree}$. The maximum mode of the operator is set to ${\ell_{\text{max}}}=20$, however, which corresponds to $\sim$9\text{\,}\mathrm{\SIUnitSymbolDegree}$$. The true sky is used as input for the full data model, which is obtained by applying the visibility response to the true $a_{\ell m}$ modes. ### 3.3 Data and noise model The data input for the analysis in this work is based on the visibility response simulations. In the future, this will be done with real visibility data instead, but for now simulations with a known diffuse sky component serve as a means of validation. Here, we use the visibility response operator $\boldsymbol{\mathbf{X}}$ to map the true-sky spherical harmonic coefficient vector $\boldsymbol{\mathbf{a}}_{\rm true}$ into visibilities and add a noise vector $\boldsymbol{\mathbf{n}}$ to the simulated visibilities, $\displaystyle\boldsymbol{\mathbf{d}}=\boldsymbol{\mathbf{X}}\boldsymbol{\mathbf{a}}_{\text{true}}+\boldsymbol{\mathbf{n}}.$ (23) The noise on the data model is added as uncorrelated Gaussian random noise, $\displaystyle\boldsymbol{\mathbf{n}}\sim\mathcal{N}(0,\boldsymbol{\mathbf{N}}),$ (24) given by the noise covariance $\boldsymbol{\mathbf{N}}$ with components $N_{ij}$ modelled by the simulated auto-correlation visibilities $V_{ii}$ and $V_{jj}$ as given by the radiometer equation, $\displaystyle N_{ij}=\sigma_{ij}^{2}=\frac{V_{ii}V_{jj}}{\Delta t\Delta\nu}.$ (25) Since the foregrounds are spectrally smooth, there is no need for very high spectral resolution and we choose $\Delta\nu=$1\text{\,}\mathrm{MHz}$$. The time resolution is set to $\Delta t=$60\text{\,}\mathrm{s}$$, both to avoid issues with sky rotation (smearing) and to still ensure good signal-to-noise ratio. Lastly, we have applied a $10\%$ prior on the ${a_{\ell m}}$ values. To implement this, we have defined a prior covariance $\boldsymbol{\mathbf{S}}$ that is diagonal, with values ${S}_{nm}=(0.1\,a_{\text{true},n})^{2}\delta_{nm}$. The prior mean, $\boldsymbol{\mathbf{a}}_{0}$, is not set equal to $\boldsymbol{\mathbf{a}}_{\text{true}}$ however, as this would effectively be inputting information about the correct answer into the inference ahead of time, which is not realistic. Instead, we draw values for $\boldsymbol{\mathbf{a}}_{0}$ from a Gaussian distribution centered on the true $a_{\ell m}$ values, with the same prior covariance, i.e. $\boldsymbol{\mathbf{a}}_{0}\sim\mathcal{N}(\boldsymbol{\mathbf{a}}_{\text{true}},\boldsymbol{\mathbf{S}})$. This ensures that the prior mean that we input in the GCR equation (Eq. 20) is set consistently with the chosen prior model (GSM2016), while deviating from the true values of the parameters (as would be the case in a real analysis). Figure 3: The real- and imaginary parts of the difference of the mean of 100 samples of $a_{\ell m}$ modes from the GCR solver and the true sky using the _standard_ configuration and parameters. Any _outliers_ from the central region is marked with $\blacktriangle$ and it is notable that these occur more frequently in the low-$\ell$ region. As described in Sec. 2.1 the $m=0$ imaginary modes ($\times$) should always be zero and the GCR solver therefore does not solve for this. Figure 4: Maps generated from the estimated spherical harmonic modes on the sky using the _standard_ configuration and parameters. _Upper left:_ The true sky given by pyGSM with ${\ell_{\text{max}}}=20$ and nside$=128$. Note that the rippled structure comes from the true sky and not from the GCR solver. _Upper right:_ The estimated sky based on the mean of 100 samples from the GCR solver. The spherical harmonic modes can be seen in Fig. 3. _Bottom left:_ Fractional difference between the mean and the true sky adjusted to show differences $<10\%$, which coincides with the beam region, see also Fig. 5 for a closeup of the region. _Bottom right:_ The standard deviation of the difference of the estimated and true sky. ## 4 Results In the following we show results of the samples obtained under various different parameter configurations of the sampler itself and by investigating the effects of different observation scenarios. The purpose of this analysis is to demonstrate the basic behaviours of the method and to validate that it can recover the true sky to a reasonable level. The comparative analysis uses simulations in order to understand the behaviour we would expect from real data as a function of different LST coverage, the effects of effectively truncating the visibility response operator at different levels of $\ell_{\text{max}}$, and the impact of increasing the FoV of the array. We also dive into a stress-test of the prior- vs likelihood levels, making sure that neither is too broad nor too narrow. Before going into the comparative analysis, we show in more detail in Sec. 4.1 the results of our chosen reference – or, _standard_ – case, using the general parameters described in Sec. 3, considering not only the fractional difference between the posterior mean and true sky map but also looking at the recovered $a_{\ell m}$ modes, followed by the full comparative analysis in Secs. 4.2–4.3. Unless otherwise stated, we draw 100 samples per scenario, as we deemed this sufficient for calculating the relevant statistics of the recovered spherical harmonic modes. As a precautionary check of this assumption, we also consider a high-sample size case of $N_{\text{samples}}\approx 3500$. Finally, Sec. 4.4 is dedicated to a closer study of the special case of increasing the field of view by decreasing the diameter of the antennas to $\theta_{\text{D}}=$3\text{\,}\mathrm{m}$$, which results in a FWHM at $100\text{\,}\mathrm{MHz}$ of $\sim$60\text{\,}\mathrm{\SIUnitSymbolDegree}$$. Having a greater sky coverage should improve on the sensitivity on lower $\ell$ modes. A large FoV can be achieved with instruments like the MWA. Figure 5: Cartesian projection of the fractional difference from the true sky using the _standard_ configuration and parameters. The RA and dec ranges have been narrowed to focus on the primary beam region. The FWHM of $12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$ is illustrated in the corner of the figure. The white horizontal lines indicate a FWHM distance from the centre of the primary beam at $-31.7\text{\,}\mathrm{\SIUnitSymbolDegree}$. The simulated LST range is $0$-$8\text{\,}\mathrm{h}$ corresponding to an RA range of $0\text{\,}\mathrm{\SIUnitSymbolDegree}$-$120\text{\,}\mathrm{\SIUnitSymbolDegree}$. The centre of the beam-region has fractional differences to the true sky of $<5\%$ and at the bounds $<10\%$. Figure 6: Approximate range of $\ell$-value sensitivity for several baselines, as a function of frequency. The baselines displayed here are the five unique baseline lengths of the 10-dish standard configuration. The width (shaded regions) is given by the FWHM of the beam $\delta\ell\simeq\pi/{\rm FWHM}$, shown here for the HERA-like case with FWHM $=$12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$$. The two horizontal lines indicate an ${\ell_{\text{max}}}=30$ (top, dark purple) and ${\ell_{\text{max}}}=20$ (bottom, magenta) and the black vertical line is set at $\nu=$100\text{\,}\mathrm{MHz}$$, the frequency used for the simulations in this work. ### 4.1 Realisations of the standard configuration Figure 7: Cartesian projection of the fractional difference between the posterior mean and true sky for multiple sets of samples with varied levels of noise covariance $\boldsymbol{\mathbf{N}}$, prior covariance $\boldsymbol{\mathbf{S}}$, prior mean $\boldsymbol{\mathbf{a}}_{0}$, and increased sample size compared to the _standard_ case. The fractional difference is defined as defined as $(\mu({\boldsymbol{\mathbf{a}}}_{\textup{cr}})-\boldsymbol{\mathbf{a}}_{\textup{true}})/\boldsymbol{\mathbf{a}}_{\textup{true}}$, where $\mu$ indicates the sample mean. The two horizontal lines indicate a FWHM distance from the centre of the beam. The results from the Gaussian constrained realisations of the $a_{\ell m}$ modes for the standard case are shown in Figs. 3, 4, and 5. First, we define the true-sky subtracted mean as the mean of the spherical harmonic coefficients sampled by the GCR method subtracted by the true sky (input model) spherical harmonic modes from pyGDSM. In Fig. 3 the true-sky subtracted mean of 100 realisations of the $a_{\ell m}$ modes is shown with corresponding error bars defined by the standard deviations of the samples. To be able to show the details of the region around a difference of zero, the figure is cropped to $\pm 100$ with any modes outside of this region are marked with a $\blacktriangle$ and their values. Since the $m=0$ imaginary modes are artificially injected back in with a value of 0 (cf. Eq. 5), they are without error bars but have been included and marked with $\times$ in the figure for clarity. It is clear that the higher order modes are easiest to constrain and that the lower order modes are far from the true sky. In Fig. 6 the $\ell$-mode sensitivity given the baseline length and frequency is shown. The shortest, and most frequent, baseline length in the standard configuration is $b=$14.6\text{\,}\mathrm{m}$$ corresponding to an $\ell_{\textup{min}}\sim(\pi b)/\lambda\sim 15$ at a frequency of $100\text{\,}\mathrm{MHz}$. Note, that at this frequency the baseline length (in metres) roughly corresponds to the $\ell$-value. Thus we are not expecting to be sensitive to $\ell\sim$ few, since modes with $\Delta\theta\gtrsim\lambda/b_{\rm min}$ are resolved out by the interferometer, and so we should not expect the data to constrain them. The recovered values of these low-$\ell$ modes are instead prior-driven. In fact, Fig. 3 is showing the spherical harmonic coefficients, each of which integrates information across the whole sky. Since the observed region is small (only $\sim$ 8.6% of the sky), the recovered $a_{\ell m}$ values are generally mostly determined by the prior even when the noise level is very low in the observed region, we suspect this is because the prior-dominated area is much larger. Hence, plotting the statistics of the recovered $a_{\ell m}$ values directly is not a very sensitive test of this method. Instead, it is much more useful to look at the recovered sky in map-space, as the difference between the observed and unobserved region is much clearer laid out. In Fig. 4 we show the maps of the true input sky (top left) given as described in Sec. 3.2 compared to the recovered map defined by the mean of the posterior (top right). It is noted that the recovered map shows lower values for the large scale modes (just as the ${a_{\ell m}}$-values in Fig. 3) compared to the true sky map, which we expect is due to a random fluctuation rather than a bias. Fig. 4 also shows the fractional difference between these two maps (bottom left) and the standard deviation of the recovered sky (bottom right). The fractional difference between the posterior mean and true sky is lowest within the primary beam region and the standard deviation is much smaller in this area too, thus demonstrating the specific patch of sky that is directly observed versus that constrained by the prior. This explains how we are still able to recover most of the galactic structure, despite the majority of it being outside the observed region of the sky. Note, that between RA$=$150\text{\,}\mathrm{\SIUnitSymbolDegree}$$ and RA$=$180\text{\,}\mathrm{\SIUnitSymbolDegree}$$ there is another small low residual region in the fractional difference map, however, this does not appear in the standard deviation map. This low residual is simply due to a random fluctuation rather than the true sky being well recovered in this region, and it is therefore ignored. Now focusing on the primary beam region alone, we show a Cartesian projection of the fractional difference between the posterior mean and true sky in Fig. 5. The region is defined to be within a FWHM on either side of the central latitude of $-31.7\text{\,}\mathrm{\SIUnitSymbolDegree}$. For the standard case with the antenna diameter of $\theta_{\text{D}}=$14\text{\,}\mathrm{m}$$ the FWHM is $12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$. For the standard case the sky is recovered to within 5% close to the centre of the beam region and to within 10% within the FWHM bounds. ### 4.2 Noise and prior levels Figure 8: The absolute value of the fractional difference between the posterior mean and true sky taken as a slice through the centre of the primary beam region at dec=$-31.7\text{\,}\mathrm{\SIUnitSymbolDegree}$. Artefacts due to the pixelisation have been smoothed away by applying a pair of moving average filters to each of these curves. For all runs in this plot the LST range is $0-$8\text{\,}\mathrm{h}$$ corresponding to RA values of $0-$120\text{\,}\mathrm{\SIUnitSymbolDegree}$$. The light shaded region covers the FWHM of $12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$ and the dark shaded region is outside of the primary beam. Within the LST range (white region) the standard and A1 case have the lowest fractional difference between the posterior mean and true sky. Figure 9: Cartesian projection of the fractional difference between the posterior mean and true sky for runs with variations to the standard observation/simulation parameters such as increasing the ${\ell_{\text{max}}}$, adding antennas for longer baselines, increasing the LST range, and increasing the cadence of observations. The fractional difference is defined as defined as $(\mu({\boldsymbol{\mathbf{a}}}_{\textup{cr}})-\boldsymbol{\mathbf{a}}_{\textup{true}})/\boldsymbol{\mathbf{a}}_{\textup{true}}$, where $\mu$ indicates the sample mean. The two horizontal lines indicate a FWHM distance from the centre of the beam. Figure 10: The absolute value of the fractional difference between the posterior mean and true sky taken as a slice through the centre of the primary beam region at dec=$-31.7\text{\,}\mathrm{\SIUnitSymbolDegree}$. Artefacts due to the pixelisation have been smoothed away by applying a pair of moving average filters to each of these curves. Results B4-B5 have different LST/RA ranges as indicated by the additional x-axes. Note that for B2, only the first half of the LST/RA range is shown, to match the $8\text{\,}\mathrm{h}$ span of the other runs. The light shaded region covers the FWHM of $12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$ and the dark shaded region is outside of the primary beam. It is noticeable that all three ${\ell_{\text{max}}}=30$ results have higher fractional differences to the true than the ${\ell_{\text{max}}}=20$ results. Table 1: Overview of all the cases analysed in this paper and their labels along with which parameters are varied. A “—” indicates that the parameter is set as in the standard case. The labels are numbered after the order of appearance in Figs. 7 and 9. | $N_{\rm samples}$ | Noise covariance | Prior covariance | Prior mean ---|---|---|---|--- A1 | $\sim 3500$ | — | — | — A2 | — | — | — | $\boldsymbol{\mathbf{a}}_{0}=0$ A3 | — | $\boldsymbol{\mathbf{N}}\times 10^{4}$ | — | $\boldsymbol{\mathbf{a}}_{0}=0$ A4 | — | $\boldsymbol{\mathbf{N}}\times 10^{2}$ | — | — A5 | — | $\boldsymbol{\mathbf{N}}\times 10^{2}$ | $\boldsymbol{\mathbf{S}}^{-1}=0$ | — A6 | — | — | $\boldsymbol{\mathbf{S}}\times 10^{2}$ | — A7 | — | $\boldsymbol{\mathbf{N}}\times 10^{4}$ | — | — A8 | — | $\boldsymbol{\mathbf{N}}\times 10^{4}$ | $\boldsymbol{\mathbf{S}}^{-1}=0$ | — A9 | — | — | $\boldsymbol{\mathbf{S}}\times 10^{4}$ | — | ${\ell_{\text{max}}}$ | Number of LSTs | LST range | B1 | ${\ell_{\text{max}}}=30$ | — | — | B2 | — | $N_{\rm LST}=20$ | $0-$16\text{\,}\mathrm{h}$$ | B3 | — | $N_{\rm LST}=20$ | $0-8\,\,\,\,$\mathrm{h}$$ | B4 | — | $N_{\rm LST}=10$ | $16-$24\text{\,}\mathrm{h}$\,\,\,$ | B5 | — | $N_{\rm LST}=10$ | $8-$16\text{\,}\mathrm{h}$$ | | $N_{\rm ants}$ | Dish diameter | FWHM | C1 | — | $3\text{\,}\mathrm{m}$ | $57.3\text{\,}\mathrm{\SIUnitSymbolDegree}$ | In this section we explore how varying the noise covariance $\mathbf{N}$, prior covariance $\mathbf{S}$, and prior mean $\boldsymbol{\mathbf{a}}_{0}$ will affect the results from the GCR-sampler compared to our chosen standard case. The motivation for this is to assure that a sufficient balance is reached between the data- and prior-information, and that the prior is not dominating the system. The results of the various noise- and prior level runs can be seen in Figs. 7 and 8. All the various runs have been labelled in order of there appearance in Fig. 7 and these labels will be used throughout the paper. A full overview of all labels can also be seen in Table 1. Before varying the noise and prior-levels, we examine whether using many more samples ($\approx 3500$) would significantly change the results (case A1), in order to check that only 100 samples is enough to produce statistically valid results. The fractional difference between the posterior mean and true sky for the A1 result is shown in Fig. 7 along with re-displaying the standard case for comparison. It is clear that the fractional difference of the two cases look almost identical. For even closer comparison we take a slice of the centre of the beam of the absolute values of the fractional difference to the true value, as shown in Fig. 8. As the figure shows, even when comparing the fractional difference slices, the two cases only have a slight difference at ${\rm RA}\sim$10\text{\,}\mathrm{\SIUnitSymbolDegree}$$, thus emphasising that these two results are almost identical, which (together with the full beam region fractional difference) justifies the use of 100 samples as adequate for our purposes. Next, we set the inverse prior covariance to zero, which may be interpreted as the case where only the likelihood term is used and the prior is neglected. We then multiply the noise covariance with a factor of $10^{2}$ and $10^{4}$, case A5 and A8 respectively. This is done in order to confirm, that removing the information provided by the likelihood and/or prior does indeed lead to no recovery of information of the sky, i.e. we demonstrate that the good recovery in other cases is not simply due to the structure of the sky model or similar. The fractional difference between the posterior mean and true sky for the A5 case ($\boldsymbol{\mathbf{S}}^{-1}=0,\mathbf{N}\times 10^{2}$) shows a slight hint of the primary beam structure, but the high noise levels washes out the data and there is no prior-information to assist the recovery of the sky. When increasing the noise further in the A8 case ($\boldsymbol{\mathbf{S}}^{-1}=0,\mathbf{N}\times 10^{4}$), the recovered sky is all gone and there is only noise left. In order to see how well the sampler can recover the sky with no prior information, an extra run (not shown) was made with standard noise level (but keeping the inverse prior covariance at zero), which resulted in a fractional difference between the posterior mean and true sky very similar to that of case A8 (i.e. where $\mathbf{S}\times 10^{4}$). Essentially the very high value of the prior covariance $\mathbf{S}$ is the same as saying the inverse prior covariance $\boldsymbol{\mathbf{S}}^{-1}$ is closer to zero. It is therefore not that surprising that these two results would be similar. For both these cases the prior is now so broad and uncertain that the only well-recovered part is the observed region. The prior drives the good recovery outside of the observed region, which is very clear when comparing for instance case A7 ($\mathbf{S}=$ standard, $\mathbf{N}\times 10^{4}$) to case A8 ($\mathbf{S}^{-1}=0,\mathbf{N}\times 10^{4}$). Without any prior information nothing is recovered – but with the prior most of the sky is recovered well to within $\sim 10\%$. However, A7 lacks the primary beam structure that we saw in the standard case, so the recovery here is due to the prior driving the solution since the broad noise covariance is washing out the data. When the noise is lowered again in case A4 to $\mathbf{N}\times 10^{2}$ (while keeping $\mathbf{S}=$ standard), the central beam structure starts to reappear. The good recovery of the sky outside of the observed region is without a doubt driven by the prior – as the data cannot say anything about this region anyway – but the _very_ good recovery of the observed region is due to the data. In the standard case the well-recovered sky is further limited to be close to the observed region. The fact that the spherical harmonics are continuous means the sky will still be constrained by the data a little outside of the observed region. This explains why the standard deviation in Fig. 4 smoothly goes from low to a high value. An important test is to make sure that our choice of prior mean, $\boldsymbol{\mathbf{a}}_{0}$, is not over-informing the sampler. Comparing the A2 ($\mathbf{a}_{0}=0,\mathbf{N}=$ standard) and A3 ($\mathbf{a}_{0}=0,\mathbf{N}\times 10^{4}$) cases to the standard case in Fig. 7, it initially looks like setting the prior mean to zero results in a poorly recovered sky. Indeed, this is the case if the noise is also increased (as in A3), although generally increasing the noise with a factor of $10^{4}$ has yielded poor results (A7, A8, A9) and does not seem to be due to the change of the prior mean alone. Instead, when the noise level is kept at standard level (case A2), the true sky is still well-recovered near the centre of the beam (i.e. recover is not biased), but the residual is larger beyond the beam FWHM than in the standard case. Again, taking a slice of the absolute fractional difference at the central declination only (around $\textup{dec}=$-31.7\text{\,}\mathrm{\SIUnitSymbolDegree}$$), Fig. 8), shows that the two $\mathbf{a}_{0}=0$ cases are not performing as good as the standard case. Although, especially for the A2 case, at RA$=60-$100\text{\,}\mathrm{\SIUnitSymbolDegree}$$ it is performing similarly or (very) slightly better than the standard case. In most of our runs we have a 10% prior that has prior means which are scattered around the true value. This is somewhat conservative (we are not assuming a ’correct’ prior mean and thereby not enforcing the true solution), but there is still the question of how sensitive the results are to prior choices. What the $\boldsymbol{\mathbf{a}}_{0}=0$ runs (A2, A3) show is what happens when our prior is typically wrong by $10\sigma$. ### 4.3 Effect of changing array and observing configurations The performance of the sampler is affected by which spherical harmonic modes are available to it. This can be tested by changing the ${\ell_{\text{max}}}$ or for instance by including longer baselines to increase sensitivity to higher $\ell$ modes, as can also be gauged from Fig. 6. Since most of the diffuse foreground power comes from the Galaxy, it is also interesting to examine how changing the LST range will affect how well the sky is recovered. All the test cases in the following section has been labelled with a similar style to the previous section, now based on the order of appearance in Fig. 9. A full overview of all labels can be seen in Table 1. To capture regions of the entire HERA-strip we examine LST ranges of $8-$16\text{\,}\mathrm{h}$$ and $16-$24\text{\,}\mathrm{h}$$ on top of the $0-$8\text{\,}\mathrm{h}$$ LST range; case B5, B4, and the standard case, respectively. The B4 case covers the brightest regions of the galactic diffuse emission, which can be seen in the RA range of $$240\text{\,}\mathrm{\SIUnitSymbolDegree}$-$360\text{\,}\mathrm{\SIUnitSymbolDegree}$$ (the readers right hand side) on the top-left plot of Fig. 4. We also examine the effects of increasing the data set to $N_{\textup{LST}}=20$ either by doubling the LST range and keeping the same cadence (case B2) – or, vice versa (case B3). The former results in a greater sky coverage and the latter results in more information of the same sky region. The standard case cadence is one integration of duration 60 sec per 48 minutes equivalent of 10 samples evenly spread over an 8 hour LST range. Note, since the sky rotates $15\text{\,}\deg\mathrm{/}\mathrm{h}$, the beam-crossing time of the HERA-like array is $\sim$49\text{\,}\mathrm{min}$$ given the FWHM of $12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$. In this section we go over the results from each of these cases and how they affect the resulting recovered sky. For the standard case we use ${\ell_{\text{max}}}=20$ because it is a reasonable minimum for testing purposes that includes the shortest (most sensitive) HERA baseline, see Fig. 6. In case B1, we see from Fig. 10 that increasing ${\ell_{\text{max}}}$ to 30 results in a significantly degraded fractional residual compared to ${\ell_{\text{max}}}$ of 20. While increasing ${\ell_{\text{max}}}$ brings the $b=$25.3\text{\,}\mathrm{m}$$ baseline group into the _directly_ constrained $\ell$ range (see Fig. 6), each baseline is actually sensitive to a broad range of $\ell$-values, as for example the primary beam contributes to smearing of the $\ell$-mode sensitivity. For instance the $b=$29.2\text{\,}\mathrm{m}$$ baseline (antennas $(0,2)$) in Fig. 1 is clearly sensitive to $\ell$-values down to around $\ell\sim 8$ and the $(0,5)$-baseline, which is a shorter $25.3\text{\,}\mathrm{m}$ baseline, contributes mostly at $\ell\gtrsim 5$ but still has some visibility response below this level. These baselines will therefore already contribute to some degree at the ${\ell_{\text{max}}}=20$ level. However, when we increase the ${\ell_{\text{max}}}$ to 30 the total number of modes more than doubles, from 441 to 961. This stretches the available signal-to-noise of the data much farther than in the ${\ell_{\text{max}}}=20$ case. Next, we vary the LST ranges. First, we keep to 10 time steps and 8 hr ranges (as in the standard case), but shift the ranges to cover different parts of the sky. Fig. 9 shows that the range of $8-$16\text{\,}\mathrm{h}$$ (B5) performs better than both the standard $0-$8\text{\,}\mathrm{h}$$ and the B4 case ($16-$24\text{\,}\mathrm{h}$$), which covers the brightest part of the galactic emission. As before, we inspect the slice at the central declination for a closer comparison, see Fig. 10. It is clear from the figure, that within their respective observed regions all three sets of LST ranges actually perform comparatively well with only small differences. As one moves closer to the edge of the observed region, however, the B5 case shows some improvement to the standard case – and, especially, when compared to the B4 case. The difference in recovery level can simply be due to the shifting of the sampling points. For all the runs the simulated data is the same and have one specific realisation of the noise. Areas that have less noise will therefore always have a higher signal-to-noise and should perform better. Increasing the data volume should also help constrain the sky better. One way to do this is to double the number of observation times from $N_{\textup{LST}}=10$ to $20$, either by doubling the range and keeping the cadence the same (case B2) or by increasing the cadence within the same LST range (case B3). At first glance both results seem to perform similar or perhaps slightly better than the standard case, see Fig. 9. The recovered signal will be worse at the edges of the observed region. The B2 case is twice the size and encompasses the far edge of the B3 case at LST $=$8\text{\,}\mathrm{h}$$, and will naturally be better recovered at this LST than the B3 case. We will therefore limit the comparison to the $0-$8\text{\,}\mathrm{h}$$ range for the fractional residual slice in Fig. 10. The B3 result mostly shows improvement to the standard case with the exception of RA $$20\text{\,}\mathrm{\SIUnitSymbolDegree}$-$40\text{\,}\mathrm{\SIUnitSymbolDegree}$$ although exceeded by the performance of the B2 case, that not only performs better than the standard case but also remains closer to true at the lower edge ($0\text{\,}\mathrm{h}$ of the observed region. This hints that increasing the observation time range (and thus increasing the size of the observable sky) has a larger impact on the recovery of the sky than doubling the data we have inside the same region. This can also be explained with a larger observable sky being sensitive to larger scales (lower spherical harmonic modes). ### 4.4 Improving sensitivity to larger scales Figure 11: The real- and imaginary parts of the difference of the mean of 100 samples of $a_{\ell m}$ modes from the GCR solver and the true sky with the diameter of the receivers set to $\theta_{D}=$3\text{\,}\mathrm{m}$$ and thereby increasing the field of view. As described in Sec. 2.1 the $m=0$ imaginary modes ($\times$) should always be zero and the GCR solver therefore does not solve for this. All residual ${a_{\ell m}}$ modes now lie within the $\pm 100$ boundaries and including the uncertainties are consistent with zero, except the zeroth $\ell$-mode that still has some loss of power. Figure 12: Resulting maps from the estimated spherical harmonic modes on the sky with the diameter of the receivers set to $\theta_{D}=$3\text{\,}\mathrm{m}$$ and thereby increasing the field of view. _Upper left:_ The true sky given by pyGSM with ${\ell_{\text{max}}}=20$ and nside$=128$. Note that the rippled structure comes from the true sky and not from the GCR solver. _Upper right:_ The estimated sky based on the mean of 100 samples from the GCR solver. The spherical harmonic modes can be seen in Fig. 11. _Bottom left:_ Fractional difference between the mean and the true sky adjusted to show differences $<10\%$ (i.e. highlighting improvements over the prior). The beam region is no longer clear cut as in the standard case. _Bottom right:_ The standard deviation of the difference of the estimated and true sky. The lowest noise levels are now larger than with the standard case, but overall noise is lower. One reason that the smaller scales are more difficult to constrain comes from the array layout itself, as discussed above. The prior helps breaking the degeneracy of the low-$\ell$ modes, that can otherwise only be determined as a linear combination (i.e. they are degenerate, as there are no sufficiently short baselines to resolve them individually). It is clear, however, from the $a_{\ell m}$ modes presented in Fig. 3 that these modes are still difficult to constrain up to $\ell\sim 5$. Moving on the same logic as above, where the observable sky is increased to increase sensitivity to the larger scales, we can decrease the size of the dishes to increase the field of view (FoV). With the HERA-like dishes the FWHM $=$12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$$. Altering the array to have a FoV more similar to that of OVRO-LWA (Eastwood et al., 2018) or MWA (Yoshiura et al., 2021), the diameter of the receivers is changed to $3\text{\,}\mathrm{m}$, now with a ${\rm FWHM}=$57.3\text{\,}\mathrm{\SIUnitSymbolDegree}$$ (case C1). A broader FoV now means that the smearing around the $\ell$ modes in Fig. 6 is much narrower ($\delta\ell\simeq\pi/{\rm FWHM}$), thus more clearly picking out the specific $\ell$ modes. Unlike when we increase the LST range, we expand the observable sky not only in the E-W direction but also N-S. As increasing the FoV with more than $4\times{\rm FWHM}$ of the standard case is a much more radical change than those made in Sec. 4.3, the C1 results are presented in more detail in Figs. 11 and 12, although now leaving out the Cartesian projections of the fractional residual, since the beam is now so wide that a close up of the primary beam region is no longer relevant. Starting with the true- subtracted mean of the recovered ${a_{\ell m}}$ modes in Fig. 11, it is noticeable that there are no longer any _outliers_ from the displayed region, as was the case in Fig. 3 with the standard configuration. The zeroth-mode is still the most difficult to constrain and is again underestimated, however, not as severely underestimated as earlier. Earlier it was argued that the ${a_{\ell m}}$ modes are not the most representative performance indicator of the code, since they are all-sky quantities and the observed region is only a narrow strip, but now that we have much larger coverage it clearly shows that full-sky features as the ${a_{\ell m}}$ modes can be well-constrained. By-eye, it is now almost impossible to tell the difference between the true sky map and the mean of the recovered sky in Fig. 12. Taking the fractional difference between the two maps reveals a very different structure than earlier. Now, almost all parts of the sky are recovered within $10\%$ with the Galaxy showing the smallest residuals. It is striking that the primary beam structure no longer shows up in either the fractional difference map or the standard deviation. Since the noise covariance $\mathbf{N}$ and subsequently the noise $n_{ij}$ on the data is defined by the auto-visibilities as per Eq. 25, the noise level will be affected by whether the Galactic emission is within the beam. Since the FoV is now so large, the noise has increased with the extra power from the Galaxy as most of this falls within the observed area. Additional runs have been made where the noise was reduced with a factor of $10^{-1}$ and $10^{-2}$ respectively (not shown), and not only does this improve the estimate of the ${a_{\ell m}}$ modes, it is also clear that the area of the sky with the largest fractional residual is the furthest region from the observed region, corresponding to the same area in Fig. 12 where the standard deviation is higher. Decreasing the noise also reduced the standard deviation inside the observed region to the same level as the standard case. Earlier, in Sec. 4.2, we showed that reducing the noise level could lead to worse recovery of the sky. Since we are dealing with constrained realisations, the solution in the prior-dominated region can still depend on the data- dominated region (and vice-versa), as the spherical harmonic basis functions connect the two. The influence of the data on the prior-dominated region is minimal if the noise covariance is very narrow. For the $3\text{\,}\mathrm{m}$ beam case it is not as big an issue, since the large FoV makes up for it. Before the beam was also very narrow, so when the noise covariance was made narrow, it did not allow the prior to contribute as much, which made it difficult to recover the sky even at the edges of the observed region. Now, that the FoV is larger, the prior is not as crucial to determine full-sky features and we can reduce the noise without losing the constraining power. ## 5 Conclusions For next-generation wide-field radio interferometer arrays, particularly those targeting the 21 cm fluctuations at low and high redshift, the biggest challenge remains proper handling of foreground contamination following its distortion/modulation by the instrument. To improve on this issue, it is crucial to be able to make accurate maps of the radio emission on both small and large scales from the measured visibilities. Traditional deconvolution algorithms such as CLEAN (and its extensions) can struggle to properly recover diffuse emission, which is the dominant foreground contaminant on large angular scales. To address this, methods such as Direct Optimal Mapping and $m$-mode analysis have been developed, both of which are maximum likelihood estimators that focus on accurately recovering wide field maps of the emission, i.e. on scales where the curvature of the sky and the shape of the primary beam becomes important. In this work we have presented an alternative approach to recovering diffuse foregrounds from visibilities – by constructing a parametric model of the emission, represented by spherical harmonics on the sky, and drawing samples from the joint posterior of the spherical harmonic coefficients given the visibility data and a choice of prior. A linear model can be constructed by writing down an operator that encodes the response of the interferometer (and thus the measured visibilities) to each spherical harmonic mode. This operator includes all the relevant instrumental degrees of freedom. We can then estimate the joint posterior distribution of all the coefficients by repeatedly solving the Gaussian constrained realisation (GCR) equation, an extension of the Wiener filter that includes random realisation terms. This can be solved efficiently even for a very large number of coefficients, making this large inference problem computationally tractable. Furthermore, each sample of coefficients is a complete realisation of the spherical harmonic coefficients, and therefore the full sky (i.e. with no gaps or mask) that is statistically-consistent with the data. This ultimately allows us to recover the diffuse emission signal in a statistically robust manner while also avoiding difficulties with missing data in subsequent steps of the data analysis pipeline. After presenting the mathematical formalism for this sampler, the primary aim of this paper was to validate the method by applying it to a suite of simulations. We tested the performance of the spherical harmonic sampler by comparing the recovered sky, defined as the mean of the sky maps of the samples, to the (known) true sky modelled using pyGSM. For the analysis we chose a standard set of parameters to act as a reference for the various tests of noise and prior uncertainty levels, sample size, ${\ell_{\text{max}}}$, the specific LST range of the observations, and the influence of the field of view. The _standard case_ is based on 100 realisations with a $10\%$ prior centred on the true ${a_{\ell m}}$-modes, and the noise is given by a Gaussian distribution with zero mean and covariance defined by the auto-visibilities through the radiometer equation. The standard case uses a HERA-like closed- packed redundant array of 10 antennas and covers the LST range $0-$8\text{\,}\mathrm{h}$$. The dishes are 14 m in diameter (FWHM $=$12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$$), which means the directly observed sky is a narrow $24.6\text{\,}\mathrm{\SIUnitSymbolDegree}$ (dec) by $120\text{\,}\mathrm{\SIUnitSymbolDegree}$ (RA) strip covering only $\sim 8.6\%$ of the full sky. The standard case performs well within the observed region and recovers the true sky to within $10\%$ within a distance of 1 FWHM from the centre and to within $5\%$ at the centre of the beam. When increasing the number of samples to $N_{\textup{samples}}\approx 3500$, we found that there were minimal changes to the result allowing us to keep the computational costs lower by continuing with 100 samples per scenario only. Next, the impact of the noise level on recovery was studied. When the noise variance is increased by a factor of $>10^{2}$, the noise will dominate over the sky signal and downgrade the recovery. At this noise level, the samples will only show diffuse sky structure if there is a prior as well to drive the solution. For the standard case we chose a $10\%$ prior, as the pyGSM model should describe the diffuse emission sky to roughly within this margin. If the prior covariance is increased, however, the prior quickly becomes too broad and uncertain and the only well-recovered part is the very centre of the observed region. Including a prior helps break the degeneracy of the low-$\ell$ modes. It was also found that even if the prior mean is set to zero (a highly pessimistic assumption), the sky can still be recovered, although to a slightly worse degree than the standard case. This shows the benefit of including a prior; the boundary region between observed and unobserved sky is constrained by both the prior and the data, while regions of missing data are still filled with a statistically plausible realisation. We conclude that, for the standard case, the prior is not overly-informative, and contributes about the same as the data to the diffuse emission recovery around the observed region. For close-packed arrays like HERA, the most abundant baselines are the shortest ones, which have lengths $<$30\text{\,}\mathrm{m}$$. This justifies choosing a relatively low ${\ell_{\text{max}}}$ for our tests. When increasing the ${\ell_{\text{max}}}$ from 20 to 30, the number of modes in the ${a_{\ell m}}$-vector increases significantly from 441 to 961, more than doubling the number of parameters we are trying to constrain, and thus stretching the available signal-to-noise across more degrees of freedom. A more informative prior would help to balance this increase in the size of the parameter space. Most of the diffuse emission on the sky originates from the Galaxy and is not uniformly distributed. This results in field-dependent effects on the recovery of the signal. For instance, the recovery was best for a $8-$16\text{\,}\mathrm{h}$$ region, whereas the region covering the brightest part of the Galaxy ($16-24$\mathrm{h}$$) was recovered worst, due largely to the extra noise given by the auto-visibilities. When increasing the data volume we expect the results to improve, for instance if we increase the cadence and take more data samples within the same LST range, or, if we keep the cadence but increase the range. When doubling the number of data points within the range, there was only a slight improvement to the recovered sky, suggesting that there is not much information to be gained by making more measurements of nearby (correlated) pointings, e.g. within a beam width of one another. In contrast, doubling the range itself not only improved the results in the centre, but also at the boundaries of the LST range, as additional sky coverage led to improved constraints on the spherical harmonic modes across the board. Ultimately, the biggest impact on our simulated results occurred by increasing the field of view. This was done by decreasing the effective antenna diameter to $3\text{\,}\mathrm{m}$, thus increasing the FWHM from $12.3\text{\,}\mathrm{\SIUnitSymbolDegree}$ to $57.3\text{\,}\mathrm{\SIUnitSymbolDegree}$. For the large-FoV case, the spherical harmonic coefficients now all coincide with the true sky within their given uncertainties except the zeroth mode. The standard deviation map of the large-FoV results no longer shows the clear strip of the observed sky, as almost all of the modes are now well-constrained. Since the noise is defined by the auto-visibilities, and this setup now directly observes most of the Galaxy, the noise level is larger than the standard case. However, as the FoV is now so large, we are also less dependent on the prior as there are fewer gaps of information to fill in, and the boundary region (between directly observed and totally unobserved regions) is wider. The results presented in this paper are based on a handful of simplified example cases, using a reduced-size array and only two frequencies for the visibility response operator. A simple extension of this work would be to rerun the analysis while simulating the visibility response for more frequencies or, alternatively, the frequency dependence could be included via a parametric model, e.g. power law with spectral index $\beta$. Likewise, we have only considered cases of ${\ell_{\text{max}}}<30$, but in reality a higher angular resolution would be desirable if a larger portion of the array is considered (i.e. including more longer baselines). Finally, the Gaussian constrained realisation method has been developed with inclusion into a Gibbs sampling framework in mind, whereby beam, 21cm signal, and point source foreground model parameters would also be sampled. A simpler and more focused Gibbs sampling scheme would also enable us to sample the signal covariance $\boldsymbol{\mathbf{S}}$ (e.g. as parametrised by the angular power spectrum), as well as the spherical harmonic coefficients themselves. ## Acknowledgements This result is part of a project that has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant agreement No. 948764; KAG, PB, JB, MJW). We acknowledge use of the following software: matplotlib (Hunter, 2007), numpy (van der Walt et al., 2011), and scipy (Virtanen et al., 2020). This work used the DiRAC@Durham facility managed by the Institute for Computational Cosmology on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). The equipment was funded by BEIS capital funding via STFC capital grants ST/P002293/1, ST/R002371/1 and ST/S002502/1, Durham University and STFC operations grant ST/R000832/1. DiRAC is part of the National e-Infrastructure. ## Data Availability The Python code used to produce the results in this paper is available from https://github.com/HydraRadio/Hydra. The Jupyter notebook used for the data analysis and to produce plots is available from https://github.com/katrinealice/sph_harm_GCR_analysis. Simulated data are available on request. ## References * Ali et al. (2015) Ali Z. S., et al., 2015, ApJ, 809, 61 * Bowman et al. (2013) Bowman J. D., et al., 2013, PASA, 30, e031 * CHIME Collaboration (2023) CHIME Collaboration 2023, The Astrophysical Journal, 947, 16 * Cornwell (2008) Cornwell T. J., 2008, IEEE Journal of Selected Topics in Signal Processing, 2, 793–801 * Cornwell, T. J. (2009) Cornwell, T. J. 2009, A&A, 500, 65 * DeBoer et al. (2017) DeBoer D. R., et al., 2017, Publications of the Astronomical Society of the Pacific, 129, 045001 * Dillon & Parsons (2016) Dillon J. S., Parsons A. R., 2016, The Astrophysical Journal, 826, 181 * Dillon et al. (2015) Dillon J. S., et al., 2015, Phys. Rev. D, 91, 023002 * Dong (1994) Dong R., 1994, Nuclear Magnetic Resonance of Liquid Crystals. Partially ordered systems, Springer-Verlag, https://books.google.co.uk/books?id=tAbwAAAAMAAJ * Dowell & Taylor (2018) Dowell J., Taylor G. B., 2018, The Astrophysical Journal Letters, 858, L9 * Eastwood et al. (2018) Eastwood M. W., et al., 2018, The Astronomical Journal, 156, 32 * Eriksen et al. (2008) Eriksen H. K., Jewell J. B., Dickinson C., Banday A. J., Górski K. M., Lawrence C. R., 2008, ApJ, 676, 10 * Evoli et al. (2014) Evoli C., Mesinger A., Ferrara A., 2014, J. Cosmology Astropart. Phys, 2014, 024 * Fixsen et al. (2011) Fixsen D. J., et al., 2011, The Astrophysical Journal, 734, 5 * Furlanetto et al. (2006) Furlanetto S. R., Oh S. P., Briggs F. H., 2006, Physics Reports, 433, 181 * Geman & Geman (1984) Geman S., Geman D., 1984, IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-6, 721 * Högbom (1974) Högbom J. A., 1974, A&AS, 15, 417 * Hunter (2007) Hunter J. D., 2007, Computing in Science Engineering, 9, 90 * Kennedy et al. (2023) Kennedy F., Bull P., Wilensky M. J., Burba J., Choudhuri S., 2023, The Astrophysical Journal Supplement Series, 266, 23 * Kittiwisit et al. (2023) Kittiwisit P., et al., 2023, arXiv e-prints, p. arXiv:2312.09763 * Kriele (2022) Kriele M. A., 2022, PhD thesis, Eindhoven University of Technology, Curtin University, ASTRO3D * Liu & Shaw (2020) Liu A., Shaw J. R., 2020, Publications of the Astronomical Society of the Pacific, 132, 062001 * Lopez-Honorez et al. (2016) Lopez-Honorez L., Mena O., Moliné Á., Palomares-Ruiz S., Vincent A. C., 2016, J. Cosmology Astropart. Phys, 2016, 004 * Morales & Wyithe (2010) Morales M. F., Wyithe J. S. B., 2010, ARA&A, 48, 127 * Offringa & Smirnov (2017) Offringa A. R., Smirnov O., 2017, Monthly Notices of the Royal Astronomical Society, 471, 301–316 * Parsons et al. (2012) Parsons A., Pober J., McQuinn M., Jacobs D., Aguirre J., 2012, The Astrophysical Journal, 753, 81 * Patil et al. (2017) Patil A. H., et al., 2017, The Astrophysical Journal, 838, 65 * Planck Collaboration et al. (2020) Planck Collaboration et al., 2020, A&A, 641, A1 * Price (2016) Price D. C., 2016, 2.0.0, Astrophysics Source Code Library, 1603.013 * Pritchard & Loeb (2012) Pritchard J. R., Loeb A., 2012, Reports on Progress in Physics, 75, 086901 * Santos et al. (2005) Santos M. G., Cooray A., Knox L., 2005, ApJ, 625, 575 * Shaw et al. (2014) Shaw J. R., Sigurdson K., Pen U.-L., Stebbins A., Sitwell M., 2014, ApJ, 781, 57 * Shaw et al. (2015) Shaw J. R., Sigurdson K., Sitwell M., Stebbins A., Pen U.-L., 2015, Phys. Rev. D, 91, 083514 * Singal et al. (2010) Singal J., Stawarz L., Lawrence A., Petrosian V., 2010, Monthly Notices of the Royal Astronomical Society, 409, 1172–1182 * The HERA Collaboration et al. (2022) The HERA Collaboration et al., 2022, The Astrophysical Journal, 924, 51 * Tingay et al. (2013) Tingay S. J., et al., 2013, Publications of the Astronomical Society of Australia, 30, e007 * Tung (1985) Tung W., 1985, Group Theory in Physics. World Scientific, https://books.google.co.uk/books?id=O89tgpOBO04C * Virtanen et al. (2020) Virtanen P., et al., 2020, Nature Methods, 17, 261 * Wayth et al. (2018) Wayth R. B., et al., 2018, Publications of the Astronomical Society of Australia, 35, e033 * Xu et al. (2022) Xu Z., et al., 2022, The Astrophysical Journal, 938, 128 * Yoshiura et al. (2021) Yoshiura S., et al., 2021, Monthly Notices of the Royal Astronomical Society, 505, 4775–4790 * Zheng et al. (2017) Zheng H., et al., 2017, MNRAS, 464, 3486 * van Haarlem et al. (2013) van Haarlem M. P., et al., 2013, A&A, 556, A2 * van der Walt et al. (2011) van der Walt S., Colbert S. C., Varoquaux G., 2011, Computing in Science Engineering, 13, 22 ## Appendix A Further “realification” Inherently visibilities are complex quantities and the spherical harmonic coefficients are as well. Many standard linear solvers are not set up to handle complex numbers however. The first step to make the system real (i.e. _realify_) has already been done in Sec. 2.1 bt reordering the spherical harmonic coefficient vector, $\boldsymbol{\mathbf{a}}_{\ell m}$, to contain the values of first the real- and then the imaginary part of the spherical harmonic coefficients. This leaves us with a fully real-valued vector to solve for, denoted $\boldsymbol{\mathbf{a}}_{\rm cr}$ (constrained realisations). A real-valued spherical harmonic coefficient vector does not automatically result in real-valued visibilities however. Instead, the visibility response is still complex valued, now containing a complex visibility response per real- and per imaginary part of the $\boldsymbol{\mathbf{a}}_{\rm cr}$-vector. Further _realification_ of the system is therefore needed, while making sure that all the mixing of real and imaginary parts that you get when multiplying complex numbers together is handled correctly. First, we define vectors (bold, lower case) and matrices (bold, upper case) in this real-valued system as $\displaystyle\widetilde{\boldsymbol{\mathbf{v}}}=\begin{pmatrix}\boldsymbol{\mathbf{v}}_{\rm re}\\\ \boldsymbol{\mathbf{v}}_{\rm im}\end{pmatrix},\quad\widetilde{\boldsymbol{\mathbf{M}}}=\begin{pmatrix}\boldsymbol{\mathbf{M}}_{\rm re}&-\boldsymbol{\mathbf{M}}_{\rm im}\\\\[3.0pt] \boldsymbol{\mathbf{M}}_{\rm im}&\boldsymbol{\mathbf{M}}_{\rm re}\end{pmatrix}$ (26) with the Hermitian conjugate of the matrix $\widetilde{\boldsymbol{\mathbf{M}}}$ given as $\displaystyle\widetilde{\boldsymbol{\mathbf{M}}}^{{\dagger}}=\begin{pmatrix}\boldsymbol{\mathbf{M}}_{\rm re}^{T}&\boldsymbol{\mathbf{M}}_{\rm im}^{T}\\\\[3.0pt] -\boldsymbol{\mathbf{M}}_{\rm im}^{T}&\boldsymbol{\mathbf{M}}_{\rm re}^{T}\end{pmatrix},$ (27) where the superscript $T$ denotes the transpose and the minus sign has been swapped due to complex conjugation. With this we can redefine the complex-valued visibility response $\boldsymbol{\mathbf{X}}$ into a purely real-valued version, $\displaystyle\widetilde{\boldsymbol{\mathbf{X}}}=\begin{pmatrix}\boldsymbol{\mathbf{X}}_{\textup{re}}&-\boldsymbol{\mathbf{X}}_{\textup{im}}\\\\[3.0pt] \boldsymbol{\mathbf{X}}_{\textup{im}}&\boldsymbol{\mathbf{X}}_{\textup{re}}\end{pmatrix},\quad\widetilde{\boldsymbol{\mathbf{X}}}^{\dagger}=\begin{pmatrix}\boldsymbol{\mathbf{X}}_{\textup{re}}^{T}&\boldsymbol{\mathbf{X}}_{\textup{im}}^{T}\\\\[3.0pt] -\boldsymbol{\mathbf{X}}_{\textup{im}}^{T}&\boldsymbol{\mathbf{X}}_{\textup{re}}^{T}\end{pmatrix}\,,$ (28) and we define the new noise covariance as a diagonal matrix $\displaystyle\widetilde{\boldsymbol{\mathbf{N}}}=\begin{pmatrix}\boldsymbol{\mathbf{N}}/2&0\\\ 0&\boldsymbol{\mathbf{N}}/2\end{pmatrix}\,.$ (29) Since the spherical harmonic coefficient vector is already constructed so that it is real-valued, as described in Sec. 2.1, so is the prior mean $\boldsymbol{\mathbf{a}}_{0}$ and $\boldsymbol{\mathbf{\omega}}_{a}$, $\displaystyle\widetilde{\boldsymbol{\mathbf{a}}}=\begin{pmatrix}\boldsymbol{\mathbf{a}}_{\rm cr}\\\ 0\end{pmatrix},\quad\widetilde{\boldsymbol{\mathbf{a}}}_{0}=\begin{pmatrix}\boldsymbol{\mathbf{a}}_{0}\\\ 0\end{pmatrix},\quad\widetilde{\boldsymbol{\mathbf{\omega}}}_{a}=\begin{pmatrix}\boldsymbol{\mathbf{\omega}}_{a}\\\ 0\end{pmatrix}\,.$ (30) Lastly, the signal covariance is defined in Sec. 3.3 by the real-valued $\boldsymbol{\mathbf{a}}_{\text{true}}$-vector and so, already, is real-valued itself. We therefore define the new signal covariance as, $\displaystyle\widetilde{\boldsymbol{\mathbf{S}}}=\begin{pmatrix}{\color[rgb]{0,0,0}\boldsymbol{\mathbf{S}}}&0\\\ 0&0\end{pmatrix}\,.$ (31) Using these definitions, we can derive the Gaussian constrained realisation equation once again, and we are left with the final _realified_ GCR equation, $\displaystyle\left[\boldsymbol{\mathbf{S}}^{-1}+2\boldsymbol{\mathbf{X}}_{\textup{re}}^{T}\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{X}}_{\textup{re}}+2\boldsymbol{\mathbf{X}}_{\textup{im}}^{T}\right.$ $\displaystyle\left.\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{X}}_{\textup{im}}\right]\boldsymbol{\mathbf{a}}_{\textup{cr}}=$ $\displaystyle\quad\boldsymbol{\mathbf{X}}_{\textup{re}}^{T}\left(2\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{d}}_{\textup{re}}+\sqrt{2}\boldsymbol{\mathbf{N}}^{-\frac{1}{2}}(\boldsymbol{\mathbf{\omega}}_{d})_{\textup{re}}\right)$ $\displaystyle\quad+\boldsymbol{\mathbf{X}}_{\textup{im}}\left(2\boldsymbol{\mathbf{N}}^{-1}\boldsymbol{\mathbf{d}}_{\textup{im}}+\sqrt{2}\boldsymbol{\mathbf{N}}^{-\frac{1}{2}}(\boldsymbol{\mathbf{\omega}}_{d})_{\textup{im}}\right)$ $\displaystyle\quad+\boldsymbol{\mathbf{S}}^{-1}\boldsymbol{\mathbf{a}}_{0}+\boldsymbol{\mathbf{S}}^{-\frac{1}{2}}\boldsymbol{\mathbf{\omega}}_{a}.$ (32)
# Ultra diffuse galaxies in the Hydra I cluster from the LEWIS Project: Phase- Space distribution and globular cluster richness Duncan A. Forbes, 1 Jonah Gannon1, Enrichetta Iodice2, Michael Hilker5,Goran Doll2,3, Chiara Buttitta2, Antonio La Marca6,7, Magda Arnaboldi5, Michele Cantiello4, G. D’Ago8, Jesus Falcon Barroso15,16, Laura Greggio9, Marco Gullieuszik9, Johanna Hartke12,13, Steffen Mieske10, Marco Mirabile4,14, Roberto Rampazzo9, Marina Rejkuba5, Marilena Spavone2, Chiara Spiniello11, Giulio Capasso2 1 Centre for Astrophysics & Supercomputing, Swinburne University, Hawthorn, VIC 3122, Australia 2 INAF - Astronomical Observatory of Capodimonte, Salita Moiariello 16, I-80131, Naples, Italy 3 University of Naples “Federico II”, C.U. Monte Sant’Angelo, Via Cinthia, 80126, Naples, Italy 4 INAF - Astronomical Observatory of Abruzzo, Via Maggini, 64100, Teramo, Italy 5 European Southern Observatory, Karl-Schwarzschild-Strasse 2, 85748 Garching bei Muenchen, Germany 6 SRON Netherlands Institute for Space Research, Landleven 12, 9747 AD Groningen, The Netherlands 7 Kapteyn Astronomical Institute, University of Groningen, Postbus 800, 9700 AV Groningen, The Netherlands 8 Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge, CB3 0HA 9 Osservatorio Astronomico di Padova, Via dell’Osservatorio 8, 36012 Asiago (VI), Italy 10 European Southern Observatory, Alonso de Cordova 3107, 7630355 Vitacura, Santiago, Chile 11 Sub-department of Astrophysics, University of Oxford, Denys Wilkinson Building, Keble Road, Oxford OX1 3RH, United Kingdom 12 Finnish Centre for Astronomy with ESO (FINCA), FI-20014 University of Turku, Finland 13 Tuorla Observatory, Department of Physics and Astronomy, FI-20014 University of Turku, Finland 14 Gran Sasso Science Institute, L’Aquila, Italy 15 INAF - Astronomical Observatory of Padova, Vicolo dell’Osservatorio 5, I-35122 Padova, Italy 15 Instituto de Astrofísica de Canarias, Vía Láctea s/n, E-38205 La Laguna, Tenerife, Spain 16 Departamento de Astrofísica, Universidad de La Laguna, E-38200 La Laguna, Tenerife, Spain E-mail<EMAIL_ADDRESS> (Accepted XXX. Received YYY; in original form ZZZ) ###### Abstract Although ultra diffuse galaxies (UDGs) are found in large numbers in clusters of galaxies, the role of the cluster environment in shaping their low surface brightness and large sizes is still uncertain. Here we examine a sample of UDGs in the Hydra I cluster (D = 51 Mpc) with new radial velocities obtained as part of the LEWIS (Looking into the faintest with MUSE) project using VLT/MUSE data. Using a phase-space, or infall diagnostic, diagram we compare the UDGs to other known galaxies in the Hydra I cluster and to UDGs in other clusters. The UDGs, along with the bulk of regular Hydra I galaxies, have low relative velocities and are located near the cluster core, and thus consistent with very early infall into the cluster. Combining with literature data, we do not find the expected trend of GC-rich UDGs associated with earlier infall times. This result suggests that quenching mechanisms other than cluster infall should be further considered, e.g. quenching by strong feedback or in cosmic sheets and filaments. Tidal stripping of GCs in the cluster environment also warrants further modelling. ###### keywords: galaxies: star clusters: general — galaxies: haloes — galaxies: structure — galaxies: photometry ††pubyear: 2023††pagerange: Ultra diffuse galaxies in the Hydra I cluster from the LEWIS Project: Phase-Space distribution and globular cluster richness–References ## 1 Introduction The possible formation pathways of ultra-diffuse galaxies (UDGs) have been a subject of an ongoing vigorous debate since 2015, when a population of these extremely diffuse galaxies was identified in the Coma cluster using the Dragonfly Telephoto Array (van Dokkum et al., 2015). Existing in all environments, they are most common in clusters with several hundred found in the Coma cluster (Yagi et al., 2016; Alabi et al., 2018). This significant contribution to our ‘census of galaxies’ has prompted numerous simulation studies and accompanying predictions (see Sales et al. (2020) and references therein). These simulations can be broadly placed in two categories; internal processes (e.g. episodic supernova feedback) or external (e.g. tidal effects in a dense environment). Some combination of both processes may be operating along with past galaxy infall (and subsequent quenching) into clusters. UDGs have low surface brightnesses (they are defined to have central values in the g band of $\mu$ $>$ 24 mag. per sq. arcsec) so that spectroscopic studies of them push even 8–10m class telescopes, with efficient low surface brightness instruments, such as KWCI on Keck or MUSE on VLT, to their limits. While strictly speaking dwarf galaxies with M∗ $<10^{9}$ M⊙, UDGs are unlike classical dwarfs as they have extreme sizes with effective radii Re $>$ 1.5 kpc (i.e. comparable to the disk of the Milky Way with Re $\sim$3.5 kpc). They also reveal another unexplained feature, with some hosting up to ten times more globular clusters (GCs) than classical dwarf galaxies of the same luminosity (Forbes et al., 2020). Their very existence in clusters and their generally old stellar populations suggests that some may be protected within an overly massive dark matter halo. The latter is supported by the correlation between GC numbers and host galaxy halo mass for normal galaxies, e.g. Burkert & Forbes (2020). In the standard picture of dwarf galaxy evolution (Mistani et al., 2016), dwarfs that fell into clusters at early times will have experienced intense star formation, prior to, or at the start of, infall (which is also expected to give rise to a high fraction of stars in bound star clusters). This is followed by quenching of any further star formation as the infall proceeds. Both of these effects would lead to a high fraction of GCs relative to their field stars (Mistani et al., 2016; Ramos-Almendares et al., 2020). Indeed trends of GC richness and [$\alpha$/Fe] ratios with clustercentric radius provide some observational support for this interpretation (Peng et al., 2008; Liu et al., 2016; Romero-Gómez et al., 2023). This early-infall, or biasing, has been invoked for UDGs by Carleton et al. (2021) who include cluster tidal effects within the IllustrisTNG simulation and simplified GC formation physics. Similar to classical dwarfs, they predict that early-infall UDGs should be rich in GCs. Based on a semi-empirical model, Trujillo-Gomez et al. (2022) also predict that galaxies near the cluster core form more GCs. Using phase-space, or infall diagnostic, diagrams of the type proposed by Rhee et al. (2017) one can investigate whether GC richness depends on UDG cluster infall time. No trend between GC richness and very early infall times might suggest that GC formation and quenching occurred before cluster infall. While low mass galaxies typically quench at late times, there is a considerable range in quenching times with some low mass galaxies quenching at z $\sim$ 2 or 10.5 Gyr ago (Moster et al., 2020). Quenching at early times via stellar feedback (Stinson et al., 2013) may be one possibility. This early quenching applied to UDGs has been described by Danieli et al. (2022). Another possibility may be quenching via the interaction with cosmic sheets or filaments (Pasha et al., 2023). A first attempt at this sort of infall analysis applied to GCs was presented in Gannon et al. (2022) for several UDGs in the Coma and Perseus clusters. No clear signal was found but the sample was small with just over a dozen UDGs and with a bias towards GC-rich UDGs. In this Letter we examine the infall diagnostic diagram for a new sample of UDGs in the Hydra I cluster (A1060; D = 51 $\pm$ 4 Mpc). The Hydra I cluster appears to be fairly dynamically relaxed (Ventimiglia et al., 2011) but also reveals hints of substructures (Lima-Dias et al., 2021; La Marca et al., 2022a), an infalling group of galaxies (Arnaboldi et al., 2012), and evidence for ram pressure stripping (Wang et al., 2021; Iodice et al., 2021). The observed UDGs are located near the cluster core and the northern subgroup, with all lying within the 0.3 virial radii (R200) of the Hydra I cluster centre. Each was observed using MUSE on the VLT as part of the ongoing LEWIS (Looking into the faintEst WIth MUSE) project. Details of the project, including galaxy radial velocities, positions, GC counts etc, are given in Paper I by Iodice et al. (2023, in press). The GC counts are based on deep, optical multi-filter imaging with the VST as part of the VEGAS project (Iodice et al., 2020) and will be updated after the full analysis of the MUSE data. Here we explore the distribution of UDGs in phase space and investigate whether they reveal any trend in this space with their GC richness. We also include similar data for UDGs in other nearby clusters. For the Hydra I cluster we adopt the same parameters as used by La Marca et al. (2022b), i.e. cz = 3683 $\pm$ 46 km s-1, $\sigma$ = 724 $\pm$ 31 km s-1, and virial parameters R200 = 1.6 Mpc, M200 = 2 $\times$ 1014 h-1 M⊙ and take its centre as NGC 3311 (RA = 159.17842, Dec = –27.528339). These values are similar to those found by Lima-Dias et al. (2021) who recently studied Hydra I galaxies out to the virial radius. Figure 1: Infall diagnostic diagram for non-UDG (giants, dwarfs and LSB galaxies) and UDGs in the Hydra I cluster. The diagram shows the relative line-of-sight velocity of each galaxy normalised by the cluster velocity dispersion against the projected radius normalised by the virial radius. Regions of the diagram are shaded according to their infall times from the cosmological simulations of Rhee et al. (2017) as indicated in the legend. The plot shows that most UDGs and non-UDG galaxies of the the Hydra I cluster lie within the very early infall zone – the simulations of indicating that around half of the galaxies in this zone were part of the cluster at least 6.45 Gyr ago. ## 2 Infall Diagnostic Diagram for Hydra I Cluster Galaxies Rhee et al. (2017) carried out cosmological simulations of several galaxy clusters and examined the resulting distribution of galaxies in phase-space (i.e. velocity of the galaxy relative to the mean cluster velocity normalised by the cluster velocity dispersion versus the galaxy’s clustercentric radius normalised by the cluster virial radius). Based on the infall time of galaxies, they divided this diagram into several infall zones, ranging from those that fell into the cluster at very early times, to those that are yet to fall in. Thus the location of galaxies in this diagram provides an ‘infall diagnostic’ which is statistical in nature and additional scatter is introduced when using 2D projected radii (as is the case for observational data). For example, the ‘very early infall’ (or ancient infaller) zone in the simulation is occupied by a slight majority (52%) of galaxies that have resided in the cluster for more than 6.45 Gyr. Projection effects mean that the true clustercentric radius for some galaxies is larger in 3D than observed in 2D. For most galaxies this effect should be less a factor of two from the projected one. In Fig. 1 we show such an infall diagnostic diagram for all galaxies in the Hydra I cluster out to half the virial radius. This includes giant and dwarf galaxies from the study of Christlein & Zabludoff (2003) plus the addition of UDGs and 3 low surface brightness (LSB) galaxies that have UDG-like sizes but are slightly brighter from Iodice et al. (2023, in press). We find that the bulk of the non-UDG Hydra I galaxies are located within the ‘very early infall’ zone. The simulation of Rhee et al. (2017) predicts that just over half of these would have been part of the cluster for at least 6.45 Gyr. There are also galaxies located in later infall zones and three galaxies that may lie outside of the cluster with large relative velocities – these could be backsplash galaxies (having passed through the cluster) or simply galaxies that are yet to fall into the cluster. If we examine giant and classical dwarf galaxies separately (divided at MR = –18 or mR = 15.5) there is no clear difference between them in terms of their infall properties. Compared to the UDGs they appear to scatter to higher relative velocities on average. A more quantitative measure of the differences in their infall properties can be obtained from the product of their relative velocity from the cluster mean and their radial position: $\Delta$V/$\sigma$ x R/R200. Restricting to 0.3R/R200, as probed by the imaging, we find mean values (and error on the mean) of $\Delta$V/$\sigma$ x R/R200 = 0.83 ($\pm$ 0.07) $\times$ 0.15 ($\pm$ 0.01) = 0.12 ($\pm$ 0.02) for giant galaxies and 0.88 ($\pm$ 0.06) $\times$ 0.16 ($\pm$ 0.01) = 0.14 ($\pm$ 0.02) for classical dwarfs. For the UDGs the mean value is $\Delta$V/$\sigma$ x R/R200 = 0.80 ($\pm$ 0.17) $\times$ 0.16 ($\pm$ 0.02) = 0.13 ($\pm$ 0.04). This indicates that UDGs are similarly concentrated in phase-space to the other cluster galaxies. Also, while UDGs have a similar distribution in clustercentric radius, their velocities are closer to the cluster mean than either giants or classical dwarfs. We note that Lima-Dias et al. (2021) also found passive early-type galaxies to be concentrated in the cluster core.The LSB galaxies in Fig. 1 are found in a range of infall zones, from early to late infall. As might be expected from their inner cluster position, our UDGs were among the earliest inhabitants of the cluster, infalling at least 6.45 Gyr ago according to simulations of Rhee et al. (2017). They would be expected to have star formation (SF) histories that indicate early quenching. A preliminary analysis by Iodice et al. (2023, submitted) for one UDG (UDG11) indicates an old age of $\sim$10 Gyr, suggestive of early quenching. Future analysis will also include the [$\alpha$/Fe] ratios which appears to be a sensitive indicators of SF histories for low mass galaxies (see Ferre-Mateu et al. 2023, submitted for results on UDGs in other clusters and Romero-Gómez et al. (2023), for dwarf galaxies in the Fornax cluster). We note that the study of Lima-Dias et al. (2021) found 88% of Hydra I galaxies (with log M∗ $>$ 8.5) to be quenched, i.e. no sign of ongoing star formation. Figure 2: Infall diagnostic diagram for only UDGs in the Hydra I, Coma, Virgo and Perseus clusters. Regions of the diagram are shaded according to their infall times from the cosmological simulations of Rhee et al. (2017). As per the legend, UDGs in different clusters are denoted by different symbols. Symbols are outlined in red (if GC-rich) or blue (if GC-poor), and without an outline if the GC properties are unknown. See main text for discussion of selection effects in the UDG samples. Globular cluster (GC) rich UDGs are not predominately found in the very early infall region, indeed the data suggest that very early infall UDGs tend to be GC-poor. ## 3 Infall Diagnostic Diagram for UDGs in Several Clusters In Fig. 2 we show the UDGs from the Hydra I cluster along with those from the literature and coded by globular cluster (GC) richness. Total GC counts for the Hydra I UDGs are determined in (Iodice et al., 2020; La Marca et al., 2022b) and listed again in Paper I (Iodice et al. 2023, submitted). Literature data comes from Gannon et al. (2022) and the recent work of Toloba et al. (2023). The GC counts are almost exclusively based on imaging (i.e. lacking radial velocities) and we follow Gannon et al. (2022) assigning a somewhat arbitrary separation between rich and poor GC systems at 20 GCs. This corresponds to a halo mass of 1011 M⊙ using the scaling relation of Burkert & Forbes (2020). Below 20 GCs the scaling relation is less predictive of halo mass due to increased scatter. By this definition, all of the UDGs in the Hydra I cluster are GC-poor (ranging from no GCs for several UDGs to 15 GCs for UDG3) and this is unlikely to change significantly when the full set of MUSE spectroscopic data is available. Given the relatively small stellar mass range of the Hydra I UDGs, a fixed GC number corresponds closely to a GC system total mass per host galaxy stellar mass. If we assume the same average mass for a GC of 105 M⊙, this ratio is $<$1.2% for all of the observed Hydra I UDGs. While some Coma cluster UDGs also have a ratio $<$1.2% the majority have much higher ratios, with up to $\sim$10% of the galaxy stellar mass in their GC system, see figure 4 of Forbes et al. (2020). Before interpreting Fig. 2 there are various caveats and selection effects that should be born in mind. Firstly, we note that some of the literature UDGs lack firm GC counts and their rich/poor status is on the basis of a visual estimate only (Gannon et al., 2022). Secondly, the literature sample is subject to sample selection effects. The Coma cluster sample of UDGs comes from studies that have focused on GC-rich galaxies or they have focused on a narrow range in clustercentric radius (i.e. around 0.12 R/R200 in the Coma cluster). Observations of the Perseus cluster UDGs have so far avoided the cluster inner regions. The Virgo UDG sample is relatively small and mostly GC- poor. In terms of a selection bias, the Hydra I UDGs are the closest to being a representative sample of UDGs in the cluster, however only the inner 0.3 R/R200 was imaged in Iodice et al. (2020). Thus, we may be missing the late infalling UDGs. We note that La Marca et al. (2022b) estimated a total UDG population out to the virial radius of 48 $\pm$ 10 and so many outer region UDGs, which may be late infallers, remain to be studied. The UDG infall diagram does not clearly show GC-rich UDGs to be located in earlier infall zones as might be expected in the standard picture of dwarf galaxy quenching due to infall which leads to richer GC systems (as described in the Introduction). Indeed, the opposite trend may be present, such that in the very early infall region there are 13 GC-poor UDGs and 5 GC-rich ones, whereas outside of this region (but within 0.5 R/R200) there are only 6 GC- poor and 6 GC-rich UDGs. Again, we caution that selection and projection effects make conclusions tentative. ## 4 Discussion Alabi et al. (2018) used the phase-space diagram to investigate the infall epoch of UDGs, classical dwarfs and other galaxies in the Coma cluster (a massive, dynamically relaxed cluster). Similar to the Hydra I cluster, they saw little difference between classical dwarfs and the giant galaxies. For the UDGs, they identified both early and late infallers. A similar situation might be present for Hydra I UDGs if outer region UDGs were probed. Alabi et al. (2018) did not include GC richness in their study. Given the lack of a clear signal for ‘infall bias’ in the GC richness of UDGs alternatives should be further investigated. As noted in the Introduction, quenching at very early times prior to cluster infall should be considered. For such UDGs, we would expect very old ages, low metallicities (similar to the metal-poor subpopulation of GCs) and high alpha overabundances (indicative of rapid star formation). A high fraction of mass in GCs relative to field stars might also be expected. A UDG in the NGC 5846 group, (NGC5846$\\_$UDG1) discovered in VEGAS imaging (Forbes et al., 2019), may be an example of such a failed galaxy having a remarkable 13% of its current stellar mass in the form of GCs (Danieli et al., 2022). As noted above, the observed Hydra I UDGs (from the inner cluster regions) all have less than 1.2% of their stellar mass in GCs. Another possibility is that the Hydra I UDGs are GC-poor because they have been tidally stripped from their host galaxy. This tidal stripping would have to remove most of the dark matter halo before any GCs, since the dark matter is more radially extended than GC systems. Continued stripping would be expected to remove GCs and stars in roughly equal proportions since the radial extent of GC systems for UDGs closely follows that of the galaxy stars. As well as operating in clusters, tidal stripping of UDGs may occur in galaxy groups. We note that UDGs in the field do tend to be GC-poor (Jones et al., 2023) however this is unlikely to be due to tidal effects and rather some internal process. The Hydra I UDGs are generally well-fit by a single Sersic profile however a few show hints of asymmetries that might point to a tidal interaction (Iodice et al., 2020; La Marca et al., 2022a). For the one UDG examined in detail by Iodice et al. (2023, submitted) there is some evidence for an isophotal twist in the MUSE data. This might indicate tidal interaction (or a triaxial potential). Furthermore, a Hydra I UDG first identified by Misgeld et al. (2008) reveals a clear S-shape indicative of ongoing tidal interaction (Koch et al., 2012). In the case of Coma cluster UDGs, Mowla et al. (2017) looked specifically for signs of tidal features via position angle twists in a stacked sample, finding no evidence for such twists. Sales et al. (2020) have simulated UDGs in clusters of similar mass to Hydra I using Illustris-TNG100. They identify two types of UDGs in clusters, i.e. Tidal-UDGs and Born-UDGs (see also Jiang et al. 2019). The Tidal-UDGs were originally massive galaxies (up to 1010 M⊙) that have been tidally stripped of stars and puffed-up by the cluster. Born-UDGs were formed as UDGs outside of the cluster and more recently entered the cluster. Thus Tidal-UDGs dominate the inner $\sim$0.5R/R200 since they were accreted at early times, while Born- UDGs dominate the outer regions with some only recently falling into the cluster. We remind the reader that we only probe out to 0.3R/R200 in Hydra I. The Sales et al. (2020) model would also predict on average higher metallicities, older ages and lower internal velocity dispersions, for a given stellar mass, for their Tidal-UDG compared to the Born-UDGs. These stellar population, kinematic, GC colours and dark matter content predicted for Tidal- UDGs can be tested when the full LEWIS dataset is available. ## 5 Conclusions As part of the LEWIS project (Iodice et al. 2023, in press) we obtained new VLT/MUSE observations of the radial velocities of UDGs in the Hydra I cluster (at D = 51 Mpc). Here we examine the location of Hydra I UDGs in infall phase- space diagrams based on simulations of cluster galaxies. We find all of the observed UDGs (and 3 low surface brightness galaxies) to be associated with the cluster. From comparison with the Rhee et al. (2017) simulations, we conclude that most giants, classical dwarfs and UDGs fell into the Hydra I cluster long ago, with UDGs being among the earliest infallers. Projection effects in observations and the statistical nature of the infall diagnostic diagram limit our ability to determine the true fraction of ancient infallers. Nevertheless we might expect UDGs in the Hydra I cluster to reveal old stellar populations consistent with early quenching. We also compare Hydra I UDGs with their counterparts in the Coma, Perseus and Virgo clusters in terms of their GC richness. If very early infall into a cluster is associated with enhanced GC richness (as has been suggested for classical dwarf galaxies) then such a trend is expected. The data from these clusters do not show a clear trend of GC richness with earlier infall times, indeed the data suggest the opposite trend. If verified by larger and more complete samples, then UDGs may be quenched by a different mechanism than that thought to operate on classical dwarf galaxies. As more data for UDGs is acquired, trends, or the lack of, may become more apparent in an infall diagnostic diagram. A future analysis of star formation histories will give an indication of when quenching occurred for the Hydra I UDGs. Once the full dataset of the LEWIS project is available we will be able to test other mechanisms, such as pre-infall quenching and/or tidal stripping, and their possible role in shaping UDGs and their globular cluster systems. ## Acknowledgements We wish to thank the anonymous referee for their comments. We thank A. Romanowsky, L. Buzzo, L. Haacke and O. Gerhard for useful suggestions. This work is based on visitor mode observations collected at the European Southern Observatory (ESO) La Silla Paranal Observatory and collected at the European Southern Observatory under ESO programmes 099.B-0560(A) and 108.222P. INAF authors acknowledge financial support for the VST project (P.I.: P. Schipani). DAF thanks the ARC for support via DP220101863. Parts of this research were supported by the Australian Research Council Centre of Excellence for All Sky Astrophysics in 3 Dimensions (ASTRO 3D), through project number CE170100013. MC acknowledges support from the INAF-EDGE program (PI Leslie K. Hunt). J. F-B acknowledges support through the RAVET project by the grant PID2019-107427GB-C32 from the Spanish Ministry of Science, Innovation and Universities (MCIU), and through the IAC project TRACES which is partially supported through the state budget and the regional budget of the Consejería de Economía, Industria, Comercio y Conocimiento of the Canary Islands Autonomous Community. ## Data Availability Raw data is available from the ESO archive. ## References * Alabi et al. (2018) Alabi A., et al., 2018, MNRAS, 479, 3308 * Arnaboldi et al. (2012) Arnaboldi M., Ventimiglia G., Iodice E., Gerhard O., Coccato L., 2012, A&A, 545, A37 * Burkert & Forbes (2020) Burkert A., Forbes D. A., 2020, AJ, 159, 56 * Carleton et al. (2021) Carleton T., Guo Y., Munshi F., Tremmel M., Wright A., 2021, MNRAS, 502, 398 * Christlein & Zabludoff (2003) Christlein D., Zabludoff A. I., 2003, ApJ, 591, 764 * Danieli et al. (2022) Danieli S., et al., 2022, ApJ, 927, L28 * Forbes et al. (2019) Forbes D. A., Gannon J., Couch W. J., Iodice E., Spavone M., Cantiello M., Napolitano N., Schipani P., 2019, A&A, 626, A66 * Forbes et al. (2020) Forbes D. A., Alabi A., Romanowsky A. J., Brodie J. P., Arimoto N., 2020, MNRAS, 492, 4874 * Gannon et al. (2022) Gannon J. S., et al., 2022, MNRAS, 510, 946 * Iodice et al. (2020) Iodice E., et al., 2020, A&A, 642, A48 * Iodice et al. (2021) Iodice E., et al., 2021, A&A, 652, L11 * Jiang et al. (2019) Jiang F., Dekel A., Freundlich J., Romanowsky A. J., Dutton A. A., Macciò A. V., Di Cintio A., 2019, MNRAS, 487, 5272 * Jones et al. (2023) Jones M. G., et al., 2023, ApJ, 942, L5 * Koch et al. (2012) Koch A., Burkert A., Rich R. M., Collins M. L. M., Black C. S., Hilker M., Benson A. J., 2012, ApJ, 755, L13 * La Marca et al. (2022a) La Marca A., et al., 2022a, A&A, 659, A92 * La Marca et al. (2022b) La Marca A., et al., 2022b, A&A, 665, A105 * Lima-Dias et al. (2021) Lima-Dias C., et al., 2021, MNRAS, 500, 1323 * Liu et al. (2016) Liu Y., et al., 2016, ApJ, 818, 179 * Misgeld et al. (2008) Misgeld I., Mieske S., Hilker M., 2008, A&A, 486, 697 * Mistani et al. (2016) Mistani P. A., et al., 2016, MNRAS, 455, 2323 * Moster et al. (2020) Moster B. P., Naab T., White S. D. M., 2020, MNRAS, 499, 4748 * Mowla et al. (2017) Mowla L., van Dokkum P., Merritt A., Abraham R., Yagi M., Koda J., 2017, ApJ, 851, 27 * Pasha et al. (2023) Pasha I., Mandelker N., van den Bosch F. C., Springel V., van de Voort F., 2023, MNRAS, 520, 2692 * Peng et al. (2008) Peng E. W., et al., 2008, ApJ, 681, 197 * Ramos-Almendares et al. (2020) Ramos-Almendares F., Sales L. V., Abadi M. G., Doppel J. E., Muriel H., Peng E. W., 2020, MNRAS, 493, 5357 * Rhee et al. (2017) Rhee J., Smith R., Choi H., Yi S. K., Jaffé Y., Candlish G., Sánchez-Jánssen R., 2017, ApJ, 843, 128 * Romero-Gómez et al. (2023) Romero-Gómez J., et al., 2023, MNRAS, 522, 130 * Sales et al. (2020) Sales L. V., Navarro J. F., Peñafiel L., Peng E. W., Lim S., Hernquist L., 2020, MNRAS, 494, 1848 * Stinson et al. (2013) Stinson G. S., Brook C., Macciò A. V., Wadsley J., Quinn T. R., Couchman H. M. P., 2013, MNRAS, 428, 129 * Toloba et al. (2023) Toloba E., et al., 2023, arXiv e-prints, p. arXiv:2305.06369 * Trujillo-Gomez et al. (2022) Trujillo-Gomez S., Kruijssen J. M. D., Reina-Campos M., 2022, MNRAS, 510, 3356 * Ventimiglia et al. (2011) Ventimiglia G., Arnaboldi M., Gerhard O., 2011, A&A, 528, A24 * Wang et al. (2021) Wang J., et al., 2021, ApJ, 915, 70 * Yagi et al. (2016) Yagi M., Koda J., Komiyama Y., Yamanoi H., 2016, ApJS, 225, 11 * van Dokkum et al. (2015) van Dokkum P. G., Abraham R., Merritt A., Zhang J., Geha M., Conroy C., 2015, ApJ, 798, L45
# Falcon 2.0: An Entity and Relation Linking Tool over Wikidata Ahmad Sakor<EMAIL_ADDRESS>L3S Research Center and TIB, University of HannoverHannoverGermany , Kuldeep Singh<EMAIL_ADDRESS>Cerence GmbH and Zerotha ResearchP.O. Box 1212AachenGermany , Anery Patel <EMAIL_ADDRESS>TIB, University of HannoverHannoverGermany and Maria- Esther Vidal<EMAIL_ADDRESS>L3S Research Center and TIB, University of HannoverHannoverGermany (2020) ###### Abstract. The Natural Language Processing (NLP) community has significantly contributed to the solutions for entity and relation recognition from a natural language text, and possibly linking them to proper matches in Knowledge Graphs (KGs). Considering Wikidata as the background KG, there are still limited tools to link knowledge within the text to Wikidata. In this paper, we present Falcon 2.0, the first joint entity and relation linking tool over Wikidata. It receives a short natural language text in the English language and outputs a ranked list of entities and relations annotated with the proper candidates in Wikidata. The candidates are represented by their Internationalized Resource Identifier (IRI) in Wikidata. Falcon 2.0 resorts to the English language model for the recognition task (e.g., N-Gram tiling and N-Gram splitting), and then an optimization approach for the linking task. We have empirically studied the performance of Falcon 2.0 on Wikidata and concluded that it outperforms all the existing baselines. Falcon 2.0 is open source and can be reused by the community; all the required instructions of Falcon 2.0 are well-documented at our GitHub repository111https://github.com/SDM-TIB/falcon2.0. We also demonstrate an online API, which can be run without any technical expertise. Falcon 2.0 and its background knowledge bases are available as resources at https://labs.tib.eu/falcon/falcon2/. NLP, Entity Linking, Relation Linking, Background Knowledge, English morphology, DBpedia, and Wikidata ††journalyear: 2020††copyright: rightsretained††conference: Proceedings of the 29th ACM International Conference on Information and Knowledge Management; October 19–23, 2020; Virtual Event, Ireland††booktitle: Proceedings of the 29th ACM International Conference on Information and Knowledge Management (CIKM ’20), October 19–23, 2020, Virtual Event, Ireland††doi: 10.1145/3340531.3412777††isbn: 978-1-4503-6859-9/20/10††ccs: Information systems Resource Description Framework (RDF)††ccs: Information systems Information extraction ## 1\. Introduction Entity Linking (EL)- also known as Named Entity Disambiguation (NED)- is a well-studied research domain for aligning unstructured text to its structured mentions in various knowledge repositories (e.g., Wikipedia, DBpedia (Auer et al., 2007), Freebase (Bollacker et al., 2008) or Wikidata (Vrandecic, 2012)). Entity linking comprises two sub-tasks. The first task is Named Entity Recognition (NER), in which an approach aims to identify entity labels (or surface forms) in an input sentence. Entity disambiguation is the second sub- task of linking entity surface forms to semi-structured knowledge repositories. With the growing popularity of publicly available knowledge graphs (KGs), researchers have developed several approaches and tools for EL task over KGs. Some of these approaches implicitly perform NER and directly provide mentions of entity surface forms in the sentences to the KG (often referred to as end-to-end EL approaches) (Delpeuch, 2019). Other attempts (e.g., Yamanda et al. (Yamada et al., 2016a), DCA (Yang et al., 2019)) consider recognized surface forms of the entities as additional inputs besides the input sentence to perform entity linking. Irrespective of the input format and underlying technologies, the majority of the existing attempts (Röder et al., 2018) in the EL research are confined to well-structured KGs such as DBpedia or Freebase222it is now depreciated and no further updates are possible. These KGs rely on a well-defined process to extract information directly from Wikipedia infoboxes. They do not provide direct access to the users to add/delete the entities or alter the KG facts. Wikidata, on the other hand, also allows users to edit Wikidata pages directly, add newer entities, and define new relations between the objects. Wikidata is hugely popular as a crowdsourced collection of knowledge. Since its launch in 2012, over 1 billion edits have been made by the users across the world333https://www.wikidata.org/wiki/Wikidata:Statistics. #### Motivation, Approach, and Contributions. We motivate our work by the fact that despite the vast popularity of Wikidata, there are limited attempts to target entity and relation linking over Wikidata. For instance, there are over 20 entity linking tools/APIs for DBpedia (Singh et al., 2018b; Röder et al., 2018), which are available as APIs. To the best of our knowledge, there exists only one open-source API for Wikidata entity linking (i.e., OpenTapioca (Delpeuch, 2019)). Furthermore, there is no tool over Wikidata for relation linking, i.e., linking predicate surface forms to their corresponding Wikidata mentions. In this paper, we focus on providing Falcon 2.0, a reusable resource API for joint entity and relation linking over Wikidata. In our previous work, we proposed Falcon (Sakor et al., 2019), a rule-based approach yet effective for entity and relation linking on short text (questions in this case) over DBpedia. In general, the Falcon approach has two novel concepts: 1) a linguistic-based approach that relies on several English morphology principles such as tokenization, and N-gram tiling; 2) a local knowledge base which serves as a source of background knowledge (BK). This knowledge base is a collection of entities from DBpedia. We resort to the Falcon approach for developing Falcon 2.0. Our aim here is to study whether or not the Falcon approach is agnostic to underlying KG; hence, we do not claim novelty in the underlying _linguistic-based approach_ for Falcon 2.0. Further, we investigate the concerns related to robustness, emerging failures, and bottlenecks. We introduce Falcon 2.0 based on the methodology employed in the first version. Our tool is the first joint entity and relation linking tool for Wikidata. Our novel contributions briefly lie in two aspects: 1. (1) Falcon 2.0: The first resource for joint entity and relation linking over Wikidata. Falcon 2.0 relies on fundamental principles of English morphology (tokenization and compounding) and links entity and relation surface forms in a short sentence to its Wikidata mentions. Falcon 2.0 is available as an online API and can be accessed at https://labs.tib.eu/falcon/falcon2/. Falcon 2.0 is also able to recognize entities in keywords such as Barack Obama, where there is no relation. We empirically evaluate Falcon 2.0 on three datasets tailored for Wikidata. According to the observed results, Falcon 2.0 significantly outperforms all the existing baselines. For the ease of use, we integrate the Falcon API444https://labs.tib.eu/falcon/ into Falcon 2.0. This option is available in case Wikipedia contains an equivalence entity (Wikidata is a superset of DBpedia) The Falcon 2.0 API already has over half a million hits from February 2020 to the time of paper acceptance, which shows its gaining usability (excluding self-access of the API while performing the evaluation). 2. (2) Falcon 2.0 Background KG: We created a new background KG of Falcon 2.0 with the Wikidata. We extracted 48,042,867 Wikidata entities from its public dump and aligned these entities with the aliases present in Wikidata. For example, Barack Obama is a Wikidata entity Wiki:Q76555https://www.wikidata.org/wiki/Q76. We created a mapping between the label (Barack Obama) of Wiki:Q76 with its aliases such as President Obama, Barack Hussein Obama, and Barry Obama and stored it in the background knowledge base. We implemented a similar alignment for 15,645 properties/relations of Wikidata. The background knowledge base is an indexed graph and can be queried. The resource is also present at a persistent URI for further reuse666https://doi.org/10.6084/m9.figshare.11362883. The rest of this paper is organized as follows: Section 2 reviews the state- of-the-art, and the following Section 3 describes our two resources and approach to build Falcon 2.0. Section 4 presents experiments to evaluate the performance of Falcon 2.0. Section 5 presents the importance and impact of this work for the research community. The availability and sustainability of resources is explained in Section 6, and its maintenance related discussion is presented in Section 7. We close with the conclusion in Section 8. ## 2\. Related Work Several surveys provide a detailed overview of the advancements of the techniques employed in entity linking over KGs (Shen et al., 2015; Balog, 2018). Various reading lists (Ji, 2019), online forums777http://nlpprogress.com/english/entity_linking.html and Github repositories888https://github.com/sebastianruder/NLP- progress/blob/master/english/entity_linking.md track the progress in the domain of entity linking. Initial attempts in EL considered Wikipedia as an underlying knowledge source. The research field has matured and the SOTA nearly matches human-level performance (Raiman and Raiman, 2018). With the advent of publicly available KGs such as DBpedia, Yago, and Freebase, the focus has shifted towards developing EL over knowledge graphs. The developments in Deep Learning have introduced a range of models that carry out both NER and NED as a single end-to-end step (Kolitsas et al., 2018; Ganea and Hofmann, 2017). NCEL (Cao et al., 2018) learns both local and global features from Wikipedia articles, hyperlinks, and entity links to derive joint embeddings of words and entities. These embeddings are used to train a deep Graph Convolutional Network (GCN) that integrates all the features through a Multi-layer Perceptron. The output is passed through a Sub-Graph Convolution Network, which finally resorts to a fully connected decoder. The decoder maps the output states to linked entities. The BI-LSTM+CRF model (Inan and Dikenelli, 2018) formulates entity linking as a sequence learning task in which the entity mentions are a sequence whose length equals the series of the output entities. Albeit precise, deep learning approaches demand _high- quality_ training annotations, which are not extensively available for Wikidata entity linking (Cetoli et al., 2019; Mulang et al., 2020). There is concrete evidence in the literature that the machine learning-based models trained over generic datasets such as WikiDisamb30 (Ferragina and Scaiella, 2010), and CoNLL (YAGO) (Hoffart et al., 2011) do not perform well when applied to short texts. Singh et al. (Singh et al., 2018b) evaluated more than 20 entity linking tools over DBpedia for short text (e.g., questions) and concluded that issues like capitalization of surface forms, implicit entities, and multi-word entities affect the performance of EL tools in a short input text. Sakor et al. (Sakor et al., 2019) addresses specific challenges of short texts by applying a rule-based approach for EL over DBpedia. In addition to linking entities to DBpedia, Sakor et al. also provides DBpedia IRIs of the relations in a short text. EARL (Banerjee et al., [n.d.]) is another tool that proposes a traveling salesman algorithm-based approach for joint entity and relation linking over DBpedia. To the best of our knowledge, EARL and Falcon are the only available tools that provide both entity and relation linking. Entity linking over Wikidata is a relatively new domain. Cetoli et al. (Cetoli et al., 2019) propose a neural network-based approach for linking entities to Wikidata. The authors also align an existing Wikipedia corpus-based dataset to Wikidata. However, this work only targets entity disambiguation and assumes that the entities are already recognized in the sentences. Arjun (Mulang et al., 2020) is the latest work for Wikidata entity linking. It uses an attention-based neural network for linking Wikidata entity labels. OpenTapioca (Delpeuch, 2019) is another attempt that performs end-to-end entity linking over Wikidata; it is the closest to our work even though OpenTapioca does not provide Wikidata Ids of relations in a sentence. OpenTapioca is also available as an API and is utilized as our baseline. S-Mart (Yang and Chang, 2015) is a tree-based structured learning framework based on multiple additive regression trees for linking entities in a tweet. The model was later adapted for linking entities in the questions. VCG (Sorokin and Gurevych, 2018) is another attempt which is a unifying network that models contexts of variable granularity to extract features for an end to end entity linking. However, Falcon 2.0 is the first tool for joint entity and relation linking over Wikidata. ## 3\. Falcon 2.0\- A Resource In this section, we describe Falcon 2.0 in detail. First the architecture of Falcon 2.0 is depicted. Next, we discuss the BK used to match the surface forms in the text to the resource in a specific KG. In the paper’s scope, we define ”short text” as grammatically correct questions (up to 15 words). ### 3.1. Architecture Figure 1. The Falcon 2.0 Architecture. The boxes highlighted in Grey are reused from Falcon (Sakor et al., 2019). Grey boxes contain a linguistic pipeline for recognizing and linking entity and relation surface forms. The boxes in White are our addition to the Falcon pipeline to build a resource for the Wikidata entity and relation linking. The white boxes constitute what we refer to as BK specific to Wikidata. The text search engine contains the alignment of Wikidata entity/relation labels along with the entity and relation aliases. It is used for generating potential candidates for entity and relation linking. RDF triple store is a local copy of Wikidata triples containing all entities and predicates. The Falcon 2.0 architecture is depicted in Figure 1. Falcon 2.0 receives short input texts and outputs a set of entities and relations extracted from the text; each entity and relation in the output is associated with a unique Internationalized Resource Identifier (IRI) in Wikidata. Falcon 2.0 resorts to BK and a catalog of rules for performing entity and relation linking. The BK combines Wikidata labels and their corresponding aliases. Additionally, it comprises alignments between nouns and entities in Wikidata. Alignments are stored in a text search engine, while the knowledge source is maintained in an RDF triple store accessible via a SPARQL endpoint. The rules that represent the English morphology are in a catalog; a forward chaining inference process is performed on top of the catalog during the extraction and linking tasks. Falcon 2.0 also comprises several modules that identify and link entities and relations to the Wikidata. These modules implement POS Tagging, Tokenization & Compounding, N-Gram Tiling, Candidate List Generation, Matching & Ranking, Query Classifier, and N-Gram Splitting and are reused from the implementation of Falcon. ### 3.2. Background Knowledge Figure 2. Falcon 2.0 Background Knowledge is built by converting labels of entities and relations in Wikidata into pairs of alignments. It is a part of search engine (cf. Figure 1). Wikidata contains over 52 million entities and 3.9 billion facts (in the form of subject-predicate-object triples). Since Falcon 2.0 background knowledge only depends on labels, a significant portion of this extensive information is not useful for our approach. Hence, we only extract all the entity and relation labels to create a local background KG, A.K.A ”alias background knowledge base.”. For example, the entity United States of America999https://www.wikidata.org/wiki/Q30 in Wikidata has the natural language label ‘United States of America’ and several other aliases (or known_as labels) of United States of America such as ”the United States of America, America, U.S.A., the U.S., United States, etc.”. We extended our background KG with this information from Wikidata. Similarly, for relation’s labels, the background KG is enriched with known_as labels to provide synonyms and derived word forms. For example, the relation spouse 101010https://www.wikidata.org/wiki/Property:P26 in Wikidata has the label spouse and the other known as labels are husband, wife, married to, wedded to, partner, etc. This variety of synonyms for each relation empowers Falcon 2.0 to match the surface form in the text to a relation in Wikidata. Figure 2 illustrates the process of building background knowledge. ### 3.3. Catalog of Rules Falcon 2.0 is a rule-based approach. A catalog of rules is predefined to extract entities and relations from the text. The rules are based on the English morphological principles and borrowed from Sakor et al. (Sakor et al., 2019). For example, Falcon 2.0 excludes all verbs from the entities candidates list based on the rule verbs are not entities. For example, the N-Gram tiling module in the Falcon 2.0 architecture resorts to the rule: entities with only stopwords between them are one entity. Another example of such rule When -> date, Where -> place solves the ambiguity of matching the correct relation in case the short text is a question by looking at the questions headword. For example, give the two questions When did Princess Diana die? and Where did Princess Diana die?, the relation died can be the death place or the death year. The question headword (When/Where) is the only insight to solve the ambiguity here. When the question word is where, Falcon 2.0 matches only relations that have a place as a range of the relation. ### 3.4. Recognition Extraction phase in Falcon 2.0 consists of three modules. POS tagging, tokenization & compounding, and N-Gram tiling. The input of this phase is a natural language text. The output of the phase is the list of surface forms related to entities or relations. Part-of-speech (POS) Tagging receives a natural language text as an input. It tags each word in the text with its related tag, e.g., noun, verb, and adverb. This module differentiates between nouns and verbs to enable the application of the morphological rules from the catalog. The output of the module is a list of the pairs of (word, tag). Tokenization & Compounding builds the tokens list by removing the stopwords from the input and splitting verbs from nouns. For example, if the input is What is the operating income for Qantas, the output of this module is a list of three tokens [operating, income, Qantas]. N-Gram Tilling module combines tokens with only stopwords between them relying on one of the rules from a catalog of rules. For example, if we consider the previous module’s output as an input for the n-gram tilling module, operating and income tokens will be combined in one token. The output of the module is a list of two tokens [operating income, Qantas]. ### 3.5. Linking This phase consists of four modules: candidate list generation, matching & ranking, relevant rule selection, and n-gram splitting. Candidate List Generation receives the output of the recognition phase. The module queries the text search engine for each token. Then, tokens will have an associated candidate list of resources. For example, the retrieved candidate list of the token operating income is [(P3362, operating income), (P2139, income), (P3362, operating profit)]; where the first element is the Wikidata predicate identifier and the second is the list of labels associated with the predicates which match the query ”operating income.” Matching & Ranking ranks the candidate list received from the candidate list generation module and matches candidates’ entities and relations. Since, in any KG, the facts are represented as triples, the matching and ranking module creates triples consisting of the entities and relationships from the candidates’ list. Then, for each pair of entity and relation, the module checks if the triple exists in the RDF triple store (Wikidata). The checking is done by executing a simple ASK query over the RDF triple store. For each triple, the module increases the rank of the involved relations and entities. The output of the module is the ranked list of the candidates. Relevant Rule Selection interacts with the matching & ranking module by suggesting increasing the ranks of some candidates relying on the catalog of rules. One of the suggestions is considering the question headword to clear the ambiguity between two relations based on the range of relationships in the KG. N-Gram Splitting module is used if none of the triples tested in the matching & ranking modules exists in the triple store, i.e., the compounding the approach did in the tokenization & compounding module led to combining two separated entities. The module splits the tokens from the right side and passes the tokens again to the candidate list generation module. Splitting the tokens from the right side resorts to one of the fundamentals of the English morphology; the compound words in English have their headword always towards the right side (Williams, 1981). Text Search Engine stores all the alignments of the labels. A simple querying technique (Gormley and Tong, 2015) is used as the text search engine over background knowledge. It receives a token as an input and then returns all the related resources with labels similar to the received token. RDF Triple store is a local copy of the Wikidata endpoint. It consists of all the RDF triples of Wikidata labeled with the English language. An RDF triple store is used to check the existence of the triples passed from the Matching & Ranking module. The RDF triple store keeps around 3.9 billion triples. ## 4\. Experimental Study We study three research questions: RQ1) What is the performance of Falcon 2.0 for entity linking over Wikidata? RQ2) What is the impact of Wikidata’s specific background knowledge on the performance of a linguistic approach? RQ3) What is the performance of Falcon 2.0 for relation linking over Wikidata? #### Metrics We report the performance using the standard metrics of Precision, Recall, and F-measure. Precision is the fraction of relevant resources among the retrieved resources. Recall is the fraction of relevant resources that have been retrieved over the total amount of relevant resources. F-Measure or F-Score is the harmonic mean of precision and recall. #### Datasets We rely on three different question answering datasets namely SimpleQuestion dataset for Wikidata (Diefenbach et al., 2017), WebQSP-WD (Sorokin and Gurevych, 2018) and LC-QuAD 2.0 (Dubey et al., 2019). The SimpleQuestion dataset contains 5,622 test questions which are answerable using Wikidata as underlying KG. WebQSP-WD contains 1639 test questions, and LC-QUAD 2.0 contains 6046 test questions. SimpleQuestion and LC-QuaD 2.0 provide the annotated gold standard for entity and relations, whereas WebQSP-WD only provides annotated gold standard for entities. Hence, we evaluated entity linking performance on three datasets and relation linking performance on two datasets. Also, SimpleQuestion and WebQSP-WD contain questions with a single entity and relation, whereas LC-QuAD 2.0 contains mostly complex questions (i.e., more than one entity and relation). #### Baselines OpenTapioca (Delpeuch, 2019): is available as a web API; it provides Wikidata URIs for entities. We run OpenTapioca API on all the three datasets. Variable Context Granularity model (VCG) (Sorokin and Gurevych, 2018): is a unifying network that models contexts of variable granularity to extract features for mention detection and entity disambiguation. We were unable to reproduce VCG using the publicly available source code. Hence, we only report its performance on WebQSP-WD from the original paper (Sorokin and Gurevych, 2018) as we are unable to run the model on the other two datasets for entity linking. For the completion of the approach, we also report the other two baselines provided by the authors, namely Heuristic Baseline and Simplified VCG. S-Mart (Yang and Chang, 2015): was initially proposed to link entities in the tweets and later adapted for question answering. The system is not open source, and we adapt its result from (Sorokin and Gurevych, 2018) for WebQSP- WD dataset. No Baseline for Relation Linking: To the best of our knowledge, there is no baseline for relation linking on Wikidata. One argument could be to run the existing DBpedia based relation linking tool on Wikidata and compare it with our performance. We contest this solely because Wikidata is extremely noisy. For example, in ”What is the longest National Highway in the world?” the entity surface form ”National Highway” matches four(4) different entities in Wikidata that share the same entity label (i.e., ”National Highway”). In comparison, 2,055 other entities contain the full mention in their labels for the surface form ”National Highway”. However, in DBpedia, there exists only one unique label for ”National Highway”. Hence, any entity linking tool or relation linking tool tailored for DBpedia will face issues on Wikidata (cf. table 3). Therefore, instead of reporting the bias and under-performance, we did not evaluate their performance for a fair comparison. Hence, we report Falcon 2.0 relation linking performance only to establish new baselines on two datasets: SimpleQuestion and LC-QuAD 2.0. Figure 3. Falcon 2.0 API Web Interface. #### Experimental Details Falcon 2.0 is extremely lightweight from an implementation point of view. A laptop machine, with eight cores and 16GB RAM running Ubuntu 18.04 is used for implementing and evaluating Falcon 2.0. We deployed its web API on a server with 723GB RAM, 96 cores (Intel(R) Xeon(R) Platinum 8160CPU with 2.10GHz) running Ubuntu 18.04. This publicly available API is used to calculate the standard evaluation metrics, namely Precision, Recall, and F-score. ### 4.1. Experimental Results Table 1. Entity linking evaluation results on LC-QuAD 2.0 & SimpleQuestion datasets. Best values are in bold. Approach | Dataset | P | R | F ---|---|---|---|--- OpenTapioca (Delpeuch, 2019) | LC-QuAD 2.0 | 0.29 | 0.42 | 0.35 Falcon 2.0 | LC-QuAD 2.0 | 0.50 | 0.56 | 0.53 OpenTapioca (Delpeuch, 2019) | SimpleQuestion | 0.01 | 0.02 | 0.01 Falcon 2.0 | SimpleQuestion | 0.56 | 0.64 | 0.60 OpenTapioca (Delpeuch, 2019) | SimpleQuestion Uppercase Entities | 0.16 | 0.28 | 0.20 Falcon 2.0 | SimpleQuestion Uppercase Entities | 0.66 | 0.75 | 0.70 Table 2. Entity linking evaluation results on the WEBQSP test dataset. Best values are in bold. Approach | P | R | F ---|---|---|--- S-MART (Yang and Chang, 2015) | 0.66 | 0.77 | 0.72 Heuristic baseline (Sorokin and Gurevych, 2018) | 0.30 | 0.61 | 0.40 Simplified VCG (Sorokin and Gurevych, 2018) | 0.84 | 0.62 | 0.71 VCG (Sorokin and Gurevych, 2018) | 0.83 | 0.65 | 0.73 OpenTapioca (Delpeuch, 2019) | 0.01 | 0.02 | 0.02 Falcon 2.0 | 0.80 | 0.84 | 0.82 #### Experimental Results 1 In the first experiment described in Table 1, we compare entity linking performance of Falcon 2.0 on SimpleQuestion and LC-QuAD 2.0 datasets. We first evaluate the performance on the SimpleQuestion dataset. Surprisingly, we observe that for the OpenTapioca baseline, the values are approximately 0.0 for Precision, Recall, and F-score. We analyzed the source of errors and found that out of 5,622 questions, only 246 have entity labels in uppercase letters. Opentapioca fails to recognize and link entity mentions written in lowercase letters. Case sensitivity is a common issue for entity linking tools over short text, as reported by Singh et al. (Singh et al., 2018a; Singh et al., 2018b) in a detailed analysis. From the remaining 246 questions, only 70 are answered correctly by OpenTapioca. Given that OpenTapioca finds limitation in linking lowercase entity surface forms. We evaluated Falcon 2.0 and OpenTapioca on the 246 questions of SimpleQuestion to provide a fair evaluation for the baseline (reported as SimpleQuestion uppercase entities in table 1). OpenTapioca reports F-score 0.20 on this subset of SimpleQuestion. On the other hand, Falcon 2.0 reports F-score 0.70 on the same dataset (cf. Table 1). For LC-QuAD 2.0, OpenTapioca reports F-score 0.35 against Falcon 2.0 with F-score 0.53 reported in Table 1. #### Experimental Results 2 We report performance of Falcon 2.0 on WebQSP-WD dataset in Table 2. Falcon 2.0 clearly outperforms all other baselines with highest F-score value 0.82. OpenTapioca demonstrates a low performance on this dataset as well. Experiment results 1 & 2 answer our first research question (RQ1). #### Ablation Study for Entity Linking and Recommendations For the second research question (RQ2), we evaluate the impact of Wikidata’s specific background knowledge on the entity linking performance. We evaluated Falcon on the WebQSP-WD dataset against Falcon 2.0. We linked Falcon predicted DBpedia IRIs with corresponding Wikidata IDs using owl:sameAs. We can see in the Table 3 that Falcon 2.0 significantly outperforms Falcon despite using the same linguistic driven approach. The jump in Falcon 2.0 performance comes from Wikidata’s specific local background knowledge, which we created by expanding Wikidata entities and relations with associated aliases. It also validates the novelty of Falcon 2.0 when compared to Falcon for the Wikidata entity linking. We observe an indifferent phenomenon in our performance for three datasets, and the performance for Falcon 2.0 differs a lot per dataset. For instance, on WebQSP-WD, our F-score is 0.82, whereas, on LC-QuAD 2.0, the F-Score drops to 0.57. The first source of error is the dataset(s) itself. In both the datasets (SimpleQuestion and LC-QuAD 2.0), many questions are grammatically incorrect. To validate our claim more robustly, we asked two native English speakers to check the grammar of 200 random questions on LC-QuAD 2.0. Annotators reported that 42 out of 200 questions are grammatically incorrect. Many questions have erroneous spellings of the entity names. For example, ”Who is the country for head of state of Mahmoud Abbas?” and ”Tell me about position held of Malcolm Fraser and elected in?” are two grammatically incorrect questions in LC-QuAD 2.0. Similarly, many questions in the SimpleQuestion dataset are also grammatically incorrect. ”where was hank cochran birthed” is one such example in the SimpleQuestion dataset. Falcon 2.0 resorts to fundamental principles of the English morphology and finds limitation in recognizing entities in many grammatically incorrect questions. We also recognize that the performance of Falcon 2.0 on sentences with minimal context is limited. For example, in the question ”when did annie open?” from the WebQSP-WD dataset, the sentential context is shallow. Also, more than one instance of ”Annie” exists in Wikidata, such as Wiki:Q566892 (correct one) and Wiki:Q181734. Falcon 2.0 wrongly predicts the entity in this case. In another example, ”which country is lamb from?”, the correct entity is Wiki:Q6481017 with label ”lamb” in Wikidata. However, Falcon 2.0 returns Wiki:13553878, which also has a label ”lamb”. In such cases, additional knowledge graph context shall prove to be useful. Approaches such as (Yang et al., 2019) introduced a concept of feeding ”entity descriptions” as an additional context in an entity linking model over Wikipedia. Suppose the extra context in the form of entity description (1985 English drama film directed by Colin Gregg) for the entity Wiki:13553878 is provided. In that case, a model may correctly predict the correct entity ”lamb.” Based on our observations, we propose the following recommendations for the community to improve the entity linking task over Wikidata: * • Wikidata has inherited challenges of vandalism and noisy entities due to crowd-authored entities (Heindorf et al., 2016). We expect the research community to come up with more robust short text datasets for the Wikidata entity linking without spelling and grammatical errors. * • Rule-based approaches come with its limitations when the sentential context is minimal. However, such methods are beneficial for the nonavailability of training data. We recommend a two-step process to target questions with minimal sentential context: 1) work towards a clean and large Wikidata dataset for entity linking of short text. This will allow more robust machine learning approaches to evolve 2) use of entity descriptions from knowledge graphs to improve the linking process (same as (Yang et al., 2019)). Table 3. Entity Linking Performance of Falcon vs Falcon 2.0 on WEBQSP-WD. Best values are in bold. Approach | P | R | F ---|---|---|--- Falcon (Sakor et al., 2019) | 0.47 | 0.45 | 0.46 Falcon 2.0 | 0.80 | 0.84 | 0.82 Table 4. Relation linking evaluation results on LC-QuAD 2.0 & SimpleQuestion datasets. Approach | Dataset | P | R | F ---|---|---|---|--- Falcon 2.0 | LC-QuAD 2.0 | 0.44 | 0.37 | 0.40 Falcon 2.0 | SimpleQuestion | 0.35 | 0.44 | 0.39 Table 5. Sample Questions from LC-QuAD 2.0 datset. The table shows five sample questions and associated gold standard relations. These sentences do not include standard sentential relations in the English language. Considering Wikidata is largely authored by the crowd, the crowd often creates such uncommon relations. Falcon 2.0 finds limitation in linking such relations, and most results are empty. Question | Gold Standard IDs | Gold Standard Labels | Predicted IDs | Predicted Labels ---|---|---|---|--- Which is the global-warming potential of dichlorodifluoromethane? | P2565 | global warming potential | [] | _ What is the AMCA Radiocommunications Licence ID for Qantas? | P2472 | ACMA Radiocommunications Client Number | P275 | copyright license What is ITIS TSN for Sphyraena? | P815 | ITIS TSN | [] | _ What is the ARICNS for Fomalhaut? | P999 | ARICNS | [] | _ Which is CIQUAL 2017 ID for cheddar? | P4696 | CIQUAL2017 ID | [] | _ #### Experimental Results 3: In the third experiment (for RQ3), we evaluate the relation linking performance of Falcon 2.0. We are not aware of any other model for relation linking over Wikidata. Table 4 summarizes relation linking performance. With this, we established new baselines over two datasets for relation linking on Wikidata. #### Ablation Study for Relation Linking and Recommendations Falcon reported an F-score of 0.43 on LC-QuAD over DBpedia in (Sakor et al., 2019) whereas Falcon 2.0 reports a comparable relation linking F-score 0.40 on LC-QuAD 2.0 for Wikidata (cf. Table 4). The wrong identification of the entities does affect the relation linking performance, and it is the major source of error in our case for relation linking. Table 5 summarizes a sample case study for relation linking on five LC-QuAD 2.0 questions. We observe that the relations present in the questions are highly uncommon and nonstandard, and it is a peculiar property of Wikidata. Falcon 2.0 finds limitations in linking such relations. We recommend the following: * • Wikidata challenges relation linking approaches by posing a new challenge: user-created nonstandard relations such as in Table 5. A rule-based approach like ours faces a clear limitation in linking such relations. Linking user- created relations in crowd-authored Wikidata is an open question for the research community. ## 5\. Impact In August 2019, Wikidata became the first Wikimedia project that crossed 1 billion edits, and over 20,000 active Wikidata editors111111https://www.wikidata.org/wiki/Wikidata:Statistics. A large subset of the information extraction community has extensively relied on its research around DBpedia and Wikidata targeting different research problems such as KG completion, question answering, entity linking, and data quality assessments (Moon et al., 2017; Reinanda et al., 2016; Yang et al., 2013). Furthermore, entity and relation linking tasks have been studied well beyond information extraction research, especially NLP and Semantic Web. Despite Wikidata being hugely popular, there are limited resources for reusing and aligning unstructured text to Wikidata mentions. However, when it comes to a short text, the performance of existing baselines are limited. We believe the availability of Falcon 2.0 as a web API along with open source access to its code will provide researchers an easy and reusable way to annotate unstructured text against Wikidata. We also believe that a rule-based approach, such as ours that does not require any training data, is beneficial for low resource languages (considering Wikidata is multilingual 121212https://www.wikidata.org/wiki/Help:Wikimedia_language_codes/lists/all). ## 6\. Adoption and Reusability Falcon 2.0 is open source. The source code is available in our public GitHub: https://github.com/SDM-TIB/Falcon2.0 for reusability and reproducibility. Falcon 2.0 is easily accessible via a simple CURL request or using our web interface. Detailed instructions are provided on our GitHub. It is currently available for the English language. However, there is no assumption in the approach or while building the background knowledge base that restricts its adaptation or extensibility to other languages. The background knowledge of Falcon 2.0 is available for the community and can be easily reused to generate candidates for entity linking (Yamada et al., 2016b) or in question answering approaches such as (Zhang and Zou, 2018). The background knowledge consists of 48,042,867 alignments of Wikidata entities and 15,645 alignments for Wikidata predicates. MIT License allows for the free distribution and re-usage of Falcon 2.0. We hope the research community and industry practitioners will use Falcon 2.0 resources for various usages such as linking entities and relations to Wikidata, annotating an unstructured text, developing new low language resources, and others. ## 7\. Maintenance and Sustainability Falcon 2.0 is a publicly available resource offering of the Scientific Data Management(SDM) group at TIB, Hannover131313https://www.tib.eu/en/research- development/scientific-data-management/. TIB is one of the largest libraries for Science and Technology in the world 141414https://www.tib.eu/en/tib/profile/. It actively promotes open access to scientific artifacts, e.g., research data, scientific literature, non-textual material, and software. Similar to other publicly maintained repositories of SDM, Falcon 2.0 will be preserved and regularly updated to fix bugs and include new features151515https://github.com/SDM-TIB. The Falcon 2.0 API will be sustained on the TIB servers to allow for unrestricted free access. ## 8\. Conclusion and Future Work We presented the resource Falcon 2.0, a rule-based entity and relation linking tool able to recognize entities & relations in a short text and link them to the existing knowledge graph, e.g., DBpedia and Wikidata. Although there are various approaches for entity & relation linking to DBpedia, Falcon 2.0 is one of the few tools targeting Wikidata. Thus, given the number of generic and domain-specific facts that compose Wikidata, Falcon 2.0 has the potential to impact researchers and practitioners that resort to NLP tools for transforming semi-structured data into structured facts. Falcon 2.0 is open source. The API is publicly accessible and maintained in the servers of the TIB labs. Falcon 2.0 has been empirically evaluated on three benchmarks, and the outcomes suggest that it is able to overcome the state of the arts. Albeit promising, the experimental results can be improved. In the future, we plan to continue researching novel techniques that enable adjusting the catalog of rules and alignments to the changes in Wikidata. We further plan to mitigate errors caused by a rule-based approach using machine learning approaches to aim towards a hybrid approach. ## 9\. Acknowledgments This work has received funding from the EU H2020 Project No. 727658 (IASIS). ## References * (1) * Auer et al. (2007) Sören Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary G. Ives. 2007. DBpedia: A Nucleus for a Web of Open Data. In _ISWC_. 722–735. * Balog (2018) Krisztian Balog. 2018\. _Entity-oriented search_. Springer Open. * Banerjee et al. ([n.d.]) Debayan Banerjee, Mohnish Dubey, Debanjan Chaudhuri, and Jens Lehmann. [n.d.]. Joint Entity and Relation Linking using EARL. ([n. d.]). * Bollacker et al. (2008) Kurt D. Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In _ACM SIGMOD_. 1247–1250. * Cao et al. (2018) Yixin Cao, Lei Hou, Juanzi Li, and Zhiyuan Liu. 2018\. Neural Collective Entity Linking. arXiv:1811.08603 http://arxiv.org/abs/1811.08603 * Cetoli et al. (2019) Alberto Cetoli, Stefano Bragaglia, Andrew D O’Harney, Marc Sloan, and Mohammad Akbari. 2019\. A Neural Approach to Entity Linking on Wikidata. In _European Conference on Information Retrieval_. Springer, 78–86. * Delpeuch (2019) Antonin Delpeuch. 2019\. OpenTapioca: Lightweight Entity Linking for Wikidata. _arXiv preprint arXiv:1904.09131_ (2019). * Diefenbach et al. (2017) Dennis Diefenbach, Thomas Tanon, Kamal Singh, and Pierre Maret. 2017\. Question answering benchmarks for wikidata. * Dubey et al. (2019) Mohnish Dubey, Debayan Banerjee, Abdelrahman Abdelkawi, and Jens Lehmann. 2019. Lc-quad 2.0: A large dataset for complex question answering over wikidata and dbpedia. In _International Semantic Web Conference_. Springer, 69–78. * Ferragina and Scaiella (2010) Paolo Ferragina and Ugo Scaiella. 2010. TAGME: on-the-fly annotation of short text fragments (by wikipedia entities). In _Proceedings of the 19th ACM Conference on Information and Knowledge Management, CIKM 2010, Toronto, Ontario, Canada, October 26-30, 2010_. 1625–1628. * Ganea and Hofmann (2017) Octavian-Eugen Ganea and Thomas Hofmann. 2017. Deep Joint Entity Disambiguation with Local Neural Attention. In _Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP 2017, Copenhagen, Denmark, September 9-11, 2017_. 2619–2629. * Gormley and Tong (2015) Clinton Gormley and Zachary Tong. 2015. _Elasticsearch: The Definitive Guide: A Distributed Real-Time Search and Analytics Engine_. ” O’Reilly Media, Inc.”. * Heindorf et al. (2016) Stefan Heindorf, Martin Potthast, Benno Stein, and Gregor Engels. 2016. Vandalism detection in wikidata. In _Proceedings of the 25th ACM International on Conference on Information and Knowledge Management_. 327–336. * Hoffart et al. (2011) Johannes Hoffart, Mohamed Amir Yosef, Ilaria Bordino, Hagen Fürstenau, Manfred Pinkal, Marc Spaniol, Bilyana Taneva, Stefan Thater, and Gerhard Weikum. 2011\. Robust Disambiguation of Named Entities in Text. In _EMNLP 2011_. 782–792. * Inan and Dikenelli (2018) Emrah Inan and Oguz Dikenelli. 2018. A Sequence Learning Method for Domain-Specific Entity Linking. In _Proceedings of the Seventh Named Entities Workshop_ (Melbourne, Australia). Association for Computational Linguistics, 14–21. http://aclweb.org/anthology/W18-2403 * Ji (2019) Heng Ji. 2019. _Entity Discovery and Linking and Wikification Reading List_. http://nlp.cs.rpi.edu/kbp/2014/elreading.html * Kolitsas et al. (2018) Nikolaos Kolitsas, Octavian-Eugen Ganea, and Thomas Hofmann. 2018. End-to-End Neural Entity Linking. In _Proceedings of the 22nd Conference on Computational Natural Language Learning_. 519–529. * Moon et al. (2017) Changsung Moon, Paul Jones, and Nagiza F Samatova. 2017\. Learning entity type embeddings for knowledge graph completion. In _Proceedings of the 2017 ACM on conference on information and knowledge management_. 2215–2218. * Mulang et al. (2020) Isaiah Onando Mulang, Kuldeep Singh, Akhilesh Vyas, Saeedeh Shekarpour, Ahmad Sakor, Maria Esther Vidal, Soren Auer, and Jens Lehmann. 2020. Encoding Knowledge Graph Entity Aliases in an Attentive Neural Networks for Wikidata Entity Linking. _In WISE (to appear)_ (2020). * Raiman and Raiman (2018) Jonathan Raphael Raiman and Olivier Michel Raiman. 2018. DeepType: multilingual entity linking by neural type system evolution. In _Thirty-Second AAAI Conference on Artificial Intelligence_. * Reinanda et al. (2016) Ridho Reinanda, Edgar Meij, and Maarten de Rijke. 2016\. Document Filtering for Long-tail Entities. In _Proceedings of the 25th ACM International Conference on Information and Knowledge Management, CIKM 2016, Indianapolis, IN, USA, October 24-28, 2016_. ACM, 771–780. https://doi.org/10.1145/2983323.2983728 * Röder et al. (2018) Michael Röder, Ricardo Usbeck, and Axel-Cyrille Ngonga Ngomo. 2018\. Gerbil–benchmarking named entity recognition and linking consistently. _Semantic Web_ 9, 5 (2018), 605–625. * Sakor et al. (2019) Ahmad Sakor, Isaiah Onando Mulang, Kuldeep Singh, Saeedeh Shekarpour, Maria Esther Vidal, Jens Lehmann, and Sören Auer. 2019\. Old is gold: linguistic driven approach for entity and relation linking of short text. In _Proceedings of the 2019 NAACL HLT (Long Papers)_. 2336–2346. * Shen et al. (2015) W. Shen, J. Wang, and J. Han. 2015. Entity Linking with a Knowledge Base: Issues, Techniques, and Solutions. _IEEE Transactions on Knowledge and Data Engineering_ 27, 2 (2015), 443–460. * Singh et al. (2018a) Kuldeep Singh, Ioanna Lytra, Arun Sethupat Radhakrishna, Saeedeh Shekarpour, Maria-Esther Vidal, and Jens Lehmann. 2018a. No One is Perfect: Analysing the Performance of Question Answering Components over the DBpedia Knowledge Graph. _arXiv:1809.10044_ (2018). * Singh et al. (2018b) Kuldeep Singh, Arun Sethupat Radhakrishna, Andreas Both, Saeedeh Shekarpour, Ioanna Lytra, Ricardo Usbeck, Akhilesh Vyas, Akmal Khikmatullaev, Dharmen Punjani, Christoph Lange, Maria-Esther Vidal, Jens Lehmann, and Sören Auer. 2018b. Why Reinvent the Wheel: Let’s Build Question Answering Systems Together. In _Web Conference_. 1247–1256. * Sorokin and Gurevych (2018) Daniil Sorokin and Iryna Gurevych. 2018. Mixing Context Granularities for Improved Entity Linking on Question Answering Data across Entity Categories. In _Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics_. 65–75. * Vrandecic (2012) Denny Vrandecic. 2012\. Wikidata: a new platform for collaborative data collection. In _Proceedings of the 21st World Wide Web Conference, WWW 2012, Lyon, France, April 16-20, 2012 (Companion Volume)_. ACM, 1063–1064. https://doi.org/10.1145/2187980.2188242 * Williams (1981) Edwin Williams. 1981\. On the notions” Lexically related” and” Head of a word”. _Linguistic inquiry_ 12, 2 (1981), 245–274. * Yamada et al. (2016a) Ikuya Yamada, Hiroyuki Shindo, Hideaki Takeda, and Yoshiyasu Takefuji. 2016a. Joint Learning of the Embedding of Words and Entities for Named Entity Disambiguation. In _CoNLL 2016_ , Yoav Goldberg and Stefan Riezler (Eds.). ACL, 250–259. * Yamada et al. (2016b) Ikuya Yamada, Hiroyuki Shindo, Hideaki Takeda, and Yoshiyasu Takefuji. 2016b. Joint Learning of the Embedding of Words and Entities for Named Entity Disambiguation. _CoRR_ abs/1601.01343 (2016). * Yang et al. (2019) Xiyuan Yang, Xiaotao Gu, Sheng Lin, Siliang Tang, Yueting Zhuang, Fei Wu, Zhigang Chen, Guoping Hu, and Xiang Ren. 2019. Learning Dynamic Context Augmentation for Global Entity Linking. In _EMNLP-IJCNLP 2019_ , Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan (Eds.). 271–281. * Yang and Chang (2015) Yi Yang and Ming-Wei Chang. 2015. S-MART: Novel Tree-based Structured Learning Algorithms Applied to Tweet Entity Linking. In _ACL- IJCNLP (Volume 1: Long Papers)_. 504–513. * Yang et al. (2013) Zi Yang, Elmer Garduño, Yan Fang, Avner Maiberg, Collin McCormack, and Eric Nyberg. 2013\. Building optimal information systems automatically: configuration space exploration for biomedical information systems. In _22nd ACM CIKM’13, San Francisco, USA_. ACM, 1421–1430. * Zhang and Zou (2018) Xinbo Zhang and Lei Zou. 2018. IMPROVE-QA: An Interactive Mechanism for RDF Question/Answering Systems. In _Proceedings of the 2018 International Conference on Management of Data, SIGMOD Conference 2018, Houston, TX, USA, June 10-15, 2018_. 1753–1756. https://doi.org/10.1145/3183713.3193555
# Signatures of fragmentation for periodically driven fermions Somsubhra Ghosh1, Indranil Paul2, and K. Sengupta1 1School of Physical Sciences, Indian Association for the Cultivation of Science, Kolkata 700032, India. 2Université Paris Cité, CNRS, Laboratoire Matériaux et Phénomènes Quantiques, 75205 Paris, France. ###### Abstract We study the possible signatures of prethermal strong Hilbert space fragmentation (HSF) for one-dimensional (1D) fermions subjected to a periodic drive. We extend the results of Phys. Rev. Lett. 130, 120401 (2023) to show the possibility of such fragmentation for a large class of experimentally relevant drive protocols. Moreover, we demonstrate the persistence of HSF when the fermion chain is taken away from half-filling. Both these analysis indicate the robustness of the fragmentation phenomenon reported earlier. We also provide an alternate derivation of the Floquet Hamiltonian of the driven chain which yields insight into the generic nested commutator structure of its higher order terms. Finally, we study the density-density out-of-time- correlators (OTOC) of the driven chain both away and near the special drive frequencies at which its first order Floquet Hamiltonian exhibits fragmentation. We show that these OTOCs, for a chain with open boundary condition, exhibit a distinct periodic unscrambling of information at special drive frequencies; such unscrambling can therefore serve as a marker of prethermal HSF. We provide an approximate analytic explanation of the role of HSF behind such periodic unscrambling and discuss experiments which can detect signatures of strong HSF in such driven chains. ## I Introduction Closed quantum systems driven out of equilibrium have become increasingly important subject of research in recent years [1, 2, 3]. One of the central questions in this field pertains to the long-time behavior of local correlation functions of these systems. In most case, the behavior of such correlation functions can be understood from the eigenstate thermalization hypothesis (ETH) [4, 5]. ETH predicts eventual thermalization, under unitary Hamiltonian dynamics, of a generic many-body quantum state which can initially be far from equilibrium; it is one of the central paradigms for understanding long-time behavior of a generic ergodic many-body system. ETH also holds, with minor modifications, for periodically driven systems, where the driven system is ultimately expected to heat up to infinite temperature [6]. ETH relies on the ergodicity of a generic quantum system and is known to fail when it is violated. Such ergodicity violation can occur in integrable models due to the presence of a large number of conserved quantities [1]. In addition, it fails in the presence of strong disorder leading to many-body localization and consequent violation of ergodicity [7, 8, 9]. A more subtle and weaker failure of ETH occurs due to emergent symmetry sectors in an otherwise generic quantum systems. The presence of such symmetries typically lead to tower of states which are protected from the other thermal states in its Hilbert space. Thus any quantum dynamics starting from an initial state which belongs to this sectors can not thermalize; such states are often called quantum scars [10, 11, 12, 13, 14]. The violation of ETH in this case is weak since it only happens if the initial state has large overlap with the states in the scar sector. The number of such states, for a one-dimensional (1D) system of length $L$, is typically ${\rm O}(L)$; they form a tiny fraction of the total number of states in the Hilbert space which is ${\rm O}(e^{L})$. Thus these systems display ETH violating dynamics for a small fraction of initial states. Another, recently found, violation of ETH occurs in non-integrable quantum systems due to the presence of kinetic constraints. The Hamiltonian of such quantum systems, expressed as a matrix in the computational basis, breaks down into an exponentially large number of dynamically disconnected fragments. The presence of such a large number of disconnected sectors is to be contrasted with those occurring from the presence of conserved global quantities; the latter only leads to an ${\rm O}(L)$ disconnected symmetry sectors. This phenomenon is known as strong Hilbert space fragmentation (HSF) [15, 16, 17, 18, 19]. Such a strong fragmentation naturally breaks ergodicity since any generic initial state, which belongs to a given fragment, can not, under the action of the system Hamiltonian, explore states in the Hilbert space that belong to other fragments. Most of the Hamiltonians studied in this context are 1D spin or fermionic models [15, 16, 17, 18]; however, more recently some higher-dimensional models have also been shown to exhibit strong HSF [19]. More recently, the generalization of strong HSF to periodically driven quantum system has been studied [20]. It has been shown that a periodically driven fermionic chain can show signatures of Hilbert space fragmentation at special drive frequencies over a large prethermal timescale; the extent of this prethermal timescale depends on the drive amplitude and can be quite large in the large drive amplitude regime. The signatures of such prethermal fragmentation can be found in entanglement entropy, autocorrelation function and equal-time correlators of such driven system; each of these quantities show departure from their counterparts in ergodic quantum systems [20] demonstrating a clear realization of prethermal strong HSF. In addition, such prethermal HSF in driven quantum systems can lead to interesting oscillatory dynamics of correlators for certain initial states which has no counterpart for HSF realized in an equilibrium setting [20]. In this work, we extend the results obtained in Ref. 20 in several ways. First, we show that such signatures of fragmentation can be obtained for a much wider class of drive protocols. This makes the prethermal fragmentation phenomenon much more relevant to standard experiments using ultracold atom platform which we discuss. Second, we provide a comprehensive analysis of the Floquet Hamiltonian. The analytical expression for the first order Floquet Hamiltonian, $H_{F}^{(1)}$, derived using Floquet perturbation theory (FPT), was presented in Ref. 20; here we provide an alternate derivation of the Floquet Hamiltonian up to second order in perturbation theory. This analysis provides insight into the commutator structure of the higher order terms that was not apparent from the previous derivation. It also provides an estimate of the frequency range over which the first order Floquet Hamiltonian provides a qualitatively accurate description of the dynamical evolution in the prethermal regime. Third, we show that the signature of fragmentation persists when the driven chain is taken away from half-filling. This shows the robustness of the prethermal fragmentation phenomenon and points out the possibility of its experimental realization for a wide range of fermion filling fraction. Finally, we study the density-density out-of-time correlator (OTOC) for the driven fermion chain. We show that the behavior of such OTOC is qualitatively different at special frequencies at which the system exhibits signatures of prethermal HSF. In particular, for a driven fermion chain with open boundary condition, we find, starting from a initial ${\mathbb{Z}}_{2}$ state, unscrambling of information manifested through periodic revival of the OTOC. We analyze this phenomenon in details, provide an analytic, albeit qualitative, understanding of its mechanism, and tie it to the fragmented structure of the first-order Floquet Hamiltonian, $H_{F}^{(1)}$, of the driven chain obtained using FPT. Our results thus demonstrate that OTOCs can server as markers for prethermal HSF in a driven system. The organization of the rest of this work is as follows. In Sec. II, we present a derivation of the Floquet Hamiltonian which brings out its nested commutator structure. Next, in Sec. III, we discuss the different classes of drive protocols which allows for signature of prethermal fragmentation and also derive the higher order Floquet Hamiltonian corresponding to them. This is followed by Sec. IV where we demonstrate signature of prethermal HSF away from half-filling. Next, in Sec. V, we discuss the behavior of OTOC in such driven system. Finally, we discuss our main results and conclude in Sec. VI. Some details of the calculation are presented in the Appendices. ## II Formalism In this section, we outline the derivation of the Floquet Hamiltonian of the driven fermion chain. Our derivation brings out the nested commutator structure of the Floquet Hamiltonian and also addresses a more general class of drive protocols for which the fermion chain exhibits prethermal HSF. ### II.1 Preliminary Consider a time dependent quantum mechanical system described by the Hamiltonian $\mathcal{H}(t)=\mathcal{H}_{0}(t)+\mathcal{H}_{1},$ (1) where all the time dependence is in the zeroth order term $\mathcal{H}_{0}$. The term $\mathcal{H}_{1}$, which in the following will be treated perturbatively, has no explicit time dependence. From Schrodinger equation $i\hbar\partial_{t}\psi(t)=[\mathcal{H}_{0}(t)+\mathcal{H}_{1}]\psi(t)$, and the definition of the time evolution operator $U(t,0)$ via $\psi(t)=U(t,0)\psi(0)$, we get $i\hbar\frac{\partial}{\partial t}U(t,0)=[\mathcal{H}_{0}(t)+\mathcal{H}_{1}]U(t,0).$ (2) The evaluation of $U(t,0)$ can be broken into two steps. To do so, we write [21, 22, 23] $U(t,0)=U_{0}(t,0)W(t,0),$ (3) where $U_{0}(t,0)$ is the exact time evolution operator in the absence of the $\mathcal{H}_{1}$ term. The first step, which is simple, is to evaluate $U_{0}(t,0)$ which is given by $U_{0}(t,0)=\exp[-\frac{i}{\hbar}\int_{0}^{t}d\tau\mathcal{H}_{0}(\tau)].$ (4) In the above the time ordering in front of the exponential can be omitted since the operator $\mathcal{H}_{0}(\tau)$ at different times commute. The second step, which is non-trivial, is to compute $W(t,0)$ that encodes the time evolution due to $\mathcal{H}_{1}$. This is performed perturbatively. Using $i\partial_{t}U_{0}(t,0)=\mathcal{H}_{0}U_{0}(t,0)$ and Eqs. (2) and (3), we get $i\hbar\frac{\partial}{\partial t}W(t,0)=\mathcal{H}_{p}(t)W(t,0),$ (5) where $\mathcal{H}_{p}(t)\equiv U_{0}(t,0)^{-1}\mathcal{H}_{1}U_{0}(t,0).$ (6) Using Eq. (5), the perturbative expansion is $\displaystyle W(t,0)$ $\displaystyle=1-\left(\frac{i}{\hbar}\right)\int_{0}^{t}d\tau\mathcal{H}_{p}(\tau)\,+$ $\displaystyle\left(-\frac{i}{\hbar}\right)^{2}\int_{0}^{t}d\tau_{1}\mathcal{H}_{p}(\tau_{1})\int_{0}^{\tau_{1}}d\tau_{2}\mathcal{H}_{p}(\tau_{2})\,+$ $\displaystyle\left(-\frac{i}{\hbar}\right)^{3}\\!\int_{0}^{t}\\!d\tau_{1}\mathcal{H}_{p}(\tau_{1})\int_{0}^{\tau_{1}}\\!d\tau_{2}\mathcal{H}_{p}(\tau_{2})\int_{0}^{\tau_{2}}\\!d\tau_{3}\mathcal{H}_{p}(\tau_{3})$ $\displaystyle+\cdots.$ (7) Note, in the above the operator $\mathcal{H}_{p}(\tau)$ at different times do not commute. The above formulation can also be viewed as a series expansion in a rotating frame for the following reason. Consider a time dependent unitary transformation $V(t)$ between a laboratory to a rotating reference frame with the initial condition $V(0)=1$. The wavefunction in the rotating frame is $\psi_{r}(t)=V^{\dagger}(t)\psi(t)$, and an operator in the same frame is $\mathcal{O}_{r}(t)=V^{\dagger}(t)\mathcal{O}V(t)$, where $\psi(t)$ and $\mathcal{O}$ are the wavefunction and the operator in the laboratory frame, respectively. Simultaneously, the Hamiltonian $\mathcal{H}(t)$ in the laboratory frame transforms to $\mathcal{H}_{r}(t)$ in the rotating frame. By demanding that $i\hbar\partial_{t}\psi_{r}(t)=\mathcal{H}_{r}(t)\psi_{r}(t)$ we get $\mathcal{H}_{r}(t)=V^{\dagger}(t)\mathcal{H}(t)V(t)-i\hbar V^{\dagger}(t)\dot{V}(t),$ (8) where $\dot{V}\equiv\partial_{t}V(t)$. Furthermore, if we define the time evolution operator $U_{r}(t_{1},t_{2})$ in the rotating frame by $i\hbar\partial_{t_{1}}U_{r}(t_{1},t_{2})=\mathcal{H}_{r}(t_{1})U_{r}(t_{1},t_{2})$, then it is related to that in the laboratory frame by $U(t_{1},t_{2})=V(t_{1})U_{r}(t_{1},t_{2})V^{\dagger}(t_{2}).$ (9) The connection between the two formulations is made if we choose the time dependent unitary transformation to be $V(t)=\exp[-\frac{i}{\hbar}\int_{0}^{t}d\tau\mathcal{H}_{0}(\tau)],$ (10) such that the $\mathcal{H}_{0}(t)$ term is “gauged out” in the rotating frame. In this case $\mathcal{H}_{r}(t)$ coincides with $\mathcal{H}_{p}(t)$ given by Eq. (6), $U_{r}(t,0)$ with $W(t,0)$, and Eqs. (3) and (9) become identical with $t_{2}=0$. However, note that the first formulation is more versatile in the sense that it can be still used when $\mathcal{H}_{1}$ is the zeroth order term and $\mathcal{H}_{0}$ is perturbative. In this case, we simply exchange $\mathcal{H}_{0}(t)\leftrightarrow\mathcal{H}_{1}$ in Eqs. (4) and (6). The resulting expansion will not match with that in the rotating frame. ### II.2 Floquet perturbation theory Until now the discussion has been general, and it applies to all time dependent problems. In the particular case of a Floquet system, where the time dependence is due a periodic external drive, we are interested in the stroboscopic time evolution operator $U(T,0)$, where $T$ is the period of the drive. The related Floquet Hamiltonian is defined by $\mathcal{H}_{F}\equiv\frac{i\hbar}{T}\log U(T,0)=\frac{i\hbar}{T}\log[U_{0}(T,0)W(T,0)].$ (11) We suppose that there is a small parameter that justifies the expansion $W(T,)=1+W_{1}(T)+W_{2}(T)+\cdots$, and correspondingly $\mathcal{H}_{F}=\mathcal{H}_{F}^{(0)}+\mathcal{H}_{F}^{(1)}+\mathcal{H}_{F}^{(2)}+\cdots$. Then, using Eq. (II.1) and after some algebra the first few terms in the expansion of the Floquet Hamiltonian are given by $\displaystyle\mathcal{H}_{F}^{(0)}$ $\displaystyle=\frac{i\hbar}{T}\log U_{0}(T,0),$ (12) $\displaystyle\mathcal{H}_{F}^{(1)}$ $\displaystyle=\frac{i\hbar}{T}W_{1}(T)=\frac{1}{T}\int_{0}^{T}d\tau\mathcal{H}_{p}(\tau),$ (13) $\displaystyle\mathcal{H}_{F}^{(2)}$ $\displaystyle=\frac{i\hbar}{T}\left[W_{2}(T)-\frac{1}{2}W_{1}(T)^{2}\right]$ $\displaystyle=\frac{-i}{2\hbar T}\int_{0}^{T}d\tau_{1}\int_{0}^{\tau_{1}}d\tau_{2}\left[\mathcal{H}_{p}(\tau_{1}),\mathcal{H}_{p}(\tau_{2})\right],$ (14) $\displaystyle\mathcal{H}_{F}^{(3)}$ $\displaystyle=\frac{i\hbar}{T}\left[W_{3}(T)-\frac{1}{2}\left(W_{1}(T)W_{2}(T)+W_{2}(T)W_{1}(T)\right)\right.$ $\displaystyle+\left.\frac{1}{3}W_{1}(T)^{3}\right]$ $\displaystyle=-\frac{1}{6\hbar^{2}T}\int_{0}^{T}d\tau_{1}\int_{0}^{\tau_{1}}d\tau_{2}\int_{0}^{\tau_{2}}d\tau_{3}\left\\{\left[\mathcal{H}_{p}(\tau_{1}),\right.\right.$ $\displaystyle\left.\left.\left[\mathcal{H}_{p}(\tau_{2}),\mathcal{H}_{p}(\tau_{3})\right]\right]+\left[\left[\mathcal{H}_{p}(\tau_{1}),\mathcal{H}_{p}(\tau_{2})\right],\mathcal{H}_{p}(\tau_{3})\right]\right\\}.$ (15) Eqs. 12-15 indicate the nested commutator structure of the higher-order terms of the Floquet Hamiltonian; we shall use them for explicit computation of $H_{F}$ in Sec. III. ## III Computation of the Floquet Hamiltonian In this section we first provide analytical results for higher order terms in the Floquet Hamiltonian for a cosine drive protocol in Sec. III.1. This is followed, in Sec. III.2, by a derivation and analysis of the first order Floquet Hamiltonian $H_{F}^{(1)}$ for a more general drive protocol. ### III.1 Cosine modulation of interaction Consider a driven system described by Eq. (1) where $\displaystyle\mathcal{H}_{0}(t)$ $\displaystyle=V_{1}\cos\omega_{D}t\sum_{i}\hat{n}_{i}\hat{n}_{i+1},$ (16) $\displaystyle\mathcal{H}_{1}$ $\displaystyle=\sum_{i}\left[-J(c^{\dagger}_{i}c_{i+1}+\rm{h.c.})+V_{0}\hat{n}_{i}\hat{n}_{i+1}+V_{2}\hat{n}_{i}\hat{n}_{i+2}\right],$ (17) with $V_{1}\gg(J,V_{0},V_{2})$. Thus, in the following we treat the $\mathcal{H}_{0}$ term exactly, and $\mathcal{H}_{1}$ perturbatively. Following Sec. II, we have, using Eq. (4), $U_{0}(t,0)=\exp[-i\lambda\hat{B}\sin\omega_{D}t],$ (18) where $\lambda\equiv V_{1}/(\hbar\omega_{D})$ is a dimensionless parameter and $\hat{B}\equiv\sum_{j}\hat{n}_{j}\hat{n}_{j+1}$. The next step is to compute $\mathcal{H}_{p}(t)$ using Eq. (6). As an intermediate step we find $\displaystyle\left[\hat{B},\mathcal{H}_{1}\right]$ $\displaystyle=-J\sum_{i}\hat{A}_{i}\left(c^{\dagger}_{i}c_{i+1}-c^{\dagger}_{i+1}c_{i}\right),$ (19) $\displaystyle\left[\hat{B},\left[\hat{B},\mathcal{H}_{1}\right]\right]$ $\displaystyle=-J\sum_{i}\hat{A}_{i}^{2}\left(c^{\dagger}_{i}c_{i+1}+c^{\dagger}_{i+1}c_{i}\right),$ (20) and so on, where $\hat{A}_{i}=\hat{n}_{i-1}-\hat{n}_{i+2}.$ (21) . Using these relations we obtain $\displaystyle\mathcal{H}_{p}(t)$ $\displaystyle=\exp[i\lambda\hat{B}\sin\omega_{D}t]\mathcal{H}_{1}\exp[-i\lambda\hat{B}\sin\omega_{D}t]$ $\displaystyle=\mathcal{H}_{1}+i\lambda\sin\omega_{D}t\left[\hat{B},\mathcal{H}_{1}\right]+\frac{1}{2!}(i\lambda\sin\omega_{D}t)^{2}$ $\displaystyle\times\left[\hat{B},\left[\hat{B},\mathcal{H}_{1}\right]\right]+\frac{1}{3!}(i\lambda\sin\omega_{D}t)^{3}\left[\hat{B},\left[\hat{B},\left[\hat{B},\mathcal{H}_{1}\right]\right]\right]$ $\displaystyle+\cdots$ $\displaystyle=\sum_{i}\left[-J\left(e^{i\lambda\hat{A}_{i}\sin\omega_{D}t}c^{\dagger}_{i}c_{i+1}+e^{-i\lambda\hat{A}_{i}\sin\omega_{D}t}c^{\dagger}_{i+1}c_{i}\right)\right.$ $\displaystyle\left.+V_{0}\hat{n}_{i}\hat{n}_{i+1}+V_{2}\hat{n}_{i}\hat{n}_{i+2}\right].$ (22) Using the explicit form of $\mathcal{H}_{p}(t)$ it is possible to compute order by order the Floquet Hamiltonian. The zeroth order Floquet Hamiltonian $\mathcal{H}_{F}^{(0)}$ vanishes because, from Eq. (18), we have $U_{0}(T,0)=1$. To compute the first order Floquet Hamiltonian we use the relation $\displaystyle I_{1}(\hat{A},\lambda)$ $\displaystyle\equiv\frac{1}{T}\int_{0}^{T}d\tau e^{i\lambda\hat{A}\sin(\omega_{D}\tau)}$ $\displaystyle=J_{0}(\lambda\hat{A})=(1-\hat{A}^{2})+\hat{A}^{2}J_{0}(\lambda),$ (23) where $J_{n}(x)$ is a Bessel function of the first kind with integer order. Using this relation and Eq. (13) we get $\displaystyle\mathcal{H}_{F}^{(1)}$ $\displaystyle=\sum_{i}\left[-JJ_{0}(\lambda\hat{A}_{i})\left(c^{\dagger}_{i}c_{i+1}+{\rm h.c.}\right)\right.$ $\displaystyle+\left.V_{0}\hat{n}_{i}\hat{n}_{i+1}+V_{2}\hat{n}_{i}\hat{n}_{i+2}\right].$ (24) If the drive frequency is tuned to $\omega_{m}$ such that $\lambda_{m}=V_{1}/(\hbar\omega_{m})$ coincides with the $m^{{\rm th}}$ zero of the Bessel function $J_{0}$, then the corresponding first order Floquet Hamiltonian is $\displaystyle\mathcal{H}_{F}^{(1)}(\lambda=\lambda_{m})$ $\displaystyle=\sum_{i}\left[-J(1-\hat{A}_{i}^{2})\left(c^{\dagger}_{i}c_{i+1}+{\rm h.c.}\right)\right.$ $\displaystyle+\left.V_{0}\hat{n}_{i}\hat{n}_{i+1}+V_{2}\hat{n}_{i}\hat{n}_{i+2}\right].$ (25) The above defines a model with constrained hopping, where only those hops are allowed which preserve the total number of nearest neighbors $\hat{N}_{D}\equiv\sum_{i}\hat{n}_{i}\hat{n}_{i+1}$. This model is known to show strong Hilbert space fragmentation [16]. The second order Floquet Hamiltonian can be broken into two parts $\mathcal{H}_{F}^{(2)}=\mathcal{H}_{F}^{(2a)}+\mathcal{H}_{F}^{(2b)}$, with $\mathcal{H}_{F}^{(2a)}=\frac{-i}{2\hbar T}\int_{0}^{T}d\tau_{1}\int_{0}^{\tau_{1}}d\tau_{2}\left[\tilde{\mathcal{H}}_{p}(\tau_{1})\,,\,\tilde{\mathcal{H}}_{p}(\tau_{2})\right],$ (26) and $\displaystyle\mathcal{H}_{F}^{(2b)}$ $\displaystyle=\frac{-i}{2\hbar T}\int_{0}^{T}d\tau_{1}\int_{0}^{\tau_{1}}d\tau_{2}\left\\{\left[\tilde{\mathcal{H}}_{p}(\tau_{1})\,,\,\hat{K}\right]\right.$ $\displaystyle+\left.\left[\hat{K}\,,\,\tilde{\mathcal{H}}_{p}(\tau_{2})\right]\right\\}.$ (27) In the above $\tilde{\mathcal{H}}_{p}(\tau)\equiv-J\sum_{i}\left(e^{i\lambda\hat{A}_{i}\sin\omega_{D}\tau}c^{\dagger}_{i}c_{i+1}+{\rm h.c.}\right),$ (28) and $\hat{K}=\sum_{i}\left(V_{0}\hat{n}_{i}\hat{n}_{i+1}+V_{2}\hat{n}_{i}\hat{n}_{i+2}\right).$ (29) The details of the evaluation of the two parts is given in the appendix. The final result is $\mathcal{H}_{F}^{(2)}=\frac{2J\mathcal{C}(\lambda)}{\hbar\omega_{D}}\left[\sum_{i}\hat{A}_{i}\left(c^{\dagger}_{i}c_{i+1}-{\rm h.c.}\right)\,,\,\mathcal{H}_{F}^{(1)}\right],$ (30) where $\mathcal{C}(\lambda)\equiv\sum_{k=0}^{\infty}\frac{J_{2k+1}(\lambda)}{2k+1}.$ This concludes our derivation of the Floquet Hamiltonian for the cosine protocol. We note that ${\mathcal{H}}_{F}^{(2)}$ does not respect the constrained hopping structure of ${\mathcal{H}}_{F}^{(1)}$ and therefore destroys HSF in the driven model beyond a prethermal timescale; below this timescale ${\mathcal{H}}_{F}^{(1)}$ dominates the dynamics leading to prethermal relaization of HSF. ### III.2 An experimentally relevant drive protocol A possible realization of a standard fermion chain where coherent quantum dynamics can be studied involves ultracold atom platforms [24, 25]. In such realizations, both the hopping amplitude and the nearest-neighbor interaction between the fermions depend on the strength of the external lasers; therefore it is difficult to dynamically alter one keeping the other fixed. Therefore an experimental realization of strong HSF would require a protocol which allows for simultaneous variation of both the hopping and the interaction strength. To take such simultaneous variations into account, we now consider a fermionic chain with the Hamiltonian $\displaystyle H$ $\displaystyle=$ $\displaystyle-J(t)\sum_{j}\left(c_{j}^{\dagger}c_{j+1}+{\rm h.c.}\right)+(V_{0}+V(t))\sum_{j}\hat{n}_{j}\hat{n}_{j+1}$ (31) $\displaystyle+V_{2}\sum_{j}\hat{n}_{j}\hat{n}_{j+2}$ where $J(t)$ and $V(t)$ are amplitudes of nearest neighbor hopping and interactions respectively, $V_{2}\ll|V(t)|$ is the amplitude of the second- neighbor interactions, $c_{j}$ denotes the fermion annihiliation operator on the $j^{\rm th}$ site of the chain, and $\hat{n}_{j}=c_{j}^{\dagger}c_{j}$ is the fermion density operator. In what follows, we choose a square pulse protocol so that $\displaystyle V(t)$ $\displaystyle=$ $\displaystyle-V_{1}\quad t\leq T/2,$ (32) $\displaystyle=$ $\displaystyle V_{1}\quad T/2<t\leq T$ $\displaystyle J(t)$ $\displaystyle=$ $\displaystyle J_{1}\quad t\leq T/2,$ (33) $\displaystyle=$ $\displaystyle J_{2}\quad T/2<t\leq T,$ with $V_{1}\gg J_{1},J_{2},V_{0},V_{2}$ so that one can reliably apply FPT to compute the Floquet Hamiltonian. We note that the protocol given by Eqs. 32 and 33 allows for simultaneous variation of the hopping and the interaction strengths of the fermions. To obtain an analytic expression for the first-order Floquet Hamiltonian, we first write the Hamitonian given by Eq. 31 as $H=H_{0}+H_{1}$ where $H_{0}=V(t)\sum_{j}\hat{n}_{j}\hat{n}_{j+1}$ and $\displaystyle H_{1}$ $\displaystyle=$ $\displaystyle-J(t)\sum_{j}\left(c_{j}^{\dagger}c_{j+1}+{\rm h.c.}\right)+V_{0}\sum_{j}\hat{n}_{j}\hat{n}_{j+1}$ (34) $\displaystyle+V_{2}\sum_{j}\hat{n}_{j}\hat{n}_{j+2}.$ We then follow the standard procedure and obtain the evolution operator corresponding to the term $H_{0}$ [21, 22, 23]. This yields $\displaystyle U_{0}(t,0)$ $\displaystyle=$ $\displaystyle e^{iV_{1}t\sum_{j}\hat{n}_{j}\hat{n}_{j+1}/\hbar}\quad\quad\quad t\leq T/2$ $\displaystyle=$ $\displaystyle e^{iV_{1}(T-t)\sum_{j}\hat{n}_{j}\hat{n}_{j+1}/\hbar}\quad T/2<t\leq T$ The Floquet Hamiltonian corresponding to $U_{0}(T,0)$ can be easily read off from Eq. III.2 to be identically $H_{F}^{(0)}=0$. Next, we consider the effect of the terms in $H_{1}$ using perturbation theory. The first order contribution to the evolution operator from $H_{1}$ is given by $\displaystyle U_{1}(T,0)$ $\displaystyle=$ $\displaystyle\frac{-i}{\hbar}\int_{0}^{T}dt\,U_{0}^{\dagger}(t,0)H_{1}U_{0}(t,0)$ (36) To obtain analytic expression of $U_{1}(T,0)$ we first note that the interaction terms in $H_{1}$ (Eq. 34) commute with $U_{0}$. Thus the contribution of this term to $U_{1}$ is trivially obtained and yields $\displaystyle U_{1a}(T,0)$ $\displaystyle=$ $\displaystyle\frac{-iT}{\hbar}\left(V_{0}\sum_{j}\hat{n}_{j}\hat{n}_{j+1}+V_{2}\sum_{j}\hat{n}_{j}\hat{n}_{j+2}\right)$ In contrast, the contribution from the hopping term in $H_{1}$ requires a more detailed analysis. To this end, using Eq. III.2, we write $\displaystyle U_{1b}(T,0)=\frac{iJ_{1}}{\hbar}\int_{0}^{T/2}e^{-iV_{1}t\sum_{j}\hat{n}_{j}\hat{n}_{j+1}/\hbar}\sum_{j}(c_{j}^{\dagger}c_{j+1}+{\rm h.c.})e^{iV_{1}t\sum_{j}\hat{n}_{j}\hat{n}_{j+1}/\hbar}$ $\displaystyle+\frac{iJ_{2}}{\hbar}\int_{T/2}^{T}e^{-iV_{1}T\sum_{j}\hat{n}_{j}\hat{n}_{j+1}/\hbar}e^{iV_{1}t\sum_{j}\hat{n}_{j}\hat{n}_{j+1}/\hbar}\sum_{j}(c_{j}^{\dagger}c_{j+1}+{\rm h.c.})e^{-iV_{1}t\sum_{j}\hat{n}_{j}\hat{n}_{j+1}/\hbar}e^{iV_{1}T\sum_{j}\hat{n}_{j}\hat{n}_{j+1}/\hbar},$ (38) where we have used Eqs. 32 and 33. To evaluate Eq. 38, we note that the hopping from site $j$ to $j+1$ costs an energy due to the nearest-neighbor interaction if it changes the number of bonds on the lattice whose both ends have sites occupied by fermions. This allows us to define an operator $\displaystyle\hat{A}_{j}$ $\displaystyle=$ $\displaystyle\hat{n}_{j+2}-\hat{n}_{j-1}$ (39) which takes values $\pm 1$ or $0$ on any site. The hopping of a fermion from a site $j$ changes the energy due to nearest-neighbor interaction only if $\hat{A}_{j}\neq 0$. This allows us to write $\displaystyle U_{1b}(T,0)=\frac{iJ_{1}}{\hbar}\int_{0}^{T/2}dt\sum_{j}(e^{-iV_{1}t\hat{A}_{j}/\hbar}c_{j}^{\dagger}c_{j+1}+{\rm h.c.})+\frac{iJ_{2}}{\hbar}\int_{T/2}^{T}dt\sum_{j}(e^{-iV_{1}(T-t)\hat{A}_{j}/\hbar}c_{j}^{\dagger}c_{j+1}+{\rm h.c.})$ (40) Carrying out the integrals in Eq. 40 and noting that $A_{j}$ can take values $0$ and $\pm 1$, we find [20] $\displaystyle U_{1b}(T,0)$ $\displaystyle=$ $\displaystyle\frac{iT}{\hbar}J_{s}\sum_{j}\left(\left[(1-\hat{A}_{j}^{2})+\hat{A}_{j}^{2}e^{-iV_{1}\hat{A}_{j}T/(4\hbar)}\frac{\sin V_{1}T/(4\hbar)}{V_{1}T/(4\hbar)}\right]c_{j}^{\dagger}c_{j+1}+{\rm h.c.}\right)=\frac{-iT\hat{J}_{c}}{\hbar}$ (41) where $J_{s}=(J_{1}+J_{2})/2$ and the expression of $\hat{J}_{c}$ can be read off from Eq. 41. Thus we find that for $\displaystyle V_{1}$ $\displaystyle=$ $\displaystyle 2m\hbar\omega_{D},$ (42) where $m\in Z$, the first order contribution to $U_{1}$ occurs only if $\hat{A}_{j}=0$. This in turn means that the first order evolution operator receives finite contribution from a constrained hopping term which propagates fermion hopping in such systems. This leads to a Floquet Hamitlonian that exhibits Hilbert space fragmentation similar to that derived in Ref. 20. Figure 1: (Color Online) (a) Plot of $S(nT)/S_{p}$ as a function of $n$ at $\omega_{D}=V_{1}/\hbar$ starting from a random Fock state for different values of the drive amplitude $V_{1}$. For all values of $V_{1}$, $S(nT)$ saturates to $S_{p}$. (b) Similar to (a) but at the special frequency $\omega_{D}=V_{1}/2\hbar$. With increase in $V_{1}$, $S(nT)$ saturates to $S_{p}^{f}$, the Page value of the fragment of $H_{F}^{(1)}$ from which the initial Fock state is chosen, for $n\leq 200$. (c) Plot of the density-density autocorrelator $C_{L}(nT)$ as a function of $n$ at $\omega_{D}=V_{1}/2\hbar$ starting from a infinite temperature thermal state. In this case too, the autocorrelator does not reach its thermal value of zero within the first 500 drive cycles, thus bearing signatures of prethermal HSF. (d) Value of $C_{L}(nT)$ after $n=5000$ drive cycles as a function of $V_{1}$ and $\hbar\omega_{D}/V_{1}$. The plot shows two special frequencies at $\hbar\omega_{D}/V_{1}=0.25$ and $\hbar\omega_{D}/V_{1}=0.5$. The time evolutions are performed using the exact unitary evolution operators, corresponding to drive protocols (32) and (33). The system sizes are $L=16$ for plots (a) and (b) and $L=14$ for (c) and (d). For all plots, $J_{1}=J_{2}/3=0.5$ and $V_{0}=V_{2}=2$. The derivation of the first order Floquet Hamiltonian from Eq. 41 and III.2 can be carried out in a straightforward manner [21] and yields $\displaystyle H_{F}^{(1)}$ $\displaystyle=$ $\displaystyle\hat{J}_{c}+V_{0}\sum_{j}\hat{n}_{j}\hat{n}_{j+1}+V_{2}\sum_{j}\hat{n}_{j}\hat{n}_{j+2}$ (43) Thus the fragmentation exhibited by $H_{F}^{(1)}$ for this protocol is identical to that found in Ref. 20. In addition, it also allows for variation of $J$ which makes the protocol much less restrictive compared to its counterpart in Ref. 20. The corresponding dynamical signatures in the half-chain entanglement entropy and the density-density autocorrelation function are shown in Fig. 1. For these plots, we use Eqs. 32 and 33 and set the hopping amplitudes to $J_{1}=0.5$ for the first half of the drive and $J_{2}=1.5$ for the next half cycle in Fig. 1. Also, we set $V_{0}=V_{2}=2$. Figs. 1(a) and 1(b) show the evolution of the half-chain entanglement entropy starting from a random Fock state from the half-filled sector in a chain of length $L=16$ with periodic boundary condition at $\omega_{D}=V_{1}/\hbar$ (generic frequency) and $\omega_{D}=V_{1}/2\hbar$ (special frequency satisfying the relation in Eq. 42) respectively. Fig. 1(a) shows that away from the special frequency, for all values of the drive amplitude, the entanglement entropy $S(nT)$ saturates to the Page value $S_{p}$ of the half-filled symmetry sector from which the initial state is chosen, as is expected of ergodic systems. In contrast to this, Fig. 1(b) shows that at the special frequency, the entanglement entropy fails to reach $S_{p}$ with increasing drive amplitude within the first $200$ drive cycles. Instead, with increase in drive amplitude, $S(nT)$ saturates to $S_{p}^{f}$, the Page value of the fragment of $H_{F}^{(1)}$, from which the initial state is chosen. Both $S_{p}$ and $S_{p}^{f}$ have been computed analytically and numerically following Ref. 20. Fig. 1(c) shows similar behavior for the time evolution of the density-density autocorrelator $\displaystyle C_{L}(nT)=\langle(n_{L}(nT)-1/2)(n_{L}(0)-1/2)\rangle$ (44) in an infinite temperature thermal state for a chain of length $L=14$ with open boundary condition at the special frequency $\omega_{D}=V_{1}/2\hbar$. A careful look at Eq. 44 reveals that $C_{L}$ also represents the connected autocorrelator since $\langle n_{L}(0)-1/2\rangle=0$. Thus, in an ergodic system, $C_{L}(nT)$ is expected to decay to zero at long enough times signifying loss of any initial memory. However, Fig. 1(c) shows that with increasing drive amplitude $V_{1}$, $C_{L}$ saturates to a value much higher than zero at long enough times. This can be explained by considering the fragmented structure of $H_{F}^{(1)}$ and the Mazur’s bound on the autocorrelator in the presence of the fragmented structure [20]. In [20], we had seen that the long-time saturation value of the autocorrelator was above the lower bound predicted by the Mazur’s bound. The autocorrelator decays down to zero when the chain is driven away from the special frequencies. Fig. 1(d) elucidates this by plotting the value of $C_{L}(nT)$ after $5000$ drive cycles as a function of the drive amplitude and the drive frequency. This plot can also serve as a “phase diagram” in the drive frequency and drive amplitude space, where non-zero saturation values of $C_{L}(nT)$ (bright regions in the color plot) indicate parameter regimes where prethermal fragmentation is observed. It is to be noted here that for a given drive amplitude, the rate of thermalization is faster as compared to that reported in [20]. This is to be attributed to the asymmetric drive protocol (different values of the hopping amplitude during the two half-cycles) used here. Due to the asymmetric nature of the protocol, the lowest non-trivial correction to the constrained Hamiltonian at the special frequency comes from the second-order Floquet Hamiltonian, $H_{F}^{(2)}$ as compared to $H_{F}^{(3)}$ in [20]. This, in turn, results in a shorter thermalization timescale. ## IV Other filling fractions Figure 2: (Color Online) (a) Plot of the ${\mathcal{D}}_{\rm sub- sector}/{\mathcal{D}}_{\rm sector}$ for $H_{F}^{(1)}$ as a function of $L$ for $N/L=1/3$ showing exponential reduction with $L$ similar to the half-filling case. (b) Similar plot for $N/L=1/4$. For both plots, $J=1$ and $V_{1}/(\hbar\omega_{D})=2m$. In this section, we study the driven fermion chain away from half-filling to demonstrate the robustness of the fragmentation signature. To this end, we consider the driven fermion chain at filling fractions $N/L=1/3,1/4$ (where $N$ is the fermion number and $L$ is the chain length). In what follows, we shall use the square-pulse protocol given by $V(t)=-(+)V_{1}$ for $t\leq(>)T/2$ in accordance with Ref. 20. We begin our study by analyzing the Hilbert space dimension (HSD) of the largest fragment of the first order Floquet Hamiltonian (Eq. 13) at $N/L=1/3$ and $1/4$. This is shown in Fig. 2 where the ratio of the HSD of the largest fragment, ${\mathcal{D}}_{{\rm sub-sector}}$, and the total HSD of the symmetry sector (one-third filled sector in (a) and one-fourth filled sector in (b)), ${\mathcal{D}}_{{\rm sector}}$, is plotted as a function of $L$ for $J=1$ and at a special frequency $V_{1}/(\hbar\omega_{D})=2m$, where $m$ is an integer. We find a clear signature of exponential decay of this ratio for both $1/3$ and $1/4$ filling fractions as a function of $L$. This indicates the possibility of the presence of signature of strong HSF in the dynamics of the driven chain at these filling. To verify this expectation, we compute the entanglement entropy $S(nT)$ as a function of $n$ and at different drive frequencies. For an ergodic driven system, $S$ is expected to increase with $n$ and eventually saturate to its Page value corresponding to the symmetry sector $S_{p}$ irrespective of the initial state chosen for the dynamics[27]. In contrast, for a chain which exhibits HSF, $S(nT)$ is expected to saturate to the Page value of the fragment to which the initial state belongs: $S\to S_{p}^{f}$. Thus the saturation value of $S$ is lower; also it depends on the initial state from which the dynamics originates. This allows one to distinguish between dynamical behavior of a driven chain with and without strong HSF. A plot of $S(nT)/S_{p}$, shown in Fig. 3(a) for $N/L=1/3$ and Fig. 3(b) for $N/L=1/4$, clearly shows the distinction between the behavior of $S$ at and away from the special frequencies. $S(nT)/S_{p}$ saturates, for both fillings, to unity at large $n$ away from the special frequencies($V_{1}/(\hbar\omega_{D})=1/2$); in contrast, at the special frequency $V_{1}/(\hbar\omega_{D})=2$, they saturate to a lower value which corresponds to $S_{p}^{f}$ of the respective sectors from which the initial states are chosen. Figure 3: (Color Online) (a) Growth of entanglement entropy from exact dynamics for $L=18$ and $N=6$, starting from a randomly chosen Fock state. The result is averaged over 10 such states chosen from the same fragment of the first order Floquet Hamiltonian. The sector dimension from which the state is chosen is $1980$. The entanglement entropy is scaled by the Page value $S_{p}$ of the symmetry sector for which $N/L=1/3$. At the special frequencies, EE saturates to a value less than $S_{p}$ but close to $S_{p}^{f}$ for the fragment from which the initial state was chosen. (b) Same as in (a) but for $L=20$ and $N=5$ corresponding to $1/4$ filling. The sector dimension of the fragment from which the initial state is chosen is $1050$. (c) Plot of $C_{L}(nT)$ as a function of $n$ at the special frequency $V_{1}/(\hbar\omega_{d})=2$ and away from it $V_{1}/(\hbar\omega_{D})=1/2$ for $N/L=1/3$ and $L=18$ The initial state is same as in (a). (d) Same as (c) but for $L=20$ and $N/L=1/4$; the initial state is same as in (b). For all plots $J=1$. In addition, we compute the density-density autocorrelation function given by Eq. 44. Fig. 3(c) and (d) show the behavior of $C_{L}(nT)$ as a function of $n$ at and away from the special frequency for $N/L=1/3$ and $1/4$ respectively. We find that in both cases, $C_{L}(nT)$ saturates to finite value at the special frequency and to zero away from it. Thus these plots confirm the existence of strong HSF similar to the half-filling sectors in these fermionic chains. Finally, in Fig. 4, we show the plot of $S(nT)/S_{p}$ starting from several initial Fock states which belong to different sectors with different values of $S_{p}^{f}$ for $V_{1}/(\hbar\omega_{D})=2$. We find that in each case, both at $N/L=1/3$ (Fig. 4(a)) and $1/4$ (Fig. 4(b)), $S(nT)/S_{p}<1$ for large $n$; moreover, $S(nT)\to S_{p}^{f}$ corresponding to the fragment of $H_{F}^{(1)}$ to which the initial state belongs. This clearly demonstrates signature of HSF at these filling fractions. Figure 4: (Color Online) Plot of $S(nT)/S_{p}$ as a function of $n$ at $V_{1}/(\hbar\omega_{D})=2$ starting from Fock states belonging to different fragments for (a) $L=18$ and $N=6$ and (b) $L=20$ and $N=5$. The numbers label the Hilbert space dimensions of the fragments and the dashed lines indicate the corresponding $S_{p}^{f}$. For all plots $J=1$. ## V Density-density out-of-time ordered correlation function In this section we analyze the out-of-time ordered correlator (OTOC) for the driven chain. The basic definitions are outlined in Sec. V.1. This is followed by numerical study of OTOC for a chain with periodic boundary condition (PBC) and starting from an infinite temperature thermal state in Sec. V.2. Finally we study the behavior of OTOC for fermion chains with open boundary condition (OBC) and starting from the ${\mathbb{Z}}_{2}=|0,1,0,1,\ldots\rangle$ Fock state in Sec. V.3. ### V.1 Preliminary The study of OTOC serves as an important tool to diagnose the rate of propagation of local information in a quantum system [29, 30, 31, 32]. Ergodic systems are known to exhibit ballistic spread of local information accompanied by a diffusive front. In case of non-ergodic systems, the behavior of information propagation, as detected using OTOC, ranges from logarithmic growth in many-body localized systems [33] to alternate scrambling and unscrambling in certain integrable systems [34]. For fragmented systems, the scrambling of information is expected to be slow since the Hamiltonian does not connect states belonging to different fragments; however, its detailed features, in the presence of a periodic drive, have not been studied earlier. To probe the rate of scrambling of information in our system, we study the temporal (in stroboscopic times) and spatial profile of the OTOC $\displaystyle F(r,nT)$ $\displaystyle=$ $\displaystyle\langle\tilde{n}_{i}(nT)\tilde{n}_{j}(0)\tilde{n}_{i}(nT)\tilde{n}_{j}(0)\rangle,$ (45) where $\tilde{n}_{i}=2n_{i}-1$ with $n_{i}$ and $n_{j}$ being the number density operators at sites $i$ and $j$ respectively and $r=|i-j|$ measures the distance between these two sites. We take the average with respect to both an infinite-temperature thermal state and the $|{\mathbb{Z}}_{2}\rangle$ state. Since the operator $\tilde{n}_{i}$ is hermitian and squares to identity, it can be shown that the function $F(r,nT)$ is related to the squared commutator $C(r,nT)$ as $\displaystyle C(r,nT)$ $\displaystyle=$ $\displaystyle\langle[\tilde{n}_{i}(nT),\tilde{n}_{j}]^{\dagger}[\tilde{n}_{i}(nT),\tilde{n}_{j}]\rangle$ (46) $\displaystyle=$ $\displaystyle 2(1-F(r,nT)).$ Cast in this form, it can be argued that as the operator $\tilde{n}_{i}$, initially localized at site $i$, spreads to the site $j$, the value of $C(r,nT)$ at this site gradually increases from zero and hence the OTOC, $F(r,nT)$ decreases from $1$. A higher value of $C(r,nT)$ (i.e. lower value of $F(r,nT)$) at a given instant of time therefore indicates larger spread of the local operator ($\tilde{n}_{i}$ in the present case). ### V.2 Infinite-temperature initial state Figure 5: (Color Online) (a) Profile of the OTOC $F(r,nT)$ in an infinite temperature thermal state for $L=14$ (half-filled sector) with PBC and $\omega_{D}=V_{1}/\hbar$. The initial operator $\tilde{n}_{i}$ is localized at the centre of the chain $i=7$. The index $j$ in the x-axis labels the site of the chain and $r=|j-7|$ for all plots. A ballistic spread of information is seen with the value of the OTOC rapidly dropping close to zero as is expected of a thermalizing system. (b) Same as (a) but at the special frequency $\omega_{1}^{*}=V_{1}/2\hbar$, where the first-order Floquet Hamiltonian $H_{F}^{(1)}$ is fragmented. Although there is some information scrambling in this case due to mixing of states within fragments, the value to which $F(r,nT)$ reaches at similar times is higher as compared to the thermalizing case. This implies that the extent of information scrambling is less in this case, compared to (a). (c) Profile of $F(r,nT)$ at site $j=14$ both at and away from the special frequency $\omega_{1}^{*}$, illustrating the same point. (d) Variation of the value of $F(r,nT)$ at the farthest site $j=14$ after $t=nT=15$ with drive amplitude $V_{1}$ and $\omega_{D}=\omega_{1}^{\ast}$. With decrease in $V_{1}$, higher-order terms in the pertubation series gain prominence and enhance information scrambling. For all plots, $J=1,V_{0}=V_{2}=2$. For (a)-(c), $V_{1}=50$. All the results are obtained using the exact time-evolution operator. In this section, we study the spread of OTOC in an infinite-temperature thermal state for a half-filled chain of length $L=14$ with PBC both at the special frequency ($\hbar\omega_{1}^{*}=V_{1}/2$, shown in Fig 5(b)) and away from it ($\omega_{D}=2\omega_{1}^{*}$, shown in Fig 5(a)). The operator is initially localized at the centre of the chain, i.e. $i=L/2$ in both the cases. Fig. 5(a) shows that at a generic frequency, the operator spreads ballistically. Such a spread can be inferred from the linear variation of $r$, for sites at which $F(r,t)$ has almost similar values, as a function of $t$. $F(r,nT)$ quickly falls to a value close to zero, implying that $C(r,nT)$ saturates to a value close to $1$. At the special frequency, however, Fig. 5(b) shows that although the local information reaches the farthest site almost at the same time as in the previous case, the OTOC saturates to a higher value as compared to its thermalizing counterpart. This is a direct consequence of the fact that the first order Floquet Hamiltonian $H_{F}^{(1)}$, which is fragmented, only allows mixing of the states within a particular fragment. Although the infinite-temperature thermal initial state (represented by a density matrix) weighs all the states equally, during time evolution they can only be connected with states belonging to the same fragment. As a result, they fail to spread out through the whole Hilbert space. The information is scrambled only due to mixing between states within individual fragments; this leads to lower scrambling in the prethermal regime than that due to ergodic evolution away from the special frequency. Fig. 5(c) illustrates this fact by plotting the value of $F(r,nT)$ for site $j=14$ both at and away from the special frequency. As the drive amplitude decreases, the higher order terms in the perturbation series start dominating, allowing mixing between different fragments. This enhances information scrambling leading to a decrease in the value of the OTOC. This is shown in Fig. 5(d) which plots the variation of the value of OTOC at the farthest site ($j=14$) at $nT=15$ as a function of $V_{1}$ for $\omega_{D}=\omega_{1}^{\ast}$. This shows that information scrambling is suppressed beyond a critical $V_{1}$ where signatures of fragmentation can be found over a long pre-thermal timescale. ### V.3 $\mathbb{Z}_{2}$ State Figure 6: (Color Online) (a) Profile of the OTOC $F(r,nT)$ in a $\mathbb{Z}_{2}$ state for L=14 with OBC and at a generic frequency $\omega_{D}=V_{1}/\hbar$. The initial operator $\tilde{n}_{i}$ is localized at one of the edges of the chain $i=1$. The information reaches to the other end of the system ballistically, bearing signature of thermalizing systems. (b) Same as (a) but at the special frequency $\omega_{1}^{*}=V_{1}/2\hbar$. The information after reaching the other end starts unscrambling again. This alternate scrambling and unscrambling of information continues over a short timescale, dictated by the quasienergy spectrum of the fragmented $H_{F}^{(1)}$ as explained in the main text. (c) Long-time oscillations in the profile of the OTOC at $\omega_{1}^{*}$. (d) Plot of $\chi_{\beta\alpha}(nT)$ (defined below Eq. 51 in the main text) as a function of $t=nT$. The color codes are: $\chi_{28}$ (blue), $\chi_{38}$ (red), $\chi_{48}$ (brown), $\chi_{58}$ (pink), $\chi_{68}$ (green) and $\chi_{78}$ (cyan). The red dashed lines mark integer multiples of $2\pi$. The first two times when these phases are very close integer multiples of $2\pi$, are marked as $t_{1}^{*}$ and $t_{2}^{*}$ respectively. These correspond to the first two recurrence times in (b). For all plots, $J=1,V_{0}=V_{2}=2,V_{1}=50$. In this section, we study the spatial and temporal profile of the OTOC in a $|{\mathbb{Z}}_{2}\rangle$ state at the special frequency with OBC. Fig. 6(a) shows that away from the special frequency, starting from one end of the chain, the information propagates ballistically to the other end, as is expected for a ergodic system; $F(r,t)$ monotonically decays to near-zero value at all sites within $nT\leq 10$. In contrast, as shown in Fig. 6(b) and (c), at the special frequency the behavior of $F(r,nT)$ is quite different and it shows signature of fragmentation. In Fig. 6(b), we find that at short time scales $nT\sim$ 10, there are initial fast oscillations which lead to alternate scrambling and unscrambling of information. Such alternate scrambling and unscrambling of information is reminiscent of the behavior of OTOC in integrable systems [34]; however, as we show below, the mechanism for this phenomenon is different in the present case. Furthermore, over longer time scales $nT\sim$ 100 - 500, we find slow oscillatory behavior as seen in Fig. 6(c). As discussed below, this is related to tunneling between two near degenerate states. Both the above oscillatory features can be related to the fact that at high drive amplitude and at the special frequencies, the dynamics is mostly governed by $H_{F}^{(1)}$ at short and intermediate timescales. To understand the behavior of $F$, we therefore focus on the fragment of $H_{F}^{(1)}$ (with OBC) to which $|{\mathbb{Z}}_{2}\rangle$ belongs. For $L=14$, there are $8$ states in this fragment namely $\displaystyle\mathcal{H}=$ $\displaystyle\\{|{\mathbb{Z}}_{2}\rangle,|j_{h}=2\rangle,|j_{h}=4\rangle,|j_{h}=6\rangle,$ $\displaystyle|j_{h}=8\rangle,|j_{h}=10\rangle,|j_{h}=12\rangle,|\bar{\mathbb{Z}}_{2}\rangle\\}$ (47) where $|\bar{\mathbb{Z}}_{2}\rangle=|1,0,1,0,\ldots\rangle,\text{ and }|j_{h}\rangle$ is a state with one hole-defect (where aa hole-defect implies two adjacent unoccupied sites) at position $j_{h}$ and zero particle defect (i.e. no two adjacent sites are occupied), viz $|j_{h}=2\rangle=|1,0,0,1,0,1,0,1,0,1,0,1,0,1\rangle$. Note, the constrained hopping introduces dynamics between these eight states, and $H_{F}^{(1)}$ in this subspace is equivalent to a nearest neigbor hopping model of a linear chain with eight sites and OBC. Here $|{\mathbb{Z}}_{2}\rangle$ and $|\bar{\mathbb{Z}}_{2}\rangle$ form the ends of the chain while $j_{h}=2,4,\ldots,12$ form the sites in between. The OTOC at a site $j$ will have the structure $F(r_{1},nT)=\langle\mathbb{Z}_{2}|\tilde{n}_{1}(nT)\tilde{n}_{j}(0)\tilde{n}_{1}(nT)\tilde{n}_{j}(0)|\mathbb{Z}_{2}\rangle$ (48) where $r_{1}=|j-1|$. Inserting the complete set of states $|m\rangle$ from this fragment and noting that the operator $\tilde{n}$ is diagonal in the Fock basis, this expression reads $F(r_{1},nT)\approx\sum_{m}(-1)^{j}f^{j}_{m}|\langle m(nT)|\tilde{n}_{1}|\mathbb{Z}_{2}(nT)\rangle|^{2}$ (49) where $f^{j}_{m}=\langle m|\tilde{n}_{j}|m\rangle$. Expanding $|\mathbb{Z}_{2}\rangle$ and $|m\rangle$ in the energy eigenstates of $H_{F}^{(1)}$: $|\mathbb{Z}_{2}\rangle=\sum_{\alpha}c_{\alpha}|\phi_{\alpha}\rangle$, $|m\rangle=\sum_{\beta}c_{\beta}^{m}|\phi_{\beta}\rangle$, Eq. 49 yields $F(r_{1},nT)\approx\sum_{m}(-1)^{j}f^{j}_{m}g_{m}(nT)$ (50) where $g_{m}(nT)=\Big{|}\sum_{\alpha,\beta}c_{\beta}^{m*}c_{\alpha}e^{-i\chi_{\beta\alpha}(t)}N^{1}_{\beta\alpha}\Big{|}^{2}$ (51) with $N^{1}_{\beta\alpha}=\langle\phi_{\beta}|\tilde{n}_{1}|\phi_{\alpha}\rangle$ being the matrix element of $\tilde{n}_{1}$ between the energy eigenstates and $\chi_{\beta\alpha}(t)=(\epsilon_{\alpha}-\epsilon_{\beta})nT/\hbar$. #### V.3.1 Short and Long time Oscillations In Eq. 50 the spatial dependence on $r_{1}$ or $j$ is factorized out from the time dependence $nT$. This implies that the time dependence of the OTOC is site-independent, i.e. the recurrence time at every site is the same and the recurrence happens when all the phases $\chi_{\beta\alpha}(t)$ are approximately close to integer multiples of $2\pi$. We arrange the spectrum $\epsilon_{1}<\epsilon_{2}<\ldots<\epsilon_{8}$. Numerically, we find that the matrix elements $N^{1}_{\beta\alpha}$ between the states $|\phi_{\beta}\rangle;\beta=2,3,\ldots,8$ and $|\phi_{\alpha}\rangle;\alpha=7,8$ are an order of magnitude higher than the rest of the off-diagonal matrix elements. This is because the states $\beta=1,2,\ldots,6$ are mostly made of the six single-hole wavefunction, while the states $\alpha=7,8$ are mostly made of the states $\mathbb{Z}_{2}$ and $\bar{\mathbb{Z}}_{2}$. Since the last two states have one extra next- nearest-neighbor interaction compared to the first six, $\epsilon_{\alpha\beta}\sim V_{2}$, and this energy scale shows up in the fast oscillations seen over timescales $nT\sim$ 10\. Thus, in this relatively short time, the recurrence is predominantly dictated by the phases $\chi_{\beta\alpha}(t)$ with $\beta=2,3,\ldots,7$ and $\alpha=8$. Fig. 6(d) plots these phases as a function of $t=nT$. It can be seen that the recurrence occurs when all these phases are close to $2\pi p_{0}$ (where $p_{0}\in Z$) as shown in Fig. 6(d). The first two of these times are marked with $t_{1}^{*}$ and $t_{2}^{*}$ in Fig.6(d). These are not exactly periodic because of involvement of multiple phases in the dynamics. It is also to be noted from Fig. 6(d) that the energies $\epsilon_{7,8}$ (which are mostly linear combinations of $\mathbb{Z}_{2}$ and $\bar{\mathbb{Z}}_{2}$ Fock states) are so close that for the short timescale involved, the phase $\chi_{87}(t)$ almost remains close to zero; it does not play much role in determining the short recurrence time. Thus, from the above discussion it is clear that the recurrences at short timescales owe their existence to two features in the model. First, the finite next-nearest-neighbor interaction energy $V_{2}$. Second, the OBC which allows the single-hole states to be included within the same fragment as to which the states $\mathbb{Z}_{2}$ and $\bar{\mathbb{Z}}_{2}$ belong (with PBC, the fragment has only the $\mathbb{Z}_{2}$ and $\bar{\mathbb{Z}}_{2}$ states). In Fig. 7(a), we show the comparison between results for $F(r_{1},nT)$ obtained from exact dynamics (solid lines) and the analytical estimate obtained from $H_{F}^{(1)}$ in Eq. 50 (dashed lines) for some representative sites $j=1,5,14$. It can be seen that the first two recurrence times at $t_{1}^{*}=8.07,t_{2}^{*}=16.21$ are well approximated by Eq. 50. The phase $\chi_{87}(t)$ manifests itself only at longer time scales of $nT\sim$ 100 - 500. As seen in Fig. 6(c), over this time scale $F(r_{1},nT)$ oscillates from values nearly one to nearly minus one with frequency $\Omega$, where $\Omega=(\epsilon_{8}-\epsilon_{7})/\hbar$. In Fig. 7(b) and (c) we show that these oscillations can be explained using Eq 50 by considering the $8$ states belonging to this fragment of $H_{F}^{(1)}$. In Appendix B, we show that a four state ansatz can be used to arrive at this result for a high next- nearest-neighbor interaction strength, when the $\mathbb{Z}_{2}$, $\bar{\mathbb{Z}}_{2}$ states are well separated in energy from the remaining $|j_{h}\rangle$ states. Figure 7: (Color Online) (a) Comparison of exact and approximate estimate (Eq. 50) of $F(r_{1},nT)$ for $j=1$ (green solid, brown dashed), $j=5$ (cyan solid, red dashed) and $j=14$ (pink solid, black dashed) sites. The solid lines represent results obtained from exact dynamics and dashed lines represent approximate estimates obtained using the fragment of $H_{F}^{(1)}$. The first two recurrence times are in good agreement, emphasizing the role of fragmentation in the scrambling and unscrambling behavior observed in Fig. 6(b). For $j=1$, the agreement between approximate and exact numerical result is nearly perfect and the green solid and the brown dashed lines are almost on top of each other. (b), (c) Similar comparison for the long time oscillations for $j=1,14$, following the color scheme used in (a). (b) Plot of exact numerical result for $j=1,14$. (c) Plot of approximate estimate for $j=1,14$. #### V.3.2 Spatial profile of OTOC The spatial dependence in the profile of the OTOC appears through the term $h_{jm}=(-1)^{j}f^{j}_{m}$ in Eq. 50. The initial linear increase in Fig. 6(b) for $nT\leq 5$ can be explained by focusing on this term. The profile of $h_{jm}$ for odd $j$’s reads $h_{jm}=\begin{pmatrix}1&-1&-1&-1&-1&-1&-1&-1\\\ 1&1&-1&-1&-1&-1&-1&-1\\\ 1&1&1&-1&-1&-1&-1&-1\\\ 1&1&1&1&-1&-1&-1&-1\\\ 1&1&1&1&1&-1&-1&-1\\\ 1&1&1&1&1&1&-1&-1\\\ 1&1&1&1&1&1&1&-1\\\ \end{pmatrix}$ (52) where the rows label the odd sites $j=1,3,5,\ldots 13$ and the columns label the different Fock states $|m\rangle$ in this fragment, in the same order as in Eq. 47. The time dependent functions $g_{m}(nT)$ are positive definite at all times. As $j$ increases, the number of $g_{m}(nT)$’s having positive weights increases linearly as is evident from Eq. 52. Thus, the shift of $F(r_{1},nT)$ from $1$ happens progressively at a later time as $j$ increases. It is also useful to note that $f^{2k-1}_{m}=-f^{2k}_{m}$ for all $k\text{ and }m$, so that $h_{2k-1,m}=h_{2k,m}$ and hence at any given instant of time, $F(2k-1,nT)=F(2k-2,nT)$. Thus the even sites $j$ have similar behavior as the odd sites. ## VI Discussion In this work, we studied the dynamics of a periodically driven Fermi chain and extended the study of prethermal HSF in these system undertaken in Ref. 20 in several ways. First, we have studied the existence of such prethermal HSF beyond half- filling in such chains. We found the existence of such prethermal HSF phase for several other filling fractions such as $N/L=1/4$ and $1/3$. This shows that the robustness of the pretheraml MBL phase in such driven chain. Second, we provide a derivation of the first and second order Floquet Hamiltonian in such driven system in an alternative manner. Our derivation brings out the commutator structure of the Floquet Hamiltonian; in particular, we find that the second order term in the Floquet Hamiltonian, $H_{F}^{2}$, can be expressed as a commutator of a constrained current operator $\sum_{j}A_{j}(c_{j}^{\dagger}c_{j}-h.c.)$ with $H_{F}^{(1)}$. We expect similar commutation relations to hold for higher order terms in $H_{F}$; this sheds light on the symmetry content of the higher order terms in the Floquet Hamiltonian for the cosine drive protocol. Third, we extend our analysis to experimentally relevant and slightly more complicated drive protocols. In a typical experiment, involving ultracold atoms, the interaction strength between fermions and their hopping strength are both controlled by intensities of the applied lasers. Consequently, experimentally relevant protocols must allow change of both the hopping amplitudes and interaction strength. We show that the prethermal HSF is stable for a large class of such drives and chart out a phase diagram for the special frequencies at which it occurs. Finally, we study the behavior of density-density OTOC for such driven systems. Our study shows that such OTOCs can serve as detectors of such prethermal HSF in two distinct ways. First, irrespective of the boundary condition used, the OTOC $F(r,t)$ for a finite fermion chain driven at the special frequency and starting from a ${\mathbb{Z}}_{2}$ initial state, exhibits a larger long-time value than when driven away from such frequencies. In addition, it also exhibits oscillations with very large periodicity at the special frequencies that have the same origin as the correlation functions discussed in Ref. 20. In contrast, no such oscillations are found when one is away from the special frequency; the OTOC monotonically decreases to zero. Second, for fermion chains with open boundary condition and driven at special frequencies, we find periodic scrambling and unscrambling of information which is in sharp contrast to standard behavior of OTOCs in driven ergodic systems. Such a behavior was found earlier for integrable spin chains [34]; however, their origin for systems with prethermal HSF quite different and can be tied to the localization of the driven system within a group of Fock states with same dipole number ($n_{d}=0$). For chains with open boundary condition, there are $O(L)$ such states which govern the dynamics up to a long prethermal time scale leading to periodic scrambling and unscrambling. This phenomenon is qualitatively different from the behavior of OTOC away from the special frequencies where it monotonically decays due to fast spread of the driven system through the Hilbert space; it is also different for a chain with PBC with two Fock states (${\mathbb{Z}}_{2}$ and ${\bar{\mathbb{Z}}_{2}}$) in the $n_{d}=0$ sector where no such unscrambling is found. Experimental verification of our result would require realization of isolated fermi chain. A possible scenario for this is a $1D$ fermion systems with nearest neighbor hopping and local interaction realized suing ultracold fermions in an optical lattice. We propose to drive this with the experimentally relevant protocol discussed in this work; this can be achieved by varying strength of lasers used to generate the lattice [24, 25]. The simplest measurement would involve measuring $\langle\hat{n}_{d}\rangle$ as a function of time. We predict that the dynamics of $\langle n_{d}\rangle$ staring from the ${\mathbb{Z}}_{2}$ state for such a chain will be approximately constant (and close to zero) for a long prethermal timescale when the system is driven at the special frequency. This is to be contrasted the behavior of $\langle n_{d}\rangle$ away from the special frequency which should exhibit rapid dynamics at short timescale. In conclusion, we have studied several aspects of prethermal HSF in a driven Fermi chain. Our results have showed the robustness of this phenomenon by confirming its existence for different, experimentally relevant, drive protocols and also when the system is away from half-filling. In addition we have demonstrated that OTOCs may serve as a marker for such prethermal HSF; they exhibit periodic scrambling and unscrambling for fermion chains with open boundary condition driven at the special frequency. Acknowledgement: SG acknowledges CSIR, India for support through project 09/080(1133)/2019-EMR-I. IP thanks Edouard Boulat for discussions. KS thanks DST, India for support through SERB project JCB/2021/000030 and Arnab Sen for discussions. ## Appendix A Computation of $\mathcal{H}_{F}^{(2)}$ for cosine protocol The second order Floquet Hamiltonian can be broken into two parts $\mathcal{H}_{F}^{(2)}=\mathcal{H}_{F}^{(2a)}+\mathcal{H}_{F}^{(2b)}$. From Eq. (26) we get $\displaystyle\mathcal{H}_{F}^{(2a)}$ $\displaystyle=\frac{-iJ^{2}}{2\hbar T}\sum_{i,j}\sum_{m,n}B_{m,n}\left[J_{m}(\lambda\hat{A}_{i})c^{\dagger}_{i}c_{i+1}+J_{m}(-\lambda\hat{A}_{i})\right.$ $\displaystyle\times\left.c^{\dagger}_{i+1}c_{i}\,,\,J_{n}(\lambda\hat{A}_{j})c^{\dagger}_{j}c_{j+1}+J_{n}(-\lambda\hat{A}_{j})c^{\dagger}_{j+1}c_{j}\right],$ (53) where $B_{m,n}\equiv\int_{0}^{T}d\tau_{1}\int_{0}^{\tau_{1}}d\tau_{2}\,e^{im\omega\tau_{1}}e^{in\omega\tau_{2}}$ (54) The evaluation of the above integrals yield $\displaystyle B_{m,n}$ $\displaystyle=\frac{T^{2}}{2}\delta_{m,0}\delta_{n,0}+\frac{T}{im\omega}\delta_{n,0}(1-\delta_{m,0})$ $\displaystyle-\frac{T}{in\omega}\delta_{m,0}(1-\delta_{n,0})+\frac{T}{in\omega}(1-\delta_{n,0})\delta_{m,-n}.$ (55) Due to the commutator structure of Eq. (A) the first and the fourth terms above do not contribute. The second and the third terms are non-zero and equal. Next, due to the $1/m$ factor in the second term, only integers $m$ contribute. We get $\displaystyle\mathcal{H}_{F}^{(2a)}$ $\displaystyle=-\frac{2J^{2}}{\hbar\omega}\sum_{k=0}^{\infty}\frac{1}{2k+1}\left[\sum_{i}J_{2k+1}(\lambda\hat{A}_{i})\right.$ $\displaystyle\times\left.\left(c^{\dagger}_{i}c_{i+1}-{\rm h.c.}\right)\,,\,\sum_{j}J_{0}(\lambda\hat{A}_{j})\left(c^{\dagger}_{j}c_{j+1}+{\rm h.c.}\right)\right].$ (56) Noting that $\hat{A}_{i}$ can only take values 0, 1, -1, we have the relation $J_{2k+1}(\lambda\hat{A}_{i})=\hat{A}_{i}J_{2k+1}(\lambda).$ Using the above we get $\displaystyle\mathcal{H}_{F}^{(2a)}$ $\displaystyle=-\frac{2J^{2}}{\hbar\omega}\mathcal{C}(\lambda)\left[\sum_{i}\hat{A}_{i}\left(c^{\dagger}_{i}c_{i+1}-{\rm h.c.}\right)\right.\,,\,$ $\displaystyle\left.\sum_{j}J_{0}(\lambda\hat{A}_{j})\left(c^{\dagger}_{j}c_{j+1}+{\rm h.c.}\right)\right],$ (57) where $\mathcal{C}(\lambda)\equiv\sum_{k=0}^{\infty}\frac{J_{2k+1}(\lambda)}{2k+1}.$ For the second term $\mathcal{H}_{F}^{(2b)}$ we note that $\int_{0}^{T}d\tau_{1}\int_{0}^{\tau_{1}}d\tau_{2}\tilde{\mathcal{H}}_{p}(\tau_{2})=\int_{0}^{T}d\tau_{1}\int_{\tau_{1}}^{T}d\tau_{2}\tilde{\mathcal{H}}_{p}(\tau_{1}).$ Using the above relation and Eq. (III.1) we get $\displaystyle\mathcal{H}_{F}^{(2b)}$ $\displaystyle=\frac{i}{2\hbar T}\int_{0}^{T}d\tau(T-2\tau)\left[\tilde{\mathcal{H}}_{p}(\tau)\,,\,\hat{K}\right]$ $\displaystyle=\frac{-iJ}{2\hbar T}\sum_{i,m}\int_{0}^{T}d\tau(T-2\tau)e^{im\omega\tau}$ $\displaystyle\times\left[J_{m}(\lambda\hat{A}_{i})c^{\dagger}_{i}c_{i+1}+J_{m}(-\lambda\hat{A}_{i})c^{\dagger}_{i+1}c_{i}\,,\,\hat{K}\right]$ For the $\tau$-integral above we use the relation $\int_{0}^{T}d\tau(T-2\tau)e^{im\omega\tau}=-\frac{2T}{im\omega}(1-\delta_{m,0}).$ The appearance of the factor $1/m$ in the above implies that, again, only the Bessel functions with odd indices contribute. This gives $\mathcal{H}_{F}^{(2b)}=\frac{2J\mathcal{C}(\lambda)}{\hbar\omega}\left[\sum_{i}\hat{A}_{i}\left(c^{\dagger}_{i}c_{i+1}-{\rm h.c.}\right)\,,\,\hat{K}\right].$ (58) The Eqs. (A) and (58) imply Eq. (30) in the main text. ## Appendix B 4-state ansatz for the long-time OTOC oscillations Figure 8: (Color Online) Comparison of exact results for OTOC $F(r_{1},nT)$ with that obtained from Eq. 61 and 63, starting from a $|\mathbb{Z}_{2}\rangle$ state. (a) Green (light-colored) line is obtained from exact dynamics for $j=1$, (b) Pink (light-colored) line is obtained from exact dynamics for $j=14$, (c) Brown (dark-colored) is obtained from the 4 state ansatz, given by Eq. 59 for $j=1$. (d)Black (dark-colored) is obtained from the 4 state ansatz, for $j=14$. For all the plots, $L=14,V_{0}=2,V_{2}=6$ and $V_{1}=50$ with $\omega_{D}=V_{1}/2\hbar$. In this section, we discuss a simplification of Eq. 50 when a high value of the next-nearest-neighbor interaction $V_{2}$ is considered. In this case, the $\\{|\mathbb{Z}_{2}\rangle,|\bar{\mathbb{Z}}_{2}\rangle\\}$ states, which have one extra next-nearest-neighbor pair, are well separated in energy from the other $|j_{h}\rangle$ states. Thus, the extent of hybridization between $\\{|\mathbb{Z}_{2}\rangle,|\bar{\mathbb{Z}}_{2}\rangle\\}$ states and the $|j_{h}\rangle$ is small and hence the eigenfunctions of $H_{F}^{(1)}$ can be written down in terms of only a few Fock states as we show below. We assume that the two highest energy levels $|\phi_{7}\rangle$ and $|\phi_{8}\rangle$ are mostly symmetric and anti-symmetric combinations of $\mathbb{Z}_{2}$ and $\bar{\mathbb{Z}}_{2}$ with very small contributions from the two “nearest” $|j_{h}\rangle$ states, i.e. $|j_{h}=2\rangle$ and $|j_{h}=12\rangle$. By “nearest”, we refer to states which can be connected to $\mathbb{Z}_{2},\bar{\mathbb{Z}}_{2}$ by one constrained hop. We also consider two more states $|\phi_{5,6}\rangle$ which are orthogonal to $|\phi_{7,8}\rangle$ and have energies $\epsilon_{5,6}$. We assume these last two states to be nearly degenerate i.e. $\epsilon_{5}\approx\epsilon_{6}=E_{0}$ and the splitting between the two highest states $\epsilon_{8}-\epsilon_{7}=\Omega\ll V_{2}$, $\epsilon_{7}-\epsilon_{6}=V_{2}$. Thus $\displaystyle|\phi_{8}\rangle$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2}}\mathcal{C}(|\mathbb{Z}_{2}\rangle-|\bar{\mathbb{Z}}_{2}\rangle)+\frac{1}{\sqrt{2}}\mathcal{S}(|2\rangle-|12\rangle)$ $\displaystyle|\phi_{7}\rangle$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2}}\mathcal{C}(|\mathbb{Z}_{2}\rangle+|\bar{\mathbb{Z}}_{2}\rangle)+\frac{1}{\sqrt{2}}\mathcal{S}(|2\rangle+|12\rangle)$ $\displaystyle|\phi_{6}\rangle$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2}}\mathcal{S}(|\mathbb{Z}_{2}\rangle-|\bar{\mathbb{Z}}_{2}\rangle)-\frac{1}{\sqrt{2}}\mathcal{C}(|2\rangle-|12\rangle)$ $\displaystyle|\phi_{5}\rangle$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2}}\mathcal{S}(|\mathbb{Z}_{2}\rangle+|\bar{\mathbb{Z}}_{2}\rangle)-\frac{1}{\sqrt{2}}\mathcal{C}(|2\rangle+|12\rangle)$ (59) where $\mathcal{C}=\cos{\theta}$ and $\mathcal{S}=\sin{\theta}$ with $\theta$ being a phenomenological parameter to be determined from diagonalization. Inverting these relations, the time evolved states read $\displaystyle|\mathbb{Z}_{2}(t)\rangle$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2}}\left[\mathcal{C}e^{-iV_{2}t}(|\phi_{7}\rangle+|\phi_{8}\rangle e^{-i\Omega t})\right.$ $\displaystyle\left.+\mathcal{S}(|\phi_{5}\rangle+|\phi_{6}\rangle)\right]$ $\displaystyle|\bar{\mathbb{Z}}_{2}(t)\rangle$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2}}\left[\mathcal{C}e^{-iV_{2}t}(|\phi_{7}\rangle\right.$ $\displaystyle\left.-|\phi_{8}\rangle e^{-i\Omega t})+\mathcal{S}(|\phi_{5}\rangle-|\phi_{6}\rangle)\right]$ $\displaystyle|2(t)\rangle$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2}}\left[\mathcal{S}e^{-iV_{2}t}(|\phi_{7}\rangle\right.$ $\displaystyle\left.+|\phi_{8}\rangle e^{-i\Omega t})-\mathcal{C}(|\phi_{5}\rangle+|\phi_{6}\rangle)\right]$ $\displaystyle|12(t)\rangle$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2}}\left[\mathcal{S}e^{-iV_{2}t}(|\phi_{7}\rangle\right.$ (60) $\displaystyle\left.-|\phi_{8}\rangle e^{-i\Omega t})-\mathcal{C}(|\phi_{5}\rangle-|\phi_{6}\rangle)\right]$ where we have set the reference energy $E_{0}=0$. Using these 4 states, Eq. 49 for sites $j=1,14$ can be simplified as $\displaystyle F(0,nT)=f_{1}(nT)-f_{2}(nT)$ $\displaystyle F(13,nT)=f_{1}(nT)+f_{2}(nT)$ (61) where $\displaystyle f_{1}(nT)$ $\displaystyle=$ $\displaystyle|\langle\mathbb{Z}_{2}(nT)|\tilde{n}_{1}|\mathbb{Z}_{2}(nT)\rangle|^{2}-|\langle\bar{\mathbb{Z}}_{2}(nT)|\tilde{n}_{1}|\mathbb{Z}_{2}(nT)\rangle|^{2}$ $\displaystyle f_{2}(nT)$ $\displaystyle=$ $\displaystyle|\langle 2(nT)|\tilde{n}_{1}|\mathbb{Z}_{2}(nT)\rangle|^{2}+|\langle 12(nT)|\tilde{n}_{1}|\mathbb{Z}_{2}(nT)\rangle|^{2}$ Using the time-evolved states in Eq. 60, we obtain $\displaystyle f_{1}(t)$ $\displaystyle=$ $\displaystyle\mathcal{C}^{8}\cos{2\Omega t}-2\mathcal{C}^{6}\mathcal{S}^{2}\sin{\Omega t}\big{(}\sin{(V_{2}+\Omega)t}-\sin{V_{2}t}\big{)}+2\mathcal{C}^{4}\mathcal{S}^{4}\left[2\cos{\Omega t}-1+2\big{(}1-\cos{(V_{2}+\Omega)t}-\cos{V_{2}t}\big{)}^{2}\right]$ $\displaystyle-$ $\displaystyle 4\mathcal{S}^{2}\mathcal{C}^{2}\left(\mathcal{S}^{4}+\mathcal{C}^{4}\cos{\Omega t}\right)\left(1-\cos{(V_{2}+\Omega)t}-\cos{V_{2}t}\right)+\mathcal{S}^{8}$ $\displaystyle f_{2}(t)$ $\displaystyle=$ $\displaystyle\mathcal{S}^{2}\mathcal{C}^{2}\left[\mathcal{S}^{4}\big{(}\cos{(V_{2}+\Omega)t}-\cos{V_{2}t}\big{)}^{2}+\left(\mathcal{C}^{2}\sin{\Omega t}+\mathcal{S}^{2}(\sin{(V_{2}+\Omega)t}-\sin{V_{2}t})\right)^{2}+\left(\sin{(V_{2}+\Omega)t}+\sin{V_{2}t}\right)^{2}\right]$ (63) $\displaystyle+$ $\displaystyle\mathcal{S}^{2}\mathcal{C}^{2}\left[\mathcal{S}^{2}-\mathcal{C}^{2}\cos{\Omega t}+(2\mathcal{C}^{2}-1)\left(\cos{(V_{2}+\Omega)t}+\cos{V_{2}t}-1\right)\right]^{2}$ For $V_{2}\gg J$, $\theta\ll 1$, we retain up to quadratic terms in $\theta$, yielding $\displaystyle F(0,t)$ $\displaystyle=$ $\displaystyle\left(1-4\theta^{2}\right)\cos{2\Omega t}+6\theta^{2}-2\theta^{2}\cos{\Omega t}$ $\displaystyle+$ $\displaystyle 4\theta^{2}\cos{V_{2}t}(\cos{\Omega t}-1)$ $\displaystyle F(13,t)$ $\displaystyle=$ $\displaystyle\left(1-4\theta^{2}\right)\cos{2\Omega t}-6\theta^{2}-6\theta^{2}\cos{\Omega t}$ (64) $\displaystyle+$ $\displaystyle 4\theta^{2}\cos{V_{2}t}(3\cos{\Omega t}+1)$ We compare in Fig. 8 the exact results and those obtained from Eq. 63 for sites $j=1$ and $j=14$ where we find good agreement between the two. The parameters chosen are $V_{0}=2,V_{2}=6,V_{1}=50$ and $\omega_{D}=V_{1}/2\hbar$. ## Appendix C OTOC in $\mathbb{Z}_{2}$ state with PBC Figure 9: (Color Online) Profile of the OTOC $F(r,nT)$ in a $\mathbb{Z}_{2}$ state for $L=14$ with PBC and $\omega_{D}=V_{1}/2\hbar$. The initial operator is localized at $i=1$. This too shows alternate scrambling and unscrambling with a period $\tau\sim 300$. The parameters are $V_{0}=10V_{2}=2,V_{1}=20$ In this appendix, we consider the evolution of the profile of the OTOC starting from a $|\mathbb{Z}_{2}\rangle$ state with PBC at the special frequency. Fig. 9 shows that in this case too, there is alternate scrambling and unscrambling for $nT\leq 1200$ and $V_{0}=10V_{2}=2,V_{1}=20$. In the following we show that these oscillations can be interpreted as tunneling back and forth between the states $\mathbb{Z}_{2}$ and $\bar{\mathbb{Z}}_{2}$ when the system is close to the fragmented limit. Such tunneling events were reported in Ref. [20]. This can be understood by noting that for PBC, the $|\mathbb{Z}_{2}\rangle$ and $|\bar{\mathbb{Z}}_{2}\rangle$ states are degenerate frozen states for $H_{F}^{(1)}$. This degeneracy is lifted by higher-order hopping terms and for exact $H_{F}$, there are two eigenstates which are symmetric and anti-symmetric combinations of $|\mathbb{Z}_{2}\rangle$ and $|\bar{\mathbb{Z}}_{2}\rangle$ states viz $\chi_{+}=\frac{1}{\sqrt{2}}(|\mathbb{Z}_{2}\rangle+|\bar{\mathbb{Z}}_{2}\rangle),\quad\chi_{-}=\frac{1}{\sqrt{2}}(|\mathbb{Z}_{2}\rangle-|\bar{\mathbb{Z}}_{2}\rangle)$ . The energy splitting between these two states is given by $\Omega=\epsilon_{-}-\epsilon_{+}$. This implies the time evolutions $\displaystyle|\mathbb{Z}_{2}(t)\rangle$ $\displaystyle=e^{i\epsilon_{-}t}\left[(e^{i\Omega t}+1)|\mathbb{Z}_{2}\rangle+(e^{i\Omega t}-1)|\bar{\mathbb{Z}}_{2}\rangle\right]/2,$ $\displaystyle|\bar{\mathbb{Z}}_{2}(t)\rangle$ $\displaystyle=e^{i\epsilon_{-}t}\left[(e^{i\Omega t}-1)|\mathbb{Z}_{2}\rangle+(e^{i\Omega t}+1)|\bar{\mathbb{Z}}_{2}\rangle\right]/2.$ (65) Inserting an approximate complete set comprising of the states $\mathbb{Z}_{2}$ and $\bar{\mathbb{Z}}_{2}$ in Eq. 48 of the main text, we obtain $\displaystyle F(r_{1},t)\approx\left|\langle\mathbb{Z}_{2}|\tilde{n}_{1}(t)|\mathbb{Z}_{2}\rangle\right|^{2}-\left|\langle\mathbb{Z}_{2}|\tilde{n}_{1}(t)|\bar{\mathbb{Z}}_{2}\rangle\right|^{2}$ $\displaystyle=\left|\langle\mathbb{Z}_{2}(t)|\tilde{n}_{1}|\mathbb{Z}_{2}(t)\rangle\right|^{2}-\left|\langle\mathbb{Z}_{2}(t)|\tilde{n}_{1}|\bar{\mathbb{Z}}_{2}(t)\rangle\right|^{2},$ (66) where the last line is going from Heisenberg to Schrodinger picture. Note, the above equation already captures an important aspect of Fig. 9, namely $F(r_{1},t)$ is mostly independent of $r_{1}$ at this timescale. Using Eq. 65,we get $F(r_{1},t)\approx\cos(2\Omega t).$ (67) Thus, at stroboscopic times $t=nT$, where $n\Omega T$ is close to a integer multiple of $\pi$, the OTOC is close to one while, when $n\Omega T$ is close to a half-integer mutiple of $\pi$, the OTOC is close to minus one. ## References * [1] A. Polkovnikov, K., Sengupta, A. Silva, A. and M. Vengalattore, Rev. Mod. Phys. 83, 863 (2011). * [2] L. D’Alessio, Y. Kafri, A. Polkovnikov, and M. Rigol, Adv. Phys. 65, 239 (2016). * [3] M. Bukov, L. D’Alessio, and A. Polkovnikov, Adv. Phys. 64, 139 (2014). * [4] J. M. Deutsch, Phys. Rev. A 43, 2046 (1991); M. Srednicki, Phys. Rev. E 50, 888 (1994); M. Srednicki, J. Phys. A 32, 1163 (1999). * [5] M. Rigol, V. Dunjko, M. Olshanii, Nature 452, 854 (2008); P. Reimann, Phys. Rev. Lett. 101, 190403 (2008). * [6] L. D’Alessio and M. Rigol, Phys. Rev. X 4, 041048 (2014). * [7] M. Basko, I. L. Aleiner, and B. L. Altshuler, Ann. Phys. 321, 1126 (2006). * [8] R. Nandkishore and D. Huse, Ann. Rev. Cond. Mat. 6, 15 (2015) * [9] T. Kohlert, S. Scherg, X. Li, H. P. Luschen, S.D. Sarma, I. Bloch, and M. Aidelsburger, Phys. Rev. Lett. 122, 170403 (2019). * [10] A. Chandran, T. Iadecola, V. Khemani, R. Moessner, Annual Review of Condensed Matter Physics 14, 443 (2023). * [11] C. J. Turner, A. A. Michailidis, D. A. Abanin, M. Serbyn, and Z. Papi´c, Nat. Phys. 14, 745 (2018); S. Maudgalya, N. Regnault, and B. A. Bernevig, Phys. Rev. B 98, 235156 (2018). * [12] W. W. Ho, S. Choi, H. Pitchler, and M. D. Lukin, Phys. Rev. Lett. 122, 040603 (2019); N. Shiraishi, J. Stat. Mech. 08313 (2019). * [13] S. Choi, C. J. Turner, H. Pichler, W. W. Ho, A. A. Michailidis, Z. Papi´c, M. Serbyn, M. D. Lukin, and D. A. Abanin, Phys. Rev. Lett. 122, 220603 (2019); T. Iadecola, M. Schecter, and S. Xu, Phys. Rev. B 100, 184312 (2019). * [14] P. A. McClarty, M. Haque, A. Sen and J. Richter, Phys. Rev. B 102, 224303(2020); D. Banerjee and A. Sen Phys. Rev. Lett. 126, 220601(2021); S. Biswas, D. Banerjee, and A. Sen, SciPost Phys. 12, 148 (2022). * [15] V. Khemani, M. Hermele and R. Nandkishore, Phys. Rev. B 101, 174204 (2020); P. Sala, T. Rakovszky, R. Verresen, M. Knap and F. Pollmann, Phys. Rev. X 10, 011047 (2020). * [16] T. Rakovszky, P. Sala, R. Verresen, M. Knap and F. Pollmann, Phys. Rev. B 101, 125126 (2020); Z.-C. Yang, F. Liu, A. V. Gorshkov and T. Iadecola, Phys. Rev. Lett. 124, 207602 (2020); G. De Tomasi, D. Hetterich, P. Sala, and F. Pollmann, Phys. Rev. B 100, 214313(2019); P. Frey, L. Hackl, and S. Rachel1, arXiv:2209.11777 (unublished); D. Vu, K. Huang, X. Li, and S. Das Sarma, Phys. Rev. Lett. 128, 146601 (2022). * [17] S. Moudgalya and O. I. Motrunich, Phys. Rev. X 12, 011050 (2022); D. T. Stephen, O. Hart, and R. M. Nandkishore, arXiv:2209.03966 (unpublished); D. Hahn, P. A. McClarty, D. J. Luitz, SciPost Phys. 11, 074 (2021); N. Regnault and B. A. Bernevig, arXiv:2210.08019 (unpublished); T. Kohlert, S. Scherg, P. Sala, F. Pollmann, B. H. Madhusudhana, I. Bloch, and M. Aidelsburger, arXiv:2106.15586 (unpublished). * [18] B. Mukherjee, D. Banerjee, K. Sengupta, and A. Sen, Phys. Rev. B 104, 155117 (2021); P. Brighi, M. Ljubotina, and M. Serbyn, arXiv:2210.15607 (unpublished). * [19] A. Chattopadhyay, B. Mukherjee, K. Sengupta, and A. Sen, SciPost Phys. 14, 146 (2023); J Lehmann, P Sala, F Pollmann, and T Rakovszky, SciPost Phys. 14, 140 (2023); Y. H. Kwan, P. H. Wilhelm, S. Biswas, and S. A. Parameswaran, arXiv:2304.02669 (unpublished). * [20] S. Ghosh, I. Paul, and K. Sengupta, Phys. Rev. Lett. 130, 120401 (2023). * [21] A. Sen, D. Sen, and K. Sengupta, J. Phys. Condens. Matter 33, 443003 (2021). * [22] A. Soori and D. Sen, Phys. Rev. B 82, 115432 (2010). * [23] T. Bilitewski and N. R. Cooper, Phys. Rev. A 91, 063611 (2015). * [24] I. Bloch, J. Dalibard, and W. Zwerger, Rev. Mod. Phys. 80, 885 (2008). * [25] For a recent review, see J. Dobrzyniecki and T. Sowinski, Adv. Quantum Technol. 3, 2000010 (2020). * [26] R. Ghosh, D. Das and K. Sengupta, JHEP 05, 172 (2021). * [27] D. N. Page, Phys. Rev. Lett. 71, 1291 (1993); L. Vidmar and M. Rigol, Phys. Rev. Lett. 119, 220603 (2017). * [28] P. Mazur, Physica 43, 533 (1969); J. Sirker, SciPost Phys. Lect. Notes 17, 1 (2020). * [29] A. I. Larkin and Y. N. Ovichinnikov, Soviet J. Exp. Theor. Phys. 28, 1200 (1969); I. L. Aleiner, L. Faoro, and L. B. Ioffe, Ann. Phys-N. Y. 375, 378 (2016). * [30] A. Nahum, S. Vijay, and J. Haah, Phys. Rev. X 8, 021014 (2018); C. W. von Keyserlingk, T. Rakovszky, F. Pollmann, and S. L. Sondhi, Phys. Rev. X 8, 021013 (2018); S. Xu, and B. Swingle, Locality, Phys. Rev. X 9, 031048 (2019). * [31] M. Garttner, J. G. Bohnet, A. Safavi-Naini, M. L. Wall, J. J. Bollinger, and A. M. Rey, Nat. Phys. 13, 781 (2017). * [32] For a recent review, see I. García-Mata, R. A. Jalabert, D. A. Wisniacki, Scholarpedia, 18, 55237 (2023) * [33] Y. Chen, arXiv:1608.02765 (unpublished); R. Fan, P. Zhang, H. Shen, and H. Zhai, Science Bulletin 62, 707 (2017). * [34] S. Sur and D. Sen, arXiv:2210.15302 (unpublished); ibid, arXiv:2310:12226 (unpublished).
# Steps Minimize Dissipation in Rapidly Driven Stochastic Systems Steven Blaber<EMAIL_ADDRESS>Dept. of Physics, Simon Fraser University, Burnaby, British Columbia V5A 1S6, Canada Miranda D. Louwerse Dept. of Chemistry, Simon Fraser University, Burnaby, British Columbia V5A 1S6, Canada David A. Sivak<EMAIL_ADDRESS>Dept. of Physics, Simon Fraser University, Burnaby, British Columbia V5A 1S6, Canada ###### Abstract Micro- and nano-scale systems driven by rapid changes in control parameters (control protocols) dissipate significant energy. In the fast-protocol limit, we find that protocols that minimize dissipation at fixed duration are universally given by a two-step process, jumping to and from a point that balances jump size with fast relaxation. Jump protocols could be exploited by molecular machines or thermodynamic computing to improve energetic efficiency, and implemented in nonequilibrium free-energy estimation to improve accuracy. The birth of thermodynamics as a modern science can be traced to Sadi Carnot’s study of the design principles for energetically efficient heat engines described in _Reflections on the Motive Power of Fire_ Carnot et al. (1960). In classical thermodynamics, minimum-dissipation protocols are important in the design of macroscopic heat engines describing, for example, adiabatic (no heat loss) and quasistatic (infinitely slow) compression of gas by a piston. Nearly 200 years later, the field of stochastic thermodynamics Jarzynski (2011); Seifert (2012) similarly studies the design principles governing the ability to dynamically vary control parameters and perform work at minimum energetic cost (minimum dissipation), but now in micro- and nano-scale fluctuating systems. Minimum-dissipation protocols inform our understanding of the design principles of biological molecular machines Brown and Sivak (2017, 2019) and are of practical use to single-molecule experiments Tafoya et al. (2019), free-energy estimation Shenfeld et al. (2009); Geiger and Dellago (2010); Blaber and Sivak (2020a); Chodera et al. (2011); Minh and Chodera (2009), and thermodynamic computing Conte et al. (2019); Proesmans et al. (2020). In contrast to macroscopic systems that are well described by averages of thermodynamic quantities, the stochastic fluctuations in microscopic systems are large relative to the mean and cannot be ignored. The work done on a stochastic system by a control protocol is a fluctuating quantity that depends on the entire protocol history, making it particularly difficult to optimize. General optimization requires minimizing over all possible paths through control-parameter space, which cannot be solved in general Schmiedl and Seifert (2007). Despite advances relating optimal-control to optimal-transport theory, even numerical optimizations are still limited to relatively simple systems Aurell et al. (2011). Although general solutions are intractable, we have gained considerable insight into minimum-dissipation protocols by considering simple systems. For example, Schmiedl and Seifert Schmiedl and Seifert (2007) showed that for a Brownian particle diffusing in a harmonic potential with time-dependent minimum or stiffness, minimum-dissipation protocols exhibit jump discontinuities. It was posited that jumps in control parameters are a general feature of minimum-dissipation protocols, and they have since been observed in a number of different systems Gomez-Marin et al. (2008); Then and Engel (2008); Esposito et al. (2010). More general insight can be gained by approximating the mean work in relevant limits. For slow protocols, linear-response theory yields an approximation for the mean work, from which the approximate minimum-dissipation protocol can be calculated Sivak and Crooks (2012). Despite its success, the linear-response formalism relies on near-equilibrium approximations that break down for fast protocols, miss key features of the exact minimum-dissipation protocol (e.g., jumps in control parameters), and for short duration can perform worse than naive (constant-velocity) protocols Blaber and Sivak (2020b). While minimum-dissipation protocols for slowly driven systems are relatively well understood, comparatively little is known about rapidly driven systems. In this work we focus on fast protocols and find a universal design principle: the minimum-dissipation protocol consists of jumps at the beginning and end of the protocol, spending the entire duration at the control-parameter value that optimally balances the _initial force-relaxation rate_ (IFRR) (7b) with the jump size (13). Our results are physically intuitive, apply to a wide range of physical systems, and generalize easily to multidimensional control. To illustrate this, we calculate the minimum-dissipation protocols in a diverse set of systems described by Fokker-Planck or master-equation dynamics with single- (Fig. 1) or multi-dimensional control (Fig. 4). Combining our results with known minimum-dissipation protocols in the slow limit Sivak and Crooks (2012), we demonstrate that a simple interpolation scheme produces protocols that reduce dissipation at all speeds (Fig. 3). _Derivation_.—Consider a thermodynamic system described by dynamics of the form $\displaystyle\frac{\partial p_{\Lambda}(x,t)}{\partial t}=L[x,\boldsymbol{\lambda}(t)]\,p_{\Lambda}(x,t)\ ,$ (1) where $p_{\Lambda}(x,t)$ is the probability distribution over microstates $x$ at time $t$ given the control protocol $\Lambda$, and $L[x,\boldsymbol{\lambda}(t)]$ is the operator describing the system’s time evolution. $L$ is the Liouvillian for Hamiltonian, drift/diffusion operator for Fokker-Planck, and the transition-rate matrix for master-equation dynamics. The system is in contact with a heat bath at temperature $T$ such that the equilibrium probability distribution over microstates $x$ at fixed control parameters $\boldsymbol{\lambda}$ is $\pi(x|\boldsymbol{\lambda})\equiv\exp\\{\beta[F(\boldsymbol{\lambda})-U(x,\boldsymbol{\lambda})]\\}$, for internal energy $U(x,\boldsymbol{\lambda})$ and free energy $F(\boldsymbol{\lambda})\equiv-k_{\rm B}T\,\ln\sum_{x}\exp[-\beta U(x,\boldsymbol{\lambda})]$, where $\beta\equiv(k_{\rm B}T)$ for Boltzmann’s constant $k_{\rm B}$. The average excess work $W_{\rm ex}\equiv W-\Delta F$ by an external agent changing control parameters $\boldsymbol{\lambda}$ according to protocol $\Lambda$ is $\displaystyle\langle W_{\rm ex}\rangle_{\Lambda}=-\int_{0}^{\Delta t}\mathrm{d}t\,\frac{\mathrm{d}\boldsymbol{\lambda}^{\rm T}}{\mathrm{d}t}\langle\delta{\bf f}(t)\rangle_{\Lambda}\ ,$ (2) where a bold symbol denotes a column vector and superscript ${\rm T}$ the transpose. ${\bf f}\equiv-\partial U/\partial\boldsymbol{\lambda}$ are the forces conjugate to the control parameters, and $\delta{\bf f}\equiv{\bf f}-\langle{\bf f}\rangle_{\rm eq}$ the deviations from the equilibrium averages. Angle brackets $\langle\cdots\rangle_{\Lambda}$ denote a nonequilibrium ensemble average over the control-parameter protocol $\Lambda$. Here we hold fixed the initial ($\boldsymbol{\lambda}_{\rm i}$) and final ($\boldsymbol{\lambda}_{\rm f}$) control parameters, consistent with nonequilibrium free-energy estimation Ritort et al. (2002); Gore et al. (2003); Shenfeld et al. (2009); Geiger and Dellago (2010); Kim et al. (2012); Blaber and Sivak (2020b); Schindler et al. (2020); Aldeghi et al. (2018); Kuhn et al. (2017); Ciordia et al. (2016); Wang et al. (2015); Gapsys et al. (2015); Chodera et al. (2011); Minh and Chodera (2009) but distinct from optimizations that constrain the initial and final probability distributions Proesmans et al. (2020); Zhang (2020). If the total duration $\Delta t$ is short compared to the system’s natural relaxation time $\tau$ (a fast protocol), expanding the probability distribution in $\Delta t/\tau$ around an initial equilibrium distribution gives $\displaystyle p_{\Lambda}(x,s)=\pi(x|\boldsymbol{\lambda}_{\rm i})+p^{1}_{\Lambda}(x,s)\frac{\Delta t}{\tau}+\mathcal{O}\left[\left(\frac{\Delta t}{\tau}\right)^{2}\right]\ ,$ (3) for $s\equiv t/\Delta t$ and first-order coefficient $p^{1}_{\Lambda}(x,s)$. Plugging (3) up to $\mathcal{O}(\Delta t/\tau)$ into (1) gives $\displaystyle\frac{\partial p^{1}_{\Lambda}(x,s)}{\partial s}\approx\mathcal{L}[x,\boldsymbol{\lambda}(s)]\,\pi(x|\boldsymbol{\lambda}_{\rm i})\ ,$ (4) with $\mathcal{L}\equiv\tau L$ the dimensionless time-evolution operator. Solving for $p^{1}_{\Lambda}(x,s)$ and substituting into (3) yields $\displaystyle p_{\Lambda}(x,s)\approx\pi(x|\boldsymbol{\lambda}_{\rm i})+\frac{\Delta t}{\tau}\int_{0}^{s}\mathrm{d}s^{\prime}\mathcal{L}[x,\boldsymbol{\lambda}(s^{\prime})]\,\pi(x|\boldsymbol{\lambda}_{\rm i})\ .$ (5) Multiplying by conjugate forces ${\bf f}$, integrating over microstates $x$, and changing the time variable back to $t$ gives $\displaystyle\langle{\bf f}(t)\rangle_{\Lambda}\approx\langle{\bf f}\rangle_{\boldsymbol{\lambda}_{\rm i}}+\int_{0}^{t}\mathrm{d}t^{\prime}\,{\bf R}_{\boldsymbol{\lambda}_{\rm i}}[\boldsymbol{\lambda}(t^{\prime})]\ ,$ (6) for the _initial force-relaxation rate_ (IFRR) $\displaystyle{\bf R}_{\boldsymbol{\lambda}_{\rm i}}[\boldsymbol{\lambda}(t)]$ $\displaystyle\equiv\int\mathrm{d}x\ {\bf f}(x)\,L[x,\boldsymbol{\lambda}(t)]\,\pi(x|\boldsymbol{\lambda}_{\rm i})$ (7a) $\displaystyle=\frac{\mathrm{d}\langle{\bf f}\rangle_{\boldsymbol{\lambda}_{\rm i}}}{\mathrm{d}t}\bigg{|}_{\boldsymbol{\lambda}(t)}\ ,$ (7b) the rate of change of the conjugate forces at the current control-parameter values (averaged over the initial equilibrium distribution). Within this approximation, the average excess work is $\displaystyle\langle W_{\rm ex}\rangle_{\Lambda}\approx\langle W_{\rm ex}\rangle_{\boldsymbol{\lambda}_{\rm i}}-\int_{0}^{\Delta t}\mathrm{d}t\,\frac{\mathrm{d}\boldsymbol{\lambda}^{\rm T}}{\mathrm{d}t}\int_{0}^{t}\mathrm{d}t^{\prime}\,{\bf R}_{\boldsymbol{\lambda}_{\rm i}}[\boldsymbol{\lambda}(t^{\prime})]\ .$ (8) The first RHS term is the excess work during an instantaneous jump between the initial and final control-parameter values, which equals the _relative entropy_ $k_{\rm B}TD(\pi_{\rm i}||\pi_{\rm f})\equiv k_{\rm B}T\int\mathrm{d}x~{}\pi_{\rm i}\ln[\pi_{\rm i}/\pi_{\rm f}]$ between the initial $\pi_{\rm i}\equiv\pi(x|\boldsymbol{\lambda}_{\rm i})$ and final $\pi_{\rm f}\equiv\pi(x|\boldsymbol{\lambda}_{\rm f})$ equilibrium probability distributions Large and Sivak (2019). Integrating (8) by parts gives our main theoretical result: for sufficiently short duration, the excess work is $\displaystyle\langle W_{\rm ex}\rangle_{\Lambda}\approx k_{\rm B}TD(\pi_{\rm i}||\pi_{\rm f})-\int_{0}^{\Delta t}\mathrm{d}t\,{\bf R}_{\boldsymbol{\lambda}_{\rm i}}^{\rm T}[\boldsymbol{\lambda}(t)]\,[\boldsymbol{\lambda}_{{\rm f}}-\boldsymbol{\lambda}(t)]\ .$ (9) The second RHS term is the first-order correction in $\Delta t$, an approximation for the saved work $W_{\rm save}\equiv k_{\rm B}TD(\pi_{\rm i}||\pi_{\rm f})-W_{\rm ex}$ compared to an instantaneous protocol. _Initial Force-Relaxation Rate_.—The IFRR can be intuitively understood by considering one-dimensional exponential relaxation. For a discrete jump from initial control-parameter value $\lambda_{\rm i}$ to an intermediate value $\lambda$, an exponentially relaxing mean conjugate force obeys $\displaystyle\langle f(t)\rangle_{\Lambda}=\langle f\rangle_{\lambda_{\rm i}}+\left(\langle f\rangle_{\lambda}-\langle f\rangle_{\lambda_{\rm i}}\right)e^{-t/\tau(\lambda)}\ ,$ (10) where $\tau(\lambda)$ is the relaxation time of the conjugate force. The IFRR is the initial slope of the mean conjugate force as it relaxes towards equilibrium (7b), which for exponential relaxation is $\displaystyle{R}_{{\lambda}_{\rm i}}(\lambda)=\frac{\langle f\rangle_{\lambda_{\rm i}}-\langle f\rangle_{\lambda}}{\tau(\lambda)}\ .$ (11) Under simple exponential relaxation, $\tau(\lambda)$ is the same relaxation time defined in Ref. Sivak and Crooks, 2012 for slow protocols, thereby connecting short- and long-duration control. For a small control-parameter jump $\lambda-\lambda_{\rm i}$, static linear- response theory, $\langle f\rangle_{\lambda_{\rm i}}-\langle f\rangle_{\lambda}\approx\beta(\lambda-\lambda_{\rm i})\langle\delta f^{2}\rangle_{\lambda_{\rm i}}$, implies that the IFRR further simplifies to $\displaystyle{R}_{{\lambda}_{\rm i}}(\lambda)\approx\frac{\langle\delta f^{2}\rangle_{\lambda_{\rm i}}(\lambda-\lambda_{\rm i})}{\tau(\lambda)}\ .$ (12) The relaxation rate is zero at the initial control-parameter value and increases with larger control-parameter jumps which drive the system further from equilibrium. _Minimum-dissipation protocols_.—Equation (9) allows for relatively straightforward optimization to determine the _short-time efficient protocol_ (STEP), satisfying the Euler-Lagrange equation $\displaystyle\frac{\partial}{\partial\boldsymbol{\lambda}}\left[{\bf R}_{\boldsymbol{\lambda}_{\rm i}}^{\rm T}\left(\boldsymbol{\lambda}\right)\left(\boldsymbol{\lambda}_{{\rm f}}-\boldsymbol{\lambda}^{\rm STEP}\right)\right]\bigg{|}_{\boldsymbol{\lambda}^{\rm STEP}}={\bf R}_{\boldsymbol{\lambda}_{\rm i}}\left(\boldsymbol{\lambda}^{\rm STEP}\right)\ .$ (13) As an algebraic equation, the solution is a point in control-parameter space, thus the STEP consists of two jumps: a jump at the start from its initial value to the optimal value $\boldsymbol{\lambda}^{\rm STEP}$, and at the end from the optimal value to the final value. The STEP is a jump protocol provided the time-evolution operator $L$ is independent of time derivatives of the control parameters. For Fokker-Planck dynamics this is satisfied if the system is driven by a (generally time-dependent) scalar potential. To illustrate the two-step minimum-dissipation protocol we have calculated the STEP for diverse model systems (Fig. 1). In the translating- and breathing- trap systems described by Fokker-Planck dynamics (Supplemental Material I SM ), the STEP jumps halfway between the two endpoints, consistent with the results of Ref. Schmiedl and Seifert, 2007. The single-spin-flip and two-state binding/unbinding reaction systems are described by master-equation dynamics (Supplemental Material II and III SM ), with STEPs that jump to points that are respectively larger and smaller than halfway between the initial and final control-parameter values. Specific jump sizes for the STEP depend on the functional form of the IFRR, but the minimum-dissipation protocol always consists of jumps to and from an intermediate control-parameter value. Figure 1: Short-time efficient protocols (STEPs) for a single spin in a time- dependent magnetic field (blue dots), Brownian particle in a translating harmonic potential (black dashed), Brownian particle in a harmonic potential with time-dependent stiffness (black dashed), and a two-state binding/unbinding reaction system with variable binding and unbinding rates controlled by the chemical-potential difference (red dash-dotted). The STEP jumps to the point in control-parameter space that maximizes the _short-time power savings_ $\displaystyle P_{\rm save}^{\rm st}(\boldsymbol{\lambda})\equiv{\bf R}_{\boldsymbol{\lambda}_{\rm i}}^{\rm T}(\boldsymbol{\lambda})(\boldsymbol{\lambda}_{{\rm f}}-\boldsymbol{\lambda})$ (14) due to relaxation at intermediate $\boldsymbol{\lambda}$. The STEP balances fast relaxation rate ${\bf R}_{\boldsymbol{\lambda}_{\rm i}}$ with large final jump $\boldsymbol{\lambda}_{{\rm f}}-\boldsymbol{\lambda}$. The STEP spends the duration $\Delta t$ relaxing at $\boldsymbol{\lambda}^{\rm STEP}$, so for short duration $P_{\rm save}^{\rm st}(\boldsymbol{\lambda}^{\rm{STEP}})\Delta t$ is the work saved relative to an instantaneous protocol. To demonstrate the energetics of the STEP, consider the thermodynamic cycle consisting of tightening and loosening a harmonic trap (Fig. 2). For a quasistatic (infinitely slow) protocol, work equals the free-energy difference, which exactly cancels for a cyclic process. An instantaneous protocol has an additional contribution, which for tightening (loosening) the trap equals the relative entropy between the open (closed) and closed (open) states. The relative entropy is dissipated as heat during equilibration between tightening and loosening the trap (outer vertical arrows). The STEP spends the duration relaxing at an intermediate control-parameter value, resulting in saved work approximated by the area of the rectangle with width given by the final jump size $\lambda_{{\rm f}}-\lambda^{\rm STEP}$ and height by ${R}_{{\lambda}_{\rm i}}(\lambda^{\rm STEP})\Delta t$. To maximize the saved work (rectangle area) the STEP optimally balances the IFRR (determining the height) with final jump size (width). Figure 2: Thermodynamic cycle in the force vs. control parameter plane for the breathing harmonic trap driven by instantaneous (red dotted), STEP (green dash-dotted), and quasistatic (black dashed) protocols. Arrows denote protocol direction for transitions between open ($\lambda_{\rm i}$) and closed ($\lambda_{\rm f}$) states, shown schematically. The area under each curve gives the average work done by the respective protocol, the area under the quasistatic curve is the free-energy difference $\Delta F_{{\rm i}\rightarrow{\rm f}}=F_{\rm f}-F_{\rm i}$, the area between the instantaneous (dotted) and quasistatic (dashed) curves is the relative entropy (e.g., $\langle W\rangle_{\lambda_{\rm i}}\ -\Delta F_{{\rm i}\rightarrow{\rm f}}=k_{\rm B}TD(\pi_{\rm i}||\pi_{\rm f})$), and the area between the STEP (dash-dotted) and instantaneous (dotted) curves is the saved work $\langle W_{\rm save}\rangle$ (shaded rectangles). Control-parameter endpoints satisfy $\lambda_{\rm i}/\lambda_{\rm f}=1/2$, with duration $\Delta t/\tau=2/5$ for fastest relaxation time $\tau=1/(2\lambda_{\rm f})$. For a single control parameter, if the duration is sufficiently short the _gain_ $G_{\rm save}\equiv{\langle W_{\rm save}\rangle_{\Lambda}^{\rm des}}/{\langle W_{\rm save}\rangle_{\Lambda}^{\rm naive}}$ in saved work by the STEP is $\displaystyle G_{\rm save}^{\rm STEP}\approx\frac{\max_{\boldsymbol{\lambda}}[P_{\rm save}^{\rm st}(\lambda)]}{\overline{P_{\rm save}^{\rm st}(\lambda)}}\ ,$ (15) where an overbar denotes the spatial average $\overline{P_{\rm save}^{\rm st}(\lambda)}\equiv(\Delta\lambda)^{-1}\int_{\lambda_{\rm i}}^{\lambda_{\rm f}}\mathrm{d}\lambda~{}P_{\rm save}^{\rm st}(\lambda)$, “naive” the constant- velocity protocol, and “des” a designed protocol. The gain from a STEP is greatest when the power savings $P_{\rm save}^{\rm st}(\lambda)$ is sharply peaked. _Interpolated protocols_.—In order to design a protocol that performs well for any duration, we combine the STEP (optimal for fast protocols) with the minimum-dissipation protocol from linear-response theory (optimal for slow protocols). The linear-response protocol is continuous and when driven by a single control parameter proceeds at velocity $\mathrm{d}\lambda/\mathrm{d}t\propto[\zeta(\lambda)]^{-1/2}$, where $\zeta(\lambda)$ is the generalized friction coefficient Sivak and Crooks (2012). We assume the shape of the protocol from linear-response theory remains unchanged (i.e., $\mathrm{d}\lambda/\mathrm{d}t\propto[\zeta(\lambda)]^{-1/2}$) but with initial jump $(\lambda^{\rm STEP}-\lambda_{\rm i})/(1+\Delta t/\tau)^{\alpha}$ and final jump $(\lambda_{\rm f}-\lambda^{\rm STEP})/(1+\Delta t/\tau)^{\alpha}$, where the constant $\alpha$ controls the crossover from slow to fast approximations. For our simple systems we empirically find $\alpha=1$ performs relatively well. Figure 3 shows the improvement from designed protocols relative to naive (constant-velocity) for the model system of a breathing harmonic trap. The difference between naive and designed work (Fig. 3a) shows the expected asymptotic behavior in the short- and long-duration limits: scaling as $\Delta t$ (slope of one) for small $\Delta t/\tau$ and $(\Delta t)^{-1}$ (slope of negative one) for large $\Delta t/\tau$. Both the fast and slow designed protocols perform worse than naive (the difference is negative) for large and small $\Delta t/\tau$, respectively. The fast-protocol approximation (9) breaks down for long duration because the conjugate-force relaxation rate decreases as the system approaches equilibrium at $\boldsymbol{\lambda}$, whereas (9) assumes a constant relaxation time. However, the interpolated protocol performs well for any duration, and the largest work saved compared to naive is for intermediate duration. The gain $G_{\rm save}$ quantifies the percent increase in saved work from designed protocols relative to naive, where a gain greater than one indicates the designed does less work than the naive. For this system, the largest gain in saved work occurs in the fast limit (small $\Delta t/\tau$) for the STEP, interpolated, and exact optimal protocols. Figure 3: Benefit in the breathing harmonic trap from designed protocols relative to the naive (constant-velocity) protocol, as a function of the duration $\Delta t$ scaled by the fastest integral relaxation time $\tau$. The different designed (“des”) protocols include the exact optimal Schmiedl and Seifert (2007) (“opt”, solid black), linear-response optimized (“slow opt”, dashed blue), STEP (“fast opt”, red dots), and interpolated optimal protocol (“inter opt”, dash-dotted green). (a) Difference between the work done by the naive (constant-velocity) and designed protocols. (b) Gain $G_{\rm save}\equiv{\langle W_{\rm save}\rangle_{\Lambda}^{\rm des}}/{\langle W_{\rm save}\rangle_{\Lambda}^{\rm naive}}$ in saved work. Solid red line in (b) denotes the short-duration limit (15). Control-parameter endpoints satisfy $\lambda_{\rm i}/\lambda_{\rm f}=16$, and the interpolated protocol uses $\alpha=1$ and fastest integral relaxation time $\tau=1/(2\lambda_{\rm i})$ Blaber and Sivak (2020b). _Multi-dimensional control_.—Optimization of multi-dimensional control protocols has seen a recent surge in interest, primarily driven by possible improvements to nonequilibrium free-energy estimates in fast-switching simulations Chipot and Lelièvre (2011); Dellago and Hummer (2014). Previous calculations of minimum-dissipation protocols which observed jumps were limited to one-dimensional optimization. A significant advantage of the present approximation is that it permits simple multidimensional control- protocol optimization. Equation (13) implies that for multidimensional control the STEP consists of jumps to and from the control-parameter point $\boldsymbol{\lambda}^{\rm STEP}$. To illustrate, we consider a nine-spin Ising model with frustrated boundary conditions (Fig. 4) Rotskoff et al. (2017); Venturoli et al. (2009). We use a 2D control parameter $\boldsymbol{h}=(h_{\rm b},h_{\rm g})$ of magnetic fields applied to the mid-edge spins (Fig. 4a) which initially hold the system in the spin-down state and reverse during the protocol, driving the system to invert its magnetization. Supplemental Material IV SM gives model details. Figure 4: a) Nine-spin ferromagnetic Ising model (internal black spins) with fixed boundary conditions (external gray spins). The multi-dimensional control parameter is two external magnetic fields, $h_{\rm b}$ (blue) applied to horizontal-edge spins and $h_{\rm g}$ (green) applied to vertical-edge spins. b) Short-time power savings (14) as function of control parameters ($h_{\rm b}$,$h_{\rm g}$). Red line: naive protocol; red star: $\boldsymbol{h}^{\rm{STEP}}$ (13). c) Work difference between designed and naive protocols (dotted red), and its short-duration approximation (9) (solid red). d) Gain $G_{\rm save}\equiv{\langle W_{\rm save}\rangle_{\Lambda}^{\rm des}}/{\langle W_{\rm save}\rangle_{\Lambda}^{\rm naive}}$ in saved work for multi-dimensional STEP relative to naive (dotted red), and its short-duration limit (15) (solid red). Control-parameter endpoints are $\boldsymbol{h}_{\rm i}=(-2,-2)$ and $\boldsymbol{h}_{\rm f}=(2,2)$, with duration $\Delta t$ and fastest relaxation time $\tau=N/k_{0}$, for $N=9$ spins and single-spin flip attempt rate $k_{0}$. The power saving (14) vanishes at initial and final control-parameter values, respectively corresponding to zero relaxation rate and zero final jump size (Fig. 4b). By jumping past control-parameter regions with small power saving, the STEP outperforms the naive protocol for short duration, as quantified by the difference between naive and designed work (Fig. 4c) and the gain in saved work (Fig. 4d). Indeed, for short duration the STEP more than doubles the work saved by the naive protocol (i.e., has gain greater than two). _Discussion_.—We have developed an approximation for work in the fast-protocol limit (9) that permits straightforward optimization (13) simply from the initial force-relaxation rate (IFRR), Eq. (7b). We find that jumps are a universal feature of minimum-dissipation protocols in this fast limit, which we illustrate with several model systems under single- (Fig. 1) or multi- dimensional control (Fig. 4). Jumps minimize dissipation for fast protocols because the relaxation rate is approximately constant, with no diminishing returns from spending the entire duration at a single control-parameter value. Therefore, the STEP jumps between the fixed control-parameter endpoints to spend the entire duration at the control-parameter value that maximizes the product of the IFRR and the subsequent jump size (14). This breaks down for slow protocols since with sufficient time at a single control-parameter value, the relaxation rate decreases over time; indeed, in the slow limit the minimum-dissipation protocol is continuous Sivak and Crooks (2012). We combine these two seemingly disparate limits with a simple interpolation scheme, producing protocols that perform well for any duration (Fig. 3). One important application of minimum-dissipation protocols is to free-energy estimation, which aids the design of novel ligands for targeted protein binding Schindler et al. (2020); Aldeghi et al. (2018); Kuhn et al. (2017); Ciordia et al. (2016); Wang et al. (2015); Gapsys et al. (2015); Chodera et al. (2011). Quite generally, the accuracy of free-energy estimates decreases with increasing dissipation Ritort et al. (2002); Gore et al. (2003); Shenfeld et al. (2009); Geiger and Dellago (2010); Kim et al. (2012); Blaber and Sivak (2020b). Based on the results of Ref. Schmiedl and Seifert, 2007, jump protocols have been used to reduce dissipation and improve free-energy estimates Geiger and Dellago (2010), but previously it was unknown whether jumps would always reduce dissipation in these more complex systems, and there was no simple procedure to find the optimal jump size. The present formalism demonstrates that jumps are a general feature and gives a simple optimization procedure applicable to multidimensional control. This makes protocol optimization tractable for a considerably expanded range of systems. Although we focused on systems with known equations of motion, the IFRR (7b) and short-time power savings (14) are easily estimated without detailed dynamical information: the system only needs to be equilibrated at a single control-parameter value; the protocol can be very short; the average converges with few samples; and the STEP is found using standard optimization techniques applied to (14). The STEP can thus be computed relatively inexpensively, easing determination of minimum-dissipation protocols in rapidly driven complex chemical and biological systems. This opens the door to improve the energetic efficiency in thermodynamic computing Conte et al. (2019); Proesmans et al. (2020) and the accuracy of nonequilibrium free-energy estimates in simulations and single-molecule experiments Tafoya et al. (2019); Blaber and Sivak (2020b); Shenfeld et al. (2009); Ritort et al. (2002). This work is supported by an SFU Graduate Deans Entrance Scholarship (SB), an NSERC CGS Doctoral fellowship (MDL), an NSERC Discovery Grant and Discovery Accelerator Supplement (DAS), and a Tier-II Canada Research Chair (DAS), and was enabled in part by support provided by WestGrid (www.westgrid.ca) and Compute Canada Calcul Canada (www.computecanada.ca). The authors thank John Bechhoefer, Jannik Ehrich, and Joseph Lucero (SFU Physics) for enlightening feedback on the manuscript. ## References * Carnot et al. (1960) S. Carnot, E. Clapeyron, and R. Clausius, in _Reflections on the Motive Power of Fire: And Others Papers on the Second Law of Thermodynamics_ , edited by E. Mendoza (New York: Dover Publications, 1960). * Jarzynski (2011) C. Jarzynski, Annu. Rev. Condens. Matter Phys. 2, 329 (2011). * Seifert (2012) U. Seifert, Rep. Prog. Phys. 75, 126001 (2012). * Brown and Sivak (2017) A. I. Brown and D. A. Sivak, Physics in Canada 73 (2017). * Brown and Sivak (2019) A. I. Brown and D. A. Sivak, Chem. Rev. 120, 434 (2019). * Tafoya et al. (2019) S. Tafoya, S. J. Large, S. Liu, C. Bustamante, and D. A. Sivak, Proc. Natl. Acad. Sci. U.S.A 116, 5920 (2019). * Shenfeld et al. (2009) D. K. Shenfeld, H. Xu, M. P. Eastwood, R. O. Dror, and D. E. Shaw, Phys. Rev. E 80, 046705 (2009). * Geiger and Dellago (2010) P. Geiger and C. Dellago, Phys. Rev. E 81, 021127 (2010). * Blaber and Sivak (2020a) S. Blaber and D. A. Sivak, Phys. Rev. E 101, 022118 (2020a). * Chodera et al. (2011) J. D. Chodera, D. L. Mobley, M. R. Shirts, R. W. Dixon, K. Branson, and V. S. Pande, Curr. Opin. Struc. Biol. 21, 150 (2011). * Minh and Chodera (2009) D. D. Minh and J. D. Chodera, J. Chem. Phys. 131, 134110 (2009). * Conte et al. (2019) T. Conte, E. DeBenedictis, N. Ganesh, T. Hylton, J. P. Strachan, R. S. Williams, A. Alemi, L. Altenberg, G. Crooks, J. Crutchfield, et al., arXiv preprint arXiv:1911.01968 (2019). * Proesmans et al. (2020) K. Proesmans, J. Ehrich, and J. Bechhoefer, Phys. Rev. Lett. 125, 100602 (2020). * Schmiedl and Seifert (2007) T. Schmiedl and U. Seifert, Phys. Rev. Lett. 98, 108301 (2007). * Aurell et al. (2011) E. Aurell, C. Mejía-Monasterio, and P. Muratore-Ginanneschi, Phys. Rev. Lett. 106, 250601 (2011). * Gomez-Marin et al. (2008) A. Gomez-Marin, T. Schmiedl, and U. Seifert, J. Chem. Phys. 129, 024114 (2008). * Then and Engel (2008) H. Then and A. Engel, Phys. Rev. E 77, 041105 (2008). * Esposito et al. (2010) M. Esposito, R. Kawai, K. Lindenberg, and C. Van den Broeck, EPL 89, 20003 (2010). * Sivak and Crooks (2012) D. A. Sivak and G. E. Crooks, Phys. Rev. Lett. 108, 190602 (2012). * Blaber and Sivak (2020b) S. Blaber and D. A. Sivak, J. Chem. Phys. 153, 244119 (2020b). * Ritort et al. (2002) F. Ritort, C. Bustamante, and I. Tinoco, Proc. Natl. Acad. Sci. U.S.A 99, 13544 (2002). * Gore et al. (2003) J. Gore, F. Ritort, and C. Bustamante, Proc. Natl. Acad. Sci. U.S.A 100, 12564 (2003). * Kim et al. (2012) S. Kim, Y. W. Kim, P. Talkner, and J. Yi, Phys. Rev. E 86, 041130 (2012). * Schindler et al. (2020) C. E. Schindler, H. Baumann, A. Blum, D. Böse, H.-P. Buchstaller, L. Burgdorf, D. Cappel, E. Chekler, P. Czodrowski, D. Dorsch, et al., J. Chem. Inf. Model (2020). * Aldeghi et al. (2018) M. Aldeghi, V. Gapsys, and B. L. de Groot, ACS Cent. Sci. 4, 1708 (2018). * Kuhn et al. (2017) B. Kuhn, M. Tichý, L. Wang, S. Robinson, R. E. Martin, A. Kuglstatter, J. Benz, M. Giroud, T. Schirmeister, R. Abel, et al., J. Med. Chem 60, 2485 (2017). * Ciordia et al. (2016) M. Ciordia, L. Pérez-Benito, F. Delgado, A. A. Trabanco, and G. Tresadern, J. Chem. Inf. Model 56, 1856 (2016). * Wang et al. (2015) L. Wang, Y. Wu, Y. Deng, B. Kim, L. Pierce, G. Krilov, D. Lupyan, S. Robinson, M. K. Dahlgren, J. Greenwood, et al., J. Am. Chem. Soc. 137, 2695 (2015). * Gapsys et al. (2015) V. Gapsys, S. Michielssens, J. H. Peters, B. L. de Groot, and H. Leonov, in _Molecular Modeling of Proteins_ , edited by A. Kukol (Springer New York, New York, NY, 2015), pp. 173–209. * Zhang (2020) Y. Zhang, EPL 128, 30002 (2020). * Large and Sivak (2019) S. J. Large and D. A. Sivak, J. Stat. Mech.: Theory Exp. 2019, 083212 (2019). * (32) See Supplemental Material at [URL will be inserted by publisher] for Model Details. * Chipot and Lelièvre (2011) C. Chipot and T. Lelièvre, SIAM J. Appl. Math 71, 1673 (2011). * Dellago and Hummer (2014) C. Dellago and G. Hummer, Entropy 16, 41 (2014). * Rotskoff et al. (2017) G. M. Rotskoff, G. E. Crooks, and E. Vanden-Eijnden, Phys. Rev. E 95, 012148 (2017). * Venturoli et al. (2009) M. Venturoli, E. Vanden-Eijnden, and G. Ciccotti, J Math Chem 45, 188 (2009). * Glauber (1963) R. J. Glauber, J. Math. Phys. 4, 294 (1963).
Zhang et al. An Efficient Stochastic Augmented Lagrangian-Type Algorithm Solving Stochastic Optimization with Expectation Constraints Efficiently by a Stochastic Augmented Lagrangian-Type Algorithm Liwei Zhang School of Mathematical Sciences, Dalian University of Technology, 116023 Dalian, China<EMAIL_ADDRESS>Yule Zhang School of Science, Dalian Martime University, 116085 Dalian, China<EMAIL_ADDRESS> Jia Wu School of Mathematical Sciences, Dalian University of Technology, 116023 Dalian, China<EMAIL_ADDRESS> Xiantao Xiao School of Mathematical Sciences, Dalian University of Technology, 116023 Dalian, China<EMAIL_ADDRESS> This paper considers the problem of minimizing a convex expectation function with a set of inequality convex expectation constraints. We propose a stochastic augmented Lagrangian-type algorithm, namely the stochastic linearized proximal method of multipliers, to solve this convex stochastic optimization problem. This algorithm can be roughly viewed as a hybrid of stochastic approximation and the traditional proximal method of multipliers. Under mild conditions, we show that this algorithm exhibits $O(K^{-1/2})$ expected convergence rates for both objective reduction and constraint violation if parameters in the algorithm are properly chosen, where $K$ denotes the number of iterations. Moreover, we show that, with high probability, the algorithm has $O(\log(K)K^{-1/2})$ constraint violation bound and $O(\log^{3/2}(K)K^{-1/2})$ objective bound. Numerical results demonstrate that the proposed algorithm is efficient. stochastic approximation; linearized proximal method of multipliers; expectation constrained stochastic program; expected convergence rate; high probability bound § INTRODUCTION In this paper, we consider the following stochastic optimization problem \begin{equation}\label{eq:1} \begin{array}{rl} \min\limits_{x \in \cC} & f(x):=\mathbb{E}[F(x,\xi)]\\[4pt] {\rm s.t.} & g_i(x):=\mathbb{E}[G_i(x,\xi)] \leq 0,\ i=1,\ldots,p.\\ \end{array} \end{equation} Here, $\cC \subset \R^n$ is a nonempty bounded closed convex set, $\xi$ is a random vector whose probability distribution is supported on $\Xi \subseteq \R^q$, $F: \cC \times \Xi \rightarrow \R$ and $G_i:\cC \times \Xi \rightarrow \R$, $i=1,\ldots, p$. Let $\Phi$ be the feasible set of problem (<ref>) as \begin{equation}\label{eq:Phi} \Phi:=\left\{x\in \cC: g_i(x) \leq 0,\ i=1,\ldots,p\right\}. \end{equation} We assume that \mathbb{E}[F(x,\xi)]= \int_{\Xi} F(x,\xi)dP(\xi),\, \mathbb{E}[G_i(x,\xi)]= \int_{\Xi} G_i(x,\xi)dP(\xi),\ i=1,\ldots, p are well defined and finite valued for every $x\in \cC$. Moreover, we assume that the functions $F(\cdot,\xi)$ and $G_i(\cdot,\xi)$ are continuous and convex on $\cC$ for almost every $\xi$. Hence, the expectation functions $f(\cdot)$ and $g_i(\cdot,\xi)$ are continuous and convex on $\cC$. Problems in the form of (<ref>) are standard in stochastic programming <cit.> and also arise frequently in many practical applications such as machine learning <cit.> and finance <cit.>. A computational difficulty of solving (<ref>) is that expectation is a multidimensional integral and it cannot be computed with a high accuracy for large dimension $q$. In order to handle this issue, a popular approach is to use stochastic approximation (SA) technique which is based on the following general assumptions: (i) it is possible to generate i.i.d. sample $\xi^1,\xi^2,\ldots,$ of realizations of random vector $\xi$; (ii) there is an oracle, which, for any point $(x,\xi)\in \cC \times \Xi$ returns stochastic subgradients $v_0(x,\xi),\ v_1(x,\xi),\ \ldots,\ v_p(x,\xi)$ of $F(x,\xi)$, $G_1(x,\xi),\ \ldots,\ G_p(x,\xi)$ such that $ v_i(x)=\mathbb{E}[v_i(x,\xi)],\ i=0,1,\ldots,p $ are well defined and are subgradients of $f(\cdot)$, $g_1(\cdot)$, $\ldots$, $g_p(\cdot)$ at $x$, respectively, i.e., $v_0(x)\in \partial f(x)$, $v_i(x) \in \partial g_i(x)$, $i=1,\ldots,p$. Since the pioneering paper <cit.>, due to low demand for computer memory and cheap computation cost at every iteration, SA type algorithms become widely used in stochastic optimization and machine learning, see, e.g. <cit.>. If $f(\cdot)$ is twice continuously differentiable and strongly convex, in the classical analysis it is shown that the SA algorithm exhibits asymptotically optimal rate of convergence $\mathbb{E}[f(x^k)-f^*]=O(k^{-1})$, where $x^k$ is $k$th iterate and $f^*$ is the optimal value. An important improvement developed by <cit.> and <cit.> suggests that, larger stepsizes of SA algorithm can be adopted by consequently averaging the obtained iterates. Moreover, <cit.> show that, without assuming smoothness and strong convexity, a properly modified SA method achieves the convergence rate $O(k^{-1/2})$ and remarkably outperforms the sample average approximation (SAA) approach for a certain class of convex stochastic problems. After the seminal work <cit.>, there are many significant results appeared, even for nonconvex stochastic optimization problems, see <cit.> and references cited therein. Among all mentioned works, the feasible set is an abstract closed convex set and none of these SA algorithms are applicable to expectation constrained problems. The main reason is that the computation of projection $\Pi_{\Phi}$ is quite easy only when $\Phi$ is of a simple structure. However, when $\Phi$ is defined by (<ref>), the computation is prohibitive. As a first attempt for solving expectation constrained stochastic optimization problems with stochastic approximation technique, <cit.> introduce a cooperative stochastic approximation (CSA) algorithm for solving (<ref>) with single expectation constraint ($p=1$), which is a stochastic counterpart of Polyak's subgradient method <cit.>. The authors show that CSA exhibits the optimal $O(1/\sqrt{K})$ rate of expected convergence, where $K$ is a fixed iteration number. In an online fashion, <cit.> propose an algorithm (simply denoted by “YNW”) that can be easily extended to solve (<ref>) with multiple expectation constraints. Under the Slater's condition and the assumption that $\cC$ is compact, they show that the algorithm can achieve $O(1/\sqrt{K})$ expected regret and $O(\log(K)/\sqrt{K})$ high probability regret. <cit.> develops a penalized stochastic gradient (PSG) method and establishes its almost sure convergence and expected convergence rates. PSG can be roughly viewed as a hybrid of the classical penalty method for nonlinear programming and the stochastic quasi-gradient method <cit.> for stochastic composition problem. A stochastic level-set method <cit.>, which ensures a feasible solution path with high probability, is proposed and analyzed. <cit.> propose a conservative stochastic optimization algorithm (CSOA), which is in the similar primal-dual framework as PSG and YNW. In addition to CSOA, the authors also propose a projection-free algorithm named as FW-CSOA which can deal with the case that the projection $\Pi_{\cC}$ is difficult to calculate. <cit.> study an adaptive primal-dual stochastic gradient method (APriD) for solving (<ref>) and establish the convergence rate of $O(1/\sqrt{K})$ in terms of the objective error and the constraint violation. All of the above mentioned methods for solving (<ref>) can be cast into the family of stochastic first-order algorithms. Although the iteration in stochastic first-order algorithms is computationally cheap and these methods perform well for certain problems, there are plenty of practical experiences and evidences of their convergence difficulties and instability with respect to the choice of parameters. the success of augmented Lagrangian methods for various kinds of functional constrained optimization problems is witnessed. <cit.> study an augmented Lagrangian method for multistage stochastic problems. For solving semidefinite programming (SDP) problems, <cit.> consider an Newton-CG augmented Lagrangian method, which is shown to be very efficient even for large-scale SDP problems. <cit.> propose several methods based on augmented Lagrangian framework for optimization problems with stochastic-order constraints and analyze their convergence. <cit.> study an augmented Lagrangian decomposition method for nonconvex chance-constrained problems, in which a convex subproblem and a 0-1 knapsack subproblem are solved at each iteration. The aim of this paper is to develop an efficient stochastic approximation-based augmented Lagrangian-type method for solving (<ref>). To the best of our knowledge, this is still limited in the literature. <cit.> propose a stochastic proximal method of multipliers (PMMSopt) for solving problem (<ref>) and show that PMMSopt exhibits $O(K^{-1/2})$ convergence rates for both objective reduction and constraint violation. PMMSopt is partially inspired by the classic proximal method of multipliers <cit.>, which is modeled through an augmented Lagrangian with an extra proximal term. However, the subproblem is difficult to solve, that makes PMMSopt an unimplementable algorithm, and hence no numerical results are given. In this paper, based on PMMSopt, we propose a stochastic linearized proximal method of multipliers (SLPMM) for solving the stochastic convex optimization problem (<ref>), and analyze its expected convergence rate as well as probability guarantee for both objective reduction and constraint violation. In specific, at the $k$th iteration in SLPMM, we consider the augmented Lagrangian function $\cL_\sigma^k(x,\lambda)$ of a linearized problem with respect to the stochastic subgradients $v_i(x^k,\xi^k)$, $i=0,1,\ldots,p$. Then, we obtain the next iterate $x^{k+1}$ by solving the problem $\min_{x\in\cC}\cL_\sigma^k(x,\lambda^k)+\frac{\alpha}{2}\|x-x^k\|^2$ and update the Lagrange multiplier. The subproblem is the minimization of a strongly convex (approximately) quadratic function and hence is relatively easy to solve. Assuming that the set $\cC$ is compact, the subgradients are bounded and the Slater's condition holds, if the parameters in SLPMM are chosen as $\alpha=\sqrt{K}$ and $\sigma=1/\sqrt{K}$, we show that SLPMM attains $O(1/\sqrt{K})$ expected convergence rate with respect to both objective reduction and constraint violation. Under certain light-tail assumptions, we also establish the large-deviation properties of SLPMM. The numerical results on some practical applications such as Neyman-Pearson classification demonstrate that SLPMM performs efficiently and has certain advantages over the existing stochastic first-order methods. The remaining parts of this paper are organized as follows. In Section <ref>, we develop some important properties of SLPMM. In Section <ref>, in the expectation sense we establish the convergence rate of SLPMM for problem (<ref>). The high probability guarantees for objective reduction and constraint violation of SLPMM are investigated in Section <ref>. In Section <ref>, we report our numerical results. Finally, we draw a conclusion in Section <ref>. § ALGORITHMIC FRAMEWORK, ASSUMPTIONS AND AUXILIARY LEMMAS In this section, we propose a stochastic linearized proximal method of multipliers (SLPMM) for solving problem (<ref>) and establish some important auxiliary lemmas. Let us define $[t]_+:=\max\{t,0\}$ for any $t\in\R$ and let $[y]_+=\Pi_{\R^p_+}[y]$ denote the projection of $y$ onto $\R^p_+$ for any $y \in \R^p$. We also define $[t]_+^2:=(\max\{t,0\})^2$. The detail of SLPMM is described in Algorithm <ref>. A stochastic linearized proximal method of multipliers alg:SA11Initialization:  Choose an initial point $x^0 \in \cC$ and select parameters $\sigma>0,\alpha>0$. Set $\lambda^0=0\in\R^p$ and $k=0$. alg:SA33Generate i.i.d. sample $\xi^k$ of $\xi$ and compute x^k+1= _x ∈ { ^k_σ(x,λ^k) +α/2x-x^k^2}, ^k_σ(x,λ) :=F(x^k,ξ^k)+⟨v_0(x^k,ξ^k),x-x^k ⟩ + 1/2σ[ ∑_i=1^p[λ_i+σ(G_i(x^k,ξ^k)+ ⟨v_i(x^k,ξ^k),x-x^k ⟩)]_+^2-λ^2] and $v_i(x^k,\xi^k)$, $i=0,1,\ldots,p$ are the corresponding stochastic subgradients. alg:SA44Update the Lagrange multipliers by λ_i^k+1=[λ_i^k+σ(G_i(x^k,ξ^k)+ ⟨v_i(x^k,ξ^k),x^k+1-x^k ⟩)]_+, i=1,…,p. alg:SA55Set $k\leftarrow k+1$. In specific, at each iteration, we first generate an i.i.d. sample $\xi^k$ and choose the stochastic subgradients $v_i(x^k,\xi^k)$, $i=0,1,\ldots,p$ of $F$ and $G_i$, respectively. Then, in (<ref>) we obtain $x^{k+1}$ by computing the proximal point of $\cL^k_{\sigma}(x,\lambda)$, which is the augmented Lagrangian function of the linearized problem \[ \begin{array}{ll} \min\limits_{x \in \cC} & F(x^k,\xi^k)+\langle v_0(x^k,\xi^k),x-x^k \rangle\\[4pt] {\rm s.t.} & G_i(x^k,\xi^k)+ \langle v_i(x^k,\xi^k),x-x^k \rangle \leq 0,\ i=1,\ldots,p.\\ \end{array} \] Finally, in (<ref>) we update the Lagrange multipliers. \[ G(x,\xi):=(G_1(x,\xi),\ldots, G_p(x,\xi))^T,\quad g(x):=(g_1(x),\ldots, g_p(x))^T. \] \[ V(x^k,\xi^k):=(v_1(x^k,\xi^k),\ldots, v_p(x^k,\xi^k))^T, \] then (<ref>) can be rewritten as In the following, we shall study the convergence of the stochastic process $\{x^k,\lambda^k\}$ generated by SLPMM with respect to the filtration $\cF_k$ (sigma-algebra) which is generated by the random information $\{(\xi^0,\ldots,\xi^{k-1})\}$. Before that, we introduce the following assumptions. Let $R>0$ be a positive parameter such that \|x'-x''\|\leq R,\ \forall x',x'' \in \cC. There exists a constant $\nu_g>0$ such that for each $\xi^k$, \|G(x,\xi^k)\| \leq \nu_g,\ \forall x \in \cC. There exist constants $\kappa_f>0$ and $\kappa_g>0$ such that for each $\xi^k$, \|v_0(x,\xi^k)\| \leq \kappa_f, \,\, \|v_i(x,\xi^k)\| \leq \kappa_g,\ i=1,\ldots, p,\ \forall x\in \cC. The Slater's condition holds, i.e., there exist $\varepsilon_0>0$ and $\widehat x \in \cC$ such that g_i(\widehat x) \leq -\varepsilon_0, \,\, i =1,\ldots, p. Assumption <ref> shows that $\cC$ is a compact convex set with diameter $R$. Assumption <ref> indicates that the constraint functions $G_i(\cdot,\xi^k)$ are bounded over $\cC$. This assumption is a bit restrictive, but it is required in the analysis of almost all existing stochastic methods for solving problem (<ref>) <cit.>. Assumption <ref> requires that the stochastic subgradients $v_i(\cdot,\xi^k)$ are bounded over $\cC$. Assumption <ref> is a standard Slater's condition for optimization problem with functional constraints. The following auxiliary lemma will be used several times in the sequel. For any $z\in \cC$, we have \begin{equation}\label{eq:opt-x-1} \begin{array}{ll} \displaystyle\langle v_0(x^k,\xi^k),x^{k+1}-x^k \rangle + \frac{1}{2\sigma}\|\lambda^{k+1}\|^2 + \frac{\alpha}{2} \|x^{k+1}-x^k\|^2 \\[5pt] \leq \displaystyle\langle v_0(x^k,\xi^k),z-x^k \rangle + \frac{1}{2\sigma}\left[ \sum_{i=1}^p[\lambda^k_i+\sigma (G_i(x^k,\xi^k)+\langle v_i(x^k,\xi^k), z-x^k \rangle)]_+^2\right]\\[15pt] \quad\quad+ \displaystyle\frac{\alpha}{2}(\|z-x^k\|^2-\|z-x^{k+1}\|^2). \end{array} \end{equation} In particular, if we take $z=x^k$, it yields \begin{equation}\label{eq:opt-x-2} \begin{array}{ll} \displaystyle\langle v_0(x^k,\xi^k),x^{k+1}-x^k \rangle + \frac{1}{2\sigma}\|\lambda^{k+1}\|^2 + \alpha \|x^{k+1}-x^k\|^2\\[10pt] \leq \displaystyle\frac{1}{2\sigma}\left[ \sum_{i=1}^p[\lambda^k_i+\sigma G_i(x^k,\xi^k)]_+^2\right]. \end{array} \end{equation} By using the optimality conditions, we have from (<ref>) that $x^{k+1}$ satisfies \begin{equation}\label{eq:aux-opt} 0\in \nabla_x\cL^k_{\sigma}(x^{k+1},\lambda^k)+\alpha (x^{k+1}-x^k)+\cN_{\cC}(x^{k+1}), \end{equation} where $\cN_{\cC}(x^{k+1})$ denotes the normal cone of $\cC$ at $x^{k+1}$ and \[ \nabla_x\cL^k_{\sigma}(x^{k+1},\lambda^k)=v_0(x^k,\xi^k)+\sum_{i=1}^pv_i(x^k,\xi^k)\cdot[\lambda_i^k+\sigma (G_i(x^k,\xi^k)+ \langle v_i(x^k,\xi^k),x^{k+1}-x^k \rangle)]_+. \] Let us now consider the following auxiliary problem \begin{equation}\label{eq:auxP} \begin{array}{ll} \min\limits_{x \in \cC} \,\langle v_0(x^k,\xi^k),x-x^k \rangle+ \frac{1}{2\sigma}\left[ \displaystyle\sum_{i=1}^p[\lambda_i^k+\sigma (G_i(x^k,\xi^k)+ \langle v_i(x^k,\xi^k),x-x^k \rangle)]_+^2\right]\\[15pt] \quad\quad\quad +\frac{\alpha}{2}(\|x-x^k\|^2-\|x-x^{k+1}\|^2). \end{array} \ee We can easily check that (\ref{eq:auxP}) is a convex optimization problem. Therefore, $\hat{x}$ is an optimal solution to (\ref{eq:auxP}) if and only if \[ \begin{array}{ll} 0\in &v_0(x^k,\xi^k)+\sum_{i=1}^pv_i(x^k,\xi^k)\cdot[\lambda_i^k+\sigma (G_i(x^k,\xi^k)+ \langle v_i(x^k,\xi^k),\hat{x}-x^k \rangle)]_+\\[6pt]&\quad+\alpha (x^{k+1}-x^k)+\cN_{\cC}(\hat{x}). \end{array} \] Hence, if follows from (\ref{eq:aux-opt}) that $x^{k+1}$ is an optimal solution to (\ref{eq:auxP}), which gives (\ref{eq:opt-x-1}) and (\ref{eq:opt-x-2}) obviously. \Halmos\endproof In what follows, we estimate an upper bound of $\|x^{k+1}-x^k\|$. \begin{lemma}\label{lem:aux3} Let Assumptions \ref{assu:compact}-\ref{assu:moment} be satisfied. Then, if the parameters satisfy $2\alpha-p\kappa_g^2\sigma>0$, we have \[ \|x^{k+1}-x^k\|\leq \frac{1}{\alpha }(\kappa_f+ \sqrt{p}\kappa_g\|\lambda^k\|+ \sqrt{p}\nu_g\kappa_g\sigma). \] \end{lemma} %\comm{Correct. Assume that $\alpha-p\kappa_g^2\sigma>0$, then we have %\|x^{k+1}-x^k\|\leq \frac{2}{\alpha %}(\kappa_f+ \sqrt{p}\kappa_g\|\lambda^k\|+ \proof{Proof.} From (\ref{eq:opt-x-2}) and Assumption \ref{assu:moment}, we have \[ \alpha\|x^{k+1}-x^k\|^2\leq \kappa_f\|x^{k+1}-x^k\|+ \frac{1}{2\sigma}\sum_{i=1}^p\left([a_i]_+^2-[b_i]_+^2\right), \] in which, for simplicity, we use \[ a_i:=\lambda_i^k+\sigma G_i(x^k,\xi^k),\quad b_i:=\lambda_i^k+\sigma (G_i(x^k,\xi^k)+\langle v_i(x^k,\xi^k),x^{k+1}-x^k\rangle). \] Noticing that \[ \begin{array}{ll} \leq (|a_i|+|b_i|)\cdot|a_i-b_i|\\[8pt] \leq (2|a_i|+|b_i-a_i|)\cdot|a_i-b_i|\\[8pt] \leq 2|\lambda_i^k+\sigma G_i(x^k,\xi^k)|\cdot\sigma\kappa_g\|x^{k+1}-x^k\|+\sigma^2\kappa_g^2\|x^{k+1}-x^k\|^2, \end{array} \] we obtain \[ 2\alpha\|x^{k+1}-x^k\|\leq 2\kappa_f+\sum_{i=1}^p(2\kappa_g|\lambda_i^k+\sigma G_i(x^k,\xi^k)|+\sigma\kappa_g^2\|x^{k+1}-x^k\|). \] If $2\alpha-p\kappa_g^2\sigma>0$, it yields \[ \|x^{k+1}-x^k\|\leq \frac{2}{2\alpha-p\kappa_g^2\sigma}\left(\kappa_f+\sum_{i=1}^p(\kappa_g|\lambda_i^k+\sigma G_i(x^k,\xi^k)|\right). \] Therefore, from the facts that $\sum_{i=1}^p|\lambda_i^k|\leq\sqrt{p}\|\lambda^k\|$ and \[\sum_{i=1}^p|G_i(x^k,\xi^k)|\leq\sqrt{p}\|G(x^k,\xi^k)\|\leq\sqrt{p}\nu_g,\] the claim is obtained. \Halmos\endproof Under the Slater's condition, we derive the following conditional expected estimate of the multipliers. \begin{lemma}\label{lem:aux4} Let Assumption \ref{assu:slater} be satisfied. Then, for any $t_2 \leq t_1-1$ where $t_1$ and $t_2$ are positive integers, \[ \mathbb{E}\left[\langle \lambda^{t_1}, G(\widehat x, \xi^{t_1}) \rangle \,|\, \cF_{t_2}\right] \leq -\varepsilon_0 \mathbb{E}\left[\|\lambda^{t_1}\| \,|\, \cF_{t_2}\right]. \] \end{lemma} \proof{Proof.} %To prove this lemma, we first show that %\mathbb{E}\left[\lambda^{t_1}_iG_i(\widehat x, \xi^{t_1}) \,|\, \xi^{[t_2]}\right] %\leq -\varepsilon_0 \mathbb{E}\left[\lambda^{t_1}_i \,|\, \xi^{[t_2]}\right],\ i=1,\ldots,p. For any $i \in \{1,\ldots,p\}$, noticing that $\lambda^{t_1}_i\in \cF_{t_1}$ and $\cF_{t_2}\subseteq \cF_{t_1}$ for $t_2 \leq t_1-1$, we have \[ \begin{array}{ll} \mathbb{E}\left[\lambda^{t_1}_iG_i(\widehat x, \xi^{t_1}) \,|\, \cF_{t_2}\right] &=\mathbb{E} \left[\mathbb{E}\left[\lambda^{t_1}_iG_i(\widehat x, \xi^{t_1}) \,|\, \cF_{t_1}\right]\,|\,\cF_{t_2}\right]\\[10pt] %&=\mathbb{E}\left[G_i(\widehat x, \xi^{t_1})\right]\left[\lambda^{t_1}_i \,|\,\xi^{[t_2]}\right]\\[10pt] &\leq -\varepsilon_0 \mathbb{E}\left[\lambda^{t_1}_i \,|\, \cF_{t_2}\right]. \end{array} \] Summing the above inequality over $i \in \{1,\ldots,p\}$ yields \mathbb{E}\left[\langle \lambda^{t_1}, G(\widehat x, \xi^{t_1}) \rangle \,|\, \cF_{t_2}\right] \leq -\varepsilon_0 \mathbb{E} \left[ \sum_{i=1}^p \lambda^{t_1}_i \,|\, \cF_{t_2}\right] \leq -\varepsilon_0 \mathbb{E}\left[\|\lambda^{t_1}\| \,|\, \cF_{t_2}\right], by using $\sum_{i=1}^p \lambda^{t_1}_i\geq \|\lambda^{t_1}\|$. \Halmos\endproof We next present some important relations of $\|\lambda^k\|$. \begin{lemma}\label{lem:aux5} Let Assumptions \ref{assu:compact}--\ref{assu:slater} be satisfied and $s > 0$ be an arbitrary integer. Define $\beta_0:=\nu_g+\sqrt{p}\kappa_gR$ and \begin{equation}\label{eq:theta9} \vartheta (\sigma,\alpha,s):= \frac{\varepsilon_0\sigma s}{2}+\sigma \beta_0(s-1)+ \frac{\alpha R^2}{\varepsilon_0s}+\frac{2\kappa_f R}{\varepsilon_0}+ \frac{\sigma \nu_g^2}{\varepsilon_0}. \end{equation} Then, the following holds: \begin{equation}\label{eq:6} |\|\lambda^{k+1}\|-\|\lambda^k\||\leq \sigma \beta_0 \end{equation} \begin{equation}\label{eq:7} \mathbb{E}\left [ \|\lambda^{k+s}\|-\|\lambda^k\| \,|\, \cF_k\right] \leq \left \{ \begin{array}{ll} s \sigma \beta_0, & \mbox{if } \|\lambda^k\| < \vartheta (\sigma,\alpha, s),\\[6pt] -s \displaystyle \frac{\sigma \varepsilon_0}{2}, & \mbox{if } \|\lambda^k\| \geq \vartheta (\sigma,\alpha,s). \end{array} \right. \end{equation} \end{lemma} \proof{Proof.} It follows from Assumptions \ref{assu:compact}--\ref{assu:moment}, (\ref{eq:xna1}) and the nonexpansion property of the projection $\Pi_{\R^p_+}(\cdot)$ that \[ \begin{array}{ll} &\leq\|\lambda^{k+1}-\lambda^k\| =\|[\lambda^k+\sigma (G(x^k,\xi^k)+V(x^k,\xi^k)(x^{k+1}-x^k))]_+-[\lambda^k]_+\|\\[6pt] & \leq \sigma \|G(x^k,\xi^k)+V(x^k,\xi^k)(x^{k+1}-x^k)\|\\[6pt] & \leq \sigma [\nu_g+\sqrt{p}\kappa_g R], \end{array} \] which implies (\ref{eq:6}). This also gives that $\|\lambda^{k+s}\|-\|\lambda^k\|\leq s \sigma \beta_0$. Hence, we only need to establish the second part in (\ref{eq:7}) under the case $\|\lambda^k\| \geq \vartheta (\sigma,\alpha,s)$. For a given positive integer $s$, suppose that $\|\lambda^k\| \geq \vartheta (\sigma,\alpha,s)$. For any $l \in \{k,k+1,\ldots,k+s-1\}$, from (\ref{eq:opt-x-1}) and the convexity of $G_i(\cdot,\xi^l)$ one has \begin{array}{l} \langle v_0(x^l,\xi^l),x^{l+1}-x^l\rangle + \frac{1}{2\sigma}\|\lambda^{l+1}\|^2+ \frac{\alpha}{2}\|x^{l+1}-x^l\|^2\\[10pt] \leq \langle v_0(x^l,\xi^l),\widehat x-x^l\rangle + \frac{1}{2\sigma}\left[ \sum_{i=1}^p[\lambda^l_i+\sigma (G_i(x^l,\xi^l)+ \langle v_i(x^l,\xi^l), \widehat x-x^l \rangle)]_+^2\right]\\[10pt] \quad\quad + \frac{\alpha}{2}(\|\widehat x-x^l\|^2-\|\widehat x-x^{l+1}\|^2)\\[10pt] \leq \langle v_0(x^l,\xi^l),\widehat x-x^l\rangle + \frac{1}{2\sigma}\|[\lambda^{l}+\sigma G(\widehat x,\xi^l)]_{+}\|^2+\frac{\alpha}{2}(\|\widehat x-x^l\|^2-\|\widehat x-x^{l+1}\|^2)\\[10pt] \leq \langle v_0(x^l,\xi^l),\widehat x-x^l\rangle + \frac{1}{2\sigma}\|\lambda^{l}+\sigma G(\widehat x,\xi^l)\|^2+\frac{\alpha}{2}(\|\widehat x-x^l\|^2-\|\widehat x-x^{l+1}\|^2). \end{array} Rearranging terms and using Assumption \ref{assu:cons} we obtain \[ \begin{array}{ll} \frac{1}{2\sigma} \left[\|\lambda^{l+1}\|^2-\|\lambda^l\|^2\right]\\[10pt] \leq \langle v_0(x^l,\xi^l),\widehat x-x^{l+1}\rangle + \langle \lambda^l,G(\widehat x,\xi^l)\rangle + \frac{\sigma}{2}\|G(\widehat x,\xi^l)\|^2\\[10pt] \quad\quad + \frac{\alpha}{2}(\|\widehat x-x^l\|^2-\|\widehat x-x^{l+1}\|^2)\\[10pt] \leq \kappa_fR+ \langle \lambda^l,G(\widehat x,\xi^l)\rangle + \frac{\sigma}{2}\nu_g^2+ \frac{\alpha}{2}(\|\widehat x-x^l\|^2-\|\widehat x-x^{l+1}\|^2). \end{array} \] Making a summation over $\{k,k+1,\ldots,k+s-1\}$ and taking the conditional expectation, we obtain from Lemma \ref{lem:aux4} that \[ \begin{array}{ll} \frac{1}{2\sigma} \mathbb{E}\left[\|\lambda^{k+s}\|^2-\|\lambda^k\|^2\,|\, \cF_k\right]\\[10pt] \leq (\kappa_f R + \frac{\sigma}{2}\nu_g^2 )s+ \sum_{l=k}^{k+s-1} \mathbb{E}\left[\langle \lambda^l,G(\widehat x,\xi^l)\rangle\,|\, \cF_k\right]+\frac{\alpha}{2}\|\widehat x-x^k\|^2\\[10pt] \leq (\kappa_f R + \frac{\sigma}{2}\nu_g^2 )s -\varepsilon_0 \sum_{l=0}^{s-1} \mathbb{E}\left[\|\lambda^{k+l}\|\,|\, \cF_k\right]+\frac{\alpha}{2}R^2 \\[10pt] \leq (\kappa_f R + \frac{\sigma}{2}\nu_g^2 )s -\varepsilon_0 \sum_{l=0}^{s-1} \mathbb{E}\left[\|\lambda^{k}\|-\sigma\beta_0l \,|\, \cF_k\right]+\frac{\alpha}{2}R^2 \\[10pt] \quad \quad (\mbox{from } \|\lambda^{k+1}\|\geq \|\lambda^k\|-\sigma \beta_0)\\[8pt] \leq (\kappa_f R + \frac{\sigma}{2}\nu_g^2 )s + \varepsilon_0\sigma\beta_0 \frac{s(s-1)}{2} -\varepsilon_0 s \|\lambda^{k}\| +\frac{\alpha}{2}R^2. \end{array} \] Further, we get from Assumption \ref{assu:cons} and (\ref{eq:theta9}) that \[ \begin{array}{l} \mathbb{E}\left[\|\lambda^{k+s}\|^2\,|\, \cF_k\right]\\[10pt] \leq \|\lambda^k\|^2+2\sigma(\kappa_f R + \frac{\sigma}{2}\nu_g^2 )s +\varepsilon_0\sigma^2\beta_0s(s-1)-2\varepsilon_0\sigma s \|\lambda^{k}\| +\sigma\alpha R^2\\[10pt] \leq(\|\lambda^k\|- \frac{\varepsilon_0\sigma}{2}s)^2 +\varepsilon_0\sigma^2 \beta_0s(s-1)+ 2\sigma(\kappa_f R + \frac{\sigma}{2}\nu_g^2 )s+\sigma\alpha R^2-\varepsilon_0\sigma s \|\lambda^{k}\| \\[10pt] \leq(\|\lambda^k\|- \frac{\varepsilon_0\sigma}{2}s)^2 %- \frac{\varepsilon_0^2\sigma^2}{4}s^2 +\varepsilon_0\sigma s[\sigma \beta_0(s-1)+ \frac{2(\kappa_f R + \frac{\sigma}{2}\nu_g^2 )}{\varepsilon_0}+\frac{\alpha R^2}{\varepsilon_0s}- \vartheta (\sigma,\alpha,s)]\\[10pt] \leq (\|\lambda^k\|- \frac{\varepsilon_0\sigma}{2}s)^2. \end{array} \] This, together with Jensen's inequality and the fact that $\|\lambda^k\|\geq \vartheta (\sigma,\alpha,s)\geq \frac{\varepsilon_0\sigma}{2}s$, implies that \mathbb{E}\left[\|\lambda^{k+s}\|\,|\, \cF_k\right]\leq \|\lambda^k\|- \frac{\varepsilon_0\sigma}{2}s. The proof is completed. \Halmos\endproof Let us make some comments on inequality (\ref{eq:7}). This result may seem a bit confusing. From the proof, we actually show that: the inequality $\mathbb{E}[\|\lambda^{k+s}-\lambda^k\| | \cF_k]\leq s\sigma\beta_0$ holds true under the conditions of Lemma 4; in addition, if $\|\lambda^k\|\geq \vartheta(\sigma,\alpha,s)$, the bound can be improved to $\mathbb{E}[\|\lambda^{k+s}-\lambda^k\| | \cF_k]\leq -s\frac{\sigma\varepsilon_0}{2}$. However, we state it in the form of (\ref{eq:7}) intentionally. Since this is only a middle result, our true purpose is to show that the conditions of the following lemma \citep[Lemma 5]{YMNeely2017} are satisfied for $\|\lambda^k\|$. \begin{lemma}\label{lem:Yu1} Let $\{Z_t, t \geq 0\}$ be a discrete time stochastic process adapted to a filtration $\{\cF_t, t\geq 0\}$ with $Z_0 = 0$ and $\cF_0 = \{\emptyset, \Omega\}$. Suppose there exist an integer $t_0 >0$, real constants $\theta>0$, $\delta_{\max}>0$ and $ 0 <\zeta \leq \delta_{\max}$ such that \[ \begin{array}{rl} |Z_{t+1}-Z_t| & \leq \delta_{\max},\\[12pt] \mathbb{E}[Z_{t+t_0}-Z_t\,|\, \cF_t] & \leq \left \{ \begin{array}{ll} t_0 \delta_{\max}, & \mbox{if } Z_t < \theta,\\[6pt] -t_0\zeta, & \mbox{if } Z_t \geq \theta, \end{array} \right. \end{array} \] hold for all $t \in \{1,2,\ldots\}.$ Then the following properties are satisfied. \begin{itemize} \item[(i)] The following inequality holds, \begin{equation}\label{eq:aux1} \mathbb{E}[Z_t] \leq \theta +t_0 \delta_{\max}+t_0 \frac{4 \delta_{\max}^2}{\zeta}\log \left[ \frac{8 \delta_{\max}^2}{\zeta^2} \right],\ \forall t \in \{1,2,\ldots\}. \end{equation} \item[(ii)] For any constant $0 < \mu <1$, we have \Pr\left[Z_t\geq z\right] \leq \mu,\ \forall t \in \{1,2,\ldots\}, \begin{equation}\label{eq:aux2} z=\theta +t_0 \delta_{\max}+t_0 \frac{4 \delta_{\max}^2}{\zeta}\log \left[ \frac{8 \delta_{\max}^2}{\zeta^2}\right]+t_0 \frac{4 \delta_{\max}^2}{\zeta}\log\left( \frac{1}{\mu} \right). \end{equation} \end{itemize} \end{lemma} It is not difficult to verify that, Lemma \ref{lem:aux5} implies that the conditions of Lemma \ref{lem:Yu1} are satisfied with respect to $\|\lambda^k\|$ if we take \[ \theta=\vartheta (\sigma,\alpha,s),\ \delta_{\max}=\sigma \beta_0,\ \zeta = \frac{\sigma}{2}\varepsilon_0,\ t_0=s. \] For simplicity, we define \[ \psi(\sigma,\alpha,s):= \kappa_0+\kappa_1\frac{\alpha}{s}+\kappa_2\sigma+\kappa_3 \sigma s,\ \phi (\sigma,\alpha,s,\mu):=\psi(\sigma,\alpha,s)+\frac{8\beta_0^2}{\varepsilon_0} \log \left( \frac{1}{\mu}\right)\sigma s, \] where $\kappa_0,\kappa_1,\kappa_2,\kappa_3$ are constants given by \be\label{eq:kappas} \kappa_0= \frac{2\kappa_f R}{\varepsilon_0},\,\, \kappa_1= \frac{ R^2}{\varepsilon_0},\,\, \kappa_2= \frac{ \nu_g^2}{\varepsilon_0}-\beta_0,\,\, \kappa_3=\left[2\beta_0 + \frac{\varepsilon_0}{2}+ \frac{8\beta_0^2}{\varepsilon_0}\log \frac{32\beta_0^2}{\varepsilon_0^2}\right]. \ee We can also observe that $\psi(\sigma,\alpha,s)$ and $\phi (\sigma,\alpha,s,\mu)$ are exactly the same as the right-hand sides of (\ref{eq:aux1}) and (\ref{eq:aux2}), respectively. Therefore, in view of Lemma \ref{lem:aux5}, the following lemma is a direct consequence of Lemma \ref{lem:Yu1}. \begin{lemma}\label{lem:lambda} Let Assumptions \ref{assu:compact}--\ref{assu:slater} be satisfied and $s > 0$ be an arbitrary integer. Then, it holds that \be\label{eq:lambda-Exp} \Exp[\|\lambda^k\|]\leq \psi(\sigma,\alpha,s). \ee Moreover, for any constant $0<\mu<1$, we have \be\label{eq:lambda-Pr} \Pr[\|\lambda^k\|\geq \phi(\sigma,\alpha,s,\mu)]\leq\mu. \ee \end{lemma} \section{Expected convergence rates}\label{sec:rates} In this section, we shall establish the expected convergence rates of SLPMM with respect to constraint violation and objective reduction. In the following lemma, we derive a bound of the constraints in terms of the averaged iterate $$\hat{x}^K=\frac{1}{K}\sum_{k=0}^{K-1}x^k,$$ where $K$ is a fixed iteration number. \begin{lemma}\label{lem:cons} Let Assumptions \ref{assu:compact}-\ref{assu:moment} be satisfied. Then, if the parameters satisfy $2\alpha-p\kappa_g^2\sigma>0$, for each $i=1,\ldots,p$ we have \[ \Exp[g_i(\hat{x}^K)]\leq\frac{1}{\sigma K}\Exp[\lambda^{K}_i]+ \frac{\kappa_g}{\alpha \sqrt{p}\nu_g\kappa_g\sigma)+\frac{\sqrt{p}\kappa_g^2}{\alpha K}\sum_{k=0}^{K-1}\Exp[\|\lambda^k\|]. \] \end{lemma} \proof{Proof.} From the definition $\lambda^{k+1}_i=[\lambda^k_i+\sigma ( G_i(x^k,\xi^k)+\langle v_i(x^k,\xi^k), x^{k+1}-x^k\rangle )]_+$, it follows that \begin{array}{ll} \lambda^{k+1}_i &\geq \lambda^k_i+\sigma( G_i(x^k,\xi^k)+\langle v_i(x^k,\xi^k), x^{k+1}-x^k\rangle )\\[10pt] & \geq \lambda^k_i+\sigma(G_i(x^k,\xi^k)- \kappa_g\|x^{k+1}-x^k\|). \end{array} Using Lemma \ref{lem:aux3}, we have \be\label{eq:a2} G_i(x^k,\xi^k)\leq\frac{1}{\sigma}(\lambda^{k+1}_i-\lambda_i^k)+\frac{\kappa_g}{\alpha}(\kappa_f+ \sqrt{p}\kappa_g\|\lambda^k\|+ \sqrt{p}\nu_g\kappa_g\sigma). \ee Taking conditional expectation with respect to $\cF_k$, it yields that \[ g_i(x^k)\leq\frac{1}{\sigma}(\Exp[\lambda^{k+1}_i|\cF_k]-\lambda_i^k)+\frac{\kappa_g}{\alpha}(\kappa_f+ \sqrt{p}\kappa_g\|\lambda^k\|+ \sqrt{p}\nu_g\kappa_g\sigma), \] which further gives that \[ \Exp[g_i(x^k)]\leq\frac{1}{\sigma}(\Exp[\lambda^{k+1}_i]-\Exp[\lambda_i^k])+\frac{\kappa_g}{\alpha}(\kappa_f+ \sqrt{p}\kappa_g\Exp[\|\lambda^k\|]+ \sqrt{p}\nu_g\kappa_g\sigma). \] Summing over $\{0,\ldots,K-1\}$ and noticing that $\lambda^0=0$, we obtain \sum_{k=0}^{K-1} \Exp[g_i(x^k)]\leq \frac{1}{\sigma}\Exp[\lambda^{K}_i]+ \frac{\kappa_g K}{\alpha \sqrt{p}\nu_g\kappa_g\sigma)+\frac{\sqrt{p}\kappa_g^2}{\alpha}\sum_{k=0}^{K-1}\Exp[\|\lambda^k\|]. Therefore, from the convexity of $g_i$ and the definition of $\hat{x}^K$ it follows \[ \begin{array}{ll} \Exp[g_i(\hat{x}^K)]&\leq\frac{1}{K}\sum_{k=0}^{K-1} \Exp[g_i(x^k)]\\[10pt] &\leq\frac{\Exp[\lambda^{K}_i]}{\sigma K}+ \frac{\kappa_g(\kappa_f+ \sqrt{p}\nu_g\kappa_g\sigma)}{\alpha }+\frac{\sqrt{p}\kappa_g^2}{\alpha K}\sum_{k=0}^{K-1}\Exp[\|\lambda^k\|]. \end{array} \] The proof is completed. \Halmos\endproof In what follows, we present the bound of the objective reduction in terms of the averaged iterate. \begin{lemma}\label{lem:obj} Let Assumptions \ref{assu:compact}-\ref{assu:moment} be satisfied. Then, for any $z \in \Phi$, \[ \Exp[f(\hat{x}^K)]-f(z)\leq\frac{\kappa_f^2}{2\alpha}+\frac{\sigma}{2}\nu_g^2+\frac{\alpha}{2K}R^2. \] \end{lemma} \proof{Proof.} For any $z \in\Phi$, since $v_0(x^k,\xi^k)\in \partial_x F(x^k,\xi^k)$, we have \langle v_0(x^k,\xi^k), z-x^k \rangle\leq F(z,\xi^k)-F(x^k,\xi^k). Then, in view of (\ref{eq:opt-x-1}), one has \begin{equation}\label{eq:h1} \begin{array}{ll} & F(x^k,\xi^k) \\[8pt] & \leq F(z,\xi^k)+\left[\langle v_0(x^k,\xi^k), x^k-x^{k+1}\rangle- \frac{\alpha}{2}\|x^{k+1}-x^k\|^2\right &\quad + \frac{1}{2\sigma}\left[\|[\lambda^{k}+\sigma (G(x^k,\xi^k)+V(x^k,\xi^k)(z-x^k))]_+\|^2-\|\lambda^k\|^2\right]\\[12pt] & \quad -\frac{1}{2\sigma}\left[\|\lambda^{k+1}\|^2-\|\lambda^k\|^2\right] + \frac{\alpha}{2}\left[\|z-x^k\|^2-\|z-x^{k+1}\|^2\right]. \end{array} \end{equation} From Assumption \ref{assu:moment} and the fact that $\langle x,y\rangle\leq \frac{\alpha}{2}\|x\|^2+\frac{1}{2\alpha}\|y\|^2$, we obtain that \begin{equation}\label{eq:h2} \langle v_0(x^k,\xi^k),x^k-x^{k+1}\rangle- \frac{\alpha}{2}\|x^{k+1}-x^k\|^2 \leq \frac{1}{2\alpha}\|v_0(x^k,\xi^k)\|^2 \leq \frac{\kappa_f^2}{2\alpha}. \end{equation} For every $i=1,\ldots,p$, we have from $v_i(x^k,\xi^k)\in \partial_x G_i(x^k,\xi^k)$ and $[a]_+^2\leq a^2$ that [\lambda^k_i+\sigma(G_i(x^k,\xi^k)+\langle v_i(x^k,\xi^k),z-x^k\rangle)]_+^2 \leq [\lambda^k_i+\sigma G_i(z,\xi^k)]^2 and hence \|[\lambda^k+\sigma(G(x^k,\xi^k)+V(x^k,\xi^k)(z-x^k))]_+\|^2\leq \|\lambda^k+\sigma G(z,\xi^k)\|^2. Then, we obtain \begin{equation}\label{eq:h3} \begin{array}{ll} \|[\lambda^k+\sigma(G(x^k,\xi^k)+V(x^k,\xi^k)(z-x^k))]_+\|^2-\|\lambda^k\|^2\\[10pt] \leq 2\sigma \langle \lambda^k,G(z,\xi^k)\rangle+\sigma^2\|G(z,\xi^k)\|^2. \end{array} \end{equation} Substituting (\ref{eq:h2}) and (\ref{eq:h3}) into (\ref{eq:h1}), we get \be\label{eq:a8} \begin{array}{ll} F(x^k,\xi^k) & \leq F(z,\xi^k)+ \frac{\kappa_f^2}{2\alpha}-\frac{1}{2\sigma}\left[\|\lambda^{k+1}\|^2-\|\lambda^k\|^2\right] + \langle \lambda^k,G(z,\xi^k)\rangle\\[8pt] &\quad +\frac{\sigma}{2}\|G(z,\xi^k)\|^2 + \frac{\alpha}{2}\left[\|z-x^k\|^2-\|z-x^{k+1}\|^2\right]. \end{array} \ee Taking conditional expectation with respect to $\cF_k$ and noticing that \[ \Exp[\langle \lambda^k,G(z,\xi^k)\rangle|\cF_k]=\langle \lambda^k,g(z)\rangle\leq 0, \] we have \[ \begin{array}{ll} f(x^k)-f(z)&\leq \frac{\kappa_f^2}{2\alpha}-\frac{1}{2\sigma}\left[\Exp[\|\lambda^{k+1}\|^2|\cF_k]-\|\lambda^k\|^2\right]\\[8pt] &\quad+\frac{\sigma\nu_g^2}{2}+ \frac{\alpha}{2}\left[\|z-x^k\|^2-\Exp[\|z-x^{k+1}\|^2|\cF_k]\right], \end{array} \] which further gives \[ \begin{array}{ll} \Exp[f(x^k)]-f(z)&\leq \frac{\kappa_f^2}{2\alpha}-\frac{1}{2\sigma}\left[\Exp[\|\lambda^{k+1}\|^2]-\Exp[\|\lambda^k\|^2]\right]\\[8pt] &\quad+\frac{\sigma\nu_g^2}{2}+ \frac{\alpha}{2}\left[\Exp[\|z-x^k\|^2]-\Exp[\|z-x^{k+1}\|^2]\right]. \end{array} \] Making a summation and noticing that $\lambda^0=0$, one has \[ \sum_{k=0}^{K-1}\Exp[f(x^k)]\leq K\left[f(z)+\frac{\kappa_f^2}{2\alpha}+\frac{\sigma}{2}\nu_g^2\right]+\frac{\alpha}{2}\|z-x^0\|^2. \] Therefore, from the convexity of $f$ and the definition of $\hat{x}^K$ it follows \[ \Exp[f(\hat{x}^K)]\leq\frac{1}{K}\sum_{k=0}^{K-1}\Exp[f(x^k)]\leq f(z)+\frac{\kappa_f^2}{2\alpha}+\frac{\sigma}{2}\nu_g^2+\frac{\alpha}{2K}R^2. \] The proof is completed. \Halmos\endproof Based on Lemma \ref{lem:cons} and Lemma \ref{lem:obj}, if we take $\alpha=\sqrt{K}$, $\sigma=1/\sqrt{K}$ and $s=\ceil{\sqrt{K}}$, where $\ceil{a}$ denotes the ceiling function that returns the least integer greater than or equal to $a$, the excepted convergence rates of SLPMM with respect to constraint violation and objective reduction are shown to be $O(1/\sqrt{K})$ in the following theorem. \begin{theorem}\label{th:rate} Let Assumptions \ref{assu:compact}-\ref{assu:slater} be satisfied. If we take $\alpha=\sqrt{K}$ and $\sigma=1/\sqrt{K}$ in Algorithm \ref{alg:SLPMM}, where $K$ is a fixed iteration number. Then, the following statements hold. \begin{itemize} \item[(i)] If $K>\max\{1,p\kappa_g^2/2\}$, then we have \[ \Exp[g_i(\hat{x}^K)]\leq\frac{(1+\sqrt{p}\kappa_g^2 )\kappa_2+\sqrt{p}\nu_g\kappa_g^2}{K},\quad i=1,\ldots,p, \] where $\bar{\kappa}:=\kappa_0+\kappa_1 +2\kappa_3$ and $\kappa_0, \kappa_1, \kappa_2, \kappa_3$ are defined in (\ref{eq:kappas}). \item[(ii)] For all $K\geq 1$, \[ \Exp[f(\hat{x}^K)]-f(x^*)\leq\frac{\kappa_f^2+\nu_g^2+R^2}{2\sqrt{K}}, \] where $x^*$ is any optimal solution to (\ref{eq:1}). \end{itemize} \end{theorem} \proof{Proof.} Consider item (i). If $K>p\kappa_g^2/2$, we have $2\alpha-p\kappa_g^2\sigma>0$, then it follows from Lemma \ref{lem:cons} that \be\label{eq:a1} \Exp[g_i(\hat{x}^K)]\leq \frac{1}{\sigma K}\Exp[\lambda^{K}_i]+ \frac{\kappa_g}{\alpha \sqrt{p}\nu_g\kappa_g\sigma)+\frac{\sqrt{p}\kappa_g^2}{\alpha K}\sum_{k=0}^{K-1}\Exp[\|\lambda^k\|]. \ee If we take $s=\ceil{\sqrt{K}}$, then from Lemma \ref{lem:lambda} one has \[ \Exp[\|\lambda^k\|]\leq \psi(\sigma,\alpha,s)=\kappa_0+\kappa_1\frac{\alpha}{s}+\kappa_2 \sigma+\kappa_3 \sigma s\leq \kappa_0+\kappa_1+\frac{\kappa_2}{\sqrt{K}}+2\kappa_3=\bar{\kappa}+\frac{\kappa_2}{\sqrt{K}}. \] Therefore, from $\alpha=\sqrt{K}, \sigma=1/\sqrt{K}$ and (\ref{eq:a1}) we have \[ \Exp[g_i(\hat{x}^K)]\leq \frac{1}{\sqrt{K}}\left(\bar{\kappa}+\frac{\kappa_2}{\sqrt{K}}\right)+\frac{\kappa_g\kappa_f}{\sqrt{K}}+\frac{\sqrt{p}\nu_g\kappa_g^2}{K}+\frac{\sqrt{p}\kappa_g^2}{\sqrt{K}}\left(\bar{\kappa}+\frac{\kappa_2}{\sqrt{K}}\right), \] which verifies item (i). By taking $z=x^*$ in Lemma \ref{lem:obj}, we derive item (ii) since \[ \Exp[f(\hat{x}^K)]-f(x^*)\leq\frac{\kappa_f^2}{2\alpha}+\frac{\sigma}{2}\nu_g^2+\frac{\alpha}{2K}R^2=\frac{\kappa_f^2+\nu_g^2+R^2}{2\sqrt{K}}. \] The proof is completed. \Halmos\endproof Let us point out that all of the algorithms \citep{YMNeely2017,LanZ2016,ABR2021} have $O(1/\sqrt{K})$ expected convergence. However, the algorithm \citep{YMNeely2017} is an extension of Zinkevich's online algorithm \citep{Zi2003}, which is a variant of the projection gradient method, and the CSA method \citep{LanZ2016} is a stochastic counterpart of Polyak's subgradient method \citep{Polyak1967}. When problem (\ref{eq:1}) reduces to a deterministic problem, these algorithms have at most linear rate of convergence. In contrast, SLPMM becomes the (linearized) proximal method of multipliers, which has an asymptotic superlinear rate of convergence. Moreover, the iteration complexity analysis \citep{LanZ2016} is based on the selection of stepsizes, which are dependent on the parameters $R$ , $\kappa_f$ and $\kappa_g$. However, these data are not known beforehand when problem (\ref{eq:1}) is put forward to solve. Note that, in SLPMM the stepsizes $\sigma$ and $\alpha$ are problem-independent. \section{High probability performance analysis}\label{sec:prob} In this section, we shall establish the large-deviation properties of SLPMM. By Theorem \ref{th:rate} and Markov's inequality, we have for all $\rho_c>0$ and $\rho_o>0$ that \be\label{eq:d1} \Pr\left[g_i(\hat{x}^K)\leq \rho_c\left(\frac{(1+\sqrt{p}\kappa_g^2 )\kappa_2+\sqrt{p}\nu_g\kappa_g^2}{K}\right)\right]\geq 1-\frac{1}{\rho_c} \ee \be\label{eq:d2} \Pr\left[f(\hat{x}^K)-f(x^*)\leq \rho_o\frac{\kappa_f^2+\nu_g^2+R^2}{2\sqrt{K}}\right]\geq 1-\frac{1}{\rho_o}. \ee However, these results are very weak. In the following, we will show that these high probability bounds can be significantly improved. We introduce the following standard ``light-tail" assumption, see \citep{Lan2016,LanZ2016,LNSY2019} for instance. \begin{assumption}\label{assu:lt-cons} There exists a constant $\sigma_c>0$ such that, for any $x\in \cC$, \[ \Exp[\exp(\|G_i(x,\xi)-g_i(x)\|^2/\sigma_c^2)]\leq\exp(1),\quad i=1,\ldots,p. \] \end{assumption} From a well-known result \citep[Lemma 4.1]{Lan2020}, under Assumption \ref{assu:lt-cons} one has for any $\rho\geq 0$ and $i=1,\ldots,p$ that \be\label{eq:a3} \Pr\left[\frac{1}{K}\sum_{k=0}^{K-1}g_i(x^k)-\frac{1}{K}\sum_{k=0}^{K-1}G_i(x^k,\xi^k)\geq\frac{\rho\sigma_c}{\sqrt{K}}\right]\leq \exp(-\rho^2/3). \ee For the sake of readability, we define the following notations, \[ \theta_1:=\sigma_c+(1+\sqrt{p}\kappa_g^2)\frac{16\beta_0}{\varepsilon_0},\quad \theta_2:=\kappa_g\kappa_f+(1+\sqrt{p}\kappa_g^2)(\kappa_0+\kappa_1+2\kappa_3) \] \[ \theta_3:=(1+\sqrt{p}\kappa_g^2)\frac{16\beta_0}{\varepsilon_0},\quad \theta_4:=\sqrt{p}\nu_g\kappa_g^2+(1+\sqrt{p}\kappa_g^2)\kappa_2, \] in which $\beta_0$ is defined in Lemma \ref{lem:aux5}, $\kappa_0, \kappa_1, \kappa_2, \kappa_3$ are defined in (\ref{eq:kappas}) and other parameters are defined in Assumptions \ref{assu:compact}-\ref{assu:lt-cons}. We are now read to state the main result on constraint violation. \begin{theorem}\label{th:cons-pr} Let Assumptions \ref{assu:compact}-\ref{assu:lt-cons} be satisfied. We take $\alpha=\sqrt{K}$ and $\sigma=1/\sqrt{K}$ in Algorithm \ref{alg:SLPMM}, where $K$ is a fixed iteration number satisfying $K>\max\{1,p\kappa_g^2/2\}$. Then, for any $\rho\geq 0$ and $i=1,\ldots,p$, \[ \Pr\left[g_i(\hat{x}^K)\leq \frac{\theta_1\rho+\theta_2+\theta_3\log(K+1)}{\sqrt{K}}+\frac{\theta_4}{K}\right]\geq 1-\exp(-\rho^2/3)-\exp(-\rho). \] \end{theorem} \proof{Proof.} Summing (\ref{eq:a2}) over $\{0,\ldots,K-1\}$, we have \[ \frac{1}{K}\sum_{k=0}^{K-1}G_i(x^k,\xi^k)\leq \frac{\lambda_i^K}{\sigma K}+\frac{\kappa_g(\kappa_f+\sqrt{p}\nu_g\kappa_g\sigma)}{\alpha}+\frac{\sqrt{p}\kappa_g^2}{\alpha K}\sum_{k=0}^{K-1}\|\lambda^k\|. \] Noticing that $\alpha=\sqrt{K}$, $\sigma=1/\sqrt{K}$ and $g_i(\hat{x}^K)\leq \frac{1}{K}\sum_{k=0}^{K-1}g_i(x^k)$, one has \be\label{eq:a4} g_i(\hat{x}^K)\leq \frac{1}{K}\sum_{k=0}^{K-1}[g_i(x^k)-G_i(x^k,\xi^k)]+\frac{\lambda_i^K}{\sqrt{K}}+\frac{\kappa_g\kappa_f}{\sqrt{K}}+\frac{\sqrt{p}\nu_g\kappa_g^2}{K}+\frac{\sqrt{p}\kappa_g^2}{K^{3/2}}\sum_{k=0}^{K-1}\|\lambda^k\|. \ee We next consider the probability bound of $\lambda^k$. From (\ref{eq:lambda-Pr}), it follows that \[ \Pr[\|\lambda^k\|\geq \phi(\sigma,\alpha,s,\mu)]\leq\mu,\quad k=0,1\ldots,K. \] If we take $s=\ceil{\sqrt{K}}$ and $\mu=\exp(-\rho)/(K+1)$, then \[ \begin{array}{ll} \phi(\sigma,\alpha,s,\mu)&= \kappa_0+\kappa_1\frac{\alpha}{s}+\kappa_2\sigma+\kappa_3 \sigma s+\frac{8\beta_0^2}{\varepsilon_0} \log \left( \frac{1}{\mu}\right)\sigma s\\[10pt] &\leq \kappa_0+\kappa_1+\frac{\kappa_2}{\sqrt{K}}+2\kappa_3+\frac{16\beta_0^2}{\varepsilon_0}(\rho+\log(K+1)) \end{array} \] and hence for all $k=0,1,\ldots,K$, \be\label{eq:a5} \Pr[\|\lambda^k\|\geq \kappa_0+\kappa_1+\frac{\kappa_2}{\sqrt{K}}+2\kappa_3+\frac{16\beta_0^2}{\varepsilon_0}(\rho+\log(K+1))]\leq\frac{\exp(-\rho)}{K+1}. \ee Using (\ref{eq:a3}) and (\ref{eq:a5}) in (\ref{eq:a4}), we conclude that \[ \begin{array}{ll} \Pr\left[ g_i(\hat{x}^K)\geq \frac{\rho(\sigma_c+(1+\sqrt{p}\kappa_g^2)\frac{16\beta_0}{\varepsilon_0})}{\sqrt{K}}+\frac{\kappa_g\kappa_f+(1+\sqrt{p}\kappa_g^2)(\kappa_0+\kappa_1+2\kappa_3)}{\sqrt{K}}\right.\\[15pt] \quad\quad\left.+\frac{(1+\sqrt{p}\kappa_g^2)\frac{16\beta_0}{\varepsilon_0}\log(K+1)}{\sqrt{K}}+\frac{\sqrt{p}\nu_g\kappa_g^2+(1+\sqrt{p}\kappa_g^2)\kappa_2}{K} \right]\leq \exp(-\rho^2/3)+\exp(-\rho). \end{array} \] The proof is completed. \Halmos\endproof In view of Theorem \ref{th:cons-pr}, if we take $\rho=\log(K)$, then we have \[ \Pr\left[g_i(\hat{x}^K)\leq O\left(\frac{\log(K)}{\sqrt{K}}\right)\right]\geq 1-\frac{1}{K^{2/3}}-\frac{1}{K}. \] We next make the following ``light-tail" assumption with respect to the objective function. \begin{assumption}\label{assu:lt-obj} There exists a constant $\sigma_o>0$ such that, for any $x\in \cC$, \[ \Exp[\exp(\|F(x,\xi)-f(x)\|^2/\sigma_o^2)]\leq\exp(1). \] \end{assumption} Similar to (\ref{eq:a3}), under Assumption \ref{assu:lt-obj} one has for any $\rho\geq 0$ that \be\label{eq:a6} \Pr\left[\frac{1}{K}\sum_{k=0}^{K-1}f(x^k)-\frac{1}{K}\sum_{k=0}^{K-1}F(x^k,\xi^k)\geq\frac{\rho\sigma_o}{\sqrt{K}}\right]\leq \exp(-\rho^2/3) \ee \be\label{eq:a7} \Pr\left[\frac{1}{K}\sum_{k=0}^{K-1}F(z,\xi^k)-\frac{1}{K}\sum_{k=0}^{K-1}f(z)\geq\frac{\rho\sigma_o}{\sqrt{K}}\right]\leq \exp(-\rho^2/3) \ee for all $z\in \cC$. The following lemma is from \citep[Lemma 9]{YMNeely2017}. \begin{lemma}\label{lem:Yu2} Let $\{Z_t,t\geq 0\}$ be a supermartingale adapted to a filtration $\{\cF_t,t\geq 0\}$ with $Z_0=0$ and $\cF_0=\{\emptyset, \Omega\}$, i.e. $\mathbb E[Z_{t+1}\,|\, \cF_t]\leq Z_t$, $\forall t \geq 0$. Suppose there exists a constant $c>0$ such that $\{|Z_{t+1}-Z_t|>c\}\subseteq \{Y_t>0\}$, $\forall t\geq 0$, where each $Y_t$ is adapted to $\cF_t$. Then, for all $z>0$, we have \Pr[Z_t\geq z] \leq e^{-z^2/(2tc^2)}+ \sum_{j=0}^{t-1} \Pr[Y_j>0],\ \forall t \geq 1. \end{lemma} For any fixed $z \in \Phi$, by taking $Z_t:= \sum_{k=0}^{t-1} \langle \lambda^k, G(z, \xi^k) \rangle$ in Lemma \ref{lem:Yu2} we obtain the following lemma. \begin{lemma}\label{lem:l13} For any fixed $z \in \Phi$ and an arbitrary constant $c>0$, let $Z_0:=0$ and $Z_t:= \sum_{k=0}^{t-1} \langle \lambda^k, G(z, \xi^k) \rangle$ for $t\geq 1$. Let $\cF_0=\{\emptyset, \Omega\}$ and $Y_t:=\|\lambda^{t}\|-c/\nu_g$ for all $t\geq 0$. Then, for all $\gamma>0$, we have \Pr[Z_t\geq \gamma] \leq e^{-\gamma^2/(2tc^2)}+ \sum_{j=0}^{t-1} \Pr[Y_j>0],\ \forall t \geq 1. \end{lemma} \proof{Proof.} It is simple to check that $\{Z_t\}$ and $\{Y_t\}$ are both adapted to $\{\cF_t, t\geq 0\}$. Now we prove that $\{Z_t\}$ is a supermartingale. Since Z_{t+1}=Z_t+\langle \lambda^{t}, G(z, \xi^{t})\rangle, we have \begin{array}{ll} \mathbb{E}[Z_{t+1}\,|\, \cF_t]&=\mathbb{E}[ Z_t+\langle \lambda^{t}, G(z, \xi^{t})\rangle\,|\, \cF_t]\\[8pt] & =Z_t+\langle \lambda^{t}, \mathbb{E}[G(z, \xi^{t})\,|\, \cF_t]\rangle\\[8pt] &=Z_t+\langle \lambda^{t}, g(z)\rangle\\[8pt] &\leq Z_t, \end{array} which follows from $\lambda^t\in\cF_t$, $\lambda^t\geq 0$ and $g(z)\leq 0$. Thus, we obtain that $\{Z_t\}$ is a supermartingale. From Assumption \ref{assu:cons}, we get |Z_{t+1}-Z_t|=|\langle \lambda^{t}, G(z, \xi^{t} \rangle| \leq \nu_g\|\lambda^{t}\|. This implies that $\|\lambda^{t}\|> c/\nu_g$ if $|Z_{t+1}-Z_t| >c$ and hence \{|Z_{t+1}-Z_t| >c\} \subseteq \{Y_t>0\}. Therefore, we can observe that the conditions of Lemma \ref{lem:Yu2} are satisfied, and hence the claim is obtained. \Halmos\endproof Finally, we establish a high probability objective reduction bound in the following theorem. \begin{theorem}\label{th:obj-pr} Let Assumptions \ref{assu:compact}-\ref{assu:slater} and \ref{assu:lt-obj} be satisfied. We take $\alpha=\sqrt{K}$ and $\sigma=1/\sqrt{K}$ in Algorithm \ref{alg:SLPMM}, where $K\geq 1$ is a fixed iteration number. Then, for any $\rho\geq 0$, \[ \begin{array}{ll} \displaystyle\Pr\left[f(\hat{x}^K)-f(x^*)\leq\sqrt{2\rho}\nu_g\left(\frac{\kappa_0+\kappa_1+2\kappa_3}{\sqrt{K}}+\frac{\frac{16\beta_0^2}{\varepsilon_0}(\rho+\log(K))}{\sqrt{K}}+\frac{\kappa_2}{K}\right)\right.\\[15pt] \quad\quad \displaystyle\left.+ \frac{2\sigma_0\rho}{\sqrt{K}}+\frac{\theta_5}{\sqrt{K}} \right]\geq 1-2\exp(-\rho^2/3)-2\exp(-\rho), \end{array} \] where $x^*$ is any fixed optimal solution to (\ref{eq:1}), \theta_5:=(\kappa_f^2+\nu_g^2+R^2)/2$, $\beta_0$ is defined in Lemma \ref{lem:aux5} and $\kappa_0, \kappa_1, \kappa_2, \kappa_3$ are defined in (\ref{eq:kappas}). \end{theorem} \proof{Proof.} For any $z\in\Phi$, summing (\ref{eq:a8}) over $\{0,\ldots,K-1\}$ and using the facts that $\lambda^0=0$, $\|G(z,\xi^k)\|^2\leq \nu_g^2$ and $\|z-x^0\|^2\leq R^2$, we have \[ \frac{1}{K}\sum_{k=0}^{K-1}F(x^k,\xi^k)\leq \frac{1}{K}\sum_{k=0}^{K-1}F(z,\xi^k)+\frac{1}{K}\sum_{k=0}^{K-1}\langle \lambda^k, G(z, \xi^k)\rangle +\frac{\kappa_f^2+\nu_g^2+R^2}{2\sqrt{K}}. \] Then, it follows from $f(\hat{x}^K)\leq \frac{1}{K}\sum_{k=0}^{K-1}f(x^k)$ that \be\label{eq:b1} \begin{array}{ll} &\leq \displaystyle\frac{1}{K}\sum_{k=0}^{K-1}[f(x^k)-F(x^k,\xi^k)]+\frac{1}{K}\sum_{k=0}^{K-1}[F(z,\xi^k)-f(z)]\\[15pt] &\quad \displaystyle+\frac{1}{K}\sum_{k=0}^{K-1}\langle \lambda^k, G(z, \xi^k)\rangle+\frac{\kappa_f^2+\nu_g^2+R^2}{2\sqrt{K}}. \end{array} \ee By Lemma \ref{lem:l13}, for any $c>0$ and $\gamma>0$ we have \[ \Pr\left[\frac{1}{K}\sum_{k=0}^{K-1}\langle \lambda^k, G(z, \xi^k)\rangle\geq \frac{\gamma}{K}\right] \leq \exp(-\gamma^2/(2Kc^2))+ \sum_{k=0}^{K-1}\Pr[\|\lambda^k\|\geq c/\nu_g]. \] Let us take $s=\ceil{\sqrt{K}}$ and $\mu=\exp(-\rho)/K$, then \[ \phi(\sigma,\alpha,s,\mu)\leq \kappa_0+\kappa_1+\frac{\kappa_2}{\sqrt{K}}+2\kappa_3+\frac{16\beta_0^2}{\varepsilon_0}(\rho+\log(K)). \] If we take $c=\nu_g\phi(\sigma,\alpha,s,\mu)$, then from (\ref{eq:lambda-Pr}) we obtain \[ \sum_{k=0}^{K-1}\Pr[\|\lambda^k\|\geq c/\nu_g]\leq K\mu=\exp(-\rho). \] Moreover, let us take $\gamma=\sqrt{2\rho K} c$, then \[ \frac{\gamma}{K} \] and hence \be\label{eq:b2} \begin{array}{ll} \Pr\left[\frac{1}{K}\sum_{k=0}^{K-1}\langle \lambda^k, G(z, \xi^k)\rangle\geq \sqrt{2\rho}\nu_g\left(\frac{\kappa_0+\kappa_1+2\kappa_3}{\sqrt{K}}+\frac{\frac{16\beta_0^2}{\varepsilon_0}(\rho+\log(K))}{\sqrt{K}}+\frac{\kappa_2}{K}\right)\right] \\[15pt] \leq 2\exp(-\rho). \end{array} \ee Using (\ref{eq:a6}), (\ref{eq:a7}) and (\ref{eq:b2}) in (\ref{eq:b1}), one has \[ \begin{array}{ll} \Pr\left[f(\hat{x}^K)-f(z)\geq\frac{2\sigma_0\rho}{\sqrt{K}}+\sqrt{2\rho}\nu_g\left(\frac{\kappa_0+\kappa_1+2\kappa_3}{\sqrt{K}}+\frac{\frac{16\beta_0^2}{\varepsilon_0}(\rho+\log(K))}{\sqrt{K}}+\frac{\kappa_2}{K}\right)\right.\\[15pt] \quad\quad\left.+\frac{\kappa_f^2+\nu_g^2+R^2}{2\sqrt{K}} \right]\leq 2\exp(-\rho^2/3)+2\exp(-\rho). \end{array} \] The claim is derived by taking $z=x^*$ in the above inequality. \Halmos\endproof In view of Theorem \ref{th:obj-pr}, if we take $\rho=\log(K)$, then we have \[ \Pr\left[f(\hat{x}^K)-f(x^*)\leq O\left(\frac{\log^{3/2}(K)}{\sqrt{K}}\right)\right]\geq 1-\frac{2}{K^{2/3}}-\frac{2}{K}. \] In contrast to (\ref{eq:d1}) and (\ref{eq:d2}), we can observe that the results in Theorem \ref{th:cons-pr} and \ref{th:obj-pr} are much finer. % Moreover, the established probability bounds in Theorem \ref{th:cons-pr} and \ref{th:obj-pr} are comparable with the related existing work \cite{YMNeely2017}. \section{Preliminary numerical experiments}\label{sec:num} In this section, we demonstrate the efficiency of the proposed stochastic linearized proximal method of multipliers on two preliminary numerical problems. All numerical experiments are carried out using MATLAB R2020a on a desktop computer with Intel(R) Xeon(R) E-2124G 3.40GHz and 32GB memory. The MATLAB code and test problems can be found on \url{https://bitbucket.org/Xiantao_Xiao/SLPMM}. All reported time is wall-clock time in seconds. \subsection{Solving subproblems}\label{sec:subp} This subsection focuses on solving the subproblem (\ref{xna}) in SLPMM, that is \[ \begin{array}{l} x^{k+1}= \argmin\limits_{x \in \cC} \,\left\{ \cL^k_{\sigma }(x,\lambda^k) +\frac{\alpha}{2}\|x-x^k\|^2\right\}. \end{array} \] This problem is equivalent to \begin{equation}\label{eq:general-subp} \min_{x \in \cC} \phi(x):=\frac{1}{2}\sum_{i=1}^p[a_i^Tx+b_i]_+^2+\frac{1}{2}\|x\|^2+c^Tx, \end{equation} \[ a_i:=\sqrt{\frac{\sigma}{\alpha}}v_i(x^k,\xi^k),\ b_i:=\frac{\lambda_i}{\sqrt{\sigma\alpha}}+\sqrt{\frac{\sigma}{\alpha}}G_i(x^k,\xi^k)-\left\langle\sqrt{\frac{\sigma}{\alpha}}v_i(x^k,\xi^k),x^k\right\rangle \] and $c:=v_0(x^k,\xi^k)/\alpha-x^k$. Since $\phi$ is obviously strongly convex, we could apply the following popular Nesterov's accelerated gradient method to solve (\ref{eq:general-subp}). \mbox{}\\[4pt] {\bf APG}: Nesterov's accelerated projected gradient method for (\ref{eq:general-subp}). \begin{description} \item[Step 0 ] Input $x^0\in \cC$ and $\eta>1$. Set $y^0=x^0$, $L_{-1}=1$ and $t:=0$. \item[Step 1] \[ \] where $T_{L}(y):=\Pi_{\cC}[y-\frac{1}{L}\nabla \phi(y)]$, the stepsize $L_t=L_{t-1}\eta^{i_t}$ and $i_t$ is the smallest nonnegative integer satisfies the following condition \[ \begin{array}{ll} \phi(T_{L_{t-1}\eta^{i_t}}(y^t))&\leq \phi(y^t)+\langle\nabla \phi(y^t),T_{L_{t-1}\eta^{i_t}}(y^t)-y^t\rangle\\[10pt] \end{array} \] \item[ Step 2] Compute \[ \] \item[ Step 3] Set $t:=t+1$ and go to Step 1. \end{description} A well-known convergence result of the above method is that, if $\phi$ is $\mu$-strongly convex and $\nabla \phi$ is $L$-Lipschitz continuous, then $\phi(x^t)-\phi(x^*)\leq O\left((1-\sqrt{\mu/L})^t\right)$. See \citep{Beck2017} for a detailed discussion on this topic. Here, we assume that the set $\cC$ is simple such that the projection $\Pi_{\cC}$ can be efficiently computed. For example, if \[\cC:=\left\{x\in\R^n:\sum_{i=1}^nx_i=1,\ x\geq 0\right\},\] the projection $\Pi_{\cC}$ can be computed by the method proposed in \citep{WL2015}. When $\cC$ is $\R^n$ or a polyhedron, the subproblem is equivalent to a convex quadratic programming (QP) problem as \[ \begin{array}{ll} \min\limits_{x,y}\quad & \displaystyle\frac{1}{2}\sum_{i=1}^py_i^2+\frac{1}{2}\|x\|^2+c^Tx\\[8pt] \textrm{s.t.}\quad & a_i^Tx+b_i-y_i\leq 0,\ i=1,2,\ldots,p,\\[5pt] & x\in\cC,\quad y\geq 0. \end{array} \] In this case, the subproblem can also be solved by a QP solver. Let us also mention that, if $p=1$, the closed form of the stationary point to the objective function in Problem (\ref{eq:general-subp}) is given by \[ \tilde{x}=\left\{ \begin{array}{ll} -c,\quad&\mbox{if}\ -a_1^Tc+b_1\leq 0,\\[8pt] \end{array} \right. \] Then, $\tilde{x}$ is the unique optimal solution if it lies in the interior of $\cC$. \subsection{Neyman-Pearson classification} For a classifier $h$ to predict $1$ and $-1$, let us define the type I error (misclassifying class -1 as 1) and type II error (misclassifying class 1 as -1) respectively by \[ \mbox{type I error}:=\mathbb{E}[\varphi(-bh(a))|b=-1],\quad\mbox{type II error}:=\mathbb{E}[\varphi(-bh(a))|b=1], \] where $\varphi$ is some merit function. Unlike the conventional binary classification in machine learning, the Neyman-Pearson (NP) classification paradigm is developed to learn a classifier by minimizing type II error with type I error being below a user-specified level $\tau>0$, see \citep{TFZ2016} and references therein. In specific, for a given class $\mathcal{H}$ of classifiers, the NP classification is to solve the following problem \[ \begin{array}{ll} \min\limits_{h\in\mathcal{H}}&\mathbb{E}[\varphi(-bh(a))|b=1]\\[5pt] \mbox{s.t.}\quad &\mathbb{E}[\varphi(-bh(a))|b=-1]\leq \tau. \end{array} \] In what follows, we consider its empirical risk minimization counterpart. Suppose that a labeled training dataset $\{a_i\}_{i=1}^N$ consists of the positive set $\{a^0_i\}_{i=1}^{N_0}$ and the negative set $\{a^1_i\}_{i=1}^{N_1}$. The associated empirical NP classification problem is \begin{equation}\label{prob:NP} \begin{array}{ll} \min\limits_x&f(x):=\frac{1}{N_0}\sum_{i=1}^{N_0}\ell(x^Ta_i^0)\\[8pt] \mbox{s.t.}\quad &g(x):=\frac{1}{N_1}\sum_{i=1}^{N_1}\ell(-x^Ta_i^1)-\tau\leq 0, \end{array} \end{equation} where $\ell(\cdot)$ is a loss function, e.g., logistic loss $\ell(y):=\log(1+\exp(-y))$. The datasets tested in our numerical comparison are summarized in Table \ref{table:datasets}. The datasets for multi-class classification have been manually divided into two types. For example, the MNIST dataset is used for classifying odd and even digits. \begin{table}[!htp] \centering \caption{Datasets used in Neyman-Pearson classification} \begin{tabular}{|c||c|c|c|c|} \hline Dataset & Data $N$ & Variable $n$ & {Density} & Reference \\ \hline %$\mathtt{rcv1}$ & 20242 & 47236 & 0.16\% &\cite{RCV1}\\[2pt] $\mathtt{gisette}$ & 6000 & 5000 & 12.97\% &\citep{gisette}\\[2pt] $\mathtt{CINA}$ & 16033 & 132 & 29.56\% &\citep{CINA}\\[2pt] $\mathtt{MNIST}$ & 60000 & 784 & 19.12\% &\citep{MNIST}\\ \hline \end{tabular} \label{table:datasets} \end{table} %In Figure \ref{figure:regret}, we show the changes of the \textit{objective regret} and the \textit{regret in constraint violation} of SLPMM for solving the Neyman-Pearson classification problem (\ref{prob:NP}). For all three datasets, it can be seen that the regrets of objective and constraint violation converge rapidly to zero. In the following experiment, we show the performance of SLPMM compared with CSA \citep{LanZ2016}, PSG \citep{Xiao2019}, YNW \citep{YMNeely2017} and APriD \citep{YX2022}. For all five methods, we use an efficient mini-batch strategy, that is, at each iteration the stochastic gradients of the objective function and the constraint function are computed, respectively, by \[ v_0^k:=\frac{1}{|\cN_0^k|}\sum_{i\in\cN_0^k}\nabla f_i(x^k),\quad v_1^k:=\frac{1}{|\cN_1^k|}\sum_{i\in\cN_1^k}\nabla g_i(x^k), \] where $f_i(x):=\ell(x^Ta_i^0), i=1,\ldots,N_0$ and $g_i(x):=\ell(-x^Ta_i^1), i=1,\ldots,N_1$. Here, the sets $\cN_0^k$ and $\cN_1^k$ are randomly chosen from the index sets $\{1,\ldots,N_0\}$ and $\{1,\ldots,N_1\}$, respectively. The batch sizes $|\cN_0^k|$ and $|\cN_1^k|$ are fixed to $1\%$ of the data sizes $N_0$ and $N_1$, respectively. We choose $x^0=0$ as the initial point. The parameter $\tau$ is set to 1. The parameters in SLPMM is chosen as $\alpha=\sqrt{K}$ and $\sigma=1/\sqrt{K}$. The maximum number of iterations is set to $K=3000$. In Figure \ref{figure:gisette}, Figure \ref{figure:CINA} and Figure \ref{figure:MNIST}, we show the performance of all methods for solving the empirical NP classification problem with logistic loss. In each figure, the pictures (a) and (b) show the changes of the objective value and the constraint value with respect to \textit{epochs}, and the pictures (c) and (d) represent the changes of the objective value and the constraint value with respect to \textit{cputime}. Here, in (a) and (c) the horizontal dashed line represents a reference optimal objective value which is computed by the built-in MATLAB function \texttt{fmincon}. Moreover, one epoch denotes a full pass over a dataset. The results are averaged over 10 independent runs. Generally, we can observe that the behaviors of CSA, PSG and YNW are similar since all of them are stochastic first-order methods. SLPMM obviously outperforms these three methods by combining the evaluations of both objective decreasing and constraint violation. In particular, the results demonstrate that SLPMM converges obviously faster than CSA and PSG both with respect to epochs and cputime. Our results also show that PSG usually generates solutions which are failed to satisfy the constraint. In contrast, CSA always gives feasible solutions, but the objective values are far from optimal. Finally, the performance of APriD is very different from the others. The total performance of APriD seems better than the others. However, the curves of APriD oscillate heavily even for the average of 10 runs, and the issue is much worse for each independent run. \begin{figure}[htp] \centering \setlength{\belowcaptionskip}{-6pt} \begin{tabular}{cccc} \subfloat[$\mathtt{objective \slash epochs}$]{ \includegraphics[width=5.5cm]{./results_logloss_gisette_epoch_obj_10_runs.eps}} & \subfloat[$\mathtt{constraint \slash epochs}$]{ \includegraphics[width=5.5cm]{./results_logloss_gisette_epoch_cons_10_runs.eps}} &\\ \subfloat[$\mathtt{objective \slash cputime}$]{ \includegraphics[width=5.5cm]{./results_logloss_gisette_time_obj_10_runs.eps}} & \subfloat[$\mathtt{constraint \slash cputime}$]{ \includegraphics[width=5.5cm]{./results_logloss_gisette_time_cons_10_runs.eps}} \end{tabular} \caption{Comparison of algorithms on $\mathtt{gisette}$ for Neyman-Pearson classification.} \label{figure:gisette} \end{figure} \begin{figure}[htp] \centering \setlength{\belowcaptionskip}{-6pt} \begin{tabular}{cccc} \subfloat[$\mathtt{objective \slash epochs}$]{ \includegraphics[width=5.5cm]{./results_logloss_CINA_epoch_obj_10_runs.eps}} & \subfloat[$\mathtt{constraint \slash epochs}$]{ \includegraphics[width=5.5cm]{./results_logloss_CINA_epoch_cons_10_runs.eps}} &\\ \subfloat[$\mathtt{objective \slash cputime}$]{ \includegraphics[width=5.5cm]{./results_logloss_CINA_time_obj_10_runs.eps}} & \subfloat[$\mathtt{constraint \slash cputime}$]{ \includegraphics[width=5.5cm]{./results_logloss_CINA_time_cons_10_runs.eps}} \end{tabular} \caption{Comparison of algorithms on $\mathtt{CINA}$ for Neyman-Pearson classification.} \label{figure:CINA} \end{figure} \begin{figure}[htp] \centering \setlength{\belowcaptionskip}{-6pt} \begin{tabular}{cccc} \subfloat[$\mathtt{objective \slash epochs}$]{ \includegraphics[width=5.5cm]{./results_logloss_MNIST_epoch_obj_10_runs.eps}} & \subfloat[$\mathtt{constraint \slash epochs}$]{ \includegraphics[width=5.5cm]{./results_logloss_MNIST_epoch_cons_10_runs.eps}} &\\ \subfloat[$\mathtt{objective \slash cputime}$]{ \includegraphics[width=5.5cm]{./results_logloss_MNIST_time_obj_10_runs.eps}} & \subfloat[$\mathtt{constraint \slash cputime}$]{ \includegraphics[width=5.5cm]{./results_logloss_MNIST_time_cons_10_runs.eps}} \end{tabular} \caption{Comparison of algorithms on $\mathtt{MNIST}$ for Neyman-Pearson classification.} \label{figure:MNIST} \end{figure} \subsection{Stochastic quadratically constrained quadratical programming} In this subsection, we consider the following stochastic quadratically constrained quadratical programming \[ \begin{array}{ll} \min\limits_{x\in\cC} &f(x):=\mathbb{E}\left[\frac{1}{2}x^TA^{(0)}x+(b^{(0)})^Tx-c^{(0)}\right]\\[8pt] \mbox{s.t.}\quad &g_i(x):=\mathbb{E}\left[\frac{1}{2}x^TA^{(i)}x+(b^{(i)})^Tx+c^{(i)}\right]\leq 0,\ i=1,2,\ldots,p,\\[5pt] \end{array} \] where $A^{(i)}\in\cS_{+}^n$, $b^{(i)}\in\R^n$, $c^{(i)}\in\R$ for $i=0,1,\ldots,p$. Here, $\cS_+^n$ denotes the set of all $n\times n$ positive semidefinite matrices. The expectations are taken with respect to the components of the parameters $\{A^{(i)},b^{(i)},c^{(i)}\}_{i=0}^p$, which are all random variables. The following numerical example is partially motivated by \cite{CZP2021}. The set $\cC:=\{x\in\R^n:\|x\|\leq R\}$, where $R>0$ is a constant. Let $\widehat{x}\in\R^n$ be a given point with its enty $\widehat{x}_i$ being uniformly generated from $\left(-\frac{R}{\sqrt{n}},\frac{R}{\sqrt{n}}\right)$. Let $I_n$ be the identity matrix. For each $i=0,1,\ldots,p$, the random matrix $A^{(i)}=I_n+\Delta_i$, where $\Delta_i$ is a symmetric matrix and its entry is uniformly distributed over $[-0.1,0.1]$. The random vector $b^{(i)}$ is uniformly distributed from $[-1,1]$. The random variable $c^{(i)}$ is constructed with a particular purpose. Let $h^{(i)}$ be a random variable uniformly distributed over $[0,2i]$, then define $c^{(i)}=-(\frac{1}{2}\widehat{x}^TA^{(i)}\widehat{x}+(b^{(i)})^T\widehat{x}+h^{(i)})$. In this setting, we can easily verify that $g_i(\widehat{x})=-i<0$ for $i=1,\ldots,p$ and hence the Slater's condition is satisfied. We can also get that the optimal solution is $0$ and the optimal value is $\frac{1}{2}\|\widehat{x}\|^2$. In this experiment, we compare the performance of SLPMM with PSG, YNW and APriD. At each iteration of the algorithms, we generate the samples of $\{A^{(i)},b^{(i)},c^{(i)}\}_{i=0}^p$ based on the above distributions for function and gradient evaluation. We set $n=100$, $p=5$, $R=2$. The maximum number of iterations is set to $K=1000$. The initial point is set to $x^0=(\sqrt{R/n},\sqrt{R/n},\ldots,\sqrt{R/n})^T$. The results in terms of time are shown in Figure \ref{figure:QCQP}. From picture (b) (plots the value of $\max_i\{g_i(x^k)\}$), we can see the iterations of all algorithms satisfy the constraints. From picture (a), we observe that SLPMM is comparable with PSG, and obviously outperforms over APriD and YNW. \begin{figure}[htp] \centering \setlength{\belowcaptionskip}{-6pt} \begin{tabular}{cccc} \subfloat[$\mathtt{objective \slash cputime}$]{ \includegraphics[width=5.5cm]{./results_QCQP_random_time_obj_1_runs.eps}} & \subfloat[$\mathtt{constraint \slash cputime}$]{ \includegraphics[width=5.5cm]{./results_QCQP_random_time_cons_1_runs.eps}} \end{tabular} \caption{Comparison of algorithms on stochastic quadratically constrained quadratical programming.} \label{figure:QCQP} \end{figure} \subsection{Second-order stochastic dominance constrained portfolio optimization} In this subsection, we consider the following second-order stochastic dominance (SSD) constrained portfolio optimization problem \[ \begin{array}{ll} \min &\mathbb{E}[-\xi^Tx]\\[5pt] \mbox{s.t.}\quad &\mathbb{E}[[\eta-\xi^Tx]_+]\leq\mathbb{E}[[\eta-Y]_+],\quad\forall \eta\in\R,\\[5pt] &x\in\cC:=\{x\in\R^n:\sum_{i=1}^nx_i=1,\ \bar{x}\geq x\geq 0\}, \end{array} \] where $\bar{x}$ is the upper bound and $Y$ stands for the random return of a benchmark portfolio dominated by the target portfolio in the SSD sense. Since it was first introduced by \cite{DR2003}, SSD has been widely used to control risk in financial portfolio \citep{KD2018,Noyan2018}. \cite{KKU2016} showed that, if $Y$ is discretely distributed with $\{y_1,y_2,\ldots,y_p\}$, the SSD constrained portfolio optimization is reduced to \begin{equation}\label{eq:port-ssd} \begin{array}{ll} \min &f(x):=\mathbb{E}[-\xi^Tx]\\[8pt] \mbox{s.t.}\quad &g_i(x):=\mathbb{E}[[y_i-\xi^Tx]_+]-\mathbb{E}[[y_i-Y]_+]\leq 0,\quad i=1,\ldots,p,\\[8pt] &x\in \cC:=\{x\in\R^n:\sum_{i=1}^nx_i=1,\ \bar{x}\geq x\geq 0\}, \end{array} \end{equation} which is an instance of Problem (\ref{eq:1}). \cite{DMW2016} proposed several methods for solving SSD constrained optimization problems based on augmented Lagrangian framework and analyze their convergence. In particular, the proposed approximate augmented Lagrangian method with exact minimization (PALEM) has some similarities to SLPMM. At each iteration in PALEM, a minimization problem with respect to the augmented Lagrangian function of a reduced problem is solved to obtain $x^k$, and the multiplier $\mu^k$ is updated. They proved that the sequences $\{x^k\}$ and $\{\mu^k\}$ converge to the optimal solution of primal and dual problem, respectively. In contrast, although SLPMM is also constructed based on augmented Lagrangian framework as PALEM, they are quite different. The subproblem at each iteration in SLPMM is a minimization problem of a linearized augmented Lagrangian function together with a proximal term, which is easier to solve. The sampling strategy is different. In PALEM, the sample set is updated at each iteration based on the calculation of the expectation of constraint function. SLPMM only simply requires one sample at each iteration. Moreover, since in our setting the expectation is assumed to be impossible to be calculated, we can not obtain the convergence of the sequence to optimal solution. In this experiment, we compare the performance of SLPMM with APriD, PSG, YNW and PALEM to solve Problem (\ref{eq:port-ssd}) on the following four datasets \[ \{``\texttt{Dax\_26\_3046}",``\texttt{DowJones\_29\_3020}",``\texttt{SP100\_90\_3020}",``\texttt{DowJones\_76\_30000}"\} \] from \citep{KKU2016}. Take ``\texttt{DowJones\_29\_3020}" for example, ``\texttt{DowJones}" stands for Dow Jones Index, 29 is the number of stocks and 3020 is the number of scenarios, i.e., $n=29, p=3020$. The initial point is set to 0. For PALEM, we use the MATLAB function \texttt{fmincon} to solve the subproblem. For SLPMM, we utilize the Nesterov's accelerated projected gradient method (APG) to solve the subproblem (\ref{eq:general-subp}), the stopping criterion of APG is set to $\|y^t-T_{L_t}(y^t)\|\leq 10^{-6}$, and the projection $\Pi_{\cC}$ is computed by the method proposed in \citep{WL2015}. In particular, since the number of the constraints of Problem (\ref{eq:port-ssd}) is large, we apply a sampling technique to reduce the computational cost. In specific, at each iteration, instead of using the whole constraint index set $\{1,\ldots,p\}$ in the augmented Lagrangian function (\ref{augL}), we first randomly sample a subset $I_k\subset\{1,\ldots,p\}$ and then replace $\sum_{i=1}^p$ with $\sum_{i\in I_k}$ in (\ref{augL}). This sampling strategy, which is also used in \citep{Xiao2019}, is proven to be very efficient in practice. Let us also remark that, by taking an extra expectation with respect to $I_k$, the expected convergence rates of SLPMM coupled with this sampling strategy can be established in a similar way as in Section \ref{sec:rates}. This is also pointed out in \citep[Section 5]{Xiao2019}. The numerical results are presented in Figure \ref{figure:ssd}. Since the maximum of $p$ constraint values are always zero (which indicates that the constraints are satisfied), we omit the presentation of constraint violation. We only report the change of the objective value with respect to cputime. The horizontal dashed line in each picture represents a reference optimal objective value which is obtained from \citep{KKU2016}. In general, we can observe that SLPMM has an obvious advantage compared with the other four algorithms. In view of dataset ``\texttt{DowJones\_76\_30000}" which refers to a large scale optimization problem with 30,000 constraints, SLPMM converges to the optimal objective value less than 4 seconds. We can also observe that SLPMM is very robust and stable for all four datasets. \begin{figure}[htp] \centering \setlength{\belowcaptionskip}{-6pt} \begin{tabular}{cccc} \subfloat[\texttt{Dax\_26\_3046}]{ \includegraphics[width=5.5cm]{./results_SSD_DAX_26_3046_time_obj_1_runs.eps}} & \subfloat[\texttt{DowJones\_29\_3020}]{ \includegraphics[width=5.5cm]{./results_SSD_DowJones_29_3020_time_obj_1_runs.eps}} &\\ \subfloat[\texttt{SP100\_90\_3020}]{ \includegraphics[width=5.5cm]{./results_SSD_SP100_90_3020_time_obj_1_runs.eps}} & \subfloat[\texttt{DowJones\_76\_30000}]{ \includegraphics[width=5.5cm]{./results_SSD_DowJones_76_30000_time_obj_1_runs.eps}} \end{tabular} \caption{Comparison of algorithms for SSD constrained portfolio optimization.} \label{figure:ssd} \end{figure} %\subsection{MIMO Transmit Signal Design with Imperfect CSI} %Suppose that a base station is equipped with $n$ antennas and it simultaneously transmits $p$ data streams to $p$ users using MIMO signaling based on the estimated channel state information (CSI) $\hat{h}_i\in\mathbb{C}^n$, $i=1,2,\ldots,p$. Let $e_i$ denote the random error between the true channel $h_i$ and the estimated channel $\hat{h}_i$. %We now consider the following MIMO transmit signal design problem with imperfect CSI \citep{DB2009,LLK2019} %\min\limits_{Q} &f(Q):=\displaystyle\sum_{i=1}^p\Tr(Q_i)\\[5pt] %\mbox{s.t.}\quad &\mathbb{E}[G_i(Q,\xi)]\geq r_i,\quad\forall i=1,2,\ldots,p,\\[5pt] %&Q_i\succeq 0,\quad i=1,2,\ldots,p, %G_i(Q,\xi):=\log\left(1+\frac{h_i^HQ_ih_i}{\sum_{j\neq i}h_i^HQ_jh_i+\sigma_i^2}\right),\forall i=1,2,\ldots,p, %$\xi=\{e_1,e_2,\ldots,e_p\}$, $Q=\{Q_1,Q_2,\ldots,Q_p\}$ with each $Q_i$ being the covariance matrix of the transmit signal for user $i$, $\sigma_i^2$ is the variance of the thermal noise at user $i$, $r_i$ is the expected rate requirement for user $i$. Here, $\Tr(Q_i)$ is the trace of the matrix $Q_i$ and $Q_i\succeq 0$ denotes that the matrix $Q_i$ is required to be positive semidefinite. Let us note that the objective function $f(Q)$ is linear and the expectation constraints $\mathbb{E}[G_i(Q,\xi)]\geq r_i$ are nonconvex. %\cite{LLK2019} proposed a constrained stochastic successive convex approximation (CSSCA) method for solving this nonconvex problem, which is the state-of-the-art algorithm in this field. Let us remark that the successive convex approximation technique is also used by \cite{HYZ2011} to deal with joint chance constrained programs, in which the constraint is also nonconvex. We try to compare SLPMM with CSSCA for a randomly generated test case which can be found on arxiv: https://arxiv.org/abs/1801.08266. In both SLPMM and CSSCA, CVX \citep{cvx} is used to solve the convex subproblem at each iteration. %The numerical results are presented in Figure \ref{figure:mimo}. Since there are $p$ constraints, we use the maximum of $p$ constraint values to show the constraint violation. Unlike the previous experiments, the results are not promising. Although the objective value of SLPMM is better than CSSCA, the constraint value in SLPMM can not decrease to zero. The possible reason is that, for nonconvex constraint, simply linearizing it and adding to the augmented Lagrangian function may not work. Other surrogate functions should be considered, which is out of the scope of this paper. We leave this as a future topic. %\includegraphics[width=5.5cm]{./results_MIMO_exam1_time_obj_1_runs.eps}} & %\caption{Comparison of algorithms for MIMO Transmit Signal Design.} \section{Conclusion}\label{sec:conclusion} We present a hybrid method of stochastic approximation technique and proximal augmented Lagrangian method. It is shown that the expected convergence rates and the large-deviation properties are comparable with the existing related stochastic methods. On the other hand, the proposed method is parametric-independent. Numerical experiments also demonstrate the superiority in comparison with the stochastic first-order methods. Thus, both theoretical and numerical results suggest that the proposed algorithm is efficient for solving convex stochastic programming with expectation constraints. However, there are still several valuable questions left to be answered. It is well-known that the deterministic augmented Lagrangian can achieve superlinear convergence. Therefore, the first question is whether the convergence rates can be improved to match the numerical performance and the rates in the deterministic setting. Secondly, it is worthwhile to consider the inexact method, that is, the subproblem is solved inexactly. Another interesting topic is how to use the techniques in this paper to deal with nonconvex stochastic optimization. The proposed algorithm in the current form is not applicable to solve nonconvex problems, such as chance constrained programs \citep{BSZ2021} and MIMO transmit signal design problem \citep{LLK2019}. Finally, let us mention that the stochastic algorithms for stochastic optimization can be easily extended to solve online problems, and vice versa, see \citep{YMNeely2017} for instance. Hence, the proposed method can be slightly revised to solve the corresponding online problems. % Samples of sectioning (and labeling) in IJOC % NOTE: (1) \section and \subsection do NOT end with a period % (2) \subsubsection and lower need end punctuation % (3) capitalization is as shown (title style). %\section{Introduction.}\label{intro} %%1. %\subsection{Duality and the Classical EOQ Problem.}\label{class-EOQ} %% 1.1. %\subsection{Outline.}\label{outline1} %% 1.2. %\subsubsection{Cyclic Schedules for the General Deterministic SMDP.} % \label{cyclic-schedules} %% 1.2.1 %\section{Problem Description.}\label{problemdescription} %% 2. % Text of your paper here % Acknowledgments here \ACKNOWLEDGMENT{% The authors would like to thank the anonymous reviewers and the associate editor for the valuable comments and suggestions that helped us to greatly improve the quality of the paper. }% Leave this (end of acknowledgment) % Appendix here % Options are (1) APPENDIX (with or without general title) or % (2) APPENDICES (if it has more than one unrelated sections) % Outcomment the appropriate case if necessary % \begin{APPENDIX}{<Title of the Appendix>} % \end{APPENDIX} % or % \begin{APPENDICES} % \section{<Title of Section A>} % \section{<Title of Section B>} % etc % \end{APPENDICES} % References here (outcomment the appropriate case) % CASE 1: BiBTeX used to constantly update the references % (while the paper is being written). \bibliographystyle{informs2014} % outcomment this and next line in Case 1 \bibliography{ref.bib} % if more than one, comma separated % CASE 2: BiBTeX used to generate mypaper.bbl (to be further fine tuned) %\input{mypaper.bbl} % outcomment this line in Case 2 %(1) This paper presents and analyzes the convergence of an efficient stochastic augmented Lagrangian-type algorithm for solving expectation constrained stochastic optimization problem. The considered problem is a standard model in operations research, especially in stochastic programming community. The topic in this paper perfectly matches the scope of IJOC. %(2) This paper designs an efficient algorithm of a type of stochastic optimization problems and analyzes its convergence rates. Thus, it is appropriate for the Design & Analysis of Algorithms-Continuous area. \end{document}
# $\delta\mathcal{N}$ formalism on the past light-cone G. Fanizza G. Marozzi and M. Medeiros ###### Abstract We apply the gradient expansion approximation to the light-cone gauge, obtaining a separate universe picture at non-linear order in perturbation theory within this framework. Thereafter, we use it to generalize the $\delta\mathcal{N}$ formalism in terms of light-cone perturbations. As a consistency check, we demonstrate the conservation of the gauge invariant curvature perturbation on uniform density hypersurface $\zeta$ at the completely non-linear level. The approach studied provides a self-consistent framework to connect at non-linear level quantities from the primordial universe, such as $\zeta$, written in terms of the light-cone parameters, to late time observables. ## 1 Introduction Advances in cosmological observations have provided us with high-precision methods to study the universe [1, 2, 3]. So far, linear cosmological perturbation theory has been the main tool to describe the early universe, particularly the primordial seeds that are believed to be produced by quantum mechanical fluctuations during inflation. These fluctuations grow during the quasi-exponential expansion epoch and freeze outside the horizon. Later on, they re-enter the horizon during a power-law expansion epoch, giving rise to the large-scale structure observed in the universe. In order to link the gauge invariant quantity that characterizes such primordial fluctuations with the observations, we need to have a good understanding of their behaviour outside the horizon. An interesting example is given by the primordial curvature perturbation on the uniform density hypersurface $\zeta$, which is expected to be of order $10^{-5}$ at the last- scattering surface and has been shown to be conserved outside the horizon, both at the linear [4] and the non-linear level [5]. The first-order treatment for the primordial fluctuations agrees with the observations of a nearly Gaussian, scale-invariant power spectrum. Although non-linearities are expected to be small, they are however unavoidable as a consequence of the non-linear evolution of the perturbations. Detection of the related non- Gaussianities can then provide important insights into early universe models, such as the inflationary ones [6, 7]. The evolution of $\zeta$ is proportional to the non-adiabatic (if any) contribution in the energy-momentum tensor, as shown in [4, 5]. In the linear regime, $\zeta$ has been successfully calculated using the $\delta\mathcal{N}$ formalism [8, 9, 10, 11, 12, 13], which has been extended to the exact non- linear level [5] by applying the first-order gradient expansion directly in the equations of motion provided by the Arnowitt-Deser-Misner (ADM) formalism [14]. The first-order gradient expansion, also known as the separate universe (SU) scheme, describes the universe as a set of FLRW geometries with independent equations of motion and is a good approximation in the regime of large comoving wavelengths compared to the horizon [8, 9, 4, 5]. The great advantage of the SU scheme is that the equations of motion within this approximation have the same form both for the background and perturbed universe with the exception of the momentum constraint, which vanishes in the background. As a consequence, one can obtain the non-linear field’s evolution from the background one by imposing non-linear initial conditions [15]. In [16], this formalism has been generalized to include stochastic effects and derive, within the framework of the stochastic approach [17] and its relation with QFT [18, 19], non-perturbative correlation functions for single-field slow-roll inflation. Further extensions have also studied ultra-slow-roll inflation [20, 21, 22, 23], allowing for the investigation of primordial black-hole production [20, 21, 22, 23, 24, 25]. The $\delta\mathcal{N}$ formalism has also been extended to the case when cosmic shear is included to describe the anisotropic expansion. In such a framework, the evolution of gravitational waves has been explored both for the case of a Bianchi I universe [26] and when couplings with external fields are present [27, 28]. A formalism that connects the picture of the primordial universe presented so far to late-time observables would be greatly welcome, especially if such a connection can account for non-linearities. Since the Geodesic Light-Cone (GLC) gauge [29] gives the chance of describing light-like observables in the late universe exactly, such as the redshift and the distance-redshift relation [29, 30, 31, 32, 33, 34, 35, 36], the galaxy number count [37, 38], the non- linear corrections to the CMB spectra [39, 40], and also Ultra-Relativistic particles [41], this is a natural framework to pursue the aforementioned program. Moreover, there has also been recent interest in the GLC gauge application to the study of backreaction effects from the primordial universe [36, 42, 43, 44]. Although these are very interesting prospects, one still must face the fact that the evolution of perturbations in the GLC gauge is quite involved already at linear order (see [44] for an analytical treatment). An alternative approach to this may be provided by numerical attempts, as done for instance in [45] for the linearized evolved solution for the gravitational potential on the past light-cone. In this manuscript, we take a different route by providing simplified equations of motion using the SU approach on the past light-cone. The connection between the primordial origin of inhomogeneities and their observations has to deal with the fact that the latter are done along our past light-cone, whereas the primordial universe is usually described using spatial hypersurfaces. During primordial epochs, these hypersurfaces are naturally described in terms of uniform field slices. In fact, in a single-field inflationary scenario, the inflaton is the only clock available. Therefore, the natural slicing which describes the dynamical space-time evolution is the one given by uniform inflaton hypersurfaces, which also fixes the time gauge mode. Another interesting fixing for the time coordinate is the one describing uniform density slices. This is an interesting fixing because it directly translates the density perturbations into curvature perturbations, providing the initial conditions for large-scale structure formation and allowing us to use the $\delta\mathcal{N}$ formalism to connect current inhomogeneities to the primordial ones. These two gauge fixings are usually called, respectively, uniform field gauge (UFG) and uniform density gauge (UDG). To make contact between these gauges and the GLC one, we recall that, although the GLC time coordinate is fixed to the time measured by a free-falling observer, a generalization of this gauge is provided in [44], the so-called Light-Cone (LC) gauge. In this generalization, the time gauge choice is left unspecified, allowing us to fix the lapse function for describing the uniform field and uniform density slicing on the past light-cone. An alternative approach could be to start from the cosmological perturbation theory on the past light-cone developed in [36, 42], and then perform the necessary gauge transformations. A description of the primordial universe in terms of the LC gauge could be a promising framework to connect non-linearly the late universe to the primordial one, given in terms of light-cone parameters. Here, we will take a first step in this direction. In particular, along this manuscript, we will discuss the gradient expansion as done on the observer’s past light-cone, which allows us to obtain a SU picture and the $\delta\mathcal{N}$ formalism in terms of light-cone perturbations. We will provide this both at the fully non-perturbative level, using the LC gauge [44], and, as a consistency check, at the linear level using the light-cone perturbation theory [36, 42]. By considering the LC gauge as a non-linear ADM decomposition (see [44] for more details), we will show that, unlike previous literature, where the shift vector was a first-order term in the gradient expansion [15, 5, 26], in the LC gauge the shift vector has to be taken into account also for the background. This is an important difference, since in this case the shift vector corresponds to the direction of propagation of the photon, and it is used to take into account inhomogeneities along the photon propagation direction. However, we will neglect spatial derivatives of such shift vector since they correspond to light-cone distortion effects which are expected to be negligible on large scales. After such implementations, we will show that the SU picture can be realized in the LC gauge (i.e., we will obtain evolution equations with the same form for both perturbed and background universe). Furthermore, within the gradient expansion approximation, we will verify at the fully non-linear level that the curvature perturbation on uniform density slices $\zeta$ is a conserved quantity (for adiabatic pressure) also when the light-like slicing of spacetime is used. This is a sanity check that confirms how the SU picture can be extended also to the case of the light-cone gauge. In summary, we will present a novel approach for connecting the primordial universe with observations along the past light-cone, by developing the $\delta\mathcal{N}$ formalism in the LC gauge. We will verify that the LC gauge allows for a non-linear description of the primordial universe in terms of light-cone parameters, and that the SU picture can be realized in this gauge by neglecting spatial derivatives of the shift vector. The manuscript is organized as follows. In Sect. 2 we obtain a generic SU description where we keep inhomogeneities along the geodesics in terms of an ADM metric. In Sect. 3 we present the set of light-cone gauges used here and, with a non-linear diffeomorphism, we show how the standard ADM formalism relates with the LC gauge. Moreover, we provide the LC gauge fixing condition in terms of the ADM variables up to non-linear order. Thereafter, we discuss the SU formalism on the past light-cone, which is also presented for the GLC gauge. Sect. 4 is devoted to the computation of the non-linear number of e-folds and its relation to spacetime perturbations, which allow us to obtain the non-linear scale factor in the LC. We then give a proof for the super- horizon conservation of $\zeta$, at both linear and non-linear order in perturbation theory, and first order in the gradient expansion, for adiabatic fluids. Finally, we provide a generalization of the $\delta\mathcal{N}$ formalism in the LC gauge. In Sect. 5, our main conclusions are summarized and discussed. In Appendix A we provide the linear $\delta\mathcal{N}$ formalism as a consistency check of the obtained results. ## 2 Separate universe Let us begin by introducing a systematic approximation scheme, which can be used when the wavelength of the perturbations is larger than the physical horizon. This approximation, widely known as the SU approach [8, 9, 4, 5], consists of employing the already mentioned gradient expansion perturbative scheme. This is based on the quantity $\epsilon\equiv k/(aH)$, rather than on the amplitude of the perturbations. This quantity compares the comoving wavenumber of a given mode $k/a$ with the expansion rate $H$. As an example, within this approximation scheme, terms with one spatial derivative will be first order111As an example in a flat space, under a Fourier transformation, spatial derivatives give rise to terms proportional to $k$. Here we are considering that for a quantity $Q$, $\frac{1}{a}\partial_{i}Q\ll\partial_{t}Q\approx HQ$ [10]. in $\epsilon$. The first order gradient expansion is known as the SU approach, since in this case the equations of motion for a local patch of the perturbed space-time have the same form of the FLRW background ones [15]. Thus, in this view the universe can be described as a collection of FLRW geometries, each one locally described by a different scale factor. This approximation can be particularly interesting to study the super-horizon evolution of the curvature perturbation $\zeta$ with a light-cone foliation of the spacetime. In fact, the SU approach is used in [4, 5] to show the conservation of $\zeta$ on super-horizon scales, for adiabatic pressure, at linear and non-linear order in perturbation theory. By pplying the gradient expansion to the non-linear ADM formalism [15] one can provide a SU scheme in the uniform curvature gauge (UCG). It has been shown that the shift vector has a decaying evolution, and therefore during the exponential expansion of the universe, it can be considered as a first order term in the gradient expansion. Also analyzing the consistency between the Hamiltonian and momentum constraints, it has been shown that, taking into account also the momentum constraints, the results differ only by a decaying solution [15]. Moreover, it is shown that the additional information in the momentum constraints should be $\mathcal{O}(\epsilon^{3})$ in the gradient expansion. Thereby, for super-horizon perturbations, the SU scheme is a good approximation. In this manuscript, we will provide a SU picture for the LC [44] and GLC [29] gauges. As we will see later, one difference with the previous works is that, when we consider the LC gauge as an ADM decomposition, the shift vector does not vanish, not even on the background (see, for instance [44]). In fact, in the LC gauge, the shift vector describes the direction of observation. On the other hand, we will neglect the divergence of the shift vector, which describes the divergence of the direction of observation in the language of $1+3$ formalism. ### 2.1 ADM formalism In this section we provide the SU set of equations for generic perturbations. Firstly, we introduce the ADM splitting and the $1+3$ evolution equations, then we obtain general conditions which allow a SU evolution of the perturbations. Thereafter, we show how also the LC gauge satisfies these conditions. Starting with the ADM metric $ds_{ADM}^{2}=-\mathcal{M}^{2}dt^{2}+f_{ij}\left(dx^{i}+N^{i}dt\right)\left(dx^{j}+N^{j}dt\right)\,,$ (2.1) one can prove that, with a suitable choice of the coordinates, $N^{i}=\mathcal{O}\left(\epsilon\right)$ and therefore $\partial_{i}N^{i}=\mathcal{O}\left(\epsilon^{2}\right)$. This condition was assumed in the references [5, 4], and it was proved in [15] considering the UCG. Let us now work with the ADM foliation of Eq. (2.1). Thanks to this description of non-linear general perturbations, made on top of a FLRW background, we will show how to recover a SU picture even if the shift vector does not vanish in the background. It rather combines with the time derivative to provide a derivative along the time-like motion. The vector $n^{\mu}$ normal to the space-like hypersurfaces $t=const$ is given by $n^{\mu}=\frac{\partial^{\mu}t}{\left(-\partial_{\nu}t\partial^{\nu}t\right)^{\frac{1}{2}}}\,,$ (2.2) which satisfies $n_{\mu}=-\mathcal{M}\delta_{\mu}^{t}\,,\qquad n^{\mu}\partial_{\mu}=\frac{1}{\mathcal{M}}\left(\partial_{t}-N^{i}\partial_{i}\right)\,,$ (2.3) with correspondent induced metric given by $f_{\mu\nu}=g_{\mu\nu}+n_{\mu}n_{\nu}\,.$ (2.4) This metric can be used to define the following induced quantities $E\equiv n^{\mu}n^{\nu}T_{\mu\nu}\,,\quad\quad p^{\mu}\equiv-f^{\mu\nu}n^{\rho}T_{\nu\rho}\,,\quad\quad S^{\mu\nu}\equiv f^{\mu\rho}f^{\nu\sigma}T_{\rho\sigma}\,,$ (2.5) where $E$ is the energy, $p^{\mu}$ is the energy flux (or momentum) and $S^{\mu\nu}$ is stress tensor. Then, the standard energy-momentum tensor can be written as $T_{\mu\nu}=\rho n_{\mu}n_{\nu}+p_{\mu}n_{\nu}+p_{\nu}n_{\mu}+S_{\mu\nu}\,,$ (2.6) where $\rho$ and $p=\frac{1}{3}f^{\mu\nu}S_{\mu\nu}$ are respectively the energy density and the pressure of the given fluid. Now we have everything that we need to present the decomposed Einstein equations. These are developed in full details in [44], where the authors have specialized to the LC gauge as a $1+1+2$ ADM foliation. As a starting point, we can extract the energy (time-time) and momentum (time-space) constraints respectively given by222From now on, for a generic tensor $C_{ij}$, we will denote its trace with $C\equiv f^{ij}C_{ij}$. ${}^{(3)}R+\Theta_{n}^{2}-K_{ij}K^{ij}=\,2E\,,\qquad\qquad- D_{j}K_{i}^{j}+D_{i}\Theta_{n}=\,p_{i}\,,$ (2.7) where we defined the extrinsic curvature as $K_{\mu\nu}\equiv\nabla_{(\mu}n_{\nu)}$. We can then also define the expansion rate $\Theta_{n}$ as $\Theta_{n}\equiv f^{\mu\nu}K_{\mu\nu}\,.$ (2.8) The evolution of the induced metric $f_{ij}$ and of $K_{ij}$ is then obtained from the space-space decomposition $\displaystyle\left(\partial_{t}-\mathcal{L}_{N}\right)f_{ij}$ $\displaystyle=\,2\mathcal{M}K_{ij}\,,$ $\displaystyle\left(\partial_{t}-\mathcal{L}_{N}\right)K_{ij}$ $\displaystyle=\,\mathcal{M}\left[2K_{ik}K_{j}^{k}-K_{ij}\Theta_{n}-^{(3)}R_{ij}+S_{ij}-\frac{1}{2}f_{ij}\left(S-E\right)\right]+D_{i}D_{j}\mathcal{M}\,,$ (2.9) where $\mathcal{L}_{N}$ is the Lie derivative along the field $N^{i}$. Finally, the equations for the matter sector $\nabla_{\mu}T^{\mu\nu}=0$ are given by $\displaystyle\left(\partial_{t}-\mathcal{L}_{N}\right)E$ $\displaystyle=\,-D_{i}\left(\mathcal{M}p^{i}\right)-\mathcal{M}\left(\Theta_{n}E+K_{ij}S^{ij}\right)\,,$ $\displaystyle\left(\partial_{t}-\mathcal{L}_{N}\right)p_{i}$ $\displaystyle=\,-D_{j}\left(S_{i}^{j}\mathcal{M}\right)-\mathcal{M}\Theta_{n}p_{i}-ED_{i}\mathcal{M}\,.$ (2.10) Let us now follow the decomposition of [26] by extracting the shape-preserving volume expansion out of the spatial metric. First, we make a conformal re- scaling of $f_{ij}$ to $f_{ij}\equiv e^{2\Xi}\hat{f}_{ij}\,,$ (2.11) by requiring that the determinant $\hat{f}=\text{det}[\hat{f}_{ij}]=1$. In this way, we can interpret $e^{\Xi}$ as the local effective scale factor. Then, we re-scale accordingly also the other quantities $\displaystyle\hat{A}_{ij}=$ $\displaystyle\,e^{-2\Xi}\left(K_{ij}-\frac{1}{3}f_{ij}\Theta_{n}\right)\,,$ $\,{}^{(3)}\hat{\mathcal{R}}_{ij}=$ $\displaystyle\,e^{-2\Xi}\left({}^{(3)}R_{ij}-\frac{1}{3}\,^{(3)}R\,f_{ij}\right)\,,$ $\displaystyle\hat{\mathcal{S}}_{ij}=$ $\displaystyle\,e^{-2\Xi}\left(S_{ij}-\frac{1}{3}f_{ij}S\right)\,.$ (2.12) Before applying the decomposition (2.12) to Eqs. (2.9) and (2.10), we define $\mathcal{M}\frac{d}{d\widetilde{\lambda}}\equiv\left(\partial_{t}-\mathcal{L}_{N}\right)$ and compute the trace of the evolution of $A_{ij}\equiv e^{2\Xi}\hat{A}_{ij}$ $\displaystyle f^{ij}\left(\partial_{t}-\mathcal{L}_{N}\right)A_{ij}=$ $\displaystyle\frac{1}{\mathcal{M}}\left[f^{ij}f_{il}\frac{d}{d\widetilde{\lambda}}\left(f_{jk}A^{lk}\right)+f^{ij}f_{jk}A^{lk}\frac{d}{d\widetilde{\lambda}}f_{il}\right]$ $\displaystyle=$ $\displaystyle\frac{1}{\mathcal{M}}\frac{d}{d\widetilde{\lambda}}\left(f_{lk}A^{lk}\right)+f^{ij}f_{jk}A^{lk}2K_{il}$ $\displaystyle=$ $\displaystyle 2A_{ij}A^{ij}=2\hat{A}_{ij}\hat{A}^{ij}\,,$ (2.13) where we have used the first of Eqs. (2.9) from the first to the second line, and the fact that $A_{ij}$ is trace-less from the second to the last line. Now we can apply the decomposition (2.12) to Eqs. (2.9), using also Eqs. (2.7) and (2.13), to obtain $\displaystyle\frac{d\Theta_{n}}{d\widetilde{\lambda}}=$ $\displaystyle-\frac{\Theta_{n}^{2}}{3}-\hat{A}_{ij}\hat{A}^{ij}-\frac{1}{2}\left(E+S\right)+\frac{1}{\mathcal{M}}D^{2}\mathcal{M}\,,$ $\displaystyle\frac{d\Xi}{d\widetilde{\lambda}}=$ $\displaystyle\frac{\Theta_{n}}{3}\,,$ $\displaystyle\frac{d\hat{f}_{ij}}{d\widetilde{\lambda}}=$ $\displaystyle 2\hat{A}_{ij}\,.$ (2.14) Then the evolution of the trace-less quantity $\hat{A}_{ij}$ is given by $\frac{d\hat{A}_{ij}}{d\widetilde{\lambda}}=-\frac{1}{3}\Theta_{n}\hat{A}_{ij}+2e^{-2\Xi}A_{ik}A_{j}^{k}+\hat{\mathcal{S}}_{ij}-\,^{(3)}\hat{\mathcal{R}}_{ij}+\frac{1}{\mathcal{M}}\left(D_{i}D_{j}-\frac{1}{3}f_{ij}D^{2}\right)\mathcal{M}\,.$ (2.15) At this point, we will apply our gradient expansion scheme without any gauge fixing. Hence, considering that $n^{\mu}=\frac{1}{\mathcal{M}}\left(1,-N^{i}\right)$, for a generic tensor $l_{ij}$, it holds $\frac{1}{\mathcal{M}}\left(\partial_{t}-\mathcal{L}_{N}\right)l_{ij}=n^{\mu}\partial_{\mu}l_{ij}+l_{ik}\partial_{j}N^{k}+l_{jk}\partial_{i}N^{k}=\frac{d}{d\lambda}l_{ij}+\mathcal{O}\left(\epsilon^{2}\right)\,,$ (2.16) where we define $n^{\mu}\partial_{\mu}\equiv\frac{d}{d\lambda}$ and then we have $\frac{d}{d\widetilde{\lambda}}=\frac{d}{d\lambda}+\mathcal{O}\left(\epsilon^{2}\right)$. This corresponds to neglect terms proportional to $\partial_{i}N^{j}=\mathcal{O}\left(\epsilon^{2}\right)$ in Eq. (2.16), as done also in [15, 26]. This is our main assumption and comes from the fact that, due to the spatial derivatives, the momentum constraint of Eq. (2.7) is a first order relation, i.e. $p_{i}=\mathcal{O}\left(\epsilon\right)$. Hence, using Eq. (2.5), we get $\displaystyle p_{i}=-\frac{1}{\mathcal{M}}\left(T_{it}-N^{j}T_{ij}\right)\,.$ (2.17) Such Eq. (2.17) can be satisfied either by the strong condition $\displaystyle T_{it}\sim N^{i}=\mathcal{O}\left(\epsilon\right)\,,$ (2.18) or by the weaker condition that only the combination on its r.h.s. is $\mathcal{O}(\epsilon)$. Along this paper, we will adopt the stronger condition to justify our claim that $\partial_{i}N^{j}=\mathcal{O}(\epsilon^{2})$, in agreement with [15, 5, 4, 26]. Moreover, just as done in some previous works [15, 5, 4], we will also neglect ${}^{(3)}R_{ij}\thicksim R\thicksim\hat{\mathcal{S}}_{ij}=\mathcal{O}\left(\epsilon^{2}\right)$ since all of these terms contain double spatial derivatives. For what concerns the anisotropic stress $\hat{\mathcal{S}}_{ij}$, this is given by combinations of double spatial derivatives acting on the scalar fields in the matter sector. The condition $\hat{\mathcal{S}}_{ij}=\mathcal{O}(\epsilon^{2})$ was relaxed in [26] only for Bianchi geometries and on [27, 28] due to the presence of gauge fields. With these considerations, we can decompose again the metric evolution provided by the first of Eqs. (2.9) thanks Eqs. (2.12). Hence, at first order in the gradient expansion, we obtain333Note that, although Eqs. (2.19) are very similar to Eqs. (2.14), after the gradient expansion we have replaced $\frac{d}{d\tilde{\lambda}}=\frac{d}{d\lambda}+\mathcal{O}(\epsilon^{2})$. $\frac{d\Xi}{d\lambda}=\frac{\Theta_{n}}{3}\,,\qquad\qquad\frac{d\hat{f}_{ij}}{d\lambda}=2\hat{A}_{ij}\,.$ (2.19) Moreover, we can prove that $\hat{A}_{ij}$ is at least a second order term in the gradient expansion at every order in perturbation theory. In fact, following [15], from Eq. (2.15) we get $\frac{d}{d\lambda}\hat{A}_{ij}=-\frac{1}{3}\Theta_{n}\hat{A}_{ij}+2\hat{A}_{ik}\hat{A}_{j}^{k}+\mathcal{O}\left(\epsilon^{2}\right)\,,$ (2.20) and we choose a coordinate system such that $A_{ij}$ vanishes on the background444This is a quite general condition for isotropic spaces and, as we will show later, this is also the case for an isotropic LC background. Hence, at $\mathcal{O}(\delta)$ in perturbation theory, we get that $\hat{A}_{ik}\hat{A}_{j}^{k}\sim\mathcal{O}(\delta^{2})$ and then Eq. (2.20), with first of Eqs. (2.19), becomes $\frac{d}{d\lambda}\hat{A}_{ij}=-\frac{d\Xi}{d\lambda}\hat{A}_{ij}+\mathcal{O}\left(\delta^{2},\epsilon^{2}\right)\,.$ (2.21) The latter equation is clearly solved by $\hat{A}_{ij}\propto e^{-\Xi}\,,$ (2.22) and proves that the $\mathcal{O}(1)$ term in the gradient expansion of $\hat{A}_{ij}$ decays when $\Xi$ grows in time. Then it can be neglected. Therefore, we obtain that $\hat{A}_{ij}$ is at least first order in $\epsilon$. As a consequence, $\hat{A}_{ik}\hat{A}_{j}^{k}$ is not only second order in $\delta$ expansion but also at least $\mathcal{O}\left(\epsilon^{2}\right)$. Thanks to this result, this proof can be repeated iteratively to any order $n$-th in perturbation theory, leading then to $\frac{d}{d\lambda}\hat{A}_{ij}=-\frac{d\Xi}{d\lambda}\hat{A}_{ij}+\mathcal{O}\left(\delta^{n+1},\epsilon^{2}\right)\,.$ (2.23) The solution in Eq. (2.22) solves also Eq. (2.23) and this proves that also the term $\hat{A}_{ij}$, which is $\mathcal{O}(\epsilon)$, is decaying and can be neglected. So we have proven our initial claim that $\hat{A}_{ij}$ is at least of order $\epsilon^{2}$. Hence, considering the evolution of the spatial metric in Eqs. (2.19), we have that $\frac{d}{d\lambda}\hat{f}_{ij}=\mathcal{O}\left(\epsilon^{2}\right)\,,$ (2.24) at any order in perturbation theory. The above analysis was performed in [15] leading also to $N^{i}=\mathcal{O}\left(\epsilon\right)$ when $N^{i}$ vanishes at the background. For the sake of clarity, we underline that in [15] the UCG has been fixed and then $\hat{f}_{ij}$ corresponds to the tensor modes. This also shows that the evolution of the tensor modes can be neglected at linear order in $\epsilon$. Finally, our complete set of equations to order $\mathcal{O}\left(\delta^{n+1},\epsilon^{2}\right)$ is given by the energy and momentum constraints $E=\frac{\Theta_{n}^{2}}{3}\,,\hskip 85.35826ptp_{i}=\frac{2}{3}D_{i}\Theta_{n}\,,$ (2.25) with their respective evolution equations given by $\frac{dE}{d\lambda}=-\Theta_{n}\left(E+\frac{1}{3}S\right)\,,\hskip 56.9055pt\frac{dp_{i}}{d\lambda}=-\frac{1}{3\mathcal{M}}D_{i}\left(\mathcal{M}S\right)-\Theta_{n}p_{i}\,.$ (2.26) In order to complete our set of SU equations we also need the decomposed spatial metric evolution given by Eqs. (2.19) and (2.24). Also we have that the expansion rate evolution is given by $\displaystyle\frac{d\Theta_{n}}{d\lambda}=$ $\displaystyle-\frac{1}{3}\Theta_{n}^{2}-\frac{1}{2}\left(S+E\right)\,.$ (2.27) As one can easily see, Eqs. (2.19) and (2.24)-(2.27), valid at first order in the gradient expansion and to all orders in perturbation theory, exactly correspond to the homogeneous and isotropic background equations if one neglects the momentum constraint. These then prove that the condition $\left(\partial_{t}-\mathcal{L}_{N}\right)f_{ij}=\frac{1}{\mathcal{M}}\frac{d}{d\lambda}f_{ij}+\mathcal{O}\left(\delta^{n+1},\,\epsilon^{2}\right)\,,$ (2.28) with the fact that $\partial_{i}N^{i},\,\hat{\mathcal{S}}_{ij},\,R_{ij}$ and $R$ are $\mathcal{O}\left(\delta^{n+1},\,\epsilon^{2}\right)$, reproduces the SU picture, and matches with previous works [15, 5, 4] if $\lambda=t$ and $N^{i}\partial_{i}=\mathcal{O}(\epsilon^{2})$. In the next section, we will specialize this construction to non-linear LC perturbations on the top of a FLRW background. Thanks to the freedom of the choice of the lapse function in the LC gauge, we will then provide a general SU formalism. Furthermore, within the synchronous fixing of the lapse function, our formalism will be extended also to the GLC gauge. ## 3 Light-Cone gauge Let us now introduce the LC gauge [44]. This is a generalization of the GLC gauge [29], where the lapse function is left unfixed and is built as a foliation of the spacetime thanks to a set of four coordinates adapted to the observed past light-cone. In particular, the proper time of a generic observer is described by the coordinate $t$. This corresponds to the proper-time of a free-falling observer when the GLC fixing of the lapse function occurs. The coordinates $w$ and $\theta^{a}$ satisfy the same properties as in the GLC gauge, i.e. $w$ describes the observer’s past light-cone and $\theta^{a}=const$ describes the light-like geodesics. Given that, the non- linear line element is [44] $ds_{LC}^{2}=\Upsilon^{2}dw^{2}-2\mathcal{M}\Upsilon dwdt+\gamma_{ab}\left(d\theta^{a}-U^{a}dw\right)\left(d\theta^{b}-U^{b}dw\right)\,.$ (3.1) In this case, the vector $n^{\mu}$ in Eq. (2.2) is given by $n^{\mu}=\left(\frac{1}{\mathcal{M}},\,\frac{1}{\Upsilon},\,\frac{U^{a}}{\Upsilon}\right)\,.$ (3.2) The advantage of the metric (3.1) is that it simplifies the description of light-like signal. For instance, the light-like geodesics are exactly solved by $k_{\mu}=-\omega\delta_{\mu}^{w}$, where $k_{\mu}$ is the four-momentum of the photon and $\omega$ is its physical frequency. Moreover, for the GLC gauge where $\mathcal{M}=1$, also the time-like geodesic is exactly solved by $u_{\mu}=-\partial_{\mu}t$. In this case $u_{\mu}=n_{\mu}$ is the four- velocity of the geodesic observer and is perpendicular to the three- dimensional hypersurfaces of $t=const$. This particular choice simplifies the description of late-time cosmological observables and allows a completely non- linear description of such observables as a factorization of the metric entries [29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40]. As relevant examples, the expressions for the cosmological redshift and the angular distance are given, in a exact way and for an arbitrary geometry, directly as [29, 33] $1+z=\frac{\left(u_{\mu}k^{\mu}\right)_{s}}{\left(u_{\nu}k^{\nu}\right)_{o}}=\frac{\Upsilon_{o}}{\Upsilon_{s}}\,,\qquad\qquad d^{2}_{A}=\frac{\sqrt{\gamma}}{\left(\frac{\text{det}\partial_{\tau}\gamma_{ab}}{4\sqrt{\gamma}}\right)_{o}}\,,$ (3.3) where $z$ is the redshift of the source, the subscript $s$ and $o$ stands for a quantity evaluated at the source and observer position, and $\gamma$ is the determinant of $\gamma_{ab}$. ### 3.1 LC gauge shift vector Here, we will use $N_{LC}^{i}$ (instead of $N^{i}$)555For the LC metric, Latin indices will always refer to the coordinates $w$ and $\theta^{a}$. to describe the shift vector of the LC gauge, thus avoiding confusion when we relate $N_{LC}^{i}$ to the standard $N^{i}$. Hereafter, we will provide a SU picture allowing for different lapse function fixings within the LC gauge [44]. This is a crucial step to obtain the $\delta\mathcal{N}$ formalism on the past light-cone. In [44] was shown that the non-linear LC gauge can be interpreted as a $1+1+2$ ADM decomposition with coordinates $x^{\mu}=\left(t,w,\theta^{a}\right)$. This proviso, the shift vector for the first $1+3$ decomposition is then given by $N_{LC}^{i}=-\mathcal{M}\left(\frac{1}{\Upsilon},\frac{U^{a}}{\Upsilon}\right)\,,\qquad\qquad N^{LC}_{j}=-\Upsilon\delta_{j}^{w}\,.$ (3.4) For what concerns the shift vector $N^{i}_{LC}$, this is orthogonal to the surface at constant $t$ and $w$. Hence, if we recall that the photon four- momentum in the LC coordinates is $k^{\mu}=\omega\mathcal{M}^{-1}\Upsilon^{-1}\delta^{\mu}_{t}$, within the 1+3 decomposition we can write the shift vector as $N^{i}_{LC}=\frac{k^{i}}{\omega}-n^{i}\,.$ (3.5) Hence, since $n^{\mu}$ is a time-like vector, $N^{i}_{LC}$ can be interpreted as the space-like component of the propagation direction of an incoming photon (see [46]). To completely fix the LC gauge, we still have to fix three conditions. These are given by the following ones $f_{ww}=\Upsilon^{2}+U^{2}\,,\qquad\qquad f_{wa}=-U_{a}\,,\qquad\qquad f_{ab}=\gamma_{ab}\,.$ (3.6) As one can see, $N_{LC}^{w}=\frac{1}{a}$ on the background level, so it does not vanish when $\epsilon\rightarrow 0$. We then have that $N_{LC}^{w}=\mathcal{O}\left(1\right)$. With this choice of coordinates, however, we will show that the condition $\partial_{i}N_{LC}^{i}=\mathcal{O}\left(\epsilon^{2}\right)$ holds. In order to do so, we first perform a finite background coordinate transformation on the metric in Eq. (2.1), from the $x^{i}=\left(r,\theta^{a}\right)$ coordinates to the light-cone ones $y^{i}=\left(w,\theta^{a}\right)$, given by $dr=dw-\frac{dt}{a}\,,\qquad\qquad d\theta^{a}=d\theta^{a}\,,\qquad\qquad dt=dt\,,$ (3.7) in order to relate $N^{i}$ to $N_{LC}^{i}$. A direct computation for the controvariant components of $N^{i}$ returns that $N^{r}=\frac{1}{a}-\frac{\mathcal{M}}{\Upsilon}\,,\qquad\qquad N^{a}=-\frac{U^{a}}{a}\left(1-\frac{1}{a}+\frac{\mathcal{M}}{\Upsilon}\right)\,,$ (3.8) or, equivalently, in a covariant form $N_{r}=-\mathcal{M}\Upsilon+\frac{\Upsilon^{2}+U^{2}}{a}\,,$ (3.9) and $N_{a}=-\frac{U_{a}}{a}\,.$ (3.10) Finally, using Eqs. (3.4) and (3.8), the gradient expansion condition of Eq. (2.18) given by $N^{i}=\mathcal{O}(\epsilon)$ returns $\partial_{i}N^{i}=\partial_{i}N_{LC}^{i}=\mathcal{O}\left(\epsilon^{2}\right)$. Within the gradient expansion, we then have $\displaystyle\partial_{r}N^{r}=$ $\displaystyle-\partial_{w}\left(\frac{\mathcal{M}}{\Upsilon}\right)=\mathcal{O}\left(\epsilon^{2}\right)\,,$ $\displaystyle\partial_{a}N^{a}=$ $\displaystyle-\frac{1}{a}\left(1-N^{r}\right)\partial_{a}U^{a}+U^{a}\partial_{a}N^{r}=\mathcal{O}\left(\epsilon^{2}\right)\,,$ (3.11) which indeed show that $\partial_{w}\Upsilon^{-1}\sim\partial_{w}\mathcal{M}\sim\partial_{a}U^{a}=\mathcal{O}\left(\epsilon^{2}\right)$. This comes from the fact that both $\mathcal{M}$ and $\Upsilon$ have background counterparts, therefore, both $\partial_{w}\Upsilon^{-1}$ and $a^{-1}\partial_{w}\mathcal{M}$ are of order $\epsilon^{2}$. The condition $\partial_{a}U^{a}=\mathcal{O}(\epsilon^{2})$ is obtained by using the fact that $N^{r}=\mathcal{O}(\epsilon)$ on the second of Eqs. (3.11). ### 3.2 Separate Light-Cones As it has been shown in Sect. 2.1, one can still obtain a SU picture when the shift vector combines to form an integral along the geodesics and only its spatial derivatives are neglected. We will prove that this is the case for $\partial_{i}N^{i}$ in Eq. (2.18) and for $\partial_{i}N_{LC}^{i}$ in Eq. (3.11). Additionally, we need the trace-less part of the extrinsic curvature $\hat{A}_{ij}$ to be negligible in order to obtain the SU scheme for the LC metric. The condition for the shift vector is given by $\partial_{i}N_{LC}^{i}=\partial_{w}\left(\frac{\mathcal{M}}{\Upsilon}\right)+\partial_{a}\left(\frac{\mathcal{M}U^{a}}{\Upsilon}\right)=\mathcal{O}\left(\epsilon^{2}\right)\,,$ (3.12) which, with Eqs. (2.9), implies $\frac{d}{d\lambda}f_{ij}=2K_{ij}+\mathcal{O}\left(\epsilon^{2}\right)\,,$ (3.13) where in $\mathcal{L}_{N}f_{ij}$ we have neglected $\partial_{i}N^{i}$ but not $N^{i}\partial_{i}f_{jk}$. In fact, the latter combines with $\partial_{t}f_{ij}$ to reconstruct $\frac{d}{d\lambda}f_{ij}$, following the general prescription given in Eq. (2.16). One may note the similarity between Eqs. (3.13) and (2.9). This is because Eq. (2.9) is the non-perturbative version in the gradient expansion of Eq. (3.13), where we consider instead the gradient expansion on the parameter $\frac{d}{d\tilde{\lambda}}=\frac{d}{d\lambda}+\mathcal{O}(\epsilon^{2})$. Using Eqs. (3.6) at the background level, i.e. $f_{ww}=a^{2}\,,\qquad\qquad f_{wa}=0\,,\qquad\qquad f_{ab}=a^{2}r^{2}\bar{q}_{ab}\,,$ (3.14) one gets that $df_{ij}/d\lambda=2Hf_{ij}$, where $H$ is the background expansion rate defined as $H\equiv\bar{\Theta}_{n}/3$, $\bar{\Theta}_{n}$ is the extrinsic curvature on the background and $\bar{q}_{ab}=\text{diag}\left(1,\sin\theta\right)$. Thus, from Eq. (3.13), we also see that $\hat{A}_{ij}$ vanishes on the background. Following the same procedure adopted in Eqs. (2.21) and (2.23), we get that $\hat{A}_{ij}=\mathcal{O}\left(\epsilon^{2}\right)$ also when the LC gauge is fixed. Now, by taking the trace of Eq. (3.13), we obtain $\displaystyle\Theta_{n}$ $\displaystyle=\,\frac{1}{\sqrt{-g}}\partial_{\mu}\left(\sqrt{-g}n^{\mu}\right)$ $\displaystyle=\,\frac{1}{\mathcal{M}\Upsilon\sqrt{\gamma}}\frac{d}{d\lambda}\left(\mathcal{M}\Upsilon\sqrt{\gamma}\right)+\partial_{\mu}n^{\mu}$ $\displaystyle=\,\frac{1}{\Upsilon\sqrt{\gamma}}\frac{d}{d\lambda}\left(\Upsilon\sqrt{\gamma}\right)+\frac{1}{\mathcal{M}}\frac{d}{d\lambda}\mathcal{M}+\partial_{\mu}n^{\mu}$ $\displaystyle=\,\frac{1}{\Upsilon\sqrt{\gamma}}\frac{d}{d\lambda}\left(\Upsilon\sqrt{\gamma}\right)+\frac{1}{\mathcal{M}}\left[\partial_{w}\left(\frac{\mathcal{M}}{\Upsilon}\right)+\partial_{a}\left(\frac{\mathcal{M}U^{a}}{\Upsilon}\right)\right]$ $\displaystyle=\,\frac{1}{\Upsilon\sqrt{\gamma}}\frac{d}{d\lambda}\left(\Upsilon\sqrt{\gamma}\right)-\frac{1}{\mathcal{M}}\partial_{i}N_{LC}^{i}\,.$ (3.15) Thanks to this last equation, within the fixing $\mathcal{M}=1$, we realize that the difference between $\Theta_{u}\equiv\nabla_{\mu}u^{\mu}$ and $\Theta_{n}$ is of order $\epsilon^{2}$. We also have from Eq. (2.19), where we neglect $\partial_{i}N^{i}_{LC}$ in $\Theta_{n}$ $\frac{d\,\Xi}{d\lambda}=\frac{1}{3\Upsilon\sqrt{\gamma}}\frac{d\left(\Upsilon\sqrt{\gamma}\right)}{d\lambda}+\mathcal{O}\left(\epsilon^{2}\right)\,,$ (3.16) which preserves the background form. Considering now the homogeneous and isotropic LC background, namely $\bar{\Upsilon}\sqrt{\bar{\gamma}}=a^{3}r^{2}$, and also identifying $\bar{E}\equiv\bar{\rho}$ and $\frac{1}{3}\bar{S}\equiv\bar{p}$, Eqs. (2.25) return $\displaystyle\frac{\bar{\Theta}_{n}^{2}}{3}=3H^{2}=3\left(\frac{\partial_{t}a}{a}\right)^{2}=\bar{\rho}\,,\hskip 170.71652pt\text{background}$ $\displaystyle\frac{{\Theta_{n}}^{2}}{3}=\frac{1}{3}\left[\frac{1}{(\Upsilon\sqrt{\gamma})}\frac{d\left(\Upsilon\sqrt{\gamma}\right)}{d\lambda}\right]^{2}=\rho+\mathcal{O}\left(\delta^{n},\epsilon^{2}\right)\,.\hskip 56.9055pt\text{non-perturbative}$ (3.17) Moreover, Eqs. (2.26) and (2.27) give at the background level $\partial_{t}\bar{\rho}=\,-3H\left(\bar{\rho}+\bar{p}\right)\,,\qquad\qquad\partial_{t}\bar{\Xi}=\,H\,,\qquad\qquad\partial_{t}H=\,-3H^{2}-\frac{1}{2}\left(\bar{\rho}+3\bar{p}\right)\,.$ (3.18) Here we remark that Eqs. (3.17) mean that the non-linear LC perturbations in a FLRW universe at first order in the gradient expansion do evolve as a set of glued background universes with a different set of $\left(a,\,\rho,\,p\right)$ in each patch. Within this picture, then $\Upsilon\sqrt{\gamma}$ can be linked to the effective local scale factor. To complete the set of equations that have the same form for the background (last of Eq. (3.18)) and the perturbed universe, we also have from Eq. (2.27) $\frac{d}{d\lambda}\left[\frac{1}{(\Upsilon\sqrt{\gamma})}\frac{d(\Upsilon\sqrt{\gamma})}{d\lambda}\right]=-\frac{1}{3}\left[\frac{1}{(\Upsilon\sqrt{\gamma})}\frac{d(\Upsilon\sqrt{\gamma})}{d\lambda}\right]^{2}-\frac{1}{2}\left(3p+\rho\right)+\mathcal{O}\left(\delta^{n},\epsilon^{2}\right)\,.$ (3.19) Therefore, the SU scheme holds with the universe evolving as a set of homogeneous and isotropic LC background, where $\lambda$ provides the evolution of inhomogeneities along $n^{\mu}$. So far we have provided a consistent SU on the past light-cone in terms of LC gauge entries, which is a fundamental step to provide the $\delta\mathcal{N}$ formalism on the past light-cone in the next Sect. 4. ### 3.3 The Geodesic Light-Cone gauge Let us now apply the SU scheme described in the previous subsection to the case of the GLC gauge given by Eq. (3.1) with $\mathcal{M}=1$, and show how it simplifies the evolution of the density perturbations on the past light-cone. This gauge automatically provides $\nabla_{\mu}u^{\mu}$=$\nabla_{\mu}n^{\mu}$, then, one can relate the expansion of the 3D-hypersurfaces orthogonal to $n^{\mu}$ to the matter content described in terms of $u^{\mu}$. Hence, we recall that the comoving four velocity is given by $u^{\mu}=\left(1,\Upsilon^{-1},\Upsilon^{-1}U^{a}\right)$ [46, 29], we then have that the expansion of the 3D-hypersurfaces is $\Theta_{u}=\nabla_{\mu}u^{\mu}=\frac{\partial_{t}\Upsilon}{\Upsilon}+\frac{\gamma^{ab}\partial_{t}\gamma_{ab}}{2}+\frac{\gamma^{ab}}{2\Upsilon}\partial_{w}\gamma_{ab}+\frac{1}{\Upsilon}\partial_{a}U^{a}+\frac{U^{a}\gamma^{bc}}{2\Upsilon}\partial_{a}\gamma_{bc}\,,$ (3.20) which can be re-written in a more suitable form as $\Theta_{u}=\frac{1}{\Upsilon\sqrt{\gamma}}\frac{d\left(\Upsilon\sqrt{\gamma}\right)}{d\lambda}+\partial_{\mu}u^{\mu}\,,$ (3.21) where $\frac{d}{d\lambda}\equiv u^{\mu}\partial_{\mu}$ accounts for inhomogeneities along the geodesics and $\partial_{\mu}u^{\mu}=\partial_{i}N^{i}$ is $\mathcal{O}(\epsilon^{2})$, as shown in Eqs. (3.11). An interesting feature of Eq. (3.21) is that the first term contributes both to the background and to the perturbative level whereas the last term contributes only to the perturbative level. We thus obtain a separate universe description using Eq. (3.21) as aimed. In order to provide the conservation of $\zeta$, when the pressure is adiabatic, we need to analyze the energy-momentum conservation in the GLC gauge for the case of a perfect fluid. Starting from $T_{\mu\nu}=\left(\rho+p\right)u_{\mu}u_{\nu}+g_{\mu\nu}p\,,$ (3.22) the conservation law along the direction of $u^{\nu}$, i.e. $u^{\nu}\nabla_{\mu}T_{\,\nu}^{\mu}=0$, exactly returns $\frac{d\rho}{d\lambda}=-\left(\rho+p\right)\Theta_{u}\,.$ (3.23) Hence, by using Eq. (3.21), we have $\displaystyle\frac{d\rho}{d\lambda}=-\left(\rho+p\right)\left[\frac{1}{\Upsilon\sqrt{\gamma}}\frac{d\left(\Upsilon\sqrt{\gamma}\right)}{d\lambda}+\partial_{\mu}u^{\mu}\right]\,.$ (3.24) Eq. (3.24) is a fully non-linear relation between geometry and matter content. As an important remark, since Eq. (3.24) is a fully non-linear equation, it can be seen as a dynamic equation for the exact density perturbations. According to what outlined so far, within the gradient expansion, where $\partial_{\mu}u^{\mu}=\partial_{i}N^{i}=\mathcal{O}\left(\epsilon^{2}\right)$, Eq. (3.24) can then be written as $\displaystyle\frac{d\rho}{d\lambda}=$ $\displaystyle-\left(\rho+p\right)\frac{1}{\Upsilon\sqrt{\gamma}}\frac{d\left(\Upsilon\sqrt{\gamma}\right)}{d\lambda}+\mathcal{O}(\epsilon^{2})\,.$ (3.25) In general, pressure and energy density are linked by an equation of state as $p=q\rho$. The value of $q$ can be time-dependent, accordingly to the specific era when inhomogeneities are evolving (as happens, for instance, during the slow-roll inflationary stage). This makes Eq. (3.25) in general quite complicated to be solved. However, during the late time epochs (e.g. radiation, matter or cosmological constant dominated universe), $q$ is constant. In this case, Eqs. (3.25) becomes $\displaystyle\frac{d\rho}{d\lambda}=$ $\displaystyle-\rho\frac{1+q}{\Upsilon\sqrt{\gamma}}\frac{d\left(\Upsilon\sqrt{\gamma}\right)}{d\lambda}+\mathcal{O}(\epsilon^{2})\,,$ (3.26) which is exactly solved by $\rho(\lambda)=A\left(\Upsilon\sqrt{\gamma}\right)^{-(1+q)}(\lambda)+\mathcal{O}(\epsilon^{2})\,,$ (3.27) where $A$ is a constant. Eq. (3.27) gives the exact link between the geometry and the energy density in terms of light-cone metric entries. It is also the starting point to describe non-linear features of inhomogeneities on super- horizon scales. In fact, in the following we will discuss how $\Upsilon\sqrt{\gamma}$ relates with the gauge invariant curvature perturbation $\zeta$ and how both can be computed using the $\delta\mathcal{N}$ formalism. Furthermore, we will show under which conditions the quantity $\zeta$ is conserved on super-horizon scales. ## 4 Curvature perturbation evolution In this section we will study the curvature perturbations along the past light-cone using the SU scheme developed in the previous section. To this aim, we will start by considering the linear order in perturbation theory, and by using the scalar-pseudoscalar decomposition developed in [36, 42]. Hence, we will first review this perturbation theory, following the approach of [4, 5] to show the conservation of $\zeta$ along the past light-cone. Thereafter, we red will generalize this proof to non-linear order in the amplitude of the perturbations and to first order in the gradient expansion, which allow us to obtain $\zeta$ at this perturbative level in terms of light-cone perturbations. Finally, we will generalize the $\delta\mathcal{N}$ formalism on the past light-cone. ### 4.1 Linear evolution and comparison with previous results In this section, we want to linearize Eqs. (3.26). To this aim, we first recall the linear perturbation theory for the GLC coordinates, presented in [36, 42]. Firstly, we consider general perturbations, i.e. without fixing the GLC gauge. The metric and its perturbed inverse are then given by $g_{\mu\nu}=\bar{g}^{\,GLC}_{\mu\nu}+\delta g_{\mu\nu}=a^{2}\left\\{\begin{pmatrix}0&-a^{-1}&\vec{0}\\\ -a^{-1}&1&\vec{0}\\\ \vec{0}^{\mathbf{T}}&\vec{0}^{\mathbf{T}}&\bar{\gamma}_{ab}\end{pmatrix}+\begin{pmatrix}L&M&V_{b}\\\ M&N&\mathcal{U}_{b}\\\ V_{a}^{\mathbf{T}}&\mathcal{U}_{a}^{\mathbf{T}}&\delta\gamma_{ab}\end{pmatrix}\right\\}\,,$ (4.1) and $\delta g^{\mu\nu}=\begin{pmatrix}-\left(a^{2}L+N+2aM\right)&-a^{-1}\left(a^{2}L+aM\right)&-a\left(aV^{a}+\mathcal{U}^{a}\right)\\\ -a^{-1}\left(a^{2}L+aM\right)&-L&aV^{a}\\\ -a\left(aV^{a}+\mathcal{U}^{a}\right)&aV^{a}&-a^{-2}\delta\gamma^{ab}\end{pmatrix}\,,$ (4.2) where, following [36, 42], we can decompose $\mathcal{U}_{a},V_{a}$ and $\delta\gamma_{ab}$ as $\displaystyle\mathcal{U}_{a}$ $\displaystyle=r^{2}\left(D_{a}u+\tilde{D}_{a}\hat{u}\right)\,,$ $\displaystyle V_{a}$ $\displaystyle=r^{2}\left(D_{a}v+\tilde{D}_{a}\hat{v}\right)\,,$ $\displaystyle\delta\gamma_{ab}$ $\displaystyle=a^{2}r^{2}\left[\left(1+2\nu\right)\bar{q}_{ab}+D_{ab}\mu+\tilde{D}_{ab}\hat{\mu}\right]\,.$ (4.3) Here, $u,\,v,\,\nu$ and $\mu$ are scalars, and $\hat{u},\,\hat{v}$ and $\hat{\mu}$ are pseudoscalar degrees of freedom under spatial rotations. Moreover, the angular derivatives are defined as $D_{ab}=D_{(a}D_{b)}-\frac{1}{2}q_{ab}D^{2}\,,\qquad\qquad\tilde{D}_{ab}=D_{(a}\tilde{D}_{b)}\,,$ (4.4) where $\tilde{D}_{a}=\epsilon_{a}^{b}D_{b}$, and $\epsilon_{a}^{b}$ is the anti-symmetric tensor. We remark that $D_{ab}$ and $\tilde{D}_{ab}$ are trace- less, so that the trace of $\delta\gamma_{ab}$ is given by the trace of $q_{ab}=\left(1+2\nu\right)\bar{q}_{ab}$. At this point, let us also linearize the energy density as $\rho=\bar{\rho}\left(1+\delta\rho\right)\,.$ (4.5) In this expansion, however, we keep both $\bar{\rho}$ and $\delta\rho$ as function of the time-like parameter $\lambda$. By doing this, we then keep track of the perturbations as projected onto the exact time-like geodesic. Starting from Eqs. (4.3), we then obtain that $\sqrt{\gamma}=a^{2}\sqrt{\bar{\gamma}}\,\left(1+2\nu\right)\,.$ (4.6) At this point, we want to use the metric in Eq. (4.1) to compute the expansion volume $\Theta_{n}=\nabla_{\mu}n^{\mu}$ of the hypersurfaces orthogonal to $t$ defined in Eq. (2.2). Then, the volume expansion will be given by $\Theta_{n}=\frac{1}{\sqrt{-g}}\partial_{\mu}\left(\sqrt{-g}\,n^{\mu}\right)=\frac{1}{2}g^{\alpha\beta}\frac{d}{d\lambda}g_{\alpha\beta}+\partial_{\mu}n^{\mu}\,,$ (4.7) where $n^{\mu}\partial_{\mu}=\frac{d}{d\lambda}$. Using this last equation, and the fact that $\Theta_{n}=\Theta_{u}+\mathcal{O}(\epsilon^{2})$, as shown in Eq. (3.15), we then have $\Theta_{u}=3H\left[1+\frac{1}{2}\left(a^{2}L+N+2aM\right)\right]+\frac{1}{2}\frac{d}{d\lambda}\left(N+4\nu\right)+\mathcal{O}(\epsilon^{2})\,,$ (4.8) where we are neglecting the terms $\left(\partial_{w}+\frac{2}{r}\right)\left(N+aM\right)=\mathcal{O}\left(\epsilon^{2}\right)\,,\qquad\qquad D_{a}\left(\mathcal{U}^{a}+aV^{a}\right)=\mathcal{O}\left(\epsilon^{2}\right)\,,$ (4.9) since they are proportional to $\nabla_{i}\mathcal{B}^{i}$ in standard perturbation theory (see [36, 42]). The same happens in the standard approach, see [4, 5]. The reason why $\nabla_{i}\mathcal{B}^{i}$ can be neglect is that, in the language of the gradient expansion applied to the ADM metric, this term is the divergence of the shift vector, hence second order in the gradient expansion. As a side remark, we underline that Eq. (4.7) is given in terms of the cosmic time by $\Theta_{n}=3H-3\partial_{t}\psi+\mathcal{O}(\epsilon^{2})\,,$ (4.10) which is in agreement with [4]. Let us now consider the energy-momentum tensor conservation projected onto $u^{\mu}$. From Eq. (3.23), we have that $\displaystyle\frac{d\rho}{d\lambda}+\left(\rho+p\right)\Theta_{u}=0\,.$ (4.11) Thanks to Eqs. (4.5) and (4.8), and by using the equation of state $p=q\rho$, Eq. (4.11) gives $\displaystyle\frac{d\bar{\rho}}{d\lambda}+3H\bar{\rho}(1+q)=0\,,$ $\displaystyle\frac{d\left(\bar{\rho}\delta\rho\right)}{d\lambda}+\bar{\rho}\left(1+q\right)\left\\{3H\left[\delta\rho+\frac{1}{2}\left(a^{2}L+N+2aM\right)\right]+\frac{1}{2}\frac{d}{d\lambda}\left(N+4\nu\right)\right\\}+\mathcal{O}(\epsilon^{2})=0\,,$ (4.12) for the background and perturbed quantities respectively. Let us now fix the GLC gauge, which satisfies the following conditions [36] $\displaystyle a^{2}L+N+2aM$ $\displaystyle=0\,,$ $\displaystyle\partial_{w}\left(N+aM\right)$ $\displaystyle=\frac{1}{2}\partial_{w}N=\mathcal{O}(\epsilon^{2})\,,$ $\displaystyle\partial_{a}\left(\mathcal{U}^{a}+aV^{a}\right)$ $\displaystyle=\partial_{a}\mathcal{U}^{a}=\mathcal{O}\left(\epsilon^{2}\right)\,.$ (4.13) Then, by inserting first of Eqs. (4.12) into the second one, by using the GLC gauge conditions given in Eqs. (4.13), and, finally, by integrating over the affine parameter $\lambda$, we get $\delta\rho(\lambda)=-\frac{1+q}{2}\left(N+4\nu\right)\,.$ (4.14) It is worth to stress that we would obtain the same result by linearizing Eqs. (3.27). Therefore, this show the self-consistency of our gradient expansion method. At first order in the gradient expansion, for a generic adiabatic equation of state, i.e. $p=p(\rho)$, and by fixing the GLC gauge, we obtain, from Eqs. (4.8), (4.11) and (4.13), that $\frac{1}{\rho+p(\rho)}\frac{d\rho}{d\lambda}=-3H-\frac{1}{2}\frac{d\left(N+4\nu\right)}{d\lambda}\,.$ (4.15) This equation can be integrated between two different hypersurfaces marked by the values $\lambda_{1}$ and $\lambda_{2}$, and gives $\frac{1}{6}\left(N+4\nu\right)|_{\lambda_{1}}^{\lambda_{2}}+\bar{\mathcal{N}}(\lambda_{2},\lambda_{1})=-\frac{1}{3}\int_{\rho(\lambda_{1},x^{i})}^{\rho(\lambda_{2},x^{i})}\frac{d\rho}{\rho+p(\rho)}\,,$ (4.16) where the dependence on $x^{i}$ in the extremes of integration denotes that we are considering inhomogeneous hyper-surfaces. At the same time, we have defined $\bar{\mathcal{N}}$ as $\displaystyle\bar{\mathcal{N}}\left(\lambda_{2},\,\lambda_{1}\right)\equiv\int^{\lambda_{2}}_{\lambda_{1}}Hd\lambda$ (4.17) Moreover, as done for standard perturbations in [5], we want to extract the background contribution of Eq. (4.16). Therefore, we start by doing this in the r.h.s. of Eq. (4.16), which can then be written as $\int_{\rho(\lambda_{1},x^{i})}^{\rho(\lambda_{2},x^{i})}\frac{d\rho}{\rho+p(\rho)}=\int_{\bar{\rho}(\lambda_{2})}^{\rho(\lambda_{2},x^{i})}\frac{d\rho}{\rho+p(\rho)}-\int_{\bar{\rho}(\lambda_{1})}^{\rho(\lambda_{1},x^{i})}\frac{d\rho}{\rho+p(\rho)}+\int_{\bar{\rho}(\lambda_{1})}^{\bar{\rho}(\lambda_{2})}\frac{d\rho}{\rho+p(\rho)}\,,$ (4.18) where $\bar{\rho}$ determines the background value of $\rho$. The last term is the background value given by $\bar{\mathcal{N}}$ (see Eq. (4.17)). Therefore, we have $\frac{1}{6}\left(N+4\nu\right)\left(\lambda_{1}\right)+\frac{1}{3}\int_{\bar{\rho}(\lambda_{1})}^{\rho(\lambda_{1},x^{i})}\frac{d\rho}{\rho+p(\rho)}=\frac{1}{6}\left(N+4\nu\right)\left(\lambda_{2}\right)+\frac{1}{3}\int_{\bar{\rho}(\lambda_{2})}^{\rho(\lambda_{2},x^{i})}\frac{d\rho}{\rho+p(\rho)}\,,$ (4.19) then, the quantity $\tilde{\zeta}=-\frac{1}{6}\left(N+4\nu\right)-\frac{1}{3}\int_{\bar{\rho}(\lambda)}^{\rho(\lambda,x^{i})}\frac{d\rho}{\rho+p(\rho)}\,,$ (4.20) is conserved. One interesting aspect of Eq. (4.20) is that the geometrical terms $N$ and $\nu$, present on the conservation of $\tilde{\zeta}$, are precisely the same terms that contribute to the linearized angular distance- redshift relation for the linearized GLC gauge [36]. Let us now link our results to the standard perturbation theory. We follow the notation of [36, 42], where the standard metric is given by $\displaystyle ds^{2}$ $\displaystyle=\,a^{2}\left[-\left(1+2\phi\right)d\eta^{2}-2\mathcal{B}_{i}dx^{i}d\eta+\left(\bar{\gamma}_{ij}+\mathcal{C}_{ij}\right)dx^{i}dx^{j}\right]\,,$ (4.21) and the SVT decomposition given by $\displaystyle\mathcal{B}_{i}$ $\displaystyle=B_{i}+\partial_{i}B\,,$ $\displaystyle C_{ij}$ $\displaystyle=\,-2\bar{\gamma}_{ij}\psi+2D_{ij}E+2\nabla_{(i}F_{j)}+2h_{ij}\,,$ (4.22) where $B_{i}$ and $F_{i}$ are divergenceless vectors and $h_{ij}$ is a trace- less and divergenceless tensor. We also have $\displaystyle D_{ij}E=\left(\nabla_{(i}\nabla_{j)}-\bar{\gamma}_{ij}\frac{\Delta_{3}}{3}\right)E\,.$ (4.23) In this case, the trace of $g_{ij}$ is proportional to $-\psi$, and then by using the relation between standard and GLC perturbations given by [36, 42] $\psi=-\frac{1}{6}\left(N+4\nu\right)\,,$ (4.24) we get that $\tilde{\zeta}=\psi-\frac{1}{3}\int_{\bar{\rho}(\lambda)}^{\rho(\lambda,x^{i})}\frac{d\rho}{\rho+p(\rho)}\,,$ (4.25) or, equivalently, with the use of Eq. (4.5), $\tilde{\zeta}=\psi-\frac{1}{3}\int_{\bar{\rho}(\lambda)}^{\rho(1+\delta\rho)}\frac{d\rho}{\rho+p(\rho)}\approx\psi-\frac{1}{3}\frac{\bar{\rho}\delta\rho}{\rho+p(\rho)}\,,$ (4.26) where we have expanded at linear order in the density perturbations in the last equality. Since we are working at first order in the gradient expansion, where the spatial gauge modes occur at the next-to-leading order, and the time gauge mode is fixed, the quantity $\tilde{\zeta}$ given in Eq. (4.20) is gauge invariant. Hence, within this approximation scheme, we may identify it with the curvature perturbation $\zeta$. For the complete expression of $\zeta$ to order $\mathcal{O}(\delta,\epsilon^{n})$, see Eqs. (2.27) of [42], where we also provide its gauge invariance proof in terms of light-cone perturbations. ### 4.2 Non-linear $\zeta$ As we have seen, our non-linear SU approach in the GLC gauge allowed us to obtain Eq. (3.25) by neglecting the last term in Eq. (3.24), which we have shown to correspond to the terms neglected in the standard perturbation theory [4] at linear order. Now, with the aim of going beyond this result, we leave the lapse function $\mathcal{M}$ unspecified and use the SU approach on the light-cone to prove the non-linear conservation of the curvature perturbation in terms of LC parameters. Just as done in the previous subsection, we start from Eq. (3.23) which gives $\displaystyle\Theta_{u}=-\frac{1}{\left(\rho+p\right)}\frac{d\rho}{d\lambda}\,.$ (4.27) Moreover, we consider Eq. (3.15) and the approximation exploited after Eq. (4.7), namely $\Theta_{n}\equiv\nabla_{\mu}n^{\mu}\simeq\Theta_{u}$. We then get that $\displaystyle\frac{1}{\Upsilon\sqrt{\gamma}}\frac{d\left(\Upsilon\sqrt{\gamma}\right)}{d\lambda}=-\frac{1}{\left(\rho+p\right)}\frac{d\rho}{d\lambda}\,.$ (4.28) We now integrate this equation along $\lambda$ and consider that the pressure is adiabatic. In this way, one obtains that $\displaystyle\ln\left[\frac{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{2}}}{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{1}}}\right]=-\int_{\rho\left(\lambda_{1},x^{i}\right)}^{\rho\left(\lambda_{2},x^{i}\right)}\frac{d\rho}{\rho+p\left(\rho\right)}\,.$ (4.29) At this point, the r.h.s. can be manipulated in the same spirit of what done in Eq. (4.18) (see also [4]). By doing so, we get that $\displaystyle\int_{\rho(\lambda_{1},x^{i})}^{\rho(\lambda_{2},x^{i})}\frac{d\rho}{\rho+p(\rho)}$ $\displaystyle=\int_{\bar{\rho}(\lambda_{2})}^{\rho(\lambda_{2},x^{i})}\frac{d\rho}{\rho+p(\rho)}-\int_{\bar{\rho}(\lambda_{1})}^{\rho(\lambda_{1},x^{i})}\frac{d\rho}{\rho+p(\rho)}+\int_{\bar{\rho}(\lambda_{1})}^{\bar{\rho}(\lambda_{2})}\frac{d\rho}{\rho+p(\rho)}$ $\displaystyle=\int_{\bar{\rho}(\lambda_{2})}^{\rho(\lambda_{2},x^{i})}\frac{d\rho}{\rho+p(\rho)}-\int_{\bar{\rho}(\lambda_{1})}^{\rho(\lambda_{1},x^{i})}\frac{d\rho}{\rho+p(\rho)}+\ln\left(\frac{\bar{\Upsilon}\sqrt{\bar{\gamma}}|_{\lambda_{2}}}{\bar{\Upsilon}\sqrt{\bar{\gamma}}|_{\lambda_{1}}}\right)\,,$ (4.30) where, from the first to the second line, we have used Eq. (4.29) at the background level. Now, thanks to Eq. (4.30), we can rewrite Eq. (4.29) as $\displaystyle\ln\left(\frac{\Upsilon\sqrt{{\gamma}}}{\bar{\Upsilon}\sqrt{\bar{\gamma}}}\right)_{\lambda_{2}}+\int_{\bar{\rho}(\lambda_{2})}^{\rho(\lambda_{2},x^{i})}\frac{d\rho}{\rho+p(\rho)}=\ln\left(\frac{\Upsilon\sqrt{\gamma}}{\bar{\Upsilon}\sqrt{\bar{\gamma}}}\right)_{\lambda_{1}}+\int_{\bar{\rho}(\lambda_{1})}^{\rho(\lambda_{1},x^{i})}\frac{d\rho}{\rho+p(\rho)}\,.$ (4.31) This shows that there is a conserved quantity at first order in the gradient expansion. This quantity corresponds to the non-linear curvature perturbation $\zeta$ $\displaystyle\zeta=\frac{1}{3}\ln\left(\frac{\Upsilon\sqrt{{\gamma}}}{\bar{\Upsilon}\sqrt{\bar{\gamma}}}\right)+\frac{1}{3}\int_{\bar{\rho}(t)}^{\rho(t,x^{i})}\frac{d\rho}{\rho+p(\rho)}+\mathcal{O}\left(\epsilon^{2}\right)\,.$ (4.32) which then generalizes the linear result in Eq. (4.20). Thus, the curvature perturbation defined in Eq. (4.32) generalizes at the non- linear level, but at the first order in the gradient expansion, the expression of the gauge invariant curvature perturbation in the LC gauge formalism given in [42]. In fact, the first term corresponds to the curvature perturbations whilst the second term corresponds to density perturbations in the same spirit as done in Eq. (4.25). Finally, we remark that we have obtained Eq. (4.32) without specifying the lapse function $\mathcal{M}$, therefore we still have the freedom to fix the time gauge mode, as we will see better later. ### 4.3 $\delta\mathcal{N}$ formalism on the LC gauge Let us begin by writing explicitly the exact expression for the expansion rate defined by the normal vector $n^{\mu}$ given by Eq. (3.2). This we will be then applied to the evaluation of the number of e-folds using the SU picture of the LC gauge. Such approach will allow us to obtain a generalization of the non-linear $\delta\mathcal{N}$ formalism in terms of LC gauge metric entries. The expansion rate in the LC gauge is given by $\Theta_{n}=\nabla_{\mu}n^{\mu}=\frac{1}{\sqrt{-g}}\partial_{\mu}\left(n^{\mu}\sqrt{-g}\right)=\frac{1}{\mathcal{M}\Upsilon\sqrt{\gamma}}\frac{d}{d\lambda}\left(\mathcal{M}\Upsilon\sqrt{\gamma}\right)+\partial_{\mu}n^{\mu}\,,$ (4.33) where, from Eqs. (3.2) and (3.12), we have that $\partial_{\mu}n^{\mu}=-\frac{\partial_{t}\mathcal{M}}{\mathcal{M}^{2}}-\partial_{i}\left(\frac{N^{i}_{LC}}{\mathcal{M}}\right)=-\frac{\partial_{t}\mathcal{M}}{\mathcal{M}^{2}}+\mathcal{O}\left(\epsilon^{2}\right)\,.$ (4.34) Now, using Eqs. (4.33) and (4.34), we obtain $\Theta_{n}=\frac{1}{\Upsilon\sqrt{\gamma}}\frac{d}{d\lambda}\left(\Upsilon\sqrt{\gamma}\right)+\mathcal{O}\left(\epsilon^{2}\right)\,.$ (4.35) An interesting thing about this result is that it is invariant in form on the past light-cone, i.e. all the dependence on the lapse function is hidden in $n^{\mu}\partial_{\mu}\equiv\frac{d}{d\lambda}$. Let us now integrate Eq. (4.35) to compute the non-linear number of e-folds at first order in the gradient expansion in terms of light-cone entries. We can then easily obtain the following result $\mathcal{N}\left(\lambda_{f},\lambda_{i}\right)\equiv\frac{1}{3}\int_{\lambda_{i}}^{\lambda_{f}}\Theta_{n}d\lambda^{\prime}=\frac{1}{3}\ln\left[\frac{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{f}}}{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{i}}}\right]+\mathcal{O}\left(\epsilon^{2}\right)\,.$ (4.36) Let us note that $\mathcal{N}(\lambda_{f},\,\lambda_{i})$ from Eq. (4.36) is a biscalar, i.e. depends on the gauge fixing both at the initial and final slicing. One possible fixing of the lapse function is given by the uniform curvature light-cone (UCLC) gauge: in this case the effective local scale factor is given by its background value666We will be using the subscript UC to describe the UCLC gauge fixing and the subscript UD to describe the UDLC gauge fixing. $\displaystyle\left(\Upsilon\sqrt{\gamma}\right)_{UC}=\bar{\Upsilon}\sqrt{\bar{\gamma}}\,.$ (4.37) Therefore, if we fix both initial and final slices on the UCLC gauge, the number of e-folds will be given by its background value, as $\mathcal{N}_{UC}\left(\lambda_{f},\,\lambda_{i}\right)=\bar{\mathcal{N}}\left(\lambda_{f},\,\lambda_{i}\right)\,.$ (4.38) Let us also introduce the uniform density light-cone (UDLC) gauge defined by $\displaystyle\rho_{UD}\left(\lambda,\,x^{i}\right)=\bar{\rho}\left(\lambda\right)\,.$ (4.39) Then, within the UDLC gauge, we have $\zeta=\,\frac{1}{3}\ln\left[\frac{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{f\,UD}}}{\left(\bar{\Upsilon}\sqrt{\bar{\gamma}}\right)_{\lambda_{i}}}\right]=\,\frac{1}{3}\ln\left[\frac{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{f\,UD}}}{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{i\,UC}}}\right]\,$ (4.40) where from the first to the second equality we have used Eq. (4.37). We can now compare Eqs. (4.36) and (4.40), and obtain that $\zeta$ can be related to the number of e-folds in the following way $\displaystyle\mathcal{N}\left(\lambda_{f\,UD},\,\lambda_{i\,UC}\right)=$ $\displaystyle\frac{1}{3}\ln\left[\frac{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{f\,UD}}}{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{i\,UC}}}\right]$ $\displaystyle=$ $\displaystyle\frac{1}{3}\ln\left[\frac{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{f\,UD}}}{\left(\bar{\Upsilon}\sqrt{\bar{\gamma}}\right)_{\lambda_{f}}}\frac{\left(\bar{\Upsilon}\sqrt{\bar{\gamma}}\right)_{\lambda_{i}}}{\left(\Upsilon\sqrt{\gamma}\right)_{\lambda_{i\,UC}}}\right]+\bar{\mathcal{N}}\left(\lambda_{f},\lambda_{i}\right)$ $\displaystyle=$ $\displaystyle\frac{1}{3}\ln\left(\frac{\Upsilon\sqrt{\gamma}}{\bar{\Upsilon}\sqrt{\bar{\gamma}}}\right)_{\lambda_{f\,UD}}+\bar{\mathcal{N}}\left(\lambda_{f},\lambda_{i}\right)$ $\displaystyle=$ $\displaystyle-\zeta(\lambda_{f})+\bar{\mathcal{N}}\left(\lambda_{f},\lambda_{i}\right)\,.$ (4.41) Hence, if we define $\delta\mathcal{N}\equiv\,\mathcal{N}\left(\lambda_{f\,UD},\,\lambda_{i\,UC}\right)-\mathcal{N}\left(\lambda_{f\,UC},\,\lambda_{i\,UC}\right)\,,$ (4.42) we straightforwardly get, from (4.38) and (4.41), that $\delta\mathcal{N}=-\zeta$, which is valid at any order in perturbation theory, in agreement with [15]. As a remark, we underline that, while Eq. (4.36) depends directly on the initial and final slice, Eq. (4.42) depends only on the difference of the perturbations of the e-fold number in the UDLC and UCLC on the final slices. Now we will give an example of how this last result can be used to evaluate the power spectrum of $\zeta$ in terms of the LC metric entries. Following the procedure used in [15], we fix the UCLC gauge and adopt the SU approximation. In this way, the density perturbations can be written in terms of the background density with perturbed initial conditions as follows $\displaystyle\rho_{UC}(\mathcal{N}_{UC},\,\mathbf{x})=\bar{\rho}(\mathcal{\bar{N}},\,\varphi_{*}^{A}(x))\,,$ (4.43) where we use Eq. (4.38). Also, $\varphi_{*}^{A}$ is the field content of the underlying inflationary model evaluated just after the horizon exit. The index $A=1,...,d$ refers to the possibility that inflation could happen with $d$ scalar fields. Instead, by considering the UDLC gauge, we would have $\displaystyle\rho_{UD}(\mathcal{N}_{UD},\,\mathbf{x})=\bar{\rho}(\mathcal{N}_{UD})\,.$ (4.44) Since $\rho$ is a scalar, we can then write $\displaystyle\rho^{\prime}(\mathcal{N}^{\prime},\,x^{\prime})=\rho(\mathcal{N},\,x)\,,$ (4.45) which holds between two generic sets of coordinates $\mathcal{N},\,x$ and $\mathcal{N}^{\prime},\,x^{\prime}$ i.e. the value of a scalar function in a given physical point does not depend on the choice of coordinates. Using Eqs. (4.43) and (4.44) on Eq. (4.45), we get $\displaystyle\bar{\rho}(\bar{\mathcal{N}},\,\varphi_{*}^{A}(x))=\bar{\rho}(\mathcal{N}_{UD})\,,$ (4.46) where the $x^{\prime}=x$ from Eq. (4.45) corresponds to the choice of the spatial threading to fix the LC gauge. Also, we choose $\mathcal{N}^{\prime}=\mathcal{N}_{UC}$, and $\mathcal{N}=\mathcal{N}_{UD}$. Then, Eq. (4.46) can be inverted as $\displaystyle\mathcal{N}_{UD}=\bar{\mathcal{N}}(\bar{\rho},\,\varphi_{*}^{A}(\mathbf{x}))\,.$ (4.47) Hence, by expanding the fields in Eq. (4.47) at linear order as $\varphi^{A}_{*}=\bar{\varphi}^{A}_{*}+\delta\varphi^{A}_{*}$, using Eq. (4.42), we obtain $\displaystyle\frac{1}{3}\ln\left[\frac{\left(\Upsilon\sqrt{\gamma}\right)_{UD}}{\left(\Upsilon\sqrt{\gamma}\right)_{UC}}\right]=$ $\displaystyle\mathcal{N}(\bar{\rho},\,\varphi_{*}^{A}(\mathbf{x}))|_{UD}-\bar{\mathcal{N}}(\bar{\rho})|_{UC}$ $\displaystyle=$ $\displaystyle\delta\varphi_{*}^{A}\partial_{A}\bar{\mathcal{N}}+\frac{1}{2}\delta\varphi_{*}^{A}\delta\varphi_{*}^{B}\partial_{A}\partial_{B}\bar{\mathcal{N}}+...\,.$ (4.48) Therefore, given an inflationary model one may link the value of $\varphi_{*}^{A}$ to the left hand side of Eq. (4.48). Altogether, Eq. (4.48) can be a starting point for the evaluation of $f_{NL}$ in terms of light-cone perturbations (see, for example, [13] for an evaluation of $f_{NL}$ using the $\delta\mathcal{N}$ formalism). Therefore, the results presented in Eqs. (4.42) and (4.48) constitute a further step to obtain non- Gaussian predictions on the past light-cone, from the primordial universe and directly in terms of the metric entries. ## 5 Conclusions In this manuscript we have developed a separate universe (SU) description in the non-linear LC gauge. We provide the non-linear conditions to fix the LC and the GLC gauges in terms of standard coordinates on the ADM formalism. The main difference with the previous works [5, 15, 26], where the SU was considered, is that for the LC and GLC gauges we cannot neglect the shift vector, since this contains information about inhomogeneities along the world- line. As an application of our results, and a consistency check, we repeated the procedures of [4] and [5] to prove the super-horizon conservation of the comoving curvature perturbation $\zeta$, when a light-like foliation of spacetime is taken. This conservation has been achieved by neglecting the non- adiabatic pressure within the SU scheme. We then generalize the $\delta\mathcal{N}$ formalism [10, 8, 9, 5, 15, 11, 12, 13], in terms of the combination of the LC metric entries $\Upsilon\sqrt{\gamma}$ within the uniform density light-cone gauge which is one of our most important results. Let us remark that the gradient expansion employed here simplifies the expression of the expansion rate at the non- linear level (see Eq. (4.33)). This could help in simplifying also the perturbative expressions (see, for instance, the one presented in Eq. (6.11) of [43]). The $\delta\mathcal{N}$ formalism provides a procedure to investigate the evolution of the perturbations for different inflationary models. The extension of the $\delta\mathcal{N}$ formalism on the past light-cone, as developed in this manuscript, allows the evaluation of such dynamical evolution directly over the past light-cone. This moves us one step forward to the evaluation of non-linear effects (such as backreaction effects and non- Gaussianities) since the primordial universe until the late-time one along such past light-cone. Indeed, as a future step, we aim to investigate primordial backreaction effects on different expansion rates using the above- mentioned formalism and well-posed averaging procedures on the past light-cone [29, 35]. Finally, for what regards the possible non-Gaussianities associated to any inflationary model, the $\delta\mathcal{N}$ formalism is a very useful tool. In fact, as shown in [13], this formalism provides very simple expressions for $f_{NL}$ in terms of $\mathcal{N}$. In other words, the overall goal of the research program where this work is nested is to obtain a formalism to compute the curvature perturbations at horizon re-entry expressed in terms of light- cone entries. We remark this point since the subsequent evolution of this metric entries could then be compared to late-time expression of cosmological observables as, for instance, the ones presented in Eq. (3.3). This would provide a self-consistent framework entirely given on the light-cone to disentangle the primordial non-Gaussianities from the ones naturally emerging during the non-linear late-time dynamics [47, 48, 49]. ## Acknowledgement GM and MM are supported in part by INFN under the program TAsP (Theoretical Astroparticle Physics). GF acknowledges support by the FCT under the program “Stimulus” with the grant no. CEECIND/04399/2017/CP1387/CT0026 and through the research project with ref. number PTDC/FIS-AST/0054/2021. GF is also member of the Gruppo Nazionale per la Fisica Matematica (GNFM) of the Istituto Nazionale di Alta Matematica (INdAM). ## Appendix A Linear $\delta\mathcal{N}$ formalism on the light-cone In this appendix we derive the $\delta\mathcal{N}$ formalism at linear order in perturbation theory using the framework developed in [36, 42] and reviewed in Sect. 4. In this way we explicitly show the consistency of our results. To begin, we linearize Eq. (4.42) and obtain $\displaystyle\delta\mathcal{N}=$ $\displaystyle\frac{1}{3}\ln\left[\frac{\left(\Upsilon\sqrt{\gamma}\right)_{UD}}{\left(\Upsilon\sqrt{\gamma}\right)_{UC}}\right]+\mathcal{O}(\epsilon^{2})$ $\displaystyle=$ $\displaystyle\frac{1}{3}\ln\left[\frac{\left(\bar{\Upsilon}\sqrt{\bar{\gamma}}\right)(1+\delta\Upsilon)(1+2\nu)_{UD}}{\left(\Upsilon\sqrt{\gamma}\right)_{UC}}\right]+\mathcal{O}(\delta^{2},\epsilon^{2})$ $\displaystyle=$ $\displaystyle\frac{1}{3}\ln\left[1+\left(\delta\Upsilon+2\nu\right)_{UD}\right]+\mathcal{O}(\delta^{2},\epsilon^{2})\,$ $\displaystyle=$ $\displaystyle\frac{1}{3}\left(\delta\Upsilon+2\nu\right)_{UD}+\mathcal{O}(\delta^{2},\epsilon^{2})$ (A.1) where we have defined $\Upsilon=\bar{\Upsilon}(1+\delta\Upsilon)$ (A.2) and we recall that $(\Upsilon\sqrt{\gamma})_{UC}$ is equal to the background value, being $\psi=0$ within the uniform curvature gauge. Also, we have used the metric in Eq. (4.1) and the scalar/pseudoscalar decomposition of Eq. (4.3). Since $\delta\Upsilon=N/2$, we then have that $\delta\mathcal{N}(\lambda_{1,}\lambda_{2,}x^{i})=\frac{1}{6}\left(N+4\nu\right)_{UD}+\mathcal{O}(\delta^{2},\epsilon^{2})\,.$ (A.3) From the relation between the light-cone perturbation and the standard ones in Eq. (4.24), (see also [42]), one gets that $\displaystyle\psi=-\frac{1}{6}(N+4\nu)\,.$ (A.4) Therefore, Eq. (A.3) together with Eq. (A.4) and the fact that $\psi_{UD}=\zeta$, gives the well known relation $\delta\mathcal{N}=-\zeta$ [15]. The result obtained in Eq. (A.3) proves that the $\delta\mathcal{N}$ formalism on the past light-cone is consistent with the light-cone perturbation theory developed in [36, 42]. Thereby, the $\delta\mathcal{N}$ formalism within the past light-cone framework, at linear order in perturbation theory, could also be obtained by starting from the results presented in Sect. 4. To this aim, one should integrate Eqs. (4.7) and (4.8), evaluating them between the uniform curvature and the uniform density slices. ## References * [1] LSST Dark Energy Science Collaboration, A. Abate et al., Large Synoptic Survey Telescope: Dark Energy Science Collaboration, arXiv:1211.0310. * [2] L. Amendola et al., Cosmology and fundamental physics with the Euclid satellite, Living Rev. Rel. 21 (2018), no. 1 2, [arXiv:1606.00180]. * [3] DESI Collaboration, A. Aghamousa et al., The DESI Experiment Part I: Science,Targeting, and Survey Design, arXiv:1611.00036. * [4] D. Wands, K. A. Malik, D. H. Lyth, and A. R. Liddle, A New approach to the evolution of cosmological perturbations on large scales, Phys. Rev. D 62 (2000) 043527, [astro-ph/0003278]. * [5] D. H. Lyth, K. A. Malik, and M. Sasaki, A General proof of the conservation of the curvature perturbation, JCAP 05 (2005) 004, [astro-ph/0411220]. * [6] R.-G. Cai, B. Hu, and H.-B. Zhang, Acoustic signatures in the Cosmic Microwave Background bispectrum from primordial magnetic fields, JCAP 08 (2010) 025, [arXiv:1006.2985]. * [7] E. Komatsu and D. N. Spergel, The Cosmic Microwave Background bispectrum as a test of the physics of inflation and probe of the astrophysics of the low-redshift Universe, in 9th Marcel Grossmann Meeting on Recent Developments in Theoretical and Experimental General Relativity, Gravitation and Relativistic Field Theories (MG 9), 12, 2000. astro-ph/0012197. * [8] M. Sasaki and E. D. Stewart, A General analytic formula for the spectral index of the density perturbations produced during inflation, Prog. Theor. Phys. 95 (1996) 71–78, [astro-ph/9507001]. * [9] M. Sasaki and T. Tanaka, Superhorizon scale dynamics of multiscalar inflation, Prog. Theor. Phys. 99 (1998) 763–782, [gr-qc/9801017]. * [10] D. S. Salopek and J. R. Bond, Nonlinear evolution of long wavelength metric fluctuations in inflationary models, Phys. Rev. D 42 (1990) 3936–3962. * [11] A. A. Starobinsky, Dynamics of Phase Transition in the New Inflationary Universe Scenario and Generation of Perturbations, Phys. Lett. B 117 (1982) 175–178. * [12] A. A. Starobinsky, Multicomponent de Sitter (Inflationary) Stages and the Generation of Perturbations, JETP Lett. 42 (1985) 152–155. * [13] D. H. Lyth and Y. Rodriguez, The Inflationary prediction for primordial non-Gaussianity, Phys. Rev. Lett. 95 (2005) 121302, [astro-ph/0504045]. * [14] R. L. Arnowitt, S. Deser, and C. W. Misner, The Dynamics of general relativity, Gen. Rel. Grav. 40 (2008) 1997–2027, [gr-qc/0405109]. * [15] N. S. Sugiyama, E. Komatsu, and T. Futamase, $\delta$N formalism, Phys. Rev. D 87 (2013), no. 2 023530, [arXiv:1208.1073]. * [16] V. Vennin and A. A. Starobinsky, Correlation Functions in Stochastic Inflation, Eur. Phys. J. C 75 (2015) 413, [arXiv:1506.04732]. * [17] A. A. Starobinsky, STOCHASTIC DE SITTER (INFLATIONARY) STAGE IN THE EARLY UNIVERSE, Lect. Notes Phys. 246 (1986) 107–126. * [18] F. Finelli, G. Marozzi, A. A. Starobinsky, G. P. Vacca, and G. Venturi, Generation of fluctuations during inflation: Comparison of stochastic and field-theoretic approaches, Phys. Rev. D 79 (2009) 044007, [arXiv:0808.1786]. * [19] F. Finelli, G. Marozzi, A. A. Starobinsky, G. P. Vacca, and G. Venturi, Stochastic growth of quantum fluctuations during slow-roll inflation, Phys. Rev. D 82 (2010) 064020, [arXiv:1003.1327]. * [20] T. Prokopec and G. Rigopoulos, $\Delta$N and the stochastic conveyor belt of ultra slow-roll inflation, Phys. Rev. D 104 (2021), no. 8 083505, [arXiv:1910.08487]. * [21] H. Firouzjahi, A. Nassiri-Rad, and M. Noorbala, Stochastic Ultra Slow Roll Inflation, JCAP 01 (2019) 040, [arXiv:1811.02175]. * [22] G. Ballesteros, J. Rey, M. Taoso, and A. Urbano, Stochastic inflationary dynamics beyond slow-roll and consequences for primordial black hole formation, JCAP 08 (2020) 043, [arXiv:2006.14597]. * [23] C. Pattison, V. Vennin, D. Wands, and H. Assadullahi, Ultra-slow-roll inflation with quantum diffusion, arXiv:2101.05741. * [24] M. Biagetti, G. Franciolini, A. Kehagias, and A. Riotto, Primordial Black Holes from Inflation and Quantum Diffusion, JCAP 07 (2018) 032, [arXiv:1804.07124]. * [25] J. M. Ezquiaga and J. García-Bellido, Quantum diffusion beyond slow-roll: implications for primordial black-hole production, JCAP 08 (2018) 018, [arXiv:1805.06731]. * [26] A. Talebian-Ashkezari, N. Ahmadi, and A. A. Abolhasani, $\delta$ M formalism: a new approach to cosmological perturbation theory in anisotropic inflation, JCAP 03 (2018) 001, [arXiv:1609.05893]. * [27] T. Tanaka and Y. Urakawa, Anisotropic separate universe and Weinberg’s adiabatic mode, JCAP 07 (2021) 051, [arXiv:2101.05707]. * [28] T. Tanaka and Y. Urakawa, Statistical anisotropy of primordial gravitational waves from generalized $\delta N$ formalism, arXiv:2309.08497. * [29] M. Gasperini, G. Marozzi, F. Nugier, and G. Veneziano, Light-cone averaging in cosmology: Formalism and applications, JCAP 07 (2011) 008, [arXiv:1104.1167]. * [30] I. Ben-Dayan, M. Gasperini, G. Marozzi, F. Nugier, and G. Veneziano, Backreaction on the luminosity-redshift relation from gauge invariant light-cone averaging, JCAP 04 (2012) 036, [arXiv:1202.1247]. * [31] I. Ben-Dayan, M. Gasperini, G. Marozzi, F. Nugier, and G. Veneziano, Do stochastic inhomogeneities affect dark-energy precision measurements?, Phys. Rev. Lett. 110 (2013), no. 2 021301, [arXiv:1207.1286]. * [32] I. Ben-Dayan, G. Marozzi, F. Nugier, and G. Veneziano, The second-order luminosity-redshift relation in a generic inhomogeneous cosmology, JCAP 11 (2012) 045, [arXiv:1209.4326]. * [33] G. Fanizza, M. Gasperini, G. Marozzi, and G. Veneziano, An exact Jacobi map in the geodesic light-cone gauge, JCAP 11 (2013) 019, [arXiv:1308.4935]. * [34] G. Marozzi, The luminosity distance–redshift relation up to second order in the Poisson gauge with anisotropic stress, Class. Quant. Grav. 32 (2015), no. 4 045004, [arXiv:1406.1135]. [Erratum: Class.Quant.Grav. 32, 179501 (2015)]. * [35] G. Fanizza, M. Gasperini, G. Marozzi, and G. Veneziano, Generalized covariant prescriptions for averaging cosmological observables, JCAP 02 (2020) 017, [arXiv:1911.09469]. * [36] G. Fanizza, G. Marozzi, M. Medeiros, and G. Schiaffino, The Cosmological Perturbation Theory on the Geodesic Light-Cone background, JCAP 02 (2021) 014, [arXiv:2009.14134]. * [37] E. Di Dio, R. Durrer, G. Marozzi, and F. Montanari, Galaxy number counts to second order and their bispectrum, JCAP 12 (2014) 017, [arXiv:1407.0376]. [Erratum: JCAP 06, E01 (2015)]. * [38] E. Di Dio, R. Durrer, G. Marozzi, and F. Montanari, The bispectrum of relativistic galaxy number counts, JCAP 01 (2016) 016, [arXiv:1510.04202]. * [39] G. Marozzi, G. Fanizza, E. Di Dio, and R. Durrer, CMB-lensing beyond the Born approximation, JCAP 09 (2016) 028, [arXiv:1605.08761]. * [40] G. Marozzi, G. Fanizza, E. Di Dio, and R. Durrer, CMB-lensing beyond the leading order: temperature and polarization anisotropies, Phys. Rev. D 98 (2018), no. 2 023535, [arXiv:1612.07263]. * [41] G. Fanizza, M. Gasperini, G. Marozzi, and G. Veneziano, Time of flight of ultra-relativistic particles in a realistic Universe: a viable tool for fundamental physics?, Phys. Lett. B 757 (2016) 505–509, [arXiv:1512.08489]. * [42] G. Fanizza, G. Marozzi, and M. Medeiros, Gauge invariance on the light-cone: curvature perturbations and radiative degrees of freedom, JCAP 06 (2023) 015, [arXiv:2303.11743]. * [43] M. B. Fröb and W. C. C. Lima, Cosmological perturbations and invariant observables in geodesic lightcone coordinates, JCAP 01 (2022), no. 01 034, [arXiv:2108.11960]. * [44] E. Mitsou, G. Fanizza, N. Grimm, and J. Yoo, Cutting out the cosmological middle man: General Relativity in the light-cone coordinates, Class. Quant. Grav. 38 (2021), no. 5 055011, [arXiv:2009.14687]. * [45] C. Tian, M. F. Carney, J. B. Mertens, and G. Starkman, Accurate relativistic observables from post-processing light cone catalogues, arXiv:2110.00893. * [46] P. Fleury, F. Nugier, and G. Fanizza, Geodesic-light-cone coordinates and the Bianchi I spacetime, JCAP 06 (2016) 008, [arXiv:1602.04461]. * [47] A. Kehagias, A. Moradinezhad Dizgah, J. Noreña, H. Perrier, and A. Riotto, A Consistency Relation for the Observed Galaxy Bispectrum and the Local non-Gaussianity from Relativistic Corrections, JCAP 08 (2015) 018, [arXiv:1503.04467]. * [48] E. Di Dio, H. Perrier, R. Durrer, G. Marozzi, A. Moradinezhad Dizgah, J. Noreña, and A. Riotto, Non-Gaussianities due to Relativistic Corrections to the Observed Galaxy Bispectrum, JCAP 03 (2017) 006, [arXiv:1611.03720]. * [49] T. Schiavone, E. Di Dio, and G. Fanizza, The skewness of the distance-redshift relation in $\Lambda$CDM, arXiv:2307.13455.
# Feature Selection Tutorial with Python Examples Pádraig Cunningham School of Computer Science University College Dublin <EMAIL_ADDRESS> &Bahavathy Kathirgamanathan School of Computer Science University College Dublin <EMAIL_ADDRESS> &Sarah Jane Delany School of Computer Science Technological University Dublin <EMAIL_ADDRESS> ###### Abstract In Machine Learning, feature selection entails selecting a subset of the available features in a dataset to use for model development. There are many motivations for feature selection, it may result in better models, it may provide insight into the data and it may deliver economies in data gathering or data processing. For these reasons feature selection has received a lot of attention in data analytics research. In this paper we provide an overview of the main methods and present practical examples with Python implementations. While the main focus is on supervised feature selection techniques, we also cover some feature transformation methods. ## 1 Introduction In data analysis, objects described using multiple features may sometimes be described using a subset of these features without loss of information. Identifying these feature subsets is termed feature selection, variable selection or feature subset selection and is a key process in data analysis. This paper provides a tutorial introduction to the main feature selection methods used in Machine Learning (ML). While excellent reviews and evaluations of feature selection methods already exist [11, 23] our main contribution is to provide examples of these methods in operation along with links to Python notebooks that implement these methods. Feature selection is important because it can deliver a number of benefits: * • Better classifiers: The obvious benefit of feature selection is that it will improve accuracy because redundant or noisy features can damage accuracy. Perhaps surprisingly, improvements in accuracy can be quite limited because powerful ML techniques are designed to be robust against noise. * • Knowledge discovery: Perhaps the most enduring benefit of feature selection is the insight it provides. Identifying influential features and features that are not useful teaches us a lot about the data. * • Data Gathering: In domains where data comes at a cost (e.g. Medical Diagnosis, Manufacturing), identifying a minimal set of features for a classification task can save money. * • Computational Cost: Identifying minimal feature subsets will allow for simpler models that will cost less to set up and run. * • The Curse of Dimensionality: In theory, according to the Curse of Dimensionality, the amount of data required to build a classifier increases exponentially with the number of features. Feature selection is most effective in the context of supervised machine learning (classification/regression) where the availability of labelled examples can drive the selection process. The methods we cover are summarised in Figure 1. Other surveys of feature selection [23, 11] divide feature selection methods into three categories and we follow the same structure: * • Wrappers are feature selection methods where the classifier is _wrapped_ in the feature selection process. This wrapping allows classification performance to drive the feature selection process. This has the advantage of tying the feature selection to classifier performance but this comes with a significant computational cost as very many classifier variants will be evaluated during the selection. * • Filters cover methods that use criteria other than classifier performance to guide feature selection. Typically a filter provides a feature ranking and then a selection policy uses this ranking to select a feature subset. * • Embedded methods refer to any method where the feature selection _emerges_ as a by-product of the classifier training process. For instance, training a decision tree will almost always select a subset of the available features to build a tree. Figure 1: An overview of the feature selection methods covered in this paper. For completeness, Principle Component Analysis and Linear Discriminant Analysis (which are not really feature selection methods) are also covered. For completeness we also cover some dimensionality reduction methods that transform the data into a reduced dimension space rather than select a subset of features. In section 7 we cover Principle Component Analysis (PCA) and Linear Discriminant Analysis (LDA) for projecting data into lower dimension spaces. These are not feature selection methods in the sense that the original feature representation is left behind. Before working through the details of the Feature Selection methods the paper begins with a short discussion on the implications of The Curse of Dimensionality. Then the evaluation methodology used in the paper is presented in Section 3. The main material on Feature Selection methods is covered under the categories of Wrapper methods (Section 4), Filter strategies (Section 5) and Embedded methods (6). Section 7 provides a brief description of PCA and LDA as already mentioned. ## 2 The Curse of Dimensionality Before proceeding it is worth briefly discussing the implications of the Curse of Dimensionality for ML. The term was coined by Richard Bellman in 1961 and refers to a number of phenomena associated with data described by many features [1]. For our purposes there are two theoretical issues. The first is that the number of samples required to build a model may increase exponentially with the number of features (dimensions). The second is that the variation in distance between arbitrary points _decreases_ as more dimensions are added. The first of these issues is illustrated in Figure 2. This figure shows plots of 20 random data points in 1D, 2D and 3D. As the dimension increases the data gets more and more sparse. A consequence of this is that the number of samples required to cover a phenomenon increases exponentially with dimension. This is less of a problem in practice because real data is unlikely to occupy all of the space, instead it will occupy a lower dimension manifold in the space. So the _intrinsic_ dimension of the data is likely to be considerably less than the full dimension [5]. The PCA example in Figure 18 also illustrates how the intrinsic dimension can be less than the number of features. Figure 2: 20 random data points in 1D, 2D and 3D space. These plots show how data becomes increasingly sparse as the dimension is increased. The second issue is that, somewhat paradoxically, the more features used to describe the data the more similar everything appears. The box plots in Figure 3 illustrate this. The plots show the distribution of Cosine similarities between a probe point and 1,000 random points in 5D, 10D and 20D spaces. As the dimension increases the variation in similarity/distance decreases. In addition to these theoretical problems with having more features than is absolutely necessary, there are also practical problems to be considered. For most ML algorithms, the computational cost will increase with the number of features. As will the cost of gathering the data in the first place. So if the number of features can be reduced then there are strong theoretical and practical reasons for doing so. Figure 3: A demonstration of how the variation in Cosine similarity reduces as the dimension of the data increases. ## 3 Methodology The datasets used in the examples in this paper are summarised in Table 1. The Segmentation data set is from the UCI repository111https://archive.ics.uci.edu/, the Penguins dataset was constructed from a larger dataset available on GitHub222https://github.com/allisonhorst/palmerpenguins and the Harry Potter dataset was constructed by the authors based on the popular Top Trumps game333https://toptrumps.us/collections/harry-potter to illustrate PCA in operation. The Segmentation data comes from an image segmentation task on outdoor images. The ground truth segments the images into seven classes and the instances are represented by 19 features. This is a good dataset for feature selection as there is some redundancy in the data. The Penguins dataset is a three class dataset described by four features [10]. The features are physical parameters (e.g. bill and flipper dimensions) and are strongly correlated so there is some redundancy. Table 1: Data Sets: Summary details. Name | Samples | Features | Classes ---|---|---|--- Segmentation | 2,310 | 19 | 7 Penguins | 333 | 4 | 3 Harry Potter | 22 | 5 | - If we wish to get an assessment of generalisation performance for an ML system that might be deployed then we need to hold back some data for testing (option (b) in Figure 4). If we wish to assess a few different feature selection alternatives as part of the model development then these should be tested within the confines of training data and cross validation is the most effective way to do this (option (c)). Most of the testing reported in this paper follows this pattern. Indeed, the evaluation testing may involve two levels of cross validation – this strategy is not used here. It should be remembered that if the objective is to perform feature selection as part of the deployment of an ML system then all the available data can be used for feature selection (option (a) in Figure 4). Figure 4: Evaluation methodology. (a) If an estimation of generalisation accuracy is not required then all data can be used for all aspects of model development. (b) Test data can be held back from training to get an estimate of generalisation accuracy. (c) Cross validation can be used within the training data for Feature Selection. Unless otherwise stated, the classifier used in the evaluations in $k$-Nearest Neighbour ($k$-NN) [5]. $k$-NN is used because it is probably the classifier in popular use most susceptible to noisy or redundant features. So the impact of feature selection will be most evident when it is used in evaluations. Before proceeding we need to introduce the notation that will be used throughout the paper. Assume we have a dataset $D$ made up of $n$ data samples. $D=\left<\mathbf{X},\mathbf{y}\right>$ where $\mathbf{y}$ are the class labels. The examples are described by a set of features $F$ where $p=|F|$ so there are $n$ objects described by $p$ features. So the dimensions are $\mathbf{X}_{n\times p}$ and $\mathbf{y}_{p}$. The objective is to identify a subset $S\subset F$ that captures the important information in the dataset. In supervised ML the classifiers would work with data represented by a reduced set of features $\left<\mathbf{X}^{\prime}_{n\times k},\mathbf{y}\right>$ where $k=|S|$. ## 4 Feature Selection using Wrappers So the objective with feature selection is to identify a feature subset $S\subset F$ to represent the data. If $|F|$ is small we could in theory try out all possible subsets of features and select the best subset. In this case _‘try out’_ would mean training and testing a classifier using the feature subset. This would follow the protocol presented in Figure 4 (c) where cross- validation on the training data would identify a good feature subset and then this could be tested on the test data. However the number of possibilities is $2^{p}$ so exhaustive search quickly becomes impossible. Figure 5: Wrappers versus Filters: (a) With Wrappers the classifier is _wrapped_ in the search process. (b) A Filter strategy uses a separate evaluation to score features Nevertheless this is how a Wrapper feature selection strategy works with the important modification that the search can be greedy or stochastic rather than exhaustive. The general idea is shown in Figure 5(a), the classifier is _wrapped_ in the feature selection process, i.e. classifiers trained using the feature subsets are used in the search process. The feature subsets will be evaluated using hold-out testing or cross-validation testing on classifiers built using the data. The main search strategies used with Wrappers are: * • Exhaustive Search evaluates every possible feature subset. If the number of features to be considered is small it will be possible to consider all feature combinations. However, if $p>20$ there will be millions of feature subsets to considered and an exhaustive search will not be practical. * • Sequential Forward Selection (SFS) starts with no features selected and all classifiers incorporating a single feature are considered (see Figure 6 (a)). The best of these is selected and then two feature combinations including this feature are evaluated. This process proceeds, adding the winning feature at each step, until no further improvements can be made. * • Backward Elimination (BE) proceeds in the opposite direction to FSS, it starts with all features selected, considers the options with one feature deleted, selects the best of these and continues to eliminate features. Again, the process is terminated when no improvements can be made. * • Stochastic Search methods such as genetic algorithms or simulated annealing can readily be applied to Wrapper feature selection. Each state can be defined by a feature mask on which crossover and mutation can be performed [22]. Given this convenient representation, the use of a stochastic search for feature selection is quite straightforward although the evaluation of the fitness function (classifier accuracy as measured by cross-validation) is expensive. Our exploration of Wrappers will focus on SFS and BE. These are greedy strategies that explore the search space of possible feature subsets as shown in Figure 6. SFS starts with an empty set and proceeds forward considering classifiers built on single features. The best of these is selected and then pairs of features incorporating this feature are considered. The process could terminate when the addition of a new feature doesn’t result in any improvement. As the name suggests, Backward Elimination works in the opposite direction. It starts with a full set of features (Figure 6 (b)) and eliminates the least useful feature at each step. For both SFS and BE, the feature subsets are evaluated using cross-validation on the training data. As stated in Section 3, the classifier used is $k$-NN. Figure 6: Examples of feature subset selection using wrappers: (a) Sequential Forward Selection (b) Backward Elimination. Both methods have their own advantages and disadvantages. SFS is inclined to require less computation as the models being evaluated are smaller, typically a classifier with a small number of features will take less time to train and test. SFS is inclined to select less features; this parsinomy is typically an advantage. On the other hand, because BE starts with larger feature sets, it can do a better job of assessing how features work in combination. In the Appendix a link is provided to code to run SFS and BE in Python. The evaluation is on the Segmentation dataset and the results are shown in Figure 7. On the left we see a plot of accuracy on the training set as the SFS proceeds. In this case the search is allowed to run to the end but it is evident that accuracy stops improving after seven features. The overall results for SFS and BE are shown in Figure 7 (b). SFS selects seven features and 11 are selected by BE. Both feature subsets result in improved accuracy on the training data but only the SFS subset results in better accuracy on the test data. Indeed the gap between train and test accuracy for BE is evidence of overfitting – the selection process has fitted too closely to the characteristics of the training data at the cost of generalisation accuracy. Indeed overfitting is recognised to be a problem with Wrapper-based feature selection [22]. (a) (b) Figure 7: Feature selection example using wrappers. (a) Accuracy on the training data as Sequential Forward Selection proceeds measured using cross- validation. (b) Accuracy estimates for feature subsets selected by SFS and BE. SFS selects 7 features and BE selects 11. ## 5 Filter Strategies Figure 5 (a) shows how Wrapper strategies use the classification algorithm in the feature selection process. Figure 5 (b) shows that Filter strategies do not use the classifier for feature selection, instead a separate evaluation function is used. The fact that Filters are independent of the classifier is a mixed blessing. It means that Filters can be much faster than Wrappers but the selected features may not be in tune with the inductive bias of the classifier. In the next subsection we provide some detail on the operation of a basic Filter strategy. Then we cover the Relief Algorithm and Correlation-Based Feature Selection, two Filter strategies that have received a lot of attention in recent years. ### 5.1 Basic Filters A basic Filter will entail a feature scoring mechanism and then a selection strategy based on these scores. The scoring mechanism needs to quantify how much information the feature has about the outcome. The selection strategy might be: * • Select the top ranked $k$ features, * • Select top 50%, * • Select features with scores $>50\%$ of the maximum score, * • Select features with non-zero scores. In this analysis we consider the Chi-square statistic and information gain for scoring. The Chi-square statistic is a measure of independence between a feature and the class label. If samples are organised into a contingency table as shown in Figure 8, how different are the cell counts to what would be observed by chance? The data in Figure 8 (a) suggests that handedness is independent of gender because the proportions are the same. The data in (b) suggests that gender is predictive of handedness. The Chi-square statistic allows us to quantify this: $\chi^{2}=\sum_{i=1}^{m}\frac{(O_{i}-E_{i})^{2}}{E_{i}}$ (1) The statistic is a sum over the $m$ cells. For each cell we consider the difference between the observed count $O_{i}$ and the expected count $E_{i}$ if the feature and the class were independent. In Figure 8 (a) this difference would be zero because the feature and the class are independent. In (b) there would be a difference so the statistic would be positive. In general, the greater the dependence the larger the statistic. If the feature values are numeric rather than categorical then the feature values can be binned to enable the construction of the contingency table [15]. Figure 8: Two contingency tables showing relationships between handedness and gender. If handedness is the class then in (a) it is independent of the gender feature, in (b) there is a dependence. Information gain is an alternative information-theoretic measure quantifying the information a feature contains about a class [16]. In Figure 8 (b) by knowing the gender we _gain_ information about handedness. In a binary classification scenario, let’s assume the probability of a positive and negative outcomes are respectively $p$ and $q$. Then the entropy of a dataset based on these proportions is: $H(D)=-p\log_{2}(p)-q\log_{2}(q)$ (2) then the information gain for any feature $f$ in the dataset in terms of the class label is: $IG(D,f)=H(D)-\sum_{v\in values(f)}\frac{|S_{v}|}{S}H(D_{v})$ (3) As with the Chi-square statistic, information gain (I-Gain) allows us to rank features for the purpose of feature selection. This is illustrated in Figure 9 (a). This shows the Segmentation features ranked by both measures. A link for this code is provided in the Appendix. The plot shows the scores sorted by I-Gain score. It is clear that the scores are well correlated (Pearson correlation score of 0.86) so feature subsets selected based on these scores should be reasonably similar. (a) (b) Figure 9: Feature selection using filters. (a) I-Gain and Chi-square scores for the 19 features in the segmentation dataset. (b) Accuracy estimates for top-$n$ features. This does prove to be the case when we look at the performance of classifiers built with feature subsets based on these rankings. In Figure 9 (b) we see the results of a range of top $k$ selection policies ($k=3,6,10,15$). At $k=10$ both scores select a feature subset that produces accuracy on the test set equivalent to that obtained with the full feature set. The evaluation strategy here conforms to pattern (b) in Figure 4, the feature scoring is done using the training data and then tested on the test set. #### 5.1.1 Combining Filters and Wrappers The results in Figure 9 show that initially performance improves as features are added based on the ranking generated by the filters. However, this improvement tails off and eventually no improvement results from the addition of ‘poorer’ features. Indeed these features may damage performance. This suggests a hybrid Filter/Wrapper strategy whereby a Filter is used to rank the features and then a Wrapper is used to identify the optimum feature subset. This hybrid strategy is shown in operation in Figure 10. The dataset has been split into train and test sets of equal size. The training set is used to estimate I-Gain scores for all features, these are shown in blue. Then classifiers are trained with feature sets of increasing size. The performance of these feature subsets are scored on the training set using cross validation and on the test set using hold-out testing. After the addition of nine features the performance on the training set stops improving (indicated with a green X). This hybrid strategy would select this as the optimum feature subset. Figure 10: A hybrid Wrapper-Filter strategy. Features are ranked using I-Gain and then the performance of subsets based on this ranking is evaluated. The training set accuracy of the best performing feature subset is marked with an X. ### 5.2 Relief Algorithm and Variants The Relief family of algorithms deserve mention because, as Filter methods, they have the advantage of speed while scoring features in the context of other features [17, 18]. Relief algorithms belong to the $k$-NN paradigm in ML. A $k$-NN analysis is used to score or weight features. The idea is to take each sample in the dataset (or a subset) and find its nearest neighbour of the same class and nearest unlike neighbour - these are termed the nearHit and the nearMiss. Then the general principle is: * • For the nearMiss ($nM$) increment the feature weights; weights for unmatching features will be incremented more. * • For the nearHit ($nH$) decrement the feature weights; weights for unmatching features will be decremented more. The idea is that this will pull matching instances closer together and push unmatching instances apart. This is illustrated in the 2D example shown in Figure 11. The weights should be adjusted to bring $x$ and $nH$ closer together while pushing $nM$ away. Comparing $x$ and $nM$, they match more on f2 than on f1 so the weight for f1 will be incremented more. Considering $x$ and $nH$ the opposite happens, f2 is decremented more than f1, so the overall effect is that f1 scores better than f2. This is achieved with the following update function where $x$ is the query, $nM$ is the nearMiss, $nH$ is the nearHit and $w_{f}$ is the weight for feature $f$. $w_{f}\leftarrow w_{f}-(x_{f}-nH_{f})^{2}+(x_{f}-nM_{f})^{2}$ (4) This achieves what we want because the feature value differences for matching features will be smaller than those for unmatching features. Figure 11: An illustration of the principles underlying Relief in a 2D space. $x$ is the query, $nM$ is the is the nearMiss, $nH$ is the nearHit. f1 should be preferred over f2. This process is repeated multiple times to produce a set of feature weights with ‘good’ features having high weights. Implementation details of the Relief family of algorithms are provided by Urbanowicz _et al._ [25]. A link to sample code for running Relief is provided in the Appendix. The results are summarised in Figure 12. On the left we see Relief scores for the segmentation dataset compared with I-Gain scores. The Spearman correlation is 0.91 so both scores are well correlated as is evident from the plot. The plot suggests that there is a clear partition between the first 11 features and the remainder so we test the accuracy of a classifier built using just these 11 features. The results on the right show that this does result in a small improvement in accuracy compared with using all features. Figure 12: Feature selection using ReliefF. (a) I-Gain and ReliefF scores for the 19 features in the segment dataset. (a) Accuracy estimates using All features and the top 11 as selected using ReliefF. ### 5.3 Correlation-Based Feature Selection Correlation Based feature selection (CFS) is a filter strategy that relies on the principle that "A good feature subset is one that contains features highly correlated with (predictive of) the class, yet uncorrelated with (not predictive of) each other" [12]. The feature-class correlation indicates how representative of the class that feature is while the feature-feature correlation indicates any redundancies between the features. CFS works by assigning a merit value based on feature-class and feature-feature correlations to each feature subset which becomes the measure by which subsets are evaluated. Figure 13: Feature Selection using CFS. The graphs on top show how the merit scores change as the selected features are added to the selected set. The bar charts show accuracy scores for the different feature subsets. In both cases the subset with the highest merit score does not have the highest accuracy. The merit score for a feature subset $S$ is: $M_{S}=\frac{k\overline{r_{cf}}}{\sqrt{k+k(k-1)\overline{r_{ff}}}}$ (5) where $k=|S|$, $\overline{r_{ff}}$ is the average correlation between the features values in the subset ($f\in S)$ and $\overline{r_{cf}}$ is the average correlation between the values for the selected features and the class label. The correlations can be measured using techniques such as symmetrical uncertainty based on information gain, feature weighting based on the Gini- index, or by using the minimum description length (MDL) principle [12]. Information gain based methods work well and hence the symmetrical uncertainty score is commonly used in CFS implementations. CFS can work with any search strategy in a similar way to wrapper strategies, but rather than evaluating based on accuracy as in Section 4, the evaluation will be based on the merit score. For example, when using sequential forward selection, all single features merit will be evaluated from which the best will be selected. All two feature combinations which include this first feature will then be evaluated using the merit score and so on until there is no more improvement in the merit score. CFS shows a tendency to favour small subsets with moderate accuracy. The CFS implementation provided by [20] utilises a Best First search which continues the search until five consecutive non-improving feature subsets are found. Therefore, even after the merit starts decreasing, the search continues. Figure 13 shows the benefit of this as it can be seen that for both the segmentation and the penguin datasets, the hold-out accuracy from using the extra features improves in comparison to evaluating the performance at the point where merit score is highest. In both datasets, CFS has led to a drop in the hold-out accuracy compared to using the full feature subset, however has selected a smaller subset than other techniques such as Relief-F. Hence CFS may be a good technique for feature reduction. A link to sample code for running CFS is provided in the Appendix. ## 6 Embedded Methods In this section we cover feature selection methods that _emerge_ naturally from the classification algorithm or arise as a side effect of the algorithm. We will see that with Decision Trees and Logistic Regression feature selection can be an integrated part of the model building process. Then with Random Forest, we will see how feature importance scores can easily be generated from the model. ### 6.1 Decision Trees The construction of a Decision Tree from a data set will very often entail feature selection as some of the features will not appear in the tree. Features not included in the tree are effectively selected out. We show an example of this on the Penguins dataset in Figure 14. Figure 14: A decision tree for the Penguins dataset. While the data is described by four features only three are selected. In this example the dataset has been divided 50:50 into train and test sets. This tree has been trained on the training data and has 93% accuracy on the test data (see links in Appendix). This dataset has four features, flipper_length, bill_length, bill_depth and body_mass. It is clear from the tree in Figure 14 that three of the four features are selected, body_mass is not selected. This tree has been constructed with the default scikit-learn parameters so there is no pruning. It is normal in Decision Tree learning to constrain (i.e. prune) the size of the tree to prevent overfitting. The use of pruning to prevent overfitting will push the feature selection further as even less features will be selected in smaller trees. ### 6.2 Logistic Regression: Lasso In multivariate linear models such as linear regression or logistic regression, feature selection can be achieved as a side effect of regularization. In ML regularization refers to mechanisms designed to simplify models in order to prevent overfitting. Thus regularization can cause features to be deselected. Elastic net and Lasso are popular regularization methods for linear models. Here we will provide an overview of how Lasso works [24] and present examples of Lasso in operation. Starting with the basics, a multivariate regression model works as follows: $\begin{split}y=f(\mathbf{x})&=\beta_{0}+\sum_{i=1}^{p}\beta_{i}x_{i}\\\ &=\beta_{0}+\mathbf{\beta x}\end{split}$ (6) The dependent variable $y$ is a linear function of the input features; for each feature $x_{i}$ the weight of that feature is determined by the corresponding $\beta_{i}$ parameter. For binary classification problems ([0,1] labels) we can use logistic regression where the dependent variable is the log odds that an outcome variable is 1. If $pr$ is the probability that the label is 1 then $\frac{pr}{1-pr}$ is the odds. $ln\left(\frac{pr}{1-pr}\right)=\beta_{0}+\mathbf{\beta x}$ (7) So logistic regression provides a class probability: $pr=\frac{1}{1-e^{(\beta_{0}+\mathbf{\beta x})}}$ (8) Regularization prevents overfitting by limiting model capacity; this is done by limiting the size of weights. The two options are L1 or L2 regularization: $\text{L}_{1}:\quad\sum_{i=1}^{p}|\beta_{i}|<t$ (9) $\text{L}_{2}:\quad\sum_{i=1}^{p}\beta_{i}^{2}<t$ (10) So the $\beta$ parameters in (8) are fitted to the training data subject to the constraints in (9) or in (10). It transpires that when an L1 regularization is used the weaker weights will go to zero, i.e. those features will be deselected. There is an excellent explanation of _why_ this happens in the original Lasso paper by Tibshirani [24]. (a) Penguins (b) Segmentation Figure 15: The impact of Lasso on the Segmentation and Penguins datasets. Lasso reduces the magnitude of the $\beta$ parameters; some parameters get reduced to zero. Results for Lasso with the default regularization (C=1) and milder regularization (C=10) are shown. To demonstrate this on our sample datasets we reduce them to binary classification problems to make the overall process more transparent. However, feature selection using Lasso also works with multiclass problems. The results are shown in Figures 15 and 16. Because the datasets have been reduced to just two classes (Cement and Window for Segmentation and Adelie and Chinstrap for Penguins) the accuracies are higher than for the multi-class scenario. The extent of the feature reduction with Lasso is controlled by the regularization parameter C. Results are included for two levels of regularization, C=10 and C=1. C=10 results in less regularization so more features are retained. In both cases the default regularization results in too much feature reduction and generalization accuracy is reduced. For the Penguins dataset just two features are retained while three are retained in the Segmentation dataset (see Figure 15). The milder regularization retains more features resulting in no loss of generalization accuracy. (a) Penguins (b) Segmentation Figure 16: The impact of Lasso on train and test accuracy. Results for Lasso with the default regularization (C=1) and milder regularization (C=10) are shown. ### 6.3 Random Forest Feature Importance Ensembles, the idea that a committee of classifiers will be more accurate than a single classifier, are central to modern ML. It has been known for hundreds of years that a committee of decision makers will be better than an individual [6]. In ML an ensemble of classifiers voting on an outcome will be more accurate than a single classifier provided some conditions are met. Two ensemble methods, gradient boosting [8] and random forests [3] represent the state of the art in supervised ML. Breiman [3] has shown that, as a supplementary benefit of the effort required to build the random forest, the ensemble offers estimates of generalisation accuracy and feature importance. In order to explain how these estimates of feature importance work we need to go into some detail on how random forests work. In order for an ensemble to be effective, there needs to be some diversity among the ensemble members. If not the ensemble will be no better than the individual members (classifiers). If we have a dataset $\mathbf{X}_{n\times p}$ of $n$ samples described by $p$ features with class labels $\mathbf{y}_{p}$ then we have two basic strategies for training diverse classifiers: * • Bagging: We can train each classifier with different training sets drawn at random with replacement from the available data. If these ‘bagged’ training sets are of size $n$ then some samples will be selected multiple times and some not at all. For each training set, roughly 37% of samples will not be selected, i.e. out-of-bag (OOB). * • Random Subspacing: The other strategy for ensuring diversity is to work with subsets of features rather than subsets of samples. With random forest, only a subset of features are available for consideration at each split point in the construction of a tree. This subset might be quite small, for instance $\sqrt{p}$. Random forest (RF) uses both of these strategies together to ensure diversity. A typical RF could contain 1,000 trees. The samples that are OOB in these trees can be used to get an estimate of generalisation accuracy for the RF without the need to hold back test data from the training process. These OOB samples can be used to generate feature importance scores. For each of the $p$ features in the OOB samples for a given tree, the values for that feature can be randomly permuted and the revised classifications for those samples can be saved. Then, for the whole ensemble, the impact of this feature value shuffling can be assessed. If a feature is not important, then the impact of this shuffling will be minimal. For predictive features this shuffling will have a significant impact. Furthermore, and this is important for feature selection, this feature importance is being assessed in the context of other features. This contrasts with the basic filter methods described in section 5.1 where features are scored in isolation. (a) (b) Figure 17: Random forest feature importance. (a) RF feature importance scores for the Penguins dataset compared with I-Gain. (b) RF and I-Gain scores for the segmentation dataset. RF feature importance scores for the Penguin and Segmentation datasets are shown in Figure 17. I-Gain scores are also shown and these correlate well with the RF scores (0.8 for Penguins and 0.92 for Segmentation). For the Penguins dataset, the RF score ranks flipper_length highest. Because RF feature importance considers features in context, this may indicate that this feature contains information not available in the other features. The same may be true for the HUE_MEAN and EXGREEN_MEAN features in the Segmentation data which are also given higher rankings. ## 7 Feature Transformations So far we have focused on feature ranking and feature subset selection methods as that is the main focus of the paper. However, it is also worth mentioning feature transformation methods such as Principal Component Analysis (PCA) as these methods are also used for dimension reduction. The dominant feature transformation technique is Principal Components Analysis (PCA) that transforms the data into a reduced space that captures most of the variance in the data. PCA is an unsupervised technique in that it does not take class labels into account. By contrast Linear Discriminant Analysis (LDA) seeks a transformation that maximises between-class separation (Section 7.2). As with feature selection we are concerned with datasets of $n$ objects described by $p$ features. Unfortunately, with feature transformation methods, it is not unusual to represent the data as a feature-object matrix $\mathbf{X}_{p\times n}$ where each column represents an object. This contrasts with the object-feature format $\mathbf{X}_{n\times p}$ popular in supervised ML where each row is an object. For consistency with the rest of this paper we will stick with the object-feature format $\mathbf{X}_{n\times p}$. The objective with Feature Transformation is to transform the data into another set of features $F^{\prime}$ where $k=|F^{\prime}|$ and $k<p$, i.e. $\mathbf{X}_{n\times p}$ is transformed to $\mathbf{X^{\prime}}_{n\times k}$. Typically this is a linear transformation $\mathbf{W}_{p\times k}$ that will transform each $\mathbf{x}_{i}$ to $\mathbf{x}^{\prime}_{i}$ in $k$ dimensions. $\mathbf{x}^{\prime}_{i}=\mathbf{x}_{i}\mathbf{W}$ (11) ### 7.1 Principal Component Analysis In PCA the transformation described in Equation 11 is achieved so that feature $f_{1}^{\prime}$ is in the dimension in which the variance on the data is maximum, $f_{2}^{\prime}$ is in an orthogonal dimension where the remaining variance is maximum and so on. Central to the whole PCA idea is the covariance matrix of the data [13]. The diagonal terms in $\mathbf{C}$ capture the variance in the individual features and the off-diagonal terms quantify the covariance between the corresponding pairs of features. The objective with PCA is to transform the data so that the covariance terms are zero, i.e. the new dimensions are independent. The overall process is as follows: 1. 1. Calculate the means of the columns of X. 2. 2. Subtract the column means from each row of X to create the _centred matrix_ Z. 3. 3. Calculate the covariance matrix $\mathbf{C}=\frac{1}{n-1}\mathbf{Z}^{\mathsf{T}}\mathbf{Z}$. 4. 4. Calculate the eigenvectors and eigenvalues of the covariance matrix $\mathbf{C}$. 5. 5. Examine the eigenvalues in descending order to determine the number of dimensions $k$ to be retained - this is the number of principle components. 6. 6. The top $k$ eigenvectors make up the columns of the transformation matrix $\mathbf{P}$ which has dimension $(p\times k)$. 7. 7. The data is transformed by $\mathbf{X}^{\prime}=\mathbf{Z}\mathbf{P}$ where $\mathbf{X}^{\prime}$ has dimension $(n\times k$). The $i^{th}$ diagonal entry in $\mathbf{C}$ quantifies the variance of the data in the direction of the corresponding principal component. Dimension reduction is achieved by discarding the lesser principal components, i.e. $\mathbf{P}$ has dimension $(p\times k)$ where $k$ is the number of principal components retained. We illustrate PCA in operation with the example shown in Table 2 and Figure 18. The code for this is linked in the Appendix. This example is based on the Top Trump childrens’ game. Each object is a card representing a Harry Potter character described by five features. In this dataset there are 22 cards so $n=22$ and $p=5$. Figure 18(a) shows the variance in the data captured by the first four principal components (PCs). The first two PCs together account for 81% of the variance in the data so when in Figure 18(b) we plot the data in terms of these two PCs most of the variance in the data is retained. Table 2: Sample Harry Potter data for use in PCA. Name | Magic | Cunning | Courage | Wisdom | Temper ---|---|---|---|---|--- Harry Potter | 62 | 21 | 42 | 26 | 7 Hermione Granger | 60 | 16 | 40 | 73 | 2 Ron Weasley | 45 | 14 | 40 | 22 | 4 Prof. Dumbledore | 105 | 24 | 39 | 82 | 0 Prof. Snape | 85 | 24 | 19 | 71 | 7 … | … | … | … | … | … (a) (b) Figure 18: PCA on the Harry Potter dataset shown in Table 2. (a) The variance explained by the first four principal components (original data has five features). (b) A 2D plot of the data in the first two principal components. While PCA is the established method for unsupervised dimension reduction there are other methods also in use. Singular Value Decomposition (SVD), a closely related matrix factorization method, is popular in text analytics [4] and Latent Dirichlet Allocation [2] is popular for topic modelling. ### 7.2 Linear Discriminant Analysis PCA on the Penguins dataset is shown in Figure 19(a). Given that PCA is designed to project the data into dimensions that capture the variance in the data it should not be surprising that it does not do a great job of separating the classes. Figure 19(b) shows Linear Discriminant Analysis (LDA) on the same dataset. LDA takes the class labels into account and seeks a projection that maximises the separation between the classes. The objective is to uncover a transformation that will maximise between-class separation and minimise within-class separation. To do this we define two scatter matrices, $\mathbf{S}_{B}$ for between-class separation and $\mathbf{S}_{W}$ for within- class separation: (a) PCA (b) LDA Figure 19: PCA and LDA on the Penguins dataset. PCA seeks to maximize the variance captured by the two components whereas LDA seeks to maximise the separation between the classes. $\mathbf{S}_{B}=\sum_{c\in C}n_{c}(\mu_{c}-\mu)(\mu_{c}-\mu)^{\mathsf{T}}$ (12) $\mathbf{S}_{W}=\sum_{c\in C}\sum_{j:y_{j}=c}n_{c}(x_{j}-\mu_{c})(x_{j}-\mu_{c})^{\mathsf{T}}$ (13) where $n_{c}$ is the number of objects in class $c$, $\mu$ is the mean of all examples and $\mu_{c}$ is the mean of all examples in class $c$: $\mu=\frac{1}{n}\sum_{i=1}^{n}x_{i}\qquad\mu_{c}=\frac{1}{n_{c}}\sum_{j:y_{j}=c}x_{j}$ (14) The components within these summations $\mu,\mu_{c},x_{j}$ are vectors of dimension $p$ so $\mathbf{S}_{B}$ and $\mathbf{S}_{W}$ are matrices of dimension $p\times p$. The objectives of maximising between-class separation and minimising within-class separation can be combined into a single maximisation called the Fisher criterion [7, 9]: $\mathbf{W}_{LDA}=\arg\max_{\mathbf{W}}\frac{|\mathbf{W}^{\mathsf{T}}\mathbf{S}_{B}\mathbf{W}|}{|\mathbf{W}^{\mathsf{T}}\mathbf{S}_{W}\mathbf{W}|}$ (15) i.e. find $\mathbf{W}\in\mathbb{R}_{p\times k}$ so that this fraction is maximised ($|A|$ denotes the determinant of matrix $A$). This matrix $\mathbf{W}_{LDA}$ provides the transformation described in equation 11. While the choice of $k$ is again open to question it is sometimes selected to be $k=|C|-1$, i.e. one less than the number of classes in the data. Solving the optimization problem presented in equation 15 is a research topic in its own right [21, 14] so we won’t explore it in more detail here. In addition to projecting the data into a reduced dimension space LDA can also be used for classification. In the LDA space each class $c\in C$ is modelled by a multivariate Gaussian distribution $P(\mathbf{X}|\mathbf{y}=c)$. So a query sample $x_{i}$ can be assigned to the class $c_{LDA}$ for which the probability $P(y_{i}=c|x_{i})$ is largest. This is effectively a Naive Bayes classifier: $\begin{split}c_{LDA}&=\arg\max_{c\in C}P(y_{i}=c|x_{i})\\\ c_{LDA}&=\arg\max_{c\in C}P(x_{i}|y_{i}=c)P(y_{i}=c)\end{split}$ (16) In the code linked in the Appendix, LDA has 97% accuracy on the hold-out set. Half the Penguins data is used to build the LDA model shown in Figure 19 and the other half is used for testing. ## 8 Discussion & Recommended Strategies If the objective is to identify an effective feature selection strategy then there are many methods to choose from. Even though these methods have different inductive biases we generally see good correlations between the feature rankings on the two datasets we consider. This shows that all these methods have some merit in identifying good features. Our objective is not to identify a single best strategy as this will depend to some extent on the data. Given this, we propose the following methodology for feature selection on a new dataset. 1. 1. Preliminary Analysis: Use RF feature importance to rank all features (e.g. as shown in Figure 17). This will provide some insight into what features are important and it may also indicate features that can be dropped from further consideration. We recommend the RF method as it considers features in context. If the objective of the exercise is to gain an insight into the data then the analysis may stop at this point. 2. 2. Subset Selection: If the objective is to identify an effective subset for classification then a subset selection strategy is required. The main argument _against_ a Wrapper strategy is the computational cost. With advances in computing resources this is now less of an issue so we recommend a Wrapper strategy as described in Section 4. If the number of features still in consideration is not high then BE should be considered. If the set of possible features is large then SFS may be the pragmatic choice. The hybrid strategy described in Section 5.1.1 could also be considered. So our overall recommendation for feature subset selection is to use a Wrapper strategy. This would be in tune with the current model selection protocols in ML [19]. However, the recommendation for Preliminary Analysis is also important. Early in the analysis of new data it will be helpful to use a Filter strategy to gain some insight into the data. This Filter analysis will be independent of any specific classifier models that might be used subsequently. ## 9 Conclusions The objective for this tutorial paper was to provide an overview of popular feature selection methods, describing how they work and providing links to Python implementations. It was not our intention to offer a comparative evaluation to identify the best methods; that is done very well elsewhere [11, 23]. Instead, our objective was to provide a stepping stone to help researchers get started with feature selection on their own datasets. In the Introduction we listed objectives for feature selection other that to improve classifier performance. Indeed, the examples in this paper suggest that feature subset selection may not have a big impact on classifier performance - this is borne out by other studies [11]. Instead feature selection can deliver other benefits through savings in data gathering an model execution. Perhaps, the biggest benefit of feature selection is the insight it offers into what is important in the data under analysis. ## Acknowledgements This publication has emanated from research conducted with the financial support of Science Foundation Ireland under Grant numbers 16/RC/3872 and 18/CRT/6183. For the purpose of Open Access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission. ## References * [1] Richard E Bellman. Adaptive Control Processes: A Guided Tour. Princeton University Press, 1961. * [2] David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. the Journal of machine Learning research, 3:993–1022, 2003. * [3] Leo Breiman. Random forests. Machine learning, 45(1):5–32, 2001. * [4] Pádraig Cunningham. Dimension reduction. In Machine learning techniques for multimedia, pages 91–112. Springer, 2008. * [5] Padraig Cunningham and Sarah Jane Delany. k-Nearest Neighbour Classifiers: 2nd Edition (with Python examples), 2020. * [6] Marie Jean Antoine Nicolas de Caritat and Marquis De Condorcet. Essai sur l’application de l’analyse à la probabilité des décisions rendues à la pluralité des voix. L’imprimerie royale, 1785. * [7] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179–188, 1936. * [8] Jerome H Friedman. Stochastic gradient boosting. Computational statistics & data analysis, 38(4):367–378, 2002\. * [9] Keinosuke Fukunaga. Introduction to statistical pattern recognition. Elsevier, 2013. * [10] Kristen B Gorman, Tony D Williams, and William R Fraser. Ecological sexual dimorphism and environmental variability within a community of antarctic penguins (genus pygoscelis). PloS one, 9(3):e90081, 2014. * [11] Isabelle Guyon and André Elisseeff. An introduction to variable and feature selection. Journal of machine learning research, 3(Mar):1157–1182, 2003. * [12] Mark A Hall. Correlation-based Feature Selection for Machine Learning. PhD thesis, The University of Waikato, 1999. * [13] Harold Hotelling. Analysis of a complex of statistical variables into principal components. Journal of educational psychology, 24(6):417, 1933. * [14] Rui Huang, Qingshan Liu, Hanqing Lu, and Songde Ma. Solving the small sample size problem of lda. In Object recognition supported by user interaction for service robots, volume 3, pages 29–32. IEEE, 2002. * [15] Xin Jin, Anbang Xu, Rongfang Bie, and Ping Guo. Machine learning techniques and chi-square feature selection for cancer classification using sage gene expression profiles. In International Workshop on Data Mining for Biomedical Applications, pages 106–115. Springer, 2006. * [16] John D Kelleher, Brian Mac Namee, and Aoife D’arcy. Fundamentals of machine learning for predictive data analytics: algorithms, worked examples, and case studies. MIT Press, 2020. * [17] Kenji Kira, Larry A Rendell, et al. The feature selection problem: Traditional methods and a new algorithm. In Aaai, volume 2, pages 129–134, 1992. * [18] Igor Kononenko. Estimating attributes: Analysis and extensions of relief. In European conference on machine learning, pages 171–182. Springer, 1994. * [19] Arun Kumar, Robert McCann, Jeffrey Naughton, and Jignesh M Patel. Model selection management systems: The next frontier of advanced analytics. ACM SIGMOD Record, 44(4):17–22, 2016. * [20] Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P Trevino, Jiliang Tang, and Huan Liu. Feature selection: A data perspective. ACM Computing Surveys (CSUR), 50(6):94, 2018. * [21] Wei Liu, Yunhong Wang, Stan Z Li, and Tieniu Tan. Null space approach of fisher discriminant analysis for face recognition. In International Workshop on Biometric Authentication, pages 32–44. Springer, 2004. * [22] John Loughrey and Pádraig Cunningham. Overfitting in wrapper-based feature subset selection: The harder you try the worse it gets. In International Conference on Innovative Techniques and Applications of Artificial Intelligence, pages 33–43. Springer, 2004. * [23] Luis Carlos Molina Félix, Luis Antonio Belanche Muñoz, and M Àngela Nebot Castells. Feature selection algorithms: a survey and experimental evaluation. In 2002 IEEE International Conference on Data Mining, ICDM 2002: 9-12 December 2002, Maebashi City, Japan: proceedings, pages 306–313. Institute of Electrical and Electronics Engineers (IEEE), 2002. * [24] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological), 58(1):267–288, 1996. * [25] Ryan J Urbanowicz, Randal S Olson, Peter Schmitt, Melissa Meeker, and Jason H Moore. Benchmarking relief-based feature selection methods for bioinformatics data mining. Journal of Biomedical Informatics, 85:168–188, 2018. ## Appendix A Appendix I: Python Code The GitHub repository444https://github.com/PadraigC/FeatSelTutorial associated with this paper contains the following Python Notebooks: * • FS-Wrappers: Code for SFS and BE Wrappers from mlxtend. * • FS-Filters: Code for using I-Gain and Chi-square Filters from scikit-learn. * • FS-ReliefF: Code for using ReliefF Filters from skrebate. * • FS-D-Tree: Building D-Trees with embedded feature selction using scikit-learn. * • FS-Lasso: Feature selection for Logistic Regression using scikit-learn. * • FS-CFS: Correlation-Based feature subset selection. * • FS-Random-Forest: Feature importance from Random Forest using scikit-learn. * • FS-PCA: Principal Component Analysis using the PCA implementation in scikit- learn. * • FS-LDA: Linear Discriminant Analysis using the LDA implementation in scikit- learn.
# A novel Fourier neural operator framework for classification of multi-sized images: Application to 3D digital porous media Ali Kashefi<EMAIL_ADDRESS>Tapan Mukerji<EMAIL_ADDRESS> ###### Abstract Fourier neural operators (FNOs) are invariant with respect to the size of input images, and thus images with any size can be fed into FNO-based frameworks without any modification of network architectures, in contrast to traditional convolutional neural networks (CNNs). Leveraging the advantage of FNOs, we propose a novel deep-learning framework for classifying images with varying sizes. Particularly, we simultaneously train the proposed network on multi-sized images. As a practical application, we consider the problem of predicting the label (e.g., permeability) of three-dimensional digital porous media. To construct the framework, an intuitive approach is to connect FNO layers to a classifier using adaptive max pooling. First, we show that this approach is only effective for porous media with fixed sizes, whereas it fails for porous media of varying sizes. To overcome this limitation, we introduce our approach: instead of using adaptive max pooling, we use static max pooling with the size of channel width of FNO layers. Since the channel width of the FNO layers is independent of input image size, the introduced framework can handle multi-sized images during training. We show the effectiveness of the introduced framework and compare its performance with the intuitive approach through the example of the classification of three-dimensional digital porous media of varying sizes. ††journal: arXiv [first]organization=Department of Civil and Environmental Engineering, Stanford University, city=Stanford, postcode=94305, state=CA, country=USA [third]organization=Department of Energy Science and Engineering, Stanford University, city=Stanford, postcode=94305, state=CA, country=USA ## 1 Introduction and motivation Since 2020, neural operators have gained extensive popularity, specifically with two versions of graph neural operators (Li et al., 2020b) and Fourier neural operators (FNOs) (Li et al., 2020a). In this article, our attention is on FNOs. From a computer science perspective, regular FNOs fall in the category of supervised deep learning framework, necessitating a large volume of labeled data for training. FNOs have demonstrated their proficiency in input-output mapping across various industrial and scientific applications such as incompressible flows (Li et al., 2022b; Bonev et al., 2023; Peng et al., 2024; Choubineh et al., 2023; Lyu et al., 2023; Gupta and Brandstetter, 2022; Peng et al., 2023a), wave equations (Zhu et al., 2023a; Zou et al., 2023; Yang et al., 2023), thermal fields (Zhao et al., 2024; Hao et al., 2023), carbon storages and sequestration (Wen et al., 2022; Jiang et al., 2023b), and other areas (Peng et al., 2023b; You et al., 2022; Kontolati et al., 2023; Zhu et al., 2023b; Hua and Lu, 2023; White et al., 2023a; Li et al., 2021; Pathak et al., 2022; Rahman et al., 2022b, a; Yang et al., 2022; Li et al., 2022a; Maust et al., 2022; Zhao et al., 2022; Renn et al., 2023; Xiong et al., 2023; Chen et al., 2023; Huang et al., 2023; Poels et al., 2023; White et al., 2023b; Thodi et al., 2023; Zhao et al., 2023; Tran et al., 2023; Lee, 2022; Brandstetter et al., 2023; Li et al., 2023; Majumdar et al., 2023; Jiang et al., 2023a; Lehmann et al., 2024). From a computer vision perspective, these are framed as segmentation problems, where an input image, such as the geometry of an airfoil, is mapped to another image, for instance, the velocity field around that airfoil. An analogous area in computer vision is classification, where an input image is mapped, for example, to a name or number. While FNOs have potential in classification tasks, there exists only a limited number of research conducted in this application as per our knowledge (Johnny et al., 2022; Xi et al., 2022; Kabri et al., 2023). Johnny et al. (2022) used the FNO architecture for classifying images in the CIFAR-10 dataset, containing ten different classes; however, they trained the network only on images with a fixed size of 32 by 32 pixels. Additionally, Kabri et al. (2023) examined the FNO architecture for image classification. Although they tested images of various sizes (e.g., 28 by 28 pixels, 112 by 112 pixels, etc.), they trained and then tested the network separately for each size, assessing its performance on the corresponding size. Xi et al. (2022) utilized the FNO architecture for the hyperspectral remote sensing image classification. Their dataset comprised images of various sizes, including 512 by 614 pixels, 610 by 340 pixels, and 512 by 217 pixels. However, they adjusted all images to a fixed size by adding patches. Consequently, although they employed the FNO architecture, in practice, they limited their analysis to images of a uniform size. In the current study, we narrow our focus on classification problems. More specifically, we consider the problem of predicting the permeability of three-dimensional digital porous media, which vary in size, as a benchmark test case. FNOs are invariant with respect to the size of input images, and this characteristic ensures that images of varying sizes can be processed by FNO- based deep learning frameworks without requiring any architectural alterations. Note that regular convolutional neural networks (CNNs) lack this feature (Goodfellow et al., 2016). Building on this strength of FNOs, we introduce a deep-learning framework for training the network simultaneously on images with varying sizes for a classification problem. To achieve this deep learning framework, FNO layers must be connected to a classifier, which is commonly a multilayer perceptron (MLP). An intuitive approach to set this would be to link FNO layers to a classifier via adaptive max pooling. Considering the application of permeability prediction of three-dimensional porous media, our machine-learning experiments show that this intuitive approach only works well for porous media with fixed sizes. Pivoting from this, we propose our novel approach. Rather than using adaptive max pooling, we implement static max pooling with the size of the channel width of the FNO layers. Given that the size of the channel width of FNO layers is independent of the size of input images, our proposed framework can be efficiently trained on various image sizes at once (see Fig. 1 and Fig. 2). To explain, at a high level, the difference between using adaptive max pooling (see Fig. 2) and static max pooling (see Fig. 1), let us consider for example a three-dimensional image being fed as an input of the deep learning framework. For both pooling methods, at the framework’s outset, FNO layers lift the input image from its three-dimensional space to a higher dimensional space, determined by the size of the channel width of the FNO layers. In the case of adaptive max pooling, after FNO layer operations, the outcome eventually is dropped into the initial three-dimensional space with the same size as the input image. This array then serves as the input of adaptive max pooling. The output of the adaptive pooling is then the input of the classifier. In the case of static max pooling, before FNO layers drop the output, we implement static max pooling, which functions within the high dimensional space and pools with the size of the channel width of FNO layers. The resulting output from this pooling then becomes the classifier’s input. A more detailed exploration of these concepts is provided in Sect. 2. The study of physical and geometric features of porous media is important in diverse scientific and industrial areas such as digital rock physics (Andra et al., 2013a, b), membrane systems (Liang and Fletcher, 2023), geological carbon storages (Blunt et al., 2013), and medicine (Kumeria, 2022; Das et al., 2018). Deep learning frameworks have been widely used for predicting the permeability of porous media (Meng et al., 2023; Xie et al., 2023; Kashefi and Mukerji, 2023, 2021; Liu et al., 2023; Hong and Liu, 2020; Wu et al., 2018; Tembely et al., 2020; Masroor et al., 2023; Sun et al., 2023), but, to the best of our knowledge, all these frameworks were trained on a fixed-size porous media. Note that training the proposed network to predict the permeability of porous media of varying sizes comes with an exclusive challenge when compared to training the network on conventional images for the purpose of classifying them by their names (like those of cats and dogs). For conventional images, one possible solution to handle images with different sizes is to equalize them by adding mini patches to the smaller images. Nevertheless, this solution is inapplicable to the porous media problem. Adding mini patches to porous media can alter their physical properties such as permeability. For instance, adding mini patches around a porous medium simulates sealing it with wall boundaries, which prohibits flow within its pore spaces, resulting in a permeability of zero. Additionally, the inherently three-dimensional nature of porous media introduces another layer of complexity compared to the two- dimensional conventional images. The remainder of this article is organized as follows. We introduce and discuss the concept of Fourier neural operators for image classification in Sect. 2, starting with the traditional strategy of adaptive max pooling, followed by our novel approach of static max pooling in the high dimension of the Fourier space channel. A brief review of theoretical aspects of FNOs is given in Sect. 2.3. Data generation and the training methodologies are respectively presented in Sect. 3 and Sect. 4. In Sect. 5, we provide results and discussion, including a comparison between traditional strategy and our novel approach. Moreover, we present a sensitivity analysis, covering the number of Fourier modes, the channel width of discrete Fourier space, the number of FNO units, and the effect of activation functions and average pooling. The deep learning model generalizability is discussed in this section as well. Finally, we summarize the work and present insight into future directions in Sect. 6. Figure 1: Schematic of the proposed FNO-based framework for multi-size image classification Figure 2: Schematic of the intuitive FNO-based framework for multi-size image classification ## 2 Fourier neural operators for image classification ### 2.1 Our novel approach: Static max pooling in channel width of FNO layers In this subsection, we introduce the architecture of our proposed deep learning framework. Our explanation heavily uses matrix notation to ensure clarity and provide a deeper understanding. As illustrated in Fig. 1, the input of the deep learning framework is a cubic binary porous medium represented as the matrix $\mathbf{A}_{n\times n\times n}$. As a first step, the matrix $\mathbf{A}_{n\times n\times n}$ is lifted to a higher dimensional space using a fully connected network. The dimension of this space is termed the channel width of an FNO layer, shown by “width” in our matrix notation. This lifting results in a four-dimensional matrix, denoted as $\mathbf{B}_{\text{width}\times n\times n\times n}^{0}$. The matrix $\mathbf{B}_{\text{width}\times n\times n\times n}^{0}$ becomes subsequently the input of an FNO layer. Within the FNO layer, two operations are applied to $\mathbf{B}_{\text{width}\times n\times n\times n}^{0}$: the kernel integration operator, denoted by $\mathbf{K}_{\text{width}\times\text{width}}^{0}$, and the linear transformation operator, denoted by $\mathbf{W}_{\text{width}\times\text{width}}^{0}$. The network computes the matrix-matrix multiplication of $\mathbf{K}_{\text{width}\times\text{width}}^{0}\mathbf{B}_{\text{width}\times n\times n\times n}^{0}$ and $\mathbf{W}_{\text{width}\times\text{width}}^{0}\mathbf{B}_{\text{width}\times n\times n\times n}^{0}$ and then sums up the resulting matrices, as depicted in Fig. 1. The output undergoes element-wise operations of the Rectified Linear Unit (ReLU) activation function (Goodfellow et al., 2016) defined as $\sigma(\gamma)=\max(0,\gamma),$ (1) Resulting in a four-dimensional matrix $\mathbf{B}_{\text{width}\times n\times n\times n}^{1}$. Mathematically, this procedure can be summarized as $\displaystyle\mathbf{B}_{\text{width}\times n\times n\times n}^{1}$ $\displaystyle=\sigma\left(\mathbf{K}_{\text{width}\times\text{width}}^{0}\mathbf{B}_{\text{width}\times n\times n\times n}^{0}\right.$ $\displaystyle\quad\left.+\mathbf{W}_{\text{width}\times\text{width}}^{0}\mathbf{B}_{\text{width}\times n\times n\times n}^{0}\right).$ (2) In scenarios where multiple FNO layers exist in the framework, the matrix $\mathbf{B}_{\text{width}\times n\times n\times n}^{1}$ serves as the input for the succeeding FNO layers, and the same sequence of operations is applied. If we assume that there are $l$ number of FNO layers, the output from the final FNO layer is the matrix $\mathbf{B}_{\text{width}\times n\times n\times n}^{l}$. In the next step, we implement static max pooling on the first dimension of matrix $\mathbf{B}_{\text{width}\times n\times n\times n}^{l}$. Because “width” is independent of the input image dimension (i.e., $n$), the static pooling works for input images with any desired size (e.g., $n=40$, $n=48$, and $n=56$). Note that “width” is a hyper parameter of FNO layers and independent of $n$, as all the matrix-matrix multiplication operates on the dimension with the size “width”, and not “$n$”. The static max pooling produces a vector of length width, representing the global features of the input images. The vector is then connected to a classifier. The classifier is a Multilayer Perceptron (MLP) composed of three layers of sizes 128, 128, and 1. The ReLU activation function is used in the initial two layers along with a dropout with a rate of 0.3. Following the third layer, a sigmoid activation function defined as $\sigma(\gamma)=\frac{1}{1+e^{-\gamma}},$ (3) is used to ensure output values are bounded between 0 and 1. ### 2.2 Intuitive approach: Adaptive max pooling in 3D spatial space In this subsection, we explain the intuitive approach (see Fig. 2). Drawing parallels to our approach elaborated in the previous subsection, we begin by considering the input porous medium, which is a three-dimensional matrix represented by $\mathbf{A}_{n\times n\times n}$. All operations outlined in Sect. 2.1 are applied to $\mathbf{A}_{n\times n\times n}$ until the network obtains the matrix $\mathbf{B}_{\text{width}\times n\times n\times n}^{l}$ at an intermediate step, as depicted in Fig. 2. As the next step, we drop (as an inverse of the lifting operator explained in Sect. 2.1) the matrix $\mathbf{B}_{\text{width}\times n\times n\times n}^{l}$ from the high dimensional space to the default space by means of a fully connected network. This transformation results in the matrix $\mathbf{Z}_{n\times n\times n}$. At this juncture, we use adaptive three-dimensional max pooling, a functionality that is available in deep learning platforms such as PyTorch (Paszke et al., 2019) or TensorFlow (Abadi et al., 2015). To ensure a fair comparison between the traditional approach and our novel approach, we keep the size of the vector of the global feature consistent across both approaches. To this end, the output of the adaptive max pooling is tailored to yield a vector of size “width”. The resulting vector represents the global features of the input images. Note that because the size of matrix $\mathbf{Z}_{n\times n\times n}$ depends on the size of the input image (i.e., $n$), the pooling must be adaptive as we plan to train the network simultaneously on input images with varying sizes (e.g., $\mathbf{A}_{40\times 40\times 40}$, $\mathbf{A}_{48\times 48\times 48}$, and $\mathbf{A}_{56\times 56\times 56}$). Subsequent to the adaptive max pooling, the global feature vector is connected to a classifier. This classifier features and architecture is precisely the same as the one elucidated in Sect. 2.1. ### 2.3 A brief review of theoretical aspects of Fourier neural operators We focused on the technical aspects and computer implementation of FNO layers in Sect. 2.1 and Sect. 2.2. Theoretical aspects of FNO layers have already been vastly explained and discussed in the literature (Li et al., 2020a). In this subsection, we briefly review the theory behind FNO layers and highlight some important features. As discussed in Sect. 2.1, an FNO layer comprises two main operators: the integral kernel operator and the linear transformation. We overview the integral kernel operator. We consider the bounded domain $D$ such that $D\subset\mathbb{R}^{d}$, where $d$ indicates the physical dimensionality of the problem and is equal to 3 (i.e., $d=3$) for the current problem since we deal with three-dimensional porous media. We further show the input of the FNO layer by $b(x)$ with $x\in D$, where $b$ is a function representing all the operators applied to $x$ when it arrives at the gate of the FNO layer. Moreover, we define the periodic kernel function $\tau:\mathbb{R}^{2(d+d_{a})}\rightarrow\mathbb{R}^{\text{width}\times\text{width}}$, where $d_{a}$ is the number of input features and is equal to 1 (i.e., $d_{a}=1$) in this study, because the input of the deep learning framework is only a cubic binary array (representing a porous medium), and this array only provides one feature, which is the geometry of the porous medium. Additionally, recall that “width” is the channel width of the FNO layer, as illustrated in Sect. 2.1 and Sect. 2.2. Following the formulation proposed by Li et al. (2020a), the operation of the integral kernel $\mathcal{K}$ on the function $b(x)$ in the continuous space is defined as $\mathcal{K}b(x)=\int_{D}\tau(x,y)b(y)\,dy,\quad\forall x\in D.$ (4) Following the original design of FNO layers by Li et al. (2020a), we introduce the condition $\tau(x,y)=\tau(x-y)$. By applying the convolution theorem as detailed in the literature (Li et al., 2020a), the following expression for the integral kernel operator is obtained: $\mathcal{K}b(x)=\mathcal{F}^{-1}\left(\mathcal{F}(\tau)\cdot\mathcal{F}(b(x))\right),\quad\forall x\in D,$ (5) where the Fourier transform and its inverse are shown by $\mathcal{F}$ and $\mathcal{F}^{-1}$, respectively. We introduce $\mathcal{R}$ as the learnable Fourier transform of $\tau$ such that $\mathcal{R}=\mathcal{F}(\tau).$ (6) Beyond the theory, we must implement these mathematical concepts in a deep learning framework. In this way, we work with discrete spaces and consequently, discrete modes of the Fourier transform. Hence, $\mathcal{R}$ is implemented as a neural network. Additionally, each porous medium is represented by $n^{3}$ discrete points such that $\\{x_{1},x_{2},\cdots,x_{n^{3}}\\}\subset D$. Moreover, Fourier series expansions are truncated at a maximum number of modes $m_{\text{max}}$ computed as $m_{\text{max}}=\left|\left\\{m\in\mathbb{Z}^{d}:|m_{j}|\leq m_{\text{max},j},\text{ for }j=1,\cdots,d\right\\}\right|,$ (7) where $m_{\text{max},j}$ is the maximum number of modes taken in the dimension $j$, and is a hyper-parameter of the FNO layer. Note that since we work on three-dimensional problems in the current study, $d=3$, and thus, there are only $m_{\text{max},1}$, $m_{\text{max},2}$, and $m_{\text{max},3}$. As a result, the components of the $\mathcal{R}\cdot\mathcal{F}(b(x))$ operation can be computed by the following formulation $\displaystyle[\mathcal{R}\cdot\mathcal{F}(b(x))]_{m,i}$ $\displaystyle=\sum_{j=1}^{\text{width}}[\mathcal{R}]_{m,i,j}[\mathcal{F}(b(x))]_{m,j},$ (8) $\displaystyle\quad m=1,\cdots,m_{\text{max}},\quad i=1,\cdots,\text{width},$ where $[\mathcal{R}]\in\mathbb{C}^{m_{\text{max}}\times\text{width}\times\text{width}}$ is the matrix representation of $\mathcal{R}$ in the discrete space. $[\mathcal{R}\cdot\mathcal{F}(b(x))]\in\mathbb{C}^{m_{\text{max}}\times\text{width}}$ and $[\mathcal{F}(b(x))]\in\mathbb{C}^{m_{\text{max}}\times\text{width}}$ are similarly defined. To increase computing efficiency and enable parallel computing, the operator $[\mathcal{R}]$, for the current three-dimensional problem, is better to be implemented as a five-dimensional matrix expressed as $\mathbf{R}_{m_{\text{max},1}\times m_{\text{max},2}\times m_{\text{max},3}\times\text{width}\times\text{width}}.$ (9) As can be seen from Eq. 9, the size of matrix $\mathbf{R}$, and thus the count of trainable parameters in the FNO layer, is a function of the number of maximum Fourier modes at each dimension and the channel width of the FNO layer. Recall that these parameters (i.e., $m_{\text{max},1}$, $m_{\text{max},2}$, $m_{\text{max},3}$, and width) are the hyperparameter of FNO layers and need to be tuned by potential users. \begin{overpic}[width=433.62pt]{CNN40.png} \put(5.0,90.0){a} \end{overpic} \begin{overpic}[width=433.62pt]{CNN48.png} \put(5.0,90.0){b} \end{overpic} \begin{overpic}[width=433.62pt]{CNN56.png} \put(5.0,90.0){c} \end{overpic} Figure 3: A few examples of synthetically generated three-dimensional digital porous media for training the proposed neural network; a an image of size $40^{3}$, b an image of size $48^{3}$, and c an image of size $56^{3}$ ## 3 Data generation To generate synthetic data to examine the deep learning framework under investigation in this study, we consider cubic porous medium domains with length L along each side, spatial correlation length of $l_{c}$, and porosity of $\phi$. We use the truncated Gaussian algorithm (Lantuejoul, 2002; Le Ravalec-Dupin et al., 2004) to generate synthetic porous media. In practice, we create three-dimensional cubic arrays of dimension $n\times n\times n$, populated with random numbers conforming to a normal distribution with the characteristics of a mean value of 0.0 and a standard deviation of 1.0. Subsequently, we filter the arrays by a three-dimensional Gaussian smoothing kernel with a standard deviation of 5.0 and a filter size commensurate with a spatial correlation length ($l_{c}$) of 17. We then subject the arrays to a binarization process via a thresholding number such that the porosity ($\phi$) of the resulting arrays lies within the range of [0.125, 0.200]. We use the MATLAB software to handle the above-described steps. We set $L$ as $n\times\delta x$, where $\delta x$ represents the length of each side of a pixel in porous media. We set $\delta x$ to 0.003 m. We generate porous media with three different sizes by considering three different values for $n$, such that $n_{1}=40$, $n_{2}=48$, and $n_{3}=56$. In this way, each cubic porous medium can be characterized by its size as $n^{3}$ (e.g., $40^{3}$, $48^{3}$, and $56^{3}$). For each $n$, we generate 1250 data. We randomly split the generated data corresponding to each size into three categories of training (80%, i.e., 1000 data), validation (10%, i.e., 125 data), and test (10%, i.e., 125 data). Hence, there are 3750 data in total, 3000 data for the training set, 375 data for the validation set, and 375 data for the test set. Figure 3 exhibits a few examples of the generated synthetic data. To stimulate the incompressible viscous Newtonian flow within the generated porous media, we apply a constant pressure gradient in the $x$ direction ($\Delta p/L$). Zero velocity boundary condition is applied at the top and bottom of the porous medium on the $y-z$ planes. Given the geometry and boundary conditions illustrated above, we use a Lattice Boltzmann solver (Keehm et al., 2004) to solve the continuity and steady-state Stokes equations, which are written as follows: $\nabla\cdot\bm{\mathit{u}}=0,\quad\text{in }V,$ (10) $-\nabla p+\mu\Delta\bm{\mathit{u}}=\textbf{0},\quad\text{in }V,$ (11) where $\mu$ is the dynamic viscosity, $\bm{\mathit{u}}$ and $p$ indicate, respectively, the velocity vector and pressure fields in the pore space of the porous medium, $V$. In the next step, we compute the permeability in the $x-$direction ($k$) using Darcy’s law (Darcy, 1856), $k=-\frac{\mu\bar{U}}{\Delta p/L},$ (12) where $\bar{U}$ shows the average velocity in the entire porous medium (i.e., including solid matrices). The computed permeabilities of our data set fall in the range [20 mD, 200 mD]. ## 4 Training To accelerate the convergence of the training procedure, the output training data (i.e., permeability) are scaled in the range of [0, 1] using the maximum and minimum values of the training set. Note that although we train a single neural network simultaneously on porous media with three different sizes (corresponding to $n_{1}$, $n_{2}$, and $n_{3}$), we normalize the permeability of porous media of each size using the maximum and minimum values of the specific size. Mathematically, it can be written as $\\{\hat{k}_{\text{truth}}\\}_{n_{j}}=\frac{\\{k\\}_{n_{j}}-\min\\{k\\}_{n_{j}}}{\max\\{k\\}_{n_{j}}-\min\\{k\\}_{n_{j}}},\quad j=1,2,\text{and }3,$ (13) where, $\hat{k}_{\text{truth}}$ shows the ground truth scaled permeability. Moreover, for instance, $\\{k\\}_{n_{1}}$ indicates the training data containing porous media with the size of $40^{3}$ (because $n_{1}=40$). Note that we eventually rescale predicted permeability ($\hat{k}_{\text{prediction}}$) to the physical domain ($k_{\text{prediction}}$) for analyzing the neural network performances. Concerning the loss function, we use the mean squared error function defined as $\text{Loss}=\frac{1}{N}\sum_{i=1}^{N}(\hat{k}_{\text{prediction}}-\hat{k}_{\text{truth}})^{2},$ (14) where $N$ is the number of data in the training set (i.e., $N=3000$). Note that using relative mean squared error as the loss function does not lead to a significant difference in the results, based on our experiments. We set the number of modes in each dimension to 2 (i.e., set $m_{\text{max},1}=2$, $m_{\text{max},2}=2$, and $m_{\text{max},3}=2$). The channel width of the discrete Fourier space is set to 64 (i.e., $\text{width}=64$). It is worth noting that both the number of modes and the channel width play pivotal roles in the network performance. Detailed discussions on their significance and implications are provided in Sect. 5.2 and 5.3, respectively. Additionally, we implement three units of FNOs in the network. The Adam optimizer (Kingma and Ba, 2014) is used. A constant learning rate of 0.001 is selected. We use the stochastic gradient descent (Goodfellow et al., 2016) with a mini-batch size of 50. As discussed in Sect. 2, the architecture of FNOs is designed to be independent of the spatial resolution of input images. During the training process, however, all the input images within a mini-batch must be the same size. In practice, each epoch of training is characterized by an inner loop that iterates through mini-batches of differing porous medium sizes (i.e., $40^{3}$, $48^{3}$, and $56^{3}$). Within this loop, the training process starts with a mini-batch of data of size $40^{3}$, followed by one of size $48^{3}$, and then continues to $56^{3}$, in sequence until all the data in the training set are covered within the epoch. Note that the trainable parameters of the network are updated only at the end of each epoch. Our deep learning experiments show that the order in which these differently sized porous media are fed within an epoch has no significant influence on the result accuracy and convergence speed, whether starting with the porous media of size $40^{3}$, followed by $48^{3}$ and $56^{3}$, or any other permutation. From a software perspective, we employ the NVIDIA A100 (SXM4) graphic card with 80 Gigabytes of RAM for training the networks. In the last paragraph of this subsection, we address the metric used for assessing the effectiveness of permeability prediction. We use the coefficient of determination, also known as the $R^{2}$ score, which can be calculated using the following formula $R^{2}=1-\frac{\sum_{i=1}^{Q}({k_{\text{truth}}}_{i}-{k_{\text{prediction}}}_{i})^{2}}{\sum_{i=1}^{Q}({k_{\text{truth}}}_{i}-\bar{k})^{2}},$ (15) where $Q$ represents the number of the data in a set (e.g., training, test, etc.) and $\bar{k}$ is the average value of the set $\\{{k_{\text{truth}}}_{i}\\}_{i=1}^{Q}$. Figure 4: $R^{2}$ plots for the test set (375 data) using the proposed approach for classification of multi-sized images ## 5 Results and discussion ### 5.1 General analysis As illustrated in Fig. 4, the success of our approach is evident in the $R^{2}$ score, 0.96809, obtained for the test set (e.g., 375 data). Additionally, Fig. 5 specifically showcases the $R^{2}$ scores for the test set but individualized for each cubic size (i.e., $40^{3}$, $48^{3}$, and $56^{3}$). As can be seen in Fig. 5, the $R^{2}$ scores obtained are equal to 0.96830, 0.96978, and 0.96607, respectively for the cubic digital porous media of sizes $40^{3}$, $48^{3}$, and $56^{3}$. The range of $R^{2}$ scores for the three different sizes remains at an excellent level, demonstrating that our FNO-based framework is robust and not overfitted to any specific size. Table 1: $R^{2}$ score of the test set for different mode numbers of the proposed FNO-based framework Number of modes in each dimension | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 ---|---|---|---|---|---|---|---|---|--- $R^{2}$ score | 0.96809 | 0.15416 | 0.26757 | 0.26361 | 0.23325 | 0.38789 | 0.40433 | 0.31773 | 0.28839 \begin{overpic}[width=433.62pt]{C1_40.pdf} \put(5.0,90.0){a} \end{overpic} \begin{overpic}[width=433.62pt]{C1_48.pdf} \put(5.0,90.0){b} \end{overpic} \begin{overpic}[width=433.62pt]{C1_56.pdf} \put(5.0,90.0){c} \end{overpic} Figure 5: $R^{2}$ plots for the test set (375 data) using the proposed approach for the classification of multi-sized images. The results are individually shown for a images of size $40^{3}$ (125 data), b images of size $48^{3}$ (125 data), and c images of size $56^{3}$ (125 data) \begin{overpic}[width=433.62pt]{Loss_mode2.pdf} \put(5.0,90.0){a} \end{overpic} \begin{overpic}[width=433.62pt]{Loss_mode7.pdf} \put(5.0,90.0){b} \end{overpic} \begin{overpic}[width=433.62pt]{Loss_mode10.pdf} \put(5.0,90.0){c} \end{overpic} Figure 6: Evolution of the loss function for the validation and training sets for the choice of a $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=2$, b $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=7$, and c $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=10$ \begin{overpic}[width=433.62pt]{C9.pdf} \put(5.0,90.0){a} \end{overpic} \begin{overpic}[width=433.62pt]{C12.pdf} \put(5.0,90.0){b} \end{overpic} \begin{overpic}[width=433.62pt]{C15.pdf} \put(5.0,90.0){c} \end{overpic} Figure 7: $R^{2}$ plots for the test set (375 data) using the proposed approach for the classification of multi-sized images for the choice of a $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=4$, b $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=7$, and c $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=10$ \begin{overpic}[width=433.62pt]{C24.pdf} \put(5.0,90.0){a} \end{overpic} \begin{overpic}[width=433.62pt]{C25.pdf} \put(5.0,90.0){b} \end{overpic} Figure 8: $R^{2}$ plots for the test set (375 data) using the proposed approach for the classification of multi-sized images for the choice of a $m_{\text{max},1}=m_{\text{max},2}=2$ and $m_{\text{max},3}=10$, and the choice of b $m_{\text{max},1}=2$, and $m_{\text{max},2}=m_{\text{max},3}=10$ ### 5.2 Number of Fourier modes in each dimension Our deep learning experiments demonstrate that there is a critical interplay between the number of modes (i.e., $m_{\text{max},1}$, $m_{\text{max},2}$, and $m_{\text{max},3}$) set in the proposed FNO framework and the tendency for overfitting during the training procedure. Accordingly, setting the number of modes beyond 2 leads to a severe divergence between the training and validation loss. This fact can be observed in Fig. 6 when we set $m_{\text{max},1}=7$, $m_{\text{max},2}=7$, and $m_{\text{max},3}=7$ or $m_{\text{max},1}=10$, $m_{\text{max},2}=10$, and $m_{\text{max},3}=10$. The reported results indicate that the number of modes plays a critical role in the FNO model generalization. A further survey of the influence of the number of modes in the FNO configuration is performed by varying the number of modes in all three principal directions, from 2 to 10, and the obtained $R^{2}$ scores are tabulated in Table 1. Accordingly, the optimal mode configuration for avoiding overfitting is 2, as the divergence between the validation and training loss is minimized. Consequently, a careful selection of the number of modes in the FNO units is necessary to make the deep learning framework robust and reliable for the image classification application. The consequence of this scenario is observable in Fig. 7, where we plot the $R^{2}$ score for the test sets, for example, for the choice of $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=4$, $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=7$, and $m_{\text{max},1}=m_{\text{max},2}=m_{\text{max},3}=10$. In all of these cases, the $R^{2}$ scores obtained for the prediction of the permeability of the porous media in the test set are less than 0.4. We perform two other experiments. In the first one, we set only one mode (e.g., $m_{\text{max},3}$) to 10 ($m_{\text{max},3}=10$) and the other two modes to 2 (i.e., $m_{\text{max},1}=2$ and $m_{\text{max},2}=2$). In the second one, we set only two modes (e.g., $m_{\text{max},2}$ and $m_{\text{max},3}$) to 10 and the reminder mode to 2 (i.e., $m_{\text{max},1}=2$). The outputs of these two experiments are illustrated in Fig. 8. As can be seen in Fig. 8, the resulting $R^{2}$ scores of the test set are equal to 0.22298 and 0.34728, respectively, for the first and second experiments. Accordingly, we conclude that even increasing one mode beyond 2 drastically negatively affects the performance of the proposed FNO framework for the current application. ### 5.3 Channel width of FNOs We further analyze the impact of different channel widths on the performance of the introduced deep learning framework. Based on our machine learning experiments, $R^{2}$ scores obtained for the channel width of 8, 27, 64, and 125 are 0.49904, 0.81618, 0.96815, and 0.94457, respectively. When the channel width decreases from 64 to 27 or to 8, a significant drop in the $R^{2}$ score is observed. Notably, increasing the channel width beyond 64 to 125 also leads to a slight decrease in the precision of permeability predictions. As discussed in Sect. 2.3, the choice of channel width is directly related to the number of trainable parameters, which are 30897, 163783, 828673, and 3096531 for each respective channel width. Moreover, the channel width also determines the size of the max pooling, representing the size of the global feature vector. Hence, optimizing channel width is critical. Small channel width leads to poor performance, whereas large channel width imposes high computational costs and memory allocation without necessarily a significant performance improvement. ### 5.4 Number of FNO units We investigate the effect of varying the number of FNO units (see Fig. 1). Deep learning experiments are conducted using one, two, three, four, and five units to assess the impact on the introduced FNO performance. By computing the $R^{2}$ score across the test set, we realize that there is no significant improvement in the prediction accuracy. $R^{2}$ score for the FNO configuration with one, two, three, four, and five units are respectively, 0.82767, 0.91703, 0.96813, 0.96759, and 0.97818. Hence, adding more units (beyond 3) and making the network deeper does not have a remarkable effect on the prediction accuracy. However, the number of trainable parameters and consequently, the computational cost and required GPU memory (e.g., RAM) escalated by adding FNO units. For example, 820353, 824513, 828673, 832833, and 836993 are, respectively, the number of trainable parameters of the model with one, two, three, four, and five layers of FNOs. ### 5.5 Activation functions We give particular attention to the effect of choosing an activation function on the prediction ability of our FNO model. In the primary setup, we configure all layers to employ the ReLU activation function, except the last layer of the classifier, where we utilize a sigmoid function. We implement two alternative setups. In the first one, we alter the activation function of the last layer to ReLU, this configuration results in a drastic reduction in the $R^{2}$ score of the test set, regardless of if the output permeability is normalized between 0 and 1. In the second setup, we replace the activation function in all layers with sigmoid. As a consequence of this setup, a slight decrease in performance is indicated, as $R^{2}$ score of 0.91478 is obtained for the test set. Note that the training procedure becomes slower in this setup, as the derivative of the sigmoid function results in a more complicated computation graph compared to that one output by the derivation of the ReLU function. ### 5.6 Static max pooling versus static average pooling Within the context of capturing global features in the proposed FNO-based framework, we explore the efficacy of implementing static average pooling as an alternative to static max pooling. Our machine learning experiment yields a $R^{2}$ score of 0.94478 in this case, demonstrating a marginal diminishment in the network performance compared to the presence of static max pooling. As supported by the literature (Qi et al., 2017a, b; Kashefi et al., 2021; Kashefi and Mukerji, 2022; Kashefi et al., 2023), max pooling is a preferred technique for classification tasks compared to average pooling. Our finding shows a similar pattern for the introduced FNO-based framework. \begin{overpic}[width=433.62pt]{CNN36.png} \put(5.0,90.0){a} \end{overpic} \begin{overpic}[width=433.62pt]{CNN44.png} \put(5.0,90.0){b} \end{overpic} \begin{overpic}[width=433.62pt]{CNN52.png} \put(5.0,90.0){c} \end{overpic} \begin{overpic}[width=433.62pt]{CNN60.png} \put(5.0,90.0){d} \end{overpic} Figure 9: A few examples of synthetically generated three-dimensional digital porous media for examining the generalizability of the proposed neural network; a an image of size $36^{3}$, b an image of size $44^{3}$, c an image of size $52^{3}$, and d an image of size $60^{3}$ \begin{overpic}[width=433.62pt]{C26.pdf} \put(5.0,90.0){a} \end{overpic} \begin{overpic}[width=433.62pt]{C27.pdf} \put(5.0,90.0){b} \end{overpic} \begin{overpic}[width=433.62pt]{C28.pdf} \put(5.0,90.0){c} \end{overpic} \begin{overpic}[width=433.62pt]{C29.pdf} \put(5.0,90.0){d} \end{overpic} Figure 10: $R^{2}$ plots demonstrating the generalizability of the proposed approach in classifying multi-sized images. The network, trained on images of sizes $40^{3}$, $48^{3}$, and $56^{3}$, is used to predict images of sizes a $36^{3}$ (375 data), b $44^{3}$ (375 data), c $52^{3}$ (375 data), and d $60^{3}$ (375 data). ### 5.7 Generalizability In this subsection, we assess the generalizability ability of the proposed FNO-based framework. Note that the concept of generalizability in the context of the present work extends to the network’s performance to predict the permeability of cubic porous media with unseen sizes. As discussed in Sect. 4, the network was initially trained using porous media with cubic geometries of sizes $40^{3}$, $48^{3}$, and $56^{3}$. To examine the network capacity to generalize, we predict the permeability of porous media with sizes $36^{3}$, $44^{3}$, $52^{3}$, and $60^{3}$ with 375 cubes for each of these sizes using our pretrained FNO-based framework. Figure 9 shows a few examples of these synthetic data, generated for the purpose of examining the network generalizability. As shown in Fig. 10, a slight decline is observed in the accuracy of permeability predictions for porous media with unseen sizes. However, the obtained $R^{2}$ scores remain in an excellent range. These scores are 0.93185, 0.91124, 0.91500, and 0.90844 for the porous media sizes of $36^{3}$, $44^{3}$, $52^{3}$, and $60^{3}$, respectively. As another observation, the performance of our approach is marginally higher in predicting the permeability of unseen porous media with smaller cubic sizes. As highlighted in Fig. 10, $R^{2}$ scores of porous media with sizes of $36^{3}$ are greater than ones with a size of $44^{3}$. A similar scenario occurs when we compare porous media of sizes $52^{3}$ and $60^{3}$. This can be attributed to the fact that, for smaller sizes, the fixed-size vector of the global feature encodes the features of smaller cubes more effectively. Moreover, note that the vector size is the same as the width channel. As a last comment in this subsection, to enhance the network’s generalizability, a potential strategy could involve expanding the training dataset to include more than the initial three sets of geometry sizes. \begin{overpic}[width=433.62pt]{Loss_mode2.pdf} \put(5.0,90.0){a} \end{overpic} \begin{overpic}[width=433.62pt]{Loss5_V2.pdf} \put(5.0,90.0){b} \end{overpic} Figure 11: Evolution of the loss function for the validation and training sets using a the proposed approach (see Fig. 1) and b the intuitive approach (see Fig. 2) ### 5.8 Comparison with intuitive approach #### 5.8.1 Classification of fixed-sized image For the comparison between the proposed approach (see Fig. 1) and the intuitive approach (see Fig. 2), we consider the problem of predicting the permeability of porous media with fixed cubic sizes. Specifically, we consider a size of $48^{3}$. Similar outputs are observed for other sizes. To ensure a fair comparison, both methodologies are investigated under similar conditions. Specifically, both methods are set to have an approximately equal number of trainable parameters (i.e., 828738 for the intuitive strategy and 828673 for our approach). Accordingly, the size of the vector representing the global feature is 64 in both methods. All other parameters such as the number of modes in each direction, the number of FNO units, the classifier architecture, and size, are the same in both methods and are set as those listed in Sect. 4 (i.e., the training section). Our results demonstrate that both methods perform proficiently, with the $R^{2}$ score of 0.99348 and 0.97360 over the test set for the intuitive approach (see Fig. 2) and the proposed approach (see Fig. 1), respectively. The evolution of the loss function for the training and validation sets indicates a convergence after approximately 3000 epochs. This deep learning experiment confirms an approximately equivalent computational cost between the two approaches. Hence, when the image size of training data is fixed, both strategies are effective for the defined image classification task and there is no significant advantage for one method over the other, according to our analysis. As a last point in this subsection, we note that one may also use static max pooling in the architecture of the traditional approach since the size of porous media is fixed in this experiment. Based on our results, the performance does not change. #### 5.8.2 Classification of multi-sized images In this subsection, we compare the performance of the proposed approach (see Fig. 1) with the intuitive approach (see Fig. 2) in predicting the permeability of porous media with varying sizes. The evolution of both training and validation losses is depicted in Fig. 11. Figure 11 indicates a divergence between the training and validation losses for the network used in the intuitive approach, which suffers from overfitting, whereas this is not the case for the proposed approach. The superiority of the proposed approach is also evident by the $R^{2}$ score obtained for the test set. Accordingly, the $R^{2}$ scores of the proposed approach and the intuitive approach are respectively 0.96809 and $-0.42632$. The negative value of the $R^{2}$ score for the intuitive approach demonstrates that its model makes worse predictions than a model that simply predicts all outputs as the mean value of the dataset. Note that changing hyper-parameters, such as the number of modes, channel width, and number of FNO layers, does not improve the model of the intuitive approach. This flaw stems from two reasons. First, using the intuitive approach, the network captures the global feature after lifting cubes into the original space, while the trainable parameters of the network are mainly defined in the Fourier space. Second, the adaptive max pooling’s size is altered depending on the size of the input cubic porous medium. These two together lead to a misrepresentation of the global feature of cubes with different sizes, when the network tends to predict the permeability of the validation and test sets. Note that in Sect. 5.8.1, we showed that the intuitive approach worked well when it was trained over porous media with fixed sizes. However, the result of our machine learning experiments illustrates that the global features of cubes with different sizes are amalgamated. In contrast, our approach uses static max pooling consistent with the channel width of Fourier neural operators before transitioning back to the original space. This approach enables the capture of global features prior to changing spaces. ## 6 Summary and future outlooks In this research study, we introduced a novel deep learning framework based on Fourier neural operators for classifying images with different sizes (see Fig. 1). Because Fourier neural operators are resolution invariant, they have the potential to be used for the task of multi-sized image classification. To reach this goal, Fourier neural operators must be connected to a classifier, ideally using a pooling operator. To this end, we proposed the novel idea of implementing a static max pooling operator, which functions in a high dimensional space with the size of Fourier channel width. We showed the efficiency and robustness of this framework by predicting the permeability of three-dimensional digital porous media with three different sizes of $40^{3}$, $48^{3}$, and $56^{3}$. We explored the effect of key parameters such as the number of Fourier modes in each dimension, the channel width of the discrete Fourier space, activation functions in different layers, and the number of Fourier units. Additionally, we showed that while the network was only trained on the porous media with the sizes of $40^{3}$, $48^{3}$, and $56^{3}$, it could successfully predict the permeability of the porous media with the sizes of $36^{3}$, $44^{3}$, $52^{3}$, and $60^{3}$, indicating its generalizability. Moreover, we demonstrated that the idea of implementing an adaptive max pooling (see Fig. 2), as an intuitive approach for connecting the FNO layers to the classifier, showed a lack of performance when predicting the permeability of porous media of varying sizes. Note that the adaptive max pooling operated in spatial spaces and that pooling had to be adaptive to handle input images with varying sizes. As a future research direction, we aim to adapt the current architecture and extend its capabilities to image classification. In contrast to the problem of permeability prediction, this approach reduces the problem’s dimensionality to two. Additionally, given that the standard dataset for image classification is usually large, we anticipate improved generalizability of the proposed framework. ## Data availability The Python code for the three-dimensional problems is available on the following GitHub repository, https://github.com/Ali- Stanford/FNOMultiSizedImages. ## Acknowledgements Financial support by the Shell-Stanford collaborative project on digital rock physics is acknowledged. Additionally, the first author would like to thank Prof. Gege Wen at Imperial College London for her helpful guidance and discussion on the software engineering aspects of this study. ## References * Abadi et al. (2015) Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G.S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg, J., Mané, D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J., Steiner, B., Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V., Viégas, F., Vinyals, O., Warden, P., Wattenberg, M., Wicke, M., Yu, Y., Zheng, X., 2015. TensorFlow: Large-scale machine learning on heterogeneous systems. URL: https://www.tensorflow.org/. software available from tensorflow.org. * Andra et al. (2013a) Andra, H., Combaret, N., Dvorkin, J., Glatt, E., Han, J., Kabel, M., Keehm, Y., Krzikalla, F., Lee, M., Madonna, C., Marsh, M., Mukerji, T., Saenger, E.H., Sain, R., Saxena, N., Ricker, S., Wiegmann, A., Zhan, X., 2013a. Digital rock physics benchmarks—part i: Imaging and segmentation. Computers & Geosciences 50, 25–32. doi:https://doi.org/10.1016/j.cageo.2012.09.005. benchmark problems, datasets and methodologies for the computational geosciences. * Andra et al. (2013b) Andra, H., Combaret, N., Dvorkin, J., Glatt, E., Han, J., Kabel, M., Keehm, Y., Krzikalla, F., Lee, M., Madonna, C., Marsh, M., Mukerji, T., Saenger, E.H., Sain, R., Saxena, N., Ricker, S., Wiegmann, A., Zhan, X., 2013b. Digital rock physics benchmarks—part ii: Computing effective properties. Computers & Geosciences 50, 33–43. doi:https://doi.org/10.1016/j.cageo.2012.09.008. benchmark problems, datasets and methodologies for the computational geosciences. * Blunt et al. (2013) Blunt, M.J., Bijeljic, B., Dong, H., Gharbi, O., Iglauer, S., Mostaghimi, P., Paluszny, A., Pentland, C., 2013. Pore-scale imaging and modelling. Advances in Water Resources 51, 197–216. doi:https://doi.org/10.1016/j.advwatres.2012.03.003. 35th Year Anniversary Issue. * Bonev et al. (2023) Bonev, B., Kurth, T., Hundt, C., Pathak, J., Baust, M., Kashinath, K., Anandkumar, A., 2023. Spherical fourier neural operators: Learning stable dynamics on the sphere. arXiv preprint arXiv:2306.03838 . * Brandstetter et al. (2023) Brandstetter, J., van den Berg, R., Welling, M., Gupta, J.K., 2023. Clifford neural layers for pde modeling. arXiv:2209.04934. * Chen et al. (2023) Chen, G., Liu, X., Li, Y., Meng, Q., Chen, L., 2023. Laplace neural operator for complex geometries. arXiv preprint arXiv:2302.08166 . * Choubineh et al. (2023) Choubineh, A., Chen, J., Wood, D.A., Coenen, F., Ma, F., 2023. Fourier neural operator for fluid flow in small-shape 2d simulated porous media dataset. Algorithms 16. URL: https://www.mdpi.com/1999-4893/16/1/24, doi:10.3390/a16010024. * Darcy (1856) Darcy, H., 1856. Les fontaines publiques de la ville de Dijon: exposition et application des principes à suivre et des formules à employer dans les questions de distribution d’eau. volume 1. Victor Dalmont. * Das et al. (2018) Das, M.K., Mukherjee, P.P., Muralidhar, K., 2018. Porous Media Applications: Biological Systems. Springer International Publishing. * Goodfellow et al. (2016) Goodfellow, I., Bengio, Y., Courville, A., 2016. Deep Learning. MIT Press. http://www.deeplearningbook.org. * Gupta and Brandstetter (2022) Gupta, J.K., Brandstetter, J., 2022. Towards multi-spatiotemporal-scale generalized pde modeling. arXiv:2209.15616. * Hao et al. (2023) Hao, Z., Wang, Z., Su, H., Ying, C., Dong, Y., Liu, S., Cheng, Z., Song, J., Zhu, J., 2023. Gnot: A general neural operator transformer for operator learning, in: International Conference on Machine Learning, PMLR. pp. 12556–12569. * Hong and Liu (2020) Hong, J., Liu, J., 2020. Rapid estimation of permeability from digital rock using 3d convolutional neural network. Computational Geosciences 24, 1523–1539. doi:10.1007/s10596-020-09941-w. * Hua and Lu (2023) Hua, N., Lu, W., 2023. Basis operator network: A neural network-based model for learning nonlinear operators via neural basis. Neural Networks 164, 21–37. * Huang et al. (2023) Huang, X., Shi, W., Meng, Q., Wang, Y., Gao, X., Zhang, J., Liu, T.Y., 2023. Neuralstagger: accelerating physics-constrained neural pde solver with spatial-temporal decomposition. arXiv preprint arXiv:2302.10255 . * Jiang et al. (2023a) Jiang, P., Yang, Z., Wang, J., Huang, C., Xue, P., Chakraborty, T.C., Chen, X., Qian, Y., 2023a. Efficient super-resolution of near-surface climate modeling using the fourier neural operator. Journal of Advances in Modeling Earth Systems 15, e2023MS003800. doi:https://doi.org/10.1029/2023MS003800. e2023MS003800 2023MS003800. * Jiang et al. (2023b) Jiang, Z., Zhu, M., Li, D., Li, Q., Yuan, Y.O., Lu, L., 2023b. Fourier-mionet: Fourier-enhanced multiple-input neural operators for multiphase modeling of geological carbon sequestration. arXiv:2303.04778. * Johnny et al. (2022) Johnny, W., Brigido, H., Ladeira, M., Souza, J.C.F., 2022. Fourier neural operator for image classification, in: 2022 17th Iberian Conference on Information Systems and Technologies (CISTI), pp. 1–6. doi:10.23919/CISTI54924.2022.9820128. * Kabri et al. (2023) Kabri, S., Roith, T., Tenbrinck, D., Burger, M., 2023. Resolution-invariant image classification based on fourier neural operators, in: Calatroni, L., Donatelli, M., Morigi, S., Prato, M., Santacesaria, M. (Eds.), Scale Space and Variational Methods in Computer Vision, Springer International Publishing, Cham. * Kashefi et al. (2023) Kashefi, A., Guibas, L.J., Mukerji, T., 2023. Physics-informed PointNet: On how many irregular geometries can it solve an inverse problem simultaneously? application to linear elasticity. Journal of Machine Learning for Modeling and Computing 4, 1–25. * Kashefi and Mukerji (2021) Kashefi, A., Mukerji, T., 2021. Point-cloud deep learning of porous media for permeability prediction. Physics of Fluids 33, 097109. * Kashefi and Mukerji (2022) Kashefi, A., Mukerji, T., 2022. Physics-informed pointnet: A deep learning solver for steady-state incompressible flows and thermal fields on multiple sets of irregular geometries. Journal of Computational Physics 468, 111510. doi:https://doi.org/10.1016/j.jcp.2022.111510. * Kashefi and Mukerji (2023) Kashefi, A., Mukerji, T., 2023. Prediction of fluid flow in porous media by sparse observations and physics-informed pointnet. Neural Networks 167, 80–91. * Kashefi et al. (2021) Kashefi, A., Rempe, D., Guibas, L.J., 2021. A point-cloud deep learning framework for prediction of fluid flow fields on irregular geometries. Physics of Fluids 33, 027104. doi:10.1063/5.0033376. * Keehm et al. (2004) Keehm, Y., Mukerji, T., Nur, A., 2004. Permeability prediction from thin sections: 3d reconstruction and lattice-boltzmann flow simulation. Geophysical Research Letters 31. doi:https://doi.org/10.1029/2003GL018761. * Kingma and Ba (2014) Kingma, D.P., Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 . * Kontolati et al. (2023) Kontolati, K., Goswami, S., Shields, M.D., Karniadakis, G.E., 2023. On the influence of over-parameterization in manifold based surrogates and deep neural operators. Journal of Computational Physics 479, 112008. * Kumeria (2022) Kumeria, T., 2022. Advances on porous nanomaterials for biomedical application (drug delivery, sensing, and tissue engineering). ACS Biomaterials Science & Engineering 8, 4025–4027. doi:10.1021/acsbiomaterials.2c01103. * Lantuejoul (2002) Lantuejoul, C., 2002. Geostatistical Simulation: Models and Algorithms. Springer. * Le Ravalec-Dupin et al. (2004) Le Ravalec-Dupin, M., Roggero, F., Froidevaux, R., 2004. Conditioning truncated gaussian realizations to static and dynamic data. SPE Journal 9, 475–480. doi:10.2118/84944-PA. * Lee (2022) Lee, S., 2022. Mesh-independent operator learning for partial differential equations, in: ICML 2022 2nd AI for Science Workshop. * Lehmann et al. (2024) Lehmann, F., Gatti, F., Bertin, M., Clouteau, D., 2024. 3d elastic wave propagation with a factorized fourier neural operator (f-fno). Computer Methods in Applied Mechanics and Engineering 420, 116718. * Li et al. (2022a) Li, Z., Huang, D.Z., Liu, B., Anandkumar, A., 2022a. Fourier neural operator with learned deformations for pdes on general geometries. arXiv preprint arXiv:2207.05209 . * Li et al. (2020a) Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A., 2020a. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895 . * Li et al. (2020b) Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A., 2020b. Neural operator: Graph kernel network for partial differential equations. arXiv preprint arXiv:2003.03485 . * Li et al. (2023) Li, Z., Kovachki, N.B., Choy, C., Li, B., Kossaifi, J., Otta, S.P., Nabian, M.A., Stadler, M., Hundt, C., Azizzadenesheli, K., Anandkumar, A., 2023. Geometry-informed neural operator for large-scale 3d pdes. arXiv:2309.00583. * Li et al. (2022b) Li, Z., Peng, W., Yuan, Z., Wang, J., 2022b. Fourier neural operator approach to large eddy simulation of three-dimensional turbulence. Theoretical and Applied Mechanics Letters 12, 100389. * Li et al. (2021) Li, Z., Zheng, H., Kovachki, N., Jin, D., Chen, H., Liu, B., Azizzadenesheli, K., Anandkumar, A., 2021. Physics-informed neural operator for learning partial differential equations. arXiv preprint arXiv:2111.03794 . * Liang and Fletcher (2023) Liang, Y., Fletcher, D., 2023. Computational fluid dynamics simulation of forward osmosis (fo) membrane systems: Methodology, state of art, challenges and opportunities. Desalination 549, 116359. doi:https://doi.org/10.1016/j.desal.2022.116359. * Liu et al. (2023) Liu, M., Ahmad, R., Cai, W., Mukerji, T., 2023. Hierarchical homogenization with deep-learning-based surrogate model for rapid estimation of effective permeability from digital rocks. Journal of Geophysical Research: Solid Earth 128, e2022JB025378. doi:https://doi.org/10.1029/2022JB025378. * Lyu et al. (2023) Lyu, Y., Zhao, X., Gong, Z., Kang, X., Yao, W., 2023. Multi-fidelity prediction of fluid flow based on transfer learning using Fourier neural operator. Physics of Fluids 35, 077118. doi:10.1063/5.0155555. * Majumdar et al. (2023) Majumdar, R., Karande, S., Vig, L., 2023. How important are specialized transforms in neural operators?, in: 1st Workshop on the Synergy of Scientific and Machine Learning Modeling @ ICML2023. URL: https://openreview.net/forum?id=DU3Z6ZdqhZ. * Masroor et al. (2023) Masroor, M., Emami Niri, M., Sharifinasab, M.H., 2023. A multiple-input deep residual convolutional neural network for reservoir permeability prediction. Geoenergy Science and Engineering 222, 211420. doi:https://doi.org/10.1016/j.geoen.2023.211420. * Maust et al. (2022) Maust, H., Li, Z., Wang, Y., Leibovici, D., Bruno, O., Hou, T., Anandkumar, A., 2022. Fourier continuation for exact derivative computation in physics-informed neural operators. arXiv preprint arXiv:2211.15960 . * Meng et al. (2023) Meng, Y., Jiang, J., Wu, J., Wang, D., 2023. Transformer-based deep learning models for predicting permeability of porous media. Authorea Preprints . * Paszke et al. (2019) Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al., 2019. Pytorch: An imperative style, high-performance deep learning library, in: Advances in Neural Information Processing Systems, pp. 8024–8035. * Pathak et al. (2022) Pathak, J., Subramanian, S., Harrington, P., Raja, S., Chattopadhyay, A., Mardani, M., Kurth, T., Hall, D., Li, Z., Azizzadenesheli, K., et al., 2022. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214 . * Peng et al. (2024) Peng, W., Qin, S., Yang, S., Wang, J., Liu, X., Wang, L.L., 2024. Fourier neural operator for real-time simulation of 3d dynamic urban microclimate. Building and Environment 248, 111063. * Peng et al. (2023a) Peng, W., Yuan, Z., Li, Z., Wang, J., 2023a. Linear attention coupled Fourier neural operator for simulation of three-dimensional turbulence. Physics of Fluids 35, 015106. doi:10.1063/5.0130334. * Peng et al. (2023b) Peng, Z., Yang, B., Liu, L., Xu, Y., 2023b. Rapid surrogate modeling of magnetotelluric in the frequency domain using physics-driven deep neural networks. Computers & Geosciences 176, 105360. * Poels et al. (2023) Poels, Y., Derks, G., Westerhof, E., Minartz, K., Wiesen, S., Menkovski, V., 2023. Fast dynamic 1d simulation of divertor plasmas with neural pde surrogates. arXiv preprint arXiv:2305.18944 . * Qi et al. (2017a) Qi, C.R., Su, H., Mo, K., Guibas, L.J., 2017a. Pointnet: Deep learning on point sets for 3d classification and segmentation, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 652–660. * Qi et al. (2017b) Qi, C.R., Yi, L., Su, H., Guibas, L.J., 2017b. Pointnet++: Deep hierarchical feature learning on point sets in a metric space, in: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc. * Rahman et al. (2022a) Rahman, M.A., Florez, M.A., Anandkumar, A., Ross, Z.E., Azizzadenesheli, K., 2022a. Generative adversarial neural operators. arXiv preprint arXiv:2205.03017 . * Rahman et al. (2022b) Rahman, M.A., Ross, Z.E., Azizzadenesheli, K., 2022b. U-no: U-shaped neural operators. arXiv preprint arXiv:2204.11127 . * Renn et al. (2023) Renn, P.I., Wang, C., Lale, S., Li, Z., Anandkumar, A., Gharib, M., 2023. Forecasting subcritical cylinder wakes with fourier neural operators. arXiv preprint arXiv:2301.08290 . * Sun et al. (2023) Sun, H., Zhou, L., Fan, D., Zhang, L., Yang, Y., Zhang, K., Yao, J., 2023. Permeability prediction of considering organic matter distribution based on deep learning. Physics of Fluids 35, 032014. doi:10.1063/5.0142574. * Tembely et al. (2020) Tembely, M., AlSumaiti, A.M., Alameri, W., 2020. A deep learning perspective on predicting permeability in porous media from network modeling to direct simulation. Computational Geosciences 24. doi:10.1007/s10596-020-09941-w. * Thodi et al. (2023) Thodi, B.T., Ambadipudi, S.V.R., Jabari, S.E., 2023. Fourier neural operator for learning solutions to macroscopic traffic flow models: Application to the forward and inverse problems. arXiv:2308.07051. * Tran et al. (2023) Tran, A., Mathews, A., Xie, L., Ong, C.S., 2023. Factorized fourier neural operators. arXiv:2111.13802. * Wen et al. (2022) Wen, G., Li, Z., Azizzadenesheli, K., Anandkumar, A., Benson, S.M., 2022. U-fno—an enhanced fourier neural operator-based deep-learning model for multiphase flow. Advances in Water Resources 163, 104180. * White et al. (2023a) White, C., Berner, J., Kossaifi, J., Elleithy, M., Pitt, D., Leibovici, D., Li, Z., Azizzadenesheli, K., Anandkumar, A., 2023a. Physics-informed neural operators with exact differentiation on arbitrary geometries, in: The Symbiosis of Deep Learning and Differential Equations III. * White et al. (2023b) White, C., Tu, R., Kossaifi, J., Pekhimenko, G., Azizzadenesheli, K., Anandkumar, A., 2023b. Speeding up fourier neural operators via mixed precision. arXiv preprint arXiv:2307.15034 . * Wu et al. (2018) Wu, J., Yin, X., Xiao, H., 2018. Seeing permeability from images: fast prediction with convolutional neural networks. Science Bulletin 63, 1215–1222. doi:https://doi.org/10.1016/j.scib.2018.08.006. * Xi et al. (2022) Xi, J., Ersoy, O.K., Cong, M., Zhao, C., Qu, W., Wu, T., 2022. Wide and deep fourier neural network for hyperspectral remote sensing image classification. Remote Sensing 14. URL: https://www.mdpi.com/2072-4292/14/12/2931, doi:10.3390/rs14122931. * Xie et al. (2023) Xie, C., Zhu, J., Yang, H., Wang, J., Liu, L., Song, H., 2023. Relative permeability curve prediction from digital rocks with variable sizes using deep learning. Physics of Fluids 35, 096605. doi:10.1063/5.0167998. * Xiong et al. (2023) Xiong, W., Ma, M., Sun, P., Tian, Y., 2023. Koopmanlab: machine learning for solving complex physics equations. arXiv preprint arXiv:2301.01104 . * Yang et al. (2022) Yang, H., Li, Z., Sastry, K., Mukhopadhyay, S., Anandkumar, A., Khailany, B., Singh, V., Ren, H., 2022. Large scale mask optimization via convolutional fourier neural operator and litho-guided self training. arXiv preprint arXiv:2207.04056 . * Yang et al. (2023) Yang, Y., Gao, A.F., Azizzadenesheli, K., Clayton, R.W., Ross, Z.E., 2023. Rapid seismic waveform modeling and inversion with neural operators. IEEE Transactions on Geoscience and Remote Sensing 61, 1–12. doi:10.1109/TGRS.2023.3264210. * You et al. (2022) You, H., Yu, Y., D’Elia, M., Gao, T., Silling, S., 2022. Nonlocal kernel network (nkn): a stable and resolution-independent deep neural network. Journal of Computational Physics 469, 111536. * Zhao et al. (2022) Zhao, J., George, R.J., Zhang, Y., Li, Z., Anandkumar, A., 2022. Incremental fourier neural operator. arXiv preprint arXiv:2211.15188 . * Zhao et al. (2024) Zhao, X., Chen, X., Gong, Z., Zhou, W., Yao, W., Zhang, Y., 2024. Recfno: a resolution-invariant flow and heat field reconstruction method from sparse observations via fourier neural operator. International Journal of Thermal Sciences 195, 108619. * Zhao et al. (2023) Zhao, X., Sun, Y., Zhang, T., Xu, B., 2023. Local convolution enhanced global fourier neural operator for multiscale dynamic spaces prediction. arXiv:2311.12902. * Zhu et al. (2023a) Zhu, M., Feng, S., Lin, Y., Lu, L., 2023a. Fourier-deeponet: Fourier-enhanced deep operator networks for full waveform inversion with improved accuracy, generalizability, and robustness. arXiv preprint arXiv:2305.17289 . * Zhu et al. (2023b) Zhu, M., Zhang, H., Jiao, A., Karniadakis, G.E., Lu, L., 2023b. Reliable extrapolation of deep neural operators informed by physics or sparse observations. Computer Methods in Applied Mechanics and Engineering 412, 116064. * Zou et al. (2023) Zou, C., Azizzadenesheli, K., Ross, Z.E., Clayton, R.W., 2023. Deep neural helmholtz operators for 3d elastic wave propagation and inversion. arXiv:2311.09608.
# Invisible, Unreadable, and Inaudible Cookie Notices: An Evaluation of Cookie Notices for Users with Visual Impairments James M Clarke University of Surrey<EMAIL_ADDRESS>& <EMAIL_ADDRESS>Maryam Mehrnezhad Royal Holloway University of London, <EMAIL_ADDRESS>Ehsan Toreini University of Surrey, <EMAIL_ADDRESS>&<EMAIL_ADDRESS> ###### Abstract This paper investigates the accessibility of cookie notices on websites for users with visual impairments (VI) via a set of system studies on top UK websites (n=46) and a user study (n=100). We use a set of methods and tools–including accessibility testing tools, text-only browsers, and screen readers, to perform our system studies. Our results demonstrate that the majority of cookie notices on these websites have some form of accessibility issues including contrast issues, not having headings, and not being read aloud immediately when the page is loaded. We discuss how such practises impact the user experience and privacy and provide a set of recommendations for multiple stakeholders for more accessible websites and better privacy practises for users with VIs. To complement our technical contribution we conduct a user study and finding that people with VIs generally have a negative view of cookie notices and believe our recommendations could help their online experience. We also find a disparity in how users wish to respond to cookie notices as apposed to how they do in reality. ## 1 Introduction Visual impairment (VI) is a term used to describe any type of vision loss, ranging from partial vision loss to someone who cannot see at all [1, 2]. People with VI have various types of assistive technologies (AT) available to help them browse the internet [56], e.g., text-only browsers and screen readers. Screen readers are installed on users’ computers or phones to read information by outputting it as sound [49, 56, 65]. They work with the browser and interpret the code that is used to build web pages [6]. Screen readers are not capable of conveying visual and spatial information, such as layout and images, to the user unless relevant meta-information is provided in the web page code through _markups_. In addition, the text is only presented line by line, making it harder to get an overview of the page [49]. This also makes it more difficult to understand the relations of a website’s different parts and identify navigation links. To ensure that AT can correctly interpret websites, there are various accessibility standards, such as the Web Content Accessibility Guidelines (WCAG) provided by the World Wide Web Consortium (W3C) [63]. WCAG aims to provide a shared standard for web content accessibility. The WCAG documents explain how to make web content more accessible to disabled people. To be included in the WCAG, issues must impact disabled people with greater effect than those without disabilities [64]. The majority of websites employ some type of tracking, using various techniques such as cookies and fingerprinting [12]. There are two types of cookies, functional and non-functional [30], with the most common use of non- functional cookies being for personalised advertising [70, 55, 59]. A simple method to counteract this type of tracking is to allow users to manage which cookies are stored on their device [53]. With the implementation of the General Data Protection Regulation (GDPR) in 2018, companies operating in the EU and the UK and/or handling EU/UK citizens data need to choose a legal basis to collect and process user data [48]. One of the most well-known of these is cookie notices to gain consent from users [33]. Alongside the GDPR, the ePrivacy Directive and the Information Commissioner Office (ICO) give specific guidance on obtaining consent through cookie notices [42, 34]. Previous research has shown that individuals want to protect themselves from online tracking [8, 47], though they are not always confident [8, 36]. Multiple studies have looked at how the function and presentation of cookie notices differ [36, 34, 24, 61, 9]. Similarly, there are studies showing that the designs of cookie notices can affect users’ interactions [61], including through dark patterns [40]. Previous research has examined the effect of GDPR and cookie notices on the number of cookies [30, 9, 24]. It has been shown that there is a disparity between the requirement of data protection laws, the practises of websites, and users behaviour regarding online tracking protection [36]. Limited research has been conducted on privacy and VIs. Users with VIs have previously been found to have concerns about being tracked online [25], similar to others [8, 47]. There has also been research looking at VI and online information credibility [6]. In the context of cookie notices and VI, research is extremely sparse [52]. There are some reports on usability issues with cookie notices while looking at the wider accessibility of websites [66]. To the best of our knowledge, there is no research on cookie notices and AT where a comprehensive range of methods is utilised. Our research questions include: * • RQ1: How do websites and cookie notices comply with the web content accessibility guidelines and the general data protection regulations? RQ1-a: How do popular websites comply with the current accessibility guidelines (e.g., WCAG) and the GDPR? RQ1-b: Does compliance necessarily mean good privacy practices for VI users? * • RQ2: Can the existing automated accessibility tools evaluate cookie notices? RQ2-a: How do the current cookie notices score with the automated accessibility tools (e.g., WAVE and Google Lighthouse)? RQ2-b: Does a high score necessarily mean good practice for VI users? * • RQ3: How do cookie notices interface with AT? RQ3-a: How does the mainstream AT (e.g., text-only browsers and screen readers) interact with cookie notices? RQ3-b: How do the current practises impact VI users’ privacy? * • RQ4: What are the general perception and practice of VI users regarding cookie notices? RQ4-a: What issues have VI users encountered with cookie notices? RQ4-b: Who do participants believe is responsible for online accessibility? This paper contributes to the body of knowledge via its system studies, user studies, and the discussions and recommendations that we provide for improving the online privacy of users with VIs. First, we provide a set of evaluation methods based on the off-the-shelf tools for AT and for users with VI. This enables us and other researchers to conduct system experiments and assess websites and cookie notices for their accessibility. Second, using these methods and tools, we run experiments on 46 popular UK websites (according to Alexa) and report a wide range of accessibility issues with their cookie notices. Table 1 presents an overview of our system studies. Third, we conduct user studies with 100 UK participants who use AT and extract their perception, practices, and preferences regarding cookies notices on websites. The results of our systems studies as well as the user studies confirm that current practises are far from ideal in protecting the privacy of users with VIs. Finally, we discuss the impact of these practises on user privacy and provide recommendations for web developers, AT designers, policymakers, and end users to improve the privacy of real-world practises. ## 2 Background and Related Work Differential Vulnerabilities recognise how different populations face different types and degrees of security and privacy risks [44]. This challenges the universalising tendencies that frame cybersecurity around an abstract or generic user who either does not exist or is only a subset of actual end users [11, 35]. This ties into social sciences research looking at models of disability such as the Critical Realist Model [14]. Both of these threads consider the real-world lived experiences of disabled people, as well as their thoughts. With differential vulnerabilities considering how different threats can arise for different user groups. Studying and evaluating the privacy of users with VIs is challenging. The common range of privacy assessment methods would not be directly useful here. Instead, such approaches should be combined with accessibility assessment methods, as defined in the accessible writing guidelines of the Association for Computing Machinery [21]. According to the Office of National Statistics in 2020, almost 11 million adults with disabilities recently used the Internet in the UK [41]. A 2016 GOV.UK survey of 712 AT users found that 29% used a screen reader to browse the Internet, while others used screen magnifiers, speech recognition or readability software [18]. They also found several different screen readers being used, the most popular being JAWS. WebAIM also found that JAWS was the most popular screen reader with 53.7% of users using it as their primary screen reader and NVDA was the second most popular with 30.7% of users using it as their primary screen reader [68]. ### 2.1 Users with VI and Privacy Evaluating the accessibility of websites is possible through a number of automatic and manual ways and through the use of a range of tools such as screen readers and text-only browsers. For example, Southwell and Slater have previously used the WCAG to evaluate university library finding aids [56]. They used an automated web-accessibility checker, WAVE 5.0 by WebAim, to perform an initial assessment of each finding aid and then manually tested each website using the WebbIE 3 Web browser, which is a non-graphical text- only browser. They also used screen readers directed by keyboard navigation including _System Access to GO_ and _NVDA_. When using the automated checker, they found that most of the websites tested (58 of 65) had at least one accessibility error. The most common errors were missing document language, missing alternative text, missing form labels, and linked images missing alternative text. They then used the non-graphical browser, finding only 68% had headings that enabled navigation to another important section of the document. Of those which had enough headings, they did not always have the headings in proper sequential order, or were missing first-level headings. Fewer sites offered links for navigation, 57% did, 43% did not, and 25% of the sites lacked both headings and links for any kind of navigation. Using the screen readers, they found that the main content of all 65 finding aids was readable; this opposes the 89% error rate noted by the automatic checker. Table 1: Overall view of our system studies Method | (I) Cookie notice | (II) General | (III) Manual | (IV) Manual ---|---|---|---|--- | & Tracking | Automated | Testing via | Testing via | Behaviour Evaluation | Accessibility Tools | Text-only Browser | Screen Readers Tools | Google Chrome, Brave | WAVE, | WebbIE | JAWS, NVDA | | Google Lighthouse | | Website Accessibility | NA | Yes | Yes | Yes Assessment | | | | Cookie Notice | Yes (General, Baseline) | Partial (Accessibility) | Yes (Accessibility) | Yes (Accessibility) Assessment | | | | There is scarce research on the security and privacy of users with VI. Brulé et al. analysed 178 papers on technologies designed for people with VI, with the aim of facilitating and stimulating future research [5]. Inan et al. surveyed 20 individuals who are visually impaired to investigate their internet use and explore their cybersecurity challenges and concerns while browsing the internet [25]. They found a number of problems, such as automatic web page refreshing and missing or improper headings. In this study, the possibility of someone tracking their internet activities was the highest- rated concern. The authors suggest that it is important to guide the user to enable security and privacy settings and to provide accessible software solutions to protect and warn this marginalised group. Hayes et al. shadowed and interviewed people with VI and their allies [23]. Finding that self- perceptions can influence someone’s behaviour, which could have privacy and security implications, such as hiding or concealing their disability due to perceived stigma. Akter et al. studied 155 people with VIs privacy concerns relating to camera-based AT [3]. Finding that users of these systems were more concerned about the privacy of others, who may inadvertently be captured in their images, than themselves. However, camera-based AT can create a lack of personal security in the lives of the people they are trying to help. Previous research reports that users with VIs often find it difficult to complete their security task and that they had moderately high levels of concern about cybersecurity [39]. Similarly, there are reports on the complications of authentication methods such as passwords and two-factor authentication for users with VIs [51]. An exploratory user study, conducted using semi- structured in-person interviews with 14 people with VIs found that participants were aware and concerned about privacy and security and faced a variety of risks [2]. More relevant to this paper is the work of Schnell and Roy [52]. They conducted an evaluation of a select group of 40 educational and financial website cookie notices using WCAG [52]. Finding that even for users without disabilities, there were challenges to accessing, understanding, and processing privacy information. Also, finding that educational websites were more accessible than financial websites, however, not all websites complied with the WCAG’s criteria chosen for their testing. In contrast to this work, we offer a comprehensive evaluation method to review website cookie notices and apply our methods to a range of websites rather than only educational and financial. Although there have been a number of user studies looking generally at users who are VI and security, to the best of our knowledge, there have been none looking specifically at cookie notices. In this paper, we aim to address this gap via a series of system studies and a dedicated user study with users who have VIs, both focusing on cookie notices. ### 2.2 AT Regulations, Standards, and Tools According to GDPR, cookie notices should be presented on all websites that use cookies and should include opt-out options, as well as opt-in options without highlighting the latter and including any privacy or cookie walls. Cookie notices should be separated from other matters such as privacy policy and terms and conditions, and the user should be able to opt-out of the previously accepted cookie settings with the same ease as they gave the consent. Enabling non-essential cookies before the user’s consent is a non-compliant practice too. Based on Article 12 et seq. GDPR [16]: “The controller shall take appropriate measures to provide any information referred to in Articles 13 and 14 and any communication under Articles 15 to 22 and 34 relating to processing to the data subject in a concise, transparent, intelligible, and easily accessible form, using clear and plain language, in particular for any information specifically addressed to a child. The information shall be provided in writing, or by other means, including, where appropriate, by electronic means. When requested by the data subject, the information may be provided orally, provided that the identity of the data subject is proven by other means.” This article interprets as the data controller (i.e. web tracker in the context of this paper), must inform every user about the nature of the data to be collected and the purposes of such collection. Hence, websites need to be fully compliant with the regulations and also offer usable practises to comply with further requirements. The ambiguity of how practises should include marginalised users has not been discussed widely, and only limited examples are available. For example, in the verdict of an Italian case in which the data controller was mandated to provide the information acoustically for video surveillance [15]. There are many aspects to the real-world implementation of accessible web technologies [43]. For instance, an accessible web design approach should support enhancing the visual characteristics of the front–end design and utilise a range of colours, while ensuring the contrast of the colours is accessible to users who are visually impaired or colour–blind. Also, they need to build an audio commentary for the page and the images. The interconnected nature of web pages (as various resources fetched from different origins in the page) could potentially increase the complexity of fully accessible web design. To harmonise such practises, the W3C has provided a comprehensive list of 167 tools to evaluate accessibility compatibility measurements111w3.org/WAI/ER/tools/. They are implemented on a number of platforms and technologies, some supporting cross-platform products. These products include 20 support APIs, 14 authoring tool plugins, 45 browser plugins, 19 command line tools, 25 desktop applications, 4 mobile applications, and 90 online tools. There are a number of standards and regulations worldwide to provide accessibility requirements for the technologies to be considered as _publicly presentable_ in a region, country, or regional-based regulations e.g., European, Italian, Irish, Israeli, Japanese, Korean, and US Federal Law, platform (web accessibility frameworks), e.g., various versions of WCAG (2.1, 2.0 and 1.0), or file formats, e.g., EPUB standards. Looking at the standards based on the support for VI, we can conclude that all of them recognise such disabilities and provide a standardised set of guidelines for the implementation of such support. In general, they offer similar suggestions to mitigate such disabilities. For example, these standards support VI by advising to provide a form of AT and non-visual access to support visually impaired users, including the proprietary agent (in this case, the dedicated hardware or special browsers), an audio description to explain the important visual detail, the high contrast visualisation, adoption of flash thresholds, magnification, reduction of the required field of vision, and control of contrast, brightness, and intensity. Following these guidelines can contribute towards meeting the minimum requirements for complying with such regulations. In our observations, most accessibility tools are based on W3C standard family (WCAG 2.1 (85 tools out of 167), WCAG 2.0 (139 tools), WCAG 1.0 (46 tools)). Moreover, some of them comply with country-specific regulations such as German standards (21 tools), French standards (12 tools), Japanese standards (18 tools), EU standards (9 tools), US federal procurement (67 tools), Irish National IT Accessibility Guidelines (16 tools), Israeli web accessibility guidelines (7 tools), Italian accessibility legislation (11 tools), Korean standards (1 tool). Finally, format-specific standards such as EPUB accessibility 1.0 are only supported in 3 tools. ### 2.3 Online Tracking Adopted in April 2016 and implemented in May 2018, the GDPR changed the rules on online tracking and consent (including consent to cookies) [48, 33]. In order to process personal data, companies must choose a legal basis for processing. One of the most well-known is consent. Valid consent must be freely given, specific, informed, unambiguous, explicit, revocable, given prior to any data collection, and requested in a readable and accessible manner [48]. The ePrivacy Directive (“ePD”, aka “cookie law”) [42], provides supplementary rules to the GDPR. According to the ePD website, publishers must rely on user consent when collecting and processing personal data using non- mandatory (not strictly necessary for the services requested by the user) cookies or other technologies [42]. This is in accordance with the guidance given by the European Data Protection Board and the ICO [13, 27]. Various studies (e.g., [61, 9, 34, 59, 36, 24]) exist on the implementation and effectiveness of cookie notices, privacy banners, and tracking practises. Examples of dark patterns include providing invalid consent, nudging the user, making opting-out difficult, not providing the user with opting-out options from previously accepted cookies settings, pre-enabling non-essential cookies, and including trackers in the cookie notice itself. For example, the top 5 consent management platforms have been reported to use dark patterns and implied consent [40]. There is a body of knowledge on the user dimensions of tracking, including concerns and negative feelings of users about tracking [47], differences between demographics such as gender and country [8], the disparity between regulations, website practises and users’ limited knowledge for protecting against tracking and their demand for more transparency and control [36, 54, 46, 37, 60]. What is lacking in the previous work is the measurement of the current practises in the wild for web tracking notices for users with visually impairments. In this paper, we aim to run experiments in order to fill this gap. ## 3 Accessibility Evaluation Methodology In this section, we present our methodology for the evaluation of the websites. Our assessment includes a number of different methods and tools including automated accessibility testing tools, a non-graphical browser and screen readers, as explained at length in this section. The overall design of our experiments and the tools used in each part is presented in Table 1. We have included a website analysis template in Appendix A. All experiments took place between April and October 2022 on a laptop PC running Windows with a screen size of 13.3 inches and a resolution of 3840 x 2160. Windows is the most commonly used desktop OS among screen reader users according to the WebAIM 2021 survey [68]. As a case study, we use Alexa’s top 50 UK websites in April 2022. We selected this sample since GDPR is a regional regulation (EU/UK). We looked at English websites for analysis in our fluent language. Based on Alexa, the popular UK websites are comparable to others in Europe, e.g., Germany and France. From this list, four websites are excluded because they redirect to another website already on the list or are down. _t.co_ is an example of a website that was excluded due to redirecting to _twitter.com_ , however, both _amazon.com_ and _amazon.co.uk_ are retained. The US version of the site (.com) does not contain a cookie notice, whereas the UK version (.co.uk) does; therefore, it was important to keep both sites on the list for comparison. These are just examples, and the full list is presented in Table 8 (Appendix). The cookie notice experiments were conducted by two researchers to ensure consistency. A researcher performed accessibility testing twice with one specialist software/tool, by recording the results in tables (Appendix). Due to the rounds of experiments taking place over the course of six months, we believe this demonstrates stability in our results. ### 3.1 Cookie Notices All 46 websites were visited using Google Chrome (Version 103.0.5060.134 (Official Build) (64-bit)) and Brave222brave.com (Version 1.41.99 Chromium: 103.0.5060.134 (Official Build) (64-bit)). Using Google Chrome without a screen reader acts as a baseline and gave an example of how sighted users would see the site and the cookie notice. Chrome is one of the most popular browsers with the highest market share in 2022 [57]. Brave is a secure browser that was created in 2016 by two former executives of Mozilla Corporation, the company that makes the Firefox browser [32]. Brave comes with a feature called Brave Shields built in, which includes several privacy-preserving features. Brave adopts various privacy-enhancing techniques which are not possible at the browser extension level (due to access restrictions and performance limitations), making it a powerful tool to observe the tracking behaviours of websites. It is commonly used for assessing the tracking behaviour of websites on PC and mobile platforms [34, 36]. We completed these experiments before the introduction of cookie notice blocking by Brave [58]. For each of the 46 websites, we open them in these two browsers and record the location and control options given to users. When recording the details, we do not interact with the website in any way, including not interacting with notifications (e.g. requesting location permission, update notifications). To ensure that no cookies had previously been cached, each website is viewed in a new private or incognito window. We also record which options are given to the user in the cookie notice according to the categories suggested in similar work, e.g. [34, 36]. These categories include: (i) _Agree or Reject_ : where two options are presented, Agree (Agree, Accept, OK, Understand, etc.) or Reject (Reject, Decline, No, etc.), with the same level of control (e.g., two buttons). These are further categorised by which option is emphasised. (ii) _Agree or Settings_ : where two options are presented, Agree or Settings (Options, Settings, Policy, Manage, Learn more, etc.), again with the same level of control. Which are further categorised by which option is emphasised. (iii) _Agree, Reject, or Settings_ : where three options are presented; Agree, Reject, and Settings. These are further categorised on the basis of which item is highlighted in the notice. (iv) _No Notice_ : The website does not display a cookie notice. ### 3.2 Accessibility Evaluation Our accessibility evaluation consists of two parts. First, we use automated testing tools, which are designed to give developers an overview of how accessible their website is [63]. This allows us to get an impression of the overall accessibility of a website, and in some cases includes information about the accessibility of the cookie notice. Second, we use software designed for individuals with VI in the real world to assess the results of the automated testing tools and to allow us to more specifically focus on cookie notices. In this section, we explain these approaches. #### 3.2.1 Automated Accessibility Testing Tools Websites are evaluated using two different automated accessibility testing tools, WebAIM WAVE 5.0 Web Accessibility Evaluation Tool333wave.webaim.org/ and Google Lighthouse444developer.chrome.com/docs/lighthouse/overview/. WAVE is an automated accessibility tool that we use to perform an initial assessment of the conformance of each website to WCAG. WAVE generates a report containing Errors, Alerts, Features, Structural elements, and ARIA landmarks. Errors indicate issues that will impact certain users with disabilities, as well as showing failures to meet the requirements of the WCAG. Whereas alerts are elements which may cause accessibility issues, but need further manual testing to determine this. Features are elements that can improve accessibility if implemented correctly. Structural elements are HTML regions and headings, and ARIA can be used to present important accessibility information to people with disabilities. WAVE has been used in previous work, e.g. Southwell and Slater, when evaluating university library finding aids [56]. During their testing, they used WAVE to perform an initial evaluation of the conformity of each finding aid to Section 508 and WCAG 2.0 guidelines. We tested the web version of WAVE 5.0 in our preliminary testing and it did not detect any cookie notices. Therefore, we use the browser extension version555wave.webaim.org/extension/ for our experiments. The WAVE extension evaluates the rendered version of the web page allowing dynamically generated content to be evaluated [69], while the WAVE Web version may not be able to apply all the scripting on the page. This is a possible reason for the cookie notices not being displayed during our preliminary tests. We use Google Lighthouse to give an overall accessibility score, as well as to record specific problems with each website. Lighthouse is an open-source automated testing tool, which can audit performance, accessibility, and more [17]. We only test accessibility using the default (navigation) mode and while representing a desktop device. We record the score out of 100 and the individual issues with each website. This score is a weighted average of all accessibility audits it performs, with weighting based on axe user impact assessments [10]. The axe user impact assessments are composed of WCAG rules with some supplementary rules added [26]. Both WAVE and Google Lighthouse give an overview of accessibility for a whole website, however, WAVE also allowed us to view where specific problems occurred. Manual Testing via Text-only Browser: To complement and verify the results of the testing tools, we apply a range of methods to manually assess the privacy practises of these websites via their cookies notices. We visit all these websites using WebbIE666webbie.org.uk/, a text-only browser for people with VI. The WebbIE Ctrl-H and Ctrl-L commands are used to examine the heading and links on a page. This approach has been used in similar work, e.g., work [56]. WebbIE was uninstalled and reinstalled for each round of testing, as it does not have a private browsing mode or cookie manager. Through this method, we examine how users navigate the page and if and how cookie notices are displayed. We assign each website to one of the following categories: (i) _No Headings_ : The website in general has no headings which can be used for navigation. (ii) _Basic Headings_ : The website has some headings but there is a limited amount which is not useful for navigation. (iii) _Full Headings_ : The website has a number of headings that are useful for navigation. Headings allow screen readers and other accessibility software to navigate around a webpage. For example, WebbIE can move easily to different headings on a website allowing for quicker navigation and locating key information, e.g. a cookie notice. The categories above are derived from previous work [56], where similar categories were used to evaluate the accessibility of library finding aids. Similarly, we observe the website’s behaviour in presenting the cookie notice and each website’s cookie notice was also put into the following categories: _(i) Headings throughout_ : Headings are available throughout the cookie notice. _(ii) Heading at the start_ : A heading is present at the start of the notice, however, there were no other headings in the body of the notice. _(iii) No headings_ : There are no headings present in the cookie notice at all. _(iv) Notice missing_ : The cookie notice is not shown when using WebbIE, however, one is present when using the graphical browsers. _(v) No notice_ : The website does not have a cookie notice when viewed with the graphical browser. The _Headings throughout_ category for cookie notices is based on the _Full Headings_ category for the website as a whole. Meaning that a user would be able to navigate the cookie notice using heading-based navigation; this is particularly useful for longer cookie notices as seen on some websites. The _Heading at the start_ category is used to classify notices that only have a heading at the start. This would allow for navigation to the notice itself but means that a user would have to rely on a different type of navigation within the notice, e.g. line-by-line or link-based navigation. Whereas _No headings_ would mean a user would not be able to use heading navigation at all within the cookie notice and would have to rely on another form of navigation. In some instances when viewed graphically, a website did display a cookie notice, however, when using WebbIE one was not present. For this reason, we include a _Notice missing_ category to signify this. Whereas with the _No notice_ , a notice was not present on the website when viewed graphically. We included different categories for headings (Basic, Full, and No headings) since we found lengthy cookie notices on some websites (e.g., google.com, facebook.com), however, headings are not always needed due to a number of cookie notices being shorter. #### 3.2.2 Manual Testing via Screen Readers In order to have more comprehensive and conclusive results, we also carry out our experiments using screen readers to manually test each website. JAWS and NVDA were chosen as the most popular according to WebAIM [68], 53.7% and 30.7%, respectively. We use these screen readers in conjunction with Google Chrome as these are the most common combinations of screen reader and browser [68], 32.5% and 16.0%, respectively. NVDA is a free OS-level screen reader with support for popular applications such as web browsers, email clients, music players, and office programmes. JAWS is another OS-level screen reader that users need to purchase. For our experiments, we purchased a home licence (£865 with the Authorisation USB Dongle). Both screen readers should have similar reliability when parsing websites [67], however, they often parse website code slightly differently [45]. It is for this reason that we use the two most popular screen readers during our testing. We categorise cookie notices based on the way these screen readers are able to read them [62, 52]. Accordingly, each website’s cookie notice was given a pass or fail for the following categories: _(1) Readable_ : The screen reader software is able to read the cookie notice. _(2) Immediately Read_ : The cookie notice is the first thing to be read from the page, excluding the page title. _(3) Keyboard navigable_ : The cookie notice of a website is navigable using a keyboard while using a screen reader. _(4) Link or button purpose_ : The purpose of a link or button can be solely determined by the link or button. _(5) Abbreviations are explained_ : All abbreviations are explained. This was either in the cookie notice or the website offered some mechanism for identifying the expanded form. _(6) Page titled_ : The page has a title that describes its topic or purpose. _(7) Cookie notice titled_ : The cookie notice has a title or heading which is readable by the screen reader software. _(8) Headings useful for navigation_ : There are useful headings for navigation present throughout the cookie notice. _(9) No timing_ : There is no timing for reading the cookie notice. The _Readable_ category is based upon WCAG 3.1, Readable, defined as “Make text content readable and understandable” by WCAG [62], with the guideline being used in previous work [52]. We created the _Immediately Read_ category to show that a cookie notice is read close to the start of a web page. This is important as a number of websites start tracking a user before they respond to the notice, and therefore users must be able to respond to the notice at the first given opportunity. Also, meaning that users do not have to actively search the website for the notice to respond. The category _Link or button purpose_ is based on WCAG 2.4.9, Link purpose (link only), which is defined as “A mechanism is available to allow the purpose of each link to be identified from the link text alone, except where the purpose of the link would be ambiguous to users in general” [62]. _Abbreviations are explained_ is based upon WCAG 3.1, Abbreviations, which W3C define as “A mechanism for identifying the expanded form or meaning of abbreviations is available”. _Page titled_ is also based on a WCAG, namely 2.4.2 Page Titled which is defined as “Web pages have titles that describe topic or purpose”. We create another category based on this called _Cookie notice titled_ , this is to judge if a cookie notice can easily be navigated to. It also aligns with previous testing for headings, as titles often consist of headings. Alongside this, we test for _Heading useful for navigation_ , which is based on the previous heading testing with WebbIE in cookie notices. It also aligns with WCAG 2.4.10 Section headings, defined as “Section headings are used to organize the content”. We also define the category _No timing_ which is based on WCAG 2.2.3 No timing. This is defined as “Timing is not an essential part of the event or activity presented by the content, except for non-interactive synchronized media and real-time events”. ### 3.3 Limitations To the best of our knowledge, this is the first work on the assessment of cookie notices on a range of websites for users with VI. We simply chose to test the 46 top websites in the UK (out of 50). In practise, the top Alexa websites may not be the most popular websites for users with VI. However, we could not find a formal report on popular websites for this group of users. We acknowledge that this is a limited sample set and more research is required to evaluate a larger number of websites. When testing websites, we only tested the first page in our experiments. Although this is a common practise for the privacy assessment of websites in general, it is not clear if all pages would present the same information and produce the same output for AT. Further, detailed work would be needed to explain how different web pages interact with AT. Previous research has demonstrated the usefulness of mobile technology for people with VI, e.g. [19, 20]. However, due to the lack of research in this area, we only generally focus on desktop web browsers, for which the majority of the accessibility and AT tools and standards are also designed. Cross- platform studies are left as future work. ## 4 Accessibility evaluation Results Our results include (1) a general assessment of the cookie notices of the websites and their tracking behaviour, and (2) an accessibility evaluation of these websites and their cookie notices. ### 4.1 Cookie Notices and Tracking Behaviour Cookie Notice Position: We observed that the majority of websites displayed a cookie notice (n = 35 or 76.1%) when using Google Chrome. Of the positions, a bottom overlay was the most common (n = 15 or 32.6%), followed by a middle overlay (n = 7 or 15.2%). When using Brave, a higher number of web pages displayed no notice (n = 15 or 32.6%). Other than this, the popularity of categories is in the same order as that of Google Chrome. While there are some papers (e.g. [61, 4]) looking at cookie notice positions and user engagement, we could not find any for users with VI. Cookie Notice Control Options: Of the options given when using Google Chrome, Agree or Settings was the most common (17 or 37.0%). The most commonly emphasised option along with Agree or Settings was Agree (13 or 28.3%). Table 2 describes the options presented to users in Chrome and Brave. These results from Brave resemble those of Google Chrome; however, when using Brave, there was a higher percentage of websites which displayed no notice. These results are consistent with previous work (e.g., [34]) when cookie notices were evaluated across platforms. The cookie notices of some websites are blocked due to the notice itself being a tracker; resulting in it being blocked by Brave. Previous research studying GDPR compliance has focused on the following requirements: consent must be explicit, accepting all is as easy as rejecting all, and there are no pre-ticked boxes [40]. It has been shown that the pre- selection of options can impact users’ choices when giving consent [61]. For this reason and to respond to RQ1-a, in Table 2, we highlight categories that are in violation of the above requirements and therefore in violation of GDPR. As you can see, three categories (14 websites) comply with the above requirements. However, we did not test them for additional GDPR compliance items, such as opting out from previously accepted cookie notices with the same ease of opting in. Table 2: Cookie notices’ user control options in Chrome and Brave, as well as GDPR violations. | Emphasised | Browser | GDPR ---|---|---|--- Options | option | Chrome | Brave | violation (i) Agree or Reject | None | 4 | 4 | No | Agree | 4 | 4 | Yes (ii) Agree | None | 4 | 4 | Yes or Settings | Agree | 13 | 9 | Yes (iii) Agree, Reject | None | 5 | 5 | No or Settings | Agree & Reject | 5 | 5 | No (iv) No Notice | 11 | 15 | Yes Tracking Behaviour: We also observed these websites regarding their tracking behaviour through Brave. Before interacting with the cookie notices only 3 of the 46 websites (6.5%) did not have at least one item blocked by the Brave Shields without any interaction with the cookie notice. The average number of items blocked was 9, the maximum was 81, 11 of the websites had more than 10 items blocked, and 6 had more than 20. The majority of items blocked were in the _trackers & ads_ category. Our results support similar work e.g., [34, 36, 33] reporting that the majority of websites start tracking the user regardless of the presence of the cookie notice before any user interaction with it. ### 4.2 Automated Accessibility Testing Tools in Websites WAVE ran on all but one website; when using it on _ebay.co.uk_ , the overlay containing the results did not appear. Of the remaining sites, 42 (93.3%) contained at least one accessibility error, with the average number of errors being 18.98. Of the websites tested 35 (77.8%) contained at least one contrast error. All websites tested contained at least one structural element with the average being 84.02. Table 3 shows a summary of the results. We further break these down into categories which could cause issues, e.g. errors, contrast errors, and alerts, and those which could improve user experience, e.g. features, structural elements, and ARIA. Table 3: Summary of WAVE 5.0 test results for 46 websites. Criteria | no. of websites | Average no. of items ---|---|--- | with at least one | across websites | item per criteria | Cookie Notice | 33 | - Errors | 42 | 18.98 Contrast Errors | 35 | 22.98 Alerts | 45 | 124 Features | 46 | 77.16 Structural Elements | 46 | 84.02 ARIA | 43 | 235.42 Errors are general issues that cause problems such as missing labels from HTML code. While a contrast error would cause issues for someone with vision loss, e.g., light text on a light background or dark text on a dark background. Alerts are criteria that need further testing to establish if they hinder or help accessibility. For example, for an image with long alternative text, a long description could be needed to fully describe the image, or it may be unjustified. Features are elements which work to improve a user’s experience. For example, a form label is present and associated with a form control. This is similar to structural elements such as headings and lists which also help the user’s experience. ARIA is a set of roles and attributes that define ways to make websites more accessible to people with VI. An example of an error could be missing alternative text or a form control that does not have a corresponding label. ARIA is only useful if implemented correctly such as when an ‘aria-label’ or ‘aria-labelledby’ attribute is present which can be interpreted by AT. To complement this, we also note whether a cookie notice was present when testing using WAVE. In some instances, we observed specific issues with the cookie notice. The most common problems were with the low contrast between the background of the cookie notice and the text, links, or buttons. Table 8 (Appendix) shows detailed results. Overall, the website with the lowest number of issues was bbc.co.uk, with 0 errors, 0 contrast errors, 134 alerts, 23 features, 119 structural elements and 371 ARIA. There were multiple websites with close to the same number of issues, namely xvideos.com, spankbang.com and xnxx.com, all of which had between 165 and 176 items which would cause issues. In addition, we used Google Lighthouse for the overall accessibility score of the website. The average score was 89% (highest: 100%, lowest: 63%). Since Google Lighthouse uses the axe user impact assessments, the overall score is affected largely in the same way as individual WAVE tests. For example, including a button that has an accessible name will improve the overall score given to a web page. In general, these popular websites had a range of good and poor accessibility practises when tested with these automated accessibility tools. There were several websites we tested that achieved the best possible Lighthouse score of 100%: bbc.co.uk, wikipedia.org, gov.uk, paypal.com, microsoft.com, linkedin.com, and doubleclick.net. The lowest score, 63%, was achieved by tiktok.com. which, according to Lighthouse, had a number of labels and names missing, as well as navigation and contrast issues. These scores for each website are shown in Table 8 (Appendix). ### 4.3 Manual Cookie Notice Accessibility Testing via AT Tools Manual Testing via Text-only Browser: By using a text-only browser, we performed an analysis on the overall accessibility of the websites and their cookie notices. When using WebbIE, 27 of the 46 (58.7%) websites contained _full headings_ which would be useful for navigation. With 7 (15.2%) of them only having _basic headings_ and 12 (26.0%) containing _no headings_. The inclusion of headings throughout the website does not directly impact privacy, and was included in this analysis to give context to the accessibility of cookie notices. Percent of websites58.7%15.2%26.1%Headings at the start (13.0%)No headings (17.4%)Notice missing (15.2%)No notice (13.0%)Headings throughout (2.2%)No headings (4.4%)Notice missing (4.4%)No notice (4.4%)No headings (2.2%)Notice missing (17.4%)No notice (6.5%)Full HeadingsBasic HeadingsNo Headings Figure 1: WebbIE accessibility testing; inner circle: the whole site, outer circle: the cookie notice. When observing the cookie notices, 17 (48.6%) of the 35 websites which previously displayed a cookie notice did not display a cookie notice when using WebbIE (_notice missing_). Furthermore, only 1 (2.9%) of the 35 websites which had previously displayed a cookie notice website had _headings throughout_ , and 6 (17.1%) had a _heading at the start_ of the notice. 11 (31.4%) of the websites’ cookie notices contained _no headings_ , although, the majority of the websites which did not contain a notice did include links to privacy and cookies. Regardless of the number of headings throughout the website, we often found that cookie notices were missing. However, when a website had full headings the cookie notice was more likely to have a heading at the start. The results of these tests are shown in Figure 1, with the inner circle representing the headings in the website as a whole and the outer circle specifically looking at the cookie notice. The use of a heading at the start of the cookie notice can make it easier to locate, due to this the lack of headings seen in our testing could lead to problems. Headings inside the notice can also make it easier to navigate within a cookie notice, especially if it is lengthy, and therefore easier to make a decision. Manual Testing via Screen Readers: When testing with NVDA, 29 (82.9%) of the 35 websites, which graphically included a cookie notice, contained a cookie notice which could be read aloud. This result is higher than was expected following the other test. However, it still means that 6 of the cookie notices could not be read at all with NVDA. Of the cookie notices that could be read, 20 of the 35 (57.1%) were read aloud immediately when the website loaded. Others were read aloud after other elements of the page had been read or had to be specifically located to be read. 27 (77.1%) of the 35 cookie notices were keyboard navigable, these were not always the same websites as those read immediately. Therefore, this leaves 8 websites which users with VI may not be able to navigate. In some cases, these cookie notices created keyboard traps that the user would not be able to leave. Only 5 (14.3%) of the 35 cookie notices contained a link or button that could be solely determined by the link or button. Hence, without allowing time for the screen reader to output the notice, the user may not understand what they are agreeing to. Although all 46 (100%) websites contained a title, only 19 of the 35 (54.2%) cookie notices contained a title. This means it would not be possible to navigate to them using the heading, it could also make it more difficult to search for the cookie notice. 35 of the 35 (100%) cookie notices did not have any type of time limit on replying to the cookie notice. This is an excellent result, meaning that users will have time to ingest the information and make a decision. None of the 7 websites which contained abbreviations explained them, meaning that if users are unfamiliar with these terms they may not understand what they are consenting to. Also, none of the 35 websites’ privacy policies contained headings which were useful for navigation, however, some did contain different links. Due to this, it may be difficult to navigate the cookie notices, which is particularly important for some of the longer cookie notices we observed. We summarise these results in Table 4 (detailed results in Table 9 (Appendix)). In comparison, JAWS enabled 34 (97.1%) of the 35 websites with a cookie notice to be read out loud. Meaning all but one of the cookie notices could be read aloud, which is a significantly better result than when using NVDA. Of these, 22 (62.9%) of the 35 were read aloud immediately when the website loaded, which again is higher than when using NVDA. 29 of the 35 (82.9%) were keyboard navigable, this is an improvement of 2 cookie notices over NVDA. The number of cookie notices with a link or button that could be solely determined by the link or button was also higher at 11 (31.4%) of the 35. All of the other results were the same for JAWS as NVDA. These results are summarised in Table 4 and detailed results are available in Table 9 (Appendix). The reason for the disparity in results is due to the fact that the screen readers parse webpages differently, resulting in differing numbers of readable notices. This underscores the importance (lack) of standardisation efforts. Table 4: Number of websites which passed and failed each criterion of the manual testings via NVDA and JAWS. | NVDA | JAWS ---|---|--- Criteria | Pass | Fail | Pass | Fail Readable | 29 | 6 | 34 | 1 Immediately read | 20 | 15 | 22 | 13 Keyboard navigable | 27 | 8 | 29 | 6 Link or button purpose | 5 | 30 | 11 | 24 Abbreviations are explained | 0 | 7 | 0 | 7 Page titled | 46 | 0 | 46 | 0 Cookie notice titled | 19 | 16 | 19 | 16 Headings useful for navigation | 0 | 35 | 2 | 33 No timing | 35 | 0 | 35 | 0 We identified poor practices on some of these websites. For instance, a news website (dailymail.co.uk) read out adverts immediately before reading anything else such as the navigation bar or the cookie notice. This is highlighted in Figure 6 (Appendix). This is despite the fact that this website’s cookie notice is displayed using a large portion of the website. Another example was an online payment site (paypal.com), which read the body of the website aloud before reading the cookie notice. This aligns with the cookie notice being visually at the bottom of the page; however, this means that a user with VI using a screen reader could easily miss the cookie notice. An example of the visual representation of this notice and a scripted output of the website while using JAWS is available in Figure 8 (Appendix). We highlight this example, however, a similar output was common across multiple websites. One social news website (reddit.com) was the only website with a cookie notice which could not be read with either screen reader, even with intervention with mouse input. Visually the cookie notice was located at the bottom of the window, however, it could not be selected with the screen readers. A visual example of the cookie notice is included in Figure 7 (Appendix). In contrast, some of the websites presented the user with reasonable options when using a screen reader. For instance, bbc.co.uk clearly presented the users with opt-in and settings options. A scripted version of such output via NDVA screen reader is provided in Figure 9 (Appendix). ## 5 User Study Methodology In this Section, we explain the design of our online survey, data collection, and analysis. ### 5.1 Questionnaire Design When designing this survey, we followed the design principles for questionnaires for people with VIs [28]. Specifically, we aimed to inform participants about the topic of the survey before beginning the questionnaire, we indicate the type of answer after each question, and we also start each question with a consecutive number. We conducted an accessibility evaluation before conducting the survey, using the tools mentioned above, during which we did not find any issues. Our questionnaire is made up of five sections—Internet and AT, Privacy- enhancing technology usage, Cookie notices, Suggestions, and Demographics—with the complete questionnaire included in Appendix B. Internet and AT: After verifying the screening questions, we ask our participants a few background questions about technology usage, such as what devices and AT they use. We list different AT technologies including screen readers, braille displays, text only browsers, magnification software, and assistive browser extensions based on our research as well as allowing participants to add additional items. PETs usage: Next, after a brief explanation of PETs, we ask participants which PETs they use listening to them according to the categorisation suggested in [8], where the authors measure the correlation of people’s feelings about tracking and their protective actions. These categories include: browsers do not track, virtual private networks, private browsing, password manger, privacy-focused web browsers, encrypted messaging apps, ad blockers, file encryption tools, we additionally allowed participants to name other tools. Cookie notices: We also ask our participants what they think cookie notices are and what they think they are supposed to do. As well as how they feel about cookie notices and how they interact with them. For the design of these questions we followed [36, 8]. We ended this section by asking if they have encountered issues with cookie notices, what they were, and why they think they happen. Suggestions: Finally, informed by the results of our website experiments, we ask participants who they believe should be responsible for ensuring the privacy and accessibility of websites. We also ask which of our suggestions would help improve their experience online. Demographics: We conclude by asking demographic questions. ### 5.2 Data Collection and Analysis We conducted our user studies via Prolific Academic777prolific.co/ among UK participants. We conducted one initial testing round of the survey with 10 participants, asking for feedback upon completion. We fixed minor typos and made a few structural changes accordingly. At this stage and throughout further participation we received no complaints relating to accessibility of the questionnaire. We then distributed the questionnaire to a further 100 participants. We chose to use Prolific Academic to distribute the survey as this user group is notoriously difficult to recruit, therefore using a paid platform allowed us to gain this sample size. We rewarded participants at a rate of £12 per hour, which is categorised as ‘Great’ by Prolific Academic. This research received full approval from the University of Surrey’s Ethical Committee before the research commenced. Our method for processing the collected data is a mix of quantitative and qualitative analysis. For our free-text style questions, we run thematic analysis [7]; taking an inductive approach and allowing the data to determine our themes. We are confident that adopting a deductive approach would have yielded comparable themes. Two researchers conducted the thematic analysis independently, and due to the small sample size all authors discussed and agreed on these themes. Our research focused on exploring potential differences between users with visual impairments and those of previous work, allowing data to determine themes. For lengthy and complex responses multiple themes were assigned to them. We also chose participant responses which represent themes for inclusion in the paper. ### 5.3 Limitations The interaction and intersection of online services and AT could be a complex research topic to be investigated by user studies. To complement the technical findings of this paper, we ran our studies on an online platform and through a survey that provided us with self-reported responses which has its own limitations. We plan to extend our user studies to one-to-one interviews as well as focus groups to gain a deeper understanding of the implications of this user group. ## 6 User Study Results In this section, we present our findings of the user study. Our study was completed by 100 participants who self certify as using AT and live in the UK. Our participants occupy different jobs ranging from students, educators, healthcare and social assistance to business, hospitality, and some not working. 59 participants identify as male, 38 as female, 2 as non-binary, and 1 choosing not to say. ### 6.1 AT and PETs Usage Half of the users surveyed use magnification software, 42 used a screen reader, 22 use an assistive browser extension, with 9 not using any AT while browsing the web, and other forms of AT having less than 15 users. Participants use various online services, with the most popular being email (98), shopping and e-commerce (93), and social media (90). Table 10 (Appendix) gives an overview of the demographic questions. In response to Q2.1, all but 3 participants use at least one of the PET categories suggested, Figure 2 shows detailed results. The most popular technology used was a password manager (67%), and the least popular was file encryption tools (11%). 010203040506070Percent of participantsPassword managerAd blockersVirtual Private Network (VPN)Manual opt-outPrivate browsingBrowsers do not trackEncrypted messaging appsPrivacy-focused web browsersFile encryption toolsNone Figure 2: Q2.1: Which of the following privacy enhancing technologies do you use? (multiple choice). We also asked about the ways these users learn about PETs. Participants reported different ways including recommendation of friends / social contacts, being informed at work / school, and news. Only 19% of participants said that they learn about these PETs via the privacy/cookie policy of a website. ### 6.2 Cookie Notices When we asked our participants about their understanding of a cookie notice (Q3.1), they described it via different terms and we observed a few themes, where one third of our participants mentioned ‘tracking’ with a negative tone. For instance, P94 said: “It is a pop up that appears on virtually every website I visit these days. Can be quite annoying since it collects data, but I tend to reject the tracking cookies if possible”. In response to Q3.2 on the feelings of the participants about cookie notices (Table 5). Around half of our participants expressed negative feelings, one third had neutral feelings, and around a quarter expressed positive feelings regarding cookie notices. For instance, P9 said: “I don’t have any specific feeling about them just something that’s there.” and P33 said “I don’t like them, they are made difficult to understand on purpose, in order to make the user click ”Accept”. They need to be made more simple.” Table 5: Q3.2: How do you feel about cookie notices? Category | Examples | N ---|---|--- Strongly negative | Don’t trust, Intrusive | 24 | Very bad, Frustrating | Negative | Dislike, Don’t understand | 19 | Confusing | Neutral | Okay, Not bothered | 31 | Don’t care | Positive | Important, Essential | 26 | Useful | In Q3.3 and Q3.4, we asked the participants how they interact with cookie notices (Table 6). The responses varied across categories including agree, decline, ignore, edit cookie settings, get rid of it, and use other PETs. Except those who said they would agree to the cookie notice (47%), all the other categories included the word “try” in some of the responses e.g., “try to decline” and “try to edit the settings and say no.” Interestingly, 7% of participants spoke about trying to get rid of the cookie notice in any possible way by e.g., responding quickly. P46 said: “generally tick as little as possible to view the page and also reject where I can if not[,] I have to accept if the page [is needed]”. Where as P13 said “I try to reject them but this can be very difficult- I find they are often deliberately set up to make it impossible to read.” Table 6: Q3.3: How do you interact with cookie notices? Category | Examples | N ---|---|--- Agree | Accept, Say yes, Agree | 47 Decline | Reject, Reject, Cancel, Disagree | 34 Ignore | Ignore, Skip it | 8 Edit cookie settings | Edit cookie setting/notice | 7 Get rid of it | Make it go away, Respond quickly | 7 Use PET | Clear cookies later/regularly | 6 In response to more questions in this category (Q3.7 and Q3.8), we found a gap between the actual handling of cookie notices vs their preferred way. For instance, 20% of participants said they agreed to cookie notices in reality, when they wanted to act differently. Figure 3 shows the differences for each category. 01020304050Percent of participantsUnable to respondOtherIgnore cookie noticeEditing the settingsDisagreeAgreeWishReality Figure 3: Q3.7: How would you like to handle cookie notices? and Q3.8: How do you actually respond to cookie notices? ### 6.3 Issues with Cookie Notices For Q3.5, the majority (59%) of participants said they had not encountered issues with cookie notices (Figure 7). The rest said they have experienced issues regarding cookie display or settings or described negative feelings such as frustration regarding them. P98 said that they had experienced “cookie notices blocking content on the page that, if not blocked, I could read and close the page without having to interact with the notice.” P50 said “some websites make it a bit difficult to reject all cookies, it’ll open up another page where you’ll have to individually select each tick box to reject.” Table 7: Q3.5: Please describe in your own words what type of issues have you experienced with cookie notices? Category | Examples | N ---|---|--- None | Nothing, None, No problem | 59 Display problems | Too big, Loading, Can’t find, Can’t read | 14 Cookie settings problems | Difficult to reject, Forced to accept | 13 Negative feelings | Too many, Tired of disabling, Annoying | 9 Other | Cookies full, Tried to disable | 2 However, when presented with a list of possible problems in Q3.6, only 20% said none. 79% of the participants said that they had experienced at least one, the most common being unclear response options in a cookie notice and being unable to leave a cookie notice. Detailed results for this question are in Figure 4. 0510152025Percent of participantsUnclear response options in cookie noticeUnable to leave cookie noticeUnable to answer cookie noticeNoneUnable to find cookie noticeUnable to enter cookie noticeLack of headings for cookie noticeCookie notice not being present via ATLow contrast cookie noticeCookie basket was too full Figure 4: Q3.6: Which of the following issues have you experienced? In a follow-up question (Q3.9), we asked what is the potential reason when participants cannot respond to a cookie notice. The responses of the participants fell into two main categories: technical issues (37%) or malicious behaviour (16%). Four participants explicitly mentioned issues with AT e.g., P27 said that “Assistive technology may not be picking up a notice that has been given.” For example, P15 said they believe its “because they’re trying to force you to accept by pretending it’s broken?” ### 6.4 Suggestions We asked our participants about the responsible stakeholders for accessibility and security/privacy of web services. In this multiple-choice question, several entities came up including: website developers (77%), policymakers (48%), end-users (24%), accessibility evaluation designers (18%), and AT designers (15%). In addition, in response to Q4.2 in which we listed a set of recommendations (based on our system studies), all participants thought that at least one of our suggestions would help to improve user experience. For example, 79% of participants believe accessibility-by-design in websites would help their experience. Figure 5 shows the popularity of other recommendations. We discuss these in Section 7 at length. 020406080Percent of participantsAccessibility-by-design in websitesAccessibility testing by web designersInclusion of more headingsImproving the related lawsImproving the related specificationsDevelopment of more specific testing toolsEnd user’s engagement with cookie noticesAccessibly testing of sections of websitesDesigning AT- friendly PETs Figure 5: Q4.2: Which of these recommendation do you think would help improve your experience online? (multiple choice) ## 7 Discussion In this section, we discuss our results across our system studies and user study. ### 7.1 Website Accessibility and User Privacy In response to RQ2-a, we found that 93.3% of websites contained at least one accessibility error and 77.8% contained at least one contrast error. This means that most websites tested are not compliant with the WCAG success criteria and, therefore, could be inaccessible, difficult for people with VI to access, or cause access issues. The most common error during our testing of cookie notices was low-contrast buttons or links. The WCAG criteria 1.4.3 and 1.4.6 give guidance for contrast, the minimum guidance is a contrast ratio of at least 4.5 to 1 with enhanced guidance of a contrast ratio of at least 7 to 1 [62]. For the websites that contained a contrast error, this means that they did not meet the minimum guidance and, therefore, could make text difficult to read for people with VI. Alongside this, we found a number of websites that had no errors in their cookie notices but contained errors elsewhere on the page. Suggesting that the overall accessibility landscape is inadequate, this aligns with previous research, e.g. [22]. Our results align with previous work, reporting that the majority of university library finding aids had at least one accessibility error [56]. These errors with contrast could affect users who have vision loss but are not fully blind. Due to this group of people being larger than those who are fully blind, this result is concerning. Low contrast could cause users to miss important links or become confused about where to give or reject consent. For example, previous research has shown that higher contrasts between text and background colour led to faster searching [31], as well as affecting reading speed [50]. It has also been shown in a requirement survey that links can cause usability issues for users with VI [71]. ### 7.2 Cookie Notice Accessibility Issues In response to RQ3-a, we have categorised cookie notice accessibility issues including text-only browser issues, keyboard traps, and visual presentation of cookie notices vs. screen reader output. We explain each category here. Text-only browser issues: WebbIE was used to manually examine the heading contained within a website and its cookie notice. We found that 58.7% of the websites contained headings that could be useful for navigation, 15.2% containing basic headings and 26.0% containing no headings at all. Only 2.9% of websites contained headings throughout their cookie notice, with 17.1% having a heading at the start. A number of cookie notices did not appear when using WebbIE, this is most likely due to WebbIE being built using the Microsoft Web Browser object which gives a program its own internal IE [29]. In June 2022, Microsoft officially ended support for IE for some OSs [38]. It is therefore likely that web pages stopped supporting IE due to it being a legacy browser, and this caused these websites not to work with WebbIE. Keyboard traps: It was found that 77.1% of websites that contained cookie notices were keyboard navigable when using NVDA. The most common problem found was having to intervene and use a mouse, an option that is not feasible for people with VI. There were two main times when a mouse was needed. Firstly, to get NVDA to read the cookie notice, as some websites required the user to click on the cookie notice to interact with it. The other issue was escaping the cookie notice as there were websites that trapped the user in the cookie notice. This directly contradicts the WCAG success criteria 2.1.2, which is rated at level A. Whereas when using JAWS 82.9% of websites that contained cookie notices were keyboard navigable. Due to how JAWS operates a higher number of privacy policies could be read, with fewer of them creating a keyboard trap. This is most likely due to how different screen readers handle CSS code differently [67]. Visual presentation vs screen reader audio output: Only 14.3% of the cookie notices contained buttons or links whose use could be determined solely by the button or link when using NVDA. Whereas 31.4% of the cookie notices contained buttons or links whose use could be determined solely by the button or link. This difference was due to JAWS reading out alternative text associated with buttons on some websites. An example of this is where a button might visually only say _Accept all_ whereas when read aloud using JAWS it says “Accept the use of cookies and other data for the purposes described”. This change gives the user significantly more context on what the button does for them and allows them to skip the reading of the cookie notice. However, it could be argued that this could be done without the additional alternative text and, therefore, benefit both users without and with VI. For example, the accept button on a website simply reads _Accept all cookies_. Therefore, it was easy to ascertain the function of the button, only from this text, without the need for additional markup. ### 7.3 Reading Aloud the Cookie Notice When using a screen reader, the content of the web page is spoken out loud in a linear order, which may differ from the visual order on the screen [56]. When using WebbIE to view the web pages non-graphically, the cookie notices were often not at the start of the web page. To combat this, screen readers can also navigate using headings to jump to different sections. However, the lack of headings at the start of cookie notices makes it difficult to locate them when using this method. Screen readers can also search for content within a web page [49], but without a clear starting heading this becomes difficult. There were multiple websites where the cookie notice was not read aloud immediately, and the cookie notice also did not include a heading. In these examples, it would be difficult to navigate to the cookie notice, without either knowing what you are searching for or visually identifying it. When using NVDA with Chrome, 82.9% of cookie notices were read aloud, with 57.1% immediately after the website title was read. There were also websites that read the cookie notice quickly after opening but not immediately; for example, elements such as navigation bars were often read aloud before cookie notices. For example, one website (ebay.co.uk) reads the title of the page, then the navigation bar, and then the cookie notice. These were normally websites that did not display the notice at the top when using a browser graphically. In another example, graphically the cookie notice was at the bottom, but was read after the heading of the website, and before the main body. A possible reason for this is the hierarchy of the underlying HTML and CSS code. When using JAWS with Google Chrome, 97.1% of cookie notices could be read aloud, with 62.9% being read aloud immediately. The main issue when a cookie notice was not read immediately was that the user had to go through the whole page to read the cookie notice. As we showed in the results section, once loaded, these websites start collecting data at a large scale and even before user interaction with the cookie notice. When the cookie notice is the last item to be read to a VI user, it can easily distract the user from engaging with it leading to missing the cookie notice altogether. The results of our user study also confirm that cookie notices accessibility issues are indeed associated with negative feelings (RQ4-a). They also highlighted that there are a range of display issues with these cookie notices such as “they can’t be read”. This contributed to the gap we identified between the way that these users handle these cookie notices vs how they would like to handle them (Figure 3). ### 7.4 Website vs Cookie Notice Accessibility 100% of websites contained a title, while only 51.4% of cookie notices contained a title, which explained what it was. This result was the same for both screen readers. This lack of titles makes it more difficult to use headings to quickly navigate to the cookie notice. It is more of a problem when the notice is not immediately read aloud and then the user has to navigate to it. The lack of a title also means the user might miss the cookie notice. None of the 6 websites which contained abbreviations explained them on both screen readers. This lack of explanation affects the understanding of all users and directly contradicts WCAG success criteria 3.1.4. Regardless of this being a high-level success criterion, it is important in the context of cookie notices. Adding some type of mechanism for understanding abbreviations when they are used would help all users understand what they are agreeing to. In response to RQ3-b, we summarise the impact of the issues we encountered on users with VI. The fact that some cookie notices were missing when using the text-only browser means that users would not be able to respond to them. This also applies to the cookie notices that were not readable using screen readers. Similarly, users could not consent when cookie notices are not being read immediately, not including headings, and generally being difficult to navigate. Such a practice might require users to apply additional effort to specifically navigate to the notice. The lack of headings, structural elements, and explanatory buttons within the cookie notice means that it could take users with VI a longer time to respond to a cookie notice than other users. All these issues mean that these users are less protected against online tracking and cookies can be placed on their devices without any possibility for the users to know or give consent. ## 8 Recommendations In this section, we provide a set of recommendations and best practices for different stakeholders. Website developers: There are a number of ways for websites to have the maximum compatibility with the tools and software used by people with VI. When including a cookie notice, it should be close to the top of the document’s code. This will allow screen readers and other accessibility tools to quickly output this to the user. Developers could then use CSS code to change the visual location, meaning that a screen reader would always be able to read it aloud quickly. For example, developers may want visually to move it to the lower left corner (on desktop) or the lower part of the screen (on mobile) to improve the number of consent decisions for users without VI [61]. In addition, developers should always include a heading at the start of important content, whether this is a cookie notice or other important information. This allows for ATs to easily and quickly navigate to this information. It also allows users to quickly understand the content of the section they are about to interact with and therefore if this information is useful to them. To assess usability, developers should aim to use a multitude of tools. Tools such as WAVE and Lighthouse are aimed at allowing developers to easily evaluate a website. However, we showed in response to RQ2-b, they do not always highlight the problems users may face and high scores do not necessarily mean that a website is accessible. This is specifically the case when it comes to cookie notices and potentially other PETs. Therefore, more manual tests should be undertaken to find more nuanced issues with a website. Such testing should be conducted in a comprehensive manner and by multiple VI tools, since using a combination of such tools is a common practice for VI users. The tools we suggest are WAVE to perform automated testing, followed up by using a screen reader such as NVDA since it is free and relatively easy to use. Designers of accessibility evaluation tools: This research shows that accessibility tools and software available do not automatically assess websites for their privacy practises. The addition of the ability to test sub- sections of a website for accessibility issues would make testing elements of a website, such as a cookie notice, a simpler process. This would allow testing just to focus on such elements. In addition, this will enable subsets of development teams to test the accessibility of their work. In our testing with automated tools, it was often not clear where the errors and alerts were without further manual evaluation. However, this practise should not replace the overall accessibility testing of the website, but would allow more focus to be given to some areas of the website. Furthermore, the creation of specific accessibility tools and tests for cookie notices and other PETs would greatly improve real-world standard practises. Such tools could not only test the accessibility of the cookie notices, privacy policies, and other PETs, but also could evaluate the law compliance across platforms e.g., software, websites, and mobile apps. Policymakers: To respond to RQ1-b, we have performed both accessibility and GDPR compliance analysis. Overall if all websites comply with WCAG, it would benefit to all users, especially users with disabilities. The question of GDPR compliance is a more complicated question in relation to users with VI. GDPR bring many benefits for the privacy of users, however, in some cases, the implementation of cookie notices has affected the overall accessibility of websites. For this reason, we make the following recommendation to policymakers. The inclusion of specific guidelines for accessibility issues of user privacy which align with those included in GDPR and the ePD would generally improve the landscape. For example, guidelines on specific matters such as the length of time after loading that a cookie notice should be read aloud, what should be included in the content of the notices, and how should the options be presented to the users. Standardisation bodies can create comprehensive specifications for website developers and dedicated privacy sections. For example, a W3C specification which includes all the information that developers need to comply with legal frameworks, such as GDPR or The California Consumer Privacy Act (CCPA), as well as guidelines, such as WCAG. Such specifications can be also offered by Google and Apple for app developers in order to improve the privacy of VI users. End users: Generally, we believe that the onus around this issue should not be pushed onto end users, who are already a marginalised group. However, there are still additional steps users with VI could take. End-users who are concerned regarding cookie notices can manually search for them. All of the browsers tested have a feature to search within websites. Though, such a practise might not be needed in a near future due to the non-effective nature of cookie notices on websites. Several papers have reported that cookie notices are not practical and even when the user opts-out the websites still track the users. Some of these cookie notices are trackers themselves [33]. As a response, Brave has recently announced that it would automatically block cookie notices altogether [58]. This option could work to improve the privacy of users, along with the privacy-preserving nature of the Brave browser. Due to the browser being based on chromium, it would likely be just as accessible as Google Chrome. However, this remains an open research problem to be tackled in the future. In response to RQ4-b, we concluded that our participants believed that our set of recommendations can improve their online experience and privacy. Figure 5 displays the popularity of each item where accessibility-by-design in websites is rated top, followed by accessibility testing by web designers, inclusion of more headings, improvement of related laws/specifications, development of more specific testing tools, end user engagement with cookie notices, accessibility testing of sections of websites (including cookie notices), and designing AT- friendly PETs. ## 9 Conclusion This paper investigated the interaction between ATs and cookie notices via a set of system studies of 46 top UK websites and a user study of 100 users with VI via Prolific Academic. We find that 22 of these websites had at least one issue with the accessibility of their cookie notice when manually tested using a screen reader. We also observed websites which did not have issues with their cookie notices when using AT but did include issues such as low contrast when viewing them graphically. These practises often created accessibility issues when trying to read and respond to cookie notices. The results of our user study revealed that users with VI overall have a negative view on cookie notices. We also find that all users believe that at least one of our recommendations would help improve their experience online. In future work, we would like to conduct cross-platform studies looking at mobile web browsers, mobile apps, and desktop web browsers and their interaction with AT. We would also like to automate our methodology and run large-scale system studies. We also would like to focus on the creation and adaptation of dedicated accessibility testing tools for privacy matters and compliance with the law. ## Acknowledgements This research project has been granted ethical approval by the University of Surrey’s ethics committee. ## References * [1] Adeyemi, I., Sanders, C., Ong, B. N., Howells, K., Quinlivan, L., Gorman, L., Giles, S., Amp, M., Monaghan, E., Naseem, S., Pearson, A., and Cheraghi-Sohi, S. Challenges and adaptations to public involvement with marginalised groups during the COVID-19 pandemic: Commentary with illustrative case studies in the context of patient safety research. Research Involvement and Engagement 8, 1 (Apr. 2022), 13. * [2] Ahmed, T., Hoyle, R., Connelly, K., Crandall, D., and Kapadia, A. Privacy Concerns and Behaviors of People with Visual Impairments. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (Seoul Republic of Korea, Apr. 2015), ACM, pp. 3523–3532. * [3] Akter, T., Dosono, B., Ahmed, T., Kapadia, A., and Semaan, B. ”i am uncomfortable sharing what i can’t see”: Privacy concerns of the visually impaired with camera based assistive applications. In 29th USENIX Security Symposium (USENIX Security 20) (online, Aug. 2020), USENIX Association, pp. 1929–1948. * [4] Bermejo Fernandez, C., Chatzopoulos, D., Papadopoulos, D., and Hui, P. This website uses nudging: Mturk workers’ behaviour on cookie consent notices. Proceedings of the ACM on Human-Computer Interaction 5, CSCW2 (2021), 1–22. * [5] Brulé, E., Tomlinson, B. J., Metatla, O., Jouffrais, C., and Serrano, M. Review of quantitative empirical evaluations of technology for people with visual impairments. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (New York, NY, USA, 2020), CHI ’20, Association for Computing Machinery, pp. 1–14. * [6] Chandrashekar, S. Is hearing believing? Perception of online information credibility by screen reader users who are blind or visually impaired. PhD thesis, University of Toronto Toronto, ON, 2010. * [7] Coopamootoo, K. P., and Groß, T. Why privacy is all but forgotten. Proceedings on Privacy Enhancing Technologies 2017, 4 (2017), 97–118. * [8] Coopamootoo, K. P., Mehrnezhad, M., and Toreini, E. ”i feel invaded, annoyed, anxious and i may protect myself”: Individuals’ feelings about online tracking and their protective behaviour across gender and country. In 31st USENIX Security Symposium (USENIX Security 22) (Boston, MA, Aug. 2022), USENIX Association, pp. 287–304. * [9] Degeling, M., Utz, C., Lentzsch, C., Hosseini, H., Schaub, F., and Holz, T. We value your privacy… now take some cookies. Informatik Spektrum 42, 5 (2019), 345–346. * [10] Developers, C. Lighthouse accessibility scoring. https://developer.chrome.com/docs/lighthouse/ accessibility/scoring/, 2019\. * [11] Egelman, S., and Peer, E. The Myth of the Average User: Improving Privacy and Security Systems through Individualization. In Proceedings of the 2015 New Security Paradigms Workshop (Twente Netherlands, Sept. 2015), ACM, pp. 16–28. * [12] Englehardt, S., and Narayanan, A. Online Tracking: A 1-million-site Measurement and Analysis. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security (Vienna Austria, Oct. 2016), ACM, pp. 1388–1401. * [13] European Data Protection Board. Guidelines on consent under regulation 2016/679. https://ec.europa.eu/newsroom/article29/items/623051, adopted on 10 april 2018. * [14] Frauenberger, C. Disability and Technology: A Critical Realist Perspective. In Proceedings of the 17th International ACM SIGACCESS Conference on Computers & Accessibility - ASSETS ’15 (Lisbon, Portugal, 2015), ACM Press, pp. 89–96. * [15] Garante per la protezione dei dati personali. Video Surveillance ’ Decision. https://garanteprivacy.it:443/web/guest/home/docweb/-/docweb-display/docweb/1734653, Apr. 2010. * [16] Gerl, A., and Meier, B. The layered privacy language art. 12–14 gdpr extension–privacy enhancing user interfaces. Datenschutz und Datensicherheit-DuD 43, 12 (2019), 747–752. * [17] Google Developers. Lighthouse Overview. https://developer.chrome.com/docs/lighthouse/overview/, May 2022. * [18] GOV.UK. Results of the 2016 GOV.UK assistive technology survey - Accessibility in government. https://accessibility.blog.gov.uk/2016/11/01/results-of-the-2016-gov-uk-assistive-technology-survey/, Nov. 2016. * [19] Griffin-Shirley, N., Banda, D. R., Ajuwon, P. M., Cheon, J., Lee, J., Park, H. R., and Lyngdoh, S. N. A survey on the use of mobile applications for people who are visually impaired. Journal of Visual Impairment & Blindness 111, 4 (2017), 307–323. * [20] Hakobyan, L., Lumsden, J., O’Sullivan, D., and Bartlett, H. Mobile assistive technologies for the visually impaired. Survey of ophthalmology 58, 6 (2013), 513–528. * [21] Hanson, V. L., Cavender, A., and Trewin, S. Writing about accessibility. Interactions 22, 6 (Oct. 2015), 62–65. * [22] Hanson, V. L., and Richards, J. T. Progress on website accessibility? ACM Transactions on the Web (TWEB) 7, 1 (2013), 1–30. * [23] Hayes, J., Kaushik, S., Price, C. E., and Wang, Y. Cooperative privacy and security: Learning from people with visual impairments and their allies. In Fifteenth Symposium on Usable Privacy and Security (SOUPS 2019) (Santa Clara, CA, Aug. 2019), USENIX Association, pp. 1–20. * [24] Hu, X., and Sastry, N. Characterising Third Party Cookie Usage in the EU after GDPR. In Proceedings of the 10th ACM Conference on Web Science (Boston Massachusetts USA, June 2019), ACM, pp. 137–141. * [25] Inan, F. A., Namin, A. S., Pogrund, R. L., and Jones, K. S. Internet use and cybersecurity concerns of individuals with visual impairments. Journal of Educational Technology & Society 19, 1 (2016), 28–40. * [26] Inc., D. S. Axe-core. https://github.com/dequelabs/axe-core, Nov. 2022. * [27] Information Commissioner’s Office. Guidance on the use of cookies and similar technologies. https://ico.org.uk/for-organisations/guide-to-pecr/guidance-on-the-use-of-cookies-and-similar-technologies/, Feb. 2022. * [28] Kaczmirek, L., and Wolff, K. G. Survey Design for Visually Impaired and Blind People. In Universal Acess in Human Computer Interaction. Coping with Diversity, C. Stephanidis, Ed., vol. 4554. Springer Berlin Heidelberg, Berlin, Heidelberg, 2007, pp. 374–381. * [29] King, A., Evans, G., and Blenkhorn, P. Blind people and the world wide web. https://www.webbie.org.uk/webbie.html, 2004. * [30] Kretschmer, M., Pennekamp, J., and Wehrle, K. Cookie Banners and Privacy Policies: Measuring the Impact of the GDPR on the Web. ACM Transactions on the Web 15, 4 (July 2021), 1–42. * [31] Ling, J., and Van Schaik, P. The effect of text and background colour on visual search of web pages. Displays 23, 5 (2002), 223–230. * [32] Lund, B. The Brave browser: A monetary opportunity for libraries in the cryptoverse. Library Hi Tech News 38, 6 (Jan. 2021), 15–16. * [33] Matte, C., Bielova, N., and Santos, C. Do cookie banners respect my choice? : Measuring legal compliance of banners from iab europe’s transparency and consent framework. In 2020 IEEE Symposium on Security and Privacy (SP) (San Francisco, CA, USA, 2020), 2020 IEEE Symposium on Security and Privacy (SP), pp. 791–809. * [34] Mehrnezhad, M. A Cross-Platform Evaluation of Privacy Notices and Tracking Practices. In 2020 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW) (Genoa, Italy, Sept. 2020), IEEE, pp. 97–106. * [35] Mehrnezhad, M., and Almeida, T. Caring for Intimate Data in Fertility Technologies. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama Japan, May 2021), ACM, pp. 1–11. * [36] Mehrnezhad, M., Coopamootoo, K., and Toreini, E. How Can and Would People Protect From Online Tracking? Proceedings on Privacy Enhancing Technologies 2022, 1 (Jan. 2022), 105–125. * [37] Melicher, W., Sharif, M., Tan, J., Bauer, L., Christodorescu, M., and Leon, P. G. Preferences for Web Tracking. Proceedings on Privacy Enhancing Technologies 2016, 2 (2016), 1–20. * [38] Microsoft. Lifecycle FAQ - Internet Explorer and Microsoft Edge. https://docs.microsoft.com/en-us/lifecycle/faq/internet-explorer-microsoft-edge, n.d. * [39] Napoli, D., Baig, K., Maqsood, S., and Chiasson, S. ” i’m literally just hoping this will $\\{$Work:’$\\}$’obstacles blocking the online security and privacy of users with visual disabilities. In Seventeenth Symposium on Usable Privacy and Security (SOUPS 2021) (2021), pp. 263–280. * [40] Nouwens, M., Liccardi, I., Veale, M., Karger, D., and Kagal, L. Dark Patterns after the GDPR: Scraping Consent Pop-ups and Demonstrating their Influence. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (Honolulu HI USA, Apr. 2020), ACM, pp. 1–13. * [41] Office for National Statistics. Internet users, UK. https://www.ons.gov.uk/businessindustryandtrade /itandinternetindustry/bulletins/internetusers/2020, 2020. * [42] Parliament, T. E., and the Council of the European Union. Directive 2002/58/EC of the European Parliament and of the Council of 12 July 2002 concerning the processing of personal data and the protection of privacy in the electronic communications sector (Directive on privacy and electronic communications), July 2002. * [43] Pernice, K., Nielsen, J., Farrell, S., Mizobuchi, S., Ishida, N., Stover, U. A., Yohay, M., Franko, E., and Richardson, A. Usability guidelines for accessible web design. Evidence–Based User Experience Research, Training, Consulting (2001). * [44] Pierce, J., Fox, S., Merrill, N., and Wong, R. Differential vulnerabilities and a diversity of tactics: What toolkits teach us about cybersecurity. Proceedings of the ACM on Human-Computer Interaction 2, CSCW (2018), 1–24. * [45] PowerMapper Software. Screen reader reliability. https://www.powermapper.com/tests/screen-readers/, Sept. 2022. * [46] Pugliese, G., Riess, C., Gassmann, F., and Benenson, Z. Long-Term Observation on Browser Fingerprinting: Users’ Trackability and Perspective. Proceedings on Privacy Enhancing Technologies 2020, 2 (Apr. 2020), 558–577. * [47] Rao, A., Schaub, F., and Sadeh, N. What do they know about me? Contents and Concerns of Online Behavioral Profiles, June 2015. * [48] Regulation, G. D. P. General Data Protection Regulation (GDPR) – Official Legal Text. https://gdpr-info.eu/, 2016. * [49] Rotard, M., Taras, C., and Ertl, T. Tactile web browsing for blind people. Multimedia Tools and Applications 37, 1 (Mar. 2008), 53–69. * [50] Rubin, G. S., and Legge, G. E. Psychophysics of reading. VI—The role of contrast in low vision. Vision Research 29, 1 (Jan. 1989), 79–91. * [51] Schmeelk, S., and Petrie, H. Digital authentication for visually disabled people: Initial results of an online survey. In Computers Helping People with Special Needs: 18th International Conference, ICCHP-AAATE 2022, Lecco, Italy, July 11–15, 2022, Proceedings, Part II (2022), Springer, pp. 41–50. * [52] Schnell, K., and Roy, K. Website privacy notification for the visually impaired. In 2021 IEEE Symposium Series on Computational Intelligence (SSCI) (Orlando, FL, USA, 2021), 2021 IEEE Symposium Series on Computational Intelligence (SSCI), pp. 1–6. * [53] Seničar, V., Jerman-Blažič, B., and Klobučar, T. Privacy-Enhancing Technologies—approaches and development. Computer Standards & Interfaces 25, 2 (May 2003), 147–158. * [54] Shirazi, F., and Volkamer, M. What deters jane from preventing identification and tracking on the web? In Proceedings of the 13th Workshop on Privacy in the Electronic Society (New York, NY, USA, 2014), WPES ’14, Association for Computing Machinery, pp. 107–116. * [55] Sørensen, J., and Kosta, S. Before and After GDPR: The Changes in Third Party Presence at Public and Private European Websites. In The World Wide Web Conference (New York, NY, USA, May 2019), WWW ’19, Association for Computing Machinery, pp. 1590–1600. * [56] Southwell, K. L., and Slater, J. An evaluation of finding aid accessibility for screen readers. Information Technology and Libraries 32, 3 (2013), 34–46. * [57] Statcounter. Browser Market Share Worldwide. https://gs.statcounter.com/browser-market-share, n.d. * [58] the Brave Privacy Team. Blocking annoying and privacy-harming cookie consent banners. https://brave.com/privacy-updates/21-blocking-cookie-notices/, Sept. 2022\. * [59] Trevisan, M., Traverso, S., Bassi, E., and Mellia, M. 4 Years of EU Cookie Law: Results and Lessons Learned. Proceedings on Privacy Enhancing Technologies 2019, 2 (Apr. 2019), 126–145. * [60] Ur, B., Leon, P. G., Cranor, L. F., Shay, R., and Wang, Y. Smart, useful, scary, creepy: Perceptions of online behavioral advertising. In Proceedings of the Eighth Symposium on Usable Privacy and Security - SOUPS ’12 (Washington, D.C., 2012), ACM Press, p. 1. * [61] Utz, C., Degeling, M., Fahl, S., Schaub, F., and Holz, T. (Un)informed Consent: Studying GDPR Consent Notices in the Field. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security (London United Kingdom, Nov. 2019), ACM, pp. 973–990. * [62] W3. How to Meet WCAG (Quickref Reference). https://www.w3.org/WAI/WCAG21/quickref/, Oct. 2019. * [63] W3. WCAG 2 Overview. https://www.w3.org/WAI/standards-guidelines/wcag/, Mar. 2022. * [64] W3. Understanding Conformance. https://www.w3.org/WAI/WCAG21/Understanding/ conformance#levels, n.d. * [65] Wang, Y., and Price, C. E. Accessible privacy. In Modern Socio-Technical Perspectives on Privacy. Springer, Cham, 2022, pp. 293–313. * [66] Washington, B. S., Feng, J. H., Ahram, T., and Falcão, C. Proper Implementation of Website Features Affecting the Use of Screen Readers, vol. 972 of Advances in Usability and User Experience. Springer International Publishing, 2020. * [67] WebAIM. Screen Readers and CSS: Are We Going Out of Style (and into Content)? https://webaim.org/blog/screen-readers-and-css/, Aug. 2017. * [68] WebAIM. WebAIM: Screen Reader User Survey #9 Results. https://webaim.org/projects/screenreadersurvey9/, June 2021. * [69] WebAIM. Wave chrome, firefox, and edge extensions. https://wave.webaim.org/extension/, n.d. * [70] whotracksme. GDPR - what happened? https://whotracks.me/blog/gdpr-what-happened.html, 2018. * [71] Yu, W., Kuber, R., Murphy, E., Strain, P., and McAllister, G. A novel multimodal interface for improving visually impaired people’s web accessibility. Virtual Reality 9, 2 (2006), 133–148. ## Appendix A Website Analysis Template For each website in our list, we followed these steps for our analysis: * • Part I: Cookie Notice (baseline) * - Step 1: Open Google Chrome incognito window and visit the homepage of the website. * - Step 2: Observe if there is a notice (cookie consent, privacy settings, banner, etc.). * - No: Write it in the file. * - Yes: Observe the location and user control options e.g. OK, Accept, Yes, Reject, No, More Options, Settings, Links to privacy-related pages, etc. Write your observations in the file. * - Step 3: Close the Google Chrome incognito window. * - Step 4: Open Brave private window and visit the homepage of the website. * - Step 5: Repeat Step - ‣ • ‣ A (cookie notice). * - Step 6: Record the number of items blocked by the Brave Shields in the file. * - Step 7: Close the Brave private window. * • Part II: Automated Accessibility Testing Tools * - Step 8: Open a new Google Chrome incognito window and visit the website’s homepage. * - Step 9: Click on WAVE extension to run the test. * - Step 10: Record the number of each of the categories in the file. * - Step 11: Close the Google Chrome incognito window. * - Step 12: Open a new Google Chrome incognito window and visit the website’s homepage. * - Step 13: Open developer tools and navigate to the Lighthouse tab. * - Step 14: Select Navigation Mode, A desktop device and the Accessibility Categories. * - Step 15: Analyse the page. * - Step 16: Record the overall accessibility score and the number of each item shown. * - Step 17: Close the Google Chrome incognito window. * • Part III: Manual Testing Via Text-only Browser * - Step 18: Open WebbIE Browser and visit the homepage of the website. * - Step 19: Record the number of headings for the website overall. * - Step 20: Record the presence of a notice, and if so the presence of headings. * - Step 21: Close WebbIE. * • Part IV: Manual Testing via Screen Readers * - Step 22 : Open NVDA screen reader. * - Step 23: Open a new Google Chrome incognito window and visit the website’s homepage. * - Step 24: Allow screen reader to read website. * - Step 25: Interact with the website using keyboard. * - Step 26: Record pass/fail for categories in 3.2.3. * - Step 27: Close the Google Chrome incognito window and screen reader. * - Step 28: Open JAWS screen reader. * - Step 29: Repeat Steps - ‣ • ‣ A-- ‣ • ‣ A. Note that we performed two rounds of testing (with identical results). We uninstalled and reinstalled WebbIE for each round since it does not support cookie management. ## Appendix B User Study Template ### B.1 Screening validation * • Do you live in the United Kingdom? (One answer is possible) * $\square$ Yes * $\square$ No * • Do you use assistive technology? (One answer is possible) * $\square$ Yes * $\square$ No ### B.2 Section 1: Internet and assistive technology This section is about your internet usage and any assistive technology you may utilise while browsing the web. * • 1.1: How do you describe your visual impairment? (Text input is possible) * • 1.2: Which devices do you use to access the internet? Please specify if other (Several answers are possible) * $\square$ Personal Computer (Desktop or Laptop) Mobile Phone * $\square$ Tablet Computer * $\square$ Smart home devices * $\square$ Wearable devices Internet-enabled TVs Gaming consoles Public computers * $\square$ Other: * • 1.3: Which device are you using to complete this questionnaire? Please specify if other (One answer possible) * $\square$ Personal Computer (Desktop or Laptop) * $\square$ Mobile Phone * $\square$ Tablet Computer * $\square$ Smart home devices * $\square$ Wearable devices * $\square$ Internet-enabled TVs * $\square$ Gaming consoles * $\square$ Public computers * $\square$ Other: * • 1.4: In an average week, how many hours do you use the internet? (One answer is possible) * $\square$ Less then 1 hour 1-5 hours * $\square$ 6-10 hours * $\square$ 11-15 hours * $\square$ 16-20 hours * $\square$ 21-15 hours * $\square$ 26-30 hours * $\square$ More than 30 hours * • 1.5: What assistive technology do you use when browsing the web? Please specify if other. (Several answers are possible) * $\square$ Screen reader (Jaws, NVDA, Voice over or other) * $\square$ Braille Display * $\square$ Text Only Browser (WebbIE or other) * $\square$ Magnification software * $\square$ Assistive browser extension * $\square$ Alternative input devices * $\square$ None * • 1.6: Which of the following screen readers do you use? Please specify if other. (Several answers are possible) * $\square$ JAWS * $\square$ NVDA * $\square$ VoiceOver * $\square$ Natural Reader * $\square$ Read&Write * $\square$ Narrator * $\square$ Talkback * $\square$ ChromeVox * $\square$ Orca * $\square$ I don’t use a screen reader * $\square$ Other: * • 1.7: How would you describe your level of expertise with a screen reader? (Text input is possible) * • 1.8: Do you use plug-ins with a screen reader? (One answer is possible) A plugin or add-on adds a specific feature or functionality to a screen reader. Allowing users to customise their software and add the features they need. * $\square$ Yes * $\square$ No * • 1.9: Which plugins do you use? (Text input is possible) ### B.3 Section 2: Privacy enhancing technology usage This section is about the measures you take to protect your privacy and security while browsing the internet and Privacy Enhancing Technologies (PETs). PETs are tools that can help protect your privacy online by limiting the collection, use, and dissemination of your personal information. * • 2.1: Which of the following privacy enhancing technologies do you use? Please specify if other. (Several answers are possible) * $\square$ Browsers do not track * $\square$ Virtual Private Network (VPN) * $\square$ Private browsing * $\square$ Password manager * $\square$ Manual cookie opt-out * $\square$ Privacy-focused web browsers * $\square$ Encrypted messaging apps * $\square$ Ad blockers * $\square$ File encryption tools * $\square$ None * $\square$ Other: * • 2.2: How did you learn about privacy enhancing technology? Please specify if other. (Several answers are possible) * $\square$ Friend / social contact recommendation * $\square$ Work / school recommendation * $\square$ Privacy/cookie policy of a website * $\square$ News * $\square$ I don’t know * $\square$ I don’t use privacy enhancing technologies * $\square$ Other: ### B.4 Section 3: Cookie notice Cookie notices appear when a user lands on a website and informs them that the website is using cookies (a data file) and other trackers that process personal data, and that the user must make a choice whether they want their personal data processes. * • 3.1: In your own words what is a cookie notice and what are they supposed to do? (Text input is possible) * • 3.2: How do you feel about cookie notices? (Text input is possible) * • 3.3: How do you interact with cookie notices? (Text input is possible) * • 3.4: Have you encountered any issues with cookie notices? (One answer is possible) * $\square$ Yes * $\square$ No * • 3.5: Please describe in your won words what type of issues have you experienced with cookie notices? (Text input is possible) * • 3.6: Which of the following issues have you experienced? Please specify if other (Several answers are possible) * $\square$ Cookie notice not being present via my assistive technology * $\square$ Unable to find cookie notice * $\square$ Unable to answer cookie notice * $\square$ Low contrast cookie notice * $\square$ Unclear response options in cookie notice * $\square$ Lack of headings for cookie notice * $\square$ Unable to leave cookie notice * $\square$ Unable to enter cookie notice * $\square$ Other: * • 3.7: How would you wish to handle cookie notices? Please specify if other. (One answer is possible) * $\square$ Agree * $\square$ Disagree * $\square$ Editing the settings * $\square$ Ignore cookie notice * $\square$ Other: * • 3.8: How do you actually respond to cookie notices? Please specify if other. (One answer is possible) * $\square$ Agree * $\square$ Disagree * $\square$ Editing the settings * $\square$ Ignore cookie notice * $\square$ Unable to respond to cookie notice * $\square$ Other: * • 3.9: If you are unable to respond to cookie notices, what do you think is the reason? (Text input is possible) ### B.5 Section 4: Suggestions * • 4.1: Who do you think should be responsible for ensuring the secure accessibility of the Internet? Please specify if other. (Several answers are possible) * $\square$ Website Developers * $\square$ Designers of Accessibility Evaluation Tools * $\square$ Policymakers * $\square$ Users of the internet * $\square$ Designers of Assistive Technologies * $\square$ Other: * • 4.2: Which of these recommendation do you think would help improve your experience online? Please specify if other (Several answers are possible) * $\square$ Website designers should design websites with accessibility in mind * $\square$ Website designers should include more headings for useful information * $\square$ Website designers should complete more accessibility testing * $\square$ Evaluation tools should allow for testing of sections of a web page * $\square$ Specific testing tools for parts of a webpage (i.e. cookie notices or other elements) * $\square$ The inclusion of laws specifying privacy and accessibility * $\square$ Specifications of achieving privacy and accessibility * $\square$ End users can searching for cookie notices * $\square$ End users can specific privacy protecting tools such as the Brave internet browser * $\square$ Other: ### B.6 Section 5: Demographic and background questions * • 5.1: What is your age? (One answer is possible) * $\square$ 18 to 24 * $\square$ 25 to 34 * $\square$ 35 to 44 * $\square$ 45 to 54 * $\square$ 55 to 64 * $\square$ 65 or over * $\square$ Prefer not to say * • 5.2: What is your gender? (One answer is possible) * $\square$ Female * $\square$ Male * $\square$ Non-binary * $\square$ Prefer not to say * $\square$ Other * • 5.3: What is your highest level of education? (One answer is possible) * $\square$ Secondary education * $\square$ Post-secondary education * $\square$ Undergraduate education * $\square$ Graduate education * $\square$ Prefer not to say * • 5.4: What is your occupation? Please specify if other. (One answer is possible) * $\square$ Healthcare and social assistance * $\square$ Education and training * $\square$ Sales and retail * $\square$ Administrative and support * $\square$ Manufacturing and production * $\square$ Information technology * $\square$ Business and finance * $\square$ Transportation and logistics * $\square$ Construction and trades * $\square$ Arts, entertainment, and media * $\square$ Prefer not to say * $\square$ Other * • 5.5: What services do you use online? Please specify if other. (Multiple answers are possible) * $\square$ Email * $\square$ Social media * $\square$ Online shopping and e-commerce * $\square$ Video Streaming * $\square$ Music/audio streaming * $\square$ Online payment/banking * $\square$ File sharing and cloud storage * $\square$ Online travel booking * $\square$ Online education and e-learning * $\square$ Online commutation and collaboration services * $\square$ Other ## Appendix C Detailed results of automated accessibility tools Table 8: Detailed results of WAVE 5.0 and Google Lighthouse Website | Privacy Policy | Errors | Contrast Errors | Alerts | Features | Structural Elements | ARIA | Lighthouse Score ---|---|---|---|---|---|---|---|--- google.com | ✓ | 1 | 10 | 4 | 5 | 7 | 350 | 97 youtube.com | ✓ | 26 | 1 | 70 | 92 | 66 | 676 | 89 yahoo.com | ✓ | 2 | 8 | 6 | 2 | 4 | 4 | 86 facebook.com | ✓ | 4 | 46 | 12 | 1 | 9 | 9 | 93 bbc.co.uk | ✓ | 0 | 0 | 134 | 23 | 119 | 371 | 100 amazon.co.uk | ✓ | 6 | 2 | 139 | 194 | 55 | 381 | 95 reddit.com | ✓ | 9 | 139 | 657 | 104 | 36 | 433 | 77 wikipedia.org | ✗ | 3 | 0 | 97 | 123 | 70 | 24 | 100 live.com | ✗ | 0 | 5 | 8 | 11 | 38 | 77 | 98 instagram.com | ✓ | 0 | 25 | 3 | 10 | 10 | 37 | 95 twitter.com | ✓ | 1 | 19 | 3 | 1 | 4 | 55 | 88 ebay.co.uk | ✗ | 0 | 0 | 0 | 0 | 0 | 0 | 93 dailymail.co.uk | ✓ | 75 | 31 | 890 | 628 | 497 | 61 | 74 bing.com | ✓ | 12 | 23 | 38 | 26 | 40 | 244 | 94 gov.uk | ✓ | 0 | 1 | 5 | 16 | 68 | 38 | 100 netflix.com | ✓ | 2 | 22 | 9 | 22 | 36 | 68 | 88 theguardian.com | ✗ | 17 | 59 | 131 | 102 | 290 | 624 | 79 pornhub.com | ✗ | 65 | 1 | 197 | 110 | 67 | 119 | 96 office.com | ✓ | 4 | 0 | 3 | 25 | 59 | 199 | 94 fandom.com | ✓ | 12 | 33 | 14 | 28 | 28 | 2 | 86 xvideos.com | ✓ | 104 | 64 | 158 | 2 | 22 | 1 | 88 paypal.com | ✓ | 12 | 33 | 14 | 28 | 28 | 2 | 100 microsoft.com | ✓ | 2 | 0 | 3 | 24 | 47 | 183 | 100 linkedin.com | ✓ | 27 | 0 | 0 | 9 | 32 | 174 | 100 xhamster.com | ✗ | 8 | 118 | 57 | 58 | 25 | 2 | 81 imdb.com | ✗ | 8 | 0 | 42 | 98 | 37 | 1552 | 88 duckduckgo.com | ✗ | 23 | 5 | 13 | 6 | 38 | 35 | 96 amazon.com | ✗ | 4 | 2 | 178 | 216 | 45 | 317 | 98 zoom.us | ✓ | 7 | 3 | 58 | 64 | 93 | 889 | 80 twitch.tv | ✓ | 71 | 0 | 150 | 130 | 48 | 300 | 86 amazonaws.com | ✓ | 4 | 14 | 140 | 228 | 377 | 87 | 86 tiktok.com | ✓ | 52 | 14 | 54 | 9 | 59 | 0 | 63 whatsapp.com | ✓ | 5 | 14 | 10 | 4 | 142 | 5 | 85 doubleclick.net | ✓ | 2 | 4 | 51 | 11 | 78 | 141 | 100 spankbang.com | ✗ | 16 | 160 | 500 | 190 | 36 | 0 | 72 sky.com | ✓ | 3 | 16 | 28 | 22 | 41 | 105 | 90 apple.com | ✗ | 11 | 0 | 15 | 47 | 62 | 217 | 92 rightmove.co.uk | ✓ | 3 | 7 | 68 | 12 | 48 | 55 | 87 booking.com | ✓ | 30 | 3 | 37 | 40 | 85 | 574 | 92 etsy.com | ✓ | 7 | 1 | 31 | 42 | 175 | 1682 | 73 indeed.com | ✓ | 2 | 0 | 8 | 20 | 25 | 134 | 90 msn.com | ✓ | 26 | 0 | 213 | 359 | 311 | 44 | 79 github.com | ✗ | 2 | 35 | 18 | 114 | 74 | 179 | 86 adobe.com | ✓ | 2 | 1 | 12 | 111 | 120 | 139 | 96 chaturbate.com | ✗ | 19 | 113 | 494 | 102 | 210 | 5 | 84 xnxx.com | ✓ | 165 | 0 | 808 | 3 | 20 | 0 | 97 ## Appendix D Detailed results of screen reader tests Table 9: NVDA and JAWS results | NVDA | JAWS ---|---|--- Website | Readable | Immediately | Keyboard Navigable | Link or button Purpose | Abbreviations are explained | Page Titled | Cookie Notice Titled | Headings useful for navigation | No Timing | Readable | Immediately | Keyboard Navigable | Link or button Purpose | Abbreviations are explained | Page Titled | Cookie Notice Titled | Headings useful for navigation | No Timing google.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ youtube.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✗ | ✗ | ✓ yahoo.com | ✓ | ✓ | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ facebook.com | ✓ | ✗ | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ bbc.co.uk | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ amazon.co.uk | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ reddit.com | ✗ | ✗ | ✗ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | - | ✓ | ✗ | ✗ | ✓ wikipedia.org | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - live.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - instagram.com | ✗ | ✗ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | - | ✓ | ✓ | ✓ | ✓ twitter.com | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ ebay.co.uk | ✓ | ✗ | ✓ | ✗ | ✗ | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✓ | ✗ | ✗ | ✓ dailymail.co.uk | ✗ | ✗ | ✗ | ✗ | ✗ | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✓ | ✗ | ✗ | ✓ bing.com | ✓ | ✗ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ gov.uk | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ netflix.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ theguardian.com | ✓ | ✗ | ✓ | ✗ | ✗ | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✗ | ✗ | ✓ | ✗ | ✗ | ✓ pornhub.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - office.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ fandom.com | ✓ | ✗ | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ xvideos.com | ✗ | ✗ | ✗ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ paypal.com | ✓ | ✗ | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ microsoft.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ linkedin.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ xhamster.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - imdb.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - duckduckgo.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - amazon.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - zoom.us | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✗ | ✗ | ✓ twitch.tv | ✓ | ✗ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ amazonaws.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ tiktok.com | ✓ | ✗ | ✗ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✗ | ✗ | - | ✓ | ✓ | ✗ | ✓ whatsapp.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ doubleclick.net | ✓ | ✗ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✗ | ✗ | - | ✓ | ✗ | ✗ | ✓ spankbang.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - sky.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ apple.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - rightmove.co.uk | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ booking.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ etsy.com | ✓ | ✓ | ✓ | - | - | ✓ | - | - | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | - | ✗ | ✓ indeed.com | ✓ | ✓ | ✓ | ✓ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✓ | - | ✓ | ✗ | ✗ | ✓ msn.com | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✓ | ✗ | ✓ github.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - adobe.com | ✗ | ✗ | ✗ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✗ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ chaturbate.com | - | - | - | - | - | ✓ | - | - | - | - | - | - | - | - | ✓ | - | - | - xnxx.com | ✗ | ✗ | ✗ | ✗ | - | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ | ✓ | ✗ | - | ✓ | ✗ | ✗ | ✓ ### D.1 Answers to demographic questions Table 10: Answers to demographic questions 1.2 Devices used | N | 1.3 To Answer | N | 1.4 Hours a Week | N | 1.5 AT | N ---|---|---|---|---|---|---|--- Personal Computer | 94 | Personal Computer | 71 | More than 30 hours | 28 | Magnification software | 50 Mobile Phone | 94 | Mobile Phone | 22 | 26-30 hours | 23 | Screen reader | 42 Tablet Computer | 47 | Tablet Computer | 5 | 6-10 hours | 14 | Assistive browser extension | 22 Gaming consoles | 36 | Other | 2 | 16-20 hours | 12 | Other | 14 Internet-enabled TV | 34 | | | 21-15 hours | 10 | None | 9 Smart home devices | 33 | | | 11-15 hours | 8 | Alternative input devices | 7 Wearable devices | 16 | | | 1-5 hours | 5 | Braille Display | 2 Public computers | 9 | | | | | Text Only Browser | 1 1.6 Screeen reader | N | 5.1 Age | N | 5.2 Gender | N | 5.3 Education | N No screen reader | 40 | 25 to 34 | 33 | Male | 59 | Undergraduate | 37 VoiceOver | 18 | 45 to 54 | 20 | Female | 38 | Graduate | 26 Narrator | 12 | 18 to 24 | 18 | Non-binary | 2 | Post-secondary | 19 ChromeVox | 11 | 35 to 44 | 14 | Prefer not to say | 1 | Secondary | 15 JAWS | 10 | 55 to 64 | 12 | | | Prefer not to say | 3 Natural Reader | 10 | 65 or over | 2 | | | | Read&Write | 9 | Prefer not to say | 1 | | | | Talkback | 6 | | | | | | NVDA | 5 | | | | | | Other | 5 | | | | | | Orca | 1 | | | | | | 5.5 Online Services | N | 1.7 Experience | N | | | | Email | 98 | Basic | 33 | | | | Shopping and e-commerce | 93 | Moderate | 29 | | | | Social media | 90 | None | 20 | | | | Payment/banking | 89 | Expert | 9 | | | | Video Streaming | 80 | Previous experience | 5 | | | | Music/audio streaming | 77 | Other | 4 | | | | Travel booking | 68 | | | | | | File sharing and cloud storage | 64 | | | | | | Education and e-learning | 62 | | | | | | Commutation and collaboration | 35 | | | | | | ## Appendix E Examples of websites, cookie notices and their outcome via screen readers --- (Enter webpage URL and press Enter Key) NVDA: | dailymail.co.uk selected NVDA: | UK home daily mail online NVDA: | Link, graphic online news, sport, celebrity, science and health stories NVDA: | List with 14 items NVDA: | Link home NVDA: | Link, news NVDA: | Link u.s NVDA: | [Advertisement read], [Same advertisement read] NVDA: | [Advertisement read], [Same advertisement read] NVDA: | [Advertisement read], [Same advertisement read] NVDA: | Discover the best black Friday deals, discover the best black Friday deals (Down arrow key pressed) NVDA: | Link home (Proceeds to read the navigation bar) (Starts to read news items on homepage) Figure 6: Top: Graphical representation of dailymail.co.uk with the highlighted cookie notice at the bottom of the page. Bottom: Example transcript via a screen reader. Figure 7: Visual representation of reddit.com cookie notice. None of the two screen readers could audibly output the cookie notice. They only read the body of the web page. --- (Enter webpage URL and press Enter Key) JAWS: | A simple and safer way to pay and get paid, vertical bar, PayPal UK JAWS: | Page has three regions, 8 headings and 33 links JAWS: | A simple and safer way to pay and get paid, vertical bar, PayPal UK JAWS: | Link PayPal JAWS: | Navigation region, list of four items (Reads aloud navigation bar) (Navigate through body of site) JAWS: | If you accept cookies, we’ll use them to improve and customise your experience and enable our partners to show you personalised PayPal ads when you visit other sites. JAWS: | Link, manage cookies and learn more JAWS: | Accept button JAWS: | Decline button Figure 8: Top: Graphical representation of paypal.com with highlighted cookie notice at the bottom of the page. Bottom: Example transcript via a screen reader. --- (Enter webpage URL and press Enter Key) NVDA: | BBC.co.uk selected NVDA: | BBC Home NVDA: | Banner landmark, Let us know you agree to cookies, Heading level two NVDA: | Clickable banner landmark, We use, link, cookies to give you the best online experience. Please let us know if you agree to all of these cookies NVDA: | Button, Yes, I agree NVDA: | Link, No, take me to settings NVDA: | BBC Navigation landmark, BBC Homepage (Proceeds to read the rest of the home page) Figure 9: Top: Graphical representation of bbc.co.uk with highlighted cookie notice at the top of the page. Bottom: Example transcript via a screen reader.
# A Geometric flow towards hamiltonian stationary submanifolds Jingyi Chen and Micah Warren Department of Mathematics The University of British Columbia, Vancouver, BC V6T 1Z2<EMAIL_ADDRESS>Department of Mathematics University of Oregon, Eugene, OR 97403<EMAIL_ADDRESS> ###### Abstract. In this paper, we introduce a geometric flow for Lagrangian submanifolds in a Kähler manifold that stays in its initial Hamiltonian isotopy class and is a gradient flow for volume. The stationary solutions are the Hamiltonian stationary Lagrangian submanifolds. The flow is not strictly parabolic but it corresponds to a fourth order strictly parabolic scalar equation in the cotangent bundle of the submanifold via Weinstein’s Lagrangian neighborhood theorem. For any compact initial Lagrangian immersion, we establish short-time existence, uniqueness, and higher order estimates when the second fundamental forms are uniformly bounded up to time $T$. Chen is partially supported by NSERC Discovery Grant GR010074 ## 1\. Introduction The objective of this paper is to introduce a fourth order flow of Lagrangian submanifolds in a Kähler manifold as a gradient flow of volume within a Hamiltonian isotopy class and establish basic properties such as short-time existence, uniqueness, and extendibility with bounded second fundamental form. Our setting includes a Kähler manifold $(M^{2n},h,\omega,J)$ with symplectic form $\omega$ and compatible Kähler metric $h$ satisfying $h(JV,W)=\omega(V,W)$ where $J$ is a complex structure, and a given compact Lagrangian immersion $\iota:L^{n}\rightarrow M^{2n}.$ We propose to find $F:L\times[0,T)\rightarrow M^{2n}$ satisfying (1.1) $\displaystyle\frac{dF}{dt}=J\nabla\operatorname{div}\left(JH\right)$ (1.2) $\displaystyle F\left(\cdot,0\right)=\iota\left(\cdot\right)$ where $H$ is the mean curvature vector of $L_{t}=F(\cdot,t)$ in $M$ and $\nabla,\operatorname{div}$ are along $L_{t}$ in the induced metric from $h$. The stationary solutions of the above evolution equation are the so-called Hamiltonian stationary submanifolds, which are fourth order generalizations of special Lagrangians, and exist in more abundance. Within a Hamiltonian isotopy class it is possible for a compact Lagrangian submanifold of $\mathbb{C}^{n}$ to minimize volume (for example, a Clifford torus). Meanwhile, compact special Lagrangian submanifolds of $\mathbb{C}^{n}$ do not exist. Recall that two manifolds are Hamiltonian isotopic if they can be joined by the flow generated by a vector field of the form $J\nabla f$ for a function $f$. We will show that the initial value problem (1.1) - (1.2): 1. (1) stays within the Hamiltonian isotopy class - that is, the flow is generated by a vector field of the form $J\nabla f,$ 2. (2) is the gradient flow of volume with respect to an appropriate metric, so decreases volume along the flow, 3. (3) enjoys short time existence given smooth initial conditions, and 4. (4) continues to exist as long as a second fundamental form bound is satisfied. The flow (1.1) is degenerate parabolic but not strictly parabolic. Our proof of existence and uniqueness involves constructing global solutions (locally in time) using Weinstein’s Lagrangian neighbourhood theorem, which results in a nice fourth order parabolic scalar equation. This equation has a good structure, satisfying conditions required in [MM12] so we can conclude uniqueness and existence within a given Lagrangian neighbourhood. We then argue that the flows described by solutions to the scalar fourth order equations are in correspondence with the normal flows of the form (1.1) leading to uniqueness and extendability provided the flow remains smooth. After proving existence and uniqueness using Weinstein neighborhoods, we turn to local Darboux charts to prove higher regularity from second fundamental form bounds. It’s not immediately clear how to extract regularity from arbitrary Weinstein neighbourhoods as the submanifolds move, but using a description of Darboux charts in [JLS11] we can fix a finite set of charts and perform the regularity theory in these charts. Unlike mean curvature flow, there is no maximum principle for fourth order equations. We deliver a regularity theory using Sobolev spaces. The newly introduced flow already exhibits nice properties in special cases: (1) For Calabi-Yau manifolds, Harvey and Lawson [HL82] showed that, for a Lagrangian submanifold, the Lagrangian angle $\theta$ generates the mean curvature via $H=J\nabla\theta.$ In this case (1.1) becomes $\frac{dF}{dt}=-J\nabla\Delta\theta,$ while $\theta$ satisfies the pleasant fourth order parabolic equation $\frac{d\theta}{dt}=-\Delta_{g}^{2}\theta$ where $g$ is the induced metric on $L_{t}$. (2) The case $n=1$ involves curves in $\mathbb{C}$. Hamiltonian isotopy classes bound a common signed area. Higher order curvature flows have been studied and are called polyharmonic heat flow or curve diffusion flow (cf. [PW16], [EI05], [May01]), dealing with evolution in the form $\gamma^{\prime}=(-1)^{p-1}\kappa_{s^{2p}}$ where $\kappa_{s^{2p}}$ is the $2p$-th order derivative of the curvature $\kappa$ of a plane curve $\gamma$ with respect to the arclength $s$. For $p=0$ it is the standard curve shortening flow. The flow discussed here corresponds to $p=1$. The case $n=1$ arises in material science [Mul99]. For more work on the pure side, see [Whe13] and the many references therein. In [Whe13] it is shown that curves near a circle have flows which exist forever and converge to a round circle, as well as a short-time existence theorem that requires only $L^{2}$ initial curvature. It is not difficult to argue that for immersed figure 8 type curves singularities must develop in finite time. It is curious to know whether there exist embedded curves which develop singularities in finite time (cf. [EI05], [PW16], [May01]). ## 2\. Gradient Flow We begin by setting up the equation as a formal gradient flow over an $L^{2}$ metric space. Given an embedded Lagrangian submanifold $L^{n}\subset M^{2n}$, we can consider Hamiltonian deformations of $L$, these will be flows of vector fields $J\nabla f$ for scalar functions $f$ on $M$. At $x\in L$, the normal component of $J\nabla f$ is given by $J\nabla_{L}f$ where $\nabla_{L}f$ is the gradient of $f$ as a function restricted to $L$; conversely, given any smooth function $f$ on an embedded $L,$ we can extend $f$ to a function on $M$ so that $\nabla_{L}f$ is not changed and is independent of the extension of $f$ (see section 2.1 below). In other words, given a family of $C^{1}$ functions $f\left(\cdot,t\right)$ along $L,$ one can construct a family of embeddings (2.1) $F:L\times\left(-\varepsilon,\varepsilon\right)\rightarrow M$ satisfying (2.2) $\frac{d}{dt}F(x,t)=J\nabla f\left(x,t\right)$ and conversely, given any path (2.1) within a Hamiltonian isotopy class, there will be a function $f$ so that (possibly after a diffeomorphism to ensure the deformation vector is normal) the condition (2.2) is satisfied. Let $\mathcal{I}_{L_{0}}$ be the set of smooth manifolds that are Hamiltonian isotopic to $L_{0}$. The (smooth) tangent space at any $L\in\mathcal{I}_{L_{0}}$ is parameterized via (2.3) $T_{L}\mathcal{I}_{L_{0}}\mathcal{=}\left\\{f\in C^{\infty}(L):\int_{L}fdV_{g}=0\right\\}.$ We use the $L^{2}$ metric on $T_{L}\mathcal{I}_{L_{0}}$: For $JX_{i}=\nabla f_{i}$ $\in T_{L}\mathcal{I}_{L_{0}},i=1,2$ (2.4) $\langle X_{1},X_{2}\rangle=\int_{L}f_{1}f_{2}\,dV_{g}.$ With volume function given by $\operatorname{Vol}(L)=\int_{L}dV_{g}$ the classical first variation formula gives $d\operatorname{Vol}(L)(W)=-\int_{L}\langle W,H\rangle dV_{g}$ where $W$ is the deformation vector field and $H$ is the mean curvature vector. In the situation where allowable deformation vectors are of the form $W=J\nabla f,$ we get $\displaystyle d\operatorname{Vol}(L)(f)$ $\displaystyle=-\int_{L}\langle J\nabla f,H\rangle dV_{g}$ $\displaystyle=\int_{L}\langle\nabla f,JH\rangle dV_{g}$ $\displaystyle=-\int_{L}f\operatorname{div}\left(JH\right)dV_{g}$ using the fact that $J$ is orthogonal, then integrating by parts. Note that $-\operatorname{div}\left(JH\right)$ belongs to $T_{L}\mathcal{I}_{L_{0}}$ since it integrates to 0 on $L$. Therefore, it is the gradient of the volume function with respect to the metric. Thus a (volume decreasing) gradient flow for volume would be a path satisfying (2.5) $\frac{dF}{dt}=J\nabla\operatorname{div}\left(JH\right).$ ###### Remark 2.1. The metric (2.3) is not the usual $L^{2}$ metric for deformations of a submanifold, which would measure the length of the tangent vector by $\int\left|J\nabla f\right|^{2}dV_{g}.$ It is better suited than the standard metric on vector fields. Suppose instead we take the “standard” $L^{2}$ metric on deformation fields: $d\operatorname{Vol}(L)(J\nabla f)=-\int_{L}\langle J\nabla f,H\rangle dV_{g}$ The gradient with respect to this metric would be $J\nabla\eta$ for some $\eta\in C^{\infty}(L)$ such that $-\int_{L}\langle J\nabla f,H\rangle=\int_{L}\langle J\nabla f,J\nabla\eta\rangle.$ While the Lagrangian angle $\theta$ (in the Calabi-Yau case) does produce this gradient locally, typically $\theta$ is not globally defined on $L$. So instead, we must find a function $\eta$ solving the equation $\Delta\eta=-\operatorname{div}(JH)$ which one can solve uniquely up to additive constants since $L$ is compact and $\nabla\eta=-JH$ \+ $X$ (divergence free vector field on $L$). A gradient flow would be $dF/dt=H-JX$ but there is no canonical way to determine $X$. It is also worth noting that gradient flow with respect to the $L^{2}$ metric (sometimes called $\mathcal{H}^{-1}$) is not new: It has been used for example in mechanics to describe the flow of curves [Fif00]. ### 2.1. Related definitions of Hamiltonian deformations Traditionally, Hamiltonian isotopies are defined as flows of the entire manifold along the direction of a time-dependent vector field $J\nabla f$ for some $f$ a smooth function on $M$. Two submanifolds are Hamiltonian isotopic if the one submanifold is transported to the other via the isotopy. In order to use the description (2.3), we note the following standard result: ###### Lemma 2.2. For a smooth flow of embedded Lagrangian submanifolds satisfying (2.6) $\frac{dF}{dt}=J\nabla f(\cdot,t)\text{ }$ for some function $f$ defined on $L$ for each $t$, there exists a function $\tilde{f}$ on $M\times[0,1]$ that defines a Hamiltonian isotopy on $M$ and determines the same Hamiltonian isotopy of the submanifolds. Conversely, given a global Hamiltonian isotopy determined by $\tilde{f}$, the function $\tilde{f}$ restricted to $L_{t}$ determines a flow of the form (2.6), possibly up to reparameterization by diffeomorphisms of $L.$ ###### Proof. The function $f$ is defined on a smooth compact submanifold of $M\times[0,1]$. We can use the Whitney extension theorem to extend a smooth function off this set in which the normal derivatives vanish. Thus along the Lagrangian submanifolds, $J\nabla f=J\nabla\tilde{f}$. Conversely, given any function $\tilde{f}$ its gradient decomposes into the normal and tangential parts on the Lagrangian submanifold. By the Lagrangian condition, $J\nabla^{T}\tilde{f}$ is normal and $J\nabla^{\perp}\tilde{f}$ is tangential with the latter component describing merely a reparameterization of $L$. So the flow is completely determined by the component $J\nabla^{T}\tilde{f}$ which is determined by the restriction to $L$. ∎ ### 2.2. Immersed Lagrangian submanifolds and their Hamiltonian deformations Along the evolution equation (1.1), it is feasible that a submanifold which is initially embedded will become merely immersed. Thus we would like the equation to behave well even when the submanifold is immersed. Weinstein’s Lagrangian neighborhood Theorem for immersed Lagrangian submanifolds [EM02, Theorem 9.3.2] states that any Lagrangian immersion $F_{0}:L\to M$ extends to an immersion $\Psi$ from a neighborhood of the 0-section in $T^{*}L$ to $M$ with $\Psi^{*}\omega_{M}=\omega_{\text{can}}$. Sections of the cotangent bundle $T^{*}L$ are clearly embedded as graphs over the 0-section of $T^{*}L$ which is identified with $L$, so by factoring the immersion through $T^{*}L$, we get immersed submanifolds in $M$, in particular, immersed Lagrangian submanifolds in $M$ for sections defined by closed 1-forms on $L$. Even though the deformation of an immersed manifold is not properly Hamiltonian (that is, velocity vector $J\nabla f$ determined by a global function $f$ on $M$) one can define deformations by using a function $f$ defined on the submanifold, and $J\nabla f$ makes sense within $T^{*}L$ as pullback by the immersion. For example, the figure 8 is not problematic because the two components of a neighborhood of the crossing point can have different velocity vectors; these are separated within the cotangent bundle. ### 2.3. The evolution equation in terms of $\theta$ By [HL82], for a Lagrangian submanifold $L$ in a Calabi-Yau manifold $(M^{n},\omega,J,\Omega)$ with a covariant constant holomorphic $n$-form $\Omega$, the mean curvature of $L$ satisfies $H=J\nabla\theta$ where $\Omega|_{L}=e^{i\theta}d\operatorname{Vol}_{L}.$ Now (2.5) leads to $\frac{dF}{dt}=J\nabla\operatorname{div}\left(JH\right)=J\nabla\operatorname{div}\left(JJ\nabla\theta\right)=-J\nabla\Delta\theta.$ Differentiating the left-hand side (cf. [Woo20, Prop 3.2.1]): $\displaystyle\frac{d}{dt}\Omega|_{L}$ $\displaystyle=\frac{d}{dt}\left(F_{t}^{\ast}\Omega\right)=F_{0}^{\ast}\mathcal{L}_{-J\nabla\Delta\theta}\Omega$ $\displaystyle=F_{0}^{\ast}d\left(\iota_{-J\nabla\Delta\theta}\Omega\right)=d\left(F_{0}^{\ast}\left(\iota_{-J\nabla\Delta\theta}\Omega\right)\right)$ $\displaystyle=d\left(F_{0}^{\ast}i\left(\iota_{-\nabla\Delta\theta}\Omega\right)\right)$ $\displaystyle=d\left(ie^{i\theta}dVol_{L}(-\nabla\Delta\theta,\cdot,...,\cdot)\right)$ $\displaystyle=-d\left(ie^{i\theta}\ast d\Delta\theta\right)$ $\displaystyle=\left(e^{i\theta}\ast d\Delta\theta-ie^{i\theta}d\left(\ast d\Delta\theta\right)\right).$ Then differentiating the right hand side: $\frac{d}{dt}e^{i\theta}d\operatorname{Vol}_{L}=e^{i\theta}\frac{d}{dt}d\operatorname{Vol}_{L}+ie^{i\theta}\frac{d\theta}{dt}d\operatorname{Vol}_{L}.$ Comparing the imaginary parts (after multiplying by $e^{-i\theta}$) of the above two gives (2.7) $\frac{d\theta}{dt}=-\Delta_{g(t)}^{2}\theta.$ ## 3\. Existence and Uniqueness via a scalar equation on a Lagrangian Neighborhood The system of equations (1.1) is not strictly parabolic as given. Our approach is to make good use of the Lagrangian property, in particular, by setting up the equation as a scalar, uniformly parabolic equation via Weinstein’s Lagrangian neighborhood theorem. For the convenience of using common terminologies, we make our discussion for embeddings but the conclusions hold for immersions in view of subsection 2.2. ### 3.1. Accompanying flow of scalar functions Let $L$ be an embedded compact Lagrangian submanifold in a symplectic manifold $(M,\omega)$. By Weinstein’s Lagrangian neighborhood [Wei71, Corollary 6.2] theorem, there is a diffeomorphism $\Psi$ from a neighborhood $U\subset T^{*}L$ of the 0-section (identified with $L$) to a neighborhood $V\subset M$ of $L$ such that $\Psi^{\ast}\omega=d\lambda_{can}$ and $\Psi$ restricts to the identity map on $L$. Let $\varphi(x,t)$ be a smooth function on $L\times[0,\delta)$ with $\varphi(\cdot,0)=0$. Then $d\varphi$ is a $t$-family of exact (hence closed) 1-forms on $L$ hence a family of sections of $T^{\ast}L$ and each is a graph over the 0-section. The symplectomorphism $\Psi$ yields a $t$-family of Lagrangian submanifolds $L_{t}$ in $M$ near $L$: (3.1) $F=\Psi(x,d\varphi(x,t))=\Psi\left(x,\frac{\partial\varphi}{\partial x^{k}}dx^{k}\right).$ ###### Proposition 3.1. Suppose that $d\varphi(x,t)$ is an exact section describing an evolution of Lagrangian submanifolds which satisfy the equation (3.2) $\left(\frac{dF}{dt}\right)^{\perp}=J\nabla\operatorname{div}\left(JH\right).$ Then there is a function $G$ (depending on $\Psi$) such that $\varphi$ satisfies $\frac{\partial\varphi}{\partial t}=-g^{ap}g^{ij}\frac{\partial^{4}\varphi}{\partial x^{a}\partial x^{j}\partial x^{i}\partial x^{p}}+G(x,D\varphi,D^{2}\varphi,D^{3}\varphi).$ The coordinate free expression is $\frac{\partial\varphi}{\partial t}=\operatorname{div}J\Delta_{g}\Psi(x,d\varphi).$ ###### Proof. Taking $(x,v)$ for coordinates of $T^{\ast}L,$ let $y^{\alpha}$ be coordinates in $M$, $\alpha=1,...,2n$. This gives a frame (3.3) $e_{i}:=\frac{\partial F}{\partial x^{i}}=\frac{\partial\Psi^{\alpha}}{\partial x^{i}}\frac{\partial}{\partial y^{\alpha}}+\frac{\partial\Psi^{\alpha}}{\partial v^{k}}\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{j}}\delta^{jk}\frac{\partial}{\partial y^{\alpha}}=\Psi_{i}+\varphi_{ij}\delta^{jk}\Psi_{k+n}$ where $\displaystyle\Psi_{i}$ $\displaystyle:=D_{x^{i}}\Psi=\frac{\partial\Psi^{\alpha}}{\partial x^{i}}\frac{\partial}{\partial y^{\alpha}}$ $\displaystyle\Psi_{j+n}$ $\displaystyle:=D_{v^{j}}\Psi=\frac{\partial\Psi^{\alpha}}{\partial v^{j}}\frac{\partial}{\partial y^{\alpha}}.$ Letting also $F_{i}^{\alpha}:=\frac{\partial\Psi^{\alpha}}{\partial x^{i}}+\frac{\partial\Psi^{\alpha}}{\partial v^{k}}\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{j}}\delta^{jk}$ we have $e_{i}=F_{i}^{\alpha}\frac{\partial}{\partial y^{\alpha}}.$ Now suppose $h=h_{\alpha\beta}dy^{a}dy^{\beta}$ is the Riemannian metric on $M$. We are also assuming $\omega(V,W)=h(JV,W)$. We compute the induced metric $g$ from the immersion: (3.4) $\displaystyle g_{ij}$ $\displaystyle=h(\partial_{i}F,\partial_{j}F)$ $\displaystyle=\left(\frac{\partial\Psi^{\alpha}}{\partial x^{i}}\frac{\partial\Psi^{\beta}}{\partial x^{j}}+\sum_{k}\left(\frac{\partial\Psi^{\alpha}}{\partial x^{i}}\frac{\partial\Psi^{\beta}}{\partial v^{k}}\frac{\partial^{2}\varphi}{\partial x^{j}\partial x^{k}}+\frac{\partial\Psi^{\alpha}}{\partial x^{j}}\frac{\partial\Psi^{\beta}}{\partial v^{k}}\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{k}}\right)+\sum_{k.l}\frac{\partial\Psi^{\alpha}}{\partial v^{k}}\partial_{ik}^{2}\varphi\frac{\partial\Psi^{\beta}}{\partial v^{l}}\partial_{jl}^{2}\varphi\right)h_{\alpha\beta}.$ $\displaystyle=h(\Psi_{i}+\varphi_{ik}\Psi_{k+n},\Psi_{j}+\varphi_{jl}\Psi_{l+n})$ $\displaystyle=h(\Psi_{i},\Psi_{j})+\sum_{k}\left(\varphi_{ik}h(\Psi_{k+n},\Psi_{j})+\varphi_{jk}h(\Psi_{i},\Psi_{k+n})\right)+\sum_{k,l}\varphi_{ik}\varphi_{jl}h(\Psi_{k+n},\Psi_{l+n}).$ Since $\Psi:T^{\ast}L\rightarrow M$ is a symplectomorphism with $\Psi^{\ast}\omega=dx\wedge dv$, we have (3.5) $\displaystyle\delta_{ij}$ $\displaystyle=dx\wedge dv(\partial/\partial x^{i},\partial/\partial v^{j})=\Psi^{\ast}\omega\,(\partial/\partial x^{i},\partial/\partial v^{j})$ $\displaystyle=h(J\Psi_{\ast}(\partial/\partial x^{i}),\Psi_{\ast}(\partial/\partial v^{j}))=h(J\Psi_{i},\Psi_{j+n}).$ Similarly (3.6) $h(J\Psi_{i+n},\Psi_{j+n})=\omega(\Psi_{\ast}(\partial/\partial v^{i}),\Psi_{\ast}(\partial/\partial v^{j}))=0.$ Now as $F$ describes a Lagrangian manifold, (summing repeated indices below) (3.7) $\displaystyle 0$ $\displaystyle=\omega(e_{i},e_{j})=h(J\Psi_{i}+\partial_{k}\varphi_{i}J\Psi_{k+n},\Psi_{j}+\partial_{l}\varphi_{j}\Psi_{l+n})$ $\displaystyle=h(J\Psi_{i},\Psi_{j})+\partial_{l}\varphi_{j}h(J\Psi_{i},\Psi_{l+n})+\partial_{k}\varphi_{i}h(J\Psi_{k+n},\Psi_{j})+\partial_{k}\varphi_{i}\partial_{l}\varphi_{j}h(J\Psi_{k+n},\Psi_{l+n})$ $\displaystyle=h(J\Psi_{i},\Psi_{j})-\varphi_{jk}h(\Psi_{i},J\Psi_{k+n})+\varphi_{ik}h(J\Psi_{k+n},\Psi_{j})+\varphi_{ik}\varphi_{jl}h(J\Psi_{k+n},\Psi_{l+n})$ $\displaystyle=h(J\Psi_{i},\Psi_{j})+\varphi_{ik}\varphi_{jl}h(J\Psi_{k+n},\Psi_{l+n})\hskip 142.26378pt\mbox{by \eqref{psi}}$ $\displaystyle=h(J\Psi_{i},\Psi_{j}).$ Now, $\\{\Psi_{i},J\Psi_{j}:1\leq i,j\leq n\\}$ is a basis for the ambient tangent space at a point in the image of $F$. So is $\\{\Psi_{i},\Psi_{j+n}:1\leq i,j\leq n\\}$ (as $\Psi$ is a local diffeomorphism). We represent the latter vectors by (3.8) $\Psi_{i+n}=a^{ij}\Psi_{j}+b^{ij}J\Psi_{j}.$ Computing the pairing $h\left(J\Psi_{j},\Psi_{i+n}\right)$ using (3.5) on the left and (3.8) on the right yields $b^{ij}=h^{ij}$ as the inverse of the positive definite matrix $h_{ij}:=h(\Psi_{i},\Psi_{j})$. Now recalling (3.1) $\frac{\partial F}{\partial t}=\left(\frac{\partial}{\partial t}\frac{\partial\varphi}{\partial x^{k}}\right)\frac{\partial\Psi}{\partial v^{k}},$ project onto the normal space: $\displaystyle\left(\frac{\partial F}{\partial t}\right)^{\perp}$ $\displaystyle=\frac{\partial\varphi_{t}}{\partial x^{k}}h(\Psi_{k+n},Je_{p})Je_{q}g^{pq}$ $\displaystyle=\frac{\partial\varphi_{t}}{\partial x^{k}}h(\Psi_{k+n},J\Psi_{p}+\varphi_{pj}J\Psi_{j+n})Je_{q}g^{pq}$ $\displaystyle=\frac{\partial\varphi_{t}}{\partial x^{k}}h(\Psi_{k+n},J\Psi_{p})Je_{q}g^{pq}$ $\displaystyle=\frac{\partial\varphi_{t}}{\partial x^{k}}\delta_{kp}Je_{q}g^{pq}\,\,\,\,\,\,{\mbox{ by \eqref{psi}}}$ $\displaystyle=\frac{\partial\varphi_{t}}{\partial x^{k}}\left(Je_{q}g^{kq}\right)$ $\displaystyle=J\nabla\varphi_{t}.$ Let $H=H^{m}Je_{m}$ for the Lagrangian $L$. As $JH$ is tangential its divergence on $L$ is (3.9) $\displaystyle\operatorname{div}(JH)$ $\displaystyle=-\operatorname{div}\left(H^{m}e_{m}\right)$ $\displaystyle=-g^{ab}h(\nabla_{e_{a}}\left(H^{m}e_{m}\right),e_{b})$ $\displaystyle=-g^{ab}h\left(\frac{\partial}{\partial x^{a}}H^{m}e_{m}+H^{m}\Gamma_{am}^{p}e_{p},e_{b}\right)$ $\displaystyle=-g^{ab}\frac{\partial}{\partial x^{a}}H^{m}g_{mb}-g^{ab}H^{m}\Gamma_{am}^{p}g_{pb}$ $\displaystyle=-\frac{\partial}{\partial x^{a}}H^{a}-H^{m}\Gamma_{am}^{a}$ where the Christoffel symbols are for the induced metric $g$. The components of $H$ are given by (3.10) $H^{a}=h\left(H,Je_{p}\right)g^{ap}=h\left(g^{ij}\left(\frac{\partial^{2}F^{\beta}}{\partial x^{i}\partial x^{j}}+F_{i}^{\alpha}F_{j}^{\gamma}\tilde{\Gamma}_{\alpha\gamma}^{\beta}\right)\frac{\partial}{\partial y^{\beta}},Je_{p}\right)g^{ap}.$ Now differentiate components of (3.3) $\frac{\partial^{2}F^{\beta}}{\partial x^{i}\partial x^{j}}=\frac{\partial\Psi^{\beta}}{\partial x^{j}\partial x^{i}}+\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial\Psi^{\beta}}{\partial v^{l}}+\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial^{2}\Psi^{\beta}}{\partial x^{j}\partial v^{l}}.$ Plug in (3.3) to get $\displaystyle h$ $\displaystyle\left(\frac{\partial^{2}F^{\beta}}{\partial x^{i}\partial x^{j}}\frac{\partial}{\partial y^{\beta}},Je_{p}\right)=\omega\left(e_{p},\frac{\partial^{2}F^{\beta}}{\partial x^{i}\partial x^{j}}\frac{\partial}{\partial y^{\beta}}\right)$ $\displaystyle=$ $\displaystyle\omega\left(\frac{\partial\Psi^{\delta}}{\partial x^{p}}\frac{\partial}{\partial y^{\delta}}+\frac{\partial^{2}\varphi}{\partial x^{p}\partial x^{q}}\delta^{qm}\frac{\partial\Psi^{\delta}}{\partial v^{m}}\frac{\partial}{\partial y^{\delta}},\frac{\partial\Psi^{\beta}}{\partial x^{j}\partial x^{i}}\frac{\partial}{\partial y^{\beta}}+\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial\Psi^{\beta}}{\partial v^{l}}\frac{\partial}{\partial y^{\beta}}+\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial^{2}\Psi^{\beta}}{\partial x^{j}\partial v^{l}}\frac{\partial}{\partial y^{\beta}}\right)$ $\displaystyle=$ $\displaystyle\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{k}}\delta^{kl}\omega\left(\frac{\partial\Psi^{\delta}}{\partial x^{p}}\frac{\partial}{\partial y^{\delta}},\frac{\partial\Psi^{\beta}}{\partial v^{l}}\frac{\partial}{\partial y^{\beta}}\right)+\frac{\partial^{2}\varphi}{\partial x^{p}\partial x^{q}}\delta^{qm}\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{k}}\delta^{kl}\omega\left(\frac{\partial\Psi^{\delta}}{\partial v^{m}}\frac{\partial}{\partial y^{\delta}},\frac{\partial\Psi^{\beta}}{\partial v^{l}}\frac{\partial}{\partial y^{\beta}}\right)$ $\displaystyle+F_{p}^{\delta}\left(\frac{\partial\Psi^{\beta}}{\partial x^{j}\partial x^{i}}+\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial^{2}\Psi^{\beta}}{\partial x^{j}\partial v^{l}}\right)\omega_{\delta\beta}\circ F$ $\displaystyle=$ $\displaystyle\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{k}}\delta^{kl}\Psi^{\ast}\omega\left(\frac{\partial}{\partial x^{p}},\frac{\partial}{\partial v^{l}}\right)+\frac{\partial^{2}\varphi}{\partial x^{p}\partial x^{q}}\delta^{qm}\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{k}}\delta^{kl}\Psi^{\ast}\omega\left(\frac{\partial}{\partial v^{m}},\frac{\partial}{\partial v^{l}}\right)$ $\displaystyle+F_{p}^{\delta}\left(\frac{\partial\Psi^{\beta}}{\partial x^{j}\partial x^{i}}+\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial^{2}\Psi^{\beta}}{\partial x^{j}\partial v^{l}}\right)\omega_{\delta\beta}\circ F$ $\displaystyle=$ $\displaystyle\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{k}}\delta^{kl}\delta_{pl}+F_{p}^{\delta}\left(\frac{\partial\Psi^{\beta}}{\partial x^{j}\partial x^{i}}+\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial^{2}\Psi^{\beta}}{\partial x^{j}\partial v^{l}}\right)\omega_{\delta\beta}\circ F.$ Now also $h\left(\frac{\partial}{\partial y^{\beta}},Je_{p}\right)=\omega\left(F_{p}^{\delta}\frac{\partial}{\partial y^{\delta}},\frac{\partial}{\partial y^{\beta}}\right)=F_{p}^{\delta}\omega_{\delta\beta}\circ F.$ Combining (3.10) and the above $\displaystyle H^{a}$ $\displaystyle=g^{ij}g^{ap}\left(\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{k}}\delta^{kl}\delta_{pl}+F_{p}^{\delta}\left(\frac{\partial\Psi^{\beta}}{\partial x^{j}\partial x^{i}}+\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial^{2}\Psi^{\beta}}{\partial x^{j}\partial v^{l}}\right)\omega_{\delta\beta}\circ F\right)$ $\displaystyle+g^{ij}g^{ap}F_{i}^{\alpha}F_{j}^{\gamma}\tilde{\Gamma}_{\alpha\gamma}^{\beta}F_{p}^{\delta}\omega_{\delta\beta}\circ F.$ Thus using the expression we derived in (3.9) (3.11) $\displaystyle\operatorname*{div}$ $\displaystyle(JH)=-g^{ap}g^{ij}\frac{\partial^{4}\varphi}{\partial x^{a}\partial x^{j}\partial x^{i}\partial x^{p}}-\left(\frac{\partial}{\partial x^{a}}\left(g^{ij}g^{ap}\right)\right)\frac{\partial^{3}\varphi}{\partial x^{j}\partial x^{i}\partial x^{p}}$ $\displaystyle-\frac{\partial}{\partial x^{a}}\left(g^{ij}g^{ap}\right)F_{p}^{\delta}\left[\left(\frac{\partial\Psi^{\beta}}{\partial x^{j}\partial x^{i}}+\frac{\partial^{2}\varphi}{\partial x^{i}\partial x^{k}}\delta^{kl}\frac{\partial^{2}\Psi^{\beta}}{\partial x^{j}\partial v^{l}}\right)+F_{i}^{\alpha}F_{j}^{\gamma}\tilde{\Gamma}_{\alpha\gamma}^{\beta}\right]\omega_{\delta\beta}\circ F-H^{m}\Gamma_{am}^{a}.$ Now recalling (3.4), we see the metric components $g_{ab}$ involve second order derivatives in terms of $\varphi,$ thus $\Gamma_{ij}^{k}$ are third order. So each term above after the first term is at most third order. ∎ ### 3.2. Short time existence ###### Proposition 3.2. Given an initial smooth immersion of a compact $L\rightarrow M,$ there exists a solution to (1.1,1.2) for some short time. ###### Proof. Choose a Weinstein neighborhood containing $L.$ Now suppose we have $\varphi$ which satisfies the fourth order equation (3.12) $\displaystyle\varphi_{t}=-g^{ap}g^{ij}\frac{\partial^{4}\varphi}{\partial x^{a}\partial x^{j}\partial x^{i}\partial x^{p}}+G(x,D\varphi,D^{2}\varphi,D^{3}\varphi)=\operatorname{div}(JH)$ $\displaystyle\varphi(\cdot,0)=0.$ Then the immersions $F$ generated from $\varphi(x,t)$ satisfy $\displaystyle\left(\frac{\partial F}{\partial t}\right)^{\perp}$ $\displaystyle=J\nabla\varphi_{t}=J\nabla\operatorname{div}(JH).$ As the normal component satisfies the appropriate equation, we may compose with diffeomorphisms to get a flow (see Claim 3.3 below) such that (3.13) $\frac{\partial F}{\partial t}=J\nabla\operatorname{div}(JH).$ Now the equation (3.12) is precisely of the form of $2p$ order quasilinear parabolic equation studied in [MM12]. By [MM12, Theorem 1.1] we have short time existence for the solution to (3.12), thus we have short-time existence for the flow (3.13). ∎ ### 3.3. Uniqueness We start with a standard observation. ###### Claim 3.3. Suppose that $F:L\times[0,T)\rightarrow M$ is a family of immersions satisfying $\left(\frac{\partial F}{\partial t}\right)^{\perp}=N(x,t)$ for some vector field $N(x,t)$ which is normal to the immersed submanifold $F(\cdot,t)(L)$. There exists a unique family of diffeomorphisms $\chi_{t}:L\rightarrow L$ such that $\frac{\partial}{\partial t}F(\chi_{t}(x),t)=N(\chi_{t}(x),t)\,\,\,\,{and}\,\,\,\,\chi_{0}=Id_{|L}.$ ###### Proof. Given the flow exists, the given velocity field will decompose orthogonally into normal and tangential components: $\frac{\partial F}{\partial t}=N(x,t)+T\left(x,t\right).$ Consider the time-dependent vector field on $L$ $V(x,t)=-D_{L}F(x,t)^{-1}T\left(x,t\right)$ By the Fundamental Theorem on Flows, (cf. [Lee13, Theorem 9.48]) there is a unique flow on $L$ starting at the identity and satisfying $\frac{\partial}{\partial t}\chi_{t}(x)=V(\chi_{t}(x),t).$ Composing this flow with the original flow $F$ yields the result. ∎ ###### Theorem 3.4. The solution to the initial value problem (1.1) - (1.2) is unique. More precisely, if $F_{1}$ and $F_{2}$ are two solutions of (1.1) such that $F_{1}(x,t_{0})=F_{2}(x,t_{0}^{\prime})$ for some $t_{0},t_{0}^{\prime}$ and all $x\in L$, then $F_{1}(x,t_{0}+\tau)=F_{2}(x,t_{0}^{\prime}+\tau)$ for all $\tau$ in an open neighborhood of $0$ where both sides above are defined. ###### Proof. Without loss of generality, we take $t_{0}=t_{0}^{\prime}=0$. Let $L=F_{1}(\cdot,0)=F_{2}(\cdot,0)$ and $\Psi:U\subset T^{\ast}L\to V\subset M$ be a Lagrangian neighborhood mapping. First, we show that the normal flow of $F_{i}(\cdot,t)$ is given in the neighborhood $\Psi$ by the graph of an exact section $d\varphi_{i}(\cdot,t)$ where $\varphi_{i}$ solves a problem of the form (3.12). To this end, note that for $\tau$ in the domain, the path $\left\\{F_{i}(\cdot,t),t\in[0,\tau]\text{ }\right\\}$ is a Hamiltonian isotopy between $F_{i}(\cdot,0)$ and $F_{i}(\cdot,\tau)$. Being a Hamiltonian isotopy is invariant under the symplectomorphism $\Psi$, so the sections $\Psi^{-1}\left(F_{i}(\cdot,0)\right)$ and $\Psi^{-1}\left(F_{i}(\cdot,\tau)\right)$ are Hamiltonian isotopic. By [Wei71, Corollary 6.2], Lagrangian submanifolds that are near to the $0$-section are given as graphs of closed sections of the cotangent bundle. As the flow is smooth, for small times the Lagrangian submanifolds are near enough to be described by closed sections. According to [MS17, Proposition 9.4.2], these sections are exact, that is $\Psi^{-1}\circ F_{i}(\cdot,\tau)\left(L\right)=\left\\{d\varphi_{i}\left(x,\tau\right):x\in L\right\\}.$ In other words, $\Psi(\left\\{d\varphi_{i}\left(x,\tau\right):x\in L_{i}\right\\})=F_{i}(\cdot,\tau)\left(L\right)$ meaning that for each $\tau$ $\Psi\circ d\varphi_{i}:L\rightarrow T^{\ast}L\rightarrow M$ is a Lagrangian immersion, which may have reparameterized the base. In particular, the flow $F_{i}$ determines a flow of scalar functions $\varphi_{i},$ which recovers the same family of submanifolds at $F_{i}$ (up to reparameterization) as do $\Psi\circ d\varphi_{i}\left(\cdot,t\right)$. By Proposition 3.1, the scalar equation (3.12) holds on $\varphi_{i}$ as does the initial condition $d\varphi_{i}\equiv 0$. It follows that $\varphi_{1}$ and $\varphi_{2}$ both satisfy the same equation (3.12) and have the same initial condition, so $\varphi_{1}=\varphi_{2}+C$ for some constant $C$. Thus the flows $F_{1}$ and $F_{2}$ are the same. ∎ Theorem 3.4 allows for seamless extension of the flow: While the Weinstein’s Lagrangian neighborhood may only exist around $L_{0}$, if another Lagrangian neighborhood of $L_{0}$ extends the flow, the two flows patch together smoothly. ## 4\. Higher order estimates based on curvature bounds The goal of this section is to show that a solution with uniformly bounded second fundamental form over $[0,T)$ enjoy estimates of all orders and can be extended. ###### Theorem 4.1. Suppose that the flow (1.1) exists on $[0,T)$ and the second fundamental form has a uniform bound on $[0,T)$. Then the flow converges smoothly as $t\rightarrow T$ so can be extended to $[0,T+\varepsilon)$ for some $\varepsilon>0$. To prove this theorem, it is essential in our approach to establish a-priori estimates from the integral estimates derived from the following differential inequality: ###### Proposition 4.2. Suppose that $F$ is a solution to (1.1) on $[0,T)$ for a compact Lagrangian submanifold $L$ inside a compact $M$. Suppose the second fundamental form has a uniform bound $K$. There exists $C$ depending on $K$, the ambient geometry of $M$ and $\operatorname{Vol}(L_{0})$ such that for all $k\geq 2$ (4.1) $\frac{d}{dt}\int_{L}\left|\nabla^{k-1}A\right|^{2}dV_{g}(t)\leq C\int_{L}\left|\nabla^{k-1}A\right|^{2}dV_{g}(t)+C\sum_{l=0}^{k-2}\int_{L}\left|\nabla^{l}A\right|^{2}dV_{g}(t).$ A Weinstein neighborhood map determines the equation the flow must satisfy, and we could derive estimates of all orders based on this particular equation. However, the flow is expected to leave a given neighborhood after some time, and we will need to take a new neighborhood. We would need to know the speed of the flow to patch estimates from one neighborhood to another, but this requires knowing the size of the Weinstein neighborhoods around the Lagrangian submanifolds at different times. We require charts with uniform geometric estimates. To obtain these we appeal to uniform local Darboux coordinates given in [JLS11]. These charts are local but are given with uniform geometric bounds. The short-time existence of the flow is already determined by the global Weinstein neighborhoods; we write the flow in these Darboux charts as a scalar equation from which we derive integral estimates for derivatives of any order. ### 4.1. Uniform Darboux charts. We record [JLS11, Prop.3.2 and Prop.3.4] on existence of Darboux coordinates with estimates on a symplectic manifold. Let $\pi:\mathcal{U}\rightarrow M$ be the $U(n)$ frame bundle of $M$. A point in $\mathcal{U}$ is a pair $(p,v)$ with $\pi(p,v)=p\in M$ and $v:\mathbb{R}^{2n}\rightarrow T_{p}M$ an isomorphism satisfying $v^{\ast}(\omega_{p})=\omega_{0}$ and $v^{\ast}(h|_{p})=h_{0}$ (the standard metric on $\mathbb{C}^{n}$). The right action of $U(n)$ on $\mathcal{U}$ is free: $\gamma(p,v)=(p,v\circ\gamma)$ for any $\gamma\in U(n)$. ###### Proposition 4.3 (Joyce-Lee-Schoen). Let $(M,\omega)$ be a real $2n$-dimensional symplectic manifold without boundary, and a Riemannian metric $h$ compatible with $\omega$ and an almost complex structure $J$. Let $\mathcal{U}$ be the $U(n)$ frame bundle of $M$. Then for small $\varepsilon>0$ we can choose a family of embeddings $\Upsilon_{p,v}:B^{2n}_{\varepsilon}\rightarrow M$ depending smoothly on $\left(p,v\right)\in U$, where $B^{2n}_{\varepsilon}$ is the ball of radius $\varepsilon$ about $0$ in $\mathbb{C}^{n},$ such that for all $\left(p,v\right)\in U$ we have 1. (1) $\Upsilon_{p,v}(0)=p$ and $d\Upsilon_{p,v}|_{0}=v:\mathbb{C}^{n}\rightarrow T_{p}M;$ 2. (2) $\Upsilon_{p,v\circ\gamma}(0)\equiv\Upsilon_{p,v}\circ\gamma$ for all $\gamma\in U(n);$ 3. (3) $\Upsilon_{p,v}^{\ast}(\omega)\equiv\omega_{0}=\frac{\sqrt{-1}}{2}\sum_{j=1}^{n}dz_{j}\wedge d\bar{z}_{j};$ and 4. (4) $\Upsilon_{p,v}^{\ast}(h)=h_{0}+O(|z|)$. Moreover, for a dilation map $\mathbf{t}:B^{2n}_{R}\to B^{2n}_{\varepsilon}$ given by $\mathbf{t}(z)=tz$ where $t\leq\varepsilon/R$, set $h^{t}_{p,v}=t^{-2}(\Upsilon_{p,v}\circ\mathbf{t})^{*}h$. Then it holds 1. (5) $\|h^{t}_{p,v}-h_{0}\|_{C^{0}}\leq C_{0}t\ \ \ \ \mbox{and}\ \ \ \|\partial h^{t}_{p,v}\|_{C^{0}}\leq C_{1}t$, where norms are taken w.r.t. $h_{0}$ and $\partial$ is the Levi-Civita connection of $h_{0}$. ###### Proposition 4.4. Suppose that $M$ is a compact symplectic manifold with a compatible Riemannian metric $h$. Suppose that $L$ is a compact Lagrangian submanifold of $M$ with second fundamental form bounded above by $K$ and volume bounded above by $V_{0}$. Given $c_{n}>0,$ there exists an $r_{0}=r_{0}(K,c_{n})>0$ and a finite cover of $L$ by Darboux charts $\Upsilon_{p_{i},v_{i}}$:$B_{r_{0}}^{2n}\rightarrow M$ centered at points $p_{i}$ on $L$ such that 1. (1) The connected component of $L\cap B_{r_{0}}^{2n}$ containing $p_{j}$ is represented by a graph $\left(x,d\varphi^{(j)}\right)$ over $B_{r_{0}}^{2n}\cap\mathbb{R}^{n}\times\left\\{0\right\\}$ for some potential $\varphi^{(j)}$. 2. (2) The tangent plane at each point of this connected component satisfies a closeness condition with respect to the planes $v_{i}(\mathbb{R}^{n}\times\\{0\\}):$ (4.2) $\max_{\begin{subarray}{c}\left|e\right|_{g}=1,\text{ }e\in T_{p}L\\\ \left|\nu\right|_{\delta_{0}}=1,\nu\in\left\\{0\right\\}\times\mathbb{R}^{n}\end{subarray}}e\cdot\nu<c_{n}$ where the dot product is in the euclidean metric $\delta_{0}$, and $c_{n}$ is a small universal constant (say $c_{n}=\frac{1}{10\sqrt{n}}$) chosen so that quantities such as the volume element and coordinate expression for $h$ are bounded by universal constants. 3. (3) The ambient metric $h$ is very close to the euclidean metric, that is $\|h-\delta_{0}\|<c_{n}$ for some $c_{n}$ (can be the same $c_{n}$ as in (2) above). 4. (4) The submanifold $L$ is covered by the charts obtained by restricting these charts to $B_{r_{0}}^{2n}(p_{j})$ 5. (5) The number of such points $\left\\{p_{j}\right\\}$ satisfies (4.3) ${N}(K,V_{0})\leq\frac{C(n)V_{0}}{r_{0}^{n}(K,c_{n})}.$ ###### Proof. At each point $p\in L$ we take a Darboux chart $\Upsilon_{p,v}$ as described above with that $T_{p}L=\mathbb{R}^{n}\times\left\\{0\right\\}$ in the given chart. Note that after some fixed re-scalings, we can assert via Proposition 4.3 that $\Upsilon_{p,v}$ exists on $B_{\varepsilon_{0}}^{2n}$ and satisfies any near euclidean metric conditions we choose to prescribe, including the closeness condition: $\left|h-\delta_{0}\right|<c$. Now we may apply Proposition 5.1 which asserts existence of a ball $B_{r_{0}}^{n}(p)\subset\mathbb{R}^{n}\times\\{0\\}$ over which $L$ is representable as a graph, with (4.2) holding. The quantity $r_{0}$ will depend on $K$. Consider the compact immersed submanifold $L$ as a metric space $(L,d)$. Taking a finite cover of metric balls $B_{r_{0}/4}(p)$ for $p\in L$ and applying Vitalli’s covering Lemma, we conclude that there is a subset of these points $\\{p_{j}\\}$ so that $L=\cup_{i}B_{3_{0}/4}(p_{i})$ and $B_{r_{0}/4}(p_{i})$ are mutually disjoint. By (4.2), $L\cap B_{3r_{0}/4}(p_{i})$ is in the image of a graph given by Proposition 5.1. In particular, the disjoint $B_{r_{0}/4}(p_{i})$’s have a minimum total volume $\omega_{n}c^{n}r_{0}^{n}$. The bound (4.3) on the number of balls follows. As $\\{B_{3r_{0}/4}(p_{i})\\}$ covers $L$ and each of these balls is contained in a graph over $B_{r_{0}}^{n},$ we take the set of the graphs as the cover. ∎ The scalar functions from the the exact sections of $T^{*}L$ are globally defined on $L$ via the abstract Weinstein map $\Psi$. We have utilized them to establish short-time existence and uniqueness for our geometric flow of $F$. However, for higher order a-priori estimates, we need to set up the flow equation in a Darboux chart with estimates on the metric as described above. Fortunately, each $\Upsilon_{p,v}$ is a symplectomorphism, which takes gradient graphs $(x,d\varphi)$ to Lagrangian submanifolds, so the computations in section 3.1 can be repeated verbatim, with $\Upsilon_{p,v}$ in place of $\Psi$. In particular, in each chart, the flow is determined by an equation (4.4) $\varphi_{t}=-g^{ap}g^{ij}\frac{\partial^{4}\varphi}{\partial x^{a}\partial x^{j}\partial x^{i}\partial x^{p}}+G(x,D\varphi,D^{2}\varphi,D^{3}\varphi).$ ###### Remark 4.5. A precise computation in Darboux coordinates of the expression (3.10) gives $\displaystyle h\left(H,Je_{p}\right)$ $\displaystyle=g^{ij}\varphi_{pij}+g^{ij}\tilde{\Gamma}_{ij}^{p+n}+g^{ij}\varphi_{kj}\delta^{km}\tilde{\Gamma}_{i,m+n}^{p+n}+g^{ij}\varphi_{ki}\delta^{km}\tilde{\Gamma}_{m+n,j}^{p+n}$ $\displaystyle+g^{ij}\varphi_{ki}\varphi_{lj}\delta^{km}\delta^{lr}\tilde{\Gamma}_{m+n,r+n}^{p+n}-g^{ij}\tilde{\Gamma}_{ij}^{q}\varphi_{pq}-g^{ij}\varphi_{kj}\delta^{km}\tilde{\Gamma}_{i,m+n}^{q}\varphi_{pq}$ $\displaystyle-g^{ij}\varphi_{ki}\delta^{km}\tilde{\Gamma}_{m+n,j}^{q}\varphi_{pq}-g^{ij}\varphi_{ki}\varphi_{lj}\delta^{km}\delta^{lr}\tilde{\Gamma}_{m+n,r+n}^{q}\varphi_{pq}$ where $\tilde{\Gamma}_{ij}^{q}$ are Christoffel symbols in the ambient metric $\left(M,h\right)$. Considering that each expression of the form $g^{ab}$ is a smooth function in terms of $D^{2}\varphi$ with dependence on zero order of $h$ and each $\tilde{\Gamma}_{ij}^{\beta}$ expression depends on $Dh$ and $h$, one may conclude (after computing $\operatorname{div}(JH)$ as in (3.11)) that $G$ can be written as a sum of expressions that are 1. (1) quadratic in $D^{3}\varphi$ and smooth in $D^{2}\varphi,h$ in a predetermined way 2. (2) linear in $D^{3}\varphi$, smooth in $D^{2}\varphi,h,Dh$ in a predetermined way 3. (3) smooth in $D^{2}\varphi,h$ and linear in $D^{2}h$ in a predetermined way 4. (4) smooth in $D^{2}\varphi,h,Dh$ in a predetermined way. This allows us to make a claim that there is uniform control on the important quantities involved in the equation we are solving. ###### Proposition 4.6. Suppose that $L$ is a compact Lagrangian manifold with volume $V_{0}$ and evolves by (1.1) on $[0,T)$. If the norm of second fundamental $A$ of $L_{t}$ satisfies $|A|_{g(t)}\leq K$ for $t\in[0,T)$, then after a fixed rescaling on $M$ there is a finite set of Darboux charts such that 1. (1) The submanifold is covered by graphs over $B_{1}^{2n}\cap\mathbb{R}^{n}\times\left\\{0\right\\}$. 2. (2) The submanifold is graphical over $B_{5}^{2n}\cap\mathbb{R}^{n}\times\left\\{0\right\\}$ in each chart. 3. (3) The slope bound (4.2) holds over $B_{5}^{n}(0).$ 4. (4) The flow (1.1) is governed by (4.4) locally in these charts. 5. (5) For each chart, the $G$ from (4.4) satisfies a uniform bound on any fixed order derivatives of $G$ (in terms of all four arguments, not with respect to $x$ coordinate before embedding.) 6. (6) The number of charts is controlled (4.5) ${N}(K,V_{0})\leq C(K)\frac{V_{0}}{r_{0}^{n}(K)}.$ ###### Proof. Rescale $M$ so that $r_{0}=5.$ Then the expression $G$ becomes predictably controlled by Remark 4.5. Choosing a cover with interior balls, as in the proof of Proposition 4.4, determines the necessary number of balls. ∎ ### 4.2. Localization Let $L_{t}$ evolve by (1.1) with time $t\in[0,T),$ and assume $\left|A\right|_{g(t)}\leq K$ for all $L_{t}.$ Our goal is to establish integral bounds for $\left|\nabla^{l}A\right|_{g(t)}^{2},$ which only depend on $k,K,M$ and the initial volume $V_{0}$ of $L_{0}.$ To derive the differential inequality (4.1) at any time $t_{0}$, we use Proposition 4.6 and express geometric quantities $g,A,\nabla^{l}A$, etc., in the (no more than $N)$ Darboux charts in terms of $\varphi(x,t)$ for $x\in B_{5}^{n}.$ By compactness of $L$ and smoothness of the flow, the flow will continue to be described by graphs of $d\varphi(x,t)$ in this open union of $N$ charts for $t\in[t_{0},t_{1})$ for some $t_{1}>t_{0}.$ To be precise, each of the Darboux charts in Lemma 4.6 has a product structure; we may assume that each chart contains coordinates $B_{4}^{n}(0)\times B_{2}^{n}(0)$ so that $L$ is graphical over $B_{5}^{n}(0)$ and further that the collection of $B_{1}^{n}(0)\times B_{1}^{n}(0)$ covers $L$. Now we may fix once and for all a function $\eta$ which is equal to $1$ on $B_{1}^{n}(0)\times B_{1}^{n}(0)$ and vanishes within $B_{2}^{n}(0)\times B_{2}^{n}(0)$. For a given chart $\Upsilon^{\alpha}$ (here $\alpha\in\left\\{1,..N\right\\}$ indexes our choice of charts) we call the function $\eta_{\alpha}$. This function will have uniformly bounded dependence on the variables $x$ and $y$ in the chart. Now once these $\eta_{\alpha}$ are chosen, we may then define a partition of unity for the union of charts which form a tubular neighborhood of $L$, which will restrict to a partition of unity for small variations of $L$: (4.6) $\rho_{\alpha}^{2}:=\frac{\eta_{\alpha}^{2}}{\sum\eta_{\alpha}^{2}}.$ By compactness of the unit frame bundle and the smoothness of the family of charts defined in Proposition 4.3, the transition functions between charts will have bounded derivatives to any order. Thus, in a fixed chart, where a piece of $L$ is represented as $\left\\{\left(x,d\varphi(x)\right):x\in B_{1}(0)\right\\}$, the dependence of $\rho_{\alpha}^{2}$ will be uniformly controlled in terms of these variables, so there is a uniform pointwise bound (4.7) $\left|D_{x}^{2}\rho_{\alpha}^{2}\right|\leq C(D^{3}\varphi,D^{2}\varphi,D\varphi,x)$ were this dependence is at most linear on $D^{3}\varphi$. We will be using the $x$ coordinates as charts for $L.$ Note also that, if we have a uniform bound on $\frac{d}{dt}D\varphi$ and $\frac{d}{dt}D^{2}\varphi$ we can conclude a positive lower bound on $t_{1}-t_{0};$ the flow will be described by graphs of $\varphi(x,t)$ in these $N$ charts, and the condition (4.2) will be satisfied for a slightly larger $c_{n}^{\prime}$ (say $c_{n}^{\prime}=\frac{1}{5\sqrt{n}}$ instead of $c_{n}=\frac{1}{10\sqrt{n}})$. #### 4.2.1. Expression for metric and second fundamental form In the Darboux charts for $M$, the manifold $L$ is expressed graphically over the $x$ coordinate via $x\mapsto F(x)=\left(x,d\varphi(x)\right).$ Thus we have a tangential frame: (4.8) $e_{i}=\partial_{x^{i}}F=E_{i}+\varphi_{ik}\delta^{km}E_{m+n}$ with $g_{ij}=h_{ij}+\varphi_{ik}\delta^{km}\varphi_{jl}\delta^{lr}h_{\left(m+n\right)\left(l+n\right)}+\varphi_{jl}\delta^{lr}h_{\left(i\right)\left(l+n\right)}+\varphi_{ik}\delta^{km}h_{\left(m+n\right)\left(j\right)}.$ Recalling (2) and (3) in Proposition 4.4 we may assume that the expression of $h$ in these coordinates is very close to $\delta_{ij}$ and that $D^{2}\varphi$ is not large. Differentiating the components of the induced metric gives (4.9) $\partial_{x^{p}}g_{ij}=\text{function of }\left(x,D\varphi\right)\\\ +\text{Terms involving up to three factors of }D^{2}\varphi\text{ but no higher}\\\ +\text{Terms involving up to two factors of }D^{2}\varphi\text{ and one factor of }D^{3}\varphi.$ ###### Lemma 4.7. In a Darboux chart, using the coordinate basis (4.8) for the tangent space and $\left\\{Je_{l}\right\\}$ for the normal space, the covariant derivatives of the second fundamental form and of the potential $\varphi$ are related by (4.10) $\nabla^{k-1}A=D^{k+2}\varphi+S_{k}$ where $S_{1}$ is a smooth controlled function depending on the chart, $h$ and $D^{2}\varphi$, $S_{2}$ depends also on $D^{3}\varphi$ and for $k\geq 3:$ 1. (1) Each $S_{k}$ is a sum of of multilinear forms of $D^{4}\varphi,...,D^{k+1}\varphi$ 2. (2) The coefficients of these forms are functions of $\left(x,D\varphi,D^{2}\varphi,D^{3}\varphi\right)$ 3. (3) The total sum of the derivatives of $D^{3}\varphi$ that occur in a given term is no more than $k-2.$ (Note that (4.10) is interpreted as literal equality of the symbols in the choice of basis, not simply “up to a smooth function”) ###### Proof. Starting with $k=1$, differentiate in the ambient space $\tilde{\nabla}_{e_{i}}e_{j}=\tilde{\Gamma}_{ji}^{\beta}E_{\beta}+\varphi_{jmi}\delta^{mk}E_{n+k}+\varphi_{jm}\delta^{mk}\tilde{\Gamma}_{n+k,i}^{\beta}E_{\beta}+\varphi_{jm}\varphi_{ri}\delta^{mk}\delta^{rl}\tilde{\Gamma}_{n+k,n+l}^{\beta}E_{\beta}.$ Using $e_{j}$ and $Je_{k}$ as frame and normal frame, $\displaystyle A_{ijl}$ $\displaystyle=\langle\tilde{\nabla}_{e_{i}}e_{j},Je_{l}\rangle$ $\displaystyle=\omega(e_{l},\varphi_{jmi}\delta^{mk}E_{n+k})+\omega(e_{l},\tilde{\Gamma}_{ji}^{\beta}E_{\beta}+\varphi_{jm}\delta^{mk}\tilde{\Gamma}_{n+k,i}^{\beta}E_{\beta}+\varphi_{jm}\varphi_{ri}\delta^{mk}\delta^{rl}\tilde{\Gamma}_{n+k,n+l}^{\beta}E_{\beta})$ $\displaystyle=\varphi_{jli}+S_{1}$ where $S_{1}$ is a smooth function involving $D^{2}\varphi$ and the ambient Christoffel symbols at $\left(x,D\varphi\right)$ and recalling $\omega(e_{l},E_{\beta})=\omega(E_{l}+\varphi_{lk}\delta^{jk}E_{j+n},E_{\beta})=\left\\{\begin{array}[c]{cl}\delta_{sl},&\ \ \ \text{ if }\beta=s+n\text{ for }s\in\left\\{1,...,n\right\\}\\\ -\varphi_{sl},&\,\,\,\,\text{ if }\beta=s\in\left\\{1,...,n\right\\}.\end{array}\right.$ Now for $k=2$ ($\nabla$ denotes covariant derivatives on the submanifold): $\left(\nabla A\right)_{pijl}=\partial_{p}A_{ijl}-A\left(\nabla_{e_{p}}e_{i},e_{j},e_{l}\right)-A\left(e_{i},\nabla_{e_{p}}e_{j},e_{l}\right)-A\left(e_{i},e_{j},\nabla_{e_{p}}e_{l}\right).$ Now we can compute the Christoffel symbols with respect to the induced metric $g$: $\nabla_{e_{p}}e_{i}=1\ast D^{3}\varphi\text{ + lower order }$ thus $\displaystyle\left(\nabla A\right)_{pijl}$ $\displaystyle=\varphi_{jlip}+\partial_{x^{p}}S_{1}-A\ast\left(D^{3}\varphi\right)+\text{lower order}$ $\displaystyle=\varphi_{jlip}+D^{3}\varphi\ast D^{3}\varphi+1\ast D^{3}\varphi+\text{ smooth in other arguments.}$ Here and in sequel, we use $A\ast B$ to denote a predictable linear combination of terms from tensors $A$ and $B$, and $1\ast T$ to be a predictable linear combination of $T.$ Now $\displaystyle\nabla^{2}A$ $\displaystyle=D^{5}\varphi+D^{4}\varphi\ast D^{3}\varphi+\text{lower order}$ $\displaystyle\nabla^{3}A$ $\displaystyle=D^{6}\varphi+D^{5}\varphi\ast D^{3}\varphi+D^{4}\varphi\ast D^{4}\varphi\text{ + lower order}$ and so forth. The result follows by inductively applying the product rule and noting $\nabla^{k-1}A=D(\nabla^{k-2}A)+D^{3}\varphi\ast\nabla^{k-2}A\text{ + lower order}$ by the formula for covariant derivative. ∎ ### 4.3. Integral inequalities We will use $\|\cdot\|_{\infty}$ for the supremum norm in the euclidean metric $\delta_{0}$ and $\left|D^{m}\varphi\right|_{g}^{2}=g^{i_{1}j_{1}}g^{i_{2}j_{2}}...g^{i_{m}j_{m}}\varphi_{i_{1}...i_{m}}\varphi_{j_{1}...j_{m}}$ to denote the norm squared with respect to $g$ for the locally defined $m$-tensor $D^{m}\varphi$ instead of the higher covariant derivative tensor $\nabla^{m}\varphi$. We find that this makes computations on the chosen Darboux chart more transparent. Note that since $g$ is close to $\delta_{0}$ on the chart (with estimates on errors) $\frac{\left|D^{m}\varphi\right|_{g}^{2}}{\left|D^{m}\varphi\right|_{\delta_{0}}^{2}}\in(1-c_{n},1+c_{n})$ and $(1-c_{n})dx\leq dV_{g}\leq\left(1+c_{n}\right)dx$ for some small $c_{n}$. Thus we may regard as equivalent estimates on integrals against $dx$ and integrals against $dV_{g}$, provided the quantities we are integrating are nonnegative. However, if the quantity being integrated is not known to be non-negative, we have to be precise in performing estimates. All estimates below implicitly depend on $c_{n},$ but $c_{n}$ need not be tracked closely: it need not be close to zero. #### 4.3.1. Interpolation inequalities We use Gagliardo-Nirenberg interpolation to derive integral inequalities that allow us to integrate multilinear combinations of higher derivatives of $\varphi$. For simplicity of notation, we will use $C$ for uniform constants with dependence indicated in its arguments. For our application, we give interpolations for different range of indices. ###### Lemma 4.8. Let $\xi$ be a smooth compactly supported vector-valued function on $\mathbb{R}^{n}$. 1. (1) If $j_{1}+j_{2}+j_{3}+...+j_{q}=m,$ then $\int\left|D^{j_{1}}\xi\ast D^{j_{2}}\xi...\ast D^{j_{q}}\xi\right|^{2}\leq C\left\|\xi\right\|_{\infty}^{2q-2}\int\left|D^{m}\xi\right|^{2}.$ 2. (2) If $j_{1}+j_{2}+j_{3}+...+j_{q}+j^{\ast}=2\tilde{m}$ where $j_{q}<\tilde{m}$ and $j^{\ast}\geq 0$, then $\int\left|D^{j_{1}}\xi\ast D^{j_{2}}\xi...\ast D^{j_{q}}\xi\right|\leq C\left\|\xi\right\|_{\infty}^{q-\left(2-j^{\ast}/2\tilde{m}\right)}\left(\frac{2\tilde{m}-j^{\ast}}{2\tilde{m}}\int\left|D^{\tilde{m}}\xi\right|^{2}+\frac{j^{\ast}}{2\tilde{m}}\left\|\chi_{\text{supp}\left(\xi\right)}\right\|_{p^{\ast}}^{2\tilde{m}/j^{\ast}}\right).$ 3. (3) If $j_{1}+...+j_{r}=2\bar{m}+1$ and all $j_{i}\leq\bar{m},$ then for $\varepsilon>0$ $\int\left|D^{j_{1}}\xi\ast D^{j_{2}}\xi...\ast D^{j_{r}}\xi\right|\leq\varepsilon\int\left|D^{\bar{m}+1}\xi\right|^{2}+C(\varepsilon,\bar{m},\|\xi\|_{\infty})\,\left(\int\left|D^{\bar{m}}\xi\right|^{2}+\int\chi_{\text{supp}\left(f\right)}\right).$ ###### Proof. For (1), use $p_{i}=\frac{m}{j_{i}}$ and apply the generalized Hölder’s inequality $\int\left|D^{j_{1}}\xi\ast D^{j_{2}}\xi...\ast D^{j_{q}}\xi\right|^{2}\leq\left\|D^{j_{1}}\xi\right\|_{2p_{1}}^{2}...\left\|D^{j_{q}}\xi\right\|_{2p_{q}}^{2}$ and then use the Gagliardo-Nirenberg interpolation inequality (cf. [FFRS21, Theorem 1.1]) with $\theta_{i}=\frac{j_{i}}{m}$. For (2), taking $p_{i}=\frac{2\tilde{m}}{j_{i}}$ and $p^{\ast}=\frac{2\tilde{m}}{j^{\ast}}$ if $j^{\ast}>0$, then $\int\left|D^{j_{1}}\xi\ast D^{j_{2}}\xi...\ast D^{j_{q}}\xi\right|\leq\left\|D^{j_{1}}\xi\right\|_{p_{1}}...\left\|D^{j_{q}}\xi\right\|_{p_{q}}\left\|\chi_{\text{supp}\left(\xi\right)}\right\|_{p^{\ast}}.$ Now apply the Gagliardo-Nirenberg interpolation inequality with $\theta_{i}=\frac{j_{i}}{\tilde{m}}$, applying Young’s inequality if $j^{\ast}>0$. For (3), we may split, with $a\leq\bar{m}\leq\bar{m}+1\leq b$ $\displaystyle j_{1}+...+j_{s}$ $\displaystyle=a$ $\displaystyle j_{s+1}+...+j_{r}$ $\displaystyle=b.$ Now for some $p,q$ conjugates to be determined, let $p_{i}=\left\\{\begin{array}[c]{cl}\frac{ap}{j_{i}},&\ \ \ \text{ if }i\in\left\\{1,...,s\right\\}\\\ \frac{bq}{j_{i}},&\,\,\,\,\text{ if }i\in\left\\{s+1,...,r\right\\}.\end{array}\right.$ Apply the generalized Hölder’s inequality (4.11) $\int\left|D^{j_{1}}\xi\ast D^{j_{2}}\xi...\ast D^{j_{r}}\xi\right|\leq\left\|D^{j_{1}}\xi\right\|_{p_{1}}...\left\|D^{j_{q}}\xi\right\|_{p_{r}}.$ We have from the Gagliardo-Nirenberg interpolation inequality $\left\|D^{j_{i}}\xi\right\|_{p_{i}}\leq\left\\{\begin{array}[c]{cl}\left\|D^{\bar{m}}\xi\right\|_{\frac{ap}{\bar{m}}}^{\frac{j_{i}}{\bar{m}}}\left\|\xi\right\|_{\infty}^{1-\frac{j_{i}}{\bar{m}}}&\ \ \ \text{ if }i\in\left\\{1,...,s\right\\}\text{ with }\theta_{i}=\frac{j_{i}}{\bar{m}}\\\ \left\|D^{\bar{m}+1}\xi\right\|_{\frac{bq}{\bar{m}+1}}^{\frac{j_{i}}{\bar{m}+1}}\left\|\xi\right\|_{\infty}^{1-\frac{j_{i}}{\bar{m}+1}},&\,\,\,\,\text{ if }i\in\left\\{s+1,...,r\right\\}\text{ with }\theta_{i}=\frac{j_{i}}{\bar{m}+1}.\end{array}\right.$ Taking the product and then applying Young’s inequality (for the same $p,q$) $\displaystyle\left\|D^{j_{1}}\xi\right\|_{p_{1}}...\left\|D^{j_{q}}\xi\right\|_{p_{r}}$ $\displaystyle\leq\left\|D^{\bar{m}}\xi\right\|_{\frac{ap}{\bar{m}}}^{\frac{a}{\bar{m}}}\left\|D^{\bar{m}+1}\xi\right\|_{\frac{bq}{\bar{m}+1}}^{\frac{b}{\bar{m}+1}}\left\|\xi\right\|_{\infty}^{r-\frac{a}{m}-\frac{b}{m+1}}$ $\displaystyle\leq C(\varepsilon,p,q,r,\left\|\xi\right\|_{\infty})\left\|D^{\bar{m}}\xi\right\|_{\frac{ap}{\bar{m}}}^{\frac{ap}{\bar{m}}}+\varepsilon\left\|D^{\bar{m}+1}\xi\right\|_{\frac{bq}{\bar{m}+1}}^{\frac{bq}{\bar{m}+1}}$ (4.12) $\displaystyle=C(\varepsilon,p,q,r,\left\|\xi\right\|_{\infty})\left\|D^{\bar{m}}\xi\right\|_{2\frac{a(\bar{m}+1)}{\bar{m}(a+1)}}^{2\frac{a(\bar{m}+1)}{\bar{m}(a+1)}}+\varepsilon\left\|D^{\bar{m}+1}\xi\right\|_{2}^{2}$ where in the last line we have made the choices $q=\frac{2(\bar{m}+1)}{b},\text{ \ }p=\frac{2(\bar{m}+1)}{a+1}.$ Since $1\leq a\leq\bar{m}$ we have $\frac{a(\bar{m}+1)}{\bar{m}(a+1)}\leq 1$ and can use Hölder’s and Young’s inequalities to get (4.13) $C(\varepsilon,p,q)\left\|D^{\bar{m}}\xi\right\|_{2\frac{a(\bar{m}+1)}{\bar{m}(a+1)}}^{2\frac{a(\bar{m}+1)}{\bar{m}(a+1)}}\leq C(a,\bar{m})\left(\int\left|D^{\bar{m}}\xi\right|^{2}+\int\chi_{\text{supp}\left(f\right)}\right)$ omitting the last term in the case $a=\bar{m}$. Chaining together (4.11, 4.12, 4.13) gives the result. ∎ ###### Lemma 4.9. Let $f\in C^{\infty}(B_{4})$ and $r_{1}<r_{2}\leq 4$. 1. (1) If $j_{1}+...+j_{s}=m,$ then $\int_{B_{r_{1}}}\left|D^{j_{i}}f\cdots D^{j_{s}}f\right|^{2}\leq C(m,r\,_{1},r_{2})\,\|f\|_{\infty}^{2s-2}\sum_{j=0}^{m}\int_{B_{r_{2}}}\left|D^{j}f\right|^{2}.$ 2. (2) If $j_{1}+j_{2}+j_{3}+...+j_{s}+j^{\ast}=2\tilde{m}$ where $j_{q}<\tilde{m}$ and $j^{\ast}\geq 0,$ then $\int_{B_{r_{1}}}\left|D^{j_{i}}f\cdots D^{j_{s}}f\right|\leq C(\tilde{m},r_{1},r_{2})\,\left\|f\right\|_{\infty}^{s-\left(2-j^{\ast}/2\tilde{m}\right)}\left(\frac{2\tilde{m}-j^{\ast}}{2\tilde{m}}\sum_{j=0}^{\tilde{m}}\int_{B_{r_{2}}}\left|D^{j}f\right|^{2}+\frac{j^{\ast}}{2\tilde{m}}\left\|\chi_{\text{supp}\left(f\right)}\right\|_{p^{\ast}}^{2\tilde{m}/j^{\ast}}\right).$ 3. (3) If $j_{1}+...+j_{r}=2\bar{m}+1$ and all $j_{i}\leq\bar{m},$ then for $\varepsilon>0$ $\int_{B_{r_{1}}}\left|D^{j_{1}}f\ast D^{j_{2}}f...\ast D^{j_{r}}f\right|\leq\varepsilon\int_{B_{r_{2}}}\left|D^{\bar{m}+1}\xi\right|^{2}+C(\varepsilon,r_{1},r_{2},\bar{m},\|f\|_{\infty})\,\left(\sum_{j=0}^{\tilde{m}}\int_{B_{r_{2}}}\left|D^{j}\xi\right|^{2}+1\right).$ ###### Proof. Set $\tilde{\eta}\in C_{0}^{\infty}(B_{3})$ that is 1 on $B_{r_{1}}$, 0 on $B_{3}(0)\backslash B_{r_{2}}(0)$, $0\leq\tilde{\eta}\leq 1$ and $\|\tilde{\eta}\|_{C^{m}}\leq C(m)$. By Lemma 4.8 (second line below) $\displaystyle\int_{B_{r_{1}}}\left|D^{i_{1}}f\cdots D^{i_{s}}f\right|^{2}$ $\displaystyle\leq\int_{B_{r_{2}}}\left|D^{i_{1}}(\tilde{\eta}f)\cdots D^{i_{s}}(\tilde{\eta}f)\right|^{2}$ $\displaystyle\leq\|\tilde{\eta}f\|_{\infty}^{2s-2}\int_{B_{r_{2}}}\left|D^{m}(\tilde{\eta}f)\right|^{2}$ $\displaystyle\leq C(m,r_{1},r_{2},\|\tilde{\eta}\|_{C^{m}})\,\|f\|_{\infty}^{2s-2}\sum_{j=0}^{m}\int_{B_{r_{2}}}\left|D^{j}f\right|^{2}.$ The second and third inequalities in the statement of the Lemma follows by applying the previous Lemma in a similar way. ∎ The following is simple but will be used repeatedly, so we explicitly note it. ###### Lemma 4.10. Suppose that $r_{1}<r_{2}.$ Then for $\tilde{\varepsilon}>0$ $\int_{B_{r_{1}}}\left|D^{k+3}\varphi\right|^{2}\leq\tilde{\varepsilon}\int_{B_{r_{2}}}\left|D^{k+4}\varphi\right|^{2}+C(\tilde{\varepsilon},r_{1},r_{2})\int_{B_{r_{2}}}\left|D^{k+2}\varphi\right|^{2}.$ ###### Proof. For some $\tilde{\eta}=1$ on $B_{r_{1}}$ supported inside $B_{r_{2}}$ $\displaystyle\int_{B_{r_{2}}}\left|D^{k+3}\varphi\right|^{2}\tilde{\eta}^{2}$ $\displaystyle=-\int_{B_{r_{2}}}\ D^{k+2}\varphi\ast\left(\tilde{\eta}^{2}D^{k+4}\varphi+2\tilde{\eta}D\tilde{\eta}D^{k+3}\varphi\right)$ $\displaystyle\leq\tilde{\varepsilon}\int_{B_{r_{2}}}\tilde{\eta}^{2}\left|D^{k+4}\varphi\right|^{2}+\frac{1}{\tilde{\varepsilon}}C\int_{B_{r_{2}}}\tilde{\eta}^{2}\left|D^{k+2}\varphi\right|^{2}$ $\displaystyle+\frac{1}{2}\int_{B_{r_{2}}}\tilde{\eta}^{2}\left|D^{k+3}\varphi\right|^{2}+C\int_{B_{r_{2}}}\left|D\tilde{\eta}\right|^{2}\left|D^{k+2}\varphi\right|^{2}.$ Thus $\int_{B_{r_{1}}}\left|D^{k+3}\varphi\right|^{2}\leq 2\tilde{\varepsilon}\int_{B_{r_{2}}}\tilde{\eta}^{2}\left|D^{k+4}\varphi\right|^{2}+\frac{2}{\tilde{\varepsilon}}C\int_{B_{r_{2}}}\left(\tilde{\eta}^{2}+\left|D\tilde{\eta}\right|^{2}\right)\left|D^{k+2}\varphi\right|^{2}.$ ∎ #### 4.3.2. Evolution inequalities ###### Proposition 4.11. Let $\rho_{\alpha}^{2}\in C_{0}^{\infty}(B_{2}(0))$ defined by (4.6). Working in Darboux charts, for $\varepsilon>0$ we have (4.14) $\displaystyle\int_{B_{2}}\frac{d}{dt}\left(\left|D^{k+2}\varphi\right|_{g}^{2}dV_{g}\right)\rho_{\alpha}^{2}$ $\displaystyle\leq-2\int_{B_{2}}\left|D^{k+4}\varphi\right|_{g}^{2}\rho_{\alpha}^{2}dV_{g}+\varepsilon\int_{B_{3}}|D^{k+4}\varphi|^{2}dV_{g}$ $\displaystyle+C(k,\varepsilon,\left\|\varphi\right\|_{C^{3}})\left(\sum_{m=3}^{k+2}\int_{B_{3}}\left|D^{m}\varphi\right|_{g}^{2}dV_{g}+1\right).$ ###### Proof. In a Darboux chart, express $dV_{g}=V_{g}dx$. We have $\displaystyle\frac{d}{dt}\left(\left|D^{k+2}\varphi\right|_{g}^{2}V_{g}\right)=\,$ $\displaystyle 2\left(\partial_{t}\varphi_{i_{1}...i_{k+2}}\right)\left(\varphi_{j_{1}...j_{k+2}}g^{i_{1}j_{1}}g^{i_{2}j_{2}}...g^{i_{k+2}j_{k+2}}V_{g}\right)$ $\displaystyle+\varphi_{i_{1}...i_{k+2}}\varphi_{j_{1}...j_{k+2}}\partial_{t}\left(g^{i_{1}j_{1}}g^{i_{2}j_{2}}...g^{i_{k+2}j_{k+2}}V_{g}\right)$ (4.15) $\displaystyle=\,$ $\displaystyle-2(g^{kl}g^{pq}\varphi_{klpq}+G)_{i_{1}...i_{k+2}}\left(\varphi_{j_{1}...j_{k+2}}g^{i_{1}j_{1}}...g^{i_{k+2}j_{k+2}}V_{g}\right)$ $\displaystyle+\varphi_{i_{1}...i_{k+2}}\varphi_{j_{1}...j_{k+2}}\partial_{t}\left(g^{i_{1}j_{1}}g^{i_{2}j_{2}}...g^{i_{k+2}j_{k+2}}V_{g}\right).$ We count the highest order of derivatives of $\varphi$ in $x^{1},...,x^{n}$ for each term below: 1. (1) $g,g^{-1},V_{g}$ are of 2nd order 2. (2) $\partial_{t}g,\partial_{t}g^{-1}$ and $\partial_{t}V_{g}=V_{g}g^{ij}\partial_{t}g_{ij}$ are of 6th order 3. (3) $(g^{kl}g^{pq}\varphi_{klpq})_{i_{1}...i_{k+2}}$ is of $(k+6)$th order and $G_{i_{1}...i_{k+2}}$ is of $(k+5)$th order. For the sake of notation, we will use 1. (1) $P=P(x,D\varphi,D^{2}\varphi,D^{3}\varphi)$, 2. (2) $Q=P(x,D\varphi,...,D^{k+3}\varphi)\ $to be described in (4.19) and below. 3. (3) Bounded second order quantities are absorbed and not explicitly stated unless necessary (in particular, $dV_{g}$ will be dropped when not being differentiated.) Multiplying $\rho_{\alpha}^{2}$ to localize in a chart then integrate on $L$, we may then perform: (A) Integration by parts twice the first term in (4.15) leads to (4.16) $\int_{B_{2}}(-2g^{kl}g^{pq}\varphi_{klpq})_{i_{1}...i_{k+2}}\left(\varphi_{j_{1}...j_{k+2}}g^{i_{1}j_{1}}...g^{i_{k+2}j_{k+2}}V_{g}\right)\rho_{\alpha}^{2}=-2\int_{B_{2}}\left|D^{k+4}\varphi\right|_{g}^{2}\rho_{\alpha}^{2}\,dV_{g}+I$ where $I=\int_{B_{2}}D^{k+4}\varphi\ast\left(D^{4}\varphi+P^{2})\ast D^{k+2}\varphi\rho_{\alpha}^{2}+D^{k+3}\varphi\ast\left(P\ast\rho_{\alpha}^{2}+D\rho_{\alpha}^{2}\right)+D^{k+2}\varphi\ast\left(P\ast D\rho_{\alpha}^{2}+D^{2}\rho_{\alpha}^{2}\right)\right).$ To deal with the first term in $I$: (4.17) $\int_{B_{2}}D^{k+4}\varphi\ast\left(D^{4}\varphi+P^{2}\right)\ast D^{k+2}\varphi\leq\varepsilon\int_{B_{2}}\rho_{\alpha}^{2}|D^{k+4}\varphi|^{2}+C(\varepsilon)\int_{B_{2}}\rho_{\alpha}^{2}\left(\left|D^{k+2}\varphi\ast\left(D^{4}\varphi+P^{2}\right)\right|^{2}\right).$ Lemma 4.9 with $m=k$, $f=$ $D^{3}\varphi$ and $r_{2}=5/2$ yields $\int_{B_{2}}\left|D^{k+2}\varphi\ast D^{4}\varphi\right|^{2}\leq C(D^{3}\varphi)\sum_{j=0}^{m}\int_{B_{5/2}}\left|D^{j+3}\varphi\right|^{2}.$ Applying Lemma (4.10) to the highest order term provides a bound of (4.17) by the positive terms in (4.14) noting also that $\int_{B_{2}}\left|D^{k+2}\varphi\ast P^{2}\right|^{2}\leq C(D^{3}\varphi)\int_{B_{2}}\left|D^{k+2}\varphi\right|^{2}.$ Next (4.18) $\int_{B_{2}}\left|(P\ast\rho_{\alpha}^{2}+D\rho_{\alpha}^{2})D^{k+4}\varphi\ast D^{k+3}\varphi\right|\leq\varepsilon\int_{B_{2}}\rho_{\alpha}^{2}|D^{k+4}\varphi|^{2}+\frac{1}{\varepsilon}C(P)\int_{B_{2}}\left|D^{k+3}\varphi\right|^{2}$ recalling that $D\rho_{\alpha}$ is bounded by uniform constants and $D^{2}\varphi.$ By Lemma 4.10 (choosing $\tilde{\varepsilon}\approx c\varepsilon^{2}$), (4.18) is bounded by $\varepsilon\int_{B_{2}}\rho_{\alpha}^{2}|D^{k+4}\varphi|^{2}+\varepsilon\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}+\frac{1}{\varepsilon^{3}}C\int_{B_{3}}\left|D^{k+2}\varphi\right|^{2}$ which is of the correct form. Finally for $I$, using (4.7), $\int_{B_{2}}\left|D^{k+4}\varphi\ast D^{k+2}\varphi\ast\left(P\ast D\rho_{\alpha}^{2}+D^{2}\rho_{\alpha}^{2}\right)\right|\leq\varepsilon\int_{B_{2}}\rho_{\alpha}^{2}|D^{k+4}\varphi|^{2}+C(\varepsilon,P)\int_{B_{2}}\left|D^{k+2}\varphi\right|^{2}.$ (B) Note that when applying the product rule successively to $G$, we will get 1. (1) A single highest order term which is linear in the highest order with coefficients involving at most order $D^{3}\varphi.$ 2. (2) Second to highest order terms that are linear in the second highest order, may have a factor of $D^{4}\varphi$, all other dependence of lower order. 3. (3) Terms of lower order, which could be multilinearly dependent on various lower orders. Thus (4.19) $G_{i_{1}...i_{k+1}}=1\ast D^{k+4}\varphi+Q.$ with $Q$ having highest order $D^{k+3}\varphi.$ This is observed by iterating the following expansion: Using $DG$ to denote a derivative in $x$ of the composition $x\mapsto G(x,D\varphi(x),D^{2}\varphi(x),D^{3}\varphi(x))$ and $\bar{D}G$ to denote derivatives in all 4 arguments of $G,$ we have $DG=\bar{D}G\ast\left(D^{4}\varphi+D^{3}\varphi+D^{2}\varphi+\phi\right)$ where $\phi$ is the term generated by $\bar{D}G/Dx.$ Continuing $\displaystyle D^{2}G$ $\displaystyle=\bar{D}G\ast\left(D^{5}\varphi+D^{4}\varphi+D^{3}\varphi+D\phi\ast\left(D^{4}\varphi+D^{3}\varphi+D^{2}\varphi\right)\right)$ $\displaystyle+\bar{D}^{2}G\ast\left(D^{4}\varphi+D^{3}\varphi+D^{2}\varphi+\phi\right)\ast\left(D^{4}\varphi+D^{3}\varphi+D^{2}\varphi+\phi\right)$ $\displaystyle...$ (4.20) $\displaystyle D^{k+1}G$ $\displaystyle=\bar{D}G\ast\left(D^{k+4}\varphi+D^{k+3}\varphi+D^{k+2}\varphi+...\right)$ $\displaystyle+\bar{D}^{2}G\ast\left(D^{k+3}\varphi+D^{k+2}\varphi+D^{k+1}\varphi+...\right)\ast\left(D^{4}\varphi+D^{3}\varphi+D^{2}\varphi+\phi\right)$ $\displaystyle+\bar{D}^{3}G\ast\left(D^{k+2}\varphi+...\right)\ast\left\\{\left(D^{5}\varphi+...\right)+\left(D^{4}\varphi+...\right)\ast\left(D^{4}\varphi+...\right)\right\\}$ $\displaystyle...$ Now integrate by parts: $\displaystyle\int_{B_{2}}$ $\displaystyle G_{i_{1}...i_{k+2}}\left(\varphi_{j_{1}...j_{k+2}}g^{i_{1}j_{1}}...g^{i_{k+2}j_{k+2}}V_{g}\right)\rho_{\alpha}^{2}=-\int_{B_{2}}G_{i_{1}...i_{k+1}}\partial_{{i_{k+2}}}\left[\rho_{\alpha}^{2}\left(\varphi_{j_{1}...j_{k+2}}g^{i_{1}j_{1}}...g^{i_{k+2}j_{k+2}}V_{g}\right)\right]$ $\displaystyle=$ $\displaystyle\int_{B_{2}}\left(D^{k+4}\varphi\ast D^{k+3}\varphi\right)\rho_{\alpha}^{2}+\int_{B_{2}}(D\rho_{\alpha}^{2}+\rho_{\alpha}^{2}P)D^{k+4}\varphi\ast D^{k+2}\varphi$ $\displaystyle+\int_{B_{2}}\left(Q\ast D^{k+3}\varphi\right)\rho_{\alpha}^{2}+\int_{B_{2}}(D\rho_{\alpha}^{2}+\rho_{\alpha}^{2}P)Q\ast D^{k+2}\varphi.$ We use Peter-Paul’s inequality we split into two types of terms: $\varepsilon\int_{B_{2}}\left|D^{k+4}\varphi\right|^{2}\,\rho_{\alpha}^{2}+\varepsilon\int_{B_{2}}Q^{2}\rho_{\alpha}^{2}$ and $C(\varepsilon,D\rho_{\alpha})\int_{B_{2}}\left(\left|D^{k+3}\varphi\right|^{2}+\left|D^{k+2}\varphi\right|^{2}\right)(P^{2}+P+1)\rho_{\alpha}^{2}.$ First, $\int_{B_{2}}\left|D^{k+3}\varphi\right|^{2}\left(P^{2}+P+1\right)\rho_{\alpha}^{2}\leq C\left(D^{3}\varphi\right)\int_{B_{2}}\left|D^{k+3}\varphi\right|^{2}$ is bounded by the argument in Lemma (4.10). One can prove by induction, observing (4.19) and (4.20), that for each term in $Q,$ the total number of derivatives of $D^{3}\varphi$ that arise will sum up to no more than $k+1$ (i.e. $D^{k-1}\varphi\ast D^{6}\varphi\ast D^{5}\varphi=D^{3+k-4}\varphi\ast D^{3+3}\varphi\ast D^{3+2}\varphi$, here $k-4+3+2=k+1.$) Applying Lemma 4.9 for $f=D^{3}\varphi$ and $m=k+1$ to each of the squared terms gives $\int_{B_{2}}Q^{2}dV_{g}\leq C\left(\|D^{3}\varphi\|\right)\sum_{i=0}^{k+1}\int_{B_{3}}|D^{i+3}\varphi|^{2}$ thus $\varepsilon\int_{B_{2}}Q^{2}dV_{g}$ has the correct bound, by Lemma 4.10. Finally we finish bounding the last term in (4.15) $\int\varphi_{i_{1}...i_{k+2}}\varphi_{j_{1}...j_{k+2}}\partial_{t}\left(g^{i_{1}j_{1}}g^{i_{2}j_{2}}...g^{i_{k+2}j_{k+2}}V_{g}\right)\rho_{\alpha}^{2}=\int\left(D^{k+2}\varphi\ast D^{k+2}\varphi\ast D^{6}\varphi\right)\rho_{\alpha}^{2}.$ Apply Lemma 4.9 for $f=D^{3}\varphi$ and $\tilde{m}=k$ $\int_{B_{2}}\left(D^{k+2}\varphi\ast D^{k+2}\varphi\ast D^{6}\varphi\right)\rho_{\alpha}^{2}\leq\varepsilon\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}+C\left(\varepsilon,k,D^{3}\varphi\right)\sum_{j=3}^{k+3}\int_{B_{3}}\left|D^{j}\varphi\right|^{2}$ We may then sweep away the $\int_{B_{2}}\left|D^{k+3}\varphi\right|^{2}$ term Lemma 4.10 (choosing $\tilde{\varepsilon}\approx c\varepsilon^{2}$) to conclude the proof. ∎ ###### Proposition 4.12. Let $\rho_{\alpha}^{2}\in C_{0}^{\infty}(B_{2}(0))$. Considering the decomposition in Lemma 4.7, for $\varepsilon>0$ we have $\displaystyle\int_{B_{2}}\frac{d}{dt}\left(\left(\left|S_{k}\right|_{g}^{2}+2\langle D^{k+2}\varphi,S_{k}\rangle_{g}\right)dV_{g}\right)\rho_{\alpha}^{2}$ $\displaystyle\leq C(k,\varepsilon,D^{3}\varphi)\,\left(\sum_{j=3}^{k+2}\int_{B_{3}}\left|D^{j}\varphi\right|^{2}+1\right)$ $\displaystyle+\varepsilon\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}.$ ###### Proof. Recall that (4.21) $S_{k}=\left(1+D^{3}\varphi\right)\ast(D^{k+1}\varphi+D^{k}\varphi\ast D^{4}\varphi+D^{k-1}\varphi\ast D^{4}\varphi\ast D^{4}\varphi+D^{k-1}\varphi\ast D^{5}\varphi+...)$ Differentiating with respect to $t$ generates product rule expansions with 4 orders of derivatives added to a factor in each term, that is (modulo lower order geometrically controlled values like $V_{g})$ (4.22) $\displaystyle\frac{d}{dt}\left|S_{k}\right|_{g}^{2}dV_{g}$ $\displaystyle=\left(1+D^{3}\varphi\right)\ast\left(D^{k+5}\varphi+D^{k+4}\varphi\ast D^{4}\varphi+D^{k}\varphi\ast D^{8}\varphi+...\right)\ast S_{k}$ $\displaystyle\left(D^{6}\varphi+D^{7}\varphi\right)\ast\left(D^{k+1}\varphi+D^{k}\varphi\ast D^{4}\varphi+D^{k-1}\varphi\ast D^{4}\varphi\ast D^{4}\varphi+...\right)\ast S_{k}.$ Integrate by parts: $\displaystyle\int_{B_{2}}$ $\displaystyle\left(D^{k+5}\varphi\ast\left(1+D^{3}\varphi\right)\ast S_{k}\right)\rho_{\alpha}^{2}=\int_{B_{2}}D^{k+4}\varphi\ast\left(D^{4}\varphi\ast S_{k}+\left(1+D^{3}\varphi\right)\ast DS_{k}\right)\rho_{\alpha}^{2}$ $\displaystyle+\int_{B_{2}}\left(D^{k+4}\varphi\ast\left(1+D^{3}\varphi\right)\ast S_{k}\right)\ast D\rho_{\alpha}^{2}$ $\displaystyle\leq\varepsilon\int_{B_{2}}\left|D^{k+4}\varphi\right|^{2}\rho_{\alpha}^{2}+C(\varepsilon)\int_{B_{2}}\left(\left|D^{4}\varphi\ast S_{k}\right|^{2}+\left|D^{3}\varphi\ast DS_{k}\right|^{2}\right)\rho_{\alpha}^{2}$ $\displaystyle+C(\varepsilon)\int_{B_{2}}\left|\left(1+D^{3}\varphi\right)\ast S_{k}\right|^{2}\left|D\rho_{\alpha}\right|^{2}.$ Now apply Lemma 4.9 with $f=D^{3}\varphi$ and $m=k-1$ (4.23) $\int_{B_{2}}\left(\left|D^{4}\varphi\ast S_{k}\right|^{2}+\left|D^{3}\varphi\ast DS_{k}\right|^{2}\right)\rho_{\alpha}^{2}\leq C(k,D^{3}\varphi)\,\sum_{j=3}^{k+2}\int_{B_{3}}\left|D^{j}\varphi\right|^{2}.$ Similarly using $\tilde{\eta}$ as in the proof of Lemma 4.9, we have $\int_{B_{2}}\left|\left(1+D^{3}\varphi\right)\ast S_{k}\right|^{2}\left|D\rho_{\alpha}\right|^{2}\leq C\left(D\rho_{\alpha},D^{3}\varphi\right)\sum_{j=3}^{k+1}\int_{B_{3}}\left|D^{j}\varphi\right|^{2}.$ Continuing with the terms in (4.22) $\int_{B_{2}}\left(1+D^{3}\varphi\right)\ast\left(D^{k+4}\varphi\ast D^{4}\varphi\ast S_{k}\right)\rho_{\alpha}^{2}\leq\varepsilon\int_{B_{2}}\left|D^{k+4}\varphi\right|^{2}\rho_{\alpha}^{2}+C(\varepsilon)\int_{B_{2}}\left|D^{3}\varphi\ast D^{4}\varphi\ast S_{k}\right|^{2}\rho_{\alpha}^{2}$ with the latter term enjoying the same bound as (4.23). The remaining terms are of the form $\int\left(D^{3+j_{1}}\varphi\ast D^{3+j_{2}}\varphi\ast..\ast D^{3+j_{q}}\varphi\right)\rho_{\alpha}^{2}$ with $j_{1}+...+j_{q}\leq 2k$ so $\int_{B_{2}}\left(D^{3+j_{1}}\varphi\ast D^{3+j_{2}}\varphi\ast..\ast D^{3+j_{q}}\varphi\right)\rho_{\alpha}^{2}\leq C(m,D^{3}\varphi)\,\left(\sum_{j=3}^{k+3}\int_{B_{3}}\left|D^{j}\varphi\right|^{2}+1\right)$ by Lemma 4.9 again. Applying Lemma 4.10 to $\int_{B_{3}}\left|D^{k+3}\varphi\right|^{2}$ completes the desired bound for the integral of the (4.22) terms. Next (4.24) $\int_{B_{2}}\frac{d}{dt}\left(2\langle D^{k+2}\varphi,S_{k}\rangle_{g}dV_{g}\right)\rho_{\alpha}^{2}dV_{g}=\int_{B_{2}}\left(D^{k+6}\varphi\ast S_{k}\right)\rho_{\alpha}^{2}dV_{g}+\int_{B_{2}}D^{k+2}\varphi\ast\frac{d}{dt}\left(S_{k}\ast V\right)\rho_{\alpha}^{2}.$ Integrating the first term by parts twice yields $\displaystyle\int_{B_{2}}$ $\displaystyle D^{k+6}\varphi\ast S_{k}\rho_{\alpha}^{2}=\int_{B_{2}}D^{k+4}\varphi\ast D^{2}\left(S_{k}\ast V\right)\rho_{\alpha}^{2}+D\rho_{\alpha}^{2}\ast D\left(S_{k}\ast V\right)+D^{2}\rho_{\alpha}^{2}\ast\left(S_{k}\ast V\right)$ $\displaystyle\leq\varepsilon\int_{B_{2}}\left|D^{k+4}\varphi\right|^{2}\rho_{\alpha}^{2}+C(\varepsilon)\int_{B_{2}}\left|D^{2}S_{k}\right|^{2}\rho_{\alpha}^{2}+C\left(\varepsilon,D^{2}\rho_{\alpha}^{2}\right)\int_{B_{2}}\left(\left|DS_{k}\right|^{2}+\left|S_{k}\right|^{2}\right).$ Again Lemma 4.9 with $f=D^{3}\varphi$ and $\tilde{m}=k$ and $r_{2}=5/2$ (4.25) $\displaystyle\int_{B_{2}}\left|D^{2}S_{k}\right|^{2}\rho_{\alpha}^{2}$ $\displaystyle\leq C(k,D^{3}\varphi)\,\sum_{j=3}^{k+3}\int_{B_{5/2}}\left|D^{j}\varphi\right|^{2}.$ (4.26) $\displaystyle\leq\varepsilon\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}+C\left(\varepsilon,k,D^{3}\varphi\right)\sum_{j=3}^{k+2}\int_{B_{3}}\left|D^{j}\varphi\right|^{2}$ using Lemma 4.10. Now look at second term in (4.24). Note that $D^{k+2}\varphi\ast\frac{d}{dt}\left(S_{k}\ast V\right)=D^{k+2}\varphi\ast D^{k+5}\varphi\ast D^{3}\varphi+D^{k+2}\varphi\ast\left(D^{k+4}\varphi\ast D^{4}\varphi+...\right).$ The highest order term can be dealt with via integration by parts away from $D^{k+5}\varphi$ and then an iterated Peter-Paul, carefully choosing smaller $\varepsilon$ and using Lemma 4.9. For the remaining terms, we need the third statement in Lemma 4.9 which gives $\displaystyle\int_{B_{2}}\left(D^{3+j_{1}}\varphi\ast D^{3+j_{2}}\varphi\ast..\ast D^{3+j_{q}}\varphi\right)\rho_{\alpha}^{2}$ $\displaystyle\leq\varepsilon\int_{B_{5/2}}\left|D^{k+4}\varphi\right|^{2}$ $\displaystyle+C(\varepsilon,k,\frac{5}{2},\|f\|_{\infty})\,\left(\sum_{j=0}^{k}\int_{B_{5/2}}\left|D^{3+j}\varphi\right|^{2}+1\right)$ as $j_{1}+...+j_{q}=2k+1$. A final application of Lemma 4.10 to $\int_{B_{5/2}}\left|D^{k+3}\varphi\right|^{2}$ completes the proof. ∎ In Proposition 4.11 we have isolated ‘good’ terms $-2\int_{B_{2}}\left|D^{k+4}\varphi\right|^{2}\rho_{\alpha}^{2}.$ We would like to use them to offset the ‘bad’ terms of the form $\varepsilon\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}dV_{g}$ that occur in Propositions 4.11 and 4.12. Because the expressions for $D^{k+4}\varphi$ are different in each chart in the cover, the difficulty arises that we cannot directly beat the terms occurring on a larger ball by terms on a smaller ball of different charts, even when the smaller balls cover the larger ball. To make an argument that the bad terms in a larger ball of one chart are offset by the good terms in a smaller ball in a different chart requires bounding the bad terms by a global, well-defined geometric quantity involving derivatives of the second fundamental form, modulo a lower order difference. This is the point of the following lemma. ###### Lemma 4.13. Take a finite cover of charts $\Upsilon^{\alpha}$, each over $B_{4}(0)$ and partition of unity $\rho_{\alpha}^{2}$ in $B_{2}(0)$ in each respective chart as described by (4.6). Then $\sum_{\alpha}\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}dV_{g}\leq 2N\sum_{m=0}^{k+1}\int_{L}\left|\nabla^{k+1-m}A\right|^{2}dV_{g}+C.$ ###### Proof. Note that from Lemma 4.7 $\left|D^{k+4}\varphi\right|_{g}^{2}\leq 2\left|\nabla^{k+1}A\right|_{g}^{2}+2\left|S_{k+2}\right|_{g}^{2}.$ Let $\iota=\frac{1}{k+2}.$ Then we have, taking $\tilde{\eta}=1$ on each $B_{3}(0),$ with $\tilde{\eta}\in C_{c}^{\infty}(B_{3+\iota}(0))$ $\displaystyle\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}dV_{g}$ $\displaystyle\leq 2\int_{B_{3}}\left(\left|\nabla^{k+1}A\right|^{2}+\left|S_{k+2}\right|_{g}^{2}\right)dV_{g}$ $\displaystyle\leq 2\int_{B_{3}}\left|\nabla^{k+1}A\right|^{2}dV_{g}+2\int_{B_{3+\iota}}\left|S_{k+2}\right|_{g}^{2}\tilde{\eta}^{2}dV_{g}$ $\displaystyle\leq 2\int_{B_{3}}\left|\nabla^{k+1}A\right|^{2}dV_{g}+C\left(\sum_{m=3}^{k+3}\int_{B_{3+\iota}}\left|D^{m}\varphi\right|^{2}dV_{g}+1\right)$ by Lemma 4.9. Iterating this argument, using $\int_{B_{3+\iota}}\left|D^{k+3}\varphi\right|^{2}dV_{g}\leq 2\int_{B_{3+\iota}}\left|\nabla^{k}A\right|^{2}dV_{g}+C\left(\sum_{m=3}^{k+2}\int_{B_{3+2\iota}}\left|D^{m}\varphi\right|^{2}dV_{g}+1\right)$ and so forth, for a total of $k+1$ steps, we have by using $\left|D^{3}\varphi\right|\leq$ $\left|A\right|+C$ that $\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}dV_{g}\leq 2\sum_{m=0}^{k+1}\int_{B_{3+\frac{k+1}{k+2}}}\left|\nabla^{k+1-m}A\right|^{2}\tilde{\eta}^{2}dV_{g}+C.$ Now for any set of functions $\tilde{\eta}_{\alpha}$ who are $1$ on $B_{r}\subset B_{4}$ on each chart $\Upsilon^{\alpha}$, we can bound $\displaystyle\sum_{\alpha}\int_{B_{4}}\left|\nabla^{m}A\right|^{2}\tilde{\eta}_{\alpha}^{2}dV_{g}$ $\displaystyle\leq\max_{x\in L}\left(\sum_{\alpha}\tilde{\eta}_{\alpha}^{2}(x)\right)\int_{L}\left|\nabla^{m}A\right|^{2}dV_{g}\leq N\int_{L}\left|\nabla^{m}A\right|^{2}dV_{g}.$ It follows that $\sum_{\alpha}\int_{B_{3}}\left|D^{k+4}\varphi\right|^{2}dV_{g}\leq 2N\sum_{m=0}^{k+1}\int_{L}\left|\nabla^{k+1-m}A\right|^{2}dV_{g}+C.$ ∎ ### 4.4. Proof of the main theorem ###### Proof of Proposition 4.2. At a fixed time $t_{0}$ we may take the ambient charts $\left\\{\Upsilon^{\alpha}\right\\}$ for a tubular neighborhood of $L$ and subordinate partition of unity $\left\\{\rho_{\alpha}^{2}\right\\}$ which restrict to charts (via the $x$ coordinate) for $L$ with the same partition of unity. Differentiate $\displaystyle\frac{d}{dt}\int_{L}\left|\nabla^{k-1}A\right|_{g}^{2}dV_{g}$ $\displaystyle=\int_{L}\frac{d}{dt}\left(\left|\nabla^{k-1}A\right|_{g}^{2}dV_{g}\right)$ $\displaystyle=\int_{B_{2}}\left(\sum_{\alpha}\rho_{\alpha}^{2}\right)\frac{d}{dt}\left[\left(\left|D^{k+2}\varphi\right|_{g}^{2}+\left|S_{k}\right|_{g}^{2}+2\langle D^{k+2}\varphi,S_{k}\rangle_{g}\right)dV_{g}\right]$ $\displaystyle=\sum_{\alpha}\int_{B_{2}}\frac{d}{dt}\left(\left|D^{k+2}\varphi\right|_{g}^{2}dV_{g}\right)\rho_{\alpha}^{2}$ $\displaystyle+\sum_{\alpha}\int_{B_{2}}\frac{d}{dt}\left[\left(\left|S_{k}\right|_{g}^{2}+2\langle D^{k+2}\varphi,S_{k}\rangle_{g}\right)dV_{g}\right]\rho_{\alpha}^{2}.$ Thus (4.27) $\displaystyle\frac{d}{dt}\int_{L}\left|\nabla^{k-1}A\right|_{g}^{2}dV_{g}$ $\displaystyle\leq-2\sum_{\alpha}\int_{B_{2}}\left|D^{k+4}\varphi\right|_{g}^{2}\rho_{\alpha}^{2}dV_{g}+\varepsilon\sum_{\alpha}\int_{B_{3}}|D^{k+4}\varphi|^{2}dV_{g}$ $\displaystyle+\sum_{\alpha}C(k,\varepsilon,\left\|\varphi\right\|_{C^{3}})\left(\sum_{m=3}^{k+2}\int_{B_{3}}\left|D^{m}\varphi\right|_{g}^{2}dV_{g}+1\right)$ by Propositions 4.11 and 4.12. Now apply Lemma 4.13 $\displaystyle\sum_{\alpha}\int_{B_{3}}|D^{k+4}\varphi|^{2}dV_{g}$ $\displaystyle\leq\left(NC\sum_{m=0}^{k+1}\int_{L}|\nabla^{m}A|^{2}dV_{g}+C\right)$ $\displaystyle=NC\sum_{m=0}^{k+1}\int_{L}|\nabla^{m}A|^{2}\left(\sum_{\alpha}\rho_{\alpha}^{2}\right)dV_{g}+C$ $\displaystyle=NC\sum_{m=0}^{k+1}\sum_{\alpha}\int_{B_{2}}|\nabla^{m}A|^{2}\rho_{\alpha}^{2}dV_{g}$ $\displaystyle\leq NC\sum_{m=0}^{k+1}\sum_{\alpha}\int_{B_{2}}2\left(\left|D^{m+3}\varphi\right|^{2}+\left|S_{m+1}\right|_{g}^{2}\right)\rho_{\alpha}^{2}dV_{g}$ $\displaystyle=2NC\sum_{\alpha}\int_{B_{2}}\left(\left|D^{k+4}\varphi\right|^{2}+\left|S_{k+2}\right|_{g}^{2}\right)\rho_{\alpha}^{2}dV_{g}$ $\displaystyle+2NC\sum_{m=0}^{k}\sum_{\alpha}\int_{B_{2}}2\left(\left|D^{m+3}\varphi\right|^{2}+\left|S_{m+1}\right|_{g}^{2}\right)\rho_{\alpha}^{2}dV_{g}.$ Note that from Lemma 4.10 $2\int_{B_{2}}\left|D^{k+3}\varphi\right|^{2}\rho_{\alpha}^{2}dV_{g}\leq\frac{1}{4NC}\int_{B_{3}}|D^{k+4}\varphi|^{2}dV_{g}+C\left(N,\left|D\rho_{\alpha}^{2}\right|\right)\int_{B_{3}}|D^{k+2}\varphi|^{2}dV_{g}.$ Note also Lemma 4.9, recalling (4.21), then Lemma 4.10 on the highest order resulting term gives $\displaystyle 2\int_{B_{2}}\left|S_{k+2}\right|_{g}^{2}\rho_{\alpha}^{2}dV_{g}$ $\displaystyle\leq C\sum_{m=3}^{k+3}\int_{B_{5/2}}\left|D^{m}\varphi\right|_{g}^{2}+C$ $\displaystyle\leq\frac{1}{4NC}\int_{B_{3}}|D^{k+4}\varphi|^{2}dV_{g}+C\left(N,\left|D\rho_{\alpha}^{2}\right|\right)\sum_{m=3}^{k+2}\int_{B_{3}}\left|D^{m}\varphi\right|_{g}^{2}+C.$ Thus $\displaystyle\sum_{\alpha}\int_{B_{3}}|D^{k+4}\varphi|^{2}dV_{g}$ $\displaystyle\leq 4NC\sum_{\alpha}\int_{B_{2}}\left|D^{k+4}\varphi\right|^{2}\rho_{\alpha}^{2}dV_{g}+8NC\sum_{m=3}^{k+2}\sum_{\alpha}\int_{B_{2}}\left|D^{m}\varphi\right|^{2}\rho_{\alpha}^{2}dV_{g}$ (4.28) $\displaystyle+8NC\sum_{m=1}^{k+1}\sum_{\alpha}\int_{B_{2}}\left|S_{m}\right|_{g}^{2}\rho_{\alpha}^{2}dV_{g}.$ Choosing $\varepsilon<(2NC)^{-1}$ in (4.27) in light of (4.4) we have $\frac{d}{dt}\int_{L}\left|\nabla^{k-1}A\right|_{g}^{2}dV_{g}\leq C(N,k,\left\|\varphi\right\|_{C^{3}})\left(\sum_{m=3}^{k+2}\int_{B_{3}}\left|D^{m}\varphi\right|_{g}^{2}+\sum_{m=1}^{k+1}\int_{B_{3}}\left|S_{m}\right|_{g}^{2}+1\right).$ Applying Lemma 4.9 to the $\int\left|S_{m}\right|_{g}^{2}$ terms and then Lemma 4.13 to the $\left|D^{m}\varphi\right|_{g}^{2}$ terms yields the result. ∎ ###### Proof of Theorem 4.1. Suppose now that $F$ is a solution to (1.1) with $\left|A\right|\leq K$ on $[0,T)$. Starting with $\int_{L}\left|A\right|^{2}dV_{g}(t)\leq K\operatorname{Vol}\left(L\right)\leq C$ we may apply Proposition 4.2 and apply differential inequalities: continuing with $\frac{d}{dt}\int_{L}\left|\nabla A\right|^{2}dV_{g}(t)\leq C\int_{L}\left|\nabla A\right|^{2}dV_{g}(t)+C\int_{L}\left|A\right|^{2}dV_{g}(t)$ and so forth, obtaining bounds of the form (4.29) $\int_{L}\left|\nabla^{k-1}A\right|^{2}dV_{g}(t)\leq C(k,K,F_{0},T)$ for arbitrary $k$. Now at any $t_{0}\in[0,T)$ we may take a cover $\Upsilon^{\alpha}$ as described in Proposition 4.6. By Lemma 4.13 and (4.29) we have (4.30) $\left\|D^{k}\varphi\right\|_{L^{2}(B_{3})}\leq C(k,K,F_{0},T)$ for all $k,$ in every chart. By Sobolev embedding theorems, we have Hölder bounds on $D^{k}\varphi$ over $B_{2}$ for each chart. In particular, there will be uniform bounds on $\frac{d}{dt}D\varphi$ and $\frac{d}{dt}D^{2}\varphi$ which control the speed of the flow in the chart and the rate of change of the slope the manifold $L_{t}$ makes with respect to the tangent plane at the origin in the chart. We conclude then the manifolds $L_{t}$ will continue to be described by the set of charts taken at $t_{0}$ for $t<\max\left\\{T,t_{0}+\tau\right\\}$ for some positive $\tau$ with an apriori lower bound. (Perhaps we take $c_{n}$ slightly larger in (4.2)). By choosing $t_{0}$ near $T$ we are assured that these fixed charts describe the flow for all values $t\in[t_{0},T)$. Now observe that with fixed speed bounds, the paths $x\mapsto F(x,t)$ of the normal flow are Lipschitz and hence the normal flow extends to a well-defined continuous map (4.31) $F:L\times[0,T]\rightarrow M.$ We claim that $F\left(\cdot,T\right)$ is a smooth immersion. While within a chart, the vertical maps (4.32) $\bar{F}(x):=(x,d\varphi(x,t))$ converge in every Hölder norm to a smooth map at $T,$ we still must argue that the charts given by the $x$ coordinates do not collapse as $t\rightarrow T$. This can be argued locally, using coordinates on $L_{t_{0}}$. For any given $x\in L_{t_{0}}$ we may choose a chart such that $x\in B_{1}(0)\subset B_{3}(0).$ We are already assuming $F$ is an immersion at $t_{0}$ so this coordinate chart gives us a coordinate chart for the abstract smooth manifold $L.$ For $t>t_{0}$ the normal flow $F$ is given by (4.33) $F(x,t)=(\chi_{t}(x),d\varphi(\chi_{t}(x),t))$ for some local diffeomorphism $\chi_{t}(x):B_{1}(0)\rightarrow B_{2}(0)$ from Claim 3.3, provided that $t_{0}$ is chosen close enough to $T$ such that ${\chi}_{t}(x)\in B_{2}(0)\text{ for all }x\in B_{1}(0)\text{ and }t\in[t_{0},T).\text{ }$ This choice of $t_{0}$ is possible given that $\chi_{t}(x)$ is controlled by the normal projection of $\frac{d\bar{F}}{dt}$ and the inverse $(d\bar{F})^{-1}$, for $\bar{F}$ defined by (4.32), both of which are universally controlled given (4.2) and (4.30). Now because (4.32) is uniformly smooth, it can be extended smoothly to $[t_{0},T+\delta),$ as well as the normal flow associated to this extension. Applying Claim 3.3 (note that we may extend the flow outside $B_{3}$ in a nice way which doesn’t affect the behavior in $B_{2}(0))$ we get a smooth diffeomorphism $\chi_{T}.$ For $x\in B_{1}(0)$ we can compute the normal flow $F$, $F(x,T)=(\chi_{T}(x),d\varphi(\chi_{T}(x),T))$ which is a smooth extension of (4.33) to $T$, by the uniform estimates on $\varphi$. Now $F(x,T)$ is a smooth immersion from $B_{1}(0)$ because $\chi_{T}$ is a diffeomorphism. As $x$ was chosen arbitrarily, we conclude the continuous extension of $F$ defined in (4.31) must be a smooth immersion from $L$ at $T$. We may now restart the flow by Proposition 3.2 with initial immersion $F(x,T)$. The time derivatives of the new flow and $F$ agree to any order at $T$. Therefore the new flow is a smooth extension of $F$ to $[0,T+\varepsilon)$ for some $\varepsilon>0$. Moreover, Theorem 3.4 asserts that this is the only smooth extension. ∎ ## 5\. Appendix ### 5.1. Submanifold with bounded second fundamental form $A$ It is a known and frequently used fact that when $|A|$ is bounded then the submanifold can be written as a graph over a controlled region in its tangent space. We provide a proof below for any dimension and codimension. ###### Proposition 5.1. Let $L^{k}$ be a compact manifold embedded in a compact Riemannian manifold $(M^{k+l},g)$. Suppose that the second fundamental form of $L$ satisfies $|A|\leq K$ for some constant $K>0$. Then $L$ is locally a graph of a vector- valued function over a ball $B_{r}(0)\subset T_{p}L$ in a normal neighbourhood of $p\in L$ in $M$ and $r>C(M,g)(K+1)^{-1}$ for some constant $C(M,g)>0$. ###### Proof. Step 1. Bound the injectivity radius of $L$ from below in terms of $K$. Assume $M$ is isometrically embedded in some euclidean space. For the embedding $F:L^{k}\overset{f}{\to}M^{k+l}\overset{\varphi}{\to}\mathbb{R}^{k+n}$, denote its second fundamental form by $\tilde{A}$ and note that $|\tilde{A}|\leq C(|A|+1)\leq C(K+1)$ where $C$ only depends on the isometric embedding $\varphi$. Let $\gamma:\mathbb{S}^{1}\to L$ be a shortest geodesic loop based at a point $p\in L$ which is parametrized by arc-length $s$. Suppose $\gamma(0)=\gamma(a),\gamma^{\prime}(0)=\gamma^{\prime}(a)$. Take a hyperplane $P$ in $\mathbb{R}^{k+n}$ such that $P$ intersects $\gamma$ at a point $p$ orthogonally. There is a point $q\in\gamma$ where $\gamma$ meets $P$ again at first time. The angle between the unit vectors $\gamma^{\prime}(p)$ and $\gamma^{\prime}(q)$ in $\mathbb{R}^{k+n}$ is at least $\frac{\pi}{2}$. Therefore $\left|\gamma^{\prime}(p)-\gamma(^{\prime}q)\right|\geq\sqrt{2}.$ Since $F\circ\gamma:\mathbb{S}^{1}\overset{\gamma}{\to}{L}\overset{F}{\to}\mathbb{R}^{k+l}$ factors through $L$ where $\gamma$ is a geodesic, we have (cf. [ES64], [EL78] for the notation of the second fundamental form $\nabla d\phi$ of a mapping $\phi$ between Riemannian manifolds), $\nabla d(F\circ\gamma)=dF\circ\nabla(d\gamma)+\nabla d(F)(d\gamma,d\gamma)=\nabla d(F)(d\gamma,d\gamma).$ Since the Christoffel symbols of $\mathbb{S}^{1}$ and of $\mathbb{R}^{n+k}$ are 0 we have $\nabla d(F\circ\gamma)=(F\circ\gamma)^{\prime\prime}.$ Therefore $(F\circ\gamma)^{\prime\prime}=\tilde{A}(F)(\gamma^{\prime},\gamma^{\prime}).$ Integrating along the portion of $\gamma$ from $p$ to $q$, we get $\sqrt{2}\leq\left|\gamma^{\prime}(p)-\gamma^{\prime}(q)\right|\leq\int^{q}_{p}\left|(F\circ\gamma)^{\prime\prime}\right|ds\leq C(K+1)\,a.$ We conclude that that the length $a$ has a lower bound $C/(K+1)$. From the Gauss equations and $|\tilde{A}|<C(K+1)$, the sectional curvatures of $L$ are bounded above by $C^{2}(K+1)^{2}$. We conclude $\mbox{inj}(L)\leq C(K+1)^{-1}$ [Pet06, p.178]. Step 2. Take a normal neighbourhood $U\subset L$ around a given point $p\in L$ and assume $U$ is contained in a normal neighbourhood $V$ of $M$ at $p$. We will use $C(g)$ for constants only depending on the ambient geometry of $(M,g)$. Now, on $V$ we will use $\delta=\langle\cdot,\cdot\rangle_{\mathbb{R}^{k+l}}$, to measure length of various geometric quantities already defined in $(V,g)$. First, $|A|_{\delta}\leq C(g)|A|_{g}\leq C(g)K.$ Identify $T_{p}L$ with $\mathbb{R}^{k}\times\\{0\\}\subset\mathbb{R}^{k+l}$. Let $e_{1}(x),...,e_{k}(x)$ be the orthonormal frame on $U$ obtained by parallel transporting an orthonormal frame $e_{1}(0),...,e_{k}(0)$ at $T_{p}L$ along the unique radial geodesic $r_{x}(s)$ in $(U,f^{\ast}g)$ from $0$ to an arbitrary point $x\in L$, and let $e_{1+k}(0),...,e_{l+k}(0)$ be the orthonormal frame of $(T_{p}L)^{\perp}$. Integrating along $\gamma_{x}(s)$ leads to $\displaystyle\left|\langle e_{i}(x),e_{j+l}(0)\rangle\right|$ $\displaystyle=\left|\langle e_{i}(x),e_{j+l}(0)\rangle-\langle e_{i}(0),e_{j+l}(0)\rangle\right|$ $\displaystyle=\left|\int_{0}^{|x|}\frac{d}{ds}\langle e_{i}(\gamma_{x}(s)),e_{j+l}(0)\rangle ds\right|$ $\displaystyle=\left|\int_{0}^{|x|}\langle e_{i}^{\prime}(s),e_{j+l}(0)\rangle ds\right|$ $\displaystyle\leq\int_{0}^{|x|}\left|\langle\nabla_{\partial_{r}}^{g}e_{i},e_{j+l}(0)\rangle\right|ds$ $\displaystyle=\int_{0}^{|x|}\left|\langle A(\partial_{r},e_{i})+\nabla_{\partial_{r}}^{L}e_{i},e_{j+l}(0)\rangle\right|ds$ $\displaystyle\leq C(g)K\,|x|$ as $\nabla_{\partial_{r}}^{L}e_{i}=0$ on $L$. Therefore, there exists $r_{0}=C(g)K^{-1}$ (where $C(g)$ may differ from the one above) such that for any $x\in B_{r_{0}}(0)$ the projection of each $e_{i}(x)$ in each fixed normal direction $e_{j+l}(0)$ is at most $c_{n}/\sqrt{l}$ and the norm of the projection is no more than some universal constant $c_{n,l}$ that we get to choose. It is known that such $T_{x}L$ projects bijectively to $T_{p}L$. Therefore, locally around any $x\in B_{r_{0}}(0)$, implicit function theorem asserts that $U$ can be written as a graph over a ball in $T_{x}L$, hence as a graph over a ball in $T_{p}L$ from the projection. The graphing functions over the fixed reference plane $T_{P}L$ must coincide on the overlap of any pair of such balls. This yields a global graphing function $\mathcal{F}$ over $B_{r_{0}}^{n}(p)\subset T_{p}L$. Moreover, $|D\mathcal{F}|\leq C(g,l)$ because $D\mathcal{F}$ is close to $T_{x}L$ which is close (measured in $l$) to $T_{p}L$ via the projection. ∎ ## References * [EI05] Joachim Escher and Kazuo Ito, _Some dynamic properties of volume preserving curvature driven flows_ , Math. Ann. 333 (2005), no. 1, 213–230. MR 2169834 * [EL78] J. Eells and L. Lemaire, _A report on harmonic maps_ , Bull. London Math. Soc. 10 (1978), no. 1, 1–68. MR 495450 * [EM02] Y. Eliashberg and N. Mishachev, _Introduction to the $h$-principle_, Graduate Studies in Mathematics, vol. 48, American Mathematical Society, Providence, RI, 2002. MR 1909245 * [ES64] James Eells, Jr. and J. H. Sampson, _Harmonic mappings of Riemannian manifolds_ , Amer. J. Math. 86 (1964), 109–160. MR 164306 * [FFRS21] Alberto Fiorenza, Maria Rosaria Formica, Tomáš G. Roskovec, and Filip Soudský, _Detailed proof of classical Gagliardo-Nirenberg interpolation inequality with historical remarks_ , Z. Anal. Anwend. 40 (2021), no. 2, 217–236. MR 4237368 * [Fif00] Paul C. Fife, _Models for phase separation and their mathematics_ , Electron. J. Differential Equations (2000), No. 48, 26. MR 1772733 * [HL82] Reese Harvey and H. Blaine Lawson, Jr., _Calibrated geometries_ , Acta Math. 148 (1982), 47–157. MR 666108 * [JLS11] Dominic Joyce, Yng-Ing Lee, and Richard Schoen, _On the existence of Hamiltonian stationary Lagrangian submanifolds in symplectic manifolds_ , Amer. J. Math. 133 (2011), no. 4, 1067–1092. MR 2823871 * [Lee13] John M. Lee, _Introduction to smooth manifolds_ , second ed., Graduate Texts in Mathematics, vol. 218, Springer, New York, 2013. MR 2954043 * [May01] Uwe F. Mayer, _A singular example for the averaged mean curvature flow_ , Experiment. Math. 10 (2001), no. 1, 103–107. MR 1822855 * [MM12] Carlo Mantegazza and Luca Martinazzi, _A note on quasilinear parabolic equations on manifolds_ , Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) 11 (2012), no. 4, 857–874. MR 3060703 * [MS17] Dusa McDuff and Dietmar Salamon, _Introduction to Symplectic Topology_ , Oxford University Press, 03 2017. * [Mul99] W. W. Mullins, _Two-dimensional motion of idealized grain boundaries [MR0078836 (17,1252g)]_ , Fundamental contributions to the continuum theory of evolving phase interfaces in solids, Springer, Berlin, 1999, pp. 70–74. MR 1770893 * [Pet06] Peter Petersen, _Riemannian geometry_ , second ed., Graduate Texts in Mathematics, vol. 171, Springer, New York, 2006. MR 2243772 * [PW16] Scott Parkins and Glen Wheeler, _The polyharmonic heat flow of closed plane curves_ , J. Math. Anal. Appl. 439 (2016), no. 2, 608–633. MR 3475940 * [Wei71] Alan Weinstein, _Symplectic manifolds and their Lagrangian submanifolds_ , Advances in Math. 6 (1971), 329–346. MR 286137 * [Whe13] Glen Wheeler, _On the curve diffusion flow of closed plane curves_ , Ann. Mat. Pura Appl. (4) 192 (2013), no. 5, 931–950. MR 3105957 * [Woo20] Albert Wood, _Singularities of lagrangian mean curvature flow_ , Doctoral thesis, UCL (University College London) (2020).
# Removing Sequential Bottleneck of Dijkstra’s Algorithm for the Shortest Path Problem††thanks: Supported by NSF CNS-1812349, CNS-1563544, and the Cullen Trust for Higher Education Endowed Professorship Vijay K. Garg, The University of Texas at Austin, Department of Electrical and Computer Engineering, Austin, TX 78712, USA ###### Abstract All traditional methods of computing shortest paths depend upon edge- relaxation where the cost of reaching a vertex from a source vertex is possibly decreased if that edge is used. We introduce a method which maintains lower bounds as well as upper bounds for reaching a vertex. This method enables one to find the optimal cost for multiple vertices in one iteration and thereby reduces the sequential bottleneck in Dijkstra’s algorithm. We present four algorithms in this paper — $SP_{1}$, $SP_{2}$, $SP_{3}$ and $SP_{4}$. $SP_{1}$ and $SP_{2}$ reduce the number of heap operations in Dijkstra’s algorithm. For directed acyclic graphs, or directed unweighted graphs they have the optimal complexity of $O(e)$ where $e$ is the number of edges in the graph which is better than that of Dijkstra’s algorithm. For general graphs, their worst case complexity matches that of Dijkstra’s algorithm for a sequential implementation but allows for greater parallelism. Algorithms $SP_{3}$ and $SP_{4}$ allow for even more parallelism but with higher work complexity. Algorithm $SP_{3}$ requires $O(n+e(\max(\log n,\Delta)))$ work where $n$ is the number of vertices and $\Delta$ is the maximum in-degree of a node. Algorithm $SP_{4}$ has the most parallelism. It requires $O(ne)$ work. These algorithms generalize the work by Crauser, Mehlhorn, Meyer, and Sanders on parallelizing Dijkstra’s algorithm. ###### Index Terms: Single Source Shortest Path Problem, Dijkstra’s Algorithm ## I Introduction The single source shortest path (SSSP) problem has wide applications in transportation, networking and many other fields. The problem takes as input a weighted directed graph with $n$ vertices and $e$ edges. We are required to find $cost[x]$, the minimum cost of a path from the source vertex $v_{0}$ to all other vertices $x$ where the cost of a path is defined as the sum of edge weights along that path. We assume that all edge weights are strictly positive throughout this paper. Most SSSP algorithms are inspired by Dijkstra’s algorithm [5] or Bellman-Ford [2, 7]. We present four algorithms in this paper in increasing order of work complexity. Algorithms $SP_{1}$, $SP_{2}$ and $SP_{3}$ are inspired by Dijkstra’s algorithm and $SP_{4}$ is inspired by Bellman-Ford algorithm. Algorithms $SP_{1}$ and $SP_{2}$ are suitable for sequential implementations. They improve upon Dijkstra’s algorithm by reducing the total number of heap operations. For acyclic graphs, $SP_{1}$ performs no heap operations (except for the insertion of the initial source vertex) and has the time complexity of $O(e)$. Hence, it unifies Dijkstra’s algorithm with the topological sorting based algorithm for acyclic graphs. $SP_{2}$ has the optimal time complexity of $O(e)$ whenever the input graph is acyclic or unweighted. For general graphs, their worst case asymptotic complexity matches that of Dijkstra’s algorithm for a sequential implementation; however, they always perform less heap operations than Dijkstra’s algorithm. Additionally, they are more suitable for a parallel implementation because they allow multiple vertices to be explored in parallel unlike Dijkstra’s algorithm which explores vertices in the order of their shortest cost. Algorithm $SP_{2}$ allows more parallelism than $SP_{1}$ at the expense of an additional $O(e)$ processing. Algorithm $SP_{3}$ allows for even more parallelism than $SP_{2}$. It uses the technique of keeping lower bounds on $cost[x]$ for all vertices $x$. Almost all algorithms for the shortest path problem are based on keeping upper bounds. Dijkstra’s algorithm keeps $D[x]$, an upper bound on the cost of the path for any vertex $x$. It maintains the invariant that $D[x]$ always reflects the cost of a feasible path in the directed graph from the source vertex to $x$. Our algorithm $SP_{3}$ extends Dijkstra’s algorithm by maintaining the variable $C[x]$ for any vertex $x$ that gives a lower bound on the cost to reach $x$. The invariant we maintain is that any path from the source vertex to $x$ must have cost at least $C[x]$. When $C[x]$ is zero, the invariant is trivially true in a directed graph with no negative weights. At each iteration of the algorithm, we increase $C[x]$ for one or more vertices till we reach a point where $C$ is also feasible and corresponds to the cost of all shortest paths. The vertices that have matching upper bounds and lower bounds are called fixed vertices and the minimum cost from the source vertex to these vertices are known. By combining the upper bounds of Dijkstra’s algorithm with the lower bounds, we present an algorithm, $SP_{3}$, for the single source shortest path algorithm that is superior to Dijkstra’s algorithm in two respects. First, Dijkstra’s algorithm suffers from the well-known sequential bottleneck (e.g. [4, 12]). Outgoing edges of only those vertices are explored (relaxed) whose distance is the minimum of all vertices whose adjacency list have not been explored. In contrast, our algorithm explores all those vertices $x$ whose upper bounds $D[x]$ and lower bounds $C[x]$ match and have not been explored before. Although the idea of marking multiple vertices fixed in a single iteration has been explored before (for e.g. [4]), this is the first paper, to the best of our knowledge, that marks vertices fixed based on the idea of lower bounds. Second, when one is interested in a shortest path to a single destination, our algorithm may determine that $D[x]$ is equal to $C[x]$ much sooner than Dijkstra’s algorithm. There are two assumptions in our algorithms. First, we assume that all weights are strictly positive. This is a minor strengthening of the assumption in Dijkstra’s algorithm where all weights are assumed to be nonnegative. The second assumption is that we have access to incoming edges for any vertex discovered during the execution of the algorithm. Dijkstra’s algorithm uses only an adjacency list of outgoing edges. This assumption is also minor in the context of static graphs. However, when the graph is used in a dynamic setting, it may be difficult to find the list of incoming edges. We assume in this paper that either the graph is static or that a vertex can be expanded in the backward direction in a dynamic graph. The single source shortest path problem has a rich history. One popular research direction is to improve the worst case complexity of Dijkstra’s algorithm by using different data structures. For example, by using Fibonacci heaps for the min-priority queue, Fredman and Tarjan [8] gave an algorithm that takes $O(e+n\log n)$. There are many algorithms that run faster when weights are small integers bounded by some constant $\gamma$. For example, Ahuja et al [1] gave an algorithm that uses Van Emde Boas tree as the priority queue to give an algorithm that takes $O(e\log\log\gamma)$ time. Thorup [14] gave an implementation that takes $O(n+e\log\log n)$ under special constraints on the weights. Raman [13] gave an algorithm with $O(e+n\sqrt{\log n\log\log n})$ time. Our algorithms do not improve the worst case sequential complexity of the problem, but reduce the sequential bottleneck. Our algorithms also reduce the number of priority queue operations in the average case. It is also interesting to compare our approach with algorithm $A^{*}$ [9]. The algorithm $A^{*}$ is applicable when there is a single target vertex and there is a heuristic function $h(x)$ for any vertex that provides the lower bound from $x$ to the target vertex. The heuristic function assumes that there is some background knowledge that provides the lower bound to the target. Our algorithms are not based on a target vertex or availability of the background knowledge. Even though $A^{*}$ also uses the notion of lower bounds, the usage is different. We use the lower bound from the source vertex to $x$ in our algorithms and not the lower bound from $x$ to the target vertex. There are many related works for parallelizing Dijkstra’s algorithm. The most closely related work is Crauser et al [4] which gives three methods to improve parallelism. These methods, in-version, out-version and in-out-version, allow multiple vertices to be marked as fixed instead of just the one with the minimum $D$ value. The in-version marks as fixed any vertex $x$ such that $D[x]\leq\min\\{D[y]~{}|~{}\neg fixed(y)\\}+\min\\{w[v,x]~{}|~{}\neg fixed(x)\\}$. This method is a special case of our algorithm $SP_{2}$. The implementation of in-version in [4] requires an additional priority queue and the total number of heap operations increases by a factor of $2$ compared to Dijkstra’s algorithm even though it allows greater parallelism. Our algorithm $SP_{2}$ uses fewer heap operations than Dijkstra’s algorithm. The out-version in [4] works as follows. Let $L$ be defined as $\min\\{D[x]+w[x,y]~{}|~{}\neg fixed(x)\\}$. Then, the out-version marks as fixed all vertices that have $D$ value less than or equal to $L$. Our method is independent of this observation and we incorporate out-version in algorithms $SP_{3}$ and $SP_{4}$. The in- out-version is just the use of in-version as well as out-version in conjunction. A popular practical parallel algorithm for SSSP is $\Delta$-stepping algorithm due to Meyer and Sanders [12]. Meyer and Sanders also provide an excellent review of prior parallel algorithms in [12]. They classify SSSP algorithms as either label-setting, or label-correcting. Label-setting algorithms, such as Dijkstra’s algorithm, relax edges only for fixed vertices. Label-correcting algorithms may relax edges even for non-fixed vertices. $\Delta$-stepping algorithm is a label-correcting algorithm in which eligible non-fixed vertices are kept in an array of buckets such that each bucket represents a distance range of $\Delta$. During each phase, the algorithm removes all vertices of the first non-empty bucket and relaxes all the edges of weight at most $\Delta$. Edges of higher weights are relaxed only when their starting vertices are fixed. The parameter $\Delta$ provides a trade-off between the number of iterations and the work complexity. For example, when $\Delta$ is $\infty$, the algorithm reduces to Bellman-Ford algorithm where any vertex that has its $D$ label changed is explored. When $\Delta$ equals $1$ for integral weights, the algorithm is a variant of Dijkstra’s algorithm. They show that by taking $\Delta=\Theta(1/d)$ where $d$ is the maximum degree of a graph on $n$ vertices, and random edge weights that are uniformly distributed in $[0,1]$, their algorithm takes $O(n+e+dM)$ where $M$ is the maximum shortest path weight from the source vertex to any other vertex. There are many practical large-scale implementations of the $\Delta$-stepping algorithm (for instance, by Madduri et al [11]) in which authors have shown the scalability of the algorithm. Chakravarthy et al [3] give another scalable implementation of an algorithm that is a hybrid of the Bellman-Ford algorithm and the $\Delta$-stepping algorithm. The $\Delta$-stepping technique is orthogonal to our method which is based on keeping lower bounds with vertices. It is possible to apply $\Delta$-stepping in conjunction with our method. In summary, we present four algorithms for SSSP in this paper in order of increasing work complexity. We only compute the cost of the shortest paths and not the actual paths because the standard method of keeping backward parent pointers is applicable to all of our algorithms. Algorithm $SP_{1}$ counts the number of incoming edges to a vertex that have been relaxed. When all incoming edges have been relaxed, we show that it is safe to mark this vertex as fixed. The algorithm $SP_{2}$ generalizes $SP_{1}$ to allow even those vertices to be marked as fixed which have incoming edges from non-fixed vertices under certain conditions. Both of these algorithms have fewer heap operations than Dijkstra’s algorithm for the sequential case and allow more parallelism when multiple cores are used. The algorithm $SP_{3}$ generalizes $SP_{2}$ further by maintaining the lower bound $C$ for each vertex. All these algorithms are label-setting. Algorithm $SP_{3}$ has the same asymptotic complexity as Dijkstra’s algorithm when the maximum in-degree of a vertex is $O(\log n)$. It allows even more parallelism than $SP_{2}$. The algorithm $SP_{4}$ is a label- correcting algorithm. It has the the most parallelism but with highest work complexity. $SP_{4}$ combines ideas from Bellman-Ford, Dijkstra, [4] and $SP_{3}$ for faster convergence of $D$ and $C$ values. ## II Background and Notation Dijkstra’s algorithm (or one of its variants) is the most popular single source shortest path algorithm used in practice. For concreteness sake we use the version shown in Fig. 1 for comparison with our algorithm. The algorithm also helps in establishing the terminology and the notation used in our algorithm. We consider a directed weighted graph $(V,E,w)$ where $V$ is the set of vertices, $E$ is the set of directed edges and $w$ is a map from the set of edges to positive reals (see Fig. 2 for a running example). To avoid trivialities, we assume that the graph is loop-free and every vertex $x$, except the source vertex $v_{0}$, has at least one incoming edge. var $D$: array[$0\ldots n-1$] of integer initially $\forall i:D[i]=\infty$; $fixed$: array[$0\ldots n-1$] of boolean initially $\forall i:fixed[i]=false$; $H$: binary heaps of $(j,d)$ initially empty; $D[0]:=0$; $H$.insert((0,D[0])); while $\neg H$.empty() do $(j,d):=H$.removeMin(); $fixed[j]:=true$; forall $k$: $\neg fixed(k)\wedge(j,k)\in E$ if ($D[k]>D[z]+w[z,k]$) then $D[k]:=D[z]+w[z,k]$; $H$.insertOrAdjust $(k,D[k])$; endwhile; Figure 1: Dijkstra’s algorithm to find the shortest costs from $v_{0}$ . $v_{0}$$v_{2}$$v_{1}$$v_{3}$$v_{4}$ Figure 2: A Weighted Directed Graph Dijkstra’s algorithm maintains $D[i]$, which is a tentative cost to reach $v_{i}$ from $v_{0}$. Every vertex $x$ in the graph has initially $D[x]$ equal to $\infty$. Whenever a vertex is discovered for the first time, its $D[x]$ becomes less than $\infty$. We use the predicate $discovered(x)\equiv D[x]<\infty$. The variable $D$ decreases for a vertex whenever a shorter path is found due to edge relaxation. In addition to the variable $D$, a boolean array fixed is maintained. Thus, every discovered vertex is either fixed or non-fixed. The invariant maintained by the algorithm is that if a vertex $x$ is fixed then $D[x]$ gives the final shortest cost from vertex $v_{0}$ to $x$. If $x$ is non-fixed, then $D[x]$ is the cost of the shortest path to $x$ that goes only through fixed vertices. A heap $H$ keeps all vertices that have been discovered but are non-fixed along with their distance estimates $D$. We view the heap as consisting of tuples of the form $(j,D[j])$ where the heap property is with respect to $D$ values. The algorithm has one main while loop that removes the vertex with the minimum distance from the heap with the method $H$.removeMin(), say $v_{j}$, and marks it as fixed. It then explores the vertex $v_{j}$ by relaxing all its adjacent edges going to non-fixed vertices $v_{k}$. The value of $D[k]$ is updated to the minimum of $D[k]$ and $D[j]+w[j,k]$. If $v_{k}$ is not in the heap, then it is inserted, else if $D[k]$ has decreased then the label associated with vertex $k$ is adjusted in the heap. We abstract this step as the method $H$.insertOrAdjust$(k,D[k])$. The algorithm terminates when the heap is empty. At this point there are no discovered non-fixed vertices and $D$ reflects the cost of the shortest path to all discovered vertices. If a vertex $j$ is not discovered then $D[j]$ is infinity reflecting that $v_{j}$ is unreachable from $v_{0}$. Observe that every vertex goes through the following states. Every vertex $x$ is initially undiscovered (i.e., $D[x]=\infty$). If $x$ is reachable from the source vertex, then it is eventually discovered (i.e., $D[x]<\infty)$. A discovered vertex is initially non-fixed, and is therefore in the heap $H$. Whenever a vertex is removed from the heap it is a fixed vertex. A fixed vertex may either be unexplored or explored. Initially, a fixed vertex is unexplored. It is considered explored when all its outgoing edges have been relaxed. The following lemma simply summarizes the well-known properties of Dijkstra’s algorithm. ###### Lemma 1. The outer loop in Dijkstra’s algorithm satisfies the following invariants. (a) For all vertices $x$: $fixed[x]\Rightarrow(D[x]=cost[x])$. (b) For all vertices $x$: $D[x]$ is equal to cost of the shortest path from $v_{0}$ to $x$ such that all vertices in the path before $x$ are fixed. (c) For all vertices $x$: $x\in H$ iff $discovered(x)\wedge\neg fixed[x]$. ## III Algorithm $SP_{1}$: Using Predecessors Dijkstra’s algorithm finds the vertex with the minimum tentative distance and marks it as a fixed vertex. This is the only mechanism by which a vertex is marked as fixed in Dijkstra’s algorithm. Finding the non-fixed vertex with the minimum $D$ value takes $O(\log n)$ time when a heap or its variant is used. Our first observation is that if for any non-fixed vertex $x$, if all the incoming edges are from fixed vertices, then the current estimate $D[x]$ is the shortest cost. To exploit this observation, we maintain with each vertex $i$, a variable $pred[i]$ that keeps the number of incoming edges that have not been relaxed. The variable $pred[i]$ is decremented whenever an incoming edge to vertex $i$ is relaxed. When $pred[i]$ becomes zero, vertex $i$ becomes fixed. Determining a vertex to be fixed by this additional method increases the rate of marking vertices as fixed in any iteration of the while loop. The second observation is that in Dijkstra’s algorithm vertices are explored only in order of their cost. $SP_{1}$ explores vertices whenever it finds one that is fixed. Hence, in addition to the heap $H$, we maintain a set $R$ of vertices which have been fixed but not explored, i.e., their adjacency lists have not been traversed. We also relax the invariant on the heap $H$. In Dijkstra’s algorithm, the heap does not contain fixed vertices. In algorithm $SP_{1}$, the heap $H$ may contain both fixed and non-fixed vertices. However, only those fixed vertices which have been explored may exist in the heap. var $D$: array[$0\ldots n-1$] of integer initially $\forall i:D[i]=\infty$; $H$: binary heap of $(j,d)$ initially empty; $fixed$: array[$0\dots n-1$] of boolean initially $\forall i:fixed[i]=false$; $Q,R$: set of vertices initially empty; $pred$: array[$0\ldots n-1$] of integer initially $\forall i:pred[i]=~{}|~{}\\{x~{}|~{}(x,v_{i})\in E\\}~{}|~{}$; $D[0]:=0$; $H$.insert$((0,D[0]))$; while $\neg H$.empty() do $(j,d):=H$.removeMin(); if ($\neg fixed[j]$) then $R$.insert($j$); $fixed[j]$ := $true$; while $R\neq\\{\\}$ do forall $z\in R$ $R$.remove($z$); forall $k:\neg fixed(k)\wedge(z,k)\in E$: processEdge1($z,k$); endwhile; forall $z\in Q$: $Q$.remove($z$); if $\neg fixed[z]$ then $H$.insertOrAdjust $(z,D[z])$; endwhile; procedure processEdge1($z,k$); var $changed$: boolean initially false; $pred[k]:=pred[k]-1$; if ($D[k]>D[z]+w[z,k]$) then $D[k]:=D[z]+w[z,k]$; $changed$ := true; if $(pred[k]=0)$ then $fixed[k]:=true$; $R$.insert($k$); else if ($changed\wedge(k\not\in Q)$) then $Q$.insert($k$); Figure 3: Algorithm $SP_{1}$ The algorithm $SP_{1}$ is shown in Fig. 3. The algorithm starts with the insertion of the source vertex with its $D$ value as $0$ in the heap. Instead of removing the minimum vertex from the heap in each iteration and then exploring it, the algorithm consists of two while loops. The outer while loop removes one vertex from the heap. If this vertex is fixed, then it has already been explored and therefore it is skipped; otherwise, it is marked as fixed and inserted in $R$ to start the inner while loop. The inner loop keeps processing the set $R$ till it becomes empty. We do not require that vertices in $R$ be explored in the order of their cost. If $R$ consists of multiple vertices then all of them can be explored in parallel. During this exploration other non-fixed vertices may become fixed. These are then added to $R$. Some vertices may initially be non-fixed but eventually while processing $R$ may become fixed. To avoid the expense of inserting these vertices in the heap, we collect all such vertices which may need to be inserted or adjusted in the heap in a separate set called $Q$. Only, when we are done processing $R$, we call $H.insertOrAdjust$ on vertices in $Q$. The vertices $z\in R$ are explored as follows. We process all out-going adjacent edges $(z,k)$ of the vertex $z$ to non-fixed vertices $k$. This step is called processEdge1 in Fig. 3. First, we decrement the count $pred[k]$ to account for its predecessor $z$ being fixed. Then, we do the standard edge- relaxation procedure by checking whether $D[k]$ can be decreased by taking this edge. If $pred[k]$ is zero, $k$ is marked as fixed. Setting $fixed[k]$ to true also removes it effectively from the heap because whenever a fixed vertex is extracted in the outer while loop it is skipped. Finally, if $D[k]$ has decreased and $pred[k]$ is greater than $0$, we insert it in $Q$ so that once $R$ becomes empty we can call $H$.insertOrAdjust() method on vertices in $Q$ Consider the graph in Fig. 2(a). Initially $(0,D[0])$ is in the heap $H$. Since there is only one vertex in the heap $H$, it is also the minimum. This vertex is removed and inserted in $R$ marking $v_{0}$ as fixed. Now, outgoing edges of $v_{0}$ are relaxed. Since $pred[1]$ becomes $0$, $v_{1}$ is marked as fixed and added to $R$. The vertex $v_{2}$ has $pred$ as $1$ and $D[2]$ as $2$ after the relaxation of edge $(v_{0},v_{2})$. The vertex $v_{2}$ is inserted in the $Q$ for later insertion in the heap. Since $R$ is not empty, outgoing edges of $v_{1}$ are relaxed. The vertex $v_{3}$ is inserted in $Q$ and its $D$ value is set to $12$. The vertex $v_{4}$ is also inserted in $Q$ and its $D$ value is set to $11$. At this point $R$ is empty and we insert vertices in $Q$ in $H$ and get back to the outer while loop. The minimum vertex $v_{2}$ is removed from the heap, marked as fixed, and inserted in $R$ for exploration. When $v_{2}$ is explored, the $D$ label of $v_{3}$ is adjusted to $8$. When edge $(v_{2},v_{4})$ is relaxed, $D[4]$ is reduced to $7$. Moreover, $pred[4]$ becomes zero and $v_{4}$ is inserted in $R$ for exploration. When $v_{4}$ is explored, $pred[3]$ also becomes zero and is also inserted in $R$. Once $v_{3}$ is explored, $R$ becomes empty. We then go to the outer while loop. All vertices in the heap are fixed and therefore the algorithm terminates with $D$ array as $[0,9,2,8,7]$. Observe that it is easy to maintain a count of the non-fixed vertices in the the heap and the method $H.empty()$ can be overloaded to return true whenever this count is zero. We now show that ###### Lemma 2. Let $v$ be any non-fixed vertex. Suppose all incoming edges of $v$ have been relaxed, then $D[v]$ equals $cost[v]$. ###### Proof. We show that whenever $pred[v]$ is zero, $D[v]$ equals $cost[v]$. We prove this lemma by contradiction. If not, let $x$ be the vertex with the smallest $D$ value such that all its incoming edges have been relaxed but $D[x]$ is greater than $cost[x]$. Let $\alpha$ be a path from $v_{0}$ to $x$ with the smallest cost (and therefore less than $D[x]$). The path $\alpha$ must go through a non-fixed vertex because $D[x]$ is the minimum cost of all paths that go through fixed vertices. Let $y$ be the last non-fixed vertex along this path. The successor of $y$ in that path cannot be $x$ because all predecessors of $x$ are fixed. Therefore, its successor is a fixed vertex $z$ because $y$ is the last non-fixed vertex along the path. The path $\alpha$ can be broken into two parts — the path from the source vertex to $z$ and then the path from $z$ to $x$. The path from $z$ to $x$ consists only of fixed vertices by the definition of $y$. It is sufficient to show that there exists a path from the source vertex to $z$ that consists only of fixed vertex with the same cost as in $\alpha$. The vertex $z$ can be fixed either because it has the minimum value of $D$ in the heap at some iteration, or because all the incoming edges to $z$ have been relaxed. In the former case, $D[z]=cost[z]$ and therefore there exists a path from the source vertex to $z$ with only fixed vertices and the minimum cost. In the latter case, when $z$ is fixed because all its incoming edges have been relaxed, then by our choice of $x$, $D[z]$ is equal to $cost[z]$ which again shows existence of a path with only fixed vertices with the minimum cost. ∎ To show the correctness of Algorithm $SP_{1}$, we make the following claims. We use the predicate $explored(x)$ that holds true iff the adjacency list of $x$ has been explored. ###### Lemma 3. The following invariants hold at the outer and the inner while loop of $SP_{1}$. (a) For all vertices $x$: $fixed[x]\Rightarrow D[x]=cost[x]$. (b) For all vertices $x$: $D[x]=$ cost of the shortest path from $v_{0}$ to $x$ such that all vertices in the path before $x$ are fixed. (c) For all vertices $x$: $x\in H\Rightarrow discovered(x)\wedge(\neg fixed[x]\vee explored(x))$. Furthermore, $\forall x:discovered(x)\wedge\neg fixed[x]\Rightarrow(x\in H)$. ###### Proof. (a,b) The only difference from Dijkstra’s algorithm is that in one iteration of the outer while loop, not only vertices with the minimum value of $D$ are fixed, but also vertices with $pred[x]$ equal to $0$. Due to Lemma 2, the invariant on $fixed$ and $D$ continues to hold. In the inner loop, whenever a vertex is discovered and is not fixed, it is inserted in the heap maintaining the invariant on $H$. (c) Whenever a vertex $x$ is discovered and it is not fixed, it is inserted in the heap. Whenever a vertex is removed from the heap it is marked as fixed. A vertex in the heap can also become fixed in the inner while loop. However, whenever a vertex becomes fixed it is inserted in $R$ for exploration and $R$ is empty at the outer while loop. Hence, any vertex that is fixed is also explored. ∎ ###### Lemma 4. The following invariant holds at the inner while loop of $SP_{1}$. For all vertices $x$: $x\in R$ iff $fixed[x]\wedge\neg explored(x)$. ###### Proof. Whenever a vertex is marked fixed initially, it is inserted in $R$. Whenever it is explored, it is removed from $R$. ∎ We now have the following Theorem. ###### Theorem 1. Algorithm $SP_{1}$ returns the weight of a shortest path from source vertices to all other vertices. ###### Proof. Consider any vertex $x$ reachable from the source vertex. We show that $x$ is eventually discovered. We use induction on $k$ equal to the number of vertices with cost that less than or equal to that of $x$. The base case is trivial. For the inductive case, $x$ has at least one predecessor. Since all weights are positive, all predecessors of $x$ have cost less than that of $x$. If all vertices are sorted based on their cost, the outer while loop marks as fixed at least one vertex with cost less than or equal to $x$. Hence, in at most $k$ iterations of the outer while loop, one of the predecessors of $x$ is marked as fixed. The algorithm terminates only when every fixed vertex is explored, and therefore $x$ is discovered. Any vertex $x$ that is discovered is either in $H$ when it is not fixed or fixed but explored, or in $R$ when it is fixed and not explored. If the vertex $x$ is $fixed$, from the invariant on $D$ and $fixed$, we have that $D[x]$ equal to $cost[x]$. If the vertex $x$ is not $fixed$, it is eventually removed from the heap $H$ and becomes fixed. Hence, any reachable vertex $x$ has its $D[x]$ set to $cost[x]$. If any vertex $x$ is not reachable, then it can never be discovered and $D[x]$ returns $\infty$ due to initialization. ∎ We now show that $SP_{1}$ cuts down the complexity of Dijkstra’s algorithm significantly for acyclic graphs whenever source vertex is the only vertex with no incoming edges. To ensure this, whenever we read the graph we create a list $L$ of all vertices other than the source vertex that have no incoming edges. All these vertices are clearly not reachable from the source vertex. We then repeatedly remove vertices from the list $L$ and their outgoing edges from the graph. If in this process, another vertex has all its incoming edges removed, it is added to the list $L$. The procedure is continued until $L$ becomes empty and we are guaranteed that the source vertex is the only vertex with no incoming edges. This procedure takes at most $O(e)$ time because any edge is processed at most once. We now have the following result. ###### Theorem 2. $SP_{1}$ takes $O(e+n\log n)$ time with Fibonacci heaps for any directed graph and takes $O(e)$ time for directed acyclic graphs in which source node is the only one with with zero incoming edges. ###### Proof. For a general directed graph, any steps taken in $SP_{1}$ is also taken in Dijkstra’s algorithm except for the constant time operations such as decrementing $pred$, and inserting or deleting a vertex from $Q$ and $R$. Both $Q$ and $R$ can be implemented as a linked lists with $O(1)$ insertions at the tail and $O(1)$ deletions at the head of the list. The membership in $Q$ can also be implemented in $O(1)$ time using a bit vector. Hence, using Fibonacci heaps, we get the time complexity of Dijkstra’s algorithm. For directed acyclic graphs, initially the source vertex is removed from the heap and inserted in $R$. Now as we explore $R$, the predecessor count for all vertices adjacent to the source vertex will decrease by $1$. Since the graph is acyclic, at least one new vertex will become fixed. As we continue processing $R$, all the nodes of the graph will become fixed (just as in the topological sort of an acyclic graph). Thus, all reachable vertices of an acyclic graph will be processed in the first iteration of the outer while loop. In this iteration, every edges is processed exactly once, giving us $O(e)$ time complexity. ∎ The worst case for $SP_{1}$ is when the vertex discovered last has outgoing edges to all other vertices. In such a worst-case scenario, $SP_{1}$ will not have any vertex becomes fixed through processing of $R$ and the algorithm will degenerate into Dijkstra’s algorithm. ## IV Algorithm $SP_{2}$: Using Weights of Incoming edges We now strengthen our mechanism to mark vertices as fixed. $SP_{2}$ requires access to incoming edges for any vertex. Let a vertex $k$ be discovered from a predecessor vertex $z$. Then, we compute $inWeight[k]$ as the minimum weight of incoming edges from all predecessors other than $z$. We exploit $inWeight$ as follows. ###### Lemma 5. Let $k$ be any non-fixed vertex discovered from the vertex $z$ in any iteration of the outer while loop with $d$. If ($D[k]\leq d+inWeight[k]$) then $D[k]$ equals $cost[k]$. ###### Proof. Since $d$ is the weight of the vertex removed from the heap $H$, we know that any predecessor vertex $v$ that is not fixed is guaranteed to have $D[v]\geq d$. Hence, $D[k]$ is guaranteed to be less than or equal to $D[v]+w[v,k]$ for any incoming edge $(v,k)$ that is relaxed. ∎ This mechanism comes at the space overhead of maintaining an additional array $inWeight[]$ indexed by vertices. $inWeight$: array [$0\ldots n-1$] of int initially $\forall i:inWeight[i]=\infty$; procedure processEdge2($z,k$); var $changed$: boolean initially false; $pred[k]:=pred[k]-1$; // Step 1: vertex $k$ has been discovered. // Compute $inWeight$ if $(D[k]=\infty)\wedge(pred[k]>0)$ then $inWeight[k]:=\min\\{w[v,k]~{}|~{}(v,k)\in E,v\neq z\\}$; // Step 2: relax $(z,k)$ edge if ($D[k]>D[z]+w[z,k]$) then $D[k]:=D[z]+w[z,k]$; $changed$ := true; // Step 3: check if vertex $k$ can be fixed. if $((pred[k]=0)\vee(D[k]\leq d+inWeight[k])$ then $fixed[k]:=true$; $R:=R$.insert$(k)$; else if ($changed\wedge(k\not\in Q)$) then $Q$.insert$(k)$; Figure 4: Algorithm $SP_{2}$: Algorithm $SP_{1}$ with processEdge2 After incorporating Lemma 5, we get the algorithm $SP_{2}$ shown in Fig. 4. It is same as $SP_{1}$ except we use the procedure $processEdge2$ instead of $processEdge1$. In step 1, we compute $inWeight[k]$ when it is discovered for the first time, i.e., when $D[k]$ is $\infty$. If there are additional incoming edges, i.e., $(pred[k]>0)$, we determine the minimum of all the incoming weights except from the vertex $z$ that discovered $k$. In step 2, we perform the standard edge-relaxation. In step 3, we check if the vertex $k$ can be fixed either because it has no more predecessors, or for any non-fixed predecessor $v$, the relaxation of the edge $(v,k)$ will not change $D[k]$. Observe that for sequential implementations, if $R$ is maintained as a queue and all edge weights are uniform, then any vertex discovered for the first time will always be marked as fixed and will never be inserted in the heap. For such inputs, $SP_{2}$ will behave as a simple breadth-first-search. Since any vertex is discovered at most once, computing $inWeight$ requires processing of all incoming edges of a vertex at most once. Hence, the cumulative time overhead is linear in the number of edges. If the graph is unweighted, then $SP_{2}$ is much faster than Dijkstra’s algorithm when $R$ is implemented as a queue. ###### Theorem 3. Suppose that $R$ is implemented as a simple queue. $SP_{2}$ takes * • $O(e+n\log n)$ time with Fibonacci heaps for any directed graph, * • $O(e)$ time for directed acyclic graphs in which only the source node has zero incoming edges, * • $O(e)$ time for any unweighted directed graph. ###### Proof. Since $SP_{2}$ retains all properties of $SP_{1}$, we only need to prove the claim on unweighted directed graphs. In unweighted directed graphs, once the source vertex is explored any vertex $k$ adjacent to the source vertex become fixed because it satisfies the condition that $D[k]\leq d+inWeight[k]$ and is inserted in $R$. Continuing in this manner, the algorithm reduces to breadth- first search by simply inserting nodes in $R$ in breadth-first manner and removing from $R$ till all reachable vertices are explored. ∎ Hence, $SP_{2}$ unifies Dijkstra’s algorithm with the topological sort for acyclic graphs as well as the breadth-first search for unweighted graphs. Consequently, it is faster than Dijkstra’s algorithm when the input graph is close to an acyclic graph (i.e., has few cycles) or close to an unweighted graph (most weights are the same). Lemma 5 is similar to the in-version method of [4]. The in-version fixes any vertex $k$ such that $D[k]\leq d+\min\\{w[j,k]~{}|~{}\neg fixed(j),(j,k)\in E\\}$. There are two differences. First, we do not include the weight of the edge that discovered $k$ in our calculation of $inWeight$. Second, in [4] the implementation is based on maintaining an additional priority queue which adds the overhead of $O(e\log n)$ to the algorithm with ordinary heap implementation. $SP_{2}$ adds a cumulative overhead of $O(e)$. In sequential implementations, the in-version increases the number of heap operations, whereas $SP_{2}$ decreases this number. Consider the graph in Fig. 2(a). Initially $(0,0)$ is in the heap $H$. It is removed and inserted in $R$ marking $v_{0}$ as fixed. Now outgoing edges of $v_{0}$ are relaxed. Since $pred[1]$ becomes $0$, $v_{1}$ is marked as fixed and added to $R$. The vertex $v_{2}$ has $pred[2]$ as $1$ after the relaxation. It is inserted in the heap with $D$ value as $2$ and $inWeight[2]$ is computed as $1$. Since $R$ is not empty, outgoing edges of $v_{1}$ are relaxed. The vertex $v_{3}$ is inserted in $Q$ with $D$ value $12$ and the vertex $v_{4}$ is inserted with $D$ value as $11$. We also compute $inWeight[3]$ as $\min\\{6,8\\}$ equal to $6$ and $inWeight[4]$ as $5$. At this point $R$ is empty and the minimum vertex $v_{2}$ is removed from the heap and marked as fixed. When $v_{2}$ is explored and the edge $(v_{2},v3)$ is relaxed, the label of $v_{3}$ is adjusted to $8$. Since $8$ is less than or equal to $d+inWeight[3]=2+6$, it is marked as fixed and inserted in $R$. When edge $(v_{2},v_{4})$ is relaxed, $D[4]$ is reduced to $7$. Moreover, $pred[4]$ becomes zero and $v_{4}$ is inserted in $R$ for exploration. At this point, all vertices are fixed. When $R$ is processed, there are no additional changes and the algorithm terminates with the $D$ array as $[0,9,2,8,7]$. ## V Algorithm $SP_{3}$: Using Lower Bounds with Upper Bounds We now generalize the mechanism of $SP_{2}$ further to determine fixed vertices based on the idea of using lower bounds. The idea of starting with the infinite cost as an estimate of the actual cost and decreasing the estimate via edge-relaxation has been the underlying principle for not only Dijkstra’s algorithm but almost all other shortest path algorithms such as Bellman-Ford, Floyd-Warshall [6] and their derivatives. In this section, we present the idea of using lower bounds $C$ associated with every vertex in addition to the upper bounds given by $D$. We keep a global array $C$ such that $C[x]$ is the lower bound associated with each vertex $x$. We maintain the invariant that there is no path of cost strictly lower than $C[x]$ from the source vertex to $x$. Just as $D[i]$ is initialized to $\infty$, $C[i]$ is initialized to $0$ for all $i$ so that the invariant is true initially. Clearly, any vertex $x$ such that $C[x]$ and $D[x]$ are equal has both of them equal to $cost[x]$. Hence, any vertex with $C[x]$ equal to $D[x]$ can be marked as fixed. Conversely, if any vertex $x$ is known to be fixed (for example, by removal from the min-heap), then $C[x]$ can be set to $D[x]$. How do we determine nontrivial $C[x]$ for non-fixed vertices? Just as the exploration of a vertex $x$ in Dijkstra’s algorithm updates $D[y]$ for all out-going edges $(x,y)$, we define a dual step that can update $C[x]$ based upon all in-coming edges. The value of $C[x]$ for the source vertices is always zero. For other vertices, we have ###### Lemma 6. Let $C[x]$ be a lower bound on the cost of the shortest path to $x$. Then, for any vertex $x$ that is not a source vertex, $C[x]\geq\min\\{C[v]+w[v,x]~{}|~{}(v,x)\in E\\}$ (1) ###### Proof. Since $x$ is not the source vertex, it must have a predecessor $v$ in a shortest path from the source vertex to $x$. The equation follows by noting that an additional cost of $w[v,x]$ would be incurred as the last edge on that path. ∎ The lemma gives an alternate short proof of Lemma 2. Consider any $x$ such that all its predecessors are fixed. Since $C[v]$ is equal to $D[v]$ for all fixed vertices, from Eqn 1, we get that $C[x]\geq\min\\{D[v]+w[v,x]~{}|~{}(v,x)\in E\\}.$ We also get that $D[x]\leq\min\\{D[v]+w[v,x]~{}|~{}(v,x)\in E\\}$ using the edge-relaxation rule. Combining these two inequalities with $C[x]\leq D[x]$, we get that $C[x]$ is equal to $D[x]$ and therefore $x$ can be marked as fixed. An additional lower bound on the cost of a vertex is determined using the global information on the graph. At any point in execution of the graph, there are two sets of vertices — fixed and non-fixed. Any path from the source vertex to a non-fixed vertex must include at least one edge from the set of edges that go from a fixed vertex to a non-fixed vertex. ###### Lemma 7. For any $x$ such that $\neg fixed[x]$, $C[x]\geq\min\\{C[u]+w[u,v]~{}|~{}(u,v)\in E\wedge fixed[u]\wedge\neg fixed[v]\\}$. ###### Proof. Consider the shortest path from $v_{0}$ to $x$. Since $fixed[v_{0}]$ and $\neg fixed[x]$ there is an edge in the path from a fixed vertex $u^{\prime}$ to a non-fixed vertex $v^{\prime}$. We get that $C[x]\geq C[v^{\prime}]$ and $C[v^{\prime}]\geq\min\\{C[u]+w[u,v]~{}|~{}(u,v)\in E\wedge fixed[u]\wedge\neg fixed[v]\\}$. ∎ Since for a fixed vertex $u$, $C[u]$ equals $D[u]$, we get that for any non- fixed vertex $x$, $C[x]\geq\min\\{D[u]+w[u,v]~{}|~{}(u,v)\in E\wedge fixed[u]\wedge\neg fixed[v]\\}$. The right hand side is simply the minimum key in the min-heap $H$. Finally, we also exploit the method of [4]. ###### Lemma 8. [4] Let $threshold=\min\\{D[u]+w[u,v]~{}|~{}(u,v)\in E\wedge\neg fixed[u]\\}$. Consider any non-fixed vertex $x$ with $D[x]\leq threshold$. Then, $x$ can be marked as a fixed vertex. ###### Proof. Since $D[x]\leq threshold$, we know that $x$ is a discovered vertex and there is a path from $v_{0}$ to $x$. We show that this path has the shortest cost. Suppose that there is another path with cost less than $D[x]$. This path must go through at least some non-fixed vertex because $D[x]$ already accounts for all paths that go through only fixed vertices. Let $u^{\prime}$ be the first non-fixed vertex on that path. Then, the cost of that path is at least $threshold$ by the definition of $threshold$ giving us the contradiction. ∎ This lemma also allows us to mark multiple vertices as fixed and therefore update $C$ for them. To exploit Lemma 8, we use two additional variables in $SP_{3}$. The variable $outWeight[x]$ keeps the weight of the minimum outgoing edge from $x$. This array is computed exactly once with the cumulative overhead of $O(e)$. We also keep an additional binary heap $G$ as proposed in [4]. This heap keeps $D[u]+outWeight[u]$ for all non-fixed vertices. Clearly, the minimum value of this heap is the required threshold. var $C,D$: array[$0\ldots n-1$] of integer initially $\forall i:(C[i]=0)\wedge(D[i]=\infty)$; $G,H$: binary heap of $(j,d)$ initially empty; $fixed$: array[$0\ldots n-1$] of boolean initially $\forall i:fixed[i]=false$; $Q,R$: set of vertices initially empty; $outWeight$: array[$0\ldots n-1$] of integer initially $\forall i:outWeight[i]=\min\\{w[i,j]~{}|~{}(i,j)\in E\\}$; $D[0]:=0$; $H$.insert$(0,D[0])$; $G$.insert$(0,D[0]+outWeight[0])$; while $\neg H$.empty() do int threshold := $G$.getMin(); while ($H$.getMin() $\leq threshold)$ do $(j,d):=H$.removeMin(); $G$.remove$(j)$; $fixed[j]$ := $true$; $C[j]:=D[j]$; $R$.insert($j$); if ($H$.empty()) break; endwhile; while $R\neq\\{\\}$ do forall $z\in R$ $R:=R-\\{z\\}$ forall $k$: $\neg fixed(k)\wedge(z,k)\in E$ processEdge3(z,k); endwhile; forall $z\in Q$: $Q$.remove($z$); if $\neg fixed[z]$ then { $H$.insertOrAdjust $(z,D[z])$; $G$.insertOrAdjust($z,D[z]+outWeight[z])$;} endwhile; procedure processEdge3($z,k$); var $changed$: boolean initially false; $minD,minU$: int initially $\infty$; // step 1: edge relaxation if ($D[k]>D[z]+w[z,k]$) then $D[k]:=D[z]+w[z,k]$; $changed$ := true; // step 2: Update $C[v]$ for all predecessors $v$ of $k$ forall $v:\neg fixed[v]\wedge((v,k)\in E)$ $C[v]:=\max(C[v],H$.getMin()); // step 3: Update $C$ via Eqn. 1 $C[k]:=\max(C[k],\min\\{C[v]+w[v,k]~{}|~{}((v,k)\in E)\\})$; // step 4: check if vertex $k$ is fixed if $(C[k]=D[k])$ then $fixed[k]:=true$; $R:=R\cup\\{k\\};$ $G$.remove($k$); $H$.remove($k$); else if ($changed\wedge(k\not\in Q)$) then $Q$.insert$(k)$; Figure 5: Algorithm $SP_{3}$: Using upper bounds as well as lower bounds Our third algorithm $SP_{3}$ is shown in Fig. 5. We first remove from the heap $H$ all those non-fixed vertices $j$ such that $D[j]\leq threshold$. All these vertices are marked as fixed. Also, whenever any vertex is added or removed from the heap $H$, we also apply the same operation on the heap $G$. In $SP_{3}$, it is more convenient to keep only the non-fixed vertices in $G$ and $H$. All the vertices that are marked as fixed are removed from both $G$ and $H$. Note that the deletion from the heap is only a virtual operation. It simply corresponds to marking that vertex as fixed. Whenever a vertex is removed from any of the heaps in the removeMin operation, and it is a fixed vertex, the algorithm simply discards that vertex and continues. Hence, vertices are physically removed only via removeMin operation. The getMin operation removes any fixed vertex via removeMin, so that getMin applies only to the non-fixed vertices. Whenever vertices are removed from $H$ via removeMin operation, they are inserted in $R$ which explores them using processEdge3. Whenever we process an edge $(z,k)$, we update $D[k]$ as well as $C[k]$. If $C[k]$ and $D[k]$ become equal then $v_{k}$ is marked as a fixed vertex; otherwise, if $D[k]$ has changed then it is inserted in $Q$ for later processing. To update $C[k]$, we first apply Lemma 7 to all the non-fixed predecessors of $k$, and then use Eqn. 1 to update $C[k]$. To apply Lemma 7, we set $C[v]$ for any non-fixed predecessor vertex $v$ as the maximum of its previous value and $H.getMin()$. The method processEdge3 takes time $O(max(\log n,\Delta))$ where $\Delta$ is the maximum in-degree of any vertex. We now show that $SP_{3}$ generalizes $SP_{2}$ (which, in turn, generalizes $SP_{1}$). ###### Theorem 4. Any vertex marked fixed by $SP_{2}$ in any iteration is also fixed by $SP_{3}$ in that iteration or earlier. ###### Proof. $SP_{2}$ fixes a vertex when $pred[k]$ equals zero, or when $D[k]\leq d+inWeight[k]$. When $pred[k]$ equals zero, all the predecessors of $v_{k}$ are fixed and their $C$ value matches their $D$ value. Therefore, $C[k]:=\max(C[k],\min\\{C[v]+w[v,k]~{}|~{}((v,k)\in E)\\})$ guarantees that $C[k]\geq\min\\{D[v]+w[v,k]~{}|~{}((v,k)\in E)=D[k]$. Therefore, vertex $k$ is marked as fixed. Now suppose that $D[k]\leq d+inWeight[k]$ in $SP_{2}$. Let $z$ be the vertex that discovered $k$ in $SP_{2}$. Then $D[k]\leq d+inWeight[k]$ implies $D[k]\leq\min(D[k],d+inWeight[k])$. Since $D[k]\leq D[z]+w[z,k]$, we get that $D[k]\leq\min((D(z)+w[z,k]),d+inWeight[k]$. From the definition of $inWeight[k]$, we get that $D[k]\leq\min((D(z)+w[z,k]),d+\min\\{w[v,k]~{}|~{}(v,k)\in E\wedge(v\neq z)\\})$. Since $z$ is a fixed vertex, we get $D[k]\leq\min((C(z)+w[z,k]),\min\\{d+w[v,k]~{}|~{}(v,k)\in E\wedge(v\neq z)\\})$. Since $C[v]$ for all predecessors of $k$ is set to at least $d$ in step 2 of $SP_{3}$, we get that $D[k]\leq\min((C(z)+w[z,k]),\min\\{C[v]+w[v,k]~{}|~{}(v,k)\in E\wedge(v\neq z)\\})$. By combining two arguments of the $\min$, we get $D[k]\leq\min\\{C[v]+w[v,k]~{}|~{}(v,k)\in E\\}$. The right hand side is $C[k]$ due to assignment of $C[k]$ at step 4. Since $D[k]\leq C[k]$, we get that vertex $k$ is marked as fixed. ∎ We now show that any vertex marked fixed by out-version or in-version of [4] is also marked fixed by $SP_{3}$. ###### Lemma 9. (a) $SP_{3}$ fixes any vertex $k$ such that $D[k]\leq\min\\{D[x]+w[x,y]~{}|~{}\neg fixed(x)\\}$. (b) $SP_{3}$ fixes any vertex $k$ such that $D[k]\leq\min\\{D[y]~{}|~{}\neg fixed(y)\\}+\min\\{w[v,k]~{}|~{}(v,k)\in E\\}$. ###### Proof. (a) follows from $threshold$ computed and marking of vertices as fixed based on that. (b) Suppose $D[k]\leq\min\\{D[y]~{}|~{}\neg fixed(y)\\}+\min\\{w[v,k]~{}|~{}(v,k)\in E\\}$. The first part of the sum is equal to $H.getMin()$ due to the property of $H$. Therefore, this expression is equal to $\min\\{H.getMin()+w[v,k]~{}|~{}(v,k)\in E\\}$. From Step 2 in $SP_{3}$, this expression is at most $\min\\{C(v)+w[v,k]~{}|~{}(v,k)\in E\\}$. From step 3, we get this expression to be at most $C[k]$. Therefore, $D[k]\leq C[k]$ and $k$ is fixed. ∎ The following Theorem summarizes properties of $SP_{3}$. ###### Theorem 5. Algorithm $SP_{3}$ computes the cost of the shortest path from the source vertex $v_{0}$ to all other vertices in $O(n+e(\max(\log n,\Delta)))$ time, where $\Delta$ is the maximum in-degree of any vertex. ## VI Algorithm $SP_{4}$: A Parallel Label-Correcting Algorithm In this section we present an algorithm when a large number of cores are available. The goal of the algorithm is to decrease the value of $D$ and increase the value of $C$ in as few iterations of the while loop as possible. All our earlier algorithms explore only fixed vertices with the motivation of avoiding multiple edge-relaxation of the same edge (in the spirit of Dijkstra’s algorithm). In contrast, $SP_{4}$ is a label-correcting algorithm that relaxes as many edges as possible in each iteration (in the spirit of Bellman-Ford algorithm). Similarly, it recomputes $C$ for as many vertices as possible and terminates faster than $SP_{3}$. var $D$: array[$0\ldots n-1$] of integer initially $\forall i:D[i]:=\infty$; $fixed$: array[$0\ldots n-1$] of boolean initially $\forall i:fixed[i]:=false$; $C$: array[$0\ldots n-1$] of integer initially $\forall i:C[i]=0$; $outWeight$: array[$0\ldots n-1$] of integer initially $\forall i:outWeight[i]=\min\\{w[i,j]~{}|~{}(i,j)\in E\\}$; $Dout$: array[$0\ldots n-1$] of integer initially $\forall i:Dout[i]=\infty$; int $threshold$; int $minD$; $D[0]:=0;$ $Dout[0]:=D[0]+outWeight[0];$ boolean $changed:=true;$ while ($changed\wedge(\exists i:\neg fixed[i]\wedge(D[i]<\infty)$) $changed:=false;$ // Step 1: find the minimum value of $D$ and $Dout[x]$ $threshold$ := min $Dout[x]$ of all non-fixed vertices; $minD$ := min $D[x]$ of all non-fixed vertices; // Step 2: Fix all vertices with $D[x]\leq threshold$ forall $x$ such that $(D[x]\leq threshold)$ in parallel $fixed[j]:=true;$ $C[j]:=D[j];$ // Step 3: Update $D$ values forall $x,y$ such that $(D[x]<\infty)$ $\wedge\neg fixed[y]\wedge((x,y)\in E)$ in parallel if $(D[y]>D[x]+w[x,y])$ then $D[y]:=D[x]+w[x,y]$; $Dout[y]:=D[y]+outWeight[y]$; $changed:=true$; // Step 4: Update $C$ values forall $y$ such that $\neg fixed[y]$ in parallel $C[y]:=\max(C[y],minD)$; forall $y$ such that $\neg fixed[y]$ in parallel $C[y]:=\max(C[y],$ $\min\\{C[x]+w[x,y],(x,y)\in E)\\})$; // Step 5: Update $fixed$ values forall $y:\neg fixed[y]\wedge(D[y]<\infty)$ in parallel if ($C[y]=D[y]$) $fixed[y]:=true$; endwhile; Figure 6: Algorithm $SP_{4}$: A Bellman-Ford Style Algorithm with both upper and lower bounds The algorithm is shown in Fig. 6. We use an outer while loop that is executed so long as $changed\wedge(\exists i:\neg fixed[i]\wedge(D[i]<\infty)$. The variable $changed$ is used to record if any vertex changed its $D$ value. This is a well-known optimization of Ford-Bellman algorithm for early termination. If $D$ did not change in the last iteration of the while loop, we have reached the fixed point for $D$ and it is equal to $cost$. Even if $D$ changed for some vertices but all vertices are fixed, then their $D$ values cannot change and we can terminate the algorithm. The conjunct $(D[i]<\infty)$ allows us to restrict the algorithm to examine only discovered vertices. In step 1, we find $threshold$ equal to the minimum of all $Dout$ values of non-fixed vertices just as in $SP_{3}$. We also find $minD$ equal to the minimum of all $D$ values for non-fixed vertices. With $n$ processors this step can be done in $O(\log\log n)$ time and $O(n)$ work on a common-CRCW PRAM with the standard technique of using a doubly logarithmic tree and cascading[10]. In step 2, we fix all the vertices that have $D$ values less than or equal to the threshold. This step can be done in $O(1)$ time and $O(n)$ work. In step 3, we first explore all the discovered vertices. All vertices adjacent to these vertices become discovered if they have not been discovered earlier. In addition, we also relax all the incoming edges to vertices that are not fixed. Clearly, this is equivalent to relaxing all edges as in the Bellman-Ford algorithm because for fixed vertices their $D$ value cannot decrease. This step can be performed in $O(1)$ time and $O(e)$ work with $e$ cores. In step 4, we compute lower bounds for all non-fixed vertices. We first update $C$ for all non-fixed vertices to be at least as large as $minD$. In the second parallel step, we simply apply Eqn. 1 to update all $C$’s for all non-fixed vertices. This step can also be performed in $O(1)$ time and $O(e)$ work with $e$ cores. In step 5, we recompute the array $fixed$ based on $C$ and $D$. This step can be performed in $O(1)$ time and $O(n)$ work. The total number of iterations is at most $n$ giving us the parallel time complexity of $O(n\log\log n)$ and work complexity of $O(ne)$. The number of iterations in $SP_{4}$ algorithm is less than or equal to the number of iteration required by $SP_{3}$. We now show the following property of $SP_{4}$. ###### Theorem 6. Algorithm $SP_{4}$ computes the cost of the shortest path from the source vertex $v_{0}$ to all other vertices in time $O(n\log\log n)$ and work $O(ne)$ with $e$ processors. ###### Proof. We first show the correctness of $SP_{4}$. It is sufficient to show that the while loop maintains the invariant that $D[x]$ is an upper bound and $C[x]$ is a lower bound on the cost to the vertex $x$. Steps 1 and 2 correctly maintain $D$ follows from [4]. Step 3 is the standard Bellman-Ford rule and it correctly maintains $D$. Step 4 correctly maintains $C$ due to Lemma 6. Step 4, simply maintains the invariant that $fixed(x)\equiv(C[x]=D[x])$. The time and work complexity follows from the earlier discussion. ∎ ## VII Conclusions and Future Work In this paper, we have presented four algorithms for the shortest path problem. We present algorithms $SP_{1}$ and $SP_{2}$ that reduce the number of heap operations required by Dijkstra’s algorithm and allow exploration of multiple vertices in parallel thereby reducing its sequential bottleneck. We also present algorithms $SP_{3}$ and $SP_{4}$ that require more work than Dijkstra’s algorithm but reduce the sequential bottleneck even further. These algorithms are the first ones that exploit both upper and lower bounds on the cost of the shortest path to increase the number of vertices that can be explored in parallel. Extending these algorithms for distributed shared memory is a future research direction. ## References * [1] Ravindra K Ahuja, Kurt Mehlhorn, James Orlin, and Robert E Tarjan. Faster algorithms for the shortest path problem. Journal of the ACM (JACM), 37(2):213–223, 1990. * [2] Richard Bellman. On a routing problem. Quarterly of applied mathematics, 16(1):87–90, 1958. * [3] Venkatesan T Chakaravarthy, Fabio Checconi, Prakash Murali, Fabrizio Petrini, and Yogish Sabharwal. Scalable single source shortest path algorithms for massively parallel systems. IEEE Transactions on Parallel and Distributed Systems, 28(7):2031–2045, 2017. * [4] Andreas Crauser, Kurt Mehlhorn, Ulrich Meyer, and Peter Sanders. A parallelization of dijkstra’s shortest path algorithm. In International Symposium on Mathematical Foundations of Computer Science, pages 722–731. Springer, 1998. * [5] E. W. Dijkstra. A note on two problems in connexion with graphs. Numerische Mathematik, 1(1):269–271, Dec 1959. URL: https://doi.org/10.1007/BF01386390, doi:10.1007/BF01386390. * [6] Robert W Floyd. Algorithm 97: shortest path. Communications of the ACM, 5(6):345, 1962. * [7] L. A. Ford. Network flow theory. Technical report, 1956. * [8] Michael L. Fredman and Robert Endre Tarjan. Fibonacci heaps and their uses in improved network optimization algorithms. J. ACM, 34(3):596–615, July 1987. URL: http://doi.acm.org/10.1145/28869.28874, doi:10.1145/28869.28874. * [9] P. E. Hart, N. J. Nilsson, and B. Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics, 4(2):100–107, July 1968. doi:10.1109/TSSC.1968.300136. * [10] Joseph JáJá. An introduction to parallel algorithms, volume 17. Addison-Wesley Reading, 1992. * [11] Kamesh Madduri, David A Bader, Jonathan W Berry, and Joseph R Crobak. An experimental study of a parallel shortest path algorithm for solving large-scale graph instances. In 2007 Proceedings of the Ninth Workshop on Algorithm Engineering and Experiments (ALENEX), pages 23–35. SIAM, 2007. * [12] Ulrich Meyer and Peter Sanders. $\delta$-stepping: a parallelizable shortest path algorithm. Journal of Algorithms, 49(1):114–152, 2003. * [13] Rajeev Raman. Recent results on the single-source shortest paths problem. SIGACT News, 28(2):81–87, June 1997. URL: http://doi.acm.org/10.1145/261342.261352, doi:10.1145/261342.261352. * [14] Mikkel Thorup. On ram priority queues. SIAM Journal on Computing, 30(1):86–109, 2000.
11institutetext: Sentic Lab, Iasi, Romania 22institutetext: Faculty of Computer Science, ”Alexandru Ioan Cuza” University of Iasi, Romania 22email<EMAIL_ADDRESS> # Analyzing domain shift when using additional data for the MICCAI KiTS23 Challenge George Stoica ✉ 1122 Mihaela Breaban 22 Vlad Barbu 11 ###### Abstract Using additional training data is known to improve the results, especially for medical image 3D segmentation where there is a lack of training material and the model needs to generalize well from few available data. However, the new data could have been acquired using other instruments and preprocessed such its distribution is significantly different from the original training data. Therefore, we study techniques which ameliorate domain shift during training so that the additional data becomes better usable for preprocessing and training together with the original data. Our results show that transforming the additional data using histogram matching has better results than using simple normalization. ###### Keywords: 3D Segmentation Domain Shift Domain Adaptation. ## 1 Introduction The segmentation of renal structures (kidney, tumor, cyst) has gained interest in the recent years, starting from the KiTS19 Challenge [4] and continuing with KiTS21, KiPA22111 https://kipa22.grand-challenge.org/home/ and currently with KiTS23. The accurate segmentation of renal tumors and renal cysts is of important clinical significance and can benefit the clinicians in preoperative surgery planning. Deep learning leverages on huge amount of training data for learning domain specific knowledge which can be used for predicting on previously unseen data. In medical image segmentation, a smaller amount of training data is available when compared to other domains of deep learning. Therefore, using additional training data has a greater impact on the end results. In domain adaptation, the usual scenario entails learning from a source distribution and predicting on a different target distribution. The change in the distribution of the training dataset and the test dataset is called domain shift. In the case of supervised domain adaptation, labeled data from the target domain is available [9]. The medical image acquisition process is not uniform across different institutions and CT images may have different HU values and various amount of noise depending on the acquisition device, the acquisition time and other external factors. As a consequence, distribution shifts are easily encountered and this affects models that perform well on validation sets but encounter different data in practice. Creating a model which is robust to different types of distributions requires training on enough data, coming from all the target domains. When training under domain shift and using two datasets with different distributions, the data should be preprocessed in order to mitigate the data mismatch error which happens due to the distribution shift. Characteristics of the target dataset have to be incorporated into the training dataset, which could be done either by collecting more data from the target distribution, or by artificial data synthesis. As the amount of training data from the target distribution (KiTS23 challenge) is limited, our solution consists of transforming the additional data (taken from the KiPA22 challenge) to the target distribution. We compare two transformations, dataset normalization, which preserves the original but different distribution, with histogram matching, which makes both the source and target distribution the same. ## 2 Methods Our approach consists of applying initial preprocessing to an additional dataset which was used for training. The aim of the preprocessing was to reduce the distribution shift between the additional data and the target domain, and will be fully-detailed in Section 2.2. After bridging the distributions of the original and additional data, we preprocess and normalize the whole data together and train a 3D U-Net [1] using multiple data augmentation techniques. Ultimately, we evaluate our model on the validation set. ### 2.1 Training and Validation Data Our submission uses the official KiTS23 training set, built upon the training and testing data from the KiTS19 [5] and KiTS21 competitions. In addition to the official KiTS23 data, our submission made use of the public KiPA22 competition training set [3] [2] [7] [8]. The KiTS23 training dataset contains 489 CTs which include at least one kidney and tumor region and usually include both kidneys and optionally one or more cyst regions. In contrast, the KiPA22 training data contains only 70 CTs in which only the diseased kidney is selected. KiPA22 images have 4 segmentation targets: kidney, tumor, artery and vein. Unlike KiTS23, benign renal cysts are segmented as part of the kidney class for KiPA22. The initial preprocessing for the KiPA22 images consists of removing the artery and vein segmentation masks and keeping only the kidney and tumor class. We have used 342 random images from KiTS23 and 70 images from KiPA22 for training and 147 random images from KiTS23 for validation. ### 2.2 Preprocessing Initial exploratory data analysis illustrate the fact that images from KiPA22 have a totally different distribution than images from the KiTS23 training set on the HU scale. (a) Both datasets before preprocessing. (b) Shifting the mean and standard deviation. (c) Applying histogram matching. Figure 1: Histograms for both datasets before and after initial preprocessing. While the values for the KiTS23 CT images are mostly centered around - 1000 and 0 on the HU scale, KiPA22 images are situated between 800 and 1500 while also having a visible different distribution (Figure 1(a)). Training under domain shift using the original distribution for the second dataset is challenging, therefore we have taken steps towards ameliorating the effects of distribution shift. To mitigate the impact of the huge distance between the values of the vertices, the simplest solution is shifting the mean and standard deviation of KiPA images to match those of KiTS (Figure 1(b)). Nevertheless, the distributions are visibly different, therefore we also applied histogram matching to transform the KiPA images to the KiTS domain (Figure 1(c)). We choose to transform the KiPA images because the test data will be made of images which are expected to match the original KiTS distribution. When training under domain shift, only data from the target distribution should be used for validation. For the KiPA dataset, shifting the mean maintains the original shape of the curve, scaled by the factor which changed the standard deviation, spreading the values evenly. Histogram matching, on the other side, is destructive and the HU values of vertices are spread unevenly to match the KiTS distribution. To choose the best transformation, we have created two datasets to evaluate them separately in order to choose the more suitable one. 1. 1. Dataset 1: 342 images from KiTS and 70 images from KiPA whose values are shifted by changing the mean and standard deviation. 2. 2. Dataset 2: the same 342 images from KiTS and 70 images from KiPA whose values are transformed by histogram matching. For both datasets, the same preprocessing steps are applied, using the nnUNet framework [6]. Values are clipped at the 0.5th and 99.5th percentile to remove outliers. Then, images are normalized to have the mean 0 and standard deviation 1 and three order-interpolation is used to resample all images into a space of $0.7636\times 0.7636\times 0.7636$ $mm^{3}$. ### 2.3 Proposed Method After preprocessing each dataset, we have trained the model using the default nnUNet configuration for training, which uses a classic 3D U-Net. Instead of using an ensemble of 5 models trained on 5 folds of the data, we have opted to train a single model on all the training data available. We have used region based training, defining the 3 learning targets: Kidney & Tumor & Cyst, Tumor & Cyst and Tumor. While this approach directly optimizes the official evaluation metrics, it does not yield good results for predicting the cyst class. We have experimented with Dice & Focal Loss and Dice & Cross Entropy Loss, the latter achieving the best results. We have trained the model using a patch size of $128\times 128\times 128$ and a batch size of two, for 1000 epochs. We started the training using SGD and Nesterov momentum with a learning rate of 0.01 and used a Polynomial Learning Rate Scheduler to decrease the learning rate evenly until it reaches a value of 0.001. To prevent overfitting, we applied multiple data augmentation techniques integrated in nnUNet: Rotation, Scaling, Gaussian noise, Gaussian blur, Random brightness, Gamma Correction and Mirroring. We did not apply any post-processing on the prediction. ## 3 Results We have trained on both Dataset 1 and Dataset 2 and have used 147 images from KiTS23 for validation. The results are displayed in Table 1. The official metrics used in competition are in italic, but we also report the Dice score for kidney and cyst segmentation. Table 1: Validation results for Dataset 1 and Dataset 2. Dataset | Dice kidney&masses | Dice masses | Dice kidney | Dice tumor | Dice cyst ---|---|---|---|---|--- Dataset 1 | 95.310904 | 79.072143 | 94.101086 | 76.891783 | 17.012944 Dataset 2 | 95.453839 | 80.760511 | 94.024749 | 78.960431 | 20.766421 Our experiments show that applying histogram equalization (Dataset 2) on the additional dataset improves the results for all the target metrics. Using simple normalization (Dataset 1) has better results only when calculating the Dice score for the kidney area. However, the Dice score for tumors and cysts is worse by 2 and 3 percent. Using the original distribution of the KiPA dataset results in a lower Dice score for cysts also because possible cysts are labeled as the kidney class. Nonetheless, since the two distributions are still very different even after normalization and preprocessing, the scores are heavily impacted. Evaluating the results on both configurations, the model does not distinguish the cyst class and many cysts are classified as tumors. There is a class imbalance between cysts and tumors, as cysts generally encompass a smaller area. In our case, the low dice score for cysts is a result of many false positives. We presume that our learning target is the culprit, because we directly minimize the Dice and Cross Entropy loss for the whole segmentation area (kidney and masses), the lesion area (masses, both tumor and cyst) and ultimately, tumor. As a consequence, the cyst area is indirectly learnt, therefore the accuracy is lower. To make the model discriminate between the two classes and reduce the false positives, we suggest changing the learning target to directly minimize the Dice and Cross Entropy loss for the cyst class. For training and inference we have used a workstation with an RTX 3090 GPU, an AMD Ryzen Threadripper 2970WX 24-Core Processor CPU, SSD and 31GB RAM memory available. Training the model took around 3.4 days. For prediction, the average inference time was 10 minutes per case. ## 4 Discussion and Conclusion We have explored suitable techniques for mitigating distribution shift when using additional data for the kidney tumor 3D segmentation task. We identify histogram matching as an initial preprocessing step of artificial data synthesis that completely transforms the additional distribution to the target domain. Compared to simple normalization, this approach has the advantage of training only on the target distribution, which improves the results, especially for the least frequent classes, cyst and tumor. We believe that more stable results can be achieved by training an ensemble, and the discriminative power between cysts and tumors can be enhanced by changing the training target and using techniques that deal with class imbalance. ## References * [1] Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424–432. Springer (2016) * [2] He, Y., Yang, G., Yang, J., Chen, Y., Kong, Y., Wu, J., Tang, L., Zhu, X., Dillenseger, J.L., Shao, P., et al.: Dense biased networks with deep priori anatomy and hard region adaptation: Semi-supervised learning for fine renal artery segmentation. Medical image analysis 63, 101722 (2020) * [3] He, Y., Yang, G., Yang, J., Ge, R., Kong, Y., Zhu, X., Zhang, S., Shao, P., Shu, H., Dillenseger, J.L., et al.: Meta grayscale adaptive network for 3d integrated renal structures segmentation. Medical image analysis 71, 102055 (2021) * [4] Heller, N., Isensee, F., Maier-Hein, K.H., Hou, X., Xie, C., Li, F., Nan, Y., Mu, G., Lin, Z., Han, M., et al.: The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical Image Analysis 67, 101821 (2021) * [5] Heller, N., Sathianathen, N., Kalapara, A., Walczak, E., Moore, K., Kaluzniak, H., Rosenberg, J., Blake, P., Rengel, Z., Oestreich, M., et al.: The kits19 challenge data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and surgical outcomes. arXiv preprint arXiv:1904.00445 (2019) * [6] Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18(2), 203–211 (2021) * [7] Shao, P., Qin, C., Yin, C., Meng, X., Ju, X., Li, J., Lv, Q., Zhang, W., Xu, Z.: Laparoscopic partial nephrectomy with segmental renal artery clamping: technique and clinical outcomes. European urology 59(5), 849–855 (2011) * [8] Shao, P., Tang, L., Li, P., Xu, Y., Qin, C., Cao, Q., Ju, X., Meng, X., Lv, Q., Li, J., et al.: Precise segmental renal artery clamping under the guidance of dual-source computed tomography angiography during laparoscopic partial nephrectomy. European urology 62(6), 1001–1008 (2012) * [9] Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135–153 (2018)
11institutetext: School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China 22institutetext: School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, University Town of Shenzhen, Nanshan, 510085 Shenzhen, China # A Subword Guided Neural Word Segmentation Model for Sindhi Wazir Ali 11 Jay Kumar 11 Zenglin Xu 1122 Congjian Luo 1122 Junyu Lu 11 Junming Shao 11 Rajesh Kumar 11 Yazhou Ren 11 ###### Abstract Deep neural networks employ multiple processing layers for learning text representations to alleviate the burden of manual feature engineering in Natural Language Processing (NLP). Such text representations are widely used to extract features from unlabeled data. The word segmentation is a fundamental and inevitable prerequisite for many languages. Sindhi is an under-resourced language, whose segmentation is challenging as it exhibits space omission, space insertion issues, and lacks the labeled corpus for segmentation. In this paper, we investigate supervised Sindhi Word Segmentation (SWS) using unlabeled data with a Subword Guided Neural Word Segmenter (SGNWS) for Sindhi. In order to learn text representations, we incorporate subword representations to recurrent neural architecture to capture word information at morphemic-level, which takes advantage of Bidirectional Long-Short Term Memory (BiLSTM), self-attention mechanism, and Conditional Random Field (CRF). Our proposed SGNWS model achieves an F1 value of 98.51% without relying on feature engineering. The empirical results demonstrate the benefits of the proposed model over the existing Sindhi word segmenters. ###### Keywords: Recurrent neural networks sequence tagging Sindhi word segmentation subword representation learning. ## 1 Introduction Word segmentation is a fundamental and challenging task in text classification and other NLP applications [6]. Word segmenter determines the boundaries of words in the shape of beginning and ending [11]. It has been largely investigated in many space-delimited languages including English [7], Arabic [4], Urdu [46] and non-space delimited languages including Chinese [45], Japanese [17], and Burmese [43]. However, the word segmentation in low- resource Sindhi language has not been studied well [16], mainly due to the lack of language resources. Sindhi word segmentation exhibits the space omission and space insertion [6, 24] problems. Although, the white spaces between words are a good sign for predicting word boundaries, the space omission and space insertion between words bring ambiguity in the segmentation process. Therefore, the SWS task is a challenging problem because of resource scarcity, lack of standard segmentation benchmark corpus, and rich morphological features [6, 25] in Sindhi language. Previously, little work has been proposed to address the SWS problem by employing dictionary-based [6] and rule-based [25, 24, 29, 12] approaches. Thus, the existing approaches lack the applicability towards open- source implementation due to following reasons, (i) inability to deal with out-of-vocabulary words, (ii) less robust on the large datasets, and (iii) lower segmentation accuracy. Our proposed novel deep SGNWS model has the capability of dealing with such issues for SWS with the Subword Representation Learning (SRL) approach. Recently, deep neural architectures have largely gained popularity in NLP community [42] by greatly simplifying the learning and decoding in a number of NLP applications [10, 23, 39, 40] including word segmentation [4, 35] with neural word embedding [8, 21] and powerful recurrent neural architectures [14, 34]. More recently, self-attention [37] has also become a popular approach to boost the performance of neural models. Therefore, we tackle the SWS problem by taking advantage of BiLSTM, self-attention, SRL, and CRF without relying on external feature engineering. In this paper, we propose a language-independent neural word segmentation model for Sindhi. The proposed model efficiently captures the character-level information with subword representation learning. We convert segmentation into a sequence tagging problem using B, I, E, S, X tagging scheme. Where B denotes [Beginning], I [Inside], E [Ending] of a word in the given corpus, S [Single] is used for the tagging of a single or special character in the unlabeled text, and X tag is used for [hard-space] between words. We train task-oriented [21] Sindhi word representations with character-level subword approach [18]. To the best of our knowledge, this is the first attempt to tackle SWS as a sequence labeling task. We provide the open-source implementation for further investigation111https://github.com/AliWazir/Neural-Sindhi-word-segmenter. Our novel contributions are listed as follows: * • To the best of our knowledge, we are the first to propose a sequence modeling based language-independent neural model to tackle the SWS problem. * • The proposed model eliminates the constraint of external feature engineering by adopting subword representation learning. * • We treat SWS as a sequence tagging problem by assigning the B, I, E, S, X tags to unlabelled corpus for the word boundary detection. * • Extensive experiments prove the dominant performance of our proposed model compared with the baselines approaches. ## 2 Related Work Recurrent Neural Network (RNN) variants has been widely adopted in a number of learning tasks [41, 30, 20, 38] including sequence tagging problems [15, 23, 42] since the inception of well-known LSTM network [14]. However, LSTM suffers from a limitation to encode the given sequence in unidirectional way. This limitation has been handled by stacking two LSTM networks as a bidirectional encoder, known as BiLSTM [34] by the integration of the simultaneous training strategy in forward and backward directions. It is an ideal solution to learn the sequences in a language because, unlike unidirectional, the bidirectional network is beneficial to access both the contexts of right and left directions. The bidirectional RNN variants have been largely employed for word segmentation [39, 40, 22] in Chinese [9, 36], Japanese [17] and Arabic [4] by achieving excellent performance without relying on any external feature engineering strategies. On the one hand, state-of-the-art sequence tagging systems rely on large amounts of task-specific knowledge [23] in the form of hand-engineered features and data pre-processing. On the other hand, the performance of neural models can be enhanced by incorporating unsupervised neural embeddings including classical [26], character-level [8], deep contextualized [2] and task-oriented [21]. Moreover, the success of deep neural architectures also relies on the optimal hyper-parameters selection [33]. More recently, an attention mechanism [37] in neural models has also yielded new state-of-the- art results in multiple NLP tasks. Furthermore, the last layer of neural models has a significant impact on performance. The CRF [19] is broadly used in the sequence classification tasks [17, 23, 15, 9] for decoding in neural models. Taking advantage of language-independent neural models for SWS, we propose a model that efficiently captures the character-level information with subword representation learning by converting the segmentation into a sequence tagging problem. Presently, Sindhi language is being written in two famous writing system of Persion-Arabic and Devanagari [27]. However, Persian-Arabic is standard script [24] as well as frequently used in online communication, literary work, and journalism. Sindhi contains rich morphology due to the frequent usage of prefixes and suffixes to express inflections and derivations, which makes it a complex morphological language. Initially, the SWS was coined [25] by introducing the first word segmentation model using several rule-based algorithms. The proposed model was evaluated on a small dataset consists of 16,601 lexicon with cumulative segmentation error rate (SER) of 9.54$\%$. Later, [24] proposed a rule-based word tokenizer with 91.76$\%$ segmentation accuracy. The segmentation is performed in three steps; the first step consists of input and segmentation with white space. The second step is used for the segmentation of simple and compound words, while the third step deals with the segmentation of complex words. In this way, different word types are separately segmented in their proposed model. Moreover, [6] proposed a word segmenter by evaluating the dataset of 1,57,509 words obtained from news corpus and dictionary lexicon. Their proposed model achieves good performance dictionary lexicon, but poorly performed in dealing with news and books corpus. Recently, [12] proposed two algorithms for stemming and lemmatization process with an opensource222https://sindhinlp.com/stemlema.php implementation. The SWS is a challenging task because it exhibits space omission and space insertion problems. This is partly because of the Arabic script, which, although cursive in nature, consists of characters that have inherent joining and non-joining attributes regardless of a word boundary. Apart from the discussed problems, there is no gold-standard benchmark corpus for Sindhi to evaluate the segmentation task. In summary, the SWS task is difficult, important, and not-studied as a sequence modeling problem. The previous approaches mainly rely on the rule-based and dictionary-based methods, which have certain limitations such as inability to deal with out of vocabulary words, less robustness for other languages, and the algorithms’ inefficiency to deal with a large amount of noisy or raw text. ## 3 Sindhi Morphology The Persian-Arabic is a standard writing script for Sindhi, which is cursive and written from right to left direction [28, 32]. It contains rich morphology [24] due to the frequent usage of prefixes and suffixes to express inflections and derivations, which makes it a complex morphological language. The alphabet of Sindhi Persian-Arabic consists of 52 basic letters, 29 derived from Arabic language, 03 from the Persian language, and 20 modified letters [16]. It also uses 03 secondary letters, 07 honorific symbols and diacritic marks [28, 32]. Interestingly, the shape of some letters in Sindhi change the form according to their position in a word [6], such letters are referred as joiners. Thus, a joiner have at most four shapes; i) initial ii) middle iii) final and iv) isolated, as Table 1 depicts an example of some letters. Whereas the position- independent letters having final or isolated form are referred as non-joiners. Specifically, white spaces are used to detect word boundaries in Sindhi. However, writers omit a hard space between two words. Therefore, a phrase or a sentence that ends with non-joiner letters becomes one token. In the first case, the words are joined with their preceding and succeeding words in the absence of white space, which leads to misspellings. In the second case, the shape of characters remains identical even in the absence of white space. Due to position-independent and space-independent letters, the SWS exhibits both challenges of space insertion and space omission [25, 32]. Table 1: Various shapes of Sindhi alphabet according to their position in words. Roman transliteration of every isolated letter is given for the ease of reading. Ending | Middle | Initial | Isolated | Roman ---|---|---|---|--- | | | | B$\bar{e}$ | | | | J$\bar{i}$m | | | | S$\bar{i}$n | | | | $\breve{g}$ain | | | | G$\bar{a}$f | | | | N$\breve{u}$n Table 2: Complete list of Sindhi joiner and non-joiner letters, (i) denote joiner letters (ii) non-joiners, and (iii) non-joiner secondary letters. ### 3.1 Space Omission The space omission is a common phenomenon in Sindhi words that end with the non-joiner letters. However, the absence of white space exhibit the correct shape of words such as Table 3 shows an example of a Sindhi sentence with and without the use of white space. But computationally, that sentence consists of one token without the use of white spaces between words. Whereas the sentence consists of eight tokens with the use of white space between words. Therefore, the omission of white space between words ending with non-joiner letters raises a computational issue. ### 3.2 Space Insertion Another challenge in SWS arises when combining two or more root words (morphemes) form a new standalone single word (see Table 4). In such cases, writers omit white space if the first morpheme ends with a joiner letter. However, white space prevents it’s joining with the next morpheme so that the word retains a valid visual form. The missing space insertion leads to the formation of compound words and often misspelling. Hence, white space is essential in this case for the ease of readability and correct spelling of Sindhi words. Table 3: An example of a Sindhi sentence, all words end with the non-joiner letters. (i) denote the words with white space (the tokens are separated with ‘-’ symbol), (ii) without white space (iii) Roman transliteration of Sindhi sentence (iv) is the English translation of a Sindhi sentence. Table 4: Sindhi word types with an example of space insertion, along with English translation. (i) represent the words with white space (‘-’ symbol represents space), and (ii) without space. The Roman transliteration is given for ease of reading. Word Type | i. | ii. | Roman | English Translation ---|---|---|---|--- Affix | | | Be-hisaab | Uncountable Reduplicate | | | haidai hodai | Here and there Compound | | | saahib-e-Qudrat | Powerful Borrowed | | | Mobile phone | Mobile Phone Abbreviation | | | Ain Cee Aich Dee | NCHD ## 4 Methodology In this section, the proposed methodology is described in detail. Firstly, we convert the segmentation as a classification problem by introducing the proposed B, I, E, S, X tagging scheme. The labels are assigned to each Sindhi character, including punctuation marks and numbers in the dataset. Afterwards, we describe the baseline as well as the proposed SGNWS models. Later, we present the experimental details and the variants of neural models, including word representations, character-level SRL to predict subwords boundaries. ### 4.1 Tagging Scheme We modeled the word segmentation as character-level sequence labelling [9]. Theoretically, word boundary can be predicted with binary classification in word segmentation, but in practice, fine-grained tag sets [44] produce high segmentation accuracy. Following the work [36], we employ four tags [B, I, E, S] to indicate the position of letters at the Beginning [B], Inside [I], Ending [E] of a word, or a Single-character/symbol [S], respectively. Additionally, [X] is used to represent the white space to delimit word boundaries. A sentence, as an example of the proposed tagging scheme is depicted in Table 5 by assigning the proposed tags to a sentence. Table 5: An example of employed character-level sequence tagging scheme for SWS task. The [X] label represents the white spaces. The given Sindhi sentence can be read from right to left, and the Roman transliteration of each Sindhi token can be read from left to right. ### 4.2 Recurrent Neural Architectures #### 4.2.1 Long-Short-Term-Memory Unit: The LSTM network [14] is an extension of RNN proposed to solve vanishing and exploding gradient problems. For a given input $x_{t}$ of a sentence $S=\left[x_{1},x_{2},x_{3}\dots x_{n}\right]$, each word is represented into $N-$dimensional vector (word representation). As we mentioned earlier that Sindhi is being written from the right-to-left direction. Thus, an LSTM network computes each representation $\overleftarrow{h}_{t}$ of the right context of the given input at each time-step $t$. The memory unit $c$ allows the network to learn when to forget the previous information and when to update memory cells given new information. The core LSTM architecture contains forget $f$, input $i$, and output $o$ gates, which regulate the information to flow-in and flow-out at current state $t$. The mathematical representation of the gates, cell update, and output in LSTM is as follows: $\displaystyle i^{t}$ $\displaystyle=\sigma\left({W}_{i}{h}_{t-1}+{U}_{i}{x}_{t}+{b}_{i}\right)$ (1) $\displaystyle f^{t}$ $\displaystyle=\sigma\left({W}_{f}{h}_{t-1}+{U}_{f}{x}_{t}+{b}_{f}\right)$ $\displaystyle\tilde{c}^{t}$ $\displaystyle=\tanh\left({W}_{c}{h}_{t}+{U}_{c}{x}_{t}+{b}_{c}\right)$ $\displaystyle c^{t}$ $\displaystyle={f}^{t}\odot{c}^{t}+{i}^{t}\odot\tilde{{c}}^{t}$ $\displaystyle{o}^{t}$ $\displaystyle=\sigma\left({W}_{o}{h}_{t-1}+{U}_{o}{x}_{t}+{b}_{o}\right)$ $\displaystyle{h}^{t}$ $\displaystyle={o}^{t}\odot\tanh\left({c}^{t}\right)$ where $\sigma$ and $\odot$ are the element-wises sigmoid function and element- wise product, $U,W,b$ denote the input $x_{t}$ weight matrix, hidden $h_{t}$ weight matrix, and bias vector for each LSTM gate, respectively. The core model is a memory cell $c$ which encodes long-term temporal dependencies of observed inputs at every time-step $t$. #### 4.2.2 Bidirectional Long-Short-Term-Memory (BiLSTM): The BiLSTM model encodes the text sequences from both left and right directions into two separate forward $\overrightarrow{h}$ and backward $\overleftarrow{h}$ hidden states to capture the right and left context information. Afterwards, both hidden states $\overrightarrow{h}$, $\overleftarrow{h}$ are concatenated for the final output. However, the LSTM hidden state ${h_{t}}$ can only encode context of one direction, such as the left direction, knowing nothing about the right direction. The BiLSTM first computes the forward $\overrightarrow{h}$ and then backward $\overleftarrow{h}$ hidden states of a given input $x_{t}$. Afterwards, both $\overrightarrow{h}$ and $\overleftarrow{h}$ are combined to generate output $y_{t}$. This process can be expressed as follows: $\displaystyle{\overrightarrow{h_{t}}=H\left(W_{\mathrm{x}\overrightarrow{h}}x_{t}+\left(W_{\overrightarrow{h}}\overrightarrow{h}_{t-1}+b_{\overrightarrow{h}}\right)\right.}$ (2) $\displaystyle{\overleftarrow{h_{t}}=H\left(W_{{x}\overleftarrow{h}}x_{t}+\left(W_{\overleftarrow{h}}\overleftarrow{h}_{t+1}+b_{\overleftarrow{h}}\right)\right.}$ $\displaystyle{y}_{t}={W}_{\overrightarrow{h}{y}}\overrightarrow{h}_{t}+W_{\overleftarrow{h}{y}}\overleftarrow{h}_{t}+b_{t}$ where, $H$ is a concatenation of the corresponding hidden outputs of both forward $\overrightarrow{h}{y}$, and backward $\overleftarrow{h}{y}$ LSTM cells. ### 4.3 Tag Inference The CRF is an effective approach for sequence tagging problems [19] because it learns the scoring function from tag pairs, such as B, I, E, S at the training stage. Thus, it is beneficial for sequence tagging tasks by considering the correlation between the corresponding neighbour tags [23] as well as efficiently decodes the best chain of tags of a given input sequence. The probability of a possible tag sequences in CRF can be formulated as: $P(Y|X)=\frac{\left.\prod_{i=2}^{n}\exp\left(s(X,i)_{yi}+b_{yi-1yi}\right)\right)}{\left.\sum_{y^{\prime}}\prod_{i=2}^{n}\exp\left(s(X,i)_{y_{i}}^{\prime}+b_{y_{i-1}}^{\prime}y_{i}^{\prime}\right)\right)}$ (3) where $y\in\\{B,I,E,S\\}$ tags, scoring function $s\left(X,i\right)_{yi}$ is an output of the hidden layer at $i_{th}$ word, and $b_{yi-1yi}$ are the trainable parameters. While decoding is the search for tag sequences $y$ with highest conditional probability. Thus, by solving the Eq. (4), we obtain optimal tag sequence: $Y^{*}=\operatorname{argmax}P(Y|X)$ (4) ### 4.4 Subword Representation Learning We use BiLSTM network [14] for SRL by representing each word $w$ from a fixed vocabulary $V$ of unlabeled Sindhi text in a sequence of forward and backward character representations. Such as, character representations $E^{c}=\left[c_{1},c_{2},c_{3},\dots,c_{i}\right]$, bigrams ${E^{B}}=\left[c_{i},c_{i+1}\right]$, and trigrams ${E^{T}}=\left[c_{i},c_{i+1},c_{i+2}\right]$ of a given word are learned to capture the structure of words at morphemic level. Afterwards, we utilize both forward and backward representations by concatenating them: $\displaystyle\overrightarrow{h}{{}_{t}}=\text{LSTM}\left(E^{C_{i}}:E^{B_{i}}:E^{T_{i}},\overrightarrow{h}{{}_{t-1}}\right),$ (5) $\displaystyle\overleftarrow{h}{{}_{t}}=\text{LSTM}\left(E^{C_{i}}:E^{B_{i}}:E^{T_{i}},\overleftarrow{h}{{}_{t+1}}\right),$ $\displaystyle\text{BiLSTM}\left({Emb_{S}}\right)=\overrightarrow{h}{{}_{|w|}}:\overleftarrow{h}{{}_{1}},$ where ${Emb_{s}}$ is the concatenated output of Bidirectional $\overrightarrow{h}{{}_{|w|}}$, $\overleftarrow{h}{{}_{1}}$ representations of LSTM layers over the sequence of character n-grams. Table 6: An example of Sindhi subword decomposition for subword representation learning ### 4.5 Proposed Model The proposed SGNWS architecture consists of five layers. We explain sequential processing of each layer as follows: * • Input layer: The model takes character-level input $x{{}_{t}}={c_{1},c_{2},c_{3},\dots c_{i}}$ of character unigrams $c_{i}$, character bigrams $c_{i},c_{i+1}$, character trigrams $c_{i},c_{i+1},c_{i+2}$, and 4-grams of each word words $w_{n}$ for SRL, as depicted in Table 6. * • Embedding layer: After the character-level input, we learn bidirectional unigram $E_{c}$, bigram ${E_{c}^{B}}$, trigram ${E_{c}^{T}}$, and 4-gram representations of the given input words $w_{n}$, other numerical features and then concatenate them into subword embeddings as formulated in Eq.(5). In the next step, embeddings are used as an input to the proposed model after passing through a non-linear bidirectional layer. * • LSTM layers: we utilize forward $\overrightarrow{h}$ and backward hidden $\overleftarrow{h}$ layers of BiLSTM to obtain high-level features from embedding layer. The n-gram based subword representations are passed through the $\overrightarrow{h}$ and $\overleftarrow{h}$ layers. * • Hidden layer: The Bidirectional output of the forward and backward hidden layers is concatenated with a hidden layer before the input to the CRF layer. * • Self-attention layer: We add a self-attention layer before the CRF classifier, which has the ability to decide how much information to use from token-level components dynamically. * • Output layer: Finally, the CRF layer is placed on the last hidden layer of proposed model to incorporate transition information between succeeding tag sequences to obtain optimal tag sequences over the entire sentence. In this way, CRF decodes the best chain of tags $Y^{*}$ of given input sequences as represented in Eq. (4). ## 5 Experimental Setup This section provides details about the experimental setting of baseline models as well as proposed SGNWS architecture. We use tenserflow [1] deep learning framework for the implementation of all neural models on GTX 1080-TITAN GPU to conduct all the experiments. ### 5.1 Dataset We utilize the recently proposed unlabeled Sindhi text corpus [3] in the experimental setting. We convert segmentation into a sequence tagging problem using B, I, E, S, X tagging scheme and split the dataset into 80% for training and 20% for development and test sets. The complete statistics of the dataset is given in Table 7. We split each sentence with punctuation marks of period, comma, question mark, colon, semicolon, exclamation mark, dash for consistency in the dataset and do not consider sentences having tokens less than 5. Moreover, we split the large sentences with white-space if the length exceeds more than 300 tokens. The regular hard-space is tagged as [X] in the dataset. However, multi-word tokens such as numerical expressions 689.0967, date 25-06-2020, money 4736$, etc., are assigned continuous tags. For example, date 25-06-2020 is assigned continuous tags of BIIIIIIIIE, respectively. Table 7: Statistics of the proposed unlabelled datasest used in the experiments. We concatenate all the datasets and represent it as SDSEG in general experiments. Domain | Sentences | Tokens | Unique | Average ---|---|---|---|--- | | | words | word length Kawish news paper | 24,212 | 601,910 | 10,721 | 3.687 Awami-Awaz news paper | 19,736 | 521,257 | 14,690 | 3.660 Wikipedia-dumps | 14,557 | 669,623 | 11,820 | 3.738 Twitter | 10,752 | 159,130 | 17,379 | 3.820 Books | 22,496 | 430,923 | 16,127 | 3.684 Total | 91,753 | 2,382,843 | 70,737 | 3.717 ### 5.2 Baseline Models To analyze and compare the performance of proposed model, we conduct several baseline experiments by training by training the LSTM, BiLSTM, and B-LSTM-CRF. We train task-specific [21] character-level word representations in baseline experiments. The brief description of each approach is defined as follows: 1. 1. LSTM: Our first baseline is the LSTM network exploited with character-level word representations using task-oriented strategy [21]. We use a softmax classifier in the last layer of the network for the decoding of tag sequences. 2. 2. BiLSTM: The BiLSTM has the advantage of encoding forward and backward sequences to efficiently capture the word information at the morphemic level. Similar to the LSTM network, we also use softmax in the last layer of the network for decoding. 3. 3. BiLSTM-CRF: The third baseline model is based on a BiLSTM-CRF network with a similar hyper-parameter setting as LSTM and BiLSTM networks. We use CRF inference in the last layer of the network for decoding purpose. All the hyper-parameters for baseline models and SGNWS are kept similar (see Table 8) for performance difference and fair comparison. ### 5.3 Evaluation metrics We report word boundary Precision, Recall and F-Score as illustrated in Eq. (6)-(7)-(8). The Precision evaluates the percentage of correctly predicted tags with respect to the predicted boundaries, and Recall measures the percentage of correctly predicted tags with respect to the true boundaries. While $F-\text{Score}$ is the harmonic mean of Precision and Recall, which can be interpreted as their weighted average. ${\text{Precision}}=\frac{\text{\\#(correctly\\_predicted\\_ tags)}}{\text{\\#(predicted tags)}}\\\ $ (6) ${\text{Recall}}=\frac{\text{\\#(correctly\\_predicted\\_tags)}}{\text{\\#(true\\_tags)}}\\\ $ (7) $F\text{-Score}=\frac{2\times{\text{Precision}}\times{\text{Recall}}}{{\text{Precision}}+{\text{Recall}}}\\\ $ (8) ### 5.4 Parameter setting and training The training procedure is to regulate all parameters of the network from training data. We train the baselines and proposed model using the log- likelihood function. The log-likelihood has already been optimized to give strong performances in our baseline experiments compared to global learning [5] to maximize F-Score. We distribute the SDSEG dataset into training, development, and test sets. We use variational dropout [13] to both input and output recurrent units. The softmax is used for label classification in baseline LSTM and BiLSTM models, CRF is added in the last layer of the BiLSTM- CRF and SGNWS models. The Gradient normalization is used to improve the performance [31], which re-scales the gradient when the norm goes over a threshold. The range of optimal hyper-parameters for SRL, baselines, and the proposed model is depicted in Table 8. Table 8: Optimal hyper-parameters for SRL, baseline neural models, and proposed SGNWS model. | Hyper-parameter | Range ---|---|--- SRL | ${E^{c}}$ dimension | 64 $E^{B}$ dimension | 64 $E^{T}$ dimension | 64 ${Emb}_{S}$ dimension | 64 | Epochs | 100 Neural models | Optimizer | Adamax Learning rate | 0.025 | Gradient normalization | 5.0 | h layers | 200 | Dropout | 0.25 | Epochs | 40 Table 9: Results of baselines and proposed SGNWS model for Sindhi word segmentation on the SDSEG development and test sets. | | Dev. | | | Test | | ---|---|---|---|---|---|---|--- RNN variant | P | R | F | P | R | F | LSTM (Baseline) | 96.38 | 95.68 | 95.29 | 94.81 | 94.57 | 94.32 | BiLSTM | 96.86 | 96.21 | 96.19 | 96.52 | 94.28 | 95.87 | BiLSTM-CRF | 97.25 | 96.38 | 96.74 | 96.11 | 95.87 | 96.18 | BiLSTM-CRF+Char | 97.76 | 97.81 | 96.38 | 96.82 | 97.26 | 96.78 | BiLSTM-CRF+bigram | 96.34 | 97.89 | 96.58 | 96.13 | 97.23 | 96.74 | BiLSTM-CRF++Trigram | 97.14 | 98.29 | 96.89 | 97.32 | 97.68 | 96.53 | SGNWS | 99.77 | 98.83 | 98.94 | 99.08 | 98.72 | 98.51 | ## 6 Results and analysis The Table 9 shows the performance comparison of all the models on SDSEG dataset. Firstly, the LSTM yields a stable baseline F-Score of 95.29% on development and 94.32% on the test set. The BiLSTM model provides better results than LSTM in both development and test sets due to the bidirectional learning states. However, the BiLSTM-CRF is superior over both baselines of LSTM and BiLSTM, respectively, which shows that CRF is dominant over softmax classifier. Moreover, the addition of character-level features in the BiLSTM- CRF model surpasses three baselines. However, BiLSTM-CRF with bigram and trigram based word embeddings yield close results to the BiLSTM-CRF+Char model, which shows the superiority of the character-level approach a performance gain. The proposed SGNWS model produced superior results over baselines, as depicted in Table 9 on development and test data. According to the results, SRL is beneficial for the SWS task. The proposed SGNWS model surpasses all the baselines on the SDSEG dataset as well as on five different datasets (see Figure 1) of Kawish, Awami-Awaz, Wikipedia-dumps, Twitter, and books. Figure 1: The performance of proposed SGNWS model on the various datasets. The F-Score is reported on the test set of multiple datasets. Our proposed SGNWS model surpasses baselines with a high F-Score of 98.94% on development set and 98.51% on the test set using the SDSEG dataset. The observation indicates that SRL is beneficial to capture more semantic information for the word segmentation of Sindhi text. ## 7 Conclusion The word segmentation is an essential and non-trivial task in Sindhi language. The white spaces between words are a good sign for predicting word boundaries, but the existence of space-omission and space-insertion bring ambiguity in the segmentation process. We proposed the SGNWS model, keeping in view the challenges related to SWS, respectively. The proposed model has the ability to learn and extract subword features automatically by eliminating the constraints such as hand-craft features for segmentation or any other type of prior domain-specific knowledge. In this paper, we empirically demonstrate that our proposed model yields the best performance in SWS because of its high efficiency and robustness for sequential modeling tasks with great ability to capture the word information at the morphemic level for the prediction of word boundaries. The SGNWS model is an effective and elegant neural solution for SWS, which can also be applied to other sequence tagging problems. ## Acknowledgement This work was funded by the National Key R&D Program of China (No. 2018YFB1005100 & No. 2018YFB1005104). ## References * [1] Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S., Irving, G., Isard, M., et al.: Tensorflow: A system for large-scale machine learning. In: 12th $USENIX$ Symposium on Operating Systems Design and Implementation $(OSDI)$ 16). pp. 265–283 (2016) * [2] Akbik, A., Blythe, D., Vollgraf, R.: Contextual string embeddings for sequence labeling. In: Proceedings of the 27th International Conference on Computational Linguistics. pp. 1638–1649 (2018) * [3] Ali, W., Kumar, J., Lu, J., Xu, Z.: Word embedding based new corpus for low-resourced language: Sindhi. arXiv preprint arXiv:1911.12579 (2019) * [4] Almuhareb, A., Alsanie, W., Al-Thubaity, A.: Arabic word segmentation with long short-term memory neural networks and word embedding. IEEE Access 7, 12879–12887 (2019) * [5] Andor, D., Alberti, C., Weiss, D., Severyn, A., Presta, A., Ganchev, K., Petrov, S., Collins, M.: Globally normalized transition-based neural networks. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics. pp. 2442–2452 (2016) * [6] Bhatti, Z., Ismaili, I.A., Soomro, W.J., Hakro, D.N.: Word segmentation model for Sindhi text. American Journal of Computing Research Repository 2(1), 1–7 (2014) * [7] Bird, S.: NLTK: the natural language toolkit. In: Proceedings of the COLING/ACL on Interactive presentation sessions. pp. 69–72 (2006) * [8] Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics 5, 135–146 (2017) * [9] Chen, X., Qiu, X., Zhu, C., Liu, P., Huang, X.: Long short-term memory neural networks for Chinese word segmentation. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 1197–1206 (2015) * [10] Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., Kuksa, P.: Natural language processing (Almost) from scratch. The Journal of Machine Learning Research 12, 2493–2537 (2011) * [11] Ding, C., Thu, Y.K., Utiyama, M., Sumita, E.: Word segmentation for Burmese (Myanmar). ACM Transactions on Asian and Low-Resource Language Information Processing 15(4), 1–10 (2016) * [12] Dootio, M.A., Wagan, A.I.: Automatic stemming and lemmatization process for Sindhi text. Journal of Social Sciences & Interdisciplinary Research 6(2), 19–28 (2017) * [13] Gal, Y., Ghahramani, Z.: A theoretically grounded application of dropout in recurrent neural networks. In: Advances in Neural Information Processing Systems. pp. 1019–1027 (2016) * [14] Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735–1780 (1997) * [15] Huang, Z., Xu, W., Yu, K.: Bidirectional LSTM-CRF models for sequence tagging. arXiv preprint arXiv:1508.01991 (2015) * [16] Jamro, W.A.: Sindhi language processing: A survey. In: International Conference on Innovations in Electrical Engineering and Computational Technologies. pp. 1–8 (2017) * [17] Kitagawa, Y., Komachi, M.: Long short-term memory for Japanese word segmentation. In: Proceedings of the 32nd Pacific Asia Conference on Language, Information and Computation (2018) * [18] Labeau, M., Allauzen, A.: Character and subword-based word representation for neural language modeling prediction. In: Proceedings of the First Workshop on Subword and Character Level Models in NLP. pp. 1–13 (2017) * [19] Lafferty, J.D., McCallum, A., Pereira, F.C.: Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In: Proceedings of the Eighteenth International Conference on Machine Learning. pp. 282–289 (2001) * [20] Liu, H., He, L., Bai, H., Dai, B., Bai, K., Xu, Z.: Structured inference for recurrent hidden semi-markov model. In: Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence. pp. 2447–2453 (2018) * [21] Liu, Q., Huang, H.Y., Gao, Y., Wei, X., Tian, Y., Liu, L.: Task-oriented word embedding for text classification. In: Proceedings of the 27th International Conference on Computational Linguistics. pp. 2023–2032 (2018) * [22] Ma, J., Ganchev, K., Weiss, D.: State-of-the-art Chinese word segmentation with bi-lstms. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 4902–4908 (2018) * [23] Ma, X., Hovy, E.: End-to-end sequence labeling via bi-directional LSTM-CNNs-CRF. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics. pp. 1064–1074 (2016) * [24] Mahar, J., Shaikh, H., Memon, G.: A model for Sindhi text segmentation into word tokens. Sindh University Research Journal (Science Series) 44(1) (2012) * [25] Mahar, J.A., Memon, G.Q., Danwar, S.H.: Algorithms for Sindhi word segmentation using lexicon-driven appraoch. International journal of academic research 3(3) (2011) * [26] Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. In: Advances in Neural Information Processing Systems. pp. 3111–3119 (2013) * [27] Motlani, R.: Developing language technology tools and resources for a resource-poor language: Sindhi. In: Proceedings of the NAACL Student Research Workshop. pp. 51–58 (2016) * [28] Narejo, W.A., Mahar, J.A.: Morphology: Sindhi morphological analysis for natural language processing applications. In: Proceedings of the International Conference on Computing, Electronic and Electrical Engineering). pp. 27–31 (2016) * [29] Narejo, W.A., Mahar, J.A., Mahar, S.A., Surahio, F.A., Jumani, A.K.: Sindhi morphological analysis: An algorithm for Sindhi word segmentation into morphemes. International Journal of Computer Science and Information Security 14(6), 293 (2016) * [30] Pan, Y., Xu, J., Wang, M., Ye, J., Wang, F., Bai, K., Xu, Z.: Compressing recurrent neural networks with tensor ring for action recognition. In: The Thirty-Third AAAI Conference on Artificial Intelligence. pp. 4683–4690 (2019) * [31] Pascanu, R., Mikolov, T., Bengio, Y.: On the difficulty of training recurrent neural networks. In: Proceedings of the 30th International Conference on International Conference on Machine Learning. vol. 28, pp. III–1310 (2013) * [32] Rahman, M.U.: Towards Sindhi corpus construction. In: Proceedings of the Conference on Language and Technology (2010) * [33] Reimers, N., Gurevych, I.: Optimal hyperparameters for deep LSTM-networks for sequence labeling tasks. arXiv preprint arXiv:1707.06799 (2017) * [34] Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673–2681 (1997) * [35] Shao, Y.: Cross-lingual word segmentation and morpheme segmentation as sequence labelling. In: The First Workshop on Multi-Language Processing in a Globalising World. pp. 75–80 (2017) * [36] Shao, Y., Hardmeier, C., Tiedemann, J., Nivre, J.: Character-based joint segmentation and POS tagging for Chinese using bidirectional RNN-CRF. In: Proceedings of the Eighth International Joint Conference on Natural Language Processing. pp. 173–183 (2017) * [37] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998–6008 (2017) * [38] Wen, L., Zhang, X., Bai, H., Xu, Z.: Structured pruning of recurrent neural networks through neuron selection. Neural Networks 123, 134–141 (2020) * [39] Yang, J., Zhang, Y., Dong, F.: Neural word segmentation with rich pretraining. In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics. pp. 839–849 (2017) * [40] Yao, Y., Huang, Z.: Bi-directional lstm recurrent neural network for Chinese word segmentation. In: International Conference on Neural Information Processing. pp. 345–353 (2016) * [41] Ye, J., Wang, L., Li, G., Chen, D., Zhe, S., Chu, X., Xu, Z.: Learning compact recurrent neural networks with block-term tensor decomposition. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 9378–9387 (2018) * [42] Young, T., Hazarika, D., Poria, S., Cambria, E.: Recent trends in deep learning based natural language processing. IEEE Computational Intelligence Magazine 13(3), 55–75 (2018) * [43] Zhang, S., Mao, C., Yu, Z., Wang, H., Li, Z., Zhang, J.: Word segmentation for Burmese based on dual-layer CRFs. ACM Transactions on Asian and Low-Resource Language Information Processing 18(1) (2018) * [44] Zhao, H., Huang, C., Li, M., Lu, B.L.: Effective tag set selection in Chinese word segmentation via conditional random field modeling. In: Proceedings of the 20th Pacific Asia Conference on Language, Information and Computation. pp. 87–94 (2006) * [45] Zheng, B., Che, W., Guo, J., Liu, T.: Enhancing LSTM-based word segmentation using unlabeled data. In: Chinese Computational Linguistics and Natural Language Processing Based on Naturally Annotated Big Data, pp. 60–70 (2017) * [46] Zia, H.B., Raza, A.A., Athar, A.: Urdu word segmentation using conditional random fields (CRFs). In: Proceedings of the 27th International Conference on Computational Linguistics. pp. 2562–2569 (2018)
# Leveraging Modality-specific Representations for Audio-visual Speech Recognition via Reinforcement Learning Chen Chen1, Yuchen Hu1, Qiang Zhang2, 3, Heqing Zou1, Beier Zhu1, and Eng Siong Chng1 ###### Abstract Audio-visual speech recognition (AVSR) has gained remarkable success for ameliorating the noise-robustness of speech recognition. Mainstream methods focus on fusing audio and visual inputs to obtain modality-invariant representations. However, such representations are prone to over-reliance on audio modality as it is much easier to recognize than video modality in clean conditions. As a result, the AVSR model underestimates the importance of visual stream in face of noise corruption. To this end, we leverage visual modality-specific representations to provide stable complementary information for the AVSR task. Specifically, we propose a reinforcement learning (RL) based framework called MSRL, where the agent dynamically harmonizes modality- invariant and modality-specific representations in the auto-regressive decoding process. We customize a reward function directly related to task- specific metrics (_i.e._ , word error rate), which encourages the MSRL to effectively explore the optimal integration strategy. Experimental results on the LRS3 dataset show that the proposed method achieves state-of-the-art in both clean and various noisy conditions. Furthermore, we demonstrate the better generality of MSRL system than other baselines when test set contains unseen noises. ## 1 Introduction Background noise is inevitable in real world that can dramatically degrade the speech quality and intelligibility, thereby increasing the difficulty of speech recognition task (Hu et al. 2022a, b). In noisy scenarios, human will unconsciously observe the mouth region of speakers, as such noise-invariant visual cues can provide useful information for the corrupted speech understanding (Ma, Petridis, and Pantic 2021). Similar to this, the audio-visual speech recognition (AVSR) technique couples the audio and visual modalities, which has attracted increasing research interest for several years (Noda et al. 2015). Recent machine learning based AVSR methods successfully demonstrate that deep neural network (DNN) can process and fuse raw acoustic and visual inputs to improve the noise- robustness for recognition through a supervised learning paradigm (Petridis et al. 2018; Zhou et al. 2019). Additionally, self-supervised representation learning has been explored to capture the correlations between audio and visual lip movements for AVSR task, which has brought remarkable performance gain in terms of word error rate (WER) metric (Shi et al. 2022). Figure 1: Research problem. “SNR” denotes the signal-to-noise ratio, and $\alpha$ is the threshold that modality-invariant representations lose the effectiveness compared with visual modality-specific representations. Mainstream AVSR methods focus on learning modality-invariant representations by fusing audio and visual modalities into a common subspace (Song, Sun, and Li 2022). However, such a fusion manner is prone to over-reliance on the audio modality, as it is much easier to recognize than video stream in clean conditions (Mittal et al. 2020). With the rise of noise levels, the importance of the video stream is increasingly underestimated in AVSR systems, and leads to sub-optimal performance since audio modality has already been corrupted by noise. We further illustrate this research problem in Fig 1. Though modality- invariant representations (green line) outperform audio modality-specific representations (yellow line) by a large margin, it is still vulnerable to low signal-to-noise ratio (SNR) conditions. It is worth noting that when SNR is lower than $\alpha$, the modality-invariant representations even perform worse than visual modality-specific representations (blue line) which are completely unaffected by noise. We argue that this problematic situation can be avoided by reasonable coordination of modality-invariant and modality-specific representations, which is shown as the oracle system (red dashed line). Although the significance of modality-specific representations has been emphasized in other multi-modal tasks, such as emotion recognition (Hazarika, Zimmermann, and Poria 2020), it still remains challenging to integrate them into AVSR system for several reasons. Firstly, the real-world noises have dynamic and non-stationary temporal distributions, which confuse the recognizer to estimate the importance of visual modality-specific representations during auto-regressive decoding. Secondly, due to the natural distinction of input modalities, a uniform training schedule probably results in the vanishing gradient when we add a further sub-network to extract visual modality-specific representations (Yao and Mihalcea 2022). Finally, with parameter growth of neural networks, the existing integration strategies for new representations are prone to over-fit to specific types of noise distribution (Fu, Wu, and Boulet 2022), thereby failing to adapt to unseen noises in the wild. In this work, we aim to improve the AVSR system by leveraging visual modality- specific representations that carry noise-invariant information from the visual stream. To this end, we employ a pre-trained vision model that takes lip movement information as input and generates independent probability distribution for sequence generation. This idea is inspired by the language model rescoring that has been widely applied in popular ASR methods (Song et al. 2022; Xu et al. 2022; Chen et al. 2022a). However, instead of applying a typical integration approach, _e.g._ , late fusion (Inaguma et al. 2019), we propose a reinforcement learning (RL) based method to dynamically harmonize the integration process in terms of the task-specific metric. RL is appropriate to play this integration role for: 1) The auto-regressive decoding of AVSR can be modeled as an RL formulation (Bahdanau et al. 2016), where the agent can consider multiple information for reasonable token prediction. 2) The customized reward function of RL bridges the training criterion and WER, thus encouraging it to effectively improve the model performance. 3) The beam search of inference step can provide a set of hypotheses for RL sampling, which allows the agent to explore the optimal policy on token level (Chen et al. 2022c). The main contributions of this paper can be summarized as following: * • We propose MSRL – a novel AVSR system that utilizes visual modality-specific representations to dynamically remedy the noise-corrupted audio modality. * • MSRL adopts an RL-based integration method, where a new reward function is designed to encourage the agent to efficiently explore the optimal strategy in terms of the task-specific metric. * • Experimental results on the largest public LRS3 dataset show that MSRL is effective and achieves state-of-the-art performance in both clean and various noisy conditions. Furthermore, the comparative experiments on unseen noises demonstrate that MSRL has better generalization and adaptability than a strong baseline. ## 2 Related work Audio-visual speech recognition. Recently, AVSR has been attracting increasing research interest as it provides a potential solution for noise-robust speech recognition. To process and fuse audio-video modalities, TM-seq2seq (Afouras, Chung, and Zisserman 2018a) applies a separated Transformer encoder for two modalities and fuses them before decoding. (Ma, Petridis, and Pantic 2021) presents a hybrid CTC/Attention model based on Resnet-18 and Conformer (Gulati et al. 2020), which can be trained in an end-to-end manner. (Tao and Busso 2018) demonstrate the importance of aligning two modalities before fusing them. Moreover, the AV-HuBERT (Shi et al. 2022) learn the correspondence of audio and video modalities in a self-supervised manner, which is further augmented in (Shi, Hsu, and Mohamed 2022) to improve noise-robustness. Modality-invariant and modality-specific representations. Despite the advanced fusion techniques in multi-modal tasks (Chen et al. 2022b), prior works suggest that the model can benefit from modality-specific representations which capture some desirable properties (Xiong et al. 2020). Nevertheless, how to effectively utilize it is still an open question to be explored. MISA (Hazarika, Zimmermann, and Poria 2020) maps the multi-modal inputs into two subspaces for modality-invariant and modality-specific representations and then fuses them for final classification. MuSE (Yao and Mihalcea 2022) employs separated encoders for multiple modalities, then harmonize them using different learning rates and late-fusion. Similarly, (Feng, Lai, and Xie 2019) constructs an individual network for each modality, as well as designing a modality-shared identity loss to facilitate the extraction of modality- invariant representation. The integration of modality-specific representations is particularly difficult in an AVSR system because the single decoder is hard to dynamically weight the representations in a sequential decision process. Reinforcement learning in sequence generation. Extensive existing works have demonstrated that RL is suitable to play an optimizing role in sequence generation tasks. In captioning tasks like image captioning (Rennie et al. 2017) and audio captioning (Mei et al. 2021), a self-critical training approach based on RL can optimize the trained model in terms of non- differentiable metrics. Such an idea is also expanded to ASR tasks with a customized reward function (Tjandra, Sakti, and Nakamura 2019; Chen et al. 2022c). Additionally, actor-critic based RL optimization algorithms are designed to improve task-specific score (_e.g._ BLEU) in machine translation task (Williams 1992; Bahdanau et al. 2016). Compared with previous work, the proposed MSRL commits to the stability of RL training, where 1) we utilize pre-trained models to provide learned representations as state space, and 2) we design a reward function to encourage the policy network to explore in trust region (Schulman et al. 2015). Figure 2: The block diagram of the proposed MSRL system. The solid box denotes such module is fixed during training, while the dashed box denotes it is trainable. The red dashed arrow denotes the process of back-propagation. Policy network considers multiple information in auto-regressive decoding and predicts the current token “fine”. ## 3 Methodology In this part, we first introduce the main structure of the proposed MSRL system in Section 3.1. Then in Section 3.2, we model the auto-regressive decoding of AVSR as an RL formulation, as well as give mathematical derivation for the optimization. Finally, the training schedule of MSRL is illustrated in Section 3.3. ### 3.1 Main Structure Given the acoustic utterance $A$ and its paired $l$-frame video stream $V=(v_{1},v_{2},...,v_{l})$, the neural network of AVSR intends to predict a hypothesis sequence $Y=(y_{1},y_{2},...,y_{T})$. As shown in Fig. 2, the audio and video streams are fed into a pre-trained AV encoder to extract hidden representation, where a ResNet block (He et al. 2016) and a linear layer are respectively served as front-end to obtain the audio and visual features. After concatenating them, a Transformer encoder (Vaswani et al. 2017) with self-attention mechanism is employed to extract hidden representations. Subsequently, we utilize a learnable Transformer decoder including cross- attention mechanism to produce modality-invariant representations $F_{i}$ corresponding to the probability distribution of predicted tokens. Meanwhile, we further utilize a pre-trained vision model which similarly consists of ResNet front-end, Transformer encoder, and Transformer decoder. It consumes the video stream as input and independently produces visual modality-specific representations $F_{v}$ with the same shape as $F_{i}$. To harmonize the modality-specific and modality-invariant representations, a linear layer-based policy network are designed in the auto-regressive decoding process. Besides the $F_{i}$ and $F_{v}$ themselves, we argue that policy networks should also be aware of audio quality which is useful to estimate the importance of representations. To this end, an MLP block with 2 linear layers is used for downsampling and provides acoustic information $I_{a}$ for policy network. Finally, a combined distribution is generated to predict the current token (“fine” in Fig. 2). It is noted that all pre-trained models are fixed in the whole training process. The pre-trained AV encoder is initialized from AV-HuBERT (Shi, Hsu, and Mohamed 2022), which captures cross-modal correlations between audio and video modalities by a self-supervised approach. The vision model (Shi et al. 2022) including encoder and decoder is pre-trained via lip-reading task on LRS3 dataset, where only the video stream is required to generate target sequence. ### 3.2 RL Policy Basic Reinforcement Learning is typically formulated as a Markov Decision Process (MDP) that includes a tuple of trajectories $\left\langle\mathcal{S},\mathcal{A},\mathcal{R},\mathcal{T}\right\rangle$. For each time step $t$, the agent consider state $s_{t}\in S$ to generate an action $a_{t}\in\mathcal{A}$ which interacts with environment. The transition dynamics $\mathcal{T}(s_{t+1}|s_{t},a_{t})$ is defined as transition probability from current state $s_{t}$ to next state $s_{t+1}$, and gain an instant reward $r_{t}(s_{t},a_{t})$. The objective of RL is to learn optimal policy to maximize the cumulative reward $\mathcal{R}$: $\mathcal{R}=\mathop{\max}_{a_{t}\in\mathcal{A}}\sum_{t=1}^{T}r_{t}$ (1) In AVSR task, we summarize the MDP tuple as: * • State $\mathcal{S}$ should contain the comprehensive information or learned patterns for decision-making. Therefore, we denoted $\mathcal{S}$ as a combination of $F_{i}$, $F_{v}$, and $I_{a}$ defined in Section 3.1, as they are related to predict current token. * • Action $\mathcal{A}$ aims to interact with the environment and update the $\mathcal{S}$. In this work, $\mathcal{A}$ is defined as a probability distribution $P_{a}$ for the current predicted token. * • Reward $\mathcal{R}$ is an instant feedback signal to evaluate the performance of $\mathcal{A}$. we define a token-level reward function for each hypothesis $Y$ as follows: $\begin{split}\mathcal{R}(Y,Y^{*})=-D_{\text{ED}}(Y||Y^{*})&-\lambda_{1}\sum_{t=0}^{T}D_{\text{KL}}(P_{a}^{t}||F_{i}^{t})\\\ &-\lambda_{2}\sum_{t=0}^{T}D_{\text{KL}}(P_{a}^{t}||F_{v}^{t})\end{split}$ (2) where $D_{\text{ED}}(\cdot||\cdot)$ denotes the edit distance between two sequence, and $Y^{*}$ denotes the ground-truth sequence. It is noted that such distance is directly related to WER. The $D_{\text{KL}}(\cdot||\cdot)$ denotes the KL-divergence between two distributions, which are used to constrain the policy network to explore in trust region (Schulman et al. 2015). $\lambda_{1}$ and $\lambda_{2}$ are the weights to balance them. * • Transition dynamics $\mathcal{T}(s_{t+1}|s_{t},a_{t})$ denotes that the predicted token $a_{t}$ will influence next state $s_{t+1}$, since the decoding of AVSR is auto-regressive generation process. In order to maximize the cumulative reward $\mathcal{R}$, the training objective of policy network is defined as: $\begin{split}\mathcal{L}_{\theta}(\left\langle A,V\right\rangle,Y^{*})&=-\mathbb{E}[\mathcal{R}(Y,Y^{*})]\\\ &=\sum_{Y}\mathcal{P}(Y|\left\langle A,V\right\rangle,\theta)\mathcal{R}(Y,Y^{*})\end{split}$ (3) where $\theta$ denotes the neural network, $\mathcal{P}(Y|\left\langle A,V\right\rangle,\theta)$ is the probability of hypothesis $Y$ determined by input $\left\langle A,V\right\rangle$ and $\theta$. The reward function $\mathcal{R}(Y,Y^{*})$ is defined in E.q (2). Since $\sum_{Y}\mathcal{P}(Y|\left\langle A,V\right\rangle,\theta)$ involves a summation over all possible sequences, we employ the REINFORCEMENT algorithm (Williams 1992) to approximate the expected $\mathcal{R}$ and calculate the gradient $\nabla_{\theta}\mathcal{L}_{\theta}$: $\nabla_{\theta}\mathcal{L}_{\theta}\\!\\!=\\!\\!-\mathbb{E}_{Y^{n}\sim\mathcal{P}(Y^{n}|\left\langle A,\\!V\right\rangle,\theta)}[\mathcal{R}(Y^{n}\\!,Y^{*}\\!)\nabla_{\theta}log\mathcal{P}(Y^{n}|\\!\left\langle A,\\!V\right\rangle\\!,\theta)]$ (4) Where $Y^{n}$ is the sampling hypothesis drawn from the current model distribution. Different from other sampling methods, we directly utilize the beam search algorithm during decoding to select the $N$-best hypothesis, which indicates the number of sampling hypotheses is equal to the beam size $N$. Furthermore, we introduce the baseline to normalize the reward as follows: $\nabla_{\theta}\mathcal{L}_{\theta}\\!=\\!-\frac{1}{N}\\!\\!\sum_{Y^{n}\in\text{Beam}}^{N}\\!\\!\\!\nabla_{\theta}log\mathcal{P}(Y^{n}|\left\langle A,\\!V\right\rangle,\theta)\ [\ \mathcal{R}(Y^{n},Y^{*})-\bar{\mathcal{R}}\ ]$ (5) where $\bar{\mathcal{R}}$ is the baseline defined as the average of reward of all hypotheses in a beam set. Subtracting $\bar{\mathcal{R}}$ does not influence the gradient, but importantly, it can reduce the variance of the gradient estimation, thus stabilizing the training process. To simplify the calculation, we assume that the probability mass is concentrated on the $N$-best list only. Consequently, the loss function can be approximated as: $\mathcal{L}\approx-\\!\\!\\!\sum_{Y^{n}\in\text{Beam}}^{N}\\!\\!log\hat{\mathcal{P}}(Y^{n}|\left\langle A,\\!V\right\rangle,\theta)\ [\ \mathcal{R}(Y^{n},Y^{*})-\bar{\mathcal{R}}\ ]$ (6) where $\hat{\mathcal{P}}(Y^{n}|\left\langle A,\\!V\right\rangle,\theta)=\frac{\mathcal{P}(Y^{n}|\left\langle A,\\!V\right\rangle,\theta)}{\sum_{Y^{n}\in\text{Beam}}\mathcal{P}(Y^{n}|\left\langle A,\\!V\right\rangle,\theta)}$ represents the re-normalized distribution over the N-best hypotheses. Accordingly, in one Beam set, those hypotheses with a higher reward than average are encouraged to be selected by increasing their possibilities. Conversely, the hypothesis that obtains a lower reward will be suppressed. By minimizing the criterion of E.q (6), the current model intends to pursue higher reward by effective exploration in a beam set. ### 3.3 Training Schedule of MSRL The training process contains two stages as shown in Algorithm 1. We first use typical cross-entropy criterion to train the randomly initialized decoder that is shown from step 1 to step 4. The best model is selected by a valid set for subsequent sampling. Then the RL optimization is applied to integrate the visual modality-specific representations according to the reward function in step 4. Considering the continuity of utterance, we adopt an online training manner and the gradient is calculated after the completion of the beam search. Consequently, to achieve higher reward, the policy network will be updated to the direction which optimizes the posterior metric. Algorithm 1 Pseudocode for MSRL Training 1:The paired audio $A$, video $V$, and corresponding ground-truth sequence $Y^{*}=(y_{1}^{*},y_{1}^{*},...,y_{T}^{*})$. 2:Initialize the pre-trained parameters for AV encoder $\theta_{av}$ and vision model $\theta_{v}$. 3:Initialize the random parameters for Transformer decoder $\theta_{d}$, MLP block $\theta_{m}$, and RL policy network $\theta_{p}$. 4:while TRUE do Freeze the parameters of encoder $\theta_{av}$ Obtain the hidden feature $h_{av}\\!=\theta_{av}(A,V)$ Train the decoder using cross- entropy loss $\mathcal{L}_{ce}$: $\mathcal{L}_{ce}=\sum_{t=1}^{T}-\log\mathcal{P}_{\theta_{d}}(y_{t}^{*}|\ y_{t-1}^{*},...,y_{1}^{*},h_{av})\vspace{-0.1cm}$ (7) end while 5:while TRUE do for hypothesis $Y^{n}$ in $N$-best list do Freeze the encoder $\theta_{av}$ and vision model $\theta_{v}$ for t in 1,2,…, T do Obtain representations $F_{i}$ and $F_{v}$: $F_{i}$ = $\theta_{d}(h_{av})$ $F_{v}$ = $\theta_{v}(V)$ Compute current action probability: $P_{a}^{t}=\theta_{p}(F_{i}^{t},F_{v}^{t},\theta_{m}(A))$ end for Compute probability $\mathcal{P}(Y^{n})=\prod_{t=1}^{T}P_{a}^{t}$ Determine accumulative reward $\mathcal{R}(Y^{n},Y^{*})$ end for Train the policy network using E.q (6)end while ## 4 Experiment Setting ### 4.1 Database We conduct the experiments on LRS3 (Afouras, Chung, and Zisserman 2018b), which is the largest publicly available dataset for audio-visual speech recognition task. It includes face tracks from over 400 hours of TED and TEDx videos from more than 5,000 speakers, along with the corresponding subtitles and word alignment boundaries. The original training set is divided into 2 partitions: pretrain (403 hours) and trainval (30 hours), which are both from the same sources with test set (1452 utterances, 1 hour). In this paper, we randomly choose 1,200 utterances (1 hour) as a valid set for hyper-parameter tuning and best model selection. For the noisy test set, we follow the categories and mixing strategy from prior work (Shi, Hsu, and Mohamed 2022). The seen noises contains categories of “ _babble_ ”, “ _music_ ” and “ _natural_ ” that are sampled from MUSAN dataset (Snyder, Chen, and Povey 2015), and “ _speech_ ” noise is sampled from utterances in LRS3. These four categories of noises are seen by both pre- trained models and the training process. For the unseen noises, we select 4 categories of “ _Cafe_ ”, “ _Meeting_ ”, “ _River_ ”, and “ _Resto_ ” from DEMAND noise set (Thiemann, Ito, and Vincent 2013) and mix them with test set. The detailed data pre-processing strategy is illustrated in appendix. ### 4.2 MSRL Set up We develop several MSRL frameworks with different settings, as shown in Table 1. The small transformer block has 768/3072/12 of embedding dimension/feed- forward dimension/attention heads, and the large transformer block increases to 2034/4096/16 respectively. Considering the task difficulty of lip reading, the encoder and decoder of vision model adopt large blocks. The labeled data is first used for the pre-trained models (Shi et al. 2022), then it is reused for the training of the decoder and RL module. According to labeled data amount, we define it as two modes. The normal-resource contains 433 hours of full training data (pretrain subset and trainval subset), and the low-resource only contains 30 hours of training data (trainval subset). ID | AV Pre-trained Encoder (# Enc. blocks ) | Decoder (# Dec. blocks ) | Vision Pre-trained Model (# Enc./ Dec. blocks) | Labeled data (hours) ---|---|---|---|--- 1 | Small (12) | Small (6) | Large (24/16) | 30 2 | Small (12) | Small (6) | Large (24/16) | 433 3 | Large (24) | Large (9) | Large (24/16) | 30 4 | Large (24) | Large (9) | Large (24/16) | 433 Table 1: Different settings of MSRL. “# Enc.” and “# Dec.” denotes the numbers of encoder and decoder blocks. ## 5 Result and Analysis In this section, we conduct extensive experiments and answer the following questions: * • What is the effect of modality-specific representations in AVSR task? We display the experimental results to prove that MSRL addresses the research problem in Fig. 1, and the problematic situation will not happen at low SNR conditions. * • What is the effect of the RL integration? We conduct comparative experiments including other integration strategies to show the superiority of RL method. * • How does the MSRL performance against other competitive methods? We carry out a series of experiments in various conditions to compare our method with previously published works. * • How is the generalization of MSRL to unseen noises? We directly test the MSRL on various conditions with unseen noise to demonstrate its ability of generalization. ### 5.1 Effect of Modality-specific Representations In this part, we first quantitatively analyze the effect that MSRL utilizes the visual modality representations. To this end, we construct three baseline systems that leverage different representations. _Audio-only_ baseline only consumes audio modality as input to generate target sequence. _Visual-only_ baseline is trained as a lip-reading task that consumes visual modality as input. _Modality-invariant_ baseline is trained as a vanilla AVSR task that consumes both audio and visual modality as input. Considering the intensity, the _babble_ noise is selected to simulate the noisy condition with different SNR levels. The WER results of MSRL and three baselines with different Transformer blocks and resource modes are shown in Table 2. Method | Block | _Babble_ Noise, SNR= | Clean ---|---|---|--- | -15 | -10 | -5 | 0 | 5 | avg | $\infty$ Normal-resource (433 hours of labeled data) _Audio-only_ | Small | 99.1 | 98.1 | 82.7 | 32.6 | 11.9 | 64.9 | 2.53 _Audio-only_ | Large | 98.6 | 97.4 | 75.8 | 24.6 | 9.01 | 61.1 | 1.95 _Visual-only_ | Large | 26.9 | 26.9 _Modality-invariant_ | Small | 55.6 | 38.0 | 19.1 | 7.24 | 4.02 | 24.8 | 1.84 _Modality-invariant_ | Large | 43.4 | 30.3 | 13.5 | 4.90 | 2.50 | 18.9 | 1.45 MSRL | Small | 26.1 | 24.7 | 14.8 | 5.92 | 3.19 | 14.9 | 1.44 MSRL | Large | 25.5 | 22.3 | 11.3 | 4.51 | 2.31 | 13.2 | 1.33 Low-resource (30 hours of labeled data) _Audio-only_ | Small | * | * | 84.2 | 36.1 | 13.9 | 66.8 | 4.69 _Audio-only_ | Large | * | 98.0 | 77.0 | 25.9 | 14.2 | 63.0 | 3.51 _Visual-only_ | Large | 27.8 | 27.8 _Modality-invariant_ | Small | 53.0 | 39.5 | 21.4 | 10.2 | 5.92 | 26.0 | 4.10 _Modality-invariant_ | Large | 44.8 | 32.3 | 16.4 | 7.37 | 4.87 | 21.1 | 3.27 MSRL | Small | 27.4 | 25.8 | 16.7 | 7.24 | 5.20 | 16.5 | 3.38 MSRL | Large | 26.5 | 24.9 | 13.0 | 6.36 | 3.97 | 14.9 | 2.82 Table 2: The WER (%) results in _babble_ noise and clean conditions.“avg” denotes the average performance across all SNR. “*” denotes the input modality can not be recognized. Method | _Babble_ Noise, SNR= | Clean ---|---|--- -15 | -10 | -5 | 0 | 5 | avg | $\infty$ Normal-resource (433 hours of labeled data) & Large Transformer block _Modality-invariant_ | 43.4 | 30.3 | 13.5 | 4.90 | 2.50 | 18.9 | 1.45 _Early fusion_ | 38.2 | 25.8 | 12.6 | 5.07 | 2.96 | 16.9 | 1.58 _Late fusion_ | 36.7 | 26.2 | 12.2 | 4.78 | 2.70 | 16.5 | 1.68 _Model ensemble_ | 31.6 | 23.4 | 11.8 | 5.36 | 3.15 | 15.1 | 2.26 MSRL | 25.5 | 22.3 | 11.3 | 4.51 | 2.31 | 13.2 | 1.33 Low-resource (30 hours of labeled data) & Large Transformer block _Modality-invariant_ | 44.8 | 32.3 | 16.4 | 7.37 | 4.87 | 21.1 | 3.27 _Early fusion_ | 40.1 | 25.6 | 15.5 | 7.40 | 5.01 | 18.7 | 3.26 _Late fusion_ | 38.4 | 25.9 | 13.3 | 6.40 | 4.19 | 17.6 | 3.31 _Model ensemble_ | 33.7 | 25.4 | 13.6 | 6.77 | 4.18 | 16.7 | 3.35 MSRL | 26.5 | 24.9 | 13.0 | 6.36 | 3.97 | 14.9 | 2.82 Table 3: The WER (%) results of MSRL and other integration methods in _babble_ noise and clean conditions. Best results are in bold. Figure 3: The visualization WER(%) results of _Audio-only_ baseline, _Modality-invariant_ baseline, and proposed MSRL with large block and 433 hours of labeled data. We observe that except _Visual-only_ baseline, the performance of other methods degrades obviously with the decrease of SNR. When SNR is lower than -5, the performance of _Modality-invariant_ baseline is even worse than _Visual-only_ baseline. However, such a problematic situation does not happen in MSRL, since the visual modality-specific representations have been increasingly effective if audio quality becomes hard to recognize. Furthermore, we observe that MSRL system achieves up to 17.6% relative WER reduction than _Modality-invariant_ baseline in clean conditions. It is out of intuition because the visual modality-specific representations are usually considered trivial when audio quality is high. We reason that 1) visual modality-specific representations add the diversity of information, and it might be helpful when some ambiguous acoustic pronunciations have similar probabilities. 2) The training objective in E.q.(6) is related to WER, thus ameliorating the mismatch problem between training and testing modes. In general, the proposed MSRL system not only guarantees the lower-bound performance in noisy conditions but also improves the upper-bound performance in clean conditions. In order to visualize the effect of modality-specific representations, we draw the histogram of _Audio-only_ baseline, _Modality-invariant_ baseline, and MSRL system in Fig. 3. The _Visual-only_ baseline is shown as the blue dashed line as the WER keeps invariant (26.9%) in all conditions. It is noted that Fig. 3 roughly reproduces the situation in Fig. 1. The _Modality-invariant_ baseline loses its effectiveness in low SNR setting, while MSRL system performs similarly to the oracle line in all conditions. We also conduct a case study to observe how the visual modality-specific representations help the MSRL system. To this end, we sample a divergent step in decoding, where the _Modality-invariant_ baseline predicts a wrong token but MSRL predict the correct one. As shown in Fig 4, three probability distributions are drawn from _Modality-invariant_ baseline, _Visual-only_ baseline, and MSRL system. The x-axis is the vocabulary size and each value denotes a BPE (Sennrich, Haddow, and Birch 2015) token. The y-axis denotes the probability of the corresponding token. For better visualization, the improbable tokens (probability$\textless$ 0.05) are not included in the figure. It is observed that _Modality-invariant_ baseline predict a wrong token (ID=48) in this decoding step, but with help of visual modality-specific representations (_i.e._ , _Visual-only_ baseline), the MSRL predicts the correct token ’ _que_ ’ (ID=582). Figure 4: Case study of a divergent decoding step, where the ground-truth token is ’ _que_ ’ (ID=582). The probabilities higher than 0.05 are displayed. Method | Hr | _Babble_ , SNR= | _Natural_ , SNR= | _Music_ , SNR= | _Speech_ , SNR= | Clean ---|---|---|---|---|---|--- -10 | -5 | 0 | 5 | avg | -10 | -5 | 0 | 5 | avg | -10 | -5 | 0 | 5 | avg | -10 | -5 | 0 | 5 | avg | $\infty$ RNN-T | 34K | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | 4.5 TM-seq2seq | 595 | - | - | 42.5 | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | 7.2 AE-MSR | 1.4K | 38.6 | 31.1 | 25.5 | 24.3 | 29.9 | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | 6.8 AV-HuBERT | 30 | 35.1 | 18.4 | 8.3 | 4.9 | 16.7 | 11.6 | 6.5 | 4.6 | 4.0 | 6.7 | 12.4 | 7.4 | 4.7 | 4.1 | 7.2 | 11.5 | 6.8 | 5.0 | 4.2 | 6.9 | 3.3 AV-HuBERT | 433 | 34.9 | 16.6 | 5.8 | 2.6 | 15.0 | 9.4 | 4.3 | 2.4 | 1.9 | 4.5 | 10.9 | 4.6 | 2.6 | 1.8 | 5.0 | 11.4 | 4.6 | 2.9 | 2.2 | 5.3 | 1.4 MSRL (ours) | 30 | 24.9 | 13.0 | 6.4 | 4.1 | 12.1 | 9.8 | 5.6 | 3.7 | 3.4 | 5.6 | 10.8 | 6.5 | 4.0 | 3.3 | 6.2 | 8.6 | 5.5 | 4.0 | 3.5 | 5.4 | 2.8 MSRL (ours) | 433 | 22.4 | 11.3 | 4.5 | 2.3 | 10.1 | 8.0 | 4.1 | 2.3 | 1.6 | 4.0 | 8.9 | 4.4 | 2.4 | 1.7 | 4.4 | 7.2 | 3.4 | 2.3 | 1.8 | 3.7 | 1.3 Table 4: The WER (%) results of MSRL and prior works on LRS3 dataset.“ _Hr_ ” denotes the the amount of labeled audio-visual speech data used in each system. “ _Babble_ ”, “ _Natural_ ”, and “ _Music_ ” are the different types of noise from MUSAN. “ _Speech_ ” are sampled from other utterances in LRS3. Method | _Cafe_ , SNR= | _Meeting_ , SNR= | _River_ , SNR= | _Resto_ , SNR= | Clean ---|---|---|---|---|--- -10 | -5 | 0 | 5 | avg | -10 | -5 | 0 | 5 | avg | -10 | -5 | 0 | 5 | avg | -10 | -5 | 0 | 5 | avg | $\infty$ Low-resource (30 hours of labeled data) & Large Transformer block AV-HuBERT | 16.4 | 7.5 | 4.7 | 4.0 | 8.2 | 13.6 | 7.3 | 4.9 | 4.1 | 7.5 | 23.6 | 11.0 | 5.9 | 4.4 | 11.2 | 36.8 | 19.9 | 8.3 | 5.1 | 17.5 | 3.3 MSRL (ours) | 13.0 | 6.1 | 3.9 | 3.1 | 6.5 | 11.1 | 6.4 | 4.4 | 3.4 | 6.3 | 18.5 | 9.5 | 5.0 | 3.7 | 9.2 | 24.5 | 16.3 | 7.0 | 4.3 | 13.0 | 2.8 Normal-resource (433 hours of labeled data) & Large Transformer block AV-HuBERT | 13.1 | 4.8 | 2.6 | 1.9 | 5.6 | 12.4 | 5.4 | 3.0 | 2.2 | 5.8 | 21.0 | 8.3 | 3.6 | 2.4 | 8.8 | 35.9 | 17.4 | 5.9 | 2.8 | 15.5 | 1.4 MSRL (ours) | 11.2 | 4.2 | 2.3 | 1.7 | 4.9 | 10.4 | 4.5 | 2.6 | 1.8 | 4.8 | 17.8 | 7.8 | 3.2 | 1.9 | 7.7 | 23.9 | 13.9 | 5.1 | 2.4 | 11.3 | 1.3 Table 5: The WER (%) results of MSRL on unseen noises. “ _Cafe_ ”, “ _Meeting_ ”, “ _River_ ”, and “ _Resto_ ” are the different types of noise from DEMAND. ### 5.2 Effect of RL Module In this part, we examine the effect of RL module by replacing it using other integration methods, which are _early fusion_ , _late fusion_ , and _Model ensemble_. Since the pre-trained vision model also has the Transformer-based encoder-decoder pipeline, the _Early fusion_ adds the hidden features from the final encoder layer of pre-trained vision model to the corresponding layer of pre-trained AV encoder. The _Late fusion_ (Inaguma et al. 2019) executes a similar operation but add features at the final layer of decoder. Both early and late fusion strategies are applied in the cross-entropy training (step 3 in Algorithms 1), where the decoder is trainable to fit the new features. The _Model ensemble_ method directly computes the average of probabilities from _Modality-invariant_ baseline and _Visual-only_ baseline for token prediction in auto-regressive decoding, without any tuning operation. From the WER results of Table 3, in noisy conditions, all integration methods can benefit from visual modality-specific representations compared with _Modality-invariant_ baseline. MSRL achieves best performance in all SNR levels. Surprisingly, the untrained _Model ensemble_ baseline beats the _Early fusion_ and _Late fusion_ baselines on average in both normal-resource and low-resource modes. When SNR is -15, except MSRL, three other baselines are not able to avoid the problematic situation that perform worse than _Visual- only_ baseline. In clean conditions, however, the visual modality-specific representations might be redundant. We observe that _Model ensemble_ baseline overestimates the importance of visual modality-specific representations, thus suffering 55.9% of performance deterioration (1.45% $\xrightarrow[]{}$ 2.26%) in normal-resource mode. _Early fusion_ and _Early fusion_ can dilute it by tunable parameters, thereby obtaining comparable WER results with _Modality- invariant_ baseline. In general, MSRL can reasonably balance the importance of modality-specific and modality-invariant representations, as the policy network always considers acoustic information in auto-regressive decoding. ### 5.3 Benchmark against Other Methods We then report the WER performance of MSRL in various conditions, as well as comparing it with other competitive methods. Four recent published methods are selected as strong baselines, which are RNN-T (Makino et al. 2019), TM-seq2seq (Afouras, Chung, and Zisserman 2018a), AE-MSR (Xu et al. 2020), and AV-HuBERT (Shi et al. 2022). Since RNN-T and TM-seq2seq methods focus on clean conditions, and the AE-MSR is only evaluated on _babble_ noise, we only report the available results from their respective papers. For AV-HuBERT, the “ _babble_ ”, “ _speech_ ”, and “clean” columns present the WER results from original paper. The “ _natural_ ” and “ _music_ ” columns were reproduced using the official code as they are not available in original paper. The comparison of WER results is shown in Table 4. In clean conditions, we observe that MSRL achieves 5% (1.4% $\xrightarrow[]{}$1.33%) relative WER reduction over the best baseline of AV- HuBERT in normal-resource mode. In low-resource mode, such superiority increases to 14.5% (3.3% $\xrightarrow[]{}$2.82%). It indicates that visual information is particularly important when training data is limited. Furthermore, the MSRL using 30 hours of labeled data even performs better than RNN-T employs 34k hours of labeled data, which shows better data efficiency. In noisy conditions, MSRL achieves the best performances in all kinds of noises and SNR levels. For the “ _babble_ ”, “ _natural_ ”, “ _music_ ” and “ _speech_ ” noises, MSRL respectively surpasses AV-HuBERT baseline by 32.5%/27.5%, 11.1%/15%, 12.0%/19.5% and 30.2%/21.7% relatively in normal- resource/low-resource mode. It is noted that the _speech_ noise is the utterance drawn from the same source of LRS3 which might confuse the recognizer, while the MSRL can address it well without any separation module. ### 5.4 Generalization on Unseen Noise Finally, we evaluate the generalization of MSRL method, as the AVSR model usually encounters unseen noises in practical applications. We test the AV- HuBERT and MSRL models on a customized test set which contains 4 types of unseen noises, and the WER results are shown in Table 5. We observe MSRL system has better generalization on all 4 kinds of noises. The visual modality-specific representations are still effective as they are unaffected by the domain shift of audio modality. Consequently, MSRL respectively surpasses the AV-HuBERT baseline by 20.7%/12.5%, 16.4%/17.2%, 17.9%/12.5% and 25.7%/27.1% relatively in low-resource/normal-resource mode. Furthermore, we notice that models show distinct adaptability to different unseen noises. Since the “ _cafe_ ” and “ _meeting_ ” noises mainly consist of human voice, both AV-HuBERT and MSRL adapt them well and achieve comparable WER results with seen “ _speech_ ” noise. However, the WER performance degrades obviously on “ _river_ ” and “ _resto_ ” noises, as there is no similar seen noise during training process. ## 6 Conclusion In this paper, we propose a reinforcement learning-based method MSRL to leverage the modality-specific representations into AVSR task. MSRL employs a pre-trained vision model to provide the visual modality-specific and a policy network to explore the optimal integrated strategy in auto-regressive decoding process. We design the experiments to examine the effects of both visual modality-specific representations and RL integration module. WER results demonstrate that MSRL achieves state-of-the-art performance on LRS3 dataset in clean and noisy conditions, as well as showing better generalization on unseen noises. ## 7 Acknowledgement This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-01-002). ## References * Afouras, Chung, and Zisserman (2018a) Afouras, T.; Chung, J. S.; and Zisserman, A. 2018a. The conversation: Deep audio-visual speech enhancement. _arXiv preprint arXiv:1804.04121_. * Afouras, Chung, and Zisserman (2018b) Afouras, T.; Chung, J. S.; and Zisserman, A. 2018b. LRS3-TED: a large-scale dataset for visual speech recognition. _arXiv preprint arXiv:1809.00496_. * Bahdanau et al. (2016) Bahdanau, D.; Brakel, P.; Xu, K.; Goyal, A.; Lowe, R.; Pineau, J.; Courville, A.; and Bengio, Y. 2016. An actor-critic algorithm for squence prediction. * Chen et al. (2022a) Chen, C.; Hou, N.; Hu, Y.; Shirol, S.; and Chng, E. S. 2022a. Noise-robust speech recognition with 10 minutes unparalleled in-domain data. In _ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)_ , 4298–4302. IEEE. * Chen et al. (2022b) Chen, C.; Hou, N.; Hu, Y.; Zou, H.; Qi, X.; and Chng, E. S. 2022b. Interactive Audio-text Representation for Automated Audio Captioning with Contrastive Learning. _arXiv preprint arXiv:2203.15526_. * Chen et al. (2022c) Chen, C.; Hu, Y.; Hou, N.; Qi, X.; Zou, H.; and Chng, E. S. 2022c. Self-Critical Sequence Training for Automatic Speech Recognition. In _ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)_ , 3688–3692. IEEE. * Feng, Lai, and Xie (2019) Feng, Z.; Lai, J.; and Xie, X. 2019. Learning modality-specific representations for visible-infrared person re-identification. _IEEE Transactions on Image Processing_ , 29: 579–590. * Fu, Wu, and Boulet (2022) Fu, Y.; Wu, D.; and Boulet, B. 2022. Reinforcement learning based dynamic model combination for time series forecasting. In _Proceedings of the AAAI Conference on Artificial Intelligence_ , volume 36, 6639–6647. * Gulati et al. (2020) Gulati, A.; Qin, J.; Chiu, C.-C.; Parmar, N.; Zhang, Y.; Yu, J.; Han, W.; Wang, S.; Zhang, Z.; Wu, Y.; et al. 2020. Conformer: Convolution-augmented transformer for speech recognition. _arXiv preprint arXiv:2005.08100_. * Hazarika, Zimmermann, and Poria (2020) Hazarika, D.; Zimmermann, R.; and Poria, S. 2020. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In _Proceedings of the 28th ACM international conference on multimedia_ , 1122–1131. * He et al. (2016) He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In _Proceedings of the IEEE conference on computer vision and pattern recognition_ , 770–778. * Hu et al. (2022a) Hu, Y.; Hou, N.; Chen, C.; and Chng, E. S. 2022a. Dual-Path Style Learning for End-to-End Noise-Robust Speech Recognition. _arXiv preprint arXiv:2203.14838_. * Hu et al. (2022b) Hu, Y.; Hou, N.; Chen, C.; and Chng, E. S. 2022b. Interactive feature fusion for end-to-end noise-robust speech recognition. In _ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)_ , 6292–6296. IEEE. * Inaguma et al. (2019) Inaguma, H.; Cho, J.; Baskar, M. K.; Kawahara, T.; and Watanabe, S. 2019. Transfer learning of language-independent end-to-end ASR with language model fusion. In _ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)_ , 6096–6100. IEEE. * Ma, Petridis, and Pantic (2021) Ma, P.; Petridis, S.; and Pantic, M. 2021. End-to-end audio-visual speech recognition with conformers. In _ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)_ , 7613–7617. IEEE. * Makino et al. (2019) Makino, T.; Liao, H.; Assael, Y.; Shillingford, B.; Garcia, B.; Braga, O.; and Siohan, O. 2019. Recurrent neural network transducer for audio-visual speech recognition. In _2019 IEEE automatic speech recognition and understanding workshop (ASRU)_ , 905–912. IEEE. * Mei et al. (2021) Mei, X.; Huang, Q.; Liu, X.; Chen, G.; Wu, J.; Wu, Y.; Zhao, J.; Li, S.; Ko, T.; Tang, H. L.; et al. 2021. An encoder-decoder based audio captioning system with transfer and reinforcement learning. _arXiv preprint arXiv:2108.02752_. * Mittal et al. (2020) Mittal, T.; Bhattacharya, U.; Chandra, R.; Bera, A.; and Manocha, D. 2020. M3er: Multiplicative multimodal emotion recognition using facial, textual, and speech cues. In _Proceedings of the AAAI conference on artificial intelligence_ , volume 34, 1359–1367. * Noda et al. (2015) Noda, K.; Yamaguchi, Y.; Nakadai, K.; Okuno, H. G.; and Ogata, T. 2015. Audio-visual speech recognition using deep learning. _Applied Intelligence_ , 42(4): 722–737. * Petridis et al. (2018) Petridis, S.; Stafylakis, T.; Ma, P.; Tzimiropoulos, G.; and Pantic, M. 2018. Audio-visual speech recognition with a hybrid ctc/attention architecture. In _2018 IEEE Spoken Language Technology Workshop (SLT)_ , 513–520. IEEE. * Rennie et al. (2017) Rennie, S. J.; Marcheret, E.; Mroueh, Y.; Ross, J.; and Goel, V. 2017. Self-critical sequence training for image captioning. In _Proceedings of the IEEE conference on computer vision and pattern recognition_ , 7008–7024. * Schulman et al. (2015) Schulman, J.; Levine, S.; Abbeel, P.; Jordan, M.; and Moritz, P. 2015. Trust region policy optimization. In _International conference on machine learning_ , 1889–1897. PMLR. * Sennrich, Haddow, and Birch (2015) Sennrich, R.; Haddow, B.; and Birch, A. 2015. Neural machine translation of rare words with subword units. _arXiv preprint arXiv:1508.07909_. * Shi et al. (2022) Shi, B.; Hsu, W.-N.; Lakhotia, K.; and Mohamed, A. 2022. Learning audio-visual speech representation by masked multimodal cluster prediction. _arXiv preprint arXiv:2201.02184_. * Shi, Hsu, and Mohamed (2022) Shi, B.; Hsu, W.-N.; and Mohamed, A. 2022. Robust Self-Supervised Audio-Visual Speech Recognition. _arXiv preprint arXiv:2201.01763_. * Snyder, Chen, and Povey (2015) Snyder, D.; Chen, G.; and Povey, D. 2015. Musan: A music, speech, and noise corpus. _arXiv preprint arXiv:1510.08484_. * Song, Sun, and Li (2022) Song, Q.; Sun, B.; and Li, S. 2022. Multimodal Sparse Transformer Network for Audio-Visual Speech Recognition. _IEEE Transactions on Neural Networks and Learning Systems_. * Song et al. (2022) Song, T.; Xu, Q.; Ge, M.; Wang, L.; Shi, H.; Lv, Y.; Lin, Y.; and Dang, J. 2022\. Language-specific Characteristic Assistance for Code-switching Speech Recognition. _arXiv preprint arXiv:2206.14580_. * Tao and Busso (2018) Tao, F.; and Busso, C. 2018. Aligning audiovisual features for audiovisual speech recognition. In _2018 IEEE International Conference on Multimedia and Expo (ICME)_ , 1–6. IEEE. * Thiemann, Ito, and Vincent (2013) Thiemann, J.; Ito, N.; and Vincent, E. 2013. The diverse environments multi-channel acoustic noise database (demand): A database of multichannel environmental noise recordings. In _Proceedings of Meetings on Acoustics ICA2013_ , volume 19, 035081\. Acoustical Society of America. * Tjandra, Sakti, and Nakamura (2019) Tjandra, A.; Sakti, S.; and Nakamura, S. 2019. End-to-end speech recognition sequence training with reinforcement learning. _IEEE Access_ , 7: 79758–79769. * Vaswani et al. (2017) Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, Ł.; and Polosukhin, I. 2017. Attention is all you need. _Advances in neural information processing systems_ , 30. * Williams (1992) Williams, R. J. 1992. Simple statistical gradient-following algorithms for connectionist reinforcement learning. _Machine learning_ , 8(3): 229–256. * Xiong et al. (2020) Xiong, H.; Ou, W.; Yan, Z.; Gou, J.; Zhou, Q.; and Wang, A. 2020. Modality-specific matrix factorization hashing for cross-modal retrieval. _Journal of Ambient Intelligence and Humanized Computing_ , 1–15. * Xu et al. (2020) Xu, B.; Lu, C.; Guo, Y.; and Wang, J. 2020. Discriminative multi-modality speech recognition. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , 14433–14442. * Xu et al. (2022) Xu, Q.; Song, T.; Wang, L.; Shi, H.; Lin, Y.; Lv, Y.; Ge, M.; Yu, Q.; and Dang, J. 2022. Self-Distillation Based on High-level Information Supervision for Compressing End-to-End ASR Model. In _Proc. Interspeech 2022_ , 1716–1720. * Yao and Mihalcea (2022) Yao, Y.; and Mihalcea, R. 2022. Modality-specific Learning Rates for Effective Multimodal Additive Late-fusion. In _Findings of the Association for Computational Linguistics: ACL 2022_ , 1824–1834. * Zhou et al. (2019) Zhou, P.; Yang, W.; Chen, W.; Wang, Y.; and Jia, J. 2019. Modality attention for end-to-end audio-visual speech recognition. In _ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)_ , 6565–6569. IEEE.
11institutetext: Center for Modeling Social Systems (CMSS), NORCE Norwegian Research Center AS, Universitetsveien 19, Kristiansand, Norway 11email<EMAIL_ADDRESS> https://www.norceresearch.no/en/research-group/cmss # A Guide to Re-Implementing Agent-based Models: Experiences from the HUMAT Model Önder Gürcan 0000-0001-6982-5658 Timo Szczepanska 0000-0003-2442-8223 Patrycja Antosz 0000-0001-6330-1597 ###### Abstract Replicating existing agent-based models poses significant challenges, particularly for those new to the field. This article presents an all- encompassing guide to re-implementing agent-based models, encompassing vital concepts such as comprehending the original model, utilizing agent-based modeling frameworks, simulation design, model validation, and more. By embracing the proposed guide, researchers and practitioners can gain a profound understanding of the entire re-implementation process, resulting in heightened accuracy and reliability of simulations for complex systems. Furthermore, this article showcases the re-implementation of the HUMAT socio- cognitive architecture, with a specific focus on designing a versatile, language-independent model. The encountered challenges and pitfalls in the re- implementation process are thoroughly discussed, empowering readers with practical insights. Embrace this guide to expedite model development while ensuring robust and precise simulations. ###### Keywords: Agent-based Models Replication Re-implementation Simulation design Model calibration Model validation Best practices ## 1 Introduction Recognizing the need to build higher quality social simulation tools, the scientific community has made numerous efforts to develop procedures that improve description [16], reusability [28], rigor and transperency [3], and increase confidence in agent-based model (ABM) outputs. One essential procedure that deserves more attention as an external model validation method is model replication - re-implementing an existing model based on a representation provided by model builders (e.g., in the form of a natural language conceptual description or source code). Even though agent-based modelers have early on recognized replication as "one of the hallmarks of cumulative science" [10] and was proposed, alongside verification and validation, as an independent test of a model’s reliability [33], it is most often brought to attention in negative instances of a failure to reproduce results of the original model (e.g.,[31]). Since ABMs provide explicit causal explanations of investigated phenomena [9], replication is vital in validating the model’s causal consistency. After all, a causal mechanism represented in the ABM is expected to produce the same effects regardless of the language/software of the implementation. However, if it fails to do so, jumping to conclusions about a widespread replication crisis in social simulation (similar to the one in psychology [22] might be premature, given how much we still have to learn about the specificity of agent-based modeling as a scientific method. Re-implementing a conceptually identical model in a novel platform can help validate the causal mechanisms explaining the model outcomes and identify software-implicit assumptions that are not an explicit part of the conceptual causal explanation but influence the model outcomes (e.g., [14]). Re-implementing existing ABMs in another programming language is a crucial task for researchers and practitioners seeking to enhance the flexibility and scalability of their simulations. Until now, various studies have emphasized the importance of re-implementing agent-based models in different programming languages [26, 14, 4, 29, 20]. Railsback [26] emphasizes the need for re- implementing models in diverse programming languages to capture and represent the complexity of real-world systems. Edmond and Hales [14] state that replication can reveal surprising flaws, even in the simplest of models. Chattoe-Brown et al. [12] emphasize that ensuring such replication becomes even more critical when the model outcomes have the potential to impact individuals’ lives significantly. Unfortunately, replication of ABMs is underused in practice. Zhong and Kim [33] elaborate on possible challenges that explain why re-implementation is still rare. They emphasize that replication is a highly resource-demanding activity with relatively low payoffs in the form of publishable articles, sometimes seen as a trivial activity given to students who take their first steps in coding. This article attempts to aid in building procedures that support replication [27, 30, 32], recognizing the importance of the original research process that starts with the conceptual model. The aim is to report on a systematic process of model replication, sharing good practices and lessons learned from re- implementing the HUMAT socio-cognitive architecture in Python (following an original implementation in NetLogo). The following section introduces a systematic guide for re-implementing agent-based models - a step-by-step process of model re-implementation. We developed this guide alongside the re- implementing HUMAT in Python case study. Effort was taken to generalize the re-implementation process. The guide proposed here serves as a starting point, aimed to be further developed. The article concludes with a short discussion. ## 2 Guide for Re-implementing Agent-Based Models Re-implementing an existing agent-based model in a different programming language involves a series of steps to ensure the new implementation is accurate, efficient, and maintainable. We propose the following systematic approach to guide the re-implementation process (Figure 1), summarized in the most important steps below. Understand the original model: Before beginning any re-implementation, it is essential to clearly understand the existing model’s functionality and designs [24]. This allows the developer to identify potential issues or limitations that should be addressed in the new implementation. Hence, we need to start by studying the original model’s documentation, code, and any related publications and gain a thorough understanding of its objectives, assumptions, agents, behaviors, interactions, environment, and other relevant aspects. Figure 1: The process for re-implementing ABM Models Design a generic model: If the original model’s documentation is tightly coupled with the original programming language, we must outline a generic model independent of a programming language and framework. The generic model should describe the objective, assumptions, agents, behaviors, interactions, and environment. In that sense, applying UML and object-oriented patterns [19] and pattern-oriented agent modeling [15] are good candidates. Choose a new programming language111Note that the initiation of this step is independent from initiation of the other steps and can start at any time.: Choosing the correct programming language can significantly impact its success and depends on several factors [25, 23]. The criteria to be considered are the target platform, target users, (if any) partners’ experience/preference, and the language’s community, library, and support strength. Common choices include Python, Java, and NetLogo [1]. Identify appropriate libraries or frameworks: Research and choose libraries or frameworks that are compatible with your chosen programming language and can facilitate agent-based modeling. For example, Mesa for Python [21], Repast for Java [13], or NetLogo’s built-in constructs/extensions. Design the new model: Based on the generic model and considering the chosen language and framework, design a new model representing agents, environments, interactions, and behaviors. Consider whether any modifications, adaption of the data structures, or optimizations should be made to the generic model based on the new programming language’s capabilities. Implement the new model: Translate the design model into the chosen programming language, adapting the structure and syntax as needed. Use the chosen libraries or frameworks to help streamline the process. Validate the new model: Test the new model against the original to ensure it produces the same or similar results [18, 17]. This may involve comparing outputs, such as agent behaviors, interactions, aggregate patterns, and any performance metrics. Address any discrepancies or issues that arise. Document the new model: Create thorough documentation for the new model, including explanations of its purpose, assumptions, agents, behaviors, interactions, and environment. In that sense, the ODD protocol [16] or UML- based specifications [19] can be used. Include information on any changes or optimizations made during the re-implementation process. Share and collaborate: Share the new model with the original model’s authors and the broader research community through platforms like CoMSES222CoMSES Model Library, https://www.comses.net/codebases/, last access on 11/05/2023., GitLab, GitHub, and through scientific journals and conferences. Solicit feedback, collaborate on improvements, and contribute to the growing body of knowledge in agent-based modeling. ## 3 Case Study: Re-Implementing HUMAT We have chosen a realistic case study to validate the effectiveness of the proposed re-implementation process. In the following, we present how we followed the abovementioned guideline (Section 2) in three subsections. ### 3.1 Choosing the Programming Language and Identifying the Libraries/Frameworks In our case, the need for re-implementation was driven by the goal of the URBANE project333URBANE, https://www.urbane-horizoneurope.eu, last access on 10/05/2023. that requires combining the elements of two different simulation models: HUMAT [5] (implemented in NetLogo) and MASS-GT [11] (implemented in Python) into a single simulation model. Since the target of the resulting model will be used by our partner who knows Python and integrating HUMAT will be easier if we have a Python version, we decided to re-implement HUMAT in Python. NetLogo is a well-documented ABM platform that uses a primary object-oriented language with primitives (predefined keywords) to control agents. Python is a general-purpose, high-level programming language. For the URBANE implementation, we used the Mesa ABM framework [21]. Mesa extends Python with several functionalities to make programming ABMs more manageable. While it is less comprehensive and well-documented than NetLogo, it offers modelers the benefit of accessing many Python libraries. ### 3.2 Understanding HUMAT and Designing its Generic Model To understand HUMAT, we used the available documents and publications [5, 6, 7, 8], and its corresponding NetLogo version (Figure 2). Figure 2: The NetLogo version of HUMAT. As a result of the understanding process, the purpose of the HUMAT model is to represent agents’ socio-cognitive process of attitude formation. The subjects of the attitude – the options an agent decides between (alternative A and alternative B) are decided by the modeler to fit the research problem that the agent-based model investigates. The model is composed mainly of HUMAT agents connected through one or several social networks (i.e., ego networks). Each HUMAT agent is characterized by a set of needs/motives that are important for the subject of the attitude that can belong to one of three groups: experiential needs, social needs, and values. HUMAT agents vary regarding the importance of each motive and how the choice alternatives satisfy each motive. When HUMAT agents form their attitude toward a choice alternative, they reflect on how satisfying that alternative is. If the alternative satisfies one motive and dissatisfies another motive (i.e., has pros and cons), a HUMAT agent experiences an unpleasant state of dissonance. Consequently, that agent faces a dilemma and employs one of two dissonance resolution strategies to maintain cognitive consistency. Suppose the dilemma is non-social (i.e., the social need to be surrounded by enough like-minded HUMATS is satisfied). In that case, the HUMAT inquires - strives to change its own beliefs by asking the most persuasive alter in the ego network for advice. If the dilemma is social (i.e., the social need is dissatisfied), the HUMAT signals to the most gullible alter, striving to persuade them to change their mind. Figure 3: The generic conceptual UML model for HUMAT. To do this effectively, each HUMAT has a representation of all alters linked to it in the ego network. An activated link between HUMAT and the targeted alter denotes a communication act - sharing information about the subject of the attitude (either inquiring or signaling). The persuasiveness of the communicating agent depends on similarity and aspirational characteristics relevant to a given research context. Based on the above understanding, we designed a programming language- independent generic model for HUMAT (Figure 3 and Figure 4). Figure 3 depicts the high-level representations of various concepts in the HUMAT domain and their relationships. Figure 4 represents an overall behavioral model for a HUMAT agent within a social network. The model initializes nodes (HUMATS) and edges in the social network, creating agent instances, and initializing their variables, motives, and choices. Then, it adds the agents to the network, initializes their representations of other agents (alters), and updates their social motives for choices. During each simulation step (tick), agents may decide to signal, inquire, or do nothing based on their current dilemmas and the dissonance strength of their chosen action. If an agent is not satisfied with their choice, they will try to become more content by signalling or inquiring. The basic version of the HUMAT architecture assumes perfect information about alter choices, meaning that all choices are visible to other agents in the ego network. Throughout the simulation, the agents continuously update their alter representations, social motives of choices and make new choices based on their evaluations of motives and dissonance strength. Figure 4: The generic behavioral UML model for HUMAT. ### 3.3 Reimplementing HUMAT in Python Re-implementing HUMAT in Python from the generic conceptual model is a straightforward process. Each concept in the generic model is translated into a Python class with related parameters and methods. The two main classes of the model describe the agents (HumatAgent) and the model (HumatModel) (see Figure 5). Figure 5: Python code for the Signal or Inquire function. The HumatModel class extends the Mesa Model class and controls methods executed during a time step. The HumatAgent class extends the Mesa Agent class and controls the methods executed by the agent. The generic model does not specify which Python data types, syntax, and packages to use. These decisions are up to the modeler and depend on their personal experiences. ### 3.4 Validating HUMAT in Python We start each validation by configuring both models identically by importing all model states of the NetLogo model into Python after initial initialization. Subsequently, automatic unit tests of agent parameters are executed at each time step. This process is repeated, considering increasing agent populations and degrees of randomization (e.g., by controlling scheduling). Throughout the testing, the methods’ functionality is reviewed and optimized. The findings of this comprehensive case study will be documented in a separate paper. ## 4 Discussion Replicating a NetLogo model directly in Python poses some specific challenges in the implementation process. A brief description of the main challenges we faced is given below. More complete comparisons between NetLogo and Python can be found in [2]. * • Object-oriented coding and methods: The NetLogo model is written as a collection of procedures: (i) a setup defines agent parameters and the model environment (e.g., patches and networks), (ii) the main go loop is then executed to run all model procedures for a defined number of time steps (ticks), (iii) the remaining procedures are non-restrictive and written anywhere below the setup and go procedure. The Python model is organized into classes with specific methods: (i) the main class contains methods defining the model inputs and the number of time steps, (ii) the model class controls methods executed during a time step, (iii) the §agent class control the methods that agents execute. In NetLogo, two types of procedures are: to and to-report. The to procedures usually contain a set of commands executed (e.g., by agents), while the to-report returns a value. Functions and methods in Python can execute commands or return values. * • Turtles and Breed vs. Agent classes: NetLogo has four predefined agent types: turtles, patches, links, and the observer. Breeds are used to define specific sub-groups of agents (e.g., HUMATS). Each agent and breed can have specific parameters assigned to it and can be controlled using NetLogo keywords, the primitives. Python, on the other hand, uses classes to define objects. One of the features of the Mesa is the Agent class. Each object created as a sub- class of Agent is automatically equipped with a unique id and a step() method and inherits features. * • Agentset, lists, and dictionaries: In NetLogo, groups of agents are organised in agentsets. These sets of agents can be created on the fly in random order. Agentsets are a very comfortable way to control or select a subset of agents using a set of primitives. While Python can create agent sets, storing agents in dictionaries is often more convenient. Due to these challenges, it is not practical to re-implement a model in Python directly from a NetLogo model. The difference in abstractions used in both languages will make it hard for the modeler to transition. Consequently, for an effective re-implementation and rapid re-implementation in other programming languages, abstracting away the programming language concept and designing a generic model is essential. For instance, thanks to this generic model, we plan to re-implement HUMAT in Java for another project, and it will be pretty rapid. ## 5 Conclusions and Future Work This paper contributes to the literature in three meaningful ways. One, previous studies agree on the importance of replicating agent-based models, however they mostly present experiences on individual models (e.g., [4, 29, 20]). Here, we add to the existing general guidelines [32] by proposing a programming language-independent systematic approach, from understanding the existing model to sharing the new implementation. Two, replications of ABMs focus on discussing the validation of the re- implemented model: to what extent the outputs of the re-implemented model are aligned with the outputs of the original model [10]. The case study of replicating HUMAT described here focuses on the re-implementation process rather than the model outcomes. Three, the authors provide a glimpse of the re-implementation process report having developed a general, conceptual model that is the basis of the original ABM [30] or a platform-independent model [27]. This is a similar approach to developing a generic model proposed here. An intermediate, generic model enables a focus on the investigated phenomenon without anchoring in the concepts present in a given programming language. Additionally, it makes further re-implementations in different languages faster and less effortful. Up until now, we closely followed the guideline until the Validate the New Model step. This remaining step involves sensitivity analysis and testing of the new model and thus requires a more detailed discussion. In future work, we will finish validating the new model implemented in Python and report the results of our experience. We hope that, in future, the guidelines will be used and perfected by the social simulation community. To make re-implementation of ABMs more common, modellers should follow the Share and collaborate step of the proposed guideline. The social simulation community can popularize such works by initiating a dedicated label in COMSES or launching a publication outlet focusing on model evaluation, replication and re-implementation. #### 5.0.1 Acknowledgements The work reported here is part of the URBANE project, which has received funding from the European Union’s Horizon Europe Innovation Action under grant agreement No. 101069782. We thank the reviewers for the thoughtful remarks, especially related to the popularization ideas. ## References * [1] Abar, S., Theodoropoulos, G.K., Lemarinier, P., O’Hare, G.M.P.: Agent Based Modelling and Simulation tools: A review of the state-of-art software. Computer Science Review 24, 13–33 (2017) * [2] Abbott, R., Lim, J.: PyLogo: A Python Reimplementation of (Much of) NetLogo:. In: Proceedings of the 11th International Conference on Simulation and Modeling Methodologies, Technologies and Applications. pp. 199–206. SCITEPRESS - Science and Technology Publications, Online Streaming (2021) * [3] Achter, S., Borit, M., Chattoe-Brown, E., Siebers, P.O.: RAT-RS: a reporting standard for improving the documentation of data use in agent-based modelling. International Journal of Social Research Methodology 25(4), 517–540 (Jul 2022) * [4] An, G., Mi, Q., Dutta-Moscato, J., Vodovotz, Y.: Agent-based models in translational systems biology. Wiley interdisciplinary reviews. Systems biology and medicine 1(2), 159–171 (2009) * [5] Antosz, P., Jager, W., Polhill, J.G., Salt, D., Alonso-Betanzos, A., Sánchez-Maroño, N., Guijarro-Berdiñas, B., Rodríguez, A.: Simulation model implementing different relevant layers of social innovation, human choice behaviour and habitual structures. Tech. Rep. D7.2 (2019) * [6] Antosz, P., Jager, W., Polhill, J.G., Salt, D., Alonso-Betanzos, A., Sánchez-Maroño, N., Guijarro-Berdiñas, B., Rodríguez, A., Scalco, A.: SMARTEES simulation implementations. Tech. Rep. D7.3 (2021) * [7] Antosz, P., Puga-Gonzalez, I., Shults, F.L., Lane, J.E., Normann, R.: Documenting Data Use in a Model of Pandemic “Emotional Contagion” Using the Rigour and Transparency Reporting Standard (RAT-RS). In: Czupryna, M., Kamiński, B. (eds.) Advances in Social Simulation, pp. 439–451. Springer, Cham (2022) * [8] Antosz, P., Puga-Gonzalez, I., Shults, F.L., Szczepanska, T.: HUM-e: An emotive-socio-cognitive agent architecture for representing human decision-making in anxiogenic contexts. In: Squazzoni, F. (ed.) Advances in Social Simulation. Springer International Publishing, Cham * [9] Antosz, P., Szczepanska, T., Bouman, L., Polhill, J.G., Jager, W.: Sensemaking of causality in agent-based models. International Journal of Social Research Methodology 25(4), 557–567 (Jul 2022). https://doi.org/10.1080/13645579.2022.2049510 * [10] Axelrod, R.: Advancing the Art of Simulation in the Social Sciences. In: Simulating Social Phenomena, Lecture Notes in Economics and Mathematical Systems, vol. 456, pp. 21–40. Springer Berlin Heidelberg, Berlin, Heidelberg (1997) * [11] de Bok, M., Tavasszy, L.: An empirical agent-based simulation system for urban goods transport (MASS-GT). Procedia Computer Science 130, 126–133 (2018) * [12] Chattoe-Brown, E., Gilbert, N., Robertson, D.A., Watts, C.: Reproduction as a Means of Evaluating Policy Models: A Case Study of a COVID-19 Simulation. medRxiv (2021). https://doi.org/10.1101/2021.01.29.21250743 * [13] Collier, N.: RePast: An Extensible Framework for Agent Simulation. The University of Chicago’s Social Science Research (2003) * [14] Edmonds, B., Hales, D.: Replication, Replication and Replication: Some Hard Lessons from Model Alignment (Oct 2003) * [15] Grimm, V., Railsback, S.F.: Pattern-oriented modelling: a ‘multi-scope’ for predictive systems ecology. Philosophical Transactions of the Royal Society B: Biological Sciences 367(1586), 298–310 (2012) * [16] Grimm, V., Railsback, S.F., Vincenot, C.E., Berger, U., Gallagher, C., DeAngelis, D.L., Edmonds, B., Ge, J., Giske, J., Groeneveld, J., Johnston, A.S.A., Milles, A., Nabe-Nielsen, J., Polhill, J.G., Radchuk, V., Rohwäder, M.S., Stillman, R.A., Thiele, J.C., Ayllón, D.: The ODD Protocol for Describing Agent-Based and Other Simulation Models: A Second Update to Improve Clarity, Replication, and Structural Realism. Journal of Artificial Societies and Social Simulation 23(2), 7 (2020) * [17] Gürcan, O., Dikenelli, O., Bernon, C.: Towards a Generic Testing Framework for Agent-Based Simulation Models. In: Ganzha, M., Maciaszek, L.A., Paprzycki, M. (eds.) FedCSIS 2011. pp. 635–642. Szczecin, Poland (Sep 2011) * [18] Gürcan, O., Dikenelli, O., Bernon, C.: A generic testing framework for agent-based simulation models. In: Agent-Based Modeling and Simulation, pp. 231–270. Springer (2014) * [19] Larman, C.: Applying UML and Patterns: An Introduction to Object-Oriented Analysis and Design and Iterative Development (3rd Edition). Prentice Hall, USA (2004) * [20] Liang, H., Fu, K.w.: Testing Propositions Derived from Twitter Studies: Generalization and Replication in Computational Social Science. PLOS ONE 10(8), 1–14 (Aug 2015). https://doi.org/10.1371/journal.pone.0134270 * [21] Masad, D., Kazil, J.: Mesa: An Agent-Based Modeling Framework. pp. 51–58. Austin, Texas (2015). https://doi.org/10.25080/Majora-7b98e3ed-009 * [22] Maxwell, S.E., Lau, M.Y., Howard, G.S.: Is psychology suffering from a replication crisis? Am Psychol. 70(6), 487–498 (2015) * [23] North, M.J., Macal, C.M.: Agent Based Modeling and Computer Languages. In: Meyers, R.A. (ed.) Encyclopedia of Complexity and Systems Science, pp. 131–148. Springer New York (2009). https://doi.org/10.1007/978-0-387-30440-3_8 * [24] Pressman, R., Maxim, B.: Software Engineering: A Practitioner’s Approach, 8th Ed (Jan 2014) * [25] Railsback, S., Grimm, V.: Agent-Based and Individual-Based Modeling: A Practical Introduction. Agent-based and Individual-based Modeling: A Practical Introduction, Princeton University Press (2019) * [26] Railsback, S.F.: Concepts from complex adaptive systems as a framework for individual-based modelling. Ecological Modelling 139(1), 47–62 (2001) * [27] Sansores, C., Pavón, J.: Agent-Based Simulation Replication: A Model Driven Architecture Approach. In: MICAI 2005: Advances in AI 4th Mexican Int. Conf. on AI, LNAI, vol. 3789, pp. 244–253. Springer (2005) * [28] Tang, W., Grimm, V., Tesfatsion, L., Shook, E., Bennett, D., An, L., Gong, Z., Ye, X.: Code Reusability and Transparency of Agent-Based Modeling: A Review from a Cyberinfrastructure Perspective. In: Tang, W., Wang, S. (eds.) High Performance Computing for Geospatial Applications, pp. 115–134. Springer, Cham (2020) * [29] Thiele, J.C., Kurth, W., Grimm, V.: Facilitating Parameter Estimation and Sensitivity Analysis of Agent-Based Models: A Cookbook Using NetLogo and ’R’. Journal of Artificial Societies and Social Simulation 17(3), 11 (2014) * [30] Wilensky, U., Rand, W.: Making Models Match: Replicating an Agent-Based Model. Journal of Artificial Societies and Social Simulation 10(4) (2007) * [31] Will, O., Hegselmann, R.: A Replication That Failed: On the Computational Model in ’Michael W. Macy and Yoshimichi Sato: Trust, Cooperation and Market Formation in the U.S. and Japan. JASSS 11(3) (2008) * [32] Zhang, J., Robinson, D.T.: Replication of an agent-based model using the Replication Standard. Environmental Modelling & Software 139, 105016 (2021) * [33] Zhong, W., Kim, Y.: Using Model Replication to Improve Reliability of Agent-Based Models. In: Chai, S.K., Salerno, J.J., Mabry, P.L., Hutchison, D., Kanade, T. (eds.) Advances in Social Computing: 3rd Int. Conf. on Social Computing, Behavioral Modeling, and Prediction, LNCS, vol. 6007, pp. 118–127. Springer (2010)
# Evolution of high-redshift quasar hosts and promotion of massive black hole seed formation Wenxiu Li (李文秀) Kavli Institute for Astronomy and Astrophysics, Peking University, Beijing 100871, China Kohei Inayoshi Kavli Institute for Astronomy and Astrophysics, Peking University, Beijing 100871, China Yu Qiu (邱宇) Kavli Institute for Astronomy and Astrophysics, Peking University, Beijing 100871, China ###### Abstract High-redshift luminous quasars powered by accreting supermassive black holes (SMBHs) with mass $\gtrsim 10^{9}~{}M_{\odot}$ constrain their formation pathways. We investigate the formation of heavy seeds of SMBHs through gas collapse in the quasar host progenitors, using merger trees to trace the halo growth in highly-biased, overdense regions of the universe. The progenitor halos are likely irradiated by intense H2-photodissociating radiation from nearby star-forming galaxies and heat the interior gas by successive mergers. The kinetic energy of the gas originating from mergers as well as baryonic streaming motion prevents gas collapse and delays prior star formation. With a streaming velocity higher than the root-mean-square value, gas clouds in nearly all $10^{4}$ realizations of merger trees enter the atomic-cooling stage and begin to collapse isothermally with $T\simeq 8000~{}{\rm K}$ via Ly$\alpha$ cooling. The fraction of trees which host isothermal gas collapse is $14\%$ and increases with streaming velocity, while the rest form H2-cooled cores after short isothermal phases. If the collapsing gas is enriched to $Z_{\rm crit}\sim 2\times 10^{-3}~{}Z_{\odot}$, requiring efficient metal mixing, this fraction could be reduced by additional cooling via metal fine- structure lines. In the massive collapsing gas, the accretion rate onto a newly-born protostar ranges between $3\times 10^{-3}-5~{}M_{\odot}~{}{\rm yr}^{-1}$, among which a large fraction exceeds the critical rate suppressing stellar radiative feedback. As a result, we expect a distribution of stellar mass (presumably BH mass) ranging from several hundred to above $10^{5}~{}M_{\odot}$, potentially forming massive BH binary mergers and yielding gravitational wave events. Supermassive black holes (1663); Quasars (1319); High-redshift galaxies (734) ## 1 Introduction Supermassive black holes (SMBHs) with masses of $10^{6-9}~{}M_{\odot}$ are one of the most fundamental ingredients on the structure formation paradigm and are believed to coevolve with their host galaxies over the cosmic timescale through gas feeding and feedback processes (Kormendy & Ho, 2013). The existence of high-redshift quasars at $z\gtrsim 6$ suggests that such monster SMBHs form in the first billion years of the cosmic age (Fan et al., 2006; Mortlock et al., 2011; Wu et al., 2015; Jiang et al., 2016; Matsuoka et al., 2018; Onoue et al., 2019; Wang et al., 2021) via rapid assembly processes, such as the formation of heavy BH seeds (initial mass), rapid mass growth via gas accretion, or a combination of the two mechanisms (see a review by Inayoshi et al. 2020). For massive seed BH formation, a sufficiently high accretion rate of gas onto stellar objects is required. In early protogalaxies where the halo virial temperature is as high as $T_{\rm vir}\simeq 10^{4}~{}{\rm K}$ and the temperature of a self-gravitating gas cloud is as warm as that value, the mass accretion rate is expected to be $\dot{M}\simeq c_{\rm s}^{3}/G\simeq 0.1~{}M_{\odot}~{}{\rm yr}^{-1}(T/10^{4}~{}{\rm K})^{3/2}$, where $c_{\rm s}$ is the sound speed of the gas and $G$ is the gravitational constant. To keep the gas warm against efficient cooling via H2 lines, several mechanisms suppressing, delaying, and counteracting H2 formation/cooling have been proposed by many previous studies in literature: photo-dissociation of H2 by Lyman-Werner (LW) radiation (Omukai, 2001; Oh & Haiman, 2002; Shang et al., 2010; Latif et al., 2013; Inayoshi et al., 2014; Sugimura et al., 2014; Regan et al., 2014; Visbal et al., 2014a; Chon et al., 2016), supersonic baryonic streaming motion relative to dark matter (Tanaka & Li, 2014; Hirano et al., 2018; Inayoshi et al., 2018; Schauer et al., 2019), and rapid halo mergers which cause heating (Yoshida et al., 2003; Wise et al., 2019; Lupi et al., 2021) as well as reduce H2 cooling through accretion shocks (Fernandez et al., 2014) All the three processes bring the gas cloud into a dense and hot region on the gas phase diagram, where collisional dissociation from the excited rovibrational levels of H2 reduces the H2 fraction (Inayoshi & Omukai, 2012). In the subsequent stage, the gas collapses almost isothermally, keeping itself as warm as $T\simeq 3000-8000~{}{\rm K}$ and avoiding vigorous gas fragmentation into smaller clumps (Bromm & Loeb, 2003; Latif et al., 2013; Inayoshi et al., 2014; Becerra et al., 2015; Chon et al., 2018). Due to global and monolithic collapse of the warm cloud, the embryonic protostar is fed by rapidly accreting gas at a rate of $\gtrsim 0.1~{}M_{\odot}~{}{\rm yr}^{-1}$ through a compact accretion disk where gas clumps could quickly migrate inward and merge with the central protostar (Inayoshi & Haiman, 2014; Sakurai et al., 2016). Moreover, since the protostar evolves with an expanding stellar envelope due to rapid entropy inject from the accreting matter, the surface temperature is limited to $T_{\rm eff}\simeq 5000~{}{\rm K}$, which is too low for the protostar to emit ionizing radiation (Hosokawa et al., 2013; Haemmerlé et al., 2018). As a result of inefficient radiative feedback, the protostar would likely reach $\sim 10^{5-6}~{}M_{\odot}$ before the end of its lifetime and collapse into a massive seed BH. However, those formation sites of mass seed BHs are expected to be as rare as the number density of high-$z$ quasars in a comoving volume ($n_{\rm SMBH}\sim 1-10~{}{\rm Gpc}^{-3}$ from Willott et al., 2010). Recent cosmological hydrodynamical simulations have suggested that the conditions required to form massive seeds should be more modest than previously considered (e.g., Wise et al., 2019). Even with a moderate level of LW radiation, streaming motion and merger heating, a high mass accretion rate is sustained at larger radii in a protogalaxy, although the isothermality of gas is not maintained at high densities ($n\gtrsim 100~{}{\rm cm}^{-3}$). Under such less stringent situations, the average mass accretion rate onto the central protostar is reduced but the peak rate can exceed the critical rate for bifurcating the protostellar evolution (Latif & Volonteri, 2015; Hirano et al., 2017; Regan et al., 2020b). As a result, the central star grows to the intermediate mass regime at $M_{\star}\simeq 100-10^{4}~{}M_{\odot}$, which is lower than originally expected the expected mass for a SMS but still massive enough to form massive seeds that will end up as high-$z$ SMBHs (Sakurai et al., 2020a, Toyouchi et al. in prep). Therefore, those environmental effects are potentially important to initiate intermediate massive BHs (IMBHs) in the high-$z$ universe by $z\sim~{}6-7$ (Inayoshi et al., 2020), and form gravitational-wave sources for the space-based GW interferometers such as LISA, Taiji, and Tianqin (Sesana et al., 2008; Amaro-Seoane et al., 2017; Bonetti et al., 2019; Dayal et al., 2019; Luo et al., 2016) However, we emphasize that the massive seed forming halos in those scenarios do not necessarily merge into high-$z$ quasar host galaxies. In this paper, we consider a new scenario of the massive seed formation in biased, over-dense regions with $\gtrsim 5$ mass variance, where high-$z$ SMBHs are expected to form (Wyithe & Padmanabhan, 2006). In such intrinsically rare patches of the universe, stronger halo clustering increases the frequency of halo mergers and boosts the mean intensity of LW radiation background in the regions. Therefore, the modest conditions required to form massive seeds with $100-10^{4}~{}M_{\odot}$ will be naturally satisfied there. We generate merger trees of the progenitor halos that end up a high-$z$ quasar host, based on the extended Press-Schechter formalism, and quantify the expected mean LW intensity irradiating the main progenitors and the merger heating rate along with the trees. By taking into account the environmental input, the thermal and dynamical evolution of a massive gas cloud in the main progenitor halo is calculated in a self-consistent way. Among previous studies in literature, Valiante et al. (2016) investigated the origin of SMBHs using semi-analytical models and found massive BHs seeded in the quasar progenitor halos, depending on their environmental effects. Recently, Lupi et al. (2021) also proposed a similar idea that massive seed BH formation would be much more efficient in a biased halo merger tree based on dark matter (DM) only N-body simulation. They found that in an overdense region, a large number of atomic-cooling halos experience successive merger heating that counteracts radiative cooling via H2 lines and potentially promote massive seed formation. However, most of the halos in their samples do not end up in the most massive DM halo that is supposed to be a high-$z$ quasar host. Instead, we study the statistical properties of the progenitor halos of a high-$z$ quasar host by generating merger trees. Moreover, we explicitly follow the evolution of gas clouds in the main progenitors, taking into account merger heating, radiative cooling, and chemical reaction networks. Thus, the two studies are complementary. This paper is organized as follows. In §2, we summarize our construction of merger histories of a quasar host, the calculation of environmental LW intensity for individual halos, and subsequent gas evolution following the underlying halo mass growth. In §3, we discuss the results of LW intensity, the fraction of promising heavy seed formation sites, and the distribution of accretion rate realized. In §4, we quantify the critical metallicity that affects thermal evolution of gas and the efficiency of metal enrichment, and discuss caveats of our model. In §5, we show the mass distribution of seed BHs formed in the high-$z$ quasar progenitors. Finally, in §6, we summarize the main conclusions of this paper. ## 2 methodology In order to investigate the evolution of luminous quasar progenitors that form in rare, overdense regions in the universe at redshift $z\sim 6$, we construct the merger history of DM halos up to $z=50$, and model the evolution of the gas properties within the DM halos along each merger tree. The processes we model consist of three parts: (1) We first construct the hierarchical merger history of a quasar host halo using the Monte Carlo merger tree algorithm. For a $10^{9}\,M_{\odot}$ SMBH powering the luminous quasar at $z\sim 6$, the halo mass is estimated to be $M_{\rm h}\sim 10^{12}~{}M_{\odot}$ by comparing the growth rate of quasar density indicated from observations with that predicted by the Press-Schechter formalism (Wyithe & Padmanabhan, 2006). We therefore focus our analysis on halos that grow to $M_{\rm h}=10^{12}~{}M_{\odot}$ at $z=6$. (2) For a given merger tree, we calculate the LW radiation background produced by the surrounding star-forming galaxies at each redshift, in order to model the radiative impact on the gas within the halo. (3) The evolution of the gas in the parent halo of each tree is studied by taking into account the injection of thermal and kinetic energy due to violent merger events, as well as LW irradiation calculated in step (2) that dissociates the gas coolants. In the following subsections, we describe in detail the three key ingredients. Throughout the paper, we adopt cosmological parameters estimated by Planck assuming a $\Lambda$CDM universe (Planck Collaboration et al., 2016), i.e., $\Omega_{\mathrm{m}}=0.307,~{}\Omega_{\Lambda}=0.693,~{}\Omega_{\mathrm{b}}=0.0486,~{}H_{0}=67.7\mathrm{~{}km}\mathrm{~{}s}^{-1}\mathrm{Mpc}^{-1}$. ### 2.1 Merger histories of progenitors We construct DM merger trees based on the extended Press-Schechter formalism (Press & Schechter, 1974; Lacey & Cole, 1993; Cole et al., 2000) using the GALFORM semi-analytic algorithm summarized in Parkinson et al. (2008). Our sample consists of $10^{4}$ merger tree realizations for the DM halos that end up as high-$z$ quasar hosts with $M_{\rm h}=10^{12}~{}M_{\odot}$ at $z=6$. For each tree, we adopt a minimum DM halo mass of $M_{\rm h,min}=10^{5}~{}M_{\odot}$. Halos smaller than this threshold do not significantly impact the gas evolution, because the critical virial temperature above which gas collapse can be induced by coolant $\mathrm{H_{2}}$ is $\sim 10^{3}~{}{\rm K}$ (see Haiman et al., 1996; Tegmark et al., 1997), corresponding to halo mass higher than $M_{\rm h,min}$ (see also Fig. 1). Reflecting the rarity of quasar host galaxies, the progenitor halos form in highly biased regions with $\gtrsim 5$ mass variance (Mo & White, 2002). Note that the fraction of all matter in such rare halos is $\lesssim 10^{-7}$. ### 2.2 Lyman-Werner background intensity Due to the photo-dissociation of H2 exposed to LW radiation, we also consider the local LW intensity $J_{\rm LW}$ (at $h\nu=12.4\rm{eV}$, hereafter in units of $\rm 10^{-21}erg~{}s^{-1}~{}cm^{-2}~{}Hz^{-1}~{}sr^{-1}$) in order to follow the gas evolution in a given progenitor halo. Along each merger tree, we calculate the cumulative $J_{\rm LW}$ from neighboring star-forming galaxies (hereafter source halos). Based on the model developed by Dijkstra et al. (2014), the basic equations and assumptions we adopt are summarized as below. We consider a DM halo with mass $M_{\rm h}$ (gas + DM) at redshift $z$, which is supposed to be the main progenitor in a merger tree. The average number of source halos (within mass range $m\pm dm/2$) that populate a surrounding spherical shell (at a physical distance $r$ with thickness $dr$) is calculated by $\displaystyle\frac{d^{2}\mathcal{N}(m,r)}{dmdr}dmdr$ $\displaystyle=4\pi r^{2}dr(1+z)^{3}~{}\frac{\mathrm{d}n_{\mathrm{ST}}(m,z)}{\mathrm{d}m}$ $\displaystyle~{}~{}\times[1+\xi(M_{\mathrm{h}},m,z,r)]\mathrm{~{}d}m,$ (1) where $\mathrm{d}n_{\mathrm{ST}}/\mathrm{d}m$ is the mass function of source halos (Sheth et al., 2001), and $\rm\xi$ denotes the non-linear bias function (Iliev et al., 2003), which gives the deviation (from random) probability of finding a halo with mass $m$ at distance $r$ from the main progenitor. We set the minimum source halo mass to be $m_{\mathrm{ac},z}\simeq 6\times 10^{6}M_{\odot}\left(T_{\rm vir}/10^{4}~{}{\rm K}\right)^{3/2}\left[\left(1+z\right)/31\right]^{-3/2}$, where the halo virial temperature is just above the hydrogen atomic-cooling threshold of $T_{\rm vir}=10^{4}~{}{\rm K}$, where radiative cooling by Ly$\alpha$ emission leads to star formation. In our model, we do not consider the production of LW radiation background by star formation activity in less-massive DM halos. The maximum mass of source halos is determined so that the LW intensity converges towards the higher mass bins, namely in terms of averaged flux, contributions from the $m_{\rm max}$ halos vanish due to their low abundance. The value of $m_{\rm max}$ ranges from $\sim 10^{6}M_{\odot}$ to $\sim 10^{10}M_{\odot}$ and is larger at lower $z$. Following Dijkstra et al. (2014), we compute the average LW radiation flux that irradiates the target halo. The time-averaged production rate of LW photons (per unit stellar mass) emitted from a surrounding source galaxy is approximated by $\left\langle Q_{\mathrm{LW}}(t)\right\rangle=Q_{0}\left[1+\left(t_{6}/4\right)\right]^{-3/2}\mathrm{e}^{-t_{6}/300},$ (2) where $Q_{0}=10^{47}\mathrm{~{}s}^{-1}M_{\odot}^{-1}$ and $t~{}(=t_{6}~{}\mathrm{Myr})$ is the time after a single star burst in the star-forming halo. Thus, the specific LW luminosity from the halo is calculated by $L_{\mathrm{LW}}(m_{\star},t)=\frac{h\langle\nu\rangle}{\Delta\nu}\left\langle Q_{\mathrm{LW}}(t)\right\rangle f_{\mathrm{esc},\mathrm{LW}}\left(\frac{m_{\star}}{M_{\odot}}\right),$ (3) where the mean frequency and frequency width of the LW band ($11.2\leq h\nu/{\rm eV}\leq 13.6$) are set to $\langle\nu\rangle=12.4~{}\mathrm{eV}/h$ and ${\Delta\nu}=2.4~{}\mathrm{eV}/h$. The total stellar mass is calculated by $m_{\star}=f_{\star}(\Omega_{\mathrm{b}}/\Omega_{\mathrm{m}})m$, assuming the star formation efficiency to be $f_{\star}=0.05$. The escape fraction of LW photons from the halo is assumed to be unity ($f_{\mathrm{esc},\mathrm{LW}}=1$). This value tends to be lower for atomic- cooling halos with $m\gtrsim 10^{7}~{}M_{\odot}$. As a reference, Schauer et al. (2015) calculated the LW escape fraction for a single PopIII star in an atomic-cooling halo with 1D simulations and found $f_{\mathrm{esc},\mathrm{LW}}\simeq 0.7$. However, this is considered to be a lower bound because the escape fraction would be higher for 3D calculations through directions with lower optical depths, besides a higher SFR is expected in our case (rather than a single massive star). We estimate the LW luminosity at one free-fall time after the burst of star formation: $t_{\mathrm{sf}}=\sqrt{3\pi/(32G\Delta_{\rm vir}\bar{\rho})}\simeq 18~{}{\rm Myr}~{}[(1+z)/31]^{-3/2}$, where $\Delta_{\rm vir}\simeq 18~{}\pi^{2}$. Using Eqs. (1)-(3), we obtain the mean LW radiation intensity in the target halo as $J_{\rm LW}(M_{\mathrm{h}},z)=\int_{m_{\mathrm{ac},z}}^{m_{\rm max}}\int_{r_{\rm min}}^{r_{\rm max}}\frac{d^{2}\mathcal{N}(m,r)}{dmdr}\cdot\frac{L_{\mathrm{LW}}}{16\pi^{2}r^{2}}~{}dmdr,$ (4) where $r_{\rm min}$ and $r_{\rm max}$ are the minimum and maximum distance of the source halo from the target halo. In the absence of metal pollution, $r_{\rm min}$ can be safely set by adding the virial radii of the target and source halos. However, metal enrichment of the main progenitor is a main obstacle in the formation scenario of massive seed BHs, because efficient metal-line cooling (and possibly dust thermal emission) will likely lead to gas fragmentation during its gravitational collapse and thus suppress massive star formation. Generically, there are two types of enrichment processes: (1) genetic enrichment due to past star formation episodes in the progenitors, and (2) environmental enrichment owing to metal bubbles created by supernova (SN) explosions in nearby galaxies. In our model, we consider the environmental enrichment process by adopting the minimum distance to source halos as $r_{\rm min}=\max\\{r_{\rm vir}(M_{\mathrm{h}})+r_{\rm vir}(m),r_{\rm s}(m)\\}$, where $r_{\rm s}$ is the size of the metal-polluted region surrounding the source halo $\displaystyle r_{\mathrm{s}}(m,t)=\left(\frac{E_{\rm SN}m_{\star}}{m_{\rm 0}\rho_{\rm s}}\right)^{1/5}t^{2/5},$ (5) where $m_{0}=100~{}M_{\odot}$ is the stellar mass budget required to form a SN progenitor and $E_{\rm SN}=10^{51}~{}{\rm erg}$ is the explosion energy of a SN. The density $\rho_{\rm s}$ of gas surrounding the wind is considered to be $\Delta\bar{\rho}_{\rm b}$, where $\bar{\rho}_{\rm b}$ is the IGM baryon density, and $\Delta=60$ corresponding to the typical baryonic overdensity of halos at their virial radius for a NFW profile Dijkstra et al. (2014). Similar to the production of LW radiation, we estimate the size of metal-enriched bubbles at $t_{\rm sf}$. We note that metal-enrichment through in-situ star formation should be subdominant because intense LW radiation suppresses star formation in low-mass progenitors (see §4). On the other hand, the maximum distance in the integration is given by $r_{\rm max}=\left(\lambda_{\mathrm{LW},1}-\lambda_{\beta}\right)c/\left[\lambda_{\beta}H(z)\right]$, where the $\lambda_{\mathrm{LW},1}=1110\AA$ and $\lambda_{\beta}$ are wavelengths of the lowest LW energy and Ly$\beta$ line, respectively (see Haiman et al., 1997). We consider the redshift effect by cosmic expansion, where $H(z)=H_{0}\left[\Omega_{\mathrm{m}}(1+z)^{3}+\Omega_{\Lambda}\right]^{1/2}$ is the Hubble constant at redshift $z$ and $c$ is the light speed. LW photons emitted at $r>r_{\rm max}$ are redshifted into one of the Lyman series resonances and are converted into low-energy photons before reaching the target halo. The $r_{\rm max}$ is thus set as an absorbing screen, i.e., we exclude the contribution of $J_{\rm LW}$ from halos located at $r>r_{\rm max}$. ### 2.3 Energy injection through halo mergers The main progenitor halo experiences vigorous halo mergers in the high-$z$ universe. Successive merger events, in particular major mergers, inject energy into the gas in the parent halo. At early phase, energy loss through radiative cooling is inefficient, i.e., the cooling timescale is comparable or longer than the Hubble timescale. Gas is heated through shock formation at the halo virial radius in an adiabatic manner. Subsequently, the energy is transported into the halo interior, leading to gas virialization with a nearly constant temperature profile ($T_{\rm gas}\sim T_{\rm vir}$) across all radii (Wise & Abel, 2007). Assuming that the virial equilibrium state is reached after a merger event, the virial theorem applies to the gas in the post-merger halo, where the internal and kinetic (turbulence) energy is balanced with the gravitational energy as $e_{\rm tot}=e_{\rm th}+e_{\rm k}+\Phi_{R_{\rm vir}}=\frac{1}{2}\Phi_{R_{\rm vir}},$ (6) where $e_{\rm tot}$, $e_{\rm th}$ and $e_{\rm k}$ are the total, thermal, and kinetic energy per unit mass, and $\Phi_{R_{\rm vir}}$ is the gravitational energy at the virial radius. In this work, we adopt the NFW potential for DM halos given by $\Phi_{R_{\rm vir}}=-\frac{2k_{\rm B}T_{\rm vir}}{\mu m_{\rm p}}\cdot\frac{\ln(1+c_{\mathrm{vir}})}{\ln(1+c_{\mathrm{vir}})-c_{\mathrm{vir}}/(1+c_{\mathrm{vir}})},$ (7) where $T_{\rm vir}$ is the halo virial temperature, the concentration parameter of the DM density profile $c_{\mathrm{vir}}=1.9~{}(M_{\rm h}/10^{7}\,M_{\odot})^{-0.13}[(1+z)/31]^{-1}$ (Bullock et al., 2001), $k_{\rm B}$ is the Boltzmann constant, $\mu=1.22$ is the mean molecular weight, and $m_{\rm p}$ is the proton mass. Therefore, the total energy change owing to the halo evolution is given by $\Gamma_{\rm mrg}=-\frac{1}{2}\Phi_{R_{\rm vir}}\left(\frac{2}{3}\frac{\dot{M_{\mathrm{h}}}}{M_{\mathrm{h}}}-\frac{1}{1+z}\frac{dz}{dt}\right),$ (8) where the first term of the right hand side denotes the energy change associated with mass growth and the second term represents the cosmic expansion effect. In the generally turbulent virialized gas, the kinetic-to- thermal energy ratio is equal to 1 around the virial radius, and decreases to $1/3$ at the center (see Wise & Abel, 2007). Adopting this branching ratio of the total injected energy, the thermal and kinetic heating rate associated with mergers are given by $\Gamma_{\rm mrg,th}=3\Gamma_{\rm mrg}/4$ and $\Gamma_{\rm mrg,kin}=\Gamma_{\rm mrg}/4$, respectively. Combining Eqs. (6)-(8), the gas temperature follows the halo virial temperature as $\frac{\dot{T}_{\rm gas}}{\dot{T}_{\rm vir}}=\frac{1}{2}\cdot\frac{\ln(1+c_{\mathrm{vir}})}{\ln(1+c_{\mathrm{vir}})-c_{\mathrm{vir}}/(1+c_{\mathrm{vir}})}.$ (9) This ratio is close to unity for a wide range of ($M_{\rm h}$, $z$) halos of interest, e.g., $\dot{T}_{\rm gas}/\dot{T}_{\rm vir}\simeq 1.3$ and $0.81$ for $c_{\mathrm{vir}}=2$ and $10$. Note that our method is different from that adopted in previous papers (e.g., Yoshida et al., 2003; Lupi et al., 2021), where $T_{\rm gas}=T_{\rm vir}$ is imposed. The treatment allows us to precisely calculate the radiative cooling rates and chemical reaction coefficients, which sensitively depend on the gas temperature. ### 2.4 Turbulence and baryonic streaming motion The kinetic energy injected through mergers is stored in the halo as turbulence. During the viliarization process, turbulence plays an important role on massive star formation (e.g., McKee & Tan, 2002). Namely, turbulence acts as a source of pressure, which stabilizes the gas against its self- gravity and delays the collapse until the cloud becomes massive enough to overcome the turbulent pressure. In addition to turbulence, the baryonic streaming motion relative to the DM produced in the epoch of cosmic recombination at $z_{\rm rec}\simeq 1100$ also significantly delays gas collapse and star formation in protogalaxies. The streaming velocity is found to follow a Maxwell-Boltzmann distribution with the root-mean-square speed of $\sigma=30~{}{\rm km~{}s}^{-1}$ at $z=z_{\rm rec}$ and decays as $\tilde{v}_{\rm bsm}=v_{\rm bsm}(1+z)/(1+z_{\mathrm{rec}})$ (Tseliakhovich & Hirata, 2010). We note that the volume fraction of the universe with streaming velocities of $v_{\rm bsm}\geq A\sigma$ is estimated as $\simeq 0.4$, $8\times 10^{-3}$, and $5.9\times 10^{-6}$ for $A=1$, $2$, and $3$, respectively. Considering both the three-dimensional turbulence and coherent baryonic streaming velocity, we approximate the effective pressure by kinetic motion of gas as $\displaystyle P_{\rm tur}\approx\frac{1}{3}\rho v_{\rm tur}^{2}+\rho\left[\alpha_{0}\tilde{v}_{\rm bsm}(z)\right]^{2},$ (10) where $v_{\rm tur}^{2}=2\int\Gamma_{\rm mrg,kin}dt$ is the kinetic specific energy accumulated through successive mergers and the coefficient of $1/3$ is required to estimate the pressure due to isotropic turbulence (Chandrasekhar, 1951a, b). With pressure support from turbulence, gas collapse is delayed to different extents, with varying strengths of the streaming motion. In this work, we adopt $\alpha_{0}=4.7$ in our fiducial model, in order to match the delay of collapse obtained from cosmological simulations (Hirano et al., 2018). The total gas pressure is therefore defined by $P_{\rm tot}=P_{\rm gas}+P_{\rm tur}$. ### 2.5 Density evolution With the energy injection processes defined above, in this section we describe our model for calculating the density evolution of a gas cloud concentrated in a DM halo that grows through successive merger episodes. There are three characteristic stages of the evolution: (1) initial adiabatic phase, (2) transition to isothermal gas due to radiative cooling, and (3) gravitationally collapsing phase in a runaway fashion. We model the gas dynamics in these stages based on a one-zone model (e.g., Omukai, 2001), which is often used to follow the physical quantities at the center of a gravitationally collapsing cloud with a self-similar density profile $\rho_{\mathrm{gas}}\propto r^{-2}$. However, this profile does not apply to gas in hydrostatic equilibrium before the onset of gravitational collapse. Therefore, we construct a new method to model the three characteristic stages in a physically motivated way. #### 2.5.1 Adiabatic Stage In the early stage, since the gas density is not high enough for radiative cooling to operate through collisionally excited transitions, the gas is adiabatically compressed in the DM halo as the underlying DM gravitational potential evolves. In the DM assembly history through mass accretion, the entropy profile $K(r)$ of the adiabatic gas is characterized by a power-law outer profile of $K(r)=K_{\rm vir}(r/R_{\rm vir})^{1.1}$, and a constant core with $K_{0}\simeq 0.1K_{\rm vir}$, where $K_{\rm vir}=k_{\rm B}T_{\rm vir}/\left[(\mu m_{\rm p})\bar{\rho}_{\rm b}^{2/3}\right]$ is the gas entropy at the virial radius (Voit et al., 2003, 2005). This self-similar entropy profile is also found to be established inside high-$z$ protogalaxies formed in DM halos more massive than $3\times 10^{6}M_{\odot}$ at $z=10$, while the core entropy for less massive halos is maintained at the IGM entropy when gas decouples from the cosmic microwave background (CMB; see more details in Visbal et al., 2014b). Motivated by both numerical simulations and galaxy cluster observations, we approximate the entropy profile as $K(r)\simeq K_{\rm vir}\left(\dfrac{r}{R_{\rm vir}}\right)+K_{0},$ (11) where $K_{0}={\rm max}(0.1K_{\rm vir},K_{\rm IGM})$. Using the entropy profile and the equation of state given by $P_{\rm gas}=K(r)\rho_{\rm gas}^{\gamma}$, where $\gamma=5/3$, we calculate the density profile by solving the hydrostatic equation (the so-called Lane-Emden equation) for the cloud embedded in the DM potential: $\frac{1}{r^{2}}\frac{d}{dr}\left[\frac{r^{2}}{\rho_{\rm gas}}\frac{d(K\rho_{\rm gas}^{\gamma}+P_{\rm tur})}{dr}\right]=-4\pi G\left(\rho_{\rm gas}+\rho_{\rm DM}\right).$ (12) Throughout this paper, we adopt the NFW density profile of dark matter halos of all masses characterized by a simple analytical form of $\rho_{\rm DM}(r)=\rho_{\rm m}(z)\frac{\delta_{0}}{\left(c_{\mathrm{vir}}r/R_{\mathrm{vir}}\right)\left(1+c_{\mathrm{vir}}r/R_{\mathrm{vir}}\right)^{2}},$ (13) where $\rho_{\rm{m}}(z)$ is the mean matter density and $\delta_{0}=\frac{200}{3}\frac{c_{\mathrm{vir}}^{3}}{\ln(1+c_{\mathrm{vir}})-c_{\mathrm{vir}}/(1+c_{\mathrm{vir}})}$ (14) is the characteristic overdensity within halo virial radius (Navarro et al., 1997). We integrate this hydrostatic equation with respect to $\rho_{\rm gas}(r)$ imposing the regularity conditions at the center, i.e., $\rho_{\rm gas}=\rho_{0}$ and $d\rho_{\rm gas}/dr=0$ at $r=0$. Since the solution for adiabatic gas generally has the radius $r_{0}$ where $\rho_{\rm gas}(r_{0})=0$, we determine the central density $\rho_{0}$ so that the enclosed gas mass at $r\leq r_{0}$ satisfies $M_{\rm gas}=f_{\rm b}M_{\rm h}$, where $f_{\rm b}=\Omega_{\rm b}/\Omega_{\rm m}$ is the baryonic fraction. #### 2.5.2 Isothermal Stage As the gas temperature increases due to gravitational compression and merger heating, radiative cooling processes begin to operate in the cloud and the adiabatic assumption no longer applies. When the radiative cooling timescale is shorter than the heating timescale, we solve the hydrostatic equation for the density profile assuming an isothermal equation of state: $\frac{1}{r^{2}}\frac{d}{dr}\left[r^{2}c_{\rm eff}^{2}\frac{d\ln\rho_{\rm gas}}{dr}\right]=-4\pi G\left(\rho_{\rm gas}+\rho_{\rm DM}\right),$ (15) where $c_{\rm eff}\equiv\sqrt{c_{\rm s}^{2}+v_{\rm tur}^{2}/3+\left(\alpha_{0}\tilde{v}_{\rm bsm}\right)^{2}}$ is the effective sound speed developed from the isothermal sound speed $c_{\rm s}\equiv\sqrt{k_{\rm B}T_{\rm gas}/(\mu m_{\rm p})}$. The solution of the isothermal Lane-Emden equation with the regularity condition does not have the radius where the density becomes zero, but connects to the external medium with a density of $\rho_{\rm ext}=f_{\rm b}\rho_{\rm DM}$. The central density is determined so that $\rho_{\rm gas}=\rho_{\rm ext}$ at the virial radius. From the analogy of the Bonnor-Ebert sphere, the isothermal gas cloud embedded in a DM halo potential has a critical mass for the onset of its gravitational collapse. Practically, for a given $T_{\rm gas}$ and $\rho_{\rm DM}(r)$, we construct the density profile with different values of the gas central density $\rho_{0}$ and thus obtain $\rho_{\rm gas}(R_{\rm vir})$ as a function of $\rho_{0}$. Since this function has a local maximum value and the value decreases with increasing halo mass, a hydrostatic equilibrium solution where $\rho_{\rm gas}(R_{\rm vir})=\rho_{\rm ext}$ no longer exists for $M_{\rm h}\geq M_{\rm h,crit}$ (see Appendix A). In this case, the gas evolution is described by the free-fall stage below. #### 2.5.3 Free-fall Stage Once the gas cloud becomes gravitationally unstable, the evolution of the gas density profile is well described by the Penston-Larson self-similar solution (Penston, 1969; Larson, 1969), which has a density profile with a flat core of the Jeans scale and an envelope with a power-law density distribution $\rho_{\mathrm{gas}}(r)\propto r^{-2}$. The central density increases over the free-fall timescale as $\frac{d\rho_{\rm gas}}{dt}=\frac{\rho_{\rm gas}}{t_{\rm ff}},$ (16) where the free-fall timescale is calculated with $t_{\mathrm{ff}}\equiv\sqrt{\frac{3\pi}{32G\left(\rho_{\rm gas}+\langle\rho_{\rm DM}\rangle\right)}},$ (17) where $\langle\rho_{\rm DM}\rangle=\rho_{\rm m}(z)\delta_{0}$ represents the averaged DM density 111 The squared density of a NFW profile averaged within the characteristic radius of $R_{\rm vir}/c_{\mathrm{vir}}$ is given by $\langle\rho^{2}\rangle=\frac{7}{8}\left[\rho_{\rm m}(z)\delta_{0}\right]^{2}$, independent of the concentration factor $c_{\mathrm{vir}}$. . In the collapsing stage, compressional heating by the self-gravitating gas is taken into account and the rate is given by $\displaystyle\Gamma_{\rm comp}\equiv\frac{P_{\rm gas}+P_{\rm tur}}{\rho_{\rm gas}^{2}}\cdot\frac{d\rho_{\rm gas}}{dt}=\frac{c_{\rm eff}^{2}}{t_{\rm ff}}.$ (18) We note that the compressional heating rate is enhanced by turbulent pressure through the effective sound speed. ### 2.6 Temperature and chemical evolution We consider the evolution of thermal and kinetic energy of the gas by solving the two energy equations: $\displaystyle\frac{de_{\rm th}}{dt}$ $\displaystyle=\Gamma_{\rm mrg,th}+\Gamma_{\rm comp}-\mathcal{L}_{\rm chem}-\mathcal{L}_{\rm rad},$ (19) where $\mathcal{L}_{\rm chem}$ is the cooling/heating rate associated with chemical reactions, and $\mathcal{L}_{\rm rad}$ is the radiative cooling rate (note that all the rates are in units of erg s-1 g-1). While the compressional heating rate is included only in the collapse stage, the other effects are taken into account to calculate the gas temperature over the three evolutionary stages. The cooling term includes radiative cooling by H, He, He+, and He++ (Glover & Jappsen, 2007), H2 (Glover & Abel, 2008; Glover, 2015a, b), and cooling/heating associated with chemical reactions. Figure 1: Merger history of the main progenitors of a high-$z$ quasar host galaxy with a DM halo mass of $M_{\rm h}=10^{12}~{}M_{\odot}$ at $z=6$. For a reference, the median halo mass among all the $10^{4}$ trees is shown with the red curve. Three representative merger trees (in terms of growth speed) are highlighted with the blue, orange, and green curves (tree id = 1, 2, and 3). The dotted curves indicate constant virial temperatures, the values of which are denoted by numbers in the figure. We solve the chemical reactions of primordial gas among the following 9 species; H, H2, e-, H+, H${}^{+}_{2}$ , H-, He, He+, and He++. In Table. 1, we show the 35 reaction rate coefficients adopted in this work. In terms of photodissociation of H2, H- and H${}_{2}^{+}$ by external radiation emitted from nearby star-forming galaxies, the reaction rate is calculated by assuming the radiation spectral energy distribution (SED) to be a blackbody spectra with $T_{\rm rad}=2\times 10^{4}~{}{\rm K}$. The SED model approximates more realistic spectra of observed metal-poor star-forming galaxies (Inoue, 2011). The dissociation rates of H- and H${}_{2}^{+}$ are calculated by a convolution with the cross section of the $i$-th chemical species ($i=$ H- and H${}_{2}^{+}$) as $k_{\mathrm{i},\mathrm{pd}}=\int_{0}^{\infty}\frac{4\pi J(\nu)}{h\nu}\sigma_{\mathrm{i}}(\nu)d\nu.$ (20) The cross sections we adopt are from references listed in Table. 1. ## 3 Results Figure 2: Time evolution of LW radiation intensity $J_{\rm LW}$ (in units of $10^{-21}$ erg s-1 cm-2 Hz-1 sr-1) irradiating the quasar progenitors for the four cases shown in Fig. 1. For the median tree, we show two cases where the metal-bubble size $r_{\rm s}$ is calculated as described in §2.2 (solid) and the twice of the fiducial value is adopted (dashed). Figure 3: Distributions of the LW background intensity $J_{\rm LW}$ (in units of $10^{-21}$ erg s-1 cm-2 Hz-1 sr-1) irradiating the quasar progenitors at different epochs ($10\leq z\leq 45$). The mean value of $J_{\rm LW}$ increases from higher redshifts, has a peak of $J_{\rm LW}\simeq 450$ at $z\simeq 25$, and decreases toward lower redshifts. The LW intensity is distributed over a wide range of $10^{-1}\lesssim J_{\rm LW}\lesssim 10^{4}$ at higher redshifts, while the dispersion of the distribution becomes smaller toward lower redshifts. --- Figure 4: Gas density and temperature evolution along with the three representative halo merger trees for the two values of baryonic streaming velocity: $v_{\rm bsm}=1\sigma$ (upper panels) and $v_{\rm bsm}=2\sigma$ (lower panels). The elapsed epochs when the parent halo mass reaches $M_{\rm h}=10^{6}$, $10^{7}$ and $3\times 10^{7}M_{\odot}$ are marked with dots in the left panels, while those when the LW intensity cross $J_{\rm LW}=1$, $10$, $100$ and $1000$ are marked in the right panels. When the halo mass grows faster and/or the streaming velocity is higher, gas collapse is significantly delayed due to pressure (thermal + kinetic) support of the gas cloud. This effect makes the gas enter the atomic-cooling stage at lower densities (H-H2 and H-H cases) owing to strong LW irradiation before the onset of gravitational collapse. ### 3.1 Merger history & evolution of LW radiation background In Fig. 1, we show the evolution of the main progenitors, i.e., the most massive halos at each epoch, for all the $10^{4}$ merger trees that grow to $M_{\rm h}=10^{12}~{}M_{\odot}$ at $z=6$. In such over-dense regions of the universe, the DM halo mass increases via rapid mergers. The median halo mass (dashed curve) reaches $M_{\rm h}\simeq 8\times 10^{10}$, $6\times 10^{8}$, $2\times 10^{7}$, and $8\times 10^{5}~{}M_{\odot}$ at $z=10$, $20$, $30$, and $40$, respectively, and the virial temperature exceeds the atomic-cooling threshold of $T_{\rm vir}\simeq 10^{4}~{}{\rm K}$ at $z\simeq 34$. Therefore, the gas cloud concentrated in the massive halo collapses at an epoch earlier than when typical first-galaxies would form in atomic-cooling halos ($M_{\rm h}\simeq 10^{7}~{}M_{\odot}$ at $z\simeq 10$; see Bromm & Yoshida, 2011), which are usually considered to be massive seed forming sites in most previous studies (e.g., Dijkstra et al. 2014). For illustration purposes, we highlight three merger trees: the blue (id 1, a less massive tree), orange (id 2, a tree comparable to the median evolution), and green curve (id 3, a more massive tree). In the following sections, we focus our analysis on these three representative cases. Following the method laid out in § 2, in Fig. 2 we present the redshift evolution of $J_{\rm LW}$ for the three representative trees and the median track. For all the cases, the LW background intensity gradually increases from higher redshifts, peaks at the intermediate redshifts, and decreases toward lower redshifts. This redshift dependence reflects the nature of the non- linear bias function which boosts the abundance of halo pairs with comparable masses (Scannapieco & Barkana, 2002). Namely, when the mass of the main progenitor is close to the atomic-cooling halo mass ($m_{\mathrm{ac},z}\sim 10^{7}M_{\odot}$), a large number of source halos form nearby owing to the halo clustering effect and thus the LW intensity is maximized. As the main progenitor grows, its mass difference from $m_{\mathrm{ac},z}$ is larger and thus the clustering effect of atomic-cooling sources becomes weaker so that their spacial distribution is approximated to be uniform (i.e., $\xi\ll 1$). As a result, the LW intensity is dominated by the contribution from a large number of atomic-cooling source halos within the absorbing screen ($r\lesssim r_{\rm max}$) and begins to decline due to the cosmic dilution effect at lower redshifts. For rapidly growing progenitor halos exceeding $m_{\mathrm{ac},z}$ earlier, the LW intensity quickly rises at higher redshifts and the peak values become higher owing to stronger clustering at earlier epochs. Namely, the peak values of LW intensity in the overdense regions are $J_{\rm LW}\simeq 60$ (id 1), $J_{\rm LW}\simeq 400$ (id 2), $J_{\rm LW}\simeq 600$ (median), and $J_{\rm LW}\simeq 6\times 10^{3}$ (id 3), which are significantly higher than the level of LW intensity irradiating typical atomic-cooling halos that are expected to form massive BH seeds (see Dijkstra et al., 2008; Agarwal et al., 2012; Johnson et al., 2013). In our semi-analytical approach, we model metal pollution of the progenitor halos due to SN explosions that occur in source halos. Although we treat this effect by replacing the minimum distance between the target and source halos with $r_{\rm s}$, there is no information on the time-dependent spatial distributions of DM halos in our framework. To examine the impact of the model assumptions, in Fig. 2 we also show the case where the size of the metal- polluted bubbles ($r_{\rm s}$) is doubled, the corresponding $t_{\rm sf}$ is comparable to the Hubble time at the redshift, or equivalent to setting $\Delta=1$ with the fiducial value of $t_{\rm sf}$. In this case, the LW intensity is overall reduced at higher redshifts, indicating a significant contribution from nearby source halos with $\gtrsim m_{\mathrm{ac},z}$ to the LW radiation background. We note that our treatment simply removes the contribution from source halos within distances of $r_{\rm s}$, but does not address how likely the main progenitor is affected by environmental metal- enrichment. Our argument nevertheless provides a conservative estimate of $J_{\rm LW}$ if the efficiency of environmental metal-enrichment is low. As discussed in §4, the efficiency should be negligibly low because metal- polluted bubbles rarely penetrate the interior of the target halo (Chiaki et al., 2018). In Fig. 3, we present the histograms of the LW background intensity that irradiates the main progenitor halos for the $10^{4}$ trees at different redshifts. For the whole sample of the target halos in highly-biased regions, the histogram resembles a probability distribution function (PDF) of $J_{\rm LW}$, with the bar height in each bin ($\Delta\log J_{\rm LW}=0.3$) represents the number fraction of halos irradiated within $\log J_{\rm LW}\to\log J_{\rm LW}+\Delta\log J_{\rm LW}$. From higher redshifts down to $z\simeq 30$, the mean value of $J_{\rm LW}$ in the PDF increases owing to a large number of clustered source halos with $\gtrsim m_{\mathrm{ac},z}$ and the $J_{\rm LW}$ distribution peaks around $\simeq 270$ at $z\simeq 25$. Towards lower redshifts, the target halo mass becomes higher than the typical mass of source halos. Therefore the abundance of sources is hardly boosted by the clustering effect (Iliev et al., 2003). Moreover, the LW intensity is diluted by the cosmic expansion, lowering the mean value. While the dispersion of the PDF is larger at higher redshifts, reflecting the diversity of the progenitor mass, the PDF peaks at $J_{\rm LW}\simeq 60$ by $z=10$ when all the $10^{4}$ trees converge to the high-$z$ quasar host. We note that our model does not consider LW radiation produced from DM minihalos with $m<m_{\mathrm{ac},z}$, where $\mathrm{H_{2}}$ is the only coolant to induce star formation. However, strong LW background radiation in the over-dense region likely suppresses its formation. Therefore, the histogram shown in Fig. 3 counts the lower bound of the LW background intensity. Figure 5: Census of merger trees which host the three types of gas collapse with different $v_{\rm bsm}$. The blue, orange and green bars correspond to the representative evolutionary tracks of the same colors in the upper panels of Fig. 4. With increasing $v_{\rm bsm}$, the cases where gas clouds enter the atomic-cooling stage (H-H2 and H-H types) dominate primarily because of the delay of gas collapse that also leads to higher values of LW intensity. Figure 6: Distributions of the halo virial temperature $T_{\rm vir}$ (upper panels) and LW intensity $J_{\rm LW}$ (middle panels) measured at the epochs when gas clouds become gravitationally unstable for the cases with different $v_{\rm bsm}$ values. The lower panels show the mass accretion rate of $\dot{M}_{\star}\equiv c_{\rm eff}^{3}/G$ measured at the minimum temperature point at $n_{\rm gas}>10^{3}{\rm cm}^{-3}$ in the collapsing stage. Overall, with $v_{\rm bsm}\geq 1\sigma$, nearly all the cases enter the atomic-cooling stage in massive halos with $T_{\rm vir}>10^{4}~{}{\rm K}$ irradiated by LW radiation with intensity of $J_{\rm LW}>10$. Since the collapsing clouds are massive, high accretion rates become high enough ($\dot{M}\gtrsim 0.1~{}M_{\odot}~{}{\rm yr}^{-1}$) to form massive seed BHs. ### 3.2 Thermal and dynamical evolution of gas clouds in the high-$z$ quasar hosts In this section we focus our analysis on the gas properties in the main progenitors along the three representative merger trees. In Fig. 4, we show the evolution of gas density (left panels) and temperature (right panels) at the central core as a function of redshift. In order to examine the impact of baryonic streaming motion, for each merger tree we assume two different $v_{\rm bsm}$ values, i.e., $v_{\rm bsm}=1\sigma$ (upper panels), and $2\sigma$ (lower panels). Each curve corresponds to the representative case highlighted in Fig. 1. Along with the three evolutionary tracks, we denote the epochs when the DM halo mass exceeds $M_{\rm h}=10^{6}~{}M_{\odot}$, $10^{7}~{}M_{\odot}$, and $3\times 10^{7}~{}M_{\odot}$ in the left panels, and when the LW background intensity first crosses $J_{\rm LW}=1$, $10$, $10^{2}$, and $10^{3}$ in the right panels. In the following paragraphs, we first describe the gas properties with $v_{\rm bsm}=1\sigma$, and then discuss the impact of the baryonic streaming motion on gas evolution in cases with $v_{\rm bsm}=2\sigma$. For the lowest mass case (blue curve, tree id 1), the gas density gradually increases with the halo mass in the early stage ($z>30$), where the gas cloud is supported by thermal and turbulent pressure against its self-gravity and DM gravitational force. After the halo mass reaches $\simeq 10^{6}~{}M_{\odot}$, the cloud becomes gravitationally unstable owing to its low temperature, and collapses over one free-fall timescale at $z\simeq 28$. The gas temperature remains at $T\lesssim 10^{3}~{}{\rm K}$ due to H2 cooling, under a modest level of LW intensity ($J_{\rm LW}\sim 1$) at $z>35$. In addition to LW radiation, the gas is heated by four major merger events around $z\simeq 31-34$, but the dynamical heating rate does not overcome the H2 cooling rate in this case. For the intermediate mass case (orange curve, tree id 2), the evolution begins from a redshift higher than in the previous case. In this case, the gas temperature is substantially higher as a result of the combination of merger heating and intense LW irradiation with $J_{\rm LW}\gtrsim 1$ in the early stage. As several episodes of halo mergers increase the halo mass to $\sim 10^{7}~{}M_{\odot}$ by $z\simeq 30$ (the corresponding halo virial temperature is $T_{\rm vir}\simeq 10^{4}~{}{\rm K}$), the gas temperature reaches $T\simeq 10^{4}~{}{\rm K}$, where the atomic cooling via Ly$\alpha$ emission begins to operate. Although the LW intensity reaches $J_{\rm LW}\gtrsim 100$ before the cloud gravitationally collapses, the level of LW intensity is not strong enough to suppress H2 formation in the dense region ($\gtrsim 10^{2}~{}{\rm cm}^{-3}$), where H2 reforms owing to its self-shielding effect. As a result of efficient H2 cooling, the gas temperature drops down to $T\simeq 10^{3}~{}{\rm K}$ in the collapsing stage. For the highest mass case (green curve, tree id 3), the gas temperature quickly rises to $T\simeq 10^{4}~{}{\rm K}$ due to frequent mergers. Owing to the clustering effect of the massive parent halo, the LW intensity reaches $J_{\rm LW}\gtrsim 10^{3}$ at $z\simeq 47$, prominently higher than those seen in the less massive cases. Although the H2 self-shielding becomes more effective as the central density increases up to $\gtrsim 10^{4}~{}{\rm cm}^{-3}$, the gas collapses keeping a nearly constant temperature of $T\simeq 8000~{}{\rm K}$. Inside the dense and warm region, H2 is collisionally dissociated and its radiative cooling does not alter the thermal evolution. In cases where $v_{\rm bsm}=2\sigma$, the gas property evolution is shown in the lower panels of Fig. 4. Overall, the collapse of gas clouds is delayed due to kinetic energy injection to the gas concentrated at the halo center. When the cloud begins to collapse, the corresponding halo masses reach $M_{\rm h}\simeq(3.5,~{}4.2,~{}5.9)\times 10^{7}~{}M_{\odot}$. For comparison, the collapse halo masses are $M_{\rm h}\simeq(0.24,~{}2.1,~{}2.2)\times 10^{7}~{}M_{\odot}$ for $v_{\rm bsm}=1\sigma$. The delay effect is more remarkable for the lower-mass cases because the halo circular velocity is lower than the effective sound speed boosted by injection of turbulence and streaming motion. As the gas collapse proceeds, $\mathrm{H_{2}}$ forms efficiently in the modest $J_{\rm LW}$ environment, and eventually its cooling reduces the gas temperature in the low- and intermediate-mass cases. ### 3.3 The statistical properties of the high-$z$ quasar progenitors As noted in §3.2 and Fig. 4, depending on the main cooling processes inducing star formation, the evolutionary tracks of the gas clouds embedded in the main progenitors of high-$z$ quasar hosts are classified into three cases: (i) $\mathrm{H_{2}}$ cooling, (ii) initial H Ly$\alpha$ cooling followed by $\mathrm{H_{2}}$ cooling after a short isothermal collapse, (iii) H Ly$\alpha$ cooling when temperature is kept above $8000~{}{\rm K}$ by compression along a wide density range. In Fig. 5, we present the number count of merger trees for the three types with different baryonic streaming velocities, denoted as (i) H2, (ii) H-H2, and (iii) H-H. Without the streaming velocity, 74% of the trees experience gas collapse via H2 cooling, while the rest ($26\%$) form atomically-cooling gas clouds (cases H-H2 and H-H). With non-zero streaming motion ($v_{\rm bsm}\neq 0$), nearly all cases enter the atomic-cooling stage because the halo mass reaches $m_{\rm ac,z}$ via mergers due to the significant delay effect. As the streaming velocity increases, the gas mass becomes higher at the onset of gravitational collapse, and thus the compressional heating rate during the collapse stage is higher owing to the accumulation of kinetic energy. Therefore, the number of trees where gas isothermally collapses with $T\simeq 8000~{}{\rm K}$ (case H-H) increases monotonically from $14\%$ to $27\%$ with increasing streaming velocity from $v_{\rm bsm}=1\sigma$ to $3\sigma$. In Fig. 6, we show the distributions of the halo virial temperature (upper panels) and LW background intensity (middle panels) for the three types of gas collapse. For each case, the values of $T_{\rm vir}$ and $J_{\rm LW}$ are measured at the epoch when the gas cloud first enters its unstable stage. In contrast to cases with $v_{\rm bsm}=0$, where gas collapse is led by H2 cooling in less massive halos with $T_{\rm vir}\sim 10^{3-4}~{}{\rm K}$, the streaming velocity delays the cloud collapse until after the halo grows across the atomic cooling threshold of $T_{\rm vir}\gtrsim 10^{4}~{}{\rm K}$. The virial temperature for the H-H cases is generally higher than that for the H-H2 cases and the mean value of $T_{\rm vir}$ for each case increases with larger streaming velocity. This trend is more clearly shown in the distributions of $J_{\rm LW}$, namely the mean LW background intensity for the H-H cases is $\langle J_{\rm LW}\rangle\gtrsim 10^{3}$, which is $\simeq 10$ times higher than that for the H-H2 cases. The higher value of $J_{\rm LW}$ is mainly caused by the delay of gas collapse until the halo mass becomes massive enough to be exposed by a larger number of LW source halos. In addition, compressional heating in collapsing clouds is stronger with larger $v_{\rm bsm}$ and the minimum LW intensity required to keep isothermal collapse is extended to lower values. In the main progenitors of high-$z$ quasar hosts, massive gas clouds form owing to the significant delay effect of cloud collapse by rapid halo mergers and intense LW irradiation from nearby star-forming galaxies. The mass accretion rate onto the central region of a gravitationally collapsing cloud is approximated as $\dot{M}\simeq M_{\rm gas}/t_{\rm ff}$, where $M_{\rm gas}$ and $t_{\rm ff}$ are the gas mass and free-fall timescale at the onset of gravitational collapse. Since the cloud is supported by thermal and kinetic energy of the gas, the accretion rate can be written as $\simeq c_{\rm eff}^{3}/G$ (Larson, 1969; Penston, 1969, etc.), which depends only on the gas thermal and kinetic temperature (see below Eq. 15). In the lower panels of Fig. 6, we show the distributions of $\dot{M}\equiv c_{\rm eff}^{3}/G$, for which we adopt the minimum temperature value in the cloud collapse stage at $n\gtrsim 10^{3}~{}{\rm cm}^{-3}$. The accretion rate is broadly distributed over $\dot{M}\simeq 3\times 10^{-3}-5~{}M_{\odot}~{}{\rm yr}^{-1}$. The vertical line in the bottom panels indicates a reference value of $0.1~{}M_{\odot}~{}{\rm yr}^{-1}$, above which the outer envelope of an accreting protostar is bloated due to rapid heat injection through mass accretion and the emission of stellar ionizing photons is strongly suppressed. For $v_{\rm bsm}=1\sigma$, the majority of the H-H2 cases yield $\dot{M}\gtrsim 0.1~{}M_{\odot}~{}{\rm yr}^{-1}$. With $v_{\rm bsm}>1\sigma$, all the cases have sufficiently high accretion rates exceeding the reference value (see more discusssion in § 5). ## 4 Effects of Metal enrichment ### 4.1 Critical Metallicity Figure 7: Evolution of the heating rate (solid) and metal fine-structure line cooling rate (dashed) with gas density for the two representative trees (id 2 and 3) with $v_{\rm bsm}=1\sigma$. The cooling rate consists of CII and OI fine-structure line emission, and the heating rate includes the effect of turbulence and halo mergers. To quantify the critical metallicity for which metal-line cooling dominates heating during the gas collapse, we turn off the H2 cooling rate. The critical metallicity is found to be $Z_{\rm crit}\simeq 1.9\times 10^{-3}~{}Z_{\odot}$ and $2.5\times 10^{-3}~{}Z_{\odot}$ for the tree 2 and 3, respectively. Metal enrichment is considered to be a major obstacle in forming massive BH seeds through star formation because efficient radiative cooling via metal fine-structure lines will induce gas fragmentation and suppress the formation of masive stars. In order to quantify the critical metallicity, we calculate the cooling rate by CII and OI, assuming that the number fractions of carbon and oxygen nuclei in the gas phase with respect to hydrogen nuclei are $x_{\rm C,gas}=0.927\times 10^{-4}(Z/Z_{\odot})$ and $x_{\rm O,gas}=3.568\times 10^{-4}(Z/Z_{\odot})$ (Pollack et al., 1994), and all the carbon and oxygen are in the form of CII and OI, respectively. This treatment is justified for warm gas with $T\simeq 8000~{}{\rm K}$ (Omukai et al., 2008). In Fig. 7, we present the metal-line cooling rate (dashed) and heating rate associated with mergers and gravitational compression (solid) as a function of the density of gas embedded in the two representative progenitor halos (tree id 2 and 3) with $v_{\rm bsm}=1\sigma$. In order to examine the cooling effect by metal lines against heating, the H2 cooling is turned off, and metal-line cooling is calculated but not included in the thermal evolution. The metallicity for each case is set so that the cooling rate is marginally balanced with the heating rate at least once during the collapse phase. Namely, the critical metallicity is estimated as $Z_{\rm crit}\simeq 1.9\times 10^{-3}~{}Z_{\odot}$ (tree id 2) and $2.5\times 10^{-3}~{}Z_{\odot}$ (tree id 3), respectively. These values are higher than the critical metallicity of $Z_{\rm cirt}\sim 3\times 10^{-4}~{}Z_{\odot}$ (in the absence of dust) obtained by Omukai et al. (2008), where the effect of turbulence and merger heating is not included. Although the critical metallicity depends on the relative abundance of metals produced in SN ejecta, we use $Z_{\rm crit}=2\times 10^{-3}~{}Z_{\odot}$ as a reference value in the following discussion. ### 4.2 Efficiency of Metal Enrichment Throughout this paper, we do not consider the genetic pollution process through mergers of metal-rich minihalos, given that the star forming efficiency is strongly suppressed by intense LW radiation in the overdense region. However, we note that this treatment is justified only when the “actual” LW intensity is as high as the average value shown in Fig. 3. Otherwise, H2 cooling induces star formation in weak LW-radiation pockets. We do not quantify this effect that reduces the number of the main progenitors where gas is kept pristine. As a reference, Lupi et al. (2021) found $\sim 30\%$ of the atomic-cooling halos in the overdense region to be polluted genetically. Since some of those polluted halos do not belong to the merger history of the final massive quasar host halo, more than 70% of our main- progenitor samples should remain pristine (or sufficiently metal poor). On the other hand, together with the metal enrichment effect, we also exclude the contribution of LW flux from such lower mass halos, making our treatment conservative. Next, we discuss the modeling of environmental pollution led by SN-driven bubbles from nearby star-forming halos. One important caveat is that the progenitor halo is assumed to be immediately enriched once the bubble front reaches the halo virial radius. However, the instantaneous enrichment process considered in many previous studies in literature may not be realistic. In fact, metals in SN ejecta cannot penetrate into the halo center but pollute the halo superficially in the outer region with low densities of $\lesssim 10~{}{\rm cm}^{-3}$ (Chen et al., 2017; Chiaki et al., 2018), leaving the gas in the halo interior un-polluted, even for low mass halos. If more energetic pair-instability SNe occur in nearby source halos, the ejecta with stronger ram pressure deeply penetrate into the target halo and induce metal mixing at the shock front (Chen et al., 2017). To consider this uncertainty, we introduce the metal mixing efficiency $f_{\rm mix}$, which is the fraction of metals mixed with the interior gas in the target halo and is treated as a free parameter below. Another important quantity is the total amount of metals carried into the target halo through multiple SN-driven bubbles. Let us consider a source halo $m$ with a distance of $r_{\rm s}$ from the target halo with a size of $r_{\rm vir}(M_{\rm h})$. The mass of metals produced by multiple SNe in the source halo is given by $m_{\rm met}=N_{\rm sn}m_{\rm ej}$, where $N_{\rm sn}\simeq m_{\star}/m_{0}$ is the number of SNe and $m_{\rm ej}$ is the average mass of metals produced by one SN. We here adopt $m_{\rm ej}=0.746~{}M_{\odot}$, which corresponds to the metal ejecta mass produced by a $13~{}M_{\odot}$ stellar progenitor (Chiaki et al., 2018). Assuming that a fraction $f_{\rm esc,m}$ of the metals is launched isotropically by the SN bubble, the mass of the metals that reach the target halo is given by $f_{\rm esc,m}m_{\rm met}(r_{\rm vir}/r_{\rm s})^{2}/4$. Therefore, due to SN bubbles produced from one source halo, the gas metallicity in the target halo increases by $\displaystyle\Delta Z$ $\displaystyle\simeq\frac{m_{\star}m_{\rm ej}}{f_{b}M_{\rm h}m_{0}}\cdot\frac{f_{\rm esc,m}f_{\rm mix}}{4}\left(\frac{r_{\rm vir}}{r_{\rm s}}\right)^{2}$ (21) $\displaystyle\simeq 9.3\times 10^{-5}~{}Z_{\odot}~{}f_{\rm mix}\left(\frac{f_{\rm esc,m}}{0.5}\right)\left(\frac{m}{M_{\rm h}}\right)\left(\frac{5r_{\rm vir}}{r_{\rm s}}\right)^{2},$ where $f_{\rm esc,m}\simeq 0.5$ is motivated by a 3D high-resolution hydrodynamical simulations of SN-driven galactic outflows (Li et al., 2017). As discussed in § 3.1, the LW intensity peaks when the target halo reaches the atomic-cooling threshold because (1) source halos with $m_{\rm ac,z}$ are the most abundant population in number and (2) two halos with comparable mass are strongly clustered. This circumstance will also maximize the efficiency of environmental enrichment. Assuming $M_{\rm h}=m_{\rm ac,z}$, we estimate the number of source halos with mass of $m\geq m_{\rm ac,z}$ located within $r_{\rm s}$ ($\simeq 5r_{\rm vir}$ typically) from the target halo for the three representative trees as $N_{\rm s}\simeq$ 0.4 (tree id 1), 6 (tree id 2), and 86 (tree id 3), respectively. As a result, the gas metallicity in the target halo is calculated as $Z=N_{\rm s}\Delta Z\simeq 9.3\times 10^{-5}~{}Z_{\odot}f_{\rm mix}N_{\rm s}$. Therefore, we obtain the conditions where the environmental enrichment process affects the thermal evolution of gas in the target halo as $Z>Z_{\rm crit}$, or equivalently $\displaystyle N_{\rm s}>21.5f_{\rm mix}^{-1}.$ (22) Since $f_{\rm mix}\leq 1$, the gas evolution in the main progenitor surrounded by $\lesssim 21$ nearby source halos within $r_{\rm s}$ is unlikely to be affected by metal-line cooling. On the other hand, if the mixing efficiency is as high as $f_{\rm mix}\gtrsim 0.25$, metal enrichment will play an important role in changing the gas evolution in rapidly growing halos (tree id 3), reducing the number fraction of H-H collapse cases (see Fig. 5). Additionally, inhomogeneous density distributions inside the source halos and non-steady SFR that form SNe in the earlier stage change the velocity and shape of expanding bubbles. Those effects could result in either an overestimation or underestimation of the bubble size. To discuss the efficiency of environmental enrichment precisely, we need to further study a variety of situations with different physical parameters as well as the metal mixing efficiency $f_{\rm mix}$. We leave this to future work. ### 4.3 Dynamical evolution of metal enriched gas We quantify the critical metallicity and discuss the impact of metal-line cooling on the thermal evolution of gas clouds. However, dynamical evolution of a collapsing cloud with $Z\gtrsim Z_{\rm crit}$ that composes of a warm outer-envelope ($T\simeq 8000~{}{\rm K}$) and a cool central core has not been fully understood; especially, longterm behavior of the mass inflow rate onto the central newly-formed protostar is still uncertain. Recent cosmological simulations suggest that rapid mass inflows may occur even with metal pollution above the critical metallicity in atomic-cooling halos (Regan et al., 2020a), but widespread star formation limits the final mass of the central star to $\lesssim 10^{4}~{}M_{\odot}$ (Regan et al., 2020b). On the other hand, when the metallicity is lower than the critical value, the collapsing gas cloud fragments only at the central region and forms a compact disk, in which a vast majority of the clumps merge with the central protostar via inward migration (Inayoshi & Haiman, 2014; Chon & Omukai, 2020). As a result, the stellar growth is not quenched by metal pollution. Future work is needed to investigate the star formation in the overdense regions where high-$z$ quasar form to quantify the impact of metal pollution on the gas dynamics. --- Figure 8: Distributions of the mass of massive stars (equivalent to seed BHs) formed in the quasar progenitors with a bin size of $\Delta\log M_{\star}=0.1$ for the two cases with $v_{\rm bsm}=0$ (left panels) and $v_{\rm bsm}=1\sigma$ (right panels). We set the accretion efficiency of $\eta=0.6$ (and $0.3$) in the upper (lower) panels. Without streaming motion, the BH mass is widely distributed from $500~{}M_{\odot}(250~{}M_{\odot})$ to $\gtrsim 2\times 10^{5}~{}M_{\odot}$ for $\eta=0.6$ ($0.3$, respectively), while for $v_{\rm bsm}=1\sigma$, the lower bound shifts to $7000~{}M_{\odot}(3500~{}M_{\odot})$. ## 5 Discussion ### 5.1 Protostellar Mass and BH Mass Distribution We apply the obtained mass accretion rate to estimate the final protostellar mass distribution at the end of star formation episodes. Due to the existing angular momentum at large scales, the rapidly accreted pristine gas settles into a disk, which becomes gravitationally unstable and thus results in fragmentation and clump formation (e.g., Oh & Haiman, 2002). Most clumps can migrate inward and merge with the central protostar before forming stars (Inayoshi & Haiman, 2014), yielding accretion rate onto the stellar surface through the disk $\dot{M}_{\star}=\eta\dot{M}$, where $\eta(<1)$ denotes the conversion efficiency from the global accretion rate to that through the accretion disk. Hydrodynamical simulations find that mass accretion through the unstable disk proceeds episodically and the time-averaged value of the efficiency is $\eta\simeq 0.6$ (Sakurai et al. 2016; Toyouchi et al. in prep). When the time-averaged accretion rate is higher than a critical rate, $\dot{M}_{\star}\gtrsim\dot{M}_{\rm crit}$, the accreting star continues to expand its envelope with a lower surface temperature of $T_{\rm eff}\simeq 5000~{}{\rm K}$, from which UV radiation hardly emits. As a result, stellar radiative feedback does not prevent the central star growing via mass accretion (Omukai & Palla, 2001; Hosokawa et al., 2012, 2013; Schleicher et al., 2013; Haemmerlé et al., 2018; Sakurai et al., 2015, 2020b). Since the value of $\dot{M}_{\rm crit}$ ranges from $0.01$ to $0.04~{}M_{\odot}~{}{\rm yr}^{-1}$, depending on the treatment of the stellar evolution calculations and their boundary conditions (Hosokawa et al., 2013; Haemmerlé et al., 2018), we adopt the highest $\dot{M}_{\rm crit}=0.04~{}M_{\odot}~{}{\rm yr}^{-1}$ as a reference value. This choice leads to a lower bound of the stellar/BH mass. With $\dot{M}_{\star}\gtrsim\dot{M}_{\rm crit}$, the final stellar mass is determined either by its finite lifetime or by the onset of stellar collapse triggered by the general-relativistic (GR) instability (Shibata et al. 2016; see a review by Woods et al. 2019 and references therein). The final mass is also affected by fuel supply through mass accretion onto the star. Woods et al. (2017) have investigated the final mass of stars accreting at a constant rate over $\simeq 0.01-10~{}M_{\odot}~{}{\rm yr}^{-1}$ (radiative feedback is neglected), and found that the final mass linearly increases with the accretion rate below $\sim 0.03~{}M_{\odot}~{}{\rm yr}^{-1}$ but is saturated around $\sim{\rm a~{}few}\times 10^{5}~{}M_{\odot}$ due to the GR instability. The relation between the critical mass and accretion rate is fitted as $M_{\rm\star,GR}\simeq\left[0.83\log\left(\frac{\dot{M}_{\star}}{~{}M_{\odot}~{}{\rm yr}^{-1}}\right)+2.48\right]\times 10^{5}~{}M_{\odot},$ (23) at $\dot{M}_{\star}\geq 0.1~{}M_{\odot}~{}{\rm yr}^{-1}$, which is used for our analysis. On the other hand, when the stellar accretion rate is lower than the critical rate, $\dot{M}_{\star}\lesssim\dot{M}_{\rm crit}$, the star evolves to the main-sequence stage and begins to emit strong ionizing radiation, quenching the stellar growth. Here, we simply consider that ionizing radiation from the star heats the disk surface and thus photoevaporation suppresses the accretion rate (McKee & Tan, 2008; Hosokawa et al., 2011; Tanaka et al., 2013). This process becomes important when the ionization front reaches the stellar gravitational influence radius for ionized gas with a temperature of $2\times 10^{4}~{}{\rm K}$ defined by $R_{\rm inf,\star}\equiv\frac{GM_{\star}}{c_{\rm s,ion}^{2}}\simeq 0.17~{}{\rm pc}\left(\frac{M_{\star}}{10^{4}~{}M_{\odot}}\right),$ (24) and the ionized gas breaks out through the neutral infalling gas. The photoevaporation rate can be expressed as $\dot{M}_{\rm pe}\simeq 2.1\times 10^{-2}~{}M_{\odot}~{}{\rm yr}^{-1}\left(\frac{\Phi_{\rm ion}}{10^{52}~{}{\rm s}^{-1}}\right)^{1/2}\left(\frac{R_{\rm disk}}{0.1~{}{\rm pc}}\right)^{1/2},$ (25) where $\Phi_{\rm ion}$ is the ionizing photon number flux and $R_{\rm disk}$ is the size of the accretion disk. The photon flux is approximated as $\Phi_{\rm ion}\simeq 1.6\times 10^{52}~{}{\rm s}^{-1}(M_{\star}/10^{4}~{}M_{\odot})$ in the range of $10^{3}\lesssim M_{\star}/M_{\odot}\lesssim 10^{5}$ for main-sequence stars (Johnson et al., 2012). We evaluate the mass outflow rate owing to photoevaporation by setting $R_{\rm disk}\simeq R_{\rm inf,\star}$ as $\dot{M}_{\rm pe,min}\simeq 3.5\times 10^{-2}~{}M_{\odot}~{}{\rm yr}^{-1}\left(\frac{M_{\star}}{10^{4}~{}M_{\odot}}\right),$ (26) which gives a lower bound for the rate because the outflow of ionized gas is mainly driven from larger radii (i.e., a lager surface area). Therefore, equating $\dot{M}_{\star}=\dot{M}_{\rm pe,min}$, we obtain the feedback- regulated stellar mass as $M_{\rm\star,fb}\simeq\dot{M}_{\star}t_{\rm pe}$ or $\dot{M}_{\rm\star,fb}\simeq 2.9\times 10^{3}~{}M_{\odot}\left(\frac{\dot{M}_{\star}}{0.01~{}M_{\odot}~{}{\rm yr}^{-1}}\right),$ (27) at $\dot{M}_{\star}\leq\dot{M}_{\rm crit}$, where $t_{\rm pe}(\simeq 2.9\times 10^{5}~{}{\rm yr})$ is the characteristic photoevaporation timescale (note that this expression is valid when the stellar lifetime is longer than $t_{\rm pe}$). The final mass at the intermediate accretion rate ($\dot{M}_{\rm crit}\leq\dot{M}_{\star}\leq 0.1~{}M_{\odot}~{}{\rm yr}^{-1}$) is estimated by performing logarithmic interpolation. In Fig. 8, we show the mass distribution of massive BH seeds formed in the high-$z$ quasar progenitor halos, calculated with the method described above (see also the bottom panels in Fig. 6). Note that the number fraction from the different types of gas evolution is stacked at the same mass bin. Without the streaming motion ($v_{\rm bsm}=0$; left panels), the BH mass is widely distributed from $500~{}M_{\odot}$ ($250~{}M_{\odot}$) to $\gtrsim 2\times 10^{5}~{}M_{\odot}$ for $\eta=0.6$ ($0.3$, respectively) with a few peaks corresponding to the virial temperatures of halos when the BHs form by gas collapse. Overall, the cases with high accretion rates $\dot{M}_{\rm in}$ (H-H2 and H-H cases) are responsible for high-mass BH formation beyond $\sim 10^{4}~{}M_{\odot}$, while the H2 case with lower accretion rates yields less massive BHs with $<10^{4}~{}M_{\odot}$. The number of BH seeds above $2\times 10^{5}~{}M_{\odot}$ is limited because the GR instability induces direct collapse of accreting supermassive stars. The shape of the mass distribution at $10^{4}\lesssim M_{\bullet}/M_{\odot}\lesssim 10^{5}$ depend on the accretion efficiency; namely, the smaller value of $\eta(=0.3)$ yields a distribution skewed toward lower masses. With non-zero streaming motion ($v_{\rm bsm}=1\sigma$; right panels), the less massive population with $<10^{4}~{}M_{\odot}$ decreases abruptly since nearly all the cases experience the atomic-cooling stage and thus the central stars accrete at high rates without strong radiative feedback. We note that the BH mass distribution for higher streaming velocities are similar to that for $v_{\rm bsm}=1\sigma$, but their contribution to the total BH mass distribution is less important because regions with $v_{\rm bsm}\geq 2\sigma$ are rarer. As discussed in §4, the number fraction of the cases with highest mass accretion rates (H-H cases) would be reduced by the effect of line cooling via atomic carbon and oxygen which are produced in nearby source halos through SNe and carried into the quasar main progenitor halos with interest. The level of reduction depends on the metal mixing efficiency in the main progenitor; namely, the enrichment effect could be neglected if the mixing efficiency is lower than $\sim 20\%$. Nevertheless, the overall shape of the BH mass distribution still holds. ### 5.2 Subsequent BH growth and evolution How do those massive seed BHs formed in overdense regions grow to be SMBHs that are observed as high-$z$ quasars at $z\simeq 6-7$? In previous studies in literature, the subsequent growth of their BHs via gas accretion and/or mergers and the required conditions have been discussed (e.g., Tanaka & Haiman, 2009; Valiante et al., 2016). Recently, large-scale cosmological simulations have been exploring the evolution of SMBHs and the coevolution of their host galaxies including various feedback processes due to SNe and AGN activity with subgrid models. These simulations have generally found that massive seed BHs formed in protogalaxies hardly grow via gas accretion because dense, cold gas is expelled by energetic SN feedback associated with star formation. However, it is worth noting that most simulations in which SN feedback quenches BH growth have focused on “typical” atomic-cooling halos that will grow to $\sim 10^{10-11}~{}M_{\odot}$ by $z\simeq 6$ (e.g., Habouzit et al., 2017; Latif et al., 2018; Smith et al., 2018) On the other hand, as pointed out by Inayoshi et al. (2020), the progenitor halos of high-$z$ quasar hosts with $M_{\rm h}\simeq 10^{12}~{}M_{\odot}$ at $z\simeq 6$ form in rarer regions and have reached $M_{\rm h}\sim 10^{8}~{}M_{\odot}$ with deeper gravitational potential by the time when star formation takes place ($z\sim 20-35$). In such massive halos, a large amount of cold gas is supplied to the nuclear region through filamentary structures of the proto-cosmic web (Di Matteo et al., 2012), and the seed BHs can be fed at high rates significantly exceeding the Eddington limit when the metallicity of inflowing gas is as low as $\lesssim 0.01~{}Z_{\odot}$ (Toyouchi et al. 2021; see also Inayoshi et al. 2016). The critical halo mass required for the onset of rapid mass accretion exceeding the Eddington rate is $M_{\rm h}\simeq 10^{9}~{}M_{\odot}$, almost independent of redshift. Most of the quasar progenitor halos of interest can reach this mass threshold after birth of seed BHs in $\simeq 20-50$ Myr, within which intense star bursts would take place and form protogalaxies. Exploring the nature of BH growth embedded in such a protogalaxy is left for future investigations. This process is a possible way to form intermediate massive BH (IMBH) populations. Observations of IMBHs in the local universe have the potential to constrain high-$z$ BH (seed) formation (see the review by Greene et al. 2020). Furthermore, if those IMBHs form binaries through galaxy mergers and dynamical processes during the cosmic history, the seed forming channel also provides a significant number of gravitational wave events (e.g., Hartwig et al., 2018; Chon et al., 2018; Regan et al., 2020b), which will be detectable by the space-based gravitational wave detectors such as the Laser Interferometer Space Antenna (LISA) (Amaro-Seoane et al., 2017) and Tianqin (Luo et al., 2016), and third-generation terrestrial instruments. ## 6 Summary In this paper, we investigate a new scenario of the formation of heavy BH seeds through collapse of warm gas in massive halos that end up in quasar hosts at $z\simeq 6-7$. In the highly biased, overdense regions of the universe, stronger halo clustering increases the frequency of halo mergers and boosts the mean intensity of LW radiation background produced from star- forming galaxies. Those effects are expected to increase the probability of massive seed formation because the conditions required for their formation (intense LW irradiation and violent merger heating) become less stringent than previous considered. Under such unique environments, we model the thermal and dynamical evolution of massive gas clouds along with $10^{4}$ merger trees of the main progenitors of high-$z$ quasar hosts using the Monte Carlo method. With those samples, we study the statistical properties of the progenitor halos of high-z quasar hosts and massive seed BHs. Our major findings can be summarized as follows. 1. 1. In the high-$z$ quasar forming regions, DM halos are likely irradiated by strong LW radiation with intensity of $J_{\rm LW}\simeq 100-10^{3}$ (in units of $10^{-21}~{}{\rm erg}~{}{\rm s}^{-1}~{}{\rm cm}^{-2}~{}{\rm Hz}^{-1}$) from nearby star-forming galaxies at $z\lesssim 30$ and gas clouds in the halo interiors are heated by successive gaseous halo mergers. Suppression of H2 cooling via LW irradiation/merger hating as well as injection of gas kinetic energy through halo mergers prevent gas collapse and delays prior star formation episodes. 2. 2. Without baryonic streaming motion, 74% of the trees experience gas collapse led by H2 cooling, while the rest (26%) form atomically-cooling gas clouds that begin to collapse isothermally with $T\simeq 8000~{}{\rm K}$ via Ly$\alpha$ cooling. With a streaming velocity higher than the root-mean-square value, gas clouds for nearly all $10^{4}$ realizations of the merger trees enter the atomic-cooling stage. 3. 3. The fraction of trees which host isothermal gas collapse is $14\%$ and increases with streaming velocity, while the rest form H2-cooled cores after short isothermal phases. However, this fraction is reduced by additional cooling via metal fine-structure lines when the collapsing gas could be enriched to $Z_{\rm crit}\sim 2\times 10^{-3}~{}Z_{\odot}$, requiring efficient metal mixing $f_{\rm mix}\gtrsim 0.25$. This high probability reflects that high-redshift quasar forming regions likely provide such peculiar environments, which hardly occur in typical high-redshift star- forming regions. 4. 4. The mass accretion rate onto a newly-born protostar is distributed over $3\times 10^{-3}-5~{}M_{\odot}~{}{\rm yr}^{-1}$, a large fraction of which exceeds the critical rate suppressing stellar radiative feedback. As a result, we expect a distribution of stellar masses (presumably BH masses) ranging from several hundred to above $10^{5}~{}M_{\odot}$. We greatly thank Gen Chiaki, Zoltán Haiman, Tilman Hartwig, Alessandro Lupi, and Daisuke Toyouchi for constructive discussions. This work is supported by the National Natural Science Foundation of China (12073003, 12003003, 11721303, 11991052, 11950410493), the National Key R&D Program of China (2016YFA0400702), and the High-Performance Computing Platform of Peking University. Y.Q acknowledges support from the China Postdoctoral Science Foundation (2020T130019). ## Appendix A The critical conditions for collapse of an isothermal gas cloud In the Appendix, we briefly describe the method of how to calculate the critical gas density at the center by solving the hydrostatic equation for an isothermal gas cloud (Eq. 15), where the gas pressure gradient force is balanced with the gas self-gravity and DM gravitational force. For demonstration purpose, in the left panel of Fig. 9, we show the radial profiles of gas with an effective sound speed of $c_{\rm eff}=8.3\mathrm{~{}km~{}s^{-1}}$ (corresponding to $T=10^{4}~{}{\rm K}$ gas in the absence of turbulence) for different values of $\rho_{0}$ in a DM halo with $M_{\mathrm{h}}=6\times 10^{6}~{}M_{\odot}$ at $z=30$. As the central density increases, the density at the virial radius $\rho_{\rm gas}(R_{\rm vir})$ does not increase monotonically but has a local maximum value around $\rho_{0}\simeq 10^{-21}~{}\rm g~{}{cm}^{-3}$. In general, the maximum value of $\rho_{\rm gas}(R_{\rm vir})$ can be found for a given combination of $M_{\mathrm{h}}$, $z$, and $c_{\rm eff}$. In the right panel of Fig. 9, we present the relation between $\rho_{\rm gas}(R_{\rm vir})$ and $\rho_{0}$ for different halo masses ($z=30$ and $c_{\rm eff}=8.3\mathrm{~{}km~{}s^{-1}}$ are fixed). As seen in the left panel, each curve has a local maximum and the maximum value decreases with $M_{\rm h}$. The density value at the outer boundary ($\rho_{\rm ext}=f_{\rm b}\rho_{\rm DM}$) weakly depends on $M_{\rm h}$ and $z$ through the concentration factor $c_{\rm vir}$, i.e., the three halos have $\rho_{\rm ext}\simeq 8\times 10^{-25}~{}{\rm cm}^{-3}$, varying within $3\%$. For $M_{\rm h}=6\times 10^{6}~{}M_{\odot}$, there exist two solutions where the boundary conditions are satisfied. Since the solution with the higher value of $\rho_{0}$ is not stable, we adopt the solution with the lower value of $\rho_{0}$ (see Ebert, 1955; Bonnor, 1958; Lynden-Bell & Wood, 1968). As the halo mass increases to $M_{\rm h}=8\times 10^{6}$ and $10^{7}~{}M_{\odot}$ , there is no hydrostatic solution of the gas cloud. In our semi-analytical model, we calculate the hydro-static density profile which satisfies the boundary conditions at each time step and quantify the critical halo mass $M_{\rm h,crit}$ above which the gas begins to collapse. We note that this method can be applied to a wide range of $c_{\rm eff}$ and $z$ of interest in our paper. --- Figure 9: Left panel: gas density profile in a halo with $M_{\mathrm{h}}=6\times 10^{6}~{}M_{\odot}$ at $z=30$, $c_{\rm eff}=8.3\mathrm{~{}km~{}s^{-1}}$, calculated from $\rho_{0}=10^{-23,-22,-21,-20}\rm g~{}{cm}^{-3}$. With increasing $\rho_{0}$, the $\rho_{\rm gas}(R_{\rm vir})$ solved first increases then decreases. Right panel: the $\rho_{\rm gas}(R_{\rm vir})$ solved as a function of $\rho_{0}$ in diffrent halo masses. The solution of $\rho_{0}$ is determined from the left intersection of $\rho_{\rm ext}$ and $\rho_{\rm gas}(R_{\rm vir})$ curves. The local maximum of $\rho_{\rm gas}(R_{\rm vir})$ decreases with increasing halo mass, thus a critical $M_{\rm h,crit}$ exists above which no solution of $\rho_{0}$ can be found. In this case, $M_{\rm h,crit}$ lies between $6\times 10^{6}$ and $8\times 10^{6}~{}M_{\odot}$. ## References * Abel et al. (1997) Abel, T., Anninos, P., Zhang, Y., & Norman, M. L. 1997, New A, 2, 181, doi: 10.1016/S1384-1076(97)00010-9 * Abel et al. (2002) Abel, T., Bryan, G. L., & Norman, M. L. 2002, Science, 295, 93, doi: 10.1126/science.295.5552.93 * Agarwal et al. (2012) Agarwal, B., Khochfar, S., Johnson, J. L., et al. 2012, MNRAS, 425, 2854, doi: 10.1111/j.1365-2966.2012.21651.x * Amaro-Seoane et al. (2017) Amaro-Seoane, P., Audley, H., Babak, S., et al. 2017, arXiv e-prints, arXiv:1702.00786. https://arxiv.org/abs/1702.00786 * Barlow (1984) Barlow, S. E. 1984, PhD thesis, UNIVERSITY OF COLORADO AT BOULDER. * Becerra et al. (2015) Becerra, F., Greif, T. H., Springel, V., & Hernquist, L. E. 2015, MNRAS, 446, 2380, doi: 10.1093/mnras/stu2284 * Bonetti et al. (2019) Bonetti, M., Sesana, A., Haardt, F., Barausse, E., & Colpi, M. 2019, MNRAS, 486, 4044, doi: 10.1093/mnras/stz903 * Bonnor (1958) Bonnor, W. B. 1958, MNRAS, 118, 523, doi: 10.1093/mnras/118.5.523 * Bromm & Loeb (2003) Bromm, V., & Loeb, A. 2003, ApJ, 596, 34, doi: 10.1086/377529 * Bromm & Yoshida (2011) Bromm, V., & Yoshida, N. 2011, ARA&A, 49, 373, doi: 10.1146/annurev-astro-081710-102608 * Bullock et al. (2001) Bullock, J. S., Kolatt, T. S., Sigad, Y., et al. 2001, MNRAS, 321, 559, doi: 10.1046/j.1365-8711.2001.04068.x * Chandrasekhar (1951a) Chandrasekhar, S. 1951a, Proceedings of the Royal Society of London Series A, 210, 18, doi: 10.1098/rspa.1951.0227 * Chandrasekhar (1951b) —. 1951b, Proceedings of the Royal Society of London Series A, 210, 26, doi: 10.1098/rspa.1951.0228 * Chen et al. (2017) Chen, K.-J., Whalen, D. J., Wollenberg, K. M. J., Glover, S. C. O., & Klessen, R. S. 2017, ApJ, 844, 111, doi: 10.3847/1538-4357/aa7b34 * Chiaki et al. (2018) Chiaki, G., Susa, H., & Hirano, S. 2018, MNRAS, 475, 4378, doi: 10.1093/mnras/sty040 * Chon et al. (2016) Chon, S., Hirano, S., Hosokawa, T., & Yoshida, N. 2016, ApJ, 832, 134, doi: 10.3847/0004-637X/832/2/134 * Chon et al. (2018) Chon, S., Hosokawa, T., & Yoshida, N. 2018, MNRAS, 475, 4104, doi: 10.1093/mnras/sty086 * Chon & Omukai (2020) Chon, S., & Omukai, K. 2020, MNRAS, 494, 2851, doi: 10.1093/mnras/staa863 * Cole et al. (2000) Cole, S., Lacey, C. G., Baugh, C. M., & Frenk, C. S. 2000, MNRAS, 319, 168, doi: 10.1046/j.1365-8711.2000.03879.x * Coppola et al. (2011) Coppola, C. M., Longo, S., Capitelli, M., Palla, F., & Galli, D. 2011, ApJS, 193, 7, doi: 10.1088/0067-0049/193/1/7 * Croft et al. (1999) Croft, H., Dickinson, A. S., & Gadea, F. X. 1999, MNRAS, 304, 327, doi: 10.1046/j.1365-8711.1999.02346.x * Dalgarno & Lepp (1987) Dalgarno, A., & Lepp, S. 1987, in Astrochemistry, ed. M. S. Vardya & S. P. Tarafdar, Vol. 120, 109–118 * Dayal et al. (2019) Dayal, P., Rossi, E. M., Shiralilou, B., et al. 2019, MNRAS, 486, 2336, doi: 10.1093/mnras/stz897 * Di Matteo et al. (2012) Di Matteo, T., Khandai, N., DeGraf, C., et al. 2012, ApJ, 745, L29, doi: 10.1088/2041-8205/745/2/L29 * Dijkstra et al. (2014) Dijkstra, M., Ferrara, A., & Mesinger, A. 2014, MNRAS, 442, 2036, doi: 10.1093/mnras/stu1007 * Dijkstra et al. (2008) Dijkstra, M., Haiman, Z., Mesinger, A., & Wyithe, J. S. B. 2008, MNRAS, 391, 1961, doi: 10.1111/j.1365-2966.2008.14031.x * Dove et al. (1987) Dove, J. E., Rusk, A. C. M., Cribb, P. H., & Martin, P. G. 1987, ApJ, 318, 379, doi: 10.1086/165375 * Ebert (1955) Ebert, R. 1955, ZAp, 37, 217 * Fan et al. (2006) Fan, X., Strauss, M. A., Richards, G. T., et al. 2006, AJ, 131, 1203, doi: 10.1086/500296 * Ferland et al. (1992) Ferland, G. J., Peterson, B. M., Horne, K., Welsh, W. F., & Nahar, S. N. 1992, ApJ, 387, 95, doi: 10.1086/171063 * Fernandez et al. (2014) Fernandez, R., Bryan, G. L., Haiman, Z., & Li, M. 2014, MNRAS, 439, 3798, doi: 10.1093/mnras/stu230 * Glover (2015a) Glover, S. C. O. 2015a, MNRAS, 451, 2082, doi: 10.1093/mnras/stv1059 * Glover (2015b) —. 2015b, MNRAS, 453, 2901, doi: 10.1093/mnras/stv1781 * Glover & Abel (2008) Glover, S. C. O., & Abel, T. 2008, MNRAS, 388, 1627, doi: 10.1111/j.1365-2966.2008.13224.x * Glover & Jappsen (2007) Glover, S. C. O., & Jappsen, A. K. 2007, ApJ, 666, 1, doi: 10.1086/519445 * Greene et al. (2020) Greene, J. E., Strader, J., & Ho, L. C. 2020, ARA&A, 58, 257, doi: 10.1146/annurev-astro-032620-021835 * Habouzit et al. (2017) Habouzit, M., Volonteri, M., & Dubois, Y. 2017, MNRAS, 468, 3935, doi: 10.1093/mnras/stx666 * Haemmerlé et al. (2018) Haemmerlé, L., Woods, T. E., Klessen, R. S., Heger, A., & Whalen, D. J. 2018, MNRAS, 474, 2757, doi: 10.1093/mnras/stx2919 * Haiman et al. (1997) Haiman, Z., Rees, M. J., & Loeb, A. 1997, ApJ, 476, 458, doi: 10.1086/303647 * Haiman et al. (1996) Haiman, Z., Thoul, A. A., & Loeb, A. 1996, ApJ, 464, 523, doi: 10.1086/177343 * Hartwig et al. (2018) Hartwig, T., Agarwal, B., & Regan, J. A. 2018, MNRAS, 479, L23, doi: 10.1093/mnrasl/sly091 * Hirano et al. (2017) Hirano, S., Hosokawa, T., Yoshida, N., & Kuiper, R. 2017, Science, 357, 1375, doi: 10.1126/science.aai9119 * Hirano et al. (2018) Hirano, S., Yoshida, N., Sakurai, Y., & Fujii, M. S. 2018, ApJ, 855, 17, doi: 10.3847/1538-4357/aaaaba * Hosokawa et al. (2012) Hosokawa, T., Omukai, K., & Yorke, H. W. 2012, ApJ, 756, 93, doi: 10.1088/0004-637X/756/1/93 * Hosokawa et al. (2011) Hosokawa, T., Omukai, K., Yoshida, N., & Yorke, H. W. 2011, Science, 334, 1250, doi: 10.1126/science.1207433 * Hosokawa et al. (2013) Hosokawa, T., Yorke, H. W., Inayoshi, K., Omukai, K., & Yoshida, N. 2013, ApJ, 778, 178, doi: 10.1088/0004-637X/778/2/178 * Hummer & Storey (1998) Hummer, D. G., & Storey, P. J. 1998, MNRAS, 297, 1073, doi: 10.1046/j.1365-8711.1998.2970041073.x * Huq et al. (1982) Huq, M. S., Doverspike, L. D., Champion, R. L., & Esaulov, V. A. 1982, Journal of Physics B Atomic Molecular Physics, 15, 951, doi: 10.1088/0022-3700/15/6/020 * Iliev et al. (2003) Iliev, I. T., Scannapieco, E., Martel, H., & Shapiro, P. R. 2003, MNRAS, 341, 81, doi: 10.1046/j.1365-8711.2003.06410.x * Inayoshi & Haiman (2014) Inayoshi, K., & Haiman, Z. 2014, MNRAS, 445, 1549, doi: 10.1093/mnras/stu1870 * Inayoshi et al. (2016) Inayoshi, K., Haiman, Z., & Ostriker, J. P. 2016, MNRAS, 459, 3738, doi: 10.1093/mnras/stw836 * Inayoshi et al. (2018) Inayoshi, K., Li, M., & Haiman, Z. 2018, MNRAS, 479, 4017, doi: 10.1093/mnras/sty1720 * Inayoshi & Omukai (2012) Inayoshi, K., & Omukai, K. 2012, MNRAS, 422, 2539, doi: 10.1111/j.1365-2966.2012.20812.x * Inayoshi et al. (2014) Inayoshi, K., Omukai, K., & Tasker, E. 2014, MNRAS, 445, L109, doi: 10.1093/mnrasl/slu151 * Inayoshi et al. (2020) Inayoshi, K., Visbal, E., & Haiman, Z. 2020, ARA&A, 58, 27, doi: 10.1146/annurev-astro-120419-014455 * Inoue (2011) Inoue, A. K. 2011, MNRAS, 415, 2920, doi: 10.1111/j.1365-2966.2011.18906.x * Jacobs et al. (1967) Jacobs, T. A., Giedt, R. R., & Cohen, N. 1967, J. Chem. Phys., 47, 54, doi: 10.1063/1.1711890 * Janev et al. (1987) Janev, R. K., Langer, W. D., Post, D. E., & Evans, K. 1987, in Elementary Processes in Hydrogen-Helium Plasmas: Cross Sections and Reaction Rate Coefficients (Berlin, Heidelberg: Springer Berlin Heidelberg), 217–231, doi: 10.1007/978-3-642-71935-6_7 * Jiang et al. (2016) Jiang, L., McGreer, I. D., Fan, X., et al. 2016, ApJ, 833, 222, doi: 10.3847/1538-4357/833/2/222 * Johnson et al. (2013) Johnson, J. L., Dalla Vecchia, C., & Khochfar, S. 2013, MNRAS, 428, 1857, doi: 10.1093/mnras/sts011 * Johnson et al. (2012) Johnson, J. L., Whalen, D. J., Fryer, C. L., & Li, H. 2012, ApJ, 750, 66, doi: 10.1088/0004-637X/750/1/66 * Karpas et al. (1979) Karpas, Z., Anicich, V., & Huntress, W. T. 1979, J. Chem. Phys., 70, 2877, doi: 10.1063/1.437823 * Kimura et al. (1993) Kimura, M., Lane, N. F., Dalgarno, A., & Dixson, R. G. 1993, ApJ, 405, 801, doi: 10.1086/172410 * Kormendy & Ho (2013) Kormendy, J., & Ho, L. C. 2013, ARA&A, 51, 511, doi: 10.1146/annurev-astro-082708-101811 * Kreckel et al. (2010) Kreckel, H., Bruhns, H., Čížek, M., et al. 2010, Science, 329, 69, doi: 10.1126/science.1187191 * Lacey & Cole (1993) Lacey, C., & Cole, S. 1993, MNRAS, 262, 627, doi: 10.1093/mnras/262.3.627 * Larson (1969) Larson, R. B. 1969, MNRAS, 145, 271, doi: 10.1093/mnras/145.3.271 * Latif et al. (2013) Latif, M. A., Schleicher, D. R. G., Schmidt, W., & Niemeyer, J. 2013, MNRAS, 433, 1607, doi: 10.1093/mnras/stt834 * Latif & Volonteri (2015) Latif, M. A., & Volonteri, M. 2015, MNRAS, 452, 1026, doi: 10.1093/mnras/stv1337 * Latif et al. (2018) Latif, M. A., Volonteri, M., & Wise, J. H. 2018, MNRAS, 476, 5016, doi: 10.1093/mnras/sty622 * Lepp & Shull (1983) Lepp, S., & Shull, J. M. 1983, ApJ, 270, 578, doi: 10.1086/161149 * Li et al. (2017) Li, M., Bryan, G. L., & Ostriker, J. P. 2017, ApJ, 841, 101, doi: 10.3847/1538-4357/aa7263 * Luo et al. (2016) Luo, J., Chen, L.-S., Duan, H.-Z., et al. 2016, Classical and Quantum Gravity, 33, 035010, doi: 10.1088/0264-9381/33/3/035010 * Lupi et al. (2021) Lupi, A., Haiman, Z., & Volonteri, M. 2021, MNRAS, doi: 10.1093/mnras/stab692 * Lynden-Bell & Wood (1968) Lynden-Bell, D., & Wood, R. 1968, MNRAS, 138, 495, doi: 10.1093/mnras/138.4.495 * Mac Low & Shull (1986) Mac Low, M. M., & Shull, J. M. 1986, ApJ, 302, 585, doi: 10.1086/164017 * Martin et al. (1998) Martin, P. G., Keogh, W. J., & Mandy, M. E. 1998, ApJ, 499, 793, doi: 10.1086/305665 * Matsuoka et al. (2018) Matsuoka, Y., Iwasawa, K., Onoue, M., et al. 2018, ApJS, 237, 5, doi: 10.3847/1538-4365/aac724 * McKee & Tan (2002) McKee, C. F., & Tan, J. C. 2002, Nature, 416, 59, doi: 10.1038/416059a * McKee & Tan (2008) —. 2008, ApJ, 681, 771, doi: 10.1086/587434 * McLaughlin et al. (2017) McLaughlin, B. M., Stancil, P. C., Sadeghpour, H. R., & Forrey, R. C. 2017, Journal of Physics B Atomic Molecular Physics, 50, 114001, doi: 10.1088/1361-6455/aa6c1f * Mo & White (2002) Mo, H. J., & White, S. D. M. 2002, MNRAS, 336, 112, doi: 10.1046/j.1365-8711.2002.05723.x * Mortlock et al. (2011) Mortlock, D. J., Warren, S. J., Venemans, B. P., et al. 2011, Nature, 474, 616, doi: 10.1038/nature10159 * Navarro et al. (1997) Navarro, J. F., Frenk, C. S., & White, S. D. M. 1997, ApJ, 490, 493, doi: 10.1086/304888 * Oh & Haiman (2002) Oh, S. P., & Haiman, Z. 2002, ApJ, 569, 558, doi: 10.1086/339393 * Omukai (2001) Omukai, K. 2001, ApJ, 546, 635, doi: 10.1086/318296 * Omukai & Palla (2001) Omukai, K., & Palla, F. 2001, ApJ, 561, L55, doi: 10.1086/324410 * Omukai et al. (2008) Omukai, K., Schneider, R., & Haiman, Z. 2008, ApJ, 686, 801, doi: 10.1086/591636 * Onoue et al. (2019) Onoue, M., Kashikawa, N., Matsuoka, Y., et al. 2019, ApJ, 880, 77, doi: 10.3847/1538-4357/ab29e9 * Orel (1987) Orel, A. E. 1987, J. Chem. Phys., 87, 314, doi: 10.1063/1.453628 * Parkinson et al. (2008) Parkinson, H., Cole, S., & Helly, J. 2008, MNRAS, 383, 557, doi: 10.1111/j.1365-2966.2007.12517.x * Peart & Hayton (1994) Peart, B., & Hayton, D. A. 1994, Journal of Physics B Atomic Molecular Physics, 27, 2551, doi: 10.1088/0953-4075/27/12/013 * Penston (1969) Penston, M. V. 1969, MNRAS, 144, 425, doi: 10.1093/mnras/144.4.425 * Planck Collaboration et al. (2016) Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2016, A&A, 594, A13, doi: 10.1051/0004-6361/201525830 * Pollack et al. (1994) Pollack, J. B., Hollenbach, D., Beckwith, S., et al. 1994, ApJ, 421, 615, doi: 10.1086/173677 * Poulaert et al. (1978) Poulaert, G., Brouillard, F., Claeys, W., McGowan, J. W., & Van Wassenhove, G. 1978, Journal of Physics B Atomic Molecular Physics, 11, L671, doi: 10.1088/0022-3700/11/21/006 * Press & Schechter (1974) Press, W. H., & Schechter, P. 1974, ApJ, 187, 425, doi: 10.1086/152650 * Regan et al. (2020a) Regan, J. A., Haiman, Z., Wise, J. H., O’Shea, B. W., & Norman, M. L. 2020a, The Open Journal of Astrophysics, 3, E9, doi: 10.21105/astro.2006.14625 * Regan et al. (2014) Regan, J. A., Johansson, P. H., & Wise, J. H. 2014, ApJ, 795, 137, doi: 10.1088/0004-637X/795/2/137 * Regan et al. (2020b) Regan, J. A., Wise, J. H., Woods, T. E., et al. 2020b, The Open Journal of Astrophysics, 3, 15, doi: 10.21105/astro.2008.08090 * Sakurai et al. (2020a) Sakurai, Y., Haiman, Z., & Inayoshi, K. 2020a, MNRAS, 499, 5960, doi: 10.1093/mnras/staa3227 * Sakurai et al. (2020b) —. 2020b, MNRAS, 499, 5960, doi: 10.1093/mnras/staa3227 * Sakurai et al. (2015) Sakurai, Y., Hosokawa, T., Yoshida, N., & Yorke, H. W. 2015, MNRAS, 452, 755, doi: 10.1093/mnras/stv1346 * Sakurai et al. (2016) Sakurai, Y., Vorobyov, E. I., Hosokawa, T., et al. 2016, MNRAS, 459, 1137, doi: 10.1093/mnras/stw637 * Savin et al. (2004) Savin, D. W., Krstić, P. S., Haiman, Z., & Stancil, P. C. 2004, ApJ, 606, L167, doi: 10.1086/421108 * Scannapieco & Barkana (2002) Scannapieco, E., & Barkana, R. 2002, ApJ, 571, 585, doi: 10.1086/340063 * Schauer et al. (2019) Schauer, A. T. P., Glover, S. C. O., Klessen, R. S., & Ceverino, D. 2019, MNRAS, 484, 3510, doi: 10.1093/mnras/stz013 * Schauer et al. (2015) Schauer, A. T. P., Whalen, D. J., Glover, S. C. O., & Klessen, R. S. 2015, MNRAS, 454, 2441, doi: 10.1093/mnras/stv2117 * Schleicher et al. (2013) Schleicher, D. R. G., Palla, F., Ferrara, A., Galli, D., & Latif, M. 2013, A&A, 558, A59, doi: 10.1051/0004-6361/201321949 * Schneider et al. (1994) Schneider, I. F., Dulieu, O., Giusti-Suzor, A., & Roueff, E. 1994, ApJ, 424, 983, doi: 10.1086/173948 * Schulz & Asundi (1967) Schulz, G. J., & Asundi, R. K. 1967, Physical Review, 158, 25, doi: 10.1103/PhysRev.158.25 * Sesana et al. (2008) Sesana, A., Vecchio, A., & Colacino, C. N. 2008, MNRAS, 390, 192, doi: 10.1111/j.1365-2966.2008.13682.x * Shang et al. (2010) Shang, C., Bryan, G. L., & Haiman, Z. 2010, MNRAS, 402, 1249, doi: 10.1111/j.1365-2966.2009.15960.x * Shapiro & Kang (1987) Shapiro, P. R., & Kang, H. 1987, ApJ, 318, 32, doi: 10.1086/165350 * Sheth et al. (2001) Sheth, R. K., Mo, H. J., & Tormen, G. 2001, MNRAS, 323, 1, doi: 10.1046/j.1365-8711.2001.04006.x * Shibata et al. (2016) Shibata, M., Sekiguchi, Y., Uchida, H., & Umeda, H. 2016, Phys. Rev. D, 94, 021501, doi: 10.1103/PhysRevD.94.021501 * Smith et al. (2018) Smith, B. D., Regan, J. A., Downes, T. P., et al. 2018, MNRAS, 480, 3762, doi: 10.1093/mnras/sty2103 * Stancil (1994) Stancil, P. C. 1994, ApJ, 430, 360, doi: 10.1086/174411 * Sugimura et al. (2014) Sugimura, K., Omukai, K., & Inoue, A. K. 2014, MNRAS, 445, 544, doi: 10.1093/mnras/stu1778 * Tanaka et al. (2013) Tanaka, K. E. I., Nakamoto, T., & Omukai, K. 2013, ApJ, 773, 155, doi: 10.1088/0004-637X/773/2/155 * Tanaka & Haiman (2009) Tanaka, T., & Haiman, Z. 2009, ApJ, 696, 1798, doi: 10.1088/0004-637X/696/2/1798 * Tanaka & Li (2014) Tanaka, T. L., & Li, M. 2014, MNRAS, 439, 1092, doi: 10.1093/mnras/stu042 * Tegmark et al. (1997) Tegmark, M., Silk, J., Rees, M. J., et al. 1997, ApJ, 474, 1, doi: 10.1086/303434 * Toyouchi et al. (2021) Toyouchi, D., Inayoshi, K., Hosokawa, T., & Kuiper, R. 2021, ApJ, 907, 74, doi: 10.3847/1538-4357/abcfc2 * Trevisan & Tennyson (2002) Trevisan, C. S., & Tennyson, J. 2002, Plasma Physics and Controlled Fusion, 44, 1263, doi: 10.1088/0741-3335/44/7/315 * Tseliakhovich & Hirata (2010) Tseliakhovich, D., & Hirata, C. 2010, Phys. Rev. D, 82, 083520, doi: 10.1103/PhysRevD.82.083520 * Valiante et al. (2016) Valiante, R., Schneider, R., Volonteri, M., & Omukai, K. 2016, MNRAS, 457, 3356, doi: 10.1093/mnras/stw225 * Visbal et al. (2014a) Visbal, E., Haiman, Z., & Bryan, G. L. 2014a, MNRAS, 445, 1056, doi: 10.1093/mnras/stu1794 * Visbal et al. (2014b) —. 2014b, MNRAS, 442, L100, doi: 10.1093/mnrasl/slu063 * Voit et al. (2003) Voit, G. M., Balogh, M. L., Bower, R. G., Lacey, C. G., & Bryan, G. L. 2003, ApJ, 593, 272, doi: 10.1086/376499 * Voit et al. (2005) Voit, G. M., Kay, S. T., & Bryan, G. L. 2005, MNRAS, 364, 909, doi: 10.1111/j.1365-2966.2005.09621.x * Walkauskas & Kaufman (1975) Walkauskas, L., & Kaufman, F. 1975, Symposium (International) on Combustion, 15, 691, doi: https://doi.org/10.1016/S0082-0784(75)80339-0 * Wang et al. (2021) Wang, F., Yang, J., Fan, X., et al. 2021, ApJ, 907, L1, doi: 10.3847/2041-8213/abd8c6 * Willott et al. (2010) Willott, C. J., Delorme, P., Reylé, C., et al. 2010, AJ, 139, 906, doi: 10.1088/0004-6256/139/3/906 * Wise & Abel (2007) Wise, J. H., & Abel, T. 2007, ApJ, 665, 899, doi: 10.1086/520036 * Wise et al. (2019) Wise, J. H., Regan, J. A., O’Shea, B. W., et al. 2019, Nature, 566, 85, doi: 10.1038/s41586-019-0873-4 * Wolcott-Green & Haiman (2011) Wolcott-Green, J., & Haiman, Z. 2011, MNRAS, 412, 2603, doi: 10.1111/j.1365-2966.2010.18080.x * Woods et al. (2017) Woods, T. E., Heger, A., Whalen, D. J., Haemmerlé, L., & Klessen, R. S. 2017, ApJ, 842, L6, doi: 10.3847/2041-8213/aa7412 * Woods et al. (2019) Woods, T. E., Agarwal, B., Bromm, V., et al. 2019, PASA, 36, e027, doi: 10.1017/pasa.2019.14 * Wu et al. (2015) Wu, X.-B., Wang, F., Fan, X., et al. 2015, Nature, 518, 512, doi: 10.1038/nature14241 * Wyithe & Padmanabhan (2006) Wyithe, J. S. B., & Padmanabhan, T. 2006, MNRAS, 366, 1029, doi: 10.1111/j.1365-2966.2005.09858.x * Yoshida et al. (2003) Yoshida, N., Abel, T., Hernquist, L., & Sugiyama, N. 2003, ApJ, 592, 645, doi: 10.1086/375810 * Zygelman et al. (1989) Zygelman, B., Dalgarno, A., Kimura, M., & Lane, N. F. 1989, Phys. Rev. A, 40, 2340, doi: 10.1103/PhysRevA.40.2340 Table 1: Chemical Reactions Number | Reaction | Reference ---|---|--- | H collisional reactions | 1 | H + e- $\rightarrow$ H+ \+ 2e- | 1 2 | H+ \+ e- $\rightarrow$ H + $\gamma$ | 2∗ 3 | H + e- $\rightarrow$ H- \+ $\gamma$ | 3 4 | H- \+ H $\rightarrow$ H2 \+ e | 4 5 | H + H+ $\rightarrow$ H${}_{2}^{+}$ \+ $\gamma$ | 5 6 | H${}_{2}^{+}$\+ H $\rightarrow$ H2 \+ H+ | 6 7 | H2 \+ H $\rightarrow$ 3H | 7 8 | H2 \+ H+ $\rightarrow$ H${}_{2}^{+}$ \+ H | 8 9 | H2 \+ e- $\rightarrow$ 2H + e- | 9 10 | H- \+ e- $\rightarrow$ H + 2e- | 10 11 | H- \+ H+ $\rightarrow$ 2H | 11 12 | H- \+ H+ $\rightarrow$ H${}_{2}^{+}$ \+ e- | 12 13 | H${}_{2}^{+}$ \+ e- $\rightarrow$ 2H | 13 14 | H${}_{2}^{+}$ \+ H- $\rightarrow$ H2 \+ H | 14 15 | 3H $\rightarrow$ H2 \+ H | 15 16 | 2H + H2 $\rightarrow$ 2H2 | 16 17 | 2H2 $\rightarrow$ 2H + H2 | 17 18 | H- \+ H $\rightarrow$ 2H + e- | 18 19 | H- \+ H${}_{2}^{+}$ $\rightarrow$ 3H | 19 20 | H2 \+ e- $\rightarrow$ H- \+ H | 20 | photo-dissociation and detatchment reactions | 21 | H2 \+ $\gamma$ $\rightarrow$ 2H | 21 22 | H- \+ $\gamma$ $\rightarrow$ H + e- | 22 23 | H${}_{2}^{+}$ \+ $\gamma$ $\rightarrow$ H + H+ | 23 | He reactions | 24 | He + e- $\rightarrow$ He+ \+ 2e- | 24 25 | He+ \+ e- $\rightarrow$ He + $\gamma$ | 25 26 | He+ \+ e- $\rightarrow$ He++ \+ 2e- | 26 27 | He++ \+ e- $\rightarrow$ He+ \+ H+ +$\gamma$ | 27 28 | H2 \+ He $\rightarrow$ 2H + He | 28 29 | H2 \+ He+ $\rightarrow$ He + H + H+ | 29 30 | H2 \+ He+ $\rightarrow$ H${}_{2}^{+}$ \+ He | 30 31 | He+ \+ H $\rightarrow$ He + H+ | 31 32 | He + H+ $\rightarrow$ He+ \+ H | 32 33 | He+ \+ H- $\rightarrow$ He + H | 33 34 | He + H- $\rightarrow$ He + H + e- | 34 35 | 2H + He $\rightarrow$ H2 \+ He | 35 (1) Abel et al. (1997); (2) Ferland et al. (1992), Case B; (3) McLaughlin et al. (2017); (4) Kreckel et al. (2010); (5) Coppola et al. (2011); (6) Karpas et al. (1979); (7) Mac Low & Shull (1986); Lepp & Shull (1983); (8) Savin et al. (2004); Coppola et al. (2011); (9) Trevisan & Tennyson (2002); (10) Janev et al. (1987); (11) Croft et al. (1999); (12) Poulaert et al. (1978); (13) Schneider et al. (1994); (14) Dalgarno & Lepp (1987); (15) Abel et al. (2002); Orel (1987); (16) Jacobs et al. (1967); (17) Martin et al. (1998); Shapiro & Kang (1987) (18) Janev et al. (1987); (19) Dalgarno & Lepp (1987); (20) Schulz & Asundi (1967); (21) Wolcott-Green & Haiman (2011); (22) McLaughlin et al. (2017); (23) Stancil (1994); (24) Janev et al. (1987); (25) Hummer & Storey (1998); (26) Janev et al. (1987); (27) Ferland et al. (1992); (28) Dove et al. (1987); (29) Barlow (1984); (30) Barlow (1984); (31) Zygelman et al. (1989); (33) Kimura et al. (1993); (33) Peart & Hayton (1994); (34) Huq et al. (1982); (35) Walkauskas & Kaufman (1975);
final portion of the segments joining ${\mathbb{X}}_{t-1}^{\epsilon}$ and ${\mathbb{X}}_{t}^{\epsilon}$. Thus, at each step, we arrive at a vertex of a hexagon. For this Exploration Process we still maintain the following properties. ###### Proposition 4.5 (Proposition 4.3, [6]). Let $\gamma_{\epsilon}([0,t])$ be the line segments formed by the process up until time $t$, and $\Gamma_{\epsilon}([0,t])$ be the hexagons revealed by the Exploration Process. Let $\partial\Omega_{\epsilon}^{t}=\partial\Omega_{\epsilon}\cup\Gamma_{\epsilon}([0,t])$ and let $\Omega_{\epsilon}^{t}=\Omega_{\epsilon}\backslash\Gamma_{\epsilon}([0,t]).$ Then, the quadruple $(\Omega_{\epsilon}^{t},\partial\Omega_{\epsilon}^{t},{\mathbb{X}}_{\epsilon}^{t},c_{\epsilon})$ is admissible. Furthermore, the Exploration Process in $\Omega_{\epsilon}^{t}$ from ${\mathbb{X}}_{\epsilon}^{t}\text{ to }c_{\epsilon}$ has the same law as the original Exploration Process from $a_{\epsilon}$ to $c_{\epsilon}$ in $\Omega_{\epsilon}$ conditioned on $\Gamma_{\epsilon}([0,t]).$ Percolation satisfies the KS Condition. It is well known that the Exploration Process produces in any critical percolation configuration $\Omega_{\epsilon}$, the unique interface connecting $a_{\epsilon}$ to $c_{\epsilon}$ denoted by $\gamma_{\epsilon}$, i.e. the unique curve which separates the blue connected cluster of the boundary from the yellow connected cluster of the boundary. Let ${\mathbb{P}}_{\Omega_{\epsilon}}$ be the law of this interface. Let $\mu_{\epsilon}$ be the probability measure on random curves induced by the Exploration Process on $\Omega_{\epsilon}$, and let us endow the space of curves with the sup-norm metric $\mathrm{dist}(\gamma_{1},\gamma_{2})=\inf\limits_{\varphi_{1},\varphi_{2}}\sup\limits_{t}|\gamma_{1}(\varphi_{1}(t))-\gamma_{2}(\varphi_{2}(t))|$ over all possible parameterizations $\varphi_{1},\varphi_{2}$. The proof of the fact that the collection $({\mathbb{P}}_{\Omega}\;:\;\Omega\text{ admissible})$ satisfies the KS Condition follows directly from [17, Proposition 4.13] since this generalized percolation model still satisfies Russo-Seymour-Welsh (RSW) type correlation inequalities. ###### Remark 4.6. As long as $a^{2}\geq 2s^{2}$, then a restricted form of Harris-FKG property holds for all paths and path type events, see [10, Lemma 6.2]. Since we have this essential ingredient in the RSW type arguments, we are indeed free to use RSW sort of correlation inequalities. ###### Proposition 4.7 (Proposition 4.13 in [17]). The collection of the laws of the interface of the modified bond percolation model described above on the hexagonal lattice. $\Sigma_{\mathrm{Percolation}}=\\{(\Omega_{\epsilon},\phi(\Omega_{\epsilon}),{\mathbb{P}}_{\Omega_{\epsilon}}):\Omega_{\epsilon}\text{ an admissible domain}\\}$ (4.1) satisfies the KS Condition. ###### Proof. First, notice that for percolation, we do not have to consider stopping times. Indeed, by Proposition (4.5) if $\gamma:[0,N]\rightarrow\Omega_{\epsilon}\cup\\{a,c\\}$ is the interface parameterized so that $\gamma(k),\;k=0,1,\cdots,N$ are vertices along the path, then $\Omega_{\epsilon}\backslash\gamma(0,k]$ is admissible for any $k=0,1,\cdots,N$ and there is no information gained during $(k,k+1)$. Also, the law of percolation satisfies the domain Markov property so the law conditioned to the vertices explored up to time $n$ is the percolation measure in the domain where $\gamma(k),\;k=0,1,\cdots,n$ is erased. Thus, the family ((4.1)) is closed under stopping. Since crossing an annuli is a translation invariant event for percolation, for any $\Omega_{\epsilon}$, we can apply a translation and consider the annuli around the origin. Let $B_{n}$ be the set of points on the triangular lattice that are graph distance less than or equal to $n$ from $0$. Consider the annulus $B_{9^{N}n}\backslash B_{n}$ for any $n,N\in{\mathbb{N}}$. We can consider concentric balls $B_{3n}$ inside the annulus $B_{9^{N}n}\backslash B_{n}$. Then for an open crossing of the annulus $B_{9^{N}n}\backslash B_{n}$, there needs to be an open path inside each annulus $A_{n}=B_{3n}\backslash B_{n},A_{3n}=B_{9n}\backslash B_{3n},\cdots$ etc. The probability that $A_{n}$ contains an open path separating $0$ from $\infty$ and $A_{3n}$ contains a closed path separating $0$ from $\infty$ are independent. Hence, by Russo- Seymour-Welsh (RSW) theory, we know that there exists a $q>0$ for any $n$ $\mu_{\epsilon}\left(\text{open path inside }A_{n}\cap\text{ closed path in }A_{3n}\text{ both separating 0 from }\infty\right)\geq q^{2}$ Since a closed path in one of the concentric annuli prohibits an open crossing of $B_{9^{N}n}\backslash B_{n}$, we conclude that ${\mathbb{P}}_{\epsilon}\left(\gamma\text{ makes an unforced crossing of }B_{9^{N}n}\backslash B_{n}\right)\leq(1-q^{2})^{N}\leq\frac{1}{2}$ for large enough $N$. ∎ The observable. Consider two addition marked points (or prime ends) b,d so that a,b,c,d are in cyclic order. Let $\Omega_{n}$ be the admissible domain described above at lattice scale $n^{-1}$ to the domain $\Omega$. More details of the construction can be found in [6, §3 and §4] and [7, §4.2]. Furthermore, the boundary arcs can be appropriately coloured and the lattice points $a_{n},b_{n},c_{n},d_{n}$ can be selected. The main objects of study for percolation is the crossing probability of the conformal rectangle $\Omega_{n}$ from $(a_{n},b_{n})$ to $(c_{n},d_{n})$, denoted by ${\mathcal{C}}_{n}$ and ${\mathcal{C}}_{\infty}$ its limit in the domain $\Omega$, i.e., Cardy’s formula in the limiting domain. Geometrically, ${\mathcal{C}}_{n}$ produces in any percolation configuration on $\Omega_{n}$, the unique interface connecting $a_{n}$ to $c_{n}$, i.e. the curve separating the blue lattice connected cluster of the boundary from the yellow. Let us temporarily forget the marked point $a_{n}$ and consider the conformal triangle $(\Omega_{n};b_{n},c_{n},d_{n})$. We will briefly recall the observable function introduced in [28] which we will denote by $S_{b},S_{c},S_{d}$. For a lattice point $z\in\Omega_{n}$, $S_{d}(z)$ is the probability of a yellow crossing from $(c_{n},d_{n})$ to $(d_{n},b_{n})$ separating $z$ from $(b_{n},c_{n})$. Notice that $S_{d}$ has boundary value $0$ on $(b_{n},c_{n})$ and $1$ at the point $d_{n}$. $S_{b}$ and $S_{c}$ are defined similarly. We define the complexified function $S_{n}:=S_{b}+\tau S_{c}+\tau^{2}S_{d}$ with $\tau=e^{2\pi i/3}$, called the Carleson-Cardy-Smirnov (CCS) function. The following lemma due to Smirnov shows that CCS observable is a martingale. ###### Lemma 4.8. The CCS observable is a martingale observable. ###### Proof. Parameterize the interface $\gamma^{\epsilon}$ and draw the exploration process up to time $t$, $\gamma^{\epsilon}[0,t]$. By convention/definition, the faces on the left and right side of the exploration process are yellow and blue, respectively. Then any open crossing (yellow crossing) from arc $bc$ to the arc $db$ inside $\Omega$ is either disjoint from $\gamma^{\epsilon}[0,t]$ or hits its ”open” (yellow) arc of $\gamma^{\epsilon}[0,t]$. Either case produces an open crossing from the arc $\gamma^{\epsilon}(t)c$ to the arc $d\gamma^{\epsilon}(t)$ inside $\Omega\backslash\gamma^{\epsilon}[0,t]$. The converse also holds. Thus, we have the following observation: crossing probabilities conditioned on $\gamma^{\epsilon}[0,t]$ coincide with crossing probabilities in the slit domain $\Omega\backslash\gamma^{\epsilon}[0,t].$ Let $Q$ denote the area above the lowest (i.e. closest to arc $bc$) open crossing from arc $cd$ to arc $db$. Then $S_{d}(z)={\mathbb{P}}(z\in Q)$. We can view the other crossing probabilities $S_{b}$ and $S_{c}$ in the same way. By the above observation, we know that this probability conditioned on $\gamma^{\epsilon}[0,t]$ will coincide with the probability in the slit domain $\Omega\backslash\gamma^{\epsilon}[0,t].$ The same holds true for $S_{b}(z)$ and $S_{c}(z)$. Figure 5: Observe that the lowest yellow crossing cannot cross the curve $\gamma$ since it is blocked by the blue side of the curve. Thus, one sees for every realization of $\gamma^{\epsilon}[0,t]$, the CCS function conditioned on $\gamma^{\epsilon}[0,t]$ coincides with the CCS function in the slit domain, an analogue of the Markov property. Stopping the curve at times $0<t<s$, say with the least discrete time such that the path has capacity $\geq t$, and using the total probability for every realization of $\gamma^{\epsilon}[0,t]$ we get the martingale property: ${\mathbb{E}}_{\mu_{\epsilon}}\left[S_{\epsilon}(\Omega_{\epsilon}\backslash\gamma^{\epsilon}[0,s],\gamma^{\epsilon}(s),b,c,d)|\gamma^{\epsilon}[0,t]\right]=S_{\epsilon}(\Omega_{\epsilon}\backslash\gamma^{\epsilon}[0,t],\gamma^{\epsilon}(t),b,c,d).$ ∎ The CCS functions $S_{n}$ are not discrete analytic but are “almost” discrete analytic in the following sense, see [8, §4]: ###### Definition 4.9 ($(\sigma,\rho)-$Holomorphic). Let $\Lambda\subseteq\mathbb{C}$ be a simply connected domain and $\Lambda_{\epsilon}$ be the (interior) discretized domain given as $\Lambda_{\epsilon}:=\bigcup_{h_{\epsilon}\subseteq\Lambda}h_{\epsilon}$ and let $(Q_{\epsilon}:\Lambda_{\epsilon}\to\mathbb{C})_{\epsilon\searrow 0}$ be a sequence of functions defined on the vertices of $\Lambda_{\epsilon}$. We say that the sequence $(Q_{\epsilon})$ is _$(\sigma,\rho)$ –holomorphic_ if there exist constants $0<\sigma,\rho\leq 1$ such that for all $\epsilon$ sufficiently small: 1. 1. $Q_{\varepsilon}$ is Hölder continuous up to $\partial\Lambda_{\epsilon}$: There exists some small $\psi>0$ and constants $c,C\in(0,\infty)$ (independent of domain and $\epsilon$) such that 1. (a) if $z_{\epsilon},w_{\epsilon}\in\Lambda_{\epsilon}\setminus N_{\psi}(\partial\Lambda_{\epsilon})$ such that $|z_{\epsilon}-w_{\epsilon}|<\psi$, then $|Q_{\epsilon}(z_{\epsilon})-Q_{\epsilon}(w_{\epsilon})|\leq c\left(\frac{|z_{\epsilon}-w_{\epsilon}|}{\psi}\right)^{\sigma}$ and 2. (b) if $z_{\epsilon}\in N_{\psi}(\partial\Lambda_{\epsilon})$, then there exists some $w_{\epsilon}^{\star}\in\partial\Lambda_{\epsilon}$ such that $|Q_{\epsilon}(z_{\epsilon})-Q_{\epsilon}(w_{\epsilon}^{\star})|\leq C\left(\frac{|z_{\epsilon}-w_{\epsilon}^{\star}|}{\psi}\right)^{\sigma}$. 2. 2. For any simply closed lattice contour $\Gamma_{\epsilon}$, $\left|\oint_{\Gamma_{\epsilon}}Q~{}dz\right|=\left|\sum_{h_{\epsilon}\subseteq\Lambda_{\epsilon}^{\prime}}\oint_{\partial h_{\epsilon}}Q~{}dz\right|\leq c\cdot|\Gamma_{\epsilon}|\cdot\epsilon^{\rho},$ (4.2) with $c\in(0,\infty)$ (independent of domain and $\epsilon$) and $\Lambda_{\epsilon}^{\prime},|\Gamma_{\epsilon}|$ denoting the region enclosed by $\Gamma_{\epsilon}$ and the Euclidean length of $\Gamma_{\epsilon}$, respectively. ###### Proposition 4.10 (Proposition 4.3, [8]). Let $\Lambda$ denote a conformal triangle with marked points (or prime ends) $b$, $c$, $d$ and let $\Lambda_{\epsilon}$ denote an interior approximation (see [7, Definition 3.1]) of $\Lambda$ with $b_{\epsilon},c_{\epsilon},d_{\epsilon}$ the associated boundary points. Let $S_{\epsilon}(z)$ denote the CCS function defined on $\Lambda_{\epsilon}$. Then for all $\epsilon$ sufficiently small, the functions $(S_{\epsilon}:\Lambda_{\epsilon}\rightarrow\mathbb{C})$ are $(\sigma,\rho)$–holomorphic for some $\sigma,\rho>0$. Polynomial convergence of the observable function to its continuous counterpart. Observe that ${\mathcal{C}}_{n}$ can be realized from $S_{d}(a_{n})$ as ${\mathcal{C}}_{n}=\frac{-2}{\sqrt{3}}\cdot\operatorname{Im}[S_{n}(a_{n})]$. Since it is already known that $S_{n}$ converges to $H:D\to T$, a conformal map to equilateral triangle $T$ which sends $(b,c,d)$ to $(1,\tau,\tau^{2})$, we can see that ${\mathcal{C}}_{\infty}=\frac{-2}{\sqrt{3}}\operatorname{Im}[H(a)]$ (see, [28], [2], and [7]). Thus, when establishing a rate of convergence of ${\mathcal{C}}_{n}$ to ${\mathcal{C}}_{\infty}$, it is sufficient to show that there exists $\psi>0$ such that $|S_{n}(a_{n})-H(a)|\leq C_{\psi}\cdot n^{-\psi}$ for some $C_{\psi}<\infty$ independent of the domain. Indeed, a polynomial rate of convergence is shown in [8, Main Theorem]. This is a slight reformulation of the theorem in which we have that the constant $\psi$ is independent of the domain $\Omega$. Indeed, a direct reconstruction of the proof in [8] gives this result. ###### Theorem 4.11. Let $\Omega$ be a domain with two marked boundary points (or prime ends) $a$ and $c$. Let $(\Omega_{n},a_{n},c_{n})$ be its admissible discretization. Consider the site percolation model or the models introduced in [10] on the domain $\Omega_{n}$. In the case of the latter we also impose the assumption that the boundary Minkowski dimension is less than 2 (in the former, this is not necessary). Let $\gamma$ be the interface between $a$ and $c$. Consider the stopping time $T:=\inf\\{t\geq 0\;:\;\gamma\text{ enters a }\Delta\text{-neighbourhood of }c\\}$ for some $\Delta>0$. Then there exists $n_{0}<\infty$ depending only on the domain $(\Omega;a,b,c,d)$ and $T$ such that the following estimate holds: There exists some $\psi>0$ (which does not depend on the domain $\Omega$) such that $\mathscr{C}_{n}$ converges to its limit with the estimate $|\mathscr{C}_{n}-\mathscr{C}_{\infty}|\leq C_{\psi}\cdot n^{-\psi},$ for some $C_{\psi}<\infty$ provided $n\geq n_{0}(\Omega)$ is sufficiently large. Polynomial convergence of critical percolation on the triangular lattice. By a straightforward computation, we can see that the martingale observable is a nondegenerate solution to BPZ equation ((1.4)). Thus, by Proposition (4.7), Lemma (4.8), Proposition (4.10), and Theorem (4.11), we can now apply Theorem (1.23) to obtain: ###### Theorem 4.12. Let $\gamma_{n}$ be the percolation Exploration Process defined above on the admissible triangular lattice domain $\Omega_{n}$. Let $\tilde{\gamma}_{n}$ be its image in $({\mathbb{H}};0,\infty)$ parameterized by capacity. There exists stopping time $T<\infty$ and $n_{1}$ such that $\displaystyle\sup_{n}\sup_{t\in[0,T]}n_{1}(\Omega_{t})<\infty$. Then if $n\geq n_{1}$, there is a coupling of $\gamma_{n}$ with Brownian motion $B(t),\;t\geq 0$ with the property that if $\tilde{\gamma}$ denotes the chordal SLE6 path in ${\mathbb{H}}$, ${\mathbb{P}}\left\\{\sup_{t\in[0,T]}|\tilde{\gamma}_{n}(t)-\tilde{\gamma}(t)\;|\;>n^{-u}\right\\}<n^{-u}$ for some $u\in(0,1)$ and where both curves are parameterized by capacity. Moreover, if $\Omega$ is an $\alpha$-Hölder domain, then under the same coupling, the SLE curve in the image is polynomially close to the original discrete curve: $\mathbb{P}\left\\{\sup_{t\in[0,T]}d_{*}\left(\gamma^{n}(t),\phi^{-1}(\tilde{\gamma}(t))\right)>n^{-v}\right\\}<n^{-v}$ where $v$ depends only on $\alpha$ and $u$. ###### Remark 4.13. The authors believe that modifications of the arguments in [8] could lead to a full convergence statement. ###### Remark 4.14. Notice that under this modified percolation model, we still maintain the reversibility of the exploration path. Let $\omega$ be a simple polygonal path from $a^{\delta}$ to $c^{\delta}$. Suppose that the corresponding path designate is the sequence $\left[H_{0,1},(\mathcal{F}_{1},h_{1}^{e},h_{1}^{x}),H_{1,2},(\mathcal{F}_{2},h_{2}^{e},h_{2}^{x}),H_{2,3},\cdots,(\mathcal{F}_{K},h_{K}^{e},h_{K}^{x}),H_{K,K+1}\right]$ where $\mathcal{F}_{1},\cdots\mathcal{F}_{K}$ are flowers in $\Omega^{\delta}$ with $h_{j}^{e}$ and $h_{j}^{x}$ are the entrance and exit petals in the $j^{th}$ flower and for $1\leq j\leq K-1,\;H_{j,j+1}$ is a path in the complement of flowers which connects $h_{j}^{x}$ to $h_{j+1}^{e}$. That is, we are not viewing the microscopic description where we have to specifying how the path got between entry and exit petals. With a small loss of generality we are also assuming that the path only visits the flower once else we would have to specify the first entrance and exit petals, the second entrance and exit petals, etc. Let $\gamma^{\delta}$ be a chordal exploration process from $a^{\delta}$ to $c^{\delta}$ in $\Omega^{\delta}$ and $\hat{\gamma}^{\delta}$ be a chordal exploration process from $c^{\delta}$ to $a^{\delta}$ in $\Omega^{\delta}$. Recall that all petal arrangements are independent, all flowers are configured independently and these in turn are independent of the background filler sites. Thus the exploration process generated by the colouring algorithm given previously, excluding colouring of flowers, is independent and flowers are independent of background filler sites. Thus, by the colouring algorithm we have: $\displaystyle{\mathbb{P}}(\gamma^{\delta}=\omega)=\left(\frac{1}{2}\right)^{l(H_{0,1})}p_{1}\left(\frac{1}{2}\right)^{l(H_{1,2})}\cdots p_{K}\left(\frac{1}{2}\right)^{l(H_{K,K+1})}$ where $l(H_{j,j+1})$ is the number of coloured hexagons in $H_{j,j+1}$ produced by the colouring algorithm on the event $\gamma^{\delta}=\omega$ and $p_{j}$ is the appropriate conditional probabilities on each flower of a petal or iris given by the colouring algorithm. Notice that on the event $\gamma^{\delta}=\hat{\gamma}^{\delta}=\omega$ for any hexagon in $\Omega^{\delta}$ either it is coloured by both the colouring algorithm for $\gamma^{\delta}$ and the colouring algorithm for $\hat{\gamma}^{\delta}$ or by neither. Therefore, we have the following lemma: ###### Lemma 4.15. Suppose $\Omega^{\delta}$ is a simply connected domain in the $\delta$-hexagonal lattice with a predetermined flower arrangement. For any simple polygonal path $\omega$ from $a^{\delta}$ to $c^{\delta}$ we have ${\mathbb{P}}(\gamma^{\delta}=\omega)={\mathbb{P}}(\hat{\gamma}^{\delta}=\omega)$ This lemma directly implies the following lemma. ###### Lemma 4.16. For any simply connected domain $\Omega^{\delta}$ with predetermined flower arrangement, the percolation exploration path from $a^{\delta}$ to $c^{\delta}$ in $\Omega^{\delta}$ has the same distribution as the time- reversal of the percolation exploration path from $c^{\delta}$ to $a^{\delta}$ in $\Omega^{\delta}$. ###### Question 4.17. Is it possible to use reversibility to extend the polynomial convergence for the whole curve percolation exploration process? ## References * [1] M. Aizenman and A. Burchard. Hölder regularity and dimension bounds for random curves. Duke Math. J., 99(3):419–453, 1999. * [2] Vincent Beffara. Cardy’s formula on the triangular lattice, the easy way. 02 2007. * [3] A. A. Belavin, A. M. Polyakov, and A. B. Zamolodchikov. Infinite conformal symmetry in two-dimensional quantum field theory. Nuclear Physics B, 241(2):333–380, July 1984. * [4] D. Beliaev. Conformal maps and geometry. 2019\. * [5] Christian Beneš, Fredrik Johansson Viklund, and Michael J. Kozdron. On the rate of convergence of loop-erased random walk to ${\rm SLE}_{2}$. Comm. Math. Phys., 318(2):307–354, 2013. * [6] I. Binder, L. Chayes, and H. K. Lei. On convergence to $\text{SLE}_{6}$ I: Conformal invariance for certain models of the bond-triangular type. J. Stat. Phys., 141(2):359–390, 2010. * [7] I. Binder, L. Chayes, and H. K. Lei. On convergence to $\text{SLE}_{6}$ II: Discrete approximations and extraction of Cardy’s formula for general domains. J. Stat. Phys., 141(2):391–408, 2010. * [8] I. Binder, L. Chayes, and H. K. Lei. On the rate of convergence for critical crossing probabilities. Ann. Inst. Henri Poincaré Probab. Stat., 51(2):672–715, 2015\. * [9] Ilia Binder, Dmitry Chelkak, and Larissa Richards. Polynomial convergence rate of harmonic explorer and ising model. Preprint. * [10] L. Chayes and H. Lei. Cardy’s formula for certain models of the bond-triangular type. Reviews in Mathematical Physics, 19, 01 2006. * [11] Dmitry Chelkak. Ising model and s-embeddings of planar graphs, 2020. * [12] Dmitry Chelkak, Hugo Duminil-Copin, Clément Hongler, Antti Kemppainen, and Stanislav Smirnov. Convergence of Ising interfaces to Schramm’s SLE curves. C. R. Math. Acad. Sci. Paris, 352(2):157–161, 2014. * [13] M. Csörgő and P. Révész. Strong Approximations in Probability and Statistics. 1981\. * [14] Richard M. Dudley. Real analysis and probability. Wadsworth & Brooks/Cole Advanced Books & Software, Pacific Grove, CA, 1989. * [15] Erich Haeusler. An exact rate of convergence in the functional central limit theorem for special martingale difference arrays. Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete, 65:523 – 534, 1984. * [16] Fredrik Johansson Viklund. Convergence rates for loop-erased random walk and other Loewner curves. Ann. Probab., 43(1):119–165, 2015. * [17] Antti Kemppainen and Stanislav Smirnov. Random curves, scaling limits and Loewner evolutions. Ann. Probab., 45(2):698–779, 2017. * [18] Gregory F. Lawler. Conformally invariant processes in the plane, volume 114 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2005. * [19] Gregory F. Lawler, Oded Schramm, and Wendelin Werner. Values of brownian intersection exponents, i: Half-plane exponents. Acta Mathematica, 187(2):237–273, 2001. * [20] Gregory F. Lawler, Oded Schramm, and Wendelin Werner. Conformal invariance of planar loop-erased random walks and uniform spanning trees [mr2044671]. In Selected works of Oded Schramm. Volume 1, 2, Sel. Works Probab. Stat., pages 931–987. Springer, New York, 2011. * [21] Gregory F. Lawler and Fredrik Viklund. Convergence of loop-erased random walk in the natural parametrization, 2016. * [22] Ch. Pommerenke. Boundary behaviour of conformal maps, volume 299 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1992. * [23] Ch. Pommerenke. Conformal maps at the boundary. In Handbook of complex analysis: geometric function theory, Vol. 1, pages 37–74. North-Holland, Amsterdam, 2002. * [24] Steffen Rohde and Oded Schramm. Basic properties of SLE [mr2153402]. In Selected works of Oded Schramm. Volume 1, 2, Sel. Works Probab. Stat., pages 989–1030. Springer, New York, 2011. * [25] Oded Schramm. Scaling limits of loop-erased random walks and uniform spanning trees. Israel J. Math., 118:221–288, 2000. * [26] Oded Schramm and Scott Sheffield. Harmonic explorer and its convergence to ${\rm SLE}_{4}$. Ann. Probab., 33(6):2127–2148, 2005. * [27] Oded Schramm and Scott Sheffield. Contour lines of the two-dimensional discrete Gaussian free field. Acta Math., 202(1):21–137, 2009. * [28] Stanislav Smirnov. Critical percolation in the plane: conformal invariance, cardy’s formula, scaling limits. Comptes Rendus de l’Académie des Sciences - Series I - Mathematics, 333(3):239 – 244, 2001. * [29] Stanislav Smirnov. Critical percolation and conformal invariance. In XIVth International Congress on Mathematical Physics, pages 99–112. World Sci. Publ., Hackensack, NJ, 2005. * [30] Stanislav Smirnov. Towards conformal invariance of 2D lattice models. In International Congress of Mathematicians. Vol. II, pages 1421–1451. Eur. Math. Soc., Zürich, 2006. * [31] S. E. Warschawski. On the degree of variation in conformal mapping of variable regions. Trans. Amer. Math. Soc., 69:335–356, 1950.
# Ethics of Generating Synthetic MRI Vocal Tract Views from the Face CVPR Responsible Generative AI Workshop Muhammad Suhaib Shahid University of Nottingham NG8 1BB, UK <EMAIL_ADDRESS>Gleb E. Yakubov University of Nottingham LE12 5RD, UK Andrew P. French University of Nottingham NG8 1BB, UK ###### Abstract Forming oral models capable of understanding the complete dynamics of the oral cavity is vital across research areas such as speech correction, designing foods for the aging population, and dentistry. Magnetic resonance imaging (MRI) technologies, capable of capturing oral data essential for creating such detailed representations, offer a powerful tool for illustrating articulatory dynamics. However, its real-time application is hindered by expense and expertise requirements. Ever advancing generative AI approaches present themselves as a way to address this barrier by leveraging multi-modal approaches for generating pseudo-MRI views. Nonetheless, this immediately sparks ethical concerns regarding the utilisation of a technology with the capability to produce MRIs from facial observations. This paper explores the ethical implications of external-to-internal correlation modeling (E2ICM). E2ICM utilises facial movements to infer internal configurations and provides a cost-effective supporting technology for MRI. In this preliminary work, we employ Pix2PixGAN to generate pseudo-MRI views from external articulatory data, demonstrating the feasibility of this approach. Ethical considerations concerning privacy, consent, and potential misuse, which are fundamental to our examination of this innovative methodology, are discussed as a result of this experimentation. ## 1 Introduction The ability to model the complete oral cavity holds significant utility across various domains, notably in dentistry, where understanding how dental prosthetics impact speech and mastication is crucial in forming personalised dental devices. However, achieving comprehensive oral cavity modeling presents challenges, including limitations in technique availability, articulator capture capacity, and associated costs. Researchers often face the dilemma of selecting a suitable technique tailored to their specific objectives. Among the available methods, magnetic resonance imaging (MRI) stands out for its capability to provide detailed representations of all articulators, particularly when augmented with real-time functionality, offering insights into dynamic movements. Nonetheless, the practicality of real-time MRI is hindered by its costliness and the need for specialised expertise, rendering it less feasible for routine use. This raises the question, is it possible to use generative AI approaches to achieve a complete model of the oral cavity, encompassing all articulators in motion, without incurring excessive expenses? This is where the concept of external-to-internal correlation modeling (E2ICM) emerges as a potential solution. By observing the external facial movements, particularly those of visible articulators such as the lips, and jaw, we investigate if we can in any way reconstruct the internal configurations of the oral cavity. This approach leverages the inherent relationship between external facial gestures and internal vocal tract configurations. Such an approach aims to address the cost and complexity concerns associated with MRI and other experimental techniques. Clearly there are limitations to this approach, but here we consider exploring the feasibility of such a technology, and bring to the fore the ethical questions such an approach might raise. As advancements in AI-based approaches continue to progress, questions regarding ethical implications become increasingly pertinent. The ability to record or photograph individuals during articulation and mastication, followed by the generation of MRI-like images of the internal oral cavity, raises several ethical considerations and potential concerns regarding privacy, consent, and misuse. This paper explores the application of generative deep learning models to create pseudo-MRI views of the oral cavity. Specifically, it employs the Pix2PixGAN network to transform external views of a participant during articulation into predicted MRI representations, in a limited speech- reconstruction scenario. For the purpose of demonstrating the feasibility of the proposed approach we briefly evaluate the challenges associated with determining the quality of the generated images. This is followed by a discussion focused on possible ethical dilemmas associated with the use of such generated data. ## 2 Background Real time MRI (RtMRI) of the vocal tract is one of the very few techniques capable of displaying, frame by frame, the movements of all articulators during speech[6]. The method allows researchers to explore a wide range of applications from articulatory studies, to oral health and food oral processing. Despite the prospects RtMRI presents, there are some underpinning issues that hinder its widespread use. These limitations are a result of the cost and expertise requirements for collecting RtMRI data on an individual subject basis [10, 3]. One possible generative solution to this is by forming predictive models capable of using external observations of the face to synthesise a representation of the vocal tract MRI view. The feasibility of such an approach relies on investigating the interrelationship between the internal vocal tract and external face views. Such research has explored correlations between the two views by linking facial movements, captured via video, with vocal tract dynamics, captured through rtMRI [7]. The main focus is to identify whether there is sufficient mutual information between the forward coronal view of the face and the sagittal MRI to make reconstruction procedures possible. Employing Principal Component Analysis (PCA), Scholes et al. (2020) simplified the data and identified key patterns of change in both modalities. Through this process, they uncovered connections between facial gestures and vocal tract configurations, showcasing the potential for mutual reconstruction between the two modalities. The findings concluded that facial information may hold sufficient data to recover certain vocal tract shapes during speech production. While the PCA-based analysis-by-synthesis technique showcases an interrelationship between the two modalities, it comes short of addressing key barriers that prevent the widespread use of MRI. In order to reconstruct an MRI representation of the vocal tract, a corresponding PCA matrix must accompany each specific external view. However, the PCA representation is derived from the MRI image, and consequently, the MRI data are still necessary each time the representation is created. Paving the way to addressing this problem are generative machine learning models capable of performing cross-modality synthesis of unseen MRI configurations when presented with a novel face view for a specific individual [11]. This technique is commonly used in computer vision and machine learning to create mappings between different visual styles, attributes, or characteristics. It is the process of transforming an input image from one domain into a corresponding output image in another domain, while preserving meaningful content and maintaining consistency between the two domains; it involves changing how an image looks while keeping its underlying meaning intact. In the application of this task, it would involve shifting from a face view to a MRI vocal tract view for any two paired frames; this pairing being key to the approach. The Pix2PixGaN framework [1] serves as the translation network chosen for this task. The architecture comprises two key components: a generator and a discriminator. The generator works to produce a realistic mapping from the input domain to the desired output, while the discriminator’s role is to determine whether an image is real or synthesised. The generator and discriminator train in an adversarial fashion, each trying to optimise ahead of the other. This approach drives the mapping of images from one domain to another in a supervised manner. Once trained, the system has the potential to operate in an autonomous manner and predict internal views based on the outside image or video only. If successful, such approaches can enable generating synthesised views without specialised equipment and a person’s consent, which raises important ethical questions and considerations that need to be addressed. Existing research has explored the (bio)ethical considerations surrounding the use of Generative Adversarial Networks (GANs) for generating medical images. The integration of AI technologies in healthcare raises complex legal, ethical, and technical challenges. In their work [5], the authors underscore the necessity for a regulatory framework to ensure the safe integration of generative technologies in medical contexts. A systematic review conducted by [2] examined recent GAN architectures utilised in medical image analysis, revealing imbalances in their capabilities, particularly with smaller datasets. These findings align with the observations of [4], regarding the imbalanced class distributions often observed in datasets, thereby raising ethical concerns. ## 3 Framework and Implementation In this study we used a dual-modal dataset used for this study, comprising registered videos captured during speech, this has been previously published[8]. Initial data collection involved 13 participants articulating a predetermined set of 10 sentences. Participants underwent two recording sessions: first, speaking the sentences in front of a camera, and second, repeating the same sentences during MRI scans. Subsequently, these video sets were then aligned. Initially, data from 13 participants were collected for the study. However, only data from 11 participants were ultimately included in the published datasets as the study focused on British English speakers. The dataset encompasses videos providing a frontal view of the face alongside sagittal MRI views. For the purposes of this preliminary study, only data from one subject was utilised, as they were the only participant for whom all 10 videos were available across all sentences. Across these 10 videos, a total of 461 frames were available, considering the videos were recorded at a frame rate of 15 frames per second (fps). The shortest video contained 30 frames, while the longest comprised 59 frames. An implementation of Pix2PixGaN framework was used as the image-to-image translation network. Based on the conditional generative adversarial network (CGaN), the architecture consists of a generator and a discriminator. The generator aims to produce realistic images based on the input, while the discriminator’s job is to distinguish between real and generated images. The generator employs a U-Net-inspired encoder-decoder architecture with skip connections. The encoder module is formed of only convolutional layers, omitting dropout. This structure forms the following sequence of layers: C64-C128-C256-C512-C512-C512-C512-C512. The decoder integrates dropout layers with a dropout rate of 0.5 in the first, second, and third layers. The decoder’s structure is as follows: CD512-CD512-CD512-C512-C256-C128-C64. This combination establishes a proficient generator capable of producing coherent translations for this dataset. The model was optimised using the Adam optimiser with hyperparameters $\alpha$ = 0.0002, $\beta_{1}$ = 0.5, $\beta_{2}$ = 0.999, and $\epsilon$ = 1e-08. The training was done with a batch size of 16, for 200 epochs. Tanh activation function was used. ## 4 Results The Fréchet Inception Distance (FID) metric and Structural Similarity Index Measure (SSIM) were used to accesses the quality of image examination. FID provides a quantitative measure of similarity between the distribution of generated and authentic images, with lower scores indicating higher quality. SSIM considers the structural information of images, accounting for spatial relationships beyond pixel values. Additionally, a qualitative evaluation was conducted by observing the movements of each articulator frame by frame, drawing conclusions regarding which articulators are constructed most effectively. The FID score for generated images compared to ground truth is 30.80, though establishing an understanding of what a ”good” FID score is when transitioning from RGB to MRI domains remains challenging. To gain some insight, an FID was calculated for various ground truth frames to assess how well the FID performs with real images but of different vocal tract views, yielding a score of 19.75. While FID offers valuable insight into image similarity, it doesn’t directly consider the spatial representation of vocal tract structure. Therefore, SSIM might be more suitable for this task. The average SSIM score for all 46 test images was 0.7961, with higher scores indicating better image quality on a scale from -1 to 1. We recognise, and highlight, that interpreting these scores in this application domain is challenging. As illustrated in Figure 1, the generative models demonstrate some proficiency in generating images with realistic appearances, particularly showcasing discernible movements in the jaw regions. However, upon closer examination, specific articulatory details are challenging to determine. Despite reasonable FID and SSIM scores indicating overall good image similarity, inconsistencies between generated articulators and ground truth are apparent in certain frames. These discrepancies could potentially lead to misleading interpretations in clinical applications, where images resembling plausible MRI scans but with incorrect articulator configurations may pose risks. Moving forward, vocal tract segmentation could serve as a promising avenue for enhancing the clinical relevance in assessing the quality of generated vocal tract views [9]. Figure 1: Still frames sample the external view (left), ground truth MRI frame (middle) and generated frame (right). ## 5 Discussion of Ethics and Responsible Use The ethical dimensions surrounding the potential of such generative medical approaches demand scrutiny. There are concerns associated with the generation and utilisation of synthetic views, the accuracy and reliability of the data, and potential misuse. ### 5.1 Synthetic dataset enlargement The fundamental principle of informed consent and participant autonomy is central. Ethical protocols governing the collection of MRI data are universally stringent, dictating both the type of data collected and its storage practices. Participants are provided with comprehensive information about the study, including potential risks and the intended utilisation of their data, to make informed decisions regarding consent. The integration of generative AI raises logistical and ethical considerations perhaps not originally conceived. In a practical application, MRI data might initially be collected for a limited set of sentences, such as the 10 here. Subsequently, using generative techniques, additional MRI data could be synthesised for sentences that were not originally captured via MRI. When using a trained model, it’s possible to employ readily available facial data, especially from public spaces where recording may be allowed by law. However, the ethical dilemma arises: Is it acceptable and responsible to generate new data modalities without obtaining explicit consent? While enlarging datasets using generative AI techniques is not novel and has been applied in various domains, the unique aspect here lies in the translation from an external view to an internal view. Accessing and utilising facial data for such a task, even if publicly available, must be carefully assessed to uphold principles of privacy. ### 5.2 Accuracy of generated images Responsible use of generative AI necessitates addressing concerns surrounding the accuracy and integrity of synthesised data. As demonstrated here, methods such as FID are employed to help assess the ”quality” of images. However, it is evident from the outset that these methods do not adequately evaluate specific spatial information in the generated images. In other words, the morphology of clinically-relevant structures is not captured well by these metrics. Often, it is also hard to identify subtle features even directly in the dataset (see Figure 1), so interpreting the quality of synthesised data in this domain is a challenge. Questions will arise regarding the potential misuse or misinterpretation of inaccurate synthetic information. Poorly performing models could lead to misdiagnosis or misinterpretation. Therefore, it is imperative to remain vigilant and implement rigorous validation procedures to ensure the reliability and accuracy of synthesised data. More work is needed to develop approaches to assess the usefulness and trustworthiness of generated images. ### 5.3 Data storage of generated images The stringent protocols governing the storage and anonymisation of MRI data are imperative to safeguard individuals’ sensitive health information. However, the creation of additional synthetic MRI data, which may still contain identifiable features or morphology, may not always undergo comparable protocol scrutiny as the original data. While large scale MRI datasets could potentially advance medical research and clinical applications, relaxed protocols for synthesised data may compromise privacy and data security. Thus, careful ethical consideration is warranted here for future research. Furthermore, there are broader societal implications to consider, particularly regarding the potential impact of synthesised medical views on areas such as identity verification. If such technology were to be deployed in contexts such as security or law enforcement, there could be implications for individuals’ rights and freedoms, including the risk of discrimination or misuse of biometric data. ### 5.4 Dataset and model biases The publicly available dataset used in this study exhibited a bias towards speakers of British English. Though it is understandable from the associated papers that this is likely an attempt to standardise an already small and challenging data set, it nevertheless highlights concerns regarding potential biases in future datasets and the models trained upon them. Certain demographic groups may be favoured in inferences, while for others the model could perform poorly. Likewise, the application of the model would likely be limited to the application studied in the dataset (e.g. speech versus chewing, for example); wrong application would lead to misleading results. The concern extends beyond data representation to the broader implications for societal equity and fairness. Though this is a problem not only relevant to this application, occurrence of such a scenario could also exacerbate existing disparities in access to resources and opportunities, with models not being tailored to regional use. Proactive measures must be implemented to mitigate bias and promote inclusivity in dataset curation and model development. Strategies may include diversifying dataset sources to encompass a broader spectrum of linguistic and cultural backgrounds, and implementing robust validation techniques to identify and mitigate bias in model predictions. ## 6 Conclusion A demonstration of an exploratory method for generating MRI images of the vocal tract has been presented. Leveraging the Pix2PixGAN architecture, this study demonstrates the application of Generative AI to synthesise previously- unseen vocal tract configurations from external facial views. The quality of the generated images has been evaluated using the Fréchet Inception Distance (FID) metric, alongside the observation of distinct articulator movements. Results are by no means conclusive at this stage, but certainly raise the question of whether this is a valid line of research in generative AI for future researchers. Furthermore, an initial discussion regarding the responsible use of generative AI in such applications has been provided. This discussion presents considerations that must be taken into account when employing such techniques. These encompass various aspects, including the enlargement of synthetic datasets, the accuracy of generated images, the storage protocols employed, and the potential biases inherent in both the dataset and the models utilised. ## References * Isola et al. [2017] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-Image Translation with Conditional Adversarial Networks. In _2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)_ , pages 5967–5976, Honolulu, HI, 2017. IEEE. * Jeong et al. [2022] Jiwoong J. Jeong, Amara Tariq, Tobiloba Adejumo, Hari Trivedi, Judy W. Gichoya, and Imon Banerjee. Systematic Review of Generative Adversarial Networks (GANs) for Medical Image Classification and Segmentation. _Journal of Digital Imaging_ , 35(2):137–152, 2022. * Kochetov [2020] Alexei Kochetov. Research methods in articulatory phonetics I: Introduction and studying oral gestures. _Language and Linguistics Compass_ , 14(4):e12368, 2020. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/lnc3.12368. * Makhlouf et al. [2023] Ahmed Makhlouf, Marina Maayah, Nada Abughanam, and Cagatay Catal. The use of generative adversarial networks in medical image augmentation. _Neural Computing and Applications_ , 35(34):24055–24068, 2023. * Paladugu et al. [2023] Phani Srivatsav Paladugu, Joshua Ong, Nicolas Nelson, Sharif Amit Kamran, Ethan Waisberg, Nasif Zaman, Rahul Kumar, Roger Daglius Dias, Andrew Go Lee, and Alireza Tavakkoli. Generative Adversarial Networks in Medicine: Important Considerations for this Emerging Innovation in Artificial Intelligence. _Annals of Biomedical Engineering_ , 51(10):2130–2142, 2023. * Ramanarayanan et al. [2018] Vikram Ramanarayanan, Sam Tilsen, Michael Proctor, Johannes Töger, Louis Goldstein, Krishna S. Nayak, and Shrikanth Narayanan. Analysis of speech production real-time MRI. _Computer Speech and Language_ , 52:1–22, 2018. * Scholes and Skipper [2020] Chris Scholes and Jeremy I. Skipper. The inter-relationship between the face and vocal-tract configuration during audio-visual speech. 2020\. Publisher: OSF. * Scholes et al. [2020] Chris Scholes, Jeremy I. Skipper, and Alan Johnston. The interrelationship between the face and vocal tract configuration during audiovisual speech. _Proceedings of the National Academy of Sciences_ , 117(51):32791–32798, 2020. Publisher: Proceedings of the National Academy of Sciences. * Shahid et al. [2024] Muhammad Suhaib Shahid, Andrew P French, Michel F. Valstar, and Gleb E. Yakubov. Research in Methodologies for Modelling the Oral Cavity. _Biomedical Physics & Engineering Express_, 2024. * Tiede et al. [2000] M.K. Tiede, Shinobu Masaki, and Eric Vatikiotis-Bateson. Contrasts in speech articulation observed in sitting and supine conditions. _Proceedings of the 5th Seminar on Speech Production_ , pages 25–28, 2000. * Xie et al. [2023] Guoyang Xie, Yawen Huang, Jinbao Wang, Jiayi Lyu, Feng Zheng, Yefeng Zheng, and Yaochu Jin. Cross-modality Neuroimage Synthesis: A Survey. _ACM Computing Surveys_ , 56(3):80:1–80:28, 2023.
∎ 11institutetext: Sachin Chauhan 22institutetext: Department of Physics, Indian Institute of Technology Bombay Powai, Mumbai, 400076, India 22email<EMAIL_ADDRESS>33institutetext: Pichai Ramadevi 44institutetext: Department of Physics, Indian Institute of Technology Bombay Powai, Mumbai, 400076, India 44email<EMAIL_ADDRESS> # $\hat{Z}$\- invariant for $SO(3)$ and $OSp(1|2)$ Groups Sachin Chauhan Pichai Ramadevi (Received: date / Accepted: date) ###### Abstract Three-manifold invariants $\hat{Z}$ (“$Z$-hat”), also known as homological blocks, are $q$-series with integer coefficients. Explicit $q$-series form for $\hat{Z}$ is known for $SU(2)$ group, supergroup $SU(2|1)$ and ortho- symplectic supergroup $OSp(2|2)$. We focus on $\hat{Z}$ for $SO(3)$ group and orthosymplectic supergroup $OSp(1|2)$ in this paper. Particularly, the change of variable relating $SU(2)$ link invariants to the $SO(3)$ & $OSp(1|2)$ link invariants plays a crucial role in explicitly writing the $q$-series. ###### Keywords: Chern-Simons theory topological field theories topological strings M-theory 3-manifold knot quantum invariant q-series colored Jones polynomial ## 1 Introduction Knot theory has attracted attention from both mathematicians and physicists during the last 40 years. The seminal work of WittenWitten:1988hf giving a three-dimensional definition for Jones polynomials of knots and links, using $SU(2)$ Chern-Simons theory on $S^{3}$, triggered a tower of new colored link invariants. Such new invariants are given by expectation value of Wilson loops carrying higher dimensional representation $R\in\mathcal{G}$ in Chern-Simons theory where $\mathcal{G}$ denotes gauge group. These link invariants are in variable ${\mathbbm{q}}$ which depends on the rank of the gauge group $\mathcal{G}$ and the Chern-Simons coupling constant $k\in\mathbb{Z}$ (For eg: when $\mathcal{G}=SU(N)$ then ${\mathbbm{q}}=\text{exp}\left(\frac{2\pi i}{k+N}\right)$). Witten’s approach also gives three-manifold invariant $Z_{k}^{\mathcal{G}}[M;{\mathbbm{q}}]$, called Chern-Simons partition function for manifold $M$, obtained from surgery of framed links on $S^{3}$(Lickorish- Wallace theorem10.2307/1970373 ; wallace_1960 ). Witten-Reshitikhin-Turaev (WRT) invariants $\tau_{k}^{\mathcal{G}}[M;{\mathbbm{q}}]$ known in the mathematics literature are proportional to the Chern-Simons partition function: $Z_{k}^{\mathcal{G}}[M;{\mathbbm{q}}]={\tau_{k}^{\mathcal{G}}[M;{\mathbbm{q}}]\over\tau_{k}^{\mathcal{G}}[S^{2}\times S^{1};{\mathbbm{q}}]}~{}.$ (1) These WRT invariants can be written in terms of the colored invariants of framed links10.2307/1970373 ; wallace_1960 ; Kaul:2000xe ; Ramadevi:1999nd . It was puzzling observation that the colored knot polynomials appear as Laurent series with integer coefficients. There must be an underlying topological interpretation of such integer coefficients. This question was answered both from mathematics and physics perspective. Initial work of Khovanovkhovanov2000categorification titled ‘cateforification’ followed by other papers on bi-graded homology theory including Khovanov-Rozansky homology led to new homological invariants. Thus the integer coefficients of the colored knot polynomials are interpreted as the dimensions of vector space of homological theory. From topological strings and intersecting branesOoguri:1999bv ; Gopakumar:1998ii ; Gopakumar:1998jq , the integers of HOMFLY-PT polynomials are interpretable as counting of BPS states. Further the connections to knot homologies within topological string context was initiated in Gukov:2004hz resulting in concrete predictions of homological invariants for some knots (see review Nawata:2015wya and references therein). Such a physics approach involving brane set up in $M$-theoryGukov:2017kmk ; Gukov:2016gkn ; Mikhaylov:2014aoa ; Ferrari:2020avq suggests the plausibility of categorification of WRT invariants $\tau_{k}^{\mathcal{G}}[M;{\mathbbm{q}}]$ for three-manifolds. However, the WRT invariants for simple three-manifolds are not a Laurent series with integer coefficients. The detailed discussion on $U(N)$ Chern-Simons partition function on Lens space $L(p,1)\equiv S^{3}/\mathbb{Z}_{p}$ (see section 6 of Gukov:2016gkn ) shows a basis transformation $Z_{k}^{\mathcal{G}}[M;{\mathbbm{q}}]\underrightarrow{~{}~{}\mathcal{S}~{}~{}}{\hat{Z}}^{\mathcal{G}}[M;q]$ so that $\hat{Z}$ are $q$-series(where variable $q$ is an arbitrary complex number inside a unit disk) with integer coefficients (GPPV conjectureGukov:2017kmk ). These $\hat{Z}$ are called the homological blocks of WRT invariants of three-manifolds $M$. Physically, the new three-manifold invariants $\hat{Z}^{\mathcal{G}}[M;q]$ is the partition function $Z_{T^{\mathcal{G}}[M]}[D^{2}\times S^{1}]$ for simple Lie groups. Here $T^{\mathcal{G}}[M]$ denote the effective 3d $\mathcal{N}=2$ theory on $D^{2}\times S^{1}$ obtained by reducing 6d $(2,0)$ theory (describing dynamics of coincident $M5$ branes) on $M$. For a class of negative definite plumbed three-manifolds as well as link complements Gukov:2017kmk ; Gukov:2019mnk ; park2020higher ; Chung:2018rea , $\hat{Z}^{SU(N)}$ has been calculated. Further, $\hat{Z}$ invariants for super unitary group $SU(n|m)$ supergroup with explicit $q$-series for $SU(2|1)$ is presented in Ferrari:2020avq . Generalisation to orthosymplectic supergroup $OSp(2|2n)$ with explicit $q$-series for $Osp(2|2)$chae2021towards motivates us to look at $\hat{Z}$ for other gauge groups. Our goal in this paper is to extract $\hat{Z}$ for the simplest orthogonal group $SO(3)$ and the simplest odd orthosymplectic supergroup $OSp(1|2)$. We take the route of relating $SU(2)$ colored link invariants to the link invariants for these two groups to obtain $\hat{Z}$ invariants. The plan of the paper is organised as follows. In section 2, we will review the developments on the invariants of knots, links and three-manifolds. We will first briefly present Chern-Simons theory and colored link invariants with explicit results for $SU(2)$ gauge group and indicate how colored $SO(3)$ and $OSp(1|2)$ link invariants can be obtained from the colored $SU(2)$ polynomials. Then, we will summarise the developments of the homological invariants. In section 3, we briefly review $\hat{Z}$-series invariant for $SU(2)$ group for the negative definite plumbed three-manifolds. This will serve as a warmup to extend to $SO(3)$ and $OSp(1|2)$ group which we will present in section 4. We summarize the results and also indicate future directions to pursue in the concluding section 5. ## 2 Knots, Links and Three-manifold Invariants In this section, we will briefly summarise new invariants in knot theory from the physics approach as well as from the mathematics approach. ### 2.1 Chern-Simons Field Theory Invariants Chern-Simons theory based on gauge group $\mathcal{G}$ is a Schwarz type topological field theory which provides a natural framework for study of knots, links and three-manifolds $M$. Chern-Simons action $S_{CS}^{\mathcal{G}}(A)$ is explicitly metric independent: $S_{CS}^{\mathcal{G}}(A)=\frac{k}{4\pi}\int_{M}Tr\left(A\wedge dA+\frac{2}{3}A\wedge A\wedge A\right)~{}.$ (2) Here $A$ is the matrix valued gauge connection based on gauge group $\mathcal{G}$ and $k\in\mathbb{Z}$ is the coupling constant. In fact, the expectation value of Wilson loop operators associated with any $m$-component link $\mathcal{L}_{m}$ are the the link invariants: $V_{R_{1},R_{2},\ldots R_{m}}^{\mathcal{G}}[\mathcal{L}_{m};{\mathbbm{q}}]=\langle W_{R_{1},R_{2},\ldots R_{m}}[\mathcal{L}_{m}]\rangle={\int{\mathcal{D}}A\exp(iS_{CS})\overbrace{P\left(\prod_{i}\Tr_{R_{i}}exp\oint_{\mathcal{K}_{i}}A\right)}^{W_{R_{1},R_{2},\ldots R_{m}}[\mathcal{L}_{m}]}\over{\underbrace{\int{\mathcal{D}}A\exp(iS_{CS})}_{Z^{\mathcal{G}}_{k}[M;{\mathbbm{q}}]}}}~{},$ (3) where $\mathcal{K}_{i}$’s denote the component knots of link $\mathcal{L}_{m}$ carrying representations $R_{i}$’s of gauge group $\mathcal{G}$ and $Z^{\mathcal{G}}_{k}[M;{\mathbbm{q}}]$ defines the Chern-Simons partition function encoding the topology of the three-manifold $M$. Exploiting the connection between Chern-Simons theory, based on group $\mathcal{G}$, and the corresponding Wess-Zumino-Witten (WZW) conformal field theory with the affine Lie algebra symmetry $\mathfrak{g}_{k}$, the invariants of these links embedded in a three-sphere $M=S^{3}$ can be explicitly written in variable ${\mathbbm{q}}$ : ${\mathbbm{q}}=\exp({2\pi i\over k+C_{v}})~{},$ (4) which depends on the coupling constant $k$ and the dual Coxeter number $C_{v}$ of the group $\mathcal{G}$. These link invariants include the well-known polynomials in the knot theory literature. $\mathcal{G}$ | $R$ | Invariant ---|---|--- $SU(2)$ | $\yng(1)$ | Jones $SU(N)$ | $\yng(1)$ | HOMFLY-PT $SO(N)$ | defining | Kauffman #### 2.1.1 Link Invariants As indicated in the above table, Jones’ polynomial corresponds to the fundamental representation $R=\tiny{\yng(1)}\equiv 1\in SU(2)$ placed on all the component knots: $V_{1,1,1,\ldots 1}^{SU(2)}[\mathcal{L}_{m};{\mathbbm{q}}]\equiv J_{2,2,\ldots 2}\left[\mathcal{L}_{m};{\mathbbm{q}}=\exp({2\pi i\over k+2})\right]~{},$ (5) where the subscript ‘2’ in Jones polynomial $J_{2,2,2,\ldots}[\mathcal{L}_{m};{\mathbbm{q}}]$ denotes the dimension of $R={\tiny{\yng(1)}}$. Higher dimensional representations placed on the component knots $R_{i}=\underbrace{\tiny{\yng(4)}}_{n_{i}-1}\equiv n_{i}-1\in SU(2)$ are the colored Jones invariants: $V_{n_{1}-1,n_{2}-1,n_{3}-1,\ldots n_{m}-1}^{SU(2)}[\mathcal{L}_{m};{\mathbbm{q}}]\equiv J_{n_{1},n_{2},\ldots n_{m}}\left[\mathcal{L}_{m};{\mathbbm{q}}=\exp({2\pi i\over k+2})\right]~{},$ (6) and the invariants with these representations belonging to $SU(N)$ ($(SO(N)$) are known as colored HOMFLY-PT (colored Kauffman) invariants. For clarity, we will restrict to $SU(2)$ group to write the invariants explicitly in terms of ${\mathbbm{q}}$ variable. We work with the following unknot ($\bigcirc$) normalisation: $J_{n+1}[\bigcirc;{\mathbbm{q}}]={\rm dim}_{\mathbbm{q}}\underbrace{\yng(4)}_{n}={{\mathbbm{q}}^{(n+1)\over 2}-{\mathbbm{q}}^{-{(n+1)\over 2}}\over{\mathbbm{q}}^{1\over 2}-{\mathbbm{q}}^{-{1\over 2}}}={\sin({\pi(n+1)\over k+2})\over\sin({\pi\over k+2})}={S_{0n}\over S_{00}},$ (7) where ${\rm dim}_{\mathbbm{q}}\underbrace{\yng(3)}_{n}$ denotes quantum dimension of the representation $\underbrace{\yng(3)}_{n}$ and $S_{n_{1}n_{2}}$ are the modular transformation matrix elements of the $\mathfrak{su}(2)_{k}$ WZW conformal field theory whose action on the characters is $\chi_{n_{1}}(\tau)~{}~{}\underrightarrow{~{}~{}~{}S~{}~{}~{}}~{}~{}\chi_{n_{2}}\left(-{1\over\tau}\right)~{},$ where $\tau$ denotes the modular parameter. These knot and link polynomials with the above unknot normalisation are referred as unnormalised colored Jones invariant. For framed unknots with framing number $f$, the invariant will be $J_{n+1}[\bigcirc_{f};{\mathbbm{q}}]={\mathbbm{q}}^{f[{(n+1)^{2}-1\over 4}]}{{\mathbbm{q}}^{(n+1)\over 2}-{\mathbbm{q}}^{-{(n+1)\over 2}}\over{\mathbbm{q}}^{1\over 2}-{\mathbbm{q}}^{-{1\over 2}}}\propto(T_{nn})^{f}{S_{0n}\over S_{00}},$ (8) where the action of the modular transformation matrix $T$ on characters is $\chi_{n}(\tau)~{}~{}\underrightarrow{~{}~{}~{}T~{}~{}~{}~{}}\chi_{n}(\tau+1)~{}.$ The colored Jones invariant for the Hopf link can also be written in terms of $S$ matrix: $J_{n_{1}+1,n_{2}+1}[H;{\mathbbm{q}}]=\left({{\mathbbm{q}}^{\frac{(n_{1}+1)(n_{2}+1)}{2}}-{\mathbbm{q}}^{-\frac{(n_{1}+1)(n_{2}+1)}{2}}\over{\mathbbm{q}}^{\frac{1}{2}}-{\mathbbm{q}}^{-\frac{1}{2}}}\right)={S_{n_{1}n_{2}}\over S_{00}}.$ (9) The invariant for a framed Hopf link $H(f_{1},f_{2})$, with framing numbers $f_{1}$ and $f_{2}$ on the two component knots, in terms of $T$ and $S$ matrices is $J_{n_{1}+1,n_{2}+1}[H(f_{1},f_{2});{\mathbbm{q}}]\propto(T_{n_{1}n_{1}})^{f_{1}}(T_{n_{2}n_{2}})^{f_{2}}{S_{n_{1}n_{2}}\over S_{00}}~{}~{}.$ (10) We will look at a class of links obtained as a connected sum of framed Hopf links. For instance, the invariant for the connected sum of two framed Hopf links $H(f_{1},f_{2})\\#H(f_{2},f_{3})$ will be $\displaystyle J_{n_{1}+1,n_{2}+1,n_{3}+1}[H(f_{1},f_{2})\\#H(f_{2},f_{3});{\mathbbm{q}}]$ $\displaystyle\propto$ $\displaystyle{\prod_{i=1}^{3}T_{n_{i}n_{i}}^{f_{i}}}{S_{n_{1}n_{2}}\over S_{00}}{S_{n_{2}n_{3}}\over S_{n_{2}0}}$ $\displaystyle=$ $\displaystyle{\prod_{i=1}^{2}J_{n_{i}+1,n_{i+1}+1}[H(f_{i},f_{i+1});{\mathbbm{q}}]\over J_{n_{2}+1}[\bigcirc;{\mathbbm{q}}]}~{}.~{}$ Such a connected sum of two framed Hopf links, which is a 3-component link, can be denoted as a linear graph $f_{1}$$f_{2}$$f_{3}$ with three vertices labeled by the framing numbers and the edges connecting the adjacent vertices. These are known as ‘plumbing graphs’. Another plumbing graph $\Gamma$ with 8 vertices denoting the link $\mathcal{L}(\Gamma)$ (the connected sum of many framed Hopf links) is illustrated in Figure 1. The colored invariant for these links $\mathcal{L}(\Gamma)$ can be written in terms of $S$ and $T$ matrices. Figure 1: An example of a plumbing graph $\Gamma$ (left) and the corresponding link ${\mathcal{L}}(\Gamma)$ of framed unknots in $S^{3}$ (right). For a general $m$ vertex plumbing graph with vertices $v_{1},v_{2},\ldots v_{m}\in V$ labelled by framing numbers $f_{1},f_{2},\ldots f_{m}$, there can be one or more edges connecting a vertex $v$ with the other vertices. The degree of any vertex $v$ ($\text{deg}(v)$) is equal to the total number of edges intersecting $v$. For the graph in Figure 1, $\text{deg}(2)=\text{deg}(4)=\text{deg}(6)=3$. The colored Jones’ invariant for any plumbing graph $\Gamma$ is $J_{n_{1}+1,n_{2}+1,\ldots n_{m}+1}[\mathcal{L};{\mathbbm{q}}]\propto{1\over S_{00}}\prod_{i=1}^{m}\\{(T_{n_{i}n_{i}})^{f_{i}}(S_{0n_{i}})^{1-\text{deg}(v_{i})}\\}\prod_{(v_{1},v_{2})\;\in\;\text{Edges}}(S_{n_{v_{1}}n_{v_{2}}})~{}.$ (12) Even though we have presented the colored Jones invariants (10, 2.1.1, 12), the formal expression of these link invariants in terms of $S$ and $T$ matrices are applicable for any arbitrary gauge group $\mathcal{G}$. $\bullet$ $SO(3)$ and $OSp(1|2)$ Link invariants Using group theory arguments, it is possible to relate colored link invariants between different groups. For instance, the representations of the $SO(3)$ can be identified with a subset of $SU(2)$ representations. As a consequence, the $SO(3)$ link invariants can be related to the colored Jones invariants as follows: $V_{n_{1},n_{2},n_{3},\ldots n_{m}}^{SO(3)}\left[\mathcal{L}_{m};Q=\exp({2\pi i\over K+1})\right]=J_{2n_{1}+1,2n_{2}+1,\ldots 2n_{m}+1}[\mathcal{L}_{m};{\mathbbm{q}}]{\big{|}}_{{\mathbbm{q}}^{2}=Q}~{},$ (13) where the level $K$ of the affine $\mathfrak{so}(3)_{K}$ Lie algebra must be an even integer( $K\in 2\mathbb{Z}$). Similarly, the representations of the orthosymplectic supergroup $OSp(1|2)$ can be related to the representations of the $SU(2)$ group from the study of $\mathfrak{osp}(1|2)_{\hat{K}}$ WZW conformal field theory and the link invariants $V_{n_{1},n_{2},n_{3},\ldots n_{m}}^{OSp(1|2)}\left[\mathcal{L}_{m};\hat{Q}=\exp({2\pi i\over 2{\hat{K}}+3})\right]$ Ennes:1997kx . Particularly, there is a precise identification of the polynomial variable $\hat{Q}$ to $SU(2)$ variable ${\mathbbm{q}}$. Further, the fusion rules of the primary fields of $\mathfrak{osp}(1|2)_{\hat{K}}$ WZW conformal field theory can be compared to integer spin primary fields of the $\mathfrak{su}(2)_{k}$. Particularly, the $\hat{S}$ and $\hat{T}$-matrices of $\mathfrak{osp}(1|2)_{\hat{K}}$ : $\displaystyle\hat{S}_{n_{1}n_{2}}$ $\displaystyle=$ $\displaystyle\sqrt{4\over 2\hat{K}+3}(-1)^{n_{1}+n_{2}}\cos\left[{(2n_{1}+1)(2n_{2}+1)\over 2(2\hat{K}+3)}\pi\right]~{}~{},$ (14) $\displaystyle\hat{T}_{n_{1},n_{2}}$ $\displaystyle\propto$ $\displaystyle\delta_{n_{1},n_{2}}{\hat{Q}}^{[\frac{(2n_{1}+1)^{2}-1]}{4}}~{}~{}~{},$ (15) are related to the $S$ and $T$ matrices of $\mathfrak{su}(2)_{k}$ in the following way: $\hat{S}_{n_{1},n_{2}}=S_{2n_{1},2n_{2}}{\big{|}}_{{\mathbbm{q}}=-\hat{Q}}~{};~{}\hat{T}_{n_{1},n_{1}}=T_{2n_{1},2n_{1}}{\big{|}}_{{\mathbbm{q}}=-\hat{Q}}$ (16) Using these relations, we can show that the $OSp(1|2)$ colored invariant match the colored Jones invariant for any arbitrary link $\mathcal{L}_{m}$ in the following way: $V_{n_{1},n_{2},n_{3},\ldots n_{m}}^{OSp(1|2)}\left[\mathcal{L}_{m};\hat{Q}=\exp({2\pi i\over 2{\hat{K}}+3})\right]=\epsilon J_{2n_{1}+1,2n_{2}+1,\ldots 2n_{m}+1}[\mathcal{L}_{m};{\mathbbm{q}}]{\big{|}}_{{\mathbbm{q}}=-\hat{Q}}~{},$ (17) where $\epsilon$ could be $\pm 1$ depending on the link $\mathcal{L}$ and the representations $n_{i}$’s. For example, the colored $OSp(1|2)$ invariant for framed Hopf link is $\displaystyle V_{n_{1},n_{2}}^{OSp(1|2)}[H(f_{1},f_{2});\hat{Q}]$ $\displaystyle=$ $\displaystyle{\hat{Q}}^{\frac{f_{1}((2n_{1}+1)^{2}-1)}{4}}{\hat{Q}}^{\frac{f_{2}((2n_{1}+1)^{2}-1)}{4}}(-1)^{(n_{1}+n_{2})}\times$ $\displaystyle\left({{\hat{Q}}^{\frac{(2n_{1}+1)(2n_{2}+1)}{2}}+{\hat{Q}}^{-\frac{(2n_{1}+1)(2n_{2}+1)}{2}}\over{\hat{Q}}^{\frac{1}{2}}+{\hat{Q}}^{-\frac{1}{2}}}\right)$ $\displaystyle=$ $\displaystyle J_{2n_{1}+1,2n_{2}+1}[H(f_{1},f_{2}),-\hat{Q}]~{}.$ In fact, for any link $\mathcal{L}(\Gamma)$ denoted by the graph $\Gamma$, the invariants will be $\displaystyle V_{n_{1},n_{2},\ldots n_{m}}^{OSp(1|2)}[\mathcal{L}(\Gamma);\hat{Q}]$ $\displaystyle=$ $\displaystyle{1\over{\hat{Q}}^{\frac{1}{2}}+{\hat{Q}}^{-\frac{1}{2}}}\prod_{i=1}^{m}(-1)^{n_{i}}{\hat{Q}}^{\frac{f_{i}((2n_{i}+1)^{2}-1)}{4}}$ $\displaystyle\left({\hat{Q}}^{\frac{2n_{i}+1}{2}}+{\hat{Q}}^{-{\frac{2n_{i}+1}{2}}}\right)^{\text{deg}(v_{i})-1}$ $\displaystyle\prod_{(v_{1},v_{2})\;\in\;{\rm Edges}}{\left({\hat{Q}}^{\frac{(2n_{v_{1}}+1)(2n_{v_{2}}+1)}{2}}+{\hat{Q}}^{-\frac{(2n_{v_{1}}+1)(2n_{v_{2}}+1)}{2}}\right)}~{}.$ (19) As three-manifolds can be constructed by a surgery procedure on any framed link, the Chern-Simons partition function/WRT invariant (1) can be written in terms of link invariants10.2307/1970373 ; wallace_1960 ; Kaul:2000xe ; Ramadevi:1999nd . We will now present the salient features of such WRT invariants. #### 2.1.2 Three-Manifold Invariants Let us confine to the three-manifold $M[\Gamma]$ obtained from surgery of framed link associated with $L$-vertex graph (an example illustrated in Figure 1). These kind of manifolds are known in the literature as plumbed three- manifolds. The linking matrix $B$ is defined as $B_{v_{1},v_{2}}=\left\\{\begin{array}[]{ll}1,&v_{1},v_{2}\text{ connected},\\\ f_{v},&v_{1}=v_{2}=v,\\\ 0,&\text{otherwise}.\end{array}\right.\qquad v_{i}\in\text{Vertices of }\Gamma\;\cong\;\\{1,\ldots,L\\}.$ (20) The algebraic expression for the WRT invariant $\tau_{k}^{\mathcal{G}}[M(\Gamma);{\mathbbm{q}}]$ is $\tau_{k}^{\mathcal{G}}[M(\Gamma);{\mathbbm{q}}]=\frac{F^{\mathcal{G}}[{\mathcal{L}}(\Gamma);{\mathbbm{q}}]}{F^{\mathcal{G}}[{\mathcal{L}}(+1\bullet);{\mathbbm{q}}]^{b_{+}}F^{\mathcal{G}}[{\mathcal{L}}(-1\bullet);{\mathbbm{q}}]^{b_{-}}}$ (21) where $b_{\pm}$ are the number of positive and negative eigenvalues of a linking matrix $B$ respectively and $F^{\mathcal{G}}[\mathcal{L}(\Gamma);{\mathbbm{q}}]$ is defined as $F^{\mathcal{G}}[{\mathcal{L}}(\Gamma);{\mathbbm{q}}]=\sum_{R_{1},R_{2},\ldots R_{L}}\left(\prod_{i=1}^{L}V_{R_{i}}^{\mathcal{G}}[\bigcirc;{\mathbbm{q}}]\right)V_{R_{1},R_{2}\ldots R_{L}}^{\mathcal{G}}[\mathcal{L}(\Gamma);{\mathbbm{q}}]~{},$ (22) where the summation indicates all the allowed integrable representations of affine $\mathfrak{g}_{k}$ Lie algebra. By construction, any two homeomorphic manifolds must share the same three-manifold invariant. There is a prescribed set of moves called Kirby moves on links which gives the same three-manifold. For framed links depicted as plumbing graphs, these moves are known as Kirby- Neumann moves as shown in Figure 2. Hence, the three-manifold invariant must obey $\tau_{k}^{\mathcal{G}}[M(\Gamma);{\mathbbm{q}}]=\tau_{k}^{\mathcal{G}}[M(\Gamma^{\prime});{\mathbbm{q}}]~{},$ (23) where the plumbing graphs $\Gamma,\Gamma^{\prime}$ are related by the Kirby- Neumann moves. Figure 2: Kirby-Neumann moves that relate plumbing graphs which result in homeomorphic 3-manifolds. Towards the end of 20th century, attempts to give a topological interpretation for the integer coefficients in the Laurent series expression for Jones polynomial (HOMFLY-PT) as well as the corresponding colored invariants for any knot $\mathcal{K}$ $J_{n}[\mathcal{K};{\mathbbm{q}}]=\sum_{s}a_{s}{\mathbbm{q}}^{s}~{},~{}\\{a_{s}\\}\in\mathbb{Z}$ (24) has resulted in developments on homology theories as well as physics explanation. We will discuss these ‘homological invariants’ and their appearance in string/M-theory in the following section. ### 2.2 Knot, Link and Three-manifold Homologies We will first review the developments on homological invariants of knots accounting for these integers $a_{s}$ (24) as dimension of the vector space $\mathcal{H}_{\mathcal{K}}$ of a homological theory. Then, we will present the topological string/M-theory approach where these integers count number of BPS states. #### 2.2.1 Homological Invariants of Knots The pioneering work of Khovanovkhovanov2000categorification on bi-graded homology theory led to categorification of the Jones polynomial. This was extended to colored $\mathfrak{sl2}$ knot homology $\mathcal{H}^{\mathfrak{sl}_{2};n}_{i,j}$ webster2017knot ; cooper2012categorification ; frenkel2012categorifying leading to new homological invariants $\mathcal{P}_{n}^{\mathfrak{sl}_{2}}[\mathcal{K},q,t]$ which categorifies the colored Jones polynomial: $\mathcal{P}_{n}^{\mathfrak{sl}_{2}}[\mathcal{K},{\mathbbm{q}},t]=\sum_{i,j}t^{j}{\mathbbm{q}}^{i}{\rm dim}\mathcal{H}^{\mathfrak{sl}_{2};n}_{i,j}~{}.$ (25) The subscripts $i$ and $j$ on the colored $\mathfrak{sl}_{2}$ homology $\mathcal{H}^{\mathfrak{sl}_{2};n}_{i,j}$ are called the polynomial grading and the homological grading respectively. In fact, the ${\mathbbm{q}}$-graded Euler characteristic of the colored $\mathfrak{sl}_{2}$ knot homology gives the colored Jones invariant: $J_{n}[\mathcal{K};{\mathbbm{q}}]=\sum_{i,j}(-1)^{j}{\mathbbm{q}}^{i}{\rm dim}\mathcal{H}^{\mathfrak{sl}_{2};n}_{i,j}~{},$ (26) explaining the reasons behind the integers $a_{s}$(24). Khovanov and Rozansky khovanov2004matrix constructed $\mathfrak{sl}_{N}$ homology using matrix factorizations. This led to the categorification of colored HOMFLY-PT polynomials of knots. There has been interesting insight on these homological invariants within topological strings context and $M$-theory. We will now discuss the essential features from physics approach. #### 2.2.2 Topological Strings and M-theory The parallel developments from topological strings and intersecting branes in $M$-theory Ooguri:1999bv ; Gopakumar:1998jq interpreted the integers of unnormalised HOMFLY-PT (24) as counting of BPS states. Invoking topological string duality in the presence of any knot $\mathcal{K}$, Ooguri-Vafa conjectured $V_{\tiny\yng(1)}^{SU(N)}[\mathcal{K};{\mathbbm{q}},\lambda={\mathbbm{q}}^{N}]={1\over({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})}\sum_{Q,s}N_{\tiny\yng(1),Q,s}\lambda^{Q}{\mathbbm{q}}^{s}~{},$ (27) where the integers $N_{\tiny\yng(1),Q,s}$ count $D4-D2$ bound states in string theory. Further the relation between the BPS spectrum and $sl_{N}$/Khovanov- Rozansky knot homology was conjectured within the topological string contextGukov:2004hz : $N_{\tiny\yng(1),Q,s}=\sum_{j}(-1)^{j}D_{Q,s,j}~{},$ (28) where the integers $D_{Q,s,j}$ are referred to as refined BPS invariants. The extra charge/ homological grading $j$ are explainable by the appearance of extra $U(1)$ symmetry in $M$-theory compactified on Calabi-Yau three-folds $CY_{3}$. The topological string duality and the dualities of physical string theories compactified on $CY_{3}$ implies that the vector space of knot homologies are the Hilbert space of BPS states(see review Nawata:2015wya and references therein): $\mathcal{H}_{\mathcal{K}}\equiv\mathcal{H}_{BPS}~{}.$ The impact of knot homology on the categorification of the WRT invariants has been studied in the last six years. We now present a concise summary of the recent developments in this direction. #### 2.2.3 Three-Manifold Homology As WRT invariants (21) of three-manifolds involves invariants of links, logically we would expect the homology of three-manifold $\mathcal{H}^{\mathcal{G};M}$ such that $\tau_{k}^{\mathcal{G}}[M;{\mathbbm{q}}]\stackrel{{\scriptstyle\text{?}}}{{=}}\sum_{i,j}(-1)^{j}{\mathbbm{q}}^{i}{\rm dim}\mathcal{H}^{\mathcal{G};M}_{i,j}~{}.$ (29) However, the WRT invariants known for many three-manifolds are not seen as ${\mathbbm{q}}$-series (29). We will now review the necessary steps Gukov:2017kmk of obtaining a new three-manifold invariant $\hat{Z}$, as ${\mathbbm{q}}$-series, from $U(N)$ Chern-Simons partition function for Lens space $M=L(p,1)\equiv S^{3}/\mathbb{Z}_{p}$. The space of flat connections $\\{a\\}$ denoted by $\pi_{1}[{S^{3}\over\mathbb{Z}_{p}}]\equiv\mathbb{Z}_{p}$. Hence $Z_{k}^{U(N)}[L[p,1];{\mathbbm{q}}]$ can be decomposed as a sum of perturbative Chern-Simons $Z_{a}^{U(N)}[L[p,1];{\mathbbm{q}}]$ around these abelian flat connections $a$ Gukov:2017kmk : $Z_{k}^{U(N)}[L[p,1];{\mathbbm{q}}]=\sum_{a}\exp[iS_{CS}^{(a)}]Z_{a}^{U(N)}[L[p,1];{\mathbbm{q}}]~{},$ (30) where $S_{CS}^{(a)}$ is the corresponding classical Chern-Simons action. The following change of basis by $\mathcal{S}$ matrix of $\mathfrak{u}(1)^{N}_{p}$ affine algebra: $Z_{a}^{U(N)}[L(p,1);{\mathbbm{q}}]=\sum_{b}\mathcal{S}_{ab}\hat{Z}_{b}^{U(N)}[\mathcal{L}[p,1];q]\Big{|}_{q\rightarrow{\mathbbm{q}}}~{},$ (31) is required so that $\hat{Z}_{b}^{U(N)}[\mathcal{L}[p,1];q]\in q^{\Delta_{b}}\mathbb{Z}[[q]]~{},~{}~{}\Delta_{b}\in\mathbb{Q}~{}.$ (32) Physically, the $\hat{Z}_{b}[\mathcal{L}[p,1];q]$ is also the vortex partition function $\hat{Z}_{T[L[p,1]]}[D^{2}\times_{q}S^{1}]$ obtained from reducing 6d $(2,0)$ theory (describing dynamics of $N$-coincident $M5$ branes on $L[p,1]\times D^{2}\times_{q}S^{1}$) on $L[p,1]$. The effective 3-d $\mathcal{N}=2$ theory on $D^{2}\times_{q}S^{1}$ (cigar geometry) is denoted as $T^{U(N)}[L[p,1]]$. For other three-manifolds $M$, $\mathcal{S}$ matrix depends only on $H_{1}(M,\mathbb{Z})$. Further the Hilbert space of BPS states $\mathcal{H}^{i,j}_{BPS}$ on the M5 brane system, in the ambient space-time $T^{*}M\times TN\times S^{1}$, where $i,j$ gradings will keep track of both spins associated with the rotational symmetry $U(1)_{q}\times U(1)_{R}$ on $D^{2}\subset TN$. The Hilbert space of states for the theory $T^{\mathcal{G}}[M]$ with boundary condition at $\partial D^{2}=S^{1}$ labeled by $a\in({\rm Tor}H_{1}(M,\mathbb{Z}))^{N}/S_{N}$ leads to bi-graded homological invariants of $M$: ${\mathcal{H}}_{a}^{U(N)}[M]={\mathcal{H}}_{T_{L[p,1]}^{U(N)}}[D^{2};a]=\bigoplus_{\begin{subarray}{c}i\in\mathbb{Z}+\Delta_{a},\\\ j\in\mathbb{Z}\end{subarray}}{\mathcal{H}}_{a}^{i,j}~{}.$ (33) Note that the grading $i$ counts the charge under $U(1)_{q}$ rotation of $D^{2}$ and homological grading $j$ is the R-charge of the $U(1)_{R}$ R-symmetry. In the following section, we will review the necessary steps of obtaining $\hat{Z}$ invariants for $SU(2)$ group. This will provide clarity of notations to investigate $\hat{Z}$ for $SO(3)$ and $OSp(1|2)$ group. ## 3 Review of $SU(2)$ $\hat{Z}$ invariant As discussed in subsection 2.2.3Gukov:2016gkn , the expression for Lens space partition function using eqns.(30-32) $\displaystyle Z_{k}^{U(N)}[L(p,1),{\mathbbm{q}}]$ $\displaystyle=$ $\displaystyle\sum_{a,b\in\mathbb{Z}_{p}}\mathcal{S}_{ab}\exp[iS_{CS}^{(a)}]\hat{Z}_{b}^{U(N)}[\mathcal{L}(p,1);q]\Big{|}_{q\rightarrow{\mathbbm{q}}}~{},$ (34) $\displaystyle{\rm where}~{}\hat{Z}_{b}[\mathcal{L}[p,1];q]$ $\displaystyle\in$ $\displaystyle q^{\Delta_{a}}\mathbb{Z}[[q]]~{},~{}~{}\Delta_{a}\in\mathbb{Q}~{}.$ (35) led to the following conjecture Gukov:2017kmk ; Gukov:2019mnk for any closed oriented three manifold $M$ known as GPPV conjecture: $\displaystyle Z^{SU(2)}_{k}[M;{\mathbbm{q}}]$ $\displaystyle=$ $\displaystyle(i\sqrt{2(k+2}))^{b_{1}(M)-1}\sum_{a,b\;\in\;\atop\text{Spin}^{c}(M)/{\mathbb{Z}}_{2}}\exp[2\pi i(k+2){\ell k}(a,a)]\,\times$ $\displaystyle~{}~{}~{}~{}~{}~{}~{}|\mathcal{W}_{b}|^{-1}\mathcal{S}_{ab}\hat{Z}^{SU(2)}_{b}[M;q]|_{q\rightarrow{\mathbbm{q}}=\exp({\frac{2\pi i}{k+2}})}$ (36) where $\hat{Z}^{SU(2)}_{b}[M;q]\in\,2^{-c}q^{\Delta_{b}}{\mathbb{Z}}[[q]]\qquad\Delta_{b}\in{\mathbb{Q}},\qquad c\in{\mathbb{Z}}_{+}$ (37) is convergent for $|q|<1$ and $\mathcal{S}_{ab}=\frac{e^{2\pi i{\ell k}(a,b)}+e^{-2\pi i{\ell k}(a,b)}}{|{\mathcal{W}}_{a}|\sqrt{|H_{1}(M,{\mathbb{Z}})|}}.$ (38) Here $\mathcal{W}_{a}$ is the stabilizer subgroup defined as ${\mathcal{W}}_{a}\;\equiv\;\text{Stab}_{{\mathbb{Z}}_{2}}(a)\;=\;\left\\{\begin{array}[]{cl}{\mathbb{Z}}_{2},&a=-a\,,\\\ 1,&\text{otherwise\,.}\end{array}\right.$ (39) and $\ell k$ denotes the linking pairing on $H_{1}(M,\mathbb{Z})$: $\begin{array}[]{cccc}{\ell k}:&H_{1}(M,\mathbb{Z})\otimes H_{1}(M,\mathbb{Z})&\longrightarrow&{\mathbb{Q}}/{\mathbb{Z}}\\\ &[a]\otimes[b]&\longmapsto&{\\#(a\cap\hat{b})}/{n}\\\ \end{array}$ (40) where $\hat{b}$ is a two-chain complex such that $\partial\hat{b}=nb$ with $n\in{\mathbb{Z}}$. Such a $\hat{b}$ and $n$ exists because $[b]\in H_{1}(M,\mathbb{Z})$. The number $\\#(a\cap\hat{b})$ counts the intersection points with signs determined by the orientation. The set of orbits is the set of $\text{Spin}^{c}$ structures on $M$, with the action of ${\mathbb{Z}}_{2}$ by conjugation. Although, the relation(36) is true for any closed oriented three-manifold $M$, the explicit $q$ series expression for $\hat{Z}$ is waiting to be discovered for a general three-manifold. In the following subsection, we will review the $\hat{Z}^{SU(2)}$ for the plumbed manifolds. We begin with the WRT invariant for a plumbing graph, of the type shown in Figure. 1, discussed in section.(2.1.2). Then analytically continue ${\mathbbm{q}}\rightarrow q$ to get the $\hat{Z}^{SU(2)}$-invariant . We will see that the analytic continuation procedure is doable only for negative definite plumbed manifolds(i.e., the signature of linking matrix $B$, $\sigma=b_{+}-b_{-}=-L$)111In principle, this procedure is also doable when $B$ is negative on a certain subspace of $\mathbb{Z}^{L}$.. Moreover, as explained inGukov:2019mnk , the $\text{Spin}^{c}$-structure in case of plumbed 3-manifold with $b_{1}(M)=0$, is given by $H_{1}(M,\mathbb{Z})\cong\text{Coker}B=\mathbb{Z}^{L}/B\mathbb{Z}^{L}$. ### 3.1 $\hat{Z}_{b}^{SU(2)(q)}$ The WRT invariant $\tau_{k}^{SU(2)}[M(\Gamma);{\mathbbm{q}}]$,222normalized such that $\tau_{k}^{\mathcal{G}}[S^{3};{\mathbbm{q}}]=1$ and $k$ is the bare level for $SU(2)$ Chern-Simons theory for plumbed three-manifold $M(\Gamma)$(21), obtained from surgery of framed link ${\mathcal{L}}(\Gamma)$ in $S^{3}$, is $\displaystyle\tau_{k}^{SU(2)}[M(\Gamma);{\mathbbm{q}}]$ $\displaystyle=$ $\displaystyle\frac{F^{SU(2)}[{\mathcal{L}}(\Gamma);{\mathbbm{q}}]}{F^{SU(2)}[{\mathcal{L}}(+1\bullet);{\mathbbm{q}}]^{b_{+}}F^{SU(2)}[{\mathcal{L}}(-1\bullet);{\mathbbm{q}}]^{b_{-}}}$ $\displaystyle{\rm where}~{}F^{SU(2)}[{\mathcal{L}}(\Gamma);{\mathbbm{q}}]$ $\displaystyle=$ $\displaystyle\sum_{{n}\in\\{1,\ldots,k+1\\}^{L}}J[{\mathcal{L}}(\Gamma)]_{n_{1},\ldots,n_{L}}\prod_{v=1}^{L}\frac{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}{{\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2}}.$ (41) Note $b_{\pm}$ are the number of positive and negative eigenvalues of a linking matrix $B$ respectively and the colored Jones polynomial of link $\mathcal{L}(\Gamma)$ (12) in variable ${\mathbbm{q}}=\exp({2i\pi/(k+2)})$ is $\displaystyle J[{\mathcal{L}}(\Gamma)]_{n_{1},\ldots,n_{L}}$ $\displaystyle=$ $\displaystyle\frac{2i}{{\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2}}\prod_{v\;\in\;\text{Vertices}\;\cong\;\\{1,\ldots,L\\}}{\mathbbm{q}}^{\frac{f_{v}(n_{v}^{2}-1)}{4}}\,\times$ (42) $\displaystyle\left(\frac{2i}{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}\right)^{\text{deg}(v)-1}\prod_{(v_{1},v_{2})\;\in\;\text{Edges}}\frac{{\mathbbm{q}}^{n_{v_{1}}n_{v_{2}}/2}-{\mathbbm{q}}^{-n_{v_{1}}n_{v_{2}}/2}}{2i}.$ Using the following Gauss sum reciprocity formula $\sum_{n\;\in\;{\mathbb{Z}}^{L}/2k{\mathbb{Z}}^{L}}\exp\left(\frac{\pi i}{2k}(n,Bn)+\frac{\pi i}{k}(\ell,n)\right)=\\\ \frac{e^{\frac{\pi i\sigma}{4}}\,(2k)^{L/2}}{|\det B|^{1/2}}\sum_{a\;\in\;{\mathbb{Z}}^{L}/B{\mathbb{Z}}^{L}}\exp\left(-2\pi ik\left(a+\frac{\ell}{2k},B^{-1}\left(a+\frac{\ell}{2k}\right)\right)\right)$ (43) where $\ell\in{\mathbb{Z}}^{L}$, $(\cdot,\cdot)$ is the standard pairing on ${\mathbb{Z}}^{L}$ and $\sigma=b_{+}-b_{-}$ is the signature of the linking matrix $B$, we can sum $F^{SU(2)}[{\mathcal{L}}(\pm 1\bullet);{\mathbbm{q}}]=\sum_{n=1}^{k+1}{\mathbbm{q}}^{\pm\frac{n^{2}-1}{4}}\,\left(\frac{{\mathbbm{q}}^{n/2}-{\mathbbm{q}}^{-n/2}}{{\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2}}\right)^{2}=\frac{[2(k+2)]^{1/2}\,e^{\pm\frac{\pi i}{4}}\,{\mathbbm{q}}^{\mp\frac{3}{4}}}{{\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2}},$ (44) for the unknot with framing $\pm 1$. Incorporating the above equation and the fact that $L-|\text{Edges}|=1$ for the framed link ${\mathcal{L}}(\Gamma)$, the WRT invariant simplifies to $\tau_{k}^{SU(2)}[M(\Gamma);{\mathbbm{q}}]=\frac{e^{-\frac{\pi i\sigma}{4}}\,{\mathbbm{q}}^{\frac{3\sigma}{4}}}{2\,(2(k+2))^{L/2}\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})}\times\\\ {\sum_{{n}\in{\mathbb{Z}}^{L}/2(k+2){\mathbb{Z}}^{L}}}^{\prime}\prod_{v\;\in\;\text{Vertices}}{\mathbbm{q}}^{\frac{f_{v}(n_{v}^{2}-1)}{4}}\,\left(\frac{1}{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}\right)^{\text{deg}(v)-2}\times\\\ \prod_{(v^{\prime},v^{\prime\prime})\;\in\;\text{Edges}}\frac{{\mathbbm{q}}^{n_{v^{\prime}}n_{v^{\prime\prime}}/2}-{\mathbbm{q}}^{-n_{v^{\prime}}n_{v^{\prime\prime}}/2}}{2}$ (45) where we used invariance of the summand under $n_{v}\rightarrow-n_{v}$. The prime ′ in the sum means that the singular values $n_{v}=0,\,k+2$ are omitted. Let us focus on the following factor for general plumbed graph: $\displaystyle\prod_{(v^{\prime},v^{\prime\prime})\;\in\;\text{Edges}}({\mathbbm{q}}^{n_{v^{\prime}}n_{v^{\prime\prime}}/2}-{\mathbbm{q}}^{-n_{v^{\prime}}n_{v^{\prime\prime}}/2})$ $\displaystyle=$ $\displaystyle\sum_{{p}\in\\{\pm 1\\}^{\text{Edges}}}\prod_{(v^{\prime},v^{\prime\prime})\;\in\;\text{Edges}}p_{(v^{\prime},v^{\prime\prime})}$ $\displaystyle{\mathbbm{q}}^{p_{(v^{\prime},v^{\prime\prime})}n_{v^{\prime}}n_{v^{\prime\prime}}/2}.$ (46) Note that, under $n_{v}\rightarrow-n_{v}$ on any vertex $v$ of degree ${\rm deg}(v)$, the factor with a given configuration of signs associated to edges (i.e., $p\in\\{\pm 1\\}^{\text{Edges}}$) will transform into a term with a different configuration times $(-1)^{\text{deg}(v)}$. For the class of graphs $\Gamma$ (like Figure. 1), the sequence of such transforms can be finally brought to the configuration with all signs $+1$. Hence, the WRT invariant (45) for these plumbed three-manifolds can be reduced to this form: $\tau_{k}^{SU(2)}[M(\Gamma)]=\frac{e^{-\frac{\pi i\sigma}{4}}\,{\mathbbm{q}}^{\frac{3\sigma-\sum_{v}f_{v}}{4}}}{2\,(2(k+2))^{L/2}\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})}\times\\\ {\sum_{{n}\in{\mathbb{Z}}^{L}/2(k+2){\mathbb{Z}}^{L}}}^{\prime}\;\;{\mathbbm{q}}^{\frac{(n,Bn)}{4}}\prod_{v\;\in\;\text{Vertices}}\,\left(\frac{1}{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}\right)^{\text{deg}(v)-2}.$ (47) In the above expression, the points $0$ and $k+2$ are excluded in the summation but in the reciprocity formula (43) no point is excluded. So, to apply the reciprocity formula we have to first regularize the sum. This is achieved by introducing the following regularising parameters: $\displaystyle\Delta_{v}\in{\mathbb{Z}}_{+}:\;\Delta_{v}$ $\displaystyle=$ $\displaystyle\text{deg}(v)-1\mod 2,\qquad\forall v\;\in\;\text{Vertices},$ (48) $\displaystyle\omega\in{\mathbb{C}}:$ $\displaystyle 0<|\omega|<1.$ so that the sum in eqn.(47) is rewritable as $\omega\rightarrow 1$: ${\sum_{{n}\in{\mathbb{Z}}^{L}/2(k+2){\mathbb{Z}}^{L}}}^{\prime}\;\;{\mathbbm{q}}^{\frac{(n,Bn)}{4}}\prod_{v\;\in\;\text{Vertices}}\,\left(\frac{1}{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}\right)^{\text{deg}(v)-2}=\\\ \lim_{\omega\rightarrow 1}\frac{1}{2^{L}}\sum_{{n}\in{\mathbb{Z}}^{L}/2(k+2){\mathbb{Z}}^{L}}{\mathbbm{q}}^{\frac{(n,Bn)}{4}}F_{\omega}(x_{1},\ldots,x_{L})|_{x_{v}={\mathbbm{q}}^{n_{v}/2}},$ (49) where $\displaystyle F_{\omega}(x_{1},\ldots,x_{L})$ $\displaystyle=$ $\displaystyle\prod_{v\;\in\;\text{Vertices}}\left({x_{v}-1/x_{v}}\right)^{\Delta_{v}}\times\,$ (50) $\displaystyle\left\\{\left(\frac{1}{x_{v}-\omega/x_{v}}\right)^{\text{deg}(v)-2+\Delta_{v}}+\left(\frac{1}{\omega x_{v}-1/x_{v}}\right)^{\text{deg}(v)-2+\Delta_{v}}\right\\}$ Note that, we can perform a binomial expansion taking $(\omega/x_{v}^{2})$ small in the first term and $(\omega x_{v}^{2})$ small in the second term to rewrite $F_{\omega}(x_{1},\ldots,x_{L})$ as a formal power series: $F_{\omega}(x_{1},\ldots,x_{L})=\sum_{\ell\in 2{\mathbb{Z}}^{L}+\delta}F_{\omega}^{\ell}\prod_{v}x_{v}^{\ell_{v}}\qquad\in{\mathbb{Z}}[\omega][[x_{1}^{\pm 1},\ldots,x_{1}^{\pm L}]]~{},$ (51) where $\delta\in{\mathbb{Z}}^{L}/2{\mathbb{Z}}^{L},~{}\delta_{v}\equiv\text{deg}(v)\mod 2$ and $F_{\omega}^{\ell}=\sum_{m:\,\ell\in{\mathcal{I}}_{m}}N_{m,\ell}\,\omega^{m}\;\in{\mathbb{Z}}[\omega]$ (52) with ${\mathcal{I}}_{m}$ being a finite set of elements from ${\mathbb{Z}}^{L}$. By definition, ${\rm lim}_{\omega\rightarrow 1}F_{\omega}^{\ell}$ is not dependent on $\Delta\in{\mathbb{Z}}^{L}$ (48). However this $\omega\rightarrow 1$ limit in eqn. (3.1) will restrict the binomial expansion range of the first term to be $x\rightarrow\infty$ and that of the second term to $x\rightarrow 0$: $\displaystyle F_{\omega\rightarrow 1}(x_{1},\ldots,x_{L})=\sum_{\ell\in 2{\mathbb{Z}}^{L}+\delta}F_{\omega\rightarrow 1}^{\ell}\prod_{v}x_{v}^{\ell_{v}}$ $\displaystyle=$ (53) $\displaystyle~{}~{}{\rm lim}_{\omega\rightarrow 1}\prod_{v\,\in\,\text{Vertices}}\left\\{{\scriptsize\begin{array}[]{c}\text{Expansion}\\\ \text{as}x\rightarrow\infty\end{array}}\frac{1}{(x_{v}-\omega/x_{v})^{\text{deg}\,v-2}}\right.$ $\displaystyle+$ $\displaystyle\left.{\scriptsize\begin{array}[]{c}\text{Expansion}\\\ \text{as }x\rightarrow 0\end{array}}\frac{1}{(\omega x_{v}-1/x_{v})^{\text{deg}\,v-2}}\right\\}.$ (58) Now let us assume that the quadratic form $B:{\mathbb{Z}}^{L}\times{\mathbb{Z}}^{L}\rightarrow{\mathbb{Z}}$ is negative definite i.e., $\sigma=-L$. Then we can define the following series in $q$ which is convergent for $|q|<1$: $\hat{Z}_{b}^{SU(2)}[M(\Gamma);q]\stackrel{{\scriptstyle\text{Def}}}{{=\joinrel=}}2^{-L}q^{-\frac{3L+\sum_{v}f_{v}}{4}}\sum_{\ell\in 2B{\mathbb{Z}}^{L}+b}F^{\ell}_{\omega\rightarrow 1}\,q^{-\frac{(\ell,B^{-1}\ell)}{4}}\;\in\;2^{-c}q^{\Delta_{b}}{\mathbb{Z}}[[q]]$ (59) where $c\in{\mathbb{Z}}_{+},c\leq L$ and $\displaystyle b$ $\displaystyle\in$ $\displaystyle(2{\mathbb{Z}}^{L}+\delta)/2B{\mathbb{Z}}^{L}\,/{\mathbb{Z}}_{2}\cong(2\text{Coker}\,B+\delta)\,/{\mathbb{Z}}_{2}\stackrel{{\scriptstyle\text{Set}}}{{\cong}}H_{1}(M_{3},{\mathbb{Z}})\,/{\mathbb{Z}}_{2},$ (60) $\displaystyle\Delta_{b}$ $\displaystyle=$ $\displaystyle-\frac{3L+\sum_{v}f_{v}}{4}-\max_{\ell\in 2M{\mathbb{Z}}^{L}+b}\frac{(\ell,B^{-1}\ell)}{4}\,\in{\mathbb{Q}}$ (61) where ${\mathbb{Z}}_{2}$ action takes $b\rightarrow-b$ and is the symmetry of (59). Using relation (49) and applying Gauss reciprocity formula (43) we arrive at the following expression for the WRT invariant: $\tau_{k}^{SU(2)}[M(\Gamma);{\mathbbm{q}}]=\\\ \\\ ~{}~{}~{}~{}~{}\frac{e^{-\frac{\pi iL}{4}}\,{\mathbbm{q}}^{-\frac{3L+\sum_{v}f_{v}}{4}}}{2\,(2(k+2))^{L/2}\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})}\,\lim_{\omega\rightarrow 1}\sum_{{n}\in{\mathbb{Z}}^{L}/2(k+2){\mathbb{Z}}^{L}}{\mathbbm{q}}^{\frac{(n,Bn)}{4}}F_{\omega}(x_{1},\ldots,x_{L})|_{x_{v}={\mathbbm{q}}^{n_{v}/2}}=\\\ \\\ ~{}~{}\frac{2^{-L}{\mathbbm{q}}^{-\frac{3L+\sum_{v}f_{v}}{4}}}{2\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})\,|\det B|^{1/2}}\sum_{\scriptsize\begin{array}[]{c}a\in\mathrm{Coker}\,B\\\ b\in 2\mathrm{Coker}\,B+\delta\end{array}}e^{-2\pi i(a,B^{-1}b)}e^{-2\pi i(k+2)(a,B^{-1}a)}\times\\\ \lim_{\omega\rightarrow 1}\sum_{\ell\in 2B{\mathbb{Z}}^{L}+b}F^{\ell}_{\omega}\,{\mathbbm{q}}^{-\frac{(\ell,B^{-1}\ell)}{4}}.$ (62) Assuming that the limit $\lim_{q\rightarrow{\mathbbm{q}}}\hat{Z}_{b}^{SU(2)}(q)$ exists, where $q$ approaches $(k+2)$-th primitive root of unity from inside of the unit disc $|q|<1$, we expect $\lim_{\omega\rightarrow 1}\sum_{\ell\in 2B{\mathbb{Z}}^{L}+b}F^{\ell}_{\omega}\,{\mathbbm{q}}^{-\frac{(\ell,B^{-1}\ell)}{4}}=\lim_{q\rightarrow{\mathbbm{q}}}\sum_{\ell\in 2B{\mathbb{Z}}^{L}+b}F^{\ell}_{\omega\rightarrow 1}\,q^{-\frac{(\ell,B^{-1}\ell)}{4}}.$ (63) Thus we obtain GPPV conjecture form: $\tau_{k}^{SU(2)}[M(\Gamma),{\mathbbm{q}}]=\frac{1}{2\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})\,|\det B|^{1/2}}\,\times\\\ \sum_{a\in\mathrm{Coker}\,B}e^{-2\pi i(k+2)(a,B^{-1}a)}\sum_{b\in 2\mathrm{Coker}\,B+\delta}e^{-2\pi i(a,B^{-1}b)}\lim_{q\rightarrow{\mathbbm{q}}}\hat{Z}_{b}^{SU(2)}[M(\Gamma);q].$ (64) There is also an equivalent contour integral form for the homological blocks(59): $\hat{Z}_{b}^{SU(2)}[M(\Gamma);q]=q^{-\frac{3L+\sum_{v}f_{v}}{4}}\cdot\text{v.p.}\int\limits_{|z_{v}|=1}\prod_{v\;\in\;\text{Vertices}}\frac{dz_{v}}{2\pi iz_{v}}\,\left({z_{v}-1/z_{v}}\right)^{2-\text{deg}(v)}\cdot\Theta^{-B}_{b}(z),$ (65) where $\Theta^{-B}_{b}(x)$ is the theta function of the lattice corresponding to minus the linking form $B$: $\Theta^{-B}_{b}(x)=\sum_{\ell\in 2B{\mathbb{Z}}^{L}+b}q^{-\frac{(\ell,B^{-1}\ell)}{4}}\prod_{i=1}^{L}x_{i}^{\ell_{i}},$ (66) and “v.p.” refers to principle value integral (i.e. take half-sum of contours $|z_{v}|=1\pm\epsilon$). This prescription corresponds to the regularization by $\omega$ made in eqn.(3.1). Thus we can obtain explicit $SU(2)$ $q$-series for any negative definite plumbed three-manifolds. For completeness, we present the $q$-series for some examples. ### 3.2 Examples $\bullet$ Poincare homology sphere is a well-studied three-manifold corresponding to the graph: (67) As $H_{1}(M,{\mathbb{Z}})=0$, we obtain only single homological block $\hat{Z}_{b_{1}}$. Solving eqns.(58,59), we get $\hat{Z}_{b_{1}}^{SU(2)}=q^{-3/2}(1-q-q^{3}-q^{7}+q^{8}+q^{14}+q^{20}+q^{29}-q^{31}-q^{42}+\cdots).$ (68) $\bullet$ The next familiar example with $H_{1}(M,{\mathbb{Z}})=0$ is Brieskorn homology sphere. A particular example of this class is $\Sigma(2,3,7)$ with the following equivalent graphs: ${\,\raisebox{-64.01869pt}{\includegraphics[width=128.0374pt]{example-3ii}}\,}\stackrel{{\scriptstyle\text{Kirby}}}{{\sim}}{\,\raisebox{-59.75095pt}{\includegraphics[width=106.69783pt]{example-3iii}}\,}$ (69) The homological block turns out to be $\hat{Z}_{b_{1}}^{SU(2)}=q^{1/2}(1-q-q^{5}+q^{10}-q^{11}+q^{18}+q^{30}-q^{41}+q^{43}-q^{56}-q^{76}\cdots).$ (70) $\bullet$ For a three-manifold with non-trivial $H_{1}(M,{\mathbb{Z}})={\mathbb{Z}}_{3}$ as drawn below, ${\,\raisebox{-56.9055pt}{\includegraphics[width=227.62204pt]{example-2a}}\,}\stackrel{{\scriptstyle\text{Kirby}}}{{\sim}}{\,\raisebox{-56.9055pt}{\includegraphics[width=113.81102pt]{example-2b}}\,}$ (71) the three homological blocks are $\hat{Z}^{SU(2)}=\left(\begin{array}[]{c}1-q+q^{6}-q^{11}+q^{13}-q^{20}+q^{35}+O\left(q^{41}\right)\\\ q^{5/3}\left(-1+q^{3}-q^{21}+q^{30}+O\left(q^{41}\right)\right)\\\ q^{5/3}\left(-1+q^{3}-q^{21}+q^{30}+O\left(q^{41}\right)\right)\\\ \end{array}\right),$ (72) where two of them are equal. Our focus is to obtain explicit $q$-series for $SO(3)$ and $OSp(1|2)$ groups. Using the relation between $SU(2)$ and $SO(3)$, $SU(2)$ and $OSp(1|2)$ link invariants(2.1.1), we will investigate the necessary steps starting from the WRT invariant for $SO(3)$ and $OSp(1|2)$ eventually leading to the $\hat{Z}$-invariant. This will be the theme of the following section. ## 4 $\hat{Z}$ for $SO(3)$ and $OSp(1|2)$ Our aim is to derive the $\hat{Z}$-invariant for $SO(3)$ and $OSp(1|2)$ groups. We will first look at the WRT invariants $\tau_{K}^{SO(3)}[M(\Gamma);Q]$ for plumbed three-manifolds written in terms of colored Jones invariants of framed links ${\mathcal{L}}[\Gamma]$ in the following subsection and then discuss $OSp(1|2)$ $\hat{Z}$ in the subsequent section. ### 4.1 $SO(3)$ WRT invariant and $\hat{Z}^{SO(3)}$ invariant Recall that the framed link invariants are written in variable ${\mathbbm{q}}$ which is dependent on Chern-Simons coupling and the rank of the gauge group $\mathcal{G}$. For $SO(3)$ Chern-Simons with coupling $K\in 2\mathbb{Z}$, the variable $Q=e^{\frac{2\pi i}{K+1}}$. Hence $F^{SO(3)}[\mathcal{L}(\Gamma);Q]$ in WRT $\tau_{K}^{SO(3)}[M(\Gamma);Q]$ is $F^{SO(3)}[\mathcal{L}(\Gamma);Q]=\\\ \sum_{n_{1},n_{2},\dots,n_{L}\in\\{0,1,\dots,K\\}}V_{n_{1},n_{2},\dots,n_{L}}^{SO(3)}(\mathcal{L}(\Gamma);Q)\prod_{v=1}^{L}V_{n_{1},n_{2},\dots,n_{L}}^{SO(3)}(\bigcirc;Q)~{}=\\\ \\\ \sum_{n_{1},n_{2},\ldots,n_{L}\in\\{1,3,\dots,2K+1\\}}J_{n_{1},n_{2},\dots,n_{L}}^{SU(2)}\left(\mathcal{L}(\Gamma);{\mathbbm{q}}=e^{\frac{2\pi i}{2K+2}}\right)\prod_{v=1}^{L}\frac{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}{{\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2}}\bigg{|}_{{\mathbbm{q}}^{2}\rightarrow Q}~{},$ (73) where we have used the relation (13) to write $SO(3)$ link invariants in terms of the colored Jones invariants. Notice that the summation is over only odd integers and hence WRT invariant for $SO(3)$ is different from the WRT for $SU(2)$ group. Further, the highest integrable representation in the summation indicates that the Chern-Simons coupling for $SU(2)$ group is $2K+2$. After performing the summation, we can convert the ${\mathbbm{q}}=Q^{1/2}$(13) to obtain $SO(3)$ WRT invariant. We need to modify the Gauss sum reciprocity formula to incorporate the summation over odd integers in $F^{SO(3)}[\mathcal{L}(\Gamma);Q]$. Using the following Gauss sum reciprocity formula $\sum_{n\;\in\;{\mathbb{Z}}^{L}/k{\mathbb{Z}}^{L}}\exp\left(\frac{2\pi i}{k}(n,Bn)+\frac{2\pi i}{k}(\ell,n)\right)=\\\ \frac{e^{\frac{\pi i\sigma}{4}}\,(k/2)^{L/2}}{|\det B|^{1/2}}\sum_{a\;\in\;{\mathbb{Z}}^{L}/2B{\mathbb{Z}}^{L}}\exp\left(\frac{-\pi ik}{2}\left(a+\frac{\ell}{k},B^{-1}\left(a+\frac{\ell}{k}\right)\right)\right)~{},$ (74) for $k=2K+2$, we can obtain the summation over odd integers by replacing $n\longrightarrow\frac{n-1}{2}$ : $\sum_{n_{1},n_{2},\dots,n_{L}\;\in\;\\{1,3,\dots,4K+3\\}}{\mathbbm{q}}^{\frac{(n,Bn)}{4}+\frac{(n,d)}{2}}=\frac{e^{\frac{\pi i\sigma}{4}}\,(K+1)^{L/2}}{|\det B|^{1/2}}{\mathbbm{q}}^{-\frac{(d,B^{-1}d)}{4}}\times\\\ \\\ \sum_{a\;\in\;{\mathbb{Z}}^{L}/2B{\mathbb{Z}}^{L}}\exp\left[-\pi i(K+1)(a,B^{-1}a)\right]\exp\left[-\pi i(a,B^{-1}(d+BI))\right],$ (75) where $d=\ell-BI$ with $I$ denoting $L$ component vector with entry $1$ on all the components. That is, the transpose of the vector $I$ is $I^{T}=[1,1,\ldots,1]~{}.$ (76) For unknot with framing $\pm 1$, the $F^{SO(3)}[{\mathcal{L}}(-1\bullet);Q={\mathbbm{q}}^{2}]$ involving summation over odd integers simplifies to $F^{SO(3)}[{\mathcal{L}}(\pm\bullet);Q={\mathbbm{q}}^{2}]=\frac{\sqrt{K+1}\;e^{\pm\pi i/4}\;{\mathbbm{q}}^{\mp 3/4}}{{\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2}}\underbrace{(1+e^{\pi iK})}_{2}~{},$ (77) as the coupling $K\in 2\mathbb{Z}$ for the $SO(3)$ Chern-Simons theory. Hence, the WRT invariant takes the following form: $\tau_{K}^{SO(3)}[M(\Gamma);Q={\mathbbm{q}}^{2}]=\frac{e^{-\frac{\pi i\sigma}{4}}\,{\mathbbm{q}}^{\frac{3\sigma}{4}}}{2^{L}(K+1)^{L/2}\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})}\times\\\ {\sum_{{n}\in\\{1,3,\ldots,2K+1\\}^{L}}}\prod_{v\;\in\;\text{Vertices}}{\mathbbm{q}}^{\frac{f_{v}(n_{v}^{2}-1)}{4}}\,\left(\frac{1}{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}\right)^{\text{deg}(v)-2}\times\\\ \prod_{(v^{\prime},v^{\prime\prime})\;\in\;\text{Edges}}\Big{(}{\mathbbm{q}}^{n_{v^{\prime}}n_{v^{\prime\prime}}/2}-{\mathbbm{q}}^{-n_{v^{\prime}}n_{v^{\prime\prime}}/2}\Big{)}.$ (78) In above equation, the terms involving edges of the graph $\Gamma$ $\prod_{(v^{\prime},v^{\prime\prime})\;\in\;\text{Edges}}\Big{(}{\mathbbm{q}}^{n_{v^{\prime}}n_{v^{\prime\prime}}/2}-{\mathbbm{q}}^{-n_{v^{\prime}}n_{v^{\prime\prime}}/2}\Big{)}=2^{L-1}\prod_{(v^{\prime},v^{\prime\prime})\;\in\;\text{Edges}}\frac{\Big{(}{\mathbbm{q}}^{n_{v^{\prime}}n_{v^{\prime\prime}}/2}-{\mathbbm{q}}^{-n_{v^{\prime}}n_{v^{\prime\prime}}/2}\Big{)}}{2},$ can also be rewritten as $\prod_{(v^{\prime},v^{\prime\prime})\in\text{Edges}}({\mathbbm{q}}^{n_{v^{\prime}}n_{v^{\prime\prime}}/2}-{\mathbbm{q}}^{-n_{v^{\prime}}n_{v^{\prime\prime}}/2})=\sum_{p\in\\{\pm 1\\}^{\text{Edges}}}\prod_{(v^{\prime},v^{\prime\prime})\in\text{Edges}}p_{(v^{\prime},v^{\prime\prime})}{\mathbbm{q}}^{p_{(v^{\prime},v^{\prime\prime})}n_{v^{\prime}}n_{v^{\prime\prime}}/2}~{}.$ Here again, if we make a change of variable as $n_{v}\longrightarrow-n_{v}$ at any vertex, a term in the sum with a given configuration of signs associated to edges (that is $p\in\\{\pm 1\\}^{\text{Edges}}$) will transform into a term with a different configuration times $(-1)^{\text{deg}(v)}$. However, for these plumbing graphs $\Gamma$, the signs of such configuration can be brought to the configuration with all signs +1. Incorporating this fact, the WRT invariant(78) simplifies to $\tau_{K}^{SO(3)}[M(\Gamma);Q={\mathbbm{q}}^{2}]=\frac{e^{-\frac{\pi i\sigma}{4}}\,{\mathbbm{q}}^{\frac{3\sigma-\sum_{v}f_{v}}{4}}}{2\,(K+1)^{L/2}\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})}\times\\\ {\sum_{{n}\in\\{1,3,\ldots,2K+1\\}^{L}}}\;\;{\mathbbm{q}}^{\frac{(n,Bn)}{4}}\prod_{v\;\in\;\text{Vertices}}\,\left(\frac{1}{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}\right)^{\text{deg}(v)-2}.$ (79) Further, we double the range of summation so as to use the reciprocity formula(75) $\tau_{K}^{SO(3)}[M(\Gamma);Q={\mathbbm{q}}^{2}]=\frac{e^{-\frac{\pi i\sigma}{4}}\,{\mathbbm{q}}^{\frac{3\sigma-\sum_{v}f_{v}}{4}}}{4\,(K+1)^{L/2}\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})}\times\\\ {\sum_{{n}\in\\{1,3,\ldots,4K+3\\}^{L}}}\;\;{\mathbbm{q}}^{\frac{(n,Bn)}{4}}\prod_{v\;\in\;\text{Vertices}}\,\left(\frac{1}{{\mathbbm{q}}^{n_{v}/2}-{\mathbbm{q}}^{-n_{v}/2}}\right)^{\text{deg}(v)-2}.$ (80) The steps discussed in the $SU(2)$ context to extract $\hat{Z}$ can be similarly followed for $SO(3)$. This procedure leads to $\tau_{K}^{SO(3)}[M(\Gamma);Q={\mathbbm{q}}^{2}]=\frac{1}{2\,({\mathbbm{q}}^{1/2}-{\mathbbm{q}}^{-1/2})\,|\det B|^{1/2}}\,\sum_{a\in\mathrm{Coker}\,B}e^{-\pi i(K+1)(a,B^{-1}a)}\\\ \\\ \sum_{b\in 2\mathrm{Coker}\,B+\delta}e^{-\pi i\big{(}a,B^{-1}(b{\color[rgb]{0,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,1}+BI})\big{)}}\lim_{q\rightarrow{\mathbbm{q}}}\hat{Z}^{SO(3)}_{b}[M(\Gamma);q]~{}.$ (81) We observe that the $SO(3)$ WRT invariant is different from the $SU(2)$ invariant due to the factor highlighted in blue color in the summand whereas the $\hat{Z}^{SO(3)}_{b}[M(\Gamma);q]$ is exactly same as the $SU(2)$ $q$-series. Even though $SO(3)\equiv SU(2)/\mathbb{Z}_{2}$, it is surprising to see that the factor group $SO(3)$ shares the same $\hat{Z}$ as that of the parent group $SU(2)$. The case of $\hat{Z}^{SO(3)}$ was also considered in Costantino:2021yfd but they took a different route by considering the refined WRT invariant which is consistent with our result. In the following subsection, we will extract $\hat{Z}$ from the WRT invariant $\tau_{\hat{K}}^{OSp(1|2)}[M(\Gamma);\hat{Q}]$ for $OSp(1|2)$ supergroup. We will see that the $OSp(1|2)$ $q$-series are related to $\hat{Z}^{SU(2)}[M(\Gamma);q]$. ### 4.2 $OSp(1|2)$ WRT and $\hat{Z}^{OSp(1|2)}$ invariant Using the relation between $OSp(1|2)$ and $SU(2)$ link invariants (17), the WRT invariant can be written for plumbed manifolds $M(\Gamma)$ as $\tau_{\hat{K}}^{OSp(1|2)}[M(\Gamma);\hat{Q}={\mathbbm{q}}]=\frac{e^{-\frac{\pi i\sigma}{4}}\,{\mathbbm{q}}^{\frac{3\sigma}{4}}}{(2\hat{K}+3)^{L/2}\,({\mathbbm{q}}^{1/2}+{\mathbbm{q}}^{-1/2})}\times\\\ {\sum_{{n_{1},n_{2},\ldots,n_{L}}\in\\{1,3,\ldots,2\hat{K}+1\\}}}\;\;\prod_{v\;\in\;\text{Vertices}}{\mathbbm{q}}^{\frac{f_{v}(n_{v}^{2}-1)}{4}}\,\left(\frac{1}{{\mathbbm{q}}^{n_{v}/2}+{\mathbbm{q}}^{-n_{v}/2}}\right)^{\text{deg}(v)-2}\times\\\ \prod_{(v^{\prime},v^{\prime\prime})\;\in\;\text{Edges}}\Big{(}{\mathbbm{q}}^{n_{v^{\prime}}n_{v^{\prime\prime}}/2}+{\mathbbm{q}}^{-n_{v^{\prime}}n_{v^{\prime\prime}}/2}\Big{)}\Big{|}_{{\mathbbm{q}}=\hat{Q}}~{}.$ (82) Here again we use the Gauss reciprocity(75) as the summation is over odd integers to work out the steps leading to $\hat{Z}^{OSp(1|2)}[M(\Gamma);q]$. Note that, the highest integrable representation $2\hat{K}+1$ which fixes the ${\mathbbm{q}}$ as $(2\hat{K}+2)$-th root of unity. However to compare the result with $OSp(1|2)$ WRT, we have to replace $\hat{K}+1\rightarrow 2\hat{K}+3$ which is equivalent to ${\mathbbm{q}}=\hat{Q}$. Following similar steps performed for $SU(2)$, we find the following expression for $OSp(1|2)$ WRT invariant: $\frac{1}{2\,({\mathbbm{q}}^{1/2}+{\mathbbm{q}}^{-1/2})\,|\det B|^{1/2}}\,\sum_{a\in\mathrm{Coker}\,B}e^{-\pi i(2\hat{K}+3)(a,B^{-1}a)}\times\\\ \sum_{b\in 2\mathrm{Coker}\,B+\delta}e^{-\pi i\big{(}a,B^{-1}(b{\color[rgb]{1,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,0}+BI})\big{)}}\lim_{q\rightarrow\hat{Q}}\hat{Z}_{b}^{OSp(1|2)}[M(\Gamma);q],$ (83) where $I$ is again the column vector (76) and $\hat{Z}_{b}^{OSp(1|2)}[M(\Gamma);q]$ is given by the following algebraic expression: $\hat{Z}_{b}^{OSp(1|2)}[M(\Gamma);q]\;\;=\;\;2^{-L}q^{-\frac{3L+\sum_{v}f_{v}}{4}}\sum_{d\;\in\;2B{\mathbb{Z}}^{L}+b}F^{d}_{1}\,q^{-\frac{(d,B^{-1}d)}{4}},$ (84) with coefficient $F_{1}^{d}$ is obtained by following relation $\sum_{d\;\in\;2{\mathbb{Z}}^{L}+\delta}F_{1}^{d}\prod_{v}x_{v}^{d_{v}}=\\\ \prod_{v\,\in\,\text{Vertices}}\left\\{{\scriptsize\begin{array}[]{c}\text{Expansion}\\\ \text{at }x\rightarrow 0\end{array}}\frac{1}{(x_{v}+1/x_{v})^{\text{deg}\,v-2}}+{\scriptsize\begin{array}[]{c}\text{Expansion}\\\ \text{at }x\rightarrow\infty\end{array}}\frac{1}{(x_{v}+1/x_{v})^{\text{deg}\,v-2}}\right\\}.$ (85) Equivalently, $\hat{Z}^{OSp(1|2)}[M(\Gamma);q]$(84) can also represented as the following contour integral: $\hat{Z}_{b}^{OSp(1|2)}[M(\Gamma);q]=q^{-\frac{3L+\sum_{v}f_{v}}{4}}\cdot\text{v.p.}\int\limits_{|z_{v}|=1}\prod_{v\;\in\;\text{Vertices}}\frac{dz_{v}}{2\pi iz_{v}}\,\left({z_{v}+1/z_{v}}\right)^{2-\text{deg}(v)}\cdot\Theta^{-B}_{b}(z)~{}.$ (86) Here $\Theta^{-B}_{b}(x)$ is the theta function of the lattice corresponding to minus the linking form $B$: $\Theta^{-B}_{b}(x)=\sum_{d\;\in\;2B{\mathbb{Z}}^{L}+b}q^{-\frac{(d,B^{-1}d)}{4}}\prod_{i=1}^{L}x_{i}^{d_{i}}$ (87) and “v.p.” again means that we take principle value integral (i.e. take half- sum of contours $|z_{v}|=1\pm\epsilon$). Comparing eqns.(84,85) with the $SU(2)$ expressions(58,59), we can see that the $\hat{Z}$ for $OSp(1|2)$ are different from $SU(2)$ q-series. We will now present explicit $q$-series for some examples. #### 4.2.1 Examples $\bullet$ For the Poincare homology sphere(67), we find the following $OSp(1|2)$ $q$-series $\hat{Z}_{b_{1}}^{OSp(1|2)}=q^{-3/2}(1+q+q^{3}+q^{7}+q^{8}+q^{14}+q^{20}-q^{29}+q^{31}-q^{42}-q^{52}+\cdots).$ (88) $\bullet$ In the case of Brieskorn homology sphere(69), the $OSp(1|2)$ $q$-series is $\hat{Z}_{b_{1}}^{OSp(1|2)}=q^{1/2}(1+q+q^{5}+q^{10}+q^{11}+q^{18}+q^{30}+q^{41}-q^{43}-q^{56}-q^{76}\cdots).$ (89) $\bullet$ For the case of plumbing graph(71), the three homological blocks are $\hat{Z}^{OSp(1|2)}=\left(\begin{array}[]{c}1+q+q^{6}+q^{11}-q^{13}-q^{20}-q^{35}+O\left(q^{41}\right)\\\ q^{5/3}\left(1+q^{3}-q^{21}-q^{30}+O\left(q^{41}\right)\right)\\\ q^{5/3}\left(1+q^{3}-q^{21}-q^{30}+O\left(q^{41}\right)\right)\end{array}\right).$ (90) After comparing the $q$-series for $SU(2)$ and $OSp(1|2)$, we noticed that these two $q$-series are related by a simple change of variable which is $q\longrightarrow-q$. This change of variable applies only to the series not to the overall coefficient outside the series. $\bullet$ Lens space $L(p,q)$ is a well studied three-manifold. For $L(-5,11)\sim L(-13,29)$ whose plumbing graph is shown below, we obtain the five homological blocks ${\,\raisebox{0.0pt}{\includegraphics[width=142.26378pt]{example4b}}\,}\;\;\stackrel{{\scriptstyle\text{Kirby}}}{{\sim}}\;\;{\,\raisebox{0.0pt}{\includegraphics[width=142.26378pt]{example4a}}\,}$ $\hat{Z}^{OSp(1|2)}=\left(\begin{array}[]{c}q^{1/10}\\\ q^{-1/10}\\\ 0\\\ q^{-1/10}\\\ q^{1/10}\end{array}\right)~{}~{}~{}~{}\text{as}~{}~{}H_{1}(M,\mathbb{Z})={\mathbb{Z}}_{5}.$ (91) $\bullet$ For the following plumbing graph, $H_{1}(M,{\mathbb{Z}})=\mathbb{Z}_{13}$ , $\hat{Z}^{OSp(1|2)}=\frac{1}{4}\left(\tiny\begin{array}[]{c}q^{-1/2}(2+2q+2q^{2}-2q^{4}-4q^{5}+6q^{10}-8q^{11}+4q^{13}+2q^{14}-4q^{15}+O\left(q^{18}\right))\\\ q^{5/26}(3+2q-2q^{2}-4q^{3}-2q^{7}+q^{8}+2q^{9}+q^{10}-2q^{12}-4q^{13}-2q^{16}+O\left(q^{18}\right))\\\ q^{5/26}(3+2q-2q^{2}-4q^{3}-2q^{7}+q^{8}+2q^{9}+q^{10}-2q^{12}-4q^{13}-2q^{16}+O\left(q^{18}\right))\\\ q^{7/26}(4-q-2q^{3}-2q^{4}-2q^{6}+3q^{7}-2q^{8}-2q^{10}+q^{11}+2q^{13}-4q^{14}+2q^{15}-4q^{16}+O\left(q^{18}\right))\\\ q^{7/26}(4-q-2q^{3}-2q^{4}-2q^{6}+3q^{7}-2q^{8}-2q^{10}+q^{11}+2q^{13}-4q^{14}+2q^{15}-4q^{16}+O\left(q^{18}\right))\\\ q^{-7/26}(3+3q^{2}-2q^{4}-2q^{5}+4q^{7}-2q^{8}+2q^{9}-2q^{10}-4q^{12}+4q^{13}-4q^{14}+2q^{15}+O\left(q^{18}\right))\\\ q^{-7/26}(3+3q^{2}-2q^{4}-2q^{5}+4q^{7}-2q^{8}+2q^{9}-2q^{10}-4q^{12}+4q^{13}-4q^{14}+2q^{15}+O\left(q^{18}\right))\\\ q^{-11/26}(1+2q+2q^{2}+4q^{3}+3q^{6}-2q^{7}-4q^{8}-2q^{9}+2q^{11}+2q^{13}-q^{14}+2q^{16}-2q^{17}+O\left(q^{18}\right))\\\ q^{-11/26}(1+2q+2q^{2}+4q^{3}+3q^{6}-2q^{7}-4q^{8}-2q^{9}+2q^{11}+2q^{13}-q^{14}+2q^{16}-2q^{17}+O\left(q^{18}\right))\\\ q^{-5/26}(2+2q^{2}+q^{3}+3q^{5}-2q^{6}-2q^{7}-4q^{8}-2q^{10}+2q^{11}-2q^{12}+2q^{13}+5q^{15}-2q^{16}+2q^{17}+O\left(q^{18}\right))\\\ q^{-5/26}(2+2q^{2}+q^{3}+3q^{5}-2q^{6}-2q^{7}-4q^{8}-2q^{10}+2q^{11}-2q^{12}+2q^{13}+5q^{15}-2q^{16}+2q^{17}+O\left(q^{18}\right))\\\ q^{-15/26}(1-2q-2q^{2}+q^{4}-2q^{6}-2q^{7}-2q^{8}-4q^{10}-2q^{12}+2q^{13}+2q^{15}+4q^{17}+O\left(q^{18}\right))\\\ q^{-15/26}(1-2q-2q^{2}+q^{4}-2q^{6}-2q^{7}-2q^{8}-4q^{10}-2q^{12}+2q^{13}+2q^{15}+4q^{17}+O\left(q^{18}\right))\end{array}\right)$ (92) We have checked for many examples that under $q\rightarrow-q$ in the $OSp(1|2)$ $q$-series(not affecting the overall coefficient), we obtain the $SU(2)$ $q$-series. ## 5 Conclusions and future directions Our goal was to investigate $\hat{Z}$ for $SO(3)$ and $OSp(1|2)$ groups for negative definite plumbed three-manifolds. The change of variable and color indeed relates invariants of framed links ${\mathcal{L}}[\Gamma]$(13,17) of $SO(3)$ and $OSp(1|2)$ to colored Jones. Such a relation allowed us to go through the steps of GPPV conjecture to extract $\hat{Z}$ from WRT invariants. Interestingly, we observe that the $\hat{Z}^{SO(3)}$ is same as $\hat{Z}^{SU(2)}$ even though the WRT invariants are different. We know that $SU(2)/\mathbb{Z}_{2}\equiv SO(3)$ and it is not at all obvious that the homological blocks are same for both the groups. It is important to explore other factor groups and the corresponding $\hat{Z}$ invariants. For the odd orthosympletic supergroup $OSp(1|2)$, we observe from our computations for many negative definite plumbing graph $\Gamma$: $\hat{Z}^{OSp(1|2)}_{b}(\Gamma;q)=2^{-c}q^{\Delta_{b}}\left(\sum_{n}a_{n}q^{n}\right)$ whereas their $SU(2)$ q-series is $\hat{Z}^{SU(2)}_{b}(\Gamma;q)=2^{-c}q^{\Delta_{b}}\left(\sum_{n}a_{n}(-q)^{n}\right)$ where $c\in\mathbb{Z}_{+}$, $\Delta_{b}\in\mathbb{Q}$. The brane setup in string theory for $U(N)$ gauge group gives a natural interpretation for these $q$-series as partition function of the theory $T^{\mathcal{G}}[M]$. In principle, there should be a natural generalisation to orthogonal $SO(N)$ and symplectic group $Sp(2n)$ involving orientifolds. It will be worth investigating such a construction to obtain $\hat{Z}$ for $SO(N)$ group and compare with our $SO(N=3)$ results. Extension of $\hat{Z}$ to the two variable series for link complementsGukov:2019mnk is another direction to pursue. We hope to report on these aspects in future. ###### Acknowledgements. PR is grateful to ICTP senior associateship funding for visit where this work was initiated with Pavel Putrov and Francesca Ferrari during summer 2019. Unfortunately due to covid, we could not pursue the collaboration through online mode. We would like to thank Pavel Putrov and Francesca Ferrari for clarifying the notations during the initial stages. SC would like to thank Sunghyuk Park for useful comments and discussions during the String-Math 2022 conference held at Univ. of Warsaw, Poland. SC is grateful to Dmitry Noshchenko for his comments on the manuscript. We would also like to thank Vivek Kumar Singh for his clarification on mathematica program. SC would also like to thank the organisers of String-Math 2022 where a part of this work was presented. SC is thankful for the MHRD fellowship from IIT Bombay providing financial support to visit. PR would like to thank SERB (MATRICS) MTR/2019/000956 funding. We would like to thank the reviewers for their valuable comments and the suggestions. ## References * (1) Chae, J.: Towards a q-series for osp (2| 2n). arXiv preprint arXiv:2106.09868 (2021) * (2) Chung, H.J.: BPS Invariants for Seifert Manifolds. JHEP 03, 113 (2020). DOI 10.1007/JHEP03(2020)113 * (3) Cooper, B., Krushkal, V.: Categorification of the jones–wenzl projectors. Quantum Topology 3(2), 139–180 (2012) * (4) Costantino, F., Gukov, S., Putrov, P.: Non-Semisimple TQFT’s and BPS $q$-Series. SIGMA 19, 010 (2023). DOI 10.3842/SIGMA.2023.010 * (5) Ennes, I.P., Ramadevi, P., Ramallo, A.V., Sanchez de Santos, J.M.: Duality in osp (1|2) conformal field theory and link invariants. Int. J. Mod. Phys. A 13, 2931–2978 (1998). DOI 10.1142/S0217751X98001487 * (6) Ferrari, F., Putrov, P.: Supergroups, q-series and 3-manifolds (2020) * (7) Frenkel, I.B., Stroppel, C., Sussan, J.: Categorifying fractional euler characteristics, jones–wenzl projectors and 3 j-symbols. Quantum Topology 3(2), 181–253 (2012) * (8) Gopakumar, R., Vafa, C.: M theory and topological strings. 1. (1998) * (9) Gopakumar, R., Vafa, C.: M theory and topological strings. 2. (1998) * (10) Gukov, S., Manolescu, C.: A two-variable series for knot complements. Quantum Topol. 12(1), 1–109 (2021). DOI 10.4171/qt/145 * (11) Gukov, S., Pei, D., Putrov, P., Vafa, C.: BPS spectra and 3-manifold invariants. J. Knot Theor. Ramifications 29(02), 2040003 (2020). DOI 10.1142/S0218216520400039 * (12) Gukov, S., Putrov, P., Vafa, C.: Fivebranes and 3-manifold homology. Journal of High Energy Physics 2017(7) (2017). DOI 10.1007/jhep07(2017)071. URL http://dx.doi.org/10.1007/JHEP07(2017)071 * (13) Gukov, S., Schwarz, A.S., Vafa, C.: Khovanov-Rozansky homology and topological strings. Lett. Math. Phys. 74, 53–74 (2005). DOI 10.1007/s11005-005-0008-8 * (14) Kaul, R.K., Ramadevi, P.: Three-manifold invariants from Chern-Simons field theory with arbitrary semi-simple gauge groups. Commun. Math. Phys. 217, 295–314 (2001). DOI 10.1007/s002200000347 * (15) Khovanov, M.: A categorification of the jones polynomial. Duke Mathematical Journal 101(3), 359–426 (2000) * (16) Khovanov, M., Rozansky, L.: Matrix factorizations and link homology. arXiv preprint math/0401268 (2004) * (17) Lickorish, W.B.R.: A representation of orientable combinatorial 3-manifolds. Annals of Mathematics 76(3), 531–540 (1962). URL http://www.jstor.org/stable/1970373 * (18) Mikhaylov, V., Witten, E.: Branes And Supergroups. Commun. Math. Phys. 340(2), 699–832 (2015). DOI 10.1007/s00220-015-2449-y * (19) Nawata, S., Oblomkov, A.: Lectures on knot homology. Contemp. Math. 680, 137 (2016). DOI 10.1090/conm/680/13702 * (20) Ooguri, H., Vafa, C.: Knot invariants and topological strings. Nucl. Phys. B 577, 419–438 (2000). DOI 10.1016/S0550-3213(00)00118-8 * (21) Park, S., et al.: Higher rank hat z and fk. SIGMA. Symmetry, Integrability and Geometry: Methods and Applications 16, 044 (2020) * (22) Ramadevi, P., Naik, S.: Computation of Lickorish’s three manifold invariants using Chern-Simons theory. Commun. Math. Phys. 209, 29–49 (2000). DOI 10.1007/s002200050014 * (23) Wallace, A.H.: Modifications and cobounding manifolds. Canadian Journal of Mathematics 12, 503–528 (1960). DOI 10.4153/CJM-1960-045-7 * (24) Webster, B.: Knot invariants and higher representation theory, vol. 250. American Mathematical Society (2017) * (25) Witten, E.: Quantum field theory and the jones polynomial. Communications in Mathematical Physics 121(3), 351–399 (1989)
Global Existence of Bi-Hamiltonian Structures on Orientable Three-Dimensional Manifolds Global Existence of Bi-Hamiltonian Structures on Orientable Three-Dimensional Manifolds Melike IŞİM EFE and Ender ABADOĞLU M. Işim Efe and E. Abadoğlu Yeditepe University, Mathematics Department, İnȯnu̇ Mah. Kayışdağı Cad. 326A, 26 Ağustos Yerleşimi, 34755 Ataşehir İstanbul, Turkey <EMAIL_ADDRESS><EMAIL_ADDRESS> Received December 21, 2016, in final form July 04, 2017; Published online July 14, 2017 In this work, we show that an autonomous dynamical system defined by a nonvanishing vector field on an orientable three-dimensional manifold is globally bi-Hamiltonian if and only if the first Chern class of the normal bundle of the given vector field vanishes. Furthermore, the bi-Hamiltonian structure is globally compatible if and only if the Bott class of the complex codimension one foliation defined by the given vector field vanishes. bi-Hamiltonian systems; Chern class; Bott class 53D17; 53D35 Dedicated to the memory of Ali Yavuz. ## 1 Introduction An autonomous dynamical system on a manifold $M$ $\displaystyle\dot{x}(t)=v(x(t))$ (1.1) is determined by a vector field $v(x)$ on a manifold up to time reparametrization. Important geometric quantities related to a dynamical system are functions $I$ which are invariant under the flow of the vector field $\displaystyle\mathcal{L}_{v}I=0.$ It is sometimes possible to relate the vector field to an invariant function via a Poisson structure $\mathcal{J},$ which is a bivector field on $M$ $\displaystyle\mathcal{J}\colon\ \Lambda^{1}(M)\rightarrow\mathfrak{X}(M)$ satisfying the Jacobi identity condition $\displaystyle[\mathcal{J},\mathcal{J}]_{\rm SN}=0,$ where $[\,,\,]_{\rm SN}$ is the Schouten–Nijenhuis bracket. The local structure of such manifolds was first introduced in [13]. The invariants satisfying the condition $\displaystyle v=\mathcal{J}({\rm d}I)$ (1.2) are called Hamiltonian functions of (1.1). Given a dynamical system on $M$ defined by the vector field $v$, the vector field $v$ is called a Hamiltonian vector field if there exists a Poisson bivector $\mathcal{J}$ and a smooth function $I$ such that equation (1.2) holds. Given a vector field $v$ on $M$, finding a Poisson structure according to which the vector field becomes Hamiltonian may not be an easy task in general. However, if a given dynamical system can be put into Hamiltonian form then, there may be more than one Poisson structure which makes it into a Hamiltonian system. In [9], a bi-Hamiltonian system is introduced for the analysis of certain infinite-dimensional soliton equations. In such a case, there arises the question of the relation between these Poisson structures, which is called compatibility. Although there are at least two different approaches to compatibility [11], by following [10] we adapt the definitions below: ###### Definition 1.1. A dynamical system is called bi-Hamiltonian if it can be written in Hamiltonian form in two distinct ways: $\displaystyle v=\mathcal{J}_{1}({\rm d}H_{2})=\mathcal{J}_{2}({\rm d}H_{1}),$ (1.3) such that $\mathcal{J}_{1}$ and $\mathcal{J}_{2}$ are nowhere multiples of each other. This bi-Hamiltonian structure is compatible if $\mathcal{J}_{1}+\mathcal{J}_{2}$ is also a Poisson structure. In this paper we confine ourselves to dynamical systems on three-dimensional orientable manifolds. For three-dimensional manifolds, where there is no symplectic structure for dimensional reasons, Poisson structures have a simple form. Poisson structures of dynamical systems on three manifolds are extensively studied first in [4] and then also in [5] and [8]. Following the definitions in [4], choosing any Riemannian metric $\boldsymbol{g}$ on $M$, a Poisson bivector field, which is skew-symmetric, can be associated to a vector field by using the Lie algebra isomorphism $\mathfrak{so}(3)\simeq\mathbb{R}^{3}$ $\displaystyle\mathcal{J}=\mathcal{J}^{mn}e_{m}\wedge e_{n}=\varepsilon_{k}^{mn}J^{k}e_{m}\wedge e_{n},$ and the vector field $\displaystyle J=J^{k}e_{k}$ is called the Poisson vector field on $M$. Then, the Jacobi identity has the form $\displaystyle J\cdot(\nabla\times J)=0,$ (1.4) and equation (1.3) becomes $\displaystyle v=J_{1}\times\nabla H_{2}=J_{2}\times\nabla H_{1}.$ (1.5) Since $J_{1}$ and $J_{2}$ are not multiples of each other by definition, we have $\displaystyle J_{1}\times J_{2}\neq 0$ (1.6) and $\displaystyle J_{i}\cdot v=0$ (1.7) for $i=1,2$. This work is focused on the bi-Hamiltonian structure of dynamical systems defined by nonvanishing vector fields on orientable three-dimensional manifolds, or equivalently vector fields on three-dimensional manifolds whose supports are orientable three-dimensional manifolds. Since all orientable three-dimensional manifolds are parallelizable [12], there is no topological obstruction to the global existence of a nonvanishing vector field. Then, by the bi-Hamiltonian form (1.5)–(1.7), $\\{v,J_{1},J_{2}\\}$ forms a local frame field. Therefore, whenever the system is globally bi-Hamiltonian, $\\{v,J_{1},J_{2}\\}$ becomes a global frame field on $M$. For example, for $M=\mathbb{R}^{3}$ and $v=\partial_{x^{0}}$ we have $J_{i}=\partial_{x^{i}}$ and $\\{\partial_{x^{0}},\partial_{x^{1}},\partial_{x^{2}}\\}$ forms such a global frame field. However, the global existence of the frame field $\\{v,J_{1},J_{2}\\}$ is by no means guaranteed. The simplest counterexample is the gradient flow of the $S^{2}$ in $\mathbb{R}^{3}\setminus\\{0\\}$. Here, a frame field $\\{v,J_{1},J_{2}\\}$ cannot be defined globally since $J_{1}$, $J_{2}$ are sections of the tangent bundle of $S^{2}$ which is not trivial and does not admit two nonvanishing linearly independent vector fields. The goal of this paper is to give necessary and sufficient conditions for a nonvanishing vector field on an orientable three-dimensional manifold to admit a compatible bi-Hamiltonian structure. The paper is organized as follows: In Section 2, the local existence of bi-Hamiltonian systems is investigated in a neighbourhood of a point, possibly refined by the existence conditions of solutions of certain ODE’s related with the problem, and it is shown in Theorem 2.7 that it is always possible to find a pair of compatible Poisson structures such that the system defined by the nonvanishing vector field becomes bi-Hamiltonian. In Section 3, obstructions to the global existence of a pair of Poisson structures are studied. In Section 3.2 the primary obstruction for the existence of a global pair of Poisson structures is investigated, and it is shown in Theorem 3.6 that such a pair, which is not necessarily compatible, exists if and only if the first Chern class of the normal bundle vanishes. Finally, the global compatibility of this pair is investigated in Section 3.3 and it is shown in Theorem 3.8 that under the assumption of global existence, the vanishing of the Bott class of the complex codimension one foliation is the necessary and sufficient condition for the global compatibility of the pair of Poisson structures. Throughout the work, bivectors are denoted by calligraphic and forms are denoted by bold letters. ## 2 Local existence of bi-Hamiltonian structure in 3D For this purpose, we will first analyze the local solutions of the equation (1.4) defining Poisson vectors, which is also studied in [6]. Let $M$ be an orientable three-dimensional manifold with an arbitrary Riemannian metric $\boldsymbol{g}$, and $v$ be a nonvanishing vector field. Let $\displaystyle\widehat{e}_{1}=\frac{v}{\|v\|}$ and extend this vector field to a local orthonormal frame field $\\{\widehat{e}_{1},\widehat{e}_{2},\widehat{e}_{3}\\}$. Define the structure functions $(C_{ij}^{k}(x))$ via the relation $\displaystyle[\widehat{e}_{i},\widehat{e}_{j}]=C_{ij}^{k}(x)\widehat{e}_{k}.$ (2.1) ###### Proposition 2.1. A nonvanishing vector field $v$ admits two independent local Poisson structures on $M$. ###### Proof. Adopting the frame defined above and using (1.7), we have the Poisson vector field $\displaystyle J=\alpha\widehat{e}_{2}+\beta\widehat{e}_{3},$ (2.2) and its curl is given by $\displaystyle\nabla\times J=\nabla\alpha\times\widehat{e}_{2}+\alpha\nabla\times\widehat{e}_{2}+\nabla\beta\times\widehat{e}_{3}+\beta\nabla\times\widehat{e}_{3}.$ (2.3) Now the Jacobi identity (1.4) is obtained by taking the dot product of (2.2) with (2.3), and using triple vector product identities we get $\displaystyle\beta\widehat{e}_{1}\cdot\nabla\alpha-\alpha\widehat{e}_{1}\cdot\nabla\beta-\alpha^{2}C_{31}^{2}-\alpha\beta\big{(}C_{31}^{3}+C_{12}^{2}\big{)}-\beta^{2}C_{12}^{3}=0.$ (2.4) If $J=0$ then $\|v\|=0$ and hence $v=0$, which contradicts with our assumption that the vector field is nonvanishing. Therefore, we assume $\displaystyle J\neq 0,$ which means that $\alpha\neq 0$ or $\beta\neq 0$. Now we assume $\alpha\neq 0$, while the case $\beta\neq 0$ is similar and amounts to a rotation of the frame fields. Defining $\displaystyle\mu=\frac{\beta}{\alpha}$ and dividing (2.4) by $\alpha^{2}$, we get $\displaystyle\widehat{e}_{1}\cdot\nabla\mu=-C_{31}^{2}-\mu\big{(}C_{31}^{3}+C_{12}^{2}\big{)}-\mu^{2}C_{12}^{3},$ (2.5) whose characteristic curve is the integral curve of (1.1) in arclength parametrization and $\displaystyle\frac{{\rm d}\mu}{{\rm d}s}=-C_{31}^{2}-\mu\big{(}C_{31}^{3}+C_{12}^{2}\big{)}-\mu^{2}C_{12}^{3}$ (2.6) in the arclength variable $s$. The Riccati equation (2.6) is equivalent to a linear second order equation and hence, possesses two linearly independent solutions leading to two Poisson vector fields. Since the vector field $v$ is assumed to be nonvanishing, for each $\boldsymbol{x}_{0}\in\mathbb{R}^{3}$ it is possible to find a neighborhood foliated by the integral curves of $v$ which are nothing but characteristic curves of (2.5). Therefore, solutions of (2.6) can be extended to a possibly smaller neighborhood on which the Riccati equation has two independent solutions which we call $\mu_{i}$ for $i=1,2$. Hence, we have two Poisson vector fields $\displaystyle J_{i}=\alpha_{i}\big{(}\widehat{e}_{2}+\mu_{i}\widehat{e}_{3}\big{)},$ (2.7) where the coefficients $\alpha_{i}$ are arbitrary. ∎ Note that, (2.5) determines $\mu_{i}$ alone, but not $\alpha_{i}$. Taking the advantage of the freedom of choosing arbitrary scaling factors we may restrict these factors by imposing compatibility on our Poisson vector fields. ###### Proposition 2.2. Two Poisson structures obtained in (2.5) are compatible iff $\displaystyle\widehat{e}_{1}\cdot\nabla\ln\frac{\alpha_{i}}{\alpha_{j}}=C_{12}^{3}(\mu_{i}-\mu_{j}).$ (2.8) ###### Proof. Let $\displaystyle J=J_{1}+J_{2}$ Using (1.4) for $J_{1}$, $J_{2}$ and $J$ $\displaystyle(\nabla\times J)\cdot J=(\nabla\times J_{2})\cdot J_{1}+(\nabla\times J_{1})\cdot J_{2}=0.$ (2.9) For the Poisson vector fields defined in (2.5), taking the dot product of both sides of (2.3) by $J_{j}$, leads to $\displaystyle(\nabla\times J_{i})\cdot J_{j}=\alpha_{i}\alpha_{j}(\mu_{i}-\mu_{j})\big{(}C_{12}^{2}+C_{12}^{3}\mu_{i}-\widehat{e}_{1}\cdot\nabla\ln\alpha_{i}\big{)}.$ (2.10) Therefore, the compatibility condition (2.9) implies that $\displaystyle C_{12}^{2}+C_{12}^{3}\mu_{i}-\widehat{e}_{1}\cdot\nabla\ln\alpha_{i}=C_{12}^{2}+C_{12}^{3}\mu_{j}-\widehat{e}_{1}\cdot\nabla\ln\alpha_{j},$ and hence, we get $\displaystyle\widehat{e}_{1}\cdot\nabla\ln\frac{\alpha_{i}}{\alpha_{j}}=C_{12}^{3}(\mu_{i}-\mu_{j}),$ (2.11) whose characteristic curve is the solution curve of (1.1) in arclength parametrization $\displaystyle\frac{{\rm d}}{{\rm d}s}\ln\frac{\alpha_{i}}{\alpha_{j}}=C_{12}^{3}(\mu_{i}-\mu_{j}).$ (2.12) By a similar line of reasoning as above, the solutions of (2.12) can also be extended to the whole neighborhood, and the proposition follows. ∎ However, having a pair of Poisson structures obtained in (2.5) and even a compatible pair satisfying (2.11) do not guarantee the existence of Hamiltonian functions even locally. ###### Proposition 2.3. The dynamical system (1.1) is locally bi-Hamiltonian with the pair of Poisson structures obtained in (2.7) if and only if $\displaystyle\widehat{e}_{1}\cdot\nabla\ln\frac{\alpha_{i}}{\|v\|}=C_{31}^{3}+\mu_{i}C_{12}^{3}.$ (2.13) ###### Proof. For this purpose we first need to write down the equations for the Hamiltonian functions. The invariance condition of Hamiltonian functions under the flow generated by $v$ implies $\displaystyle\widehat{e}_{1}\cdot\nabla H_{i}=0,$ (2.14) so the gradients of Hamiltonian functions are linear combinations of $\widehat{e}_{2}$ and $\widehat{e}_{3}$. Then, inserting (2.7) into (1.5) we get another condition $\displaystyle\widehat{e}_{3}\cdot\nabla H_{j}-\mu_{i}\widehat{e}_{2}\cdot\nabla H_{j}=\frac{\|v\|}{\alpha_{i}}$ (2.15) or by defining $\displaystyle u_{i}=-\mu_{i}\widehat{e}_{2}+\widehat{e}_{3}$ (2.15) can be written as $\displaystyle u_{i}\cdot\nabla H_{j}=\frac{\|v\|}{\alpha_{i}}.$ (2.16) Equations (2.14) and (2.16) for Hamiltonian functions are subject to the integrability condition $\displaystyle\widehat{e}_{1}(u_{i}(H_{j}))-u_{i}\big{(}\widehat{e}_{1}(H_{j})\big{)}=\big{[}\widehat{e}_{1},u_{i}\big{]}(H_{j}).$ Using the commutation relations given in (2.1) and (2.5), we obtain $\displaystyle[\widehat{e}_{1},u_{i}]=-\big{(}C_{31}^{1}+\mu_{i}C_{12}^{1}\big{)}\widehat{e}_{1}-\big{(}C_{31}^{3}+\mu_{i}C_{12}^{3}\big{)}u_{i}.$ (2.17) Applying $H_{j}$ to both sides of (2.17) and using two equations (2.14) and (2.16) for Hamiltonian functions, we get $\displaystyle\big{[}\widehat{e}_{1},u_{i}\big{]}(H_{j})=-\big{(}C_{31}^{3}+\mu_{i}C_{12}^{3}\big{)}\frac{\|v\|}{\alpha_{i}}.$ Therefore, our integrability condition for Hamiltonian functions becomes $\displaystyle\widehat{e}_{1}\cdot\nabla\left(\frac{\|v\|}{\alpha_{i}}\right)=-\big{(}C_{31}^{3}+\mu_{i}C_{12}^{3}\big{)}\frac{\|v\|}{\alpha_{i}},$ hence, $\displaystyle\widehat{e}_{1}\cdot\nabla\ln\left(\frac{\alpha_{i}}{\|v\|}\right)=\mu_{i}C_{12}^{3}+C_{31}^{3}$ (2.18) and the proposition follows. ∎ ###### Corollary 2.4. The pair of Poisson structures $J_{i}=\alpha_{i}\big{(}\widehat{e}_{2}+\mu_{i}\widehat{e}_{3}\big{)}$ where $\alpha_{i}$’s are defined by (2.18) and $\mu_{i}$’s are defined by (2.5) are compatible. ###### Proof. What we need is to show that (2.8) is satisfied. Indeed, writing (2.18) for $\alpha_{i}$ and $\alpha_{j}$ and subtracting the second from the first, the corollary follows. ∎ Note that, for a pair of compatible Poisson structures, $J_{1}$ and $J_{2}$, the dilatation symmetry $J\rightarrow fJ$ and the additive symmetry $J_{1}+J_{2}$ do not imply that $J_{1}+fJ_{2}$ is a Poisson structure. Indeed, if we apply the Jacobi identity condition and using triple vector identity $\displaystyle(J_{1}+fJ_{2})\cdot\nabla\times(J_{1}+fJ_{2})=-\nabla f\cdot(J_{1}\times J_{2})=0,$ which implies that $\displaystyle\widehat{e}_{1}\cdot\nabla f=0.$ Now we try to describe the relation between the pair of compatible Poisson structures and Hamiltonian functions. But first, we need the following lemma to describe this relation. ###### Lemma 2.5. For the bi-Hamiltonian system with a pair of compatible Poisson structures defined above, $\displaystyle\nabla\cdot\widehat{e}_{1}=\widehat{e}_{1}\cdot\nabla\ln\frac{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}{\|v\|^{2}}.$ ###### Proof. Adding the equations for integrability conditions of Hamiltonian functions (2.18) for $i=1,2$, we get $\displaystyle\widehat{e}_{1}\cdot\nabla\ln(\alpha_{1}\alpha_{2})=\widehat{e}_{1}\cdot\nabla\ln\big{(}\|v\|^{2}\big{)}+2C_{31}^{3}+(\mu_{1}+\mu_{2})C_{12}^{3}.$ (2.19) On the other hand, subtracting the equations (2.5) satisfied by $\mu_{1}$ and $\mu_{2}$, and dividing by $(\mu_{2}-\mu_{1})$, $\displaystyle\widehat{e}_{1}\cdot\nabla\ln(\mu_{2}-\mu_{1})=-\big{(}C_{31}^{3}+C_{12}^{2}\big{)}-(\mu_{1}+\mu_{2})C_{12}^{3}.$ (2.20) Adding (2.19) to (2.20) and using $\displaystyle\nabla\cdot\widehat{e}_{1}=C_{i1}^{i},$ we get $\displaystyle\widehat{e}_{1}\cdot\nabla\ln(\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1}))=\widehat{e}_{1}\cdot\nabla\ln\big{(}\|v\|^{2}\big{)}+\nabla\cdot\widehat{e}_{1},$ and the lemma follows. ∎ ###### Proposition 2.6. Given a bi-Hamiltonian system with a pair of compatible Poisson structures, there exists a canonical pair of compatible Poisson structures $K_{1}$, $K_{2}$ with the same Hamiltonian functions $H_{1}$, $H_{2}$ such that $\displaystyle K_{i}=(-1)^{i+1}\phi\nabla H_{i},$ where $\displaystyle\phi=\frac{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}{\|v\|}.$ ###### Proof. Since Poisson vector fields are linearly independent, one could write Hamiltonians in terms of Poisson vector fields as $\displaystyle\nabla H_{i}=\sigma_{i}^{j}J_{j}.$ By using (1.5), we get $\displaystyle\sigma_{2}^{2}=-\sigma_{1}^{1}=\frac{\|v\|}{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}.$ On the other hand, we have $\displaystyle\nabla\times\nabla H_{i}=\nabla\sigma_{i}^{j}\times J_{j}+\sigma_{i}^{j}\nabla\times J_{j}=0.$ Taking the dot product of both sides with $J_{1}$ and $J_{2}$, and using the compatibility condition, we obtain $\displaystyle\widehat{e}_{1}\cdot\nabla\ln\sigma_{j}^{i}=\frac{J_{1}\cdot(\nabla\times J_{2})}{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}.$ (2.21) Inserting (2.13) into (2.10) and using (2.21), $\displaystyle\widehat{e}_{1}\cdot\nabla\ln\sigma_{j}^{i}=-\widehat{e}_{1}\cdot\nabla\ln\phi,$ which leads to $\displaystyle\sigma_{j}^{i}=\frac{\Psi_{j}^{i}}{\phi},$ where $\displaystyle\widehat{e}_{1}\cdot\nabla\Psi_{j}^{i}=0.$ Therefore, we have $\displaystyle\nabla H_{1}=\frac{1}{\phi}\big{(}\Psi_{1}^{1}J_{1}+\Psi_{1}^{2}J_{2}\big{)},\qquad\nabla H_{2}=\frac{1}{\phi}\big{(}\Psi_{2}^{1}J_{1}-\Psi_{1}^{1}J_{2}\big{)}.$ (2.22) Inserting (2.22) into (1.5), we get $\displaystyle\Psi_{1}^{1}=-1,$ and finally, $\displaystyle\nabla H_{1}=-\frac{\|v\|}{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}\big{(}J_{1}-\Psi_{1}^{2}J_{2}\big{)},\qquad\nabla H_{2}=\frac{\|v\|}{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}\big{(}\Psi_{2}^{1}J_{1}+J_{2}\big{)}.$ Note that, $\displaystyle\nabla H_{1}\times\nabla H_{2}=-\big{(}1+\Psi_{2}^{1}\Psi_{1}^{2}\big{)}\frac{\|v\|^{2}}{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}\widehat{e}_{1}.$ (2.23) For the Hamiltonians to be functionally independent, r.h.s. of (2.23) must not vanish, i.e., $\displaystyle 1+\Psi_{2}^{1}\Psi_{1}^{2}\neq 0.$ Now let us define $\displaystyle K_{1}=\frac{J_{1}-\Psi_{1}^{2}J_{2}}{1+\Psi_{2}^{1}\Psi_{1}^{2}}=-\frac{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}{\big{(}1+\Psi_{2}^{1}\Psi_{1}^{2}\big{)}\|v\|}\nabla H_{1},\qquad K_{2}=\frac{J_{2}+\Psi_{2}^{1}J_{1}}{1+\Psi_{2}^{1}\Psi_{1}^{2}}=\frac{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}{\big{(}1+\Psi_{2}^{1}\Psi_{1}^{2}\big{)}\|v\|}\nabla H_{2}.$ By (1.5), we get $\displaystyle K_{1}\times\nabla H_{1}=K_{2}\times\nabla H_{2}=0,\qquad K_{2}\times\nabla H_{1}=K_{1}\times\nabla H_{2}=v.$ Choosing $K_{i}$’s to be our new Poisson vector fields, the proposition follows. ∎ Consequently, we can write the local existence theorem of bi-Hamiltonian systems in three dimensions. ###### Theorem 2.7. Any three-dimensional dynamical system $\displaystyle\dot{x}(t)=v(x(t))$ (2.24) has a pair of compatible Poisson structures $\displaystyle J_{i}=\alpha_{i}\big{(}\widehat{e}_{2}+\mu_{i}\widehat{e}_{3}\big{)},$ in which $\mu_{i}$’s are determined by the equation $\displaystyle\widehat{e}_{1}\cdot\nabla\mu_{i}=-C_{31}^{2}-\mu_{i}\big{(}C_{31}^{3}+C_{12}^{2}\big{)}-\mu_{i}^{2}C_{12}^{3},$ and $\alpha_{i}$’s are determined by the equation $\displaystyle\widehat{e}_{1}\cdot\nabla\ln\frac{\alpha_{i}}{\|v\|}=C_{31}^{3}+\mu_{i}C_{12}^{3}.$ Furthermore, (2.24) is a locally bi-Hamiltonian system with a pair of local Hamiltonian functions determined by $\displaystyle J_{i}=(-1)^{i+1}\phi\nabla H_{i},$ (2.25) where $\displaystyle\phi=\frac{\alpha_{1}\alpha_{2}(\mu_{2}-\mu_{1})}{\|v\|}.$ (2.26) ## 3 Global existence of compatible bi-Hamiltonian structure in 3D In this section, we investigate the conditions for which the local existence theorem holds globally. To study the global properties of the vector field $\boldsymbol{v}$ by topological means, we relate the vector field with its normal bundle. Let $E$ be the one-dimensional subbundle of $TM$ generated by $v$. Let $Q=TM/E$ be the normal bundle of $v$. By using the cross product with $\widehat{e}_{1}$, we can define a complex structure $\Lambda$ on the fibers of $Q\rightarrow M$, and $Q$ becomes a complex line bundle over $M$. ### 3.1 Bi-Hamiltonian structure in 3D with differential forms In order to obtain and express the obstructions to the global existence of bi- Hamiltonian structures on orientable three manifolds by certain cohomology groups and characteristic classes, we will reformulate the problem by using differential forms. For this purpose, let $\boldsymbol{\Omega}$ be the volume form associated to the Riemannian metric $\boldsymbol{g}$ of $M$. Then, there is a local one-form $\boldsymbol{J}$ associated with a local Poisson bivector field $\mathcal{J}$, $\displaystyle\boldsymbol{J}=\imath_{\mathcal{J}}\boldsymbol{\Omega},$ which is called the local Poisson one-form. The bi-Hamiltonian system (1.5) can be written as $\displaystyle\iota_{v}\boldsymbol{\Omega}=\boldsymbol{J}_{1}\wedge{\rm d}H_{2}=\boldsymbol{J}_{2}\wedge{\rm d}H_{1}.$ (3.1) Note that, although the l.h.s. of this equality is globally defined, r.h.s. is defined only locally, therefore it holds only locally. Now the Jacobi identity is given by $\displaystyle\boldsymbol{J}_{i}\wedge{\rm d}\boldsymbol{J}_{i}=0\qquad\text{for}\quad i=1,2,$ (3.2) and compatibility amounts to $\displaystyle\boldsymbol{J}_{1}\wedge{\rm d}\boldsymbol{J}_{2}=-\boldsymbol{J}_{2}\wedge{\rm d}\boldsymbol{J}_{1}.$ By (2.25), $\boldsymbol{J}_{1}$ and $\boldsymbol{J}_{2}$ can be chosen to be proportional to ${\rm d}H_{1}$ and ${\rm d}H_{2}$, respectively, and hence (3.1) takes the form $\displaystyle\iota_{v}\boldsymbol{\Omega}=\phi{\rm d}H_{1}\wedge{\rm d}H_{2}.$ The Jacobi identity for Poisson 1-forms (3.2) implies the existence of 1-forms $\boldsymbol{\beta}_{i}$ such that $\displaystyle{\rm d}\boldsymbol{J}_{i}=\boldsymbol{\beta}_{i}\wedge\boldsymbol{J}_{i}$ (3.3) for each $i=1,2$. In the next proposition we are going to show that the compatibility of Poisson structures allows us to combine $\boldsymbol{\beta}_{1}$ and $\boldsymbol{\beta}_{2}$ into a single one. ###### Proposition 3.1. There is a $1$-form $\boldsymbol{\beta}$ such that $\displaystyle{\rm d}\boldsymbol{J}_{i}=\boldsymbol{\beta}\wedge\boldsymbol{J}_{i}$ for each $i=1,2$. ###### Proof. Applying (3.3) to the compatibility condition $\displaystyle\boldsymbol{J}_{1}\wedge{\rm d}\boldsymbol{J}_{2}+\boldsymbol{J}_{2}\wedge{\rm d}\boldsymbol{J}_{1}=0,$ we get $\displaystyle(\boldsymbol{\beta}_{1}-\boldsymbol{\beta}_{2})\wedge\boldsymbol{J}_{1}\wedge\boldsymbol{J}_{2}=0,$ which implies that $\displaystyle\boldsymbol{\beta}_{1}-\boldsymbol{\beta}_{2}=b_{1}\boldsymbol{J}_{1}+b_{2}\boldsymbol{J}_{2},$ and therefore, we define $\displaystyle\boldsymbol{\beta}=\boldsymbol{\beta}_{1}-b_{1}\boldsymbol{J}_{1}=\boldsymbol{\beta}_{2}+b_{2}\boldsymbol{J}_{2}.$ Hence $\displaystyle\boldsymbol{\beta}\wedge\boldsymbol{J}_{i}=\boldsymbol{\beta}_{i}\wedge\boldsymbol{J}_{i}={\rm d}\boldsymbol{J}_{i},$ and the proposition follows. ∎ Note that $\boldsymbol{\beta}$ is a $TM$-valued 1-form. Namely, $\displaystyle\iota_{\widehat{e}_{1}}\boldsymbol{\beta}\neq 0$ in general. Now we are going to show that by an appropriate change of Poisson forms, we may reduce it to a connection 1-form on $Q$. ###### Lemma 3.2. $\displaystyle\iota_{\widehat{e}_{1}}\boldsymbol{\beta}=\iota_{\widehat{e}_{1}}({\rm d}\ln\phi),$ where $\phi$ is the function defined in (2.26). ###### Proof. For the proof, we carry out the computation with Poisson vector fields, then transform the result to differential forms. The Jacobi identity (1.4) implies that $\nabla\times J_{i}$ is orthogonal to $J_{i}$ and therefore, we get $\displaystyle\nabla\times J_{i}=a_{i1}\widehat{e}_{1}+a_{i2}\widehat{e}_{1}\times J_{i}.$ (3.4) By the definition of Poisson vector fields, we have $\displaystyle J_{1}\times J_{2}=\phi\|v\|\widehat{e}_{1}.$ We can rewrite (3.4) in the form $\displaystyle\nabla\times J_{i}=\frac{a_{i1}}{\phi\|v\|}J_{1}\times J_{2}+a_{i2}\widehat{e}_{1}\times J_{i}.$ (3.5) Using the compatibility condition (2.9), we obtain $\displaystyle a_{i1}=(\nabla\times J_{i})\cdot\widehat{e}_{1},\qquad a_{i2}=\frac{(\nabla\times J_{1})\cdot J_{2}}{\phi\|v\|}.$ Now we define $\displaystyle\xi=\frac{a_{21}J_{1}-a_{11}J_{2}+((\nabla\times J_{1})\cdot J_{2})\widehat{e}_{1}}{\phi\|\overrightarrow{v}\|},$ and (3.5) becomes $\displaystyle\nabla\times J_{i}=\xi\times J_{i}.$ After a bit of computation it is possible to show that $\displaystyle\xi=\nabla\ln\phi+\widehat{e}_{1}\times\left(\frac{[\widehat{e}_{1}\times J_{1},\widehat{e}_{1}\times J_{2}]}{\phi\|v\|}-\widehat{e}_{1}\times\nabla\ln\|v\|\right).$ Hence, we have $\displaystyle\widehat{e}_{1}\cdot\xi=\widehat{e}_{1}\cdot\nabla\ln\phi$ and defining $\displaystyle\boldsymbol{\beta}=\ast\iota_{\xi}\boldsymbol{\Omega},$ the lemma follows. ∎ Now we define new Poisson 1-forms $K_{i}$ $\displaystyle\boldsymbol{J}_{i}=\phi\boldsymbol{K}_{i}.$ Taking the exterior derivatives of both sides $\displaystyle{\rm d}\boldsymbol{J}_{i}={\rm d}\phi\wedge\boldsymbol{K}_{i}+\phi{\rm d}\boldsymbol{K}_{i}=\boldsymbol{\beta}\wedge\phi\boldsymbol{K}_{i}$ and dividing both sides by $\phi$, $\displaystyle{\rm d}\boldsymbol{K}_{i}=(\boldsymbol{\beta}-{\rm d}\ln\phi)\wedge\boldsymbol{K}_{i}.$ Let $\displaystyle\boldsymbol{\gamma}=\boldsymbol{\beta}-{\rm d}\ln\phi.$ Now, by the lemma above, $\displaystyle\iota_{\widehat{e}_{1}}\boldsymbol{\gamma}=\iota_{\widehat{e}_{1}}\boldsymbol{\beta}-\iota_{\widehat{e}_{1}}({\rm d}\ln\phi)=0,$ (3.6) therefore, $\displaystyle{\rm d}\boldsymbol{K}_{i}=\boldsymbol{\gamma}\wedge\boldsymbol{K}_{i},$ (3.7) where $\gamma$ is a connection on $Q$. ### 3.2 The first obstruction: the Chern class of $\boldsymbol{Q}$ Now we try to find conditions for which a nonvanishing vector field $v$ satisfies $\displaystyle\boldsymbol{w}=\iota_{v}\boldsymbol{\Omega}=\phi{\rm d}H_{1}\wedge{\rm d}H_{2}$ (3.8) for some globally defined functions $\phi$, $H_{1}$ and $H_{2}$. For a two- form to be decomposed into the form (3.8), first of all, the two-form must be written as a product of two globally defined, linearly independent nonvanishing factors. However, such a decomposition may not exist globally. Then, the question is to decompose $\boldsymbol{w}$ into a product of two globally defined one forms $\boldsymbol{\rho}_{1}$ and $\boldsymbol{\rho}_{2}$ $\displaystyle\boldsymbol{w}=\boldsymbol{\rho}_{1}\wedge\boldsymbol{\rho}_{2}.$ (3.9) Since $v$ is a nonvanishing vector field, $\boldsymbol{w}$ is a $2$-form of constant rank $2$. If we let $S_{\boldsymbol{w}}$ to be the sub-bundle of $TM$ on which $\boldsymbol{w}$ is of maximal rank, then we have $S_{\boldsymbol{w}}\cong Q$ defined above. The following theorem states the necessary and sufficient conditions for the decomposition of a two-form of constant rank $2s$ in the large. ###### Theorem 3.3. Let $\Sigma$ be an $\mathbb{R}^{n}$-bundle over a connected base space $M$. Let $\boldsymbol{w}$ be a $2$-form on $\Sigma$ of constant rank $2s$. Let $S_{\boldsymbol{w}}$ be the subbundle of $\Sigma$ on which $\boldsymbol{w}$ is of maximal rank. $w$ decomposes if and only if * $i)$ $S_{\boldsymbol{w}}$ is a trivial bundle. * $ii)$ The representation of its normalization as a map $w_{1}\colon M\rightarrow{\rm SO}(2s)/{\rm U}(s)$ arising from any trivialization of $S_{\boldsymbol{w}}$ lifts to ${\rm SO}(2s)$ [3]. In our case, when $s=1$, since ${\rm U}(1)\cong{\rm SO}(2)$, then ${\rm SO}(2)/{\rm U}(1)$ is a point and it lifts to ${\rm SO}(2)$ trivially, therefore the second condition in the theorem is satisfied. Hence, the necessary and sufficient condition of decomposition is the triviality of $S_{\boldsymbol{w}}\cong Q$. Since $Q$ is a complex line bundle, it is trivial if and only if $\boldsymbol{c}_{1}(Q)=0$, or equivalently it has a global section. Since the decomposition of the 2-form $\boldsymbol{w}$ into globally defined 1-forms $\boldsymbol{\rho}_{1}$ and $\boldsymbol{\rho}_{2}$ is a necessary condition for the existence of a global bi-Hamiltonian structure, the vanishing of the first Chern class of $Q$ becomes a necessary condition. However, this may not be sufficient since the existence of a decomposition in the form (3.9) may not imply that the factors $\boldsymbol{\rho}_{i}$ satisfy $\displaystyle\boldsymbol{\rho}_{i}\wedge{\rm d}\boldsymbol{\rho}_{i}=0.$ In order to determine the effect of vanishing Chern class on the constructions made so far, we are going to investigate the equation (2.5) defining the Poisson one-forms. Since our Poisson one-forms and related integrability conditions are determined by the local solutions of (2.5), they are defined locally on each chart. Let $\big{\\{}J_{i}^{p}\big{\\}}$ and $\big{\\{}J_{i}^{q}\big{\\}}$ be the Poisson vector fields in charts $(U_{p},x_{p})$ and $(U_{q},x_{q})$ around points $p\in M$ and $q\in M$, respectively. Around the point $p\in M$, the Poisson vectors $\big{\\{}J_{i}^{p}\big{\\}}$ are determined by $\mu_{i}^{p},\alpha_{i}^{p}$ and the local frame $\big{\\{}\widehat{e}_{2}^{p},\widehat{e}_{3}^{p}\big{\\}}$. Given the local frame, we can write (2.5) whose solutions are $\mu_{i}^{p}$’s, and using $\mu_{i}^{p}$’s we can determine $\alpha_{i}^{p}$’s by the equation (2.13). Now, if $\boldsymbol{c}_{1}(Q)=0$, which is a necessary condition for the existence of global bi-Hamiltonian structure, then we have a global section of $Q$, i.e., global vector fields normal to $v$. By using the metric on $M$, normalize this global section of $Q$ and take it as $\widehat{e}_{2}$, then define $\widehat{e}_{3}=\widehat{e}_{1}\times\widehat{e}_{2}$. So we have the global orthonormal frame field $\\{\widehat{e}_{1},\widehat{e}_{2},\widehat{e}_{3}\\}$. In order to understand the relation between local Poisson one-forms obtained in two different coordinate neighborhoods, we first need the following lemmas: ###### Lemma 3.4. If two solutions $\mu_{1}(s)$ and $\mu_{2}(s)$ of the Riccati equation $\displaystyle\frac{{\rm d}\mu_{i}}{{\rm d}s}=-C_{31}^{2}-\mu_{i}\big{(}C_{31}^{3}+C_{12}^{2}\big{)}-\mu_{i}^{2}C_{12}^{3}$ are known, then the general solution $\mu(s)$ is given by $\displaystyle\mu-\mu_{1}=K(\mu-\mu_{2})e^{\int C_{12}^{3}(\mu_{2}-\mu_{1}){\rm d}s},$ where $K$ is an arbitrary constant [7]. ###### Lemma 3.5. If $\boldsymbol{c}_{1}(Q)=0,$ then two pairs of compatible Poisson vector fields $\big{\\{}J_{i}^{p}\big{\\}}$ and $\big{\\{}J_{i}^{q}\big{\\}}$ on $U_{p}$ and $U_{q}$ respectively, are related on $U_{p}\cap U_{q}$ by $\displaystyle\frac{J_{i}^{q}}{\big{\|}J_{i}^{q}\big{\|}}=\frac{J_{i}^{p}}{\big{\|}J_{i}^{p}\big{\|}}.$ ###### Proof. Given the global frame field $\\{\widehat{e}_{2},\widehat{e}_{3}\\}$ defined on coordinate neighborhoods $U_{p}$ and $U_{q}$, Riccati equations for $\mu_{i}$’s can be written as $\displaystyle\widehat{e}_{1}\cdot\nabla\mu_{i}^{r}=(\nabla\times\widehat{e}_{2})\cdot\widehat{e}_{2}+\mu_{i}^{r}\big{(}(\nabla\times\widehat{e}_{2})\cdot\widehat{e}_{3}+(\nabla\times\widehat{e}_{3})\cdot\widehat{e}_{2}\big{)}+\big{(}\mu_{i}^{r}\big{)}^{2}\big{(}\nabla\times\widehat{e}_{3}\big{)}\cdot\widehat{e}_{3}$ for $r=p,q$. Therefore, on $U_{p}\cap U_{q}$, $\mu_{i}^{p}$ and $\mu_{i}^{q}$ are four solutions of the same Riccati equation for $i=1,2$. By the lemma above we have $\displaystyle\mu_{i}^{q}-\mu_{1}^{p}=K_{i}^{pq}\big{(}\mu_{i}^{q}-\mu_{2}^{p}\big{)}e^{\int C_{12}^{3}\big{(}\mu_{2}^{p}-\mu_{1}^{p}\big{)}{\rm d}s}.$ (3.10) Now, using the compatibility condition (2.8), $\displaystyle C_{12}^{3}\big{(}\mu_{2}^{p}-\mu_{1}^{p}\big{)}=\widehat{e}_{1}\cdot\nabla\ln\frac{\alpha_{2}^{p}}{\alpha_{1}^{p}},$ (3.10) becomes $\displaystyle\mu_{i}^{q}-\mu_{1}^{p}=K_{i}^{pq}\big{(}\mu_{i}^{q}-\mu_{2}^{p}\big{)}\frac{\alpha_{2}^{p}}{\alpha_{1}^{p}},$ (3.11) where $\displaystyle\widehat{e}_{1}\cdot\nabla K_{i}^{pq}=0.$ (3.12) Multiplying both sides by $\alpha_{1}^{p}\alpha_{i}^{q}$ in (3.11), gives $\displaystyle J_{i}^{q}\times J_{1}^{p}=K_{i}^{pq}J_{i}^{q}\times J_{2}^{p}.$ (3.13) Rearranging (3.13), we obtain $\displaystyle J_{i}^{q}\times\big{(}J_{1}^{p}-K_{i}^{pq}J_{2}^{p}\big{)}=0.$ Using (3.12) and the compatibility, we can take $\displaystyle\widetilde{J}_{i}^{p}=J_{1}^{p}-K_{i}^{pq}J_{2}^{p}$ to be our new Poisson vector fields on the neighborhood $U_{p}$, and obtain $\displaystyle J_{i}^{q}\times\widetilde{J}_{i}^{p}=0.$ By compatibility these new Poisson vector fields $\widetilde{J}_{i}^{p}$ produce functionally dependent Hamiltonians and therefore, for the simplicity of notation, we will assume without restriction of generality that $\displaystyle\widetilde{J}_{i}^{p}=J_{i}^{p}$ and the lemma follows. ∎ Then, we have the following result: ###### Theorem 3.6. There exist two linearly independent global sections $\widehat{j}_{i}$ of $Q$ satisfying $\displaystyle\widehat{j}_{i}\cdot\big{(}\nabla\times\widehat{j}_{i}\big{)}=0$ (3.14) if and only if $\boldsymbol{c}_{1}(Q)=0$. ###### Proof. The forward part is trivial since the existence of a global section of the complex line bundle $Q$ implies that $Q$ is trivial, and hence $\boldsymbol{c}_{1}(Q)$ vanishes. For the converse, we define $\displaystyle\widehat{j}_{i}^{p}=\frac{J_{i}^{p}}{\big{\|}J_{i}^{p}\big{\|}}$ and the lemma implies that $j_{i}^{p}=j_{i}^{q}$ on $U_{p}\cap U_{q}$ and the theorem follows. ∎ The lemma above states the reason why one may fail to extend a local pair of compatible Poisson vector fields into a global one, even if $\boldsymbol{c}_{1}(Q)=0$. In order to do so one should have $J_{i}^{q}=J_{i}^{p}$ on $U_{p}\cap U_{q}$. However, not the Poisson vector fields but their unit vector fields can be globalized. Since $\displaystyle\widehat{e}_{1}\cdot\nabla\frac{\big{\|}J_{2}^{p}\big{\|}}{\big{\|}J_{1}^{p}\big{\|}}\neq 0$ in general, they may not lead to a pair of compatible Poisson structures. Now we take $\widehat{j}_{1}$ as our first global Poisson vector field, and check whether we can find another global Poisson vector field compatible with this one by rescaling $\widehat{j}_{2}$. ### 3.3 Second obstruction: Bott class of the complex codimension 1 foliation Since $v$ is a nonvanishing vector field on $M$, it defines a real codimension two foliation on $M$ by orbits of $v$. Since $Q=TM/E$ is a complex line bundle on $M$, this foliation has complex codimension one. Now, by assuming our primary obstruction which is the vanishing of the Chern class, we compute the Bott class of the complex codimension one foliation as defined in [2], which is studied in detail in [1], and then show that the system admits two globally defined compatible Poisson structures if and only if the Bott Class is trivial. For the rest of our work, we will assume that $Q$ and its dual $Q^{\ast}$ are trivial bundles. By (3.14), $Q^{\ast}$ has two global sections $\widehat{\boldsymbol{j}}_{i}=(^{\ast}\imath_{\widehat{j}_{i}}\boldsymbol{\Omega})$ satisfying $\displaystyle{\rm d}\widehat{\boldsymbol{j}}_{i}=\boldsymbol{\Gamma}_{i}\wedge\widehat{\boldsymbol{j}}_{i}$ (3.15) for globally defined $\boldsymbol{\Gamma}_{i}$’s. These $\widehat{\boldsymbol{j}}_{i}$’s are related with the local Poisson one-forms $\boldsymbol{J}_{i}^{p}$ by $\displaystyle\boldsymbol{J}_{i}^{p}=\big{\|}\boldsymbol{J}_{i}^{p}\big{\|}\widehat{\boldsymbol{j}}_{i}.$ (3.16) By (3.7), we have $\displaystyle{\rm d}\boldsymbol{J}_{i}^{p}=\boldsymbol{\gamma}^{p}\wedge\boldsymbol{J}_{i}^{p}.$ (3.17) Inserting (3.16) and (3.17) into (3.15), we also have $\displaystyle{\rm d}\widehat{\boldsymbol{j}}_{i}=\big{(}\boldsymbol{\gamma}^{p}-{\rm d}\ln\big{\|}\boldsymbol{J}_{i}^{p}\big{\|}\big{)}\wedge\widehat{\boldsymbol{j}}_{i}.$ (3.18) Redefining $\boldsymbol{\Gamma}_{i}$’s if necessary, comparing (3.15) with (3.18), we get $\displaystyle\boldsymbol{\Gamma}_{i}=\boldsymbol{\gamma}^{p}-{\rm d}\ln\big{\|}\boldsymbol{J}_{i}^{p}\big{\|}.$ (3.19) ###### Proposition 3.7. Let $\boldsymbol{\kappa}$ be the curvature two-form of $Q$. There exists a compatible pair of global Poisson structures if and only if $\displaystyle\boldsymbol{\Xi}=(\boldsymbol{\Gamma}_{1}-\boldsymbol{\Gamma}_{2})\wedge\boldsymbol{\kappa}$ is exact. ###### Proof. Since $\widehat{\boldsymbol{j}}_{1}$ and $\widehat{\boldsymbol{j}}_{2}$ may not be compatible, we introduce a local Poisson form $\boldsymbol{j}^{p}$ defined on the coordinate neighborhood $U_{p}$ of $p\in M$, which is compatible with $\widehat{\boldsymbol{j}}_{1}$ and parallel to $\widehat{\boldsymbol{j}}_{2}$ i.e., $\displaystyle\boldsymbol{j}^{p}=f^{p}\widehat{\boldsymbol{j}}_{2}$ (3.20) and $\displaystyle\widehat{\boldsymbol{j}}_{1}\wedge{\rm d}\boldsymbol{j}^{p}+\boldsymbol{j}^{p}\wedge{\rm d}\widehat{\boldsymbol{j}}_{1}=0.$ (3.21) Now (3.20) implies that $\displaystyle{\rm d}\boldsymbol{j}^{p}=\big{(}\boldsymbol{\Gamma}_{2}+{\rm d}\ln f^{p}\big{)}\wedge\boldsymbol{j}^{p}.$ (3.22) Putting (3.15) and (3.22) into (3.21) and using (3.20), we get $\displaystyle\big{(}\boldsymbol{\Gamma}_{1}-\boldsymbol{\Gamma}_{2}-{\rm d}\ln f^{p}\big{)}\wedge\widehat{\boldsymbol{j}}_{1}\wedge\boldsymbol{j}^{p}=0$ which implies $\displaystyle(\boldsymbol{\Gamma}_{1}-\boldsymbol{\Gamma}_{2})\wedge\widehat{\boldsymbol{j}}_{1}\wedge\widehat{\boldsymbol{j}}_{2}={\rm d}\ln f^{p}\wedge\widehat{\boldsymbol{j}}_{1}\wedge\widehat{\boldsymbol{j}}_{2}.$ (3.23) Our aim here is to find the obstruction to extending $f^{p}$ to $M$, or for (3.23) to hold globally. For this purpose, we consider the connections on $Q$ defined by $\Gamma_{i}$’s. By (3.19), we define the curvature of these connections to be $\displaystyle\boldsymbol{\kappa}={\rm d}\boldsymbol{\Gamma}_{i}={\rm d}\boldsymbol{\gamma}^{p}.$ Taking the exterior derivative of (3.17) and using (3.16), we get $\displaystyle{\rm d}\boldsymbol{\gamma}^{p}\wedge\boldsymbol{J}_{i}^{p}={\rm d}\boldsymbol{\gamma}^{p}\wedge\widehat{\boldsymbol{j}}_{i}=0,$ which leads to $\displaystyle\kappa={\rm d}\boldsymbol{\gamma}^{p}=\varphi\widehat{\boldsymbol{j}}_{1}\wedge\widehat{\boldsymbol{j}}_{2}.$ (3.24) Now multiplying both sides of (3.23) with $\varphi$, $\displaystyle(\boldsymbol{\Gamma}_{1}-\boldsymbol{\Gamma}_{2})\wedge\boldsymbol{\kappa}={\rm d}\ln f^{p}\wedge\boldsymbol{\kappa}={\rm d}\big{(}\big{(}\ln f^{p}\big{)}\boldsymbol{\kappa}\big{)}$ and the proposition follows. ∎ Now we are going to show that the cohomology class of $\boldsymbol{\Xi}$ vanishes if and only if the Bott class of the complex codimension 1 foliation vanishes. Since $Q$ is a complex line bundle we have $\displaystyle\boldsymbol{c}_{1}(Q)=[\boldsymbol{\kappa}]$ and the vanishing of $\boldsymbol{c}_{1}(Q)$ is a necessary condition $\displaystyle\boldsymbol{c}_{1}={\rm d}\boldsymbol{h}_{1}.$ So we have $\displaystyle\boldsymbol{c}_{1}=[\boldsymbol{\kappa}]=\big{[}{\rm d}\boldsymbol{\gamma}^{p}\big{]},$ which implies that on $U_{p}$ $\displaystyle\boldsymbol{h}_{1}=\boldsymbol{\gamma}^{p}+{\rm d}\ln h^{p}.$ Then, the Bott class [2] becomes $\displaystyle\boldsymbol{h}_{1}\wedge\boldsymbol{c}_{1}=\big{(}\boldsymbol{\gamma}^{p}+{\rm d}\ln h^{p}\big{)}\wedge{\rm d}\boldsymbol{\gamma}^{p}={\rm d}\ln h^{p}\wedge\boldsymbol{\kappa}+\boldsymbol{\gamma}^{p}\wedge{\rm d}\boldsymbol{\gamma}^{p}.$ Now by (3.6) and (3.24) we have $\displaystyle\boldsymbol{\gamma}^{p}\wedge{\rm d}\boldsymbol{\gamma}^{p}=0,$ and therefore, $\displaystyle\boldsymbol{h}_{1}\wedge\boldsymbol{c}_{1}={\rm d}\big{(}\big{(}\ln h^{p}\big{)}\kappa\big{)}.$ Since $\boldsymbol{h}_{1}$ is globally defined, on $U_{p}\cap U_{q}$ we have $\displaystyle\boldsymbol{h}_{1}=\boldsymbol{\gamma}^{p}+{\rm d}\ln h^{p}=\boldsymbol{\gamma}^{q}+{\rm d}\ln h^{q}$ and $\displaystyle\boldsymbol{\gamma}^{p}-\boldsymbol{\gamma}^{q}={\rm d}\ln\frac{h^{q}}{h^{p}}.$ (3.25) Now we have the following theorem: ###### Theorem 3.8. The cohomology class of $\boldsymbol{\Xi}$ vanishes if and only if the Bott class of the complex codimension one foliation defined by the nonvanishing vector field vanishes. ###### Proof. If the Bott class vanishes, then we have a globally defined function $h$ such that $\displaystyle{\rm d}((\ln h)\boldsymbol{\kappa})=0.$ Then, choosing $f=h$ leads to a compatible pair of global Poisson structures. Conversely, if there is a pair of globally defined compatible Poisson structures, then $\boldsymbol{\gamma}$ becomes a global form, and by (3.25) we have $\displaystyle{\rm d}\ln\frac{h^{q}}{h^{p}}=0$ on $U_{p}\cap U_{q}$. Therefore, $\displaystyle\ln h^{q}-\ln h^{p}=c^{qp},$ where $c^{qp}$ is a constant on $U_{p}\cap U_{q}$. Now, fixing a point $x_{0}\in U_{p}\cap U_{q}$ $\displaystyle c^{qp}=\ln h^{q}(x_{0})-\ln h^{p}(x_{0})=\ln c^{q}-\ln c^{p},$ we obtain $\displaystyle\frac{h^{p}}{c^{p}}=\frac{h^{q}}{c^{q}}=h,$ where $h$ is a globally defined function, and $\displaystyle{\rm d}\ln h={\rm d}\ln h^{p}.$ Therefore, $\displaystyle[\boldsymbol{h}_{1}\wedge\boldsymbol{c}_{1}]=[{\rm d}((\ln h)\boldsymbol{\kappa})]=0$ and the theorem follows. ∎ ### Acknowledgements We are indebted to Professor Turgut Önder for his help during this work. We also thank to the anonymous referees for their comments and corrections. ## References * [1] Asuke T., A remark on the Bott class, Ann. Fac. Sci. Toulouse Math. 10 (2001), 5–21. * [2] Bott R., Lectures on characteristic classes and foliations, in Lectures on Algebraic and Differential Topology (Second Latin American School in Math., Mexico City, 1971), Lecture Notes in Math., Vol. 279, Springer, Berlin, 1972, 1–94. * [3] Dibag I., Decomposition in the large of two-forms of constant rank, Ann. Inst. Fourier (Grenoble) 24 (1974), 317–335. * [4] Gümral H., Nutku Y., Poisson structure of dynamical systems with three degrees of freedom, J. Math. Phys. 34 (1993), 5691–5723. * [5] Haas F., Goedert J., On the generalized Hamiltonian structure of $3$D dynamical systems, Phys. Lett. A 199 (1995), 173–179, math-ph/0211035. * [6] Hernández-Bermejo B., New solutions of the Jacobi equations for three-dimensional Poisson structures, J. Math. Phys. 42 (2001), 4984–4996. * [7] Ince E.L., Ordinary differential equations, Dover Publications, New York, 1944. * [8] Laurent-Gengoux C., Pichereau A., Vanhaecke P., Poisson structures, Grundlehren der Mathematischen Wissenschaften, Vol. 347, Springer, Heidelberg, 2013. * [9] Magri F., A simple model of the integrable Hamiltonian equation, J. Math. Phys. 19 (1978), 1156–1162. * [10] Olver P.J., Canonical forms and integrability of bi-Hamiltonian systems, Phys. Lett. A 148 (1990), 177–187. * [11] Santoprete M., On the relationship between two notions of compatibility for bi-Hamiltonian systems, SIGMA 11 (2015), 089, 11 pages, arXiv:1506.08675. * [12] Stiefel E., Richtungsfelder und Fernparallelismus in $n$-dimensionalen Mannigfaltigkeiten, Comment. Math. Helv. 8 (1935), 305–353. * [13] Weinstein A., The local structure of Poisson manifolds, J. Differential Geom. 18 (1983), 523–557.
# A unified framework for coordination of thermostatically controlled loads Austin Coffman<EMAIL_ADDRESS>Ana Bušić<EMAIL_ADDRESS>Prabir Barooah <EMAIL_ADDRESS>University of Florida, Gainesville, FL, USA Inria, Paris, France University of Florida, Gainesville, FL, USA ###### Abstract A collection of thermostatically controlled loads (TCLs) – such as air conditioners and water heaters – can vary their power consumption within limits to help the balancing authority of a power grid maintain demand supply balance. Doing so requires loads to coordinate their on/off decisions so that the aggregate power consumption profile tracks a grid-supplied reference. At the same time, each consumer’s quality of service (QoS) must be maintained. While there is a large body of work on TCL coordination, there are several limitations. One is that they do not provide guarantees on the reference tracking performance and QoS maintenance. A second limitation of past work is that they do not provide a means to compute a suitable reference signal for power demand of a collection of TCLs. In this work we provide a framework that addresses these weaknesses. The framework enables coordination of an arbitrary number of TCLs that: (i) is computationally efficient, (ii) is implementable at the TCLs with local feedback and low communication, and (iii) enables reference tracking by the collection while ensuring that temperature and cycling constraints are satisfied at every TCL at all times. The framework is based on a Markov model obtained by discretizing a pair of Fokker-Planck equations derived in earlier work by Malhame and Chong [21]. We then use this model to design randomized policies for TCLs. The balancing authority broadcasts the same policy to all TCLs, and each TCL implements this policy which requires only local measurement to make on/off decisions. Simulation results are provided to support these claims. ###### keywords: Distributed control, Grid support, Randomized control, Thermostatically controlled loads. ††thanks: This paper was not presented at any IFAC meeting. Corresponding author A. Coffman. The research reported here has been partially supported by the NSF through awards 1646229 (CPS-ECCS) and 1934322 (CPS-ECCS), and the French National Research Agency grant ANR-16-CE05-0008. , , ###### Contents 1. 1 Introduction 1. 1.1 Literature review and contribution 2. 1.2 Notation 2. 2 Modeling: Individual TCL 1. 2.1 Temperature dynamics of TCLs 1. 2.1.1 Policy (at the TCL) 2. 2.2 PDE model 3. 3 Markov model from PDE Discretization 1. 3.1 Spatial discretization 2. 3.2 Temporal discretization 4. 4 Discrete space model of a TCL: structure and grid friendly policies 1. 4.1 Discrete state space 2. 4.2 Conditional independence in $P_{k}$ 1. 4.2.1 Constructing the factorization 3. 4.3 BA control command $=$ policy 5. 5 Proposed framework 1. 5.1 Individual TCL model with cycling 2. 5.2 Aggregate model of a collection of TCLs 1. 5.2.1 Evaluating the aggregate model 3. 5.3 Grid support Policy design 1. 5.3.1 Convex control synthesis 2. 5.3.2 Computational considerations 3. 5.3.3 Communication burden 6. 6 Numerical experiments 1. 6.1 Planning 2. 6.2 Real time control 7. 7 Conclusion 8. A Proofs 1. A.1 Proof of Lemma 1 2. A.2 Proof of Lemma 3 3. A.3 Proof of Lemma 4 4. A.4 Proof of Theorem 1 1. A.4.1 $\eta^{*}_{\text{CVX}}\leq\eta^{*}_{\text{NCVX}}$ 2. A.4.2 $\eta^{*}_{\text{NCVX}}\leq\eta^{*}_{\text{CVX}}$ 9. B PDE discretization 1. B.1 Internal CV’s 2. B.2 Boundary CV’s 1. B.2.1 Additional conditions 3. B.3 Overall system ## 1 Introduction Many loads are flexible in their power demand: they can vary their demand around a baseline without adversely affecting consumers’ quality of service (QoS). The flexibility can be used by a balancing authority (BA) to balance supply and demand in a power grid. The baseline demand refers to the power demand under normal operation, when each load operates only to meet its consumer’s QoS without any interference from the BA. Since the rated power of each load is small, it is necessary to use a collection of loads. To provide grid support, the collection has to vary its demand from its baseline. It is envisioned that the BA would supply a reference signal for power demand and the actions of the loads in a collection would be coordinated so that their total demand tracks this reference. Thermostatically controlled loads (TCLs) - such as residential air conditioners, heat pumps, and water heaters - are recognized to be valuable sources of flexible demand [4, 6, 23, 18]. For an air conditioner or a heat pump, baseline demand is largely dictated by ambient weather conditions. There are at least two QoS requirements: the indoor temperature must be maintained within a prespecified range and compressor short-cycling must be avoided, meaning, once the compressor turns on it cannot turn off until a prespecified time period elapses, and vice versa. Coordination of TCLs involves two conflicting requirements: (i) the TCLs collectively need to track the reference power demand signal, and (ii) every TCL’s QoS need to be maintained. The actuation at each TCL is discrete: it can either be on or off. Direct load control [7], in which a centralized controller at the BA directly commands on/off status of each TCL is not scalable to large populations. A more scalable idea, that subsequent works on TCL coordination use, is for the BA to broadcast a low dimensional control command to all TCLs, which is translated by each TCL into its actuation command with a local policy. To avoid confusion between the decision making at the BA and a TCL, we use the word “policy” to mean the algorithm at a TCL that makes on/off decisions. The literature on decentralized coordination of TCLs differ in their choice of the broadcast signal (i.e., BA’s control command) and the policy at the TCL that translates this broadcast to on/off decisions. Coordination architectures can be divided into two broad categories based on these choices: (i) thermostat set point change [4, 1] and (ii) probabilistic control [23, 20, 6, 9]. These are discussed in more detail in Section 1.1. A framework for coordinating TCLs needs two parts. The coordination scheme is one part. The other part is reference computation: the framework must provide the BA with a method to determine a suitable reference signal for the TCLs. That is, the reference must be such that the TCLs can collectively track the signal while each TCL maintains its QoS. Otherwise, even the best coordination scheme will fail to meet either the BA’s need, which is reference tracking, or the consumers’ need, which is maintaining indoor temperature etc., or both. This work presents a unified framework for coordination of a collection of TCLs for providing grid support services. The framework enables both of the above mentioned components, i.e., (i) planning a suitable reference for a collection of TCLs and (ii) designing a randomized policy for coordination of the individual TCLs, so that both the BA’s requirement and consumers’ QoS are satisfied. In the proposed framework, the BA computes randomized control policies for the TCLs and broadcasts them to all the TCLs. Each TCL receives the same policy and implements it using locally measurable information. The framework is computationally tractable for an arbitrary number of TCLs. The communication burden is low: only a few numbers need to be broadcast by the BA at every sampling instant. Feedback from TCLs to the BA can be infrequent. Underlying the framework is: (i) a Markov chain model that is derived from partial differential equations developed in the early work of Malhame and Chong [21], (ii) state augmentation to incorporate cycling constraints, and (iii) convexification of the non-convex problem that appears in the design of the randomized control policy for the individual TCL. Additionally, we show that the assumption made about the effect of weather in earlier work [3] on randomized control, under certain conditions, is in fact true. ### 1.1 Literature review and contribution Before reviewing coordination methods, we discuss two interrelated modeling approaches that underpin many of the ideas in the TCL control architectures. These are the Markov chain and partial differential equation (PDE) models [16, 21, 26, 17, 28, 24], which stem from the early work of Malhame and Chong [21]. In [21] a pair of coupled Fokker-Planck equations are developed to model a collection of TCLs under thermostat control. The Fokker-Planck equations are PDEs that describe the time evolution of a certain probability density functions (pdf) over the state space of temperature and on/off mode. The PDEs can be used to model the entire collection or a single TCL: the probability that a single TCL is “on” is approximately the fraction of TCLs that are “on”. Discretizing the PDE yields a Markov chain model, though some works have obtained Markov models without using the PDEs. Hence, _one_ set of PDEs can model a collection of TCLs. Thus, methods that base control design on the PDE or Markov chain framework scales well with the number of TCLs. Due to the lack of scalability of direct load control, we limit our attention to the two broad classes mentioned earlier: (i) thermostat set point changes, (ii) probabilistic policy. There are many forms of probabilistic policy, which can be roughly subdivided into two sub categories: (ii-A) “bin switching” and (ii-B) “randomized policy”. We discuss these in detail below. In the thermostat setpoint change coordination architecture, a time-varying thermostat set point is broadcast to all TCLs, and each TCL makes on/off decisions based on this new setpoint [4, 1]. This approach may ask for an extremely small change in thermostat setpoint, far below the resolution of the temperature sensor at each TCL. Or it may ask for large changes in thermostat setpoint which will violate occupant comfort. In a probabilistic policy architecture, the TCL policy - the mapping from BA’s broadcast command to a TCL’s on/off decision - is a non-deterministic mapping. Works in this category typically first model the population of TCLs under thermostat control, which is a deterministic policy, as a Markov chain. The continuous temperature range is divided into a number of discrete bins. A finite dimensional state vector, a probability mass function, is then defined. Each entry of the state vector represents “the fraction of TCLs that are on (or off) and has temperature in a certain range.” Since the basic Markov model is derived for the thermostat policy, introduction of the BA’s control to manipulate TCLs’ on/off state is somewhat ad-hoc. In the the bin switching literature, the control command from the BA is chosen so as to affect the fraction of TCLs in the temperature bins directly. In [23], the BA’s control command is chosen to be another vector, whose $i^{\text{th}}$ entry represents “the fraction of TCLs in bin $i$ to increase/decrease”. A policy is then proposed to translate this command to on/off action at each TCL, which requires knowledge of the state of the Markov model. In [20], BA’s control command is chosen to be a scalar. The probability of a TCL turning on or off is proportional to this scalar. Subsequent works have proposed various refinements, such as BA’s command affecting the rate of fractions to switch instead of fraction to switch [26]. Providing performance guarantees with bin switching architecture has proved challenging, either on reference tracking or on QoS maintenance for individual TCLs. An alternative to bin switching that still uses probabilistic on/off decision making is randomized policy [3, 6]. A randomized policy is a specification of the conditional probability of turning on or off given the current state. On/off decisions are computed with the help of a random number generator and the policy. In this architecture it is envisioned that the thermostat policy at the TCL is replaced with a randomized policy. In [3, 6], the policy is parameterized by a scalar $\zeta(t)$. Coordination of the population is then achieved by appropriate design of $\zeta(t)$, which is computed and broadcast by the BA. This architecture also uses a Markov model of the evolution of binned temperature, but assumes a certain factorization: the next values of the temperature and mode are conditionally independent given the current joint pair of temperature and mode values under the effects of the randomized policy and exogenous disturbances, especially weather. That is, the transition matrix of the state process is a point wise product of two controlled transition matrices. In an optimal control setting, computation of the BA’s control command, $\zeta(t)$, for reference tracking is a non-convex optimization problem [11]. The probability of turning on when temperature exceeds the upper limit, or off when temperature dips below the lower limit, is set to 1 by design. This will ensure the temperature constraint is maintained. Attempts have been made to maintain the cycling constraint [9]. But a formal design method to incorporate the cycling constraint has been lacking. A complete framework for coordination of TCL needs not only a control algorithm to make decisions at TCLs, but also a method to compute a _feasible_ reference signal for the collection’s power demand. Feasible means that no TCL needs to violate local constraints in order for the collection to track the reference. The topic is sometimes described as “flexibility capacity” and has been examined in many recent works, with various definitions of flexibility [15, 25, 13, 10]. A unified treatment of reference design and coordination algorithm design that would provide a complete framework is lacking. In short, existing work on TCL coordination has a number of scattered disadvantages. Direct load control suffers from scalability/privacy issues and thermostat set-point methods have implementation issues. Bin switching does not provide guarantees on reference tracking and often requires solving a challenging state estimation problem. Prior work on randomized control requires non-convex optimization and is based on an assumed conditional independence. Finally, there is a lack of unified treatment of the reference design and policy design problems. In this work we develop a unified framework for coordination of TCLs that addresses the weaknesses of prior work described above. Our major contributions are as follows. 1. 1. We provide a complete framework that allows the BA to compute (a) an optimal reference signal that is feasible for the collection and (b) optimal randomized policies for the TCLs. When the TCLs implement these policies, their total power demand collectively tracks the reference signal and the policies guarantee that temperature and cycling QoS requirements at each TCL are satisfied. Optimal reference means it is closest to what the BA wants while being feasible for the TCLs. Implementation of the policy at a TCL is easy; it requires only local measurements. The communication burden for coordination is also low. At each sampling time, a randomized control policy - parameterized by a few numbers - is broadcast to all TCLs. Feedback from TCLs to the BA can be infrequent. 2. 2. Our framework is based on a careful discretization of the partial differential equation (PDE) model described in [21]. This discretization shows that a certain “conditional independence” that was assumed in [3] indeed holds. The conditional independence separates the effects of the policy at the TCL (control) and weather (disturbance) on the transition matrix, and greatly facilitates computation of policies. 3. 3. Numerical experiments are provided to illustrate the efficacy of the framework. Simulations show that TCLs are able to track the optimal reference collectively while each TCL is able to maintain both temperature and cycling constraints. Matlab implementation is made publicly available at [8]. Figure 1 illustrates the two parts of the proposed framework. Figure 1: Coordination architecture with the proposed framework. The Markov model obtained by discretizing a PDE was presented in [12]. For completeness, we include the discretization in this paper as an Appendix. ### 1.2 Notation The symbol $\mathbb{1}$ denotes the vector of all ones, $\mathbf{e}_{i}$ denotes the ith canonical basis vector, and $\mathbf{0}$ denotes the zero matrix or vector, all of appropriate dimension. For a vector $v$, $\text{diag}(v)$ denotes the diagonal matrix with entries of $v$, i.e., $\text{diag}(v)\mathbb{1}=v$. Further, $\otimes$ denotes matrix Kronecker product and $\mathbf{I}_{A}(\cdot)$ the indicator function of the set $A$. ## 2 Modeling: Individual TCL A thermostatically controlled load (TCL) is an on/off device that ensures the temperature of a given environment remains within a specified region, e.g., an air conditioner. During its operation, the TCL must adhere to certain operational requirements (QoS constraints). We consider two: the temperature constraint and the cycling constraint. The temperature constraint is that the TCL’s temperature must remain within a prespecified deadband, $[\lambda^{\min},\lambda^{\max}]$. This is achieved by switching the TCL on or off when it is too hot or cold. The cycling constraint is that the TCL can only change from “on” to “off” or vice versa once every $\tau$ (discrete) time instants, where $\tau$ is a prespecified constant. The cycling constraint is to ensure the mechanical hardware is not damaged. In both cases, ensuring the two constraints amounts to appropriately deciding when to switch the TCL on or off. ### 2.1 Temperature dynamics of TCLs The typical model for the TCL’s temperature $\theta(t)$ in the literature is the following ordinary differential equation (ODE), $\displaystyle\begin{split}\frac{d}{dt}\theta(t)&=f_{m}(\theta,t),\quad\text{with}\\\ f_{m}(\theta,t)&=-\frac{1}{RC}\left(\theta-\theta^{a}(t)\right)-m(t)\frac{\eta P_{0}}{C}.\end{split}$ (1) The rated electrical power consumption is denoted $P_{0}$ with coefficient of performance (COP) $\eta$. The parameters $R$ and $C$ denote thermal resistance and capacitance, respectively. The signal $\theta^{a}(t)$ is the ambient temperature. The quantity $m(t)$ is the on/off mode, and in the following we identify $m(t)=1$ and $m(t)=$ on, as well as $m(t)=0$ and $m(t)=$ off. We denote arbitrary temperature values through the variable $\lambda$, and the thermostat setpoint as $\lambda^{\text{set}}$. The values $\lambda^{\text{max}}$ and $\lambda^{\text{min}}$ set the upper and lower limit for the temperature deadband. A model for the temperature state that accounts for modeling errors in (1) and will be crucial in developing the content in Section 2.2 is the following Itô stochastic differential equation (SDE), $\displaystyle d\theta(t)=f_{m}(\theta,t)dt+\sigma dB(t).$ (2) The term $B(t)$ is Brownian motion with parameter $\sigma>0$, and the quantity $\sigma dB(t)$ captures modeling errors in (1). In either model, the baseline power for the TCL is the value of $P$ so that $f_{1}(\lambda^{\text{set}},t)=0$, solving yields: $\displaystyle\text{Baseline Power:}\quad\bar{P}^{\text{ind}}(t)=\frac{\theta^{a}(t)-\lambda^{\text{set}}}{\eta R}.$ (3) For ${\sf{N_{tcl}}}$ TCLs the baseline power $\bar{P}(t)$ and maximum power $P_{\text{\footnotesize{agg}}}$ are, $\displaystyle\bar{P}(t)\triangleq{\sf{N_{tcl}}}\bar{P}^{\text{ind}}(t),\quad\text{and}\quad P_{\text{\footnotesize{agg}}}\triangleq{\sf{N_{tcl}}}P_{0}.$ (4) The total electrical power consumption of the collection, whether with thermostat policy or some other policy, is denoted by $y(t)$: $\displaystyle y(t)\triangleq P_{0}\sum_{\ell=1}^{{\sf{N_{tcl}}}}m^{\ell}(t)$ (5) where $m^{\ell}(t)$ is the on/off state of the $\ell$-th TCL. #### 2.1.1 Policy (at the TCL) The mode state of a TCL evolves according to a policy. The following policy, which we denote as the _thermostat policy_ , ensures the temperature constraint: $\displaystyle\lim_{\epsilon\rightarrow 0}\ m(t+\epsilon)=\begin{cases}1,&\theta(t)\geq\lambda^{\text{max}}.\\\ 0,&\theta(t)\leq\lambda^{\text{min}}.\\\ m(t),&\text{o.w.}\end{cases}$ (6) We add the following set of assumptions about the individual TCL discussed so far. 1. A.1 The thermostat policy does not violate the cycling constraint. 2. A.2 For all $t\geq 0$ and $\theta\in[\lambda^{\text{min}},\lambda^{\text{max}}]$, $f_{\text{\footnotesize{on}}}(\theta,t)\leq 0$ and $f_{\text{\footnotesize{off}}}(\theta,t)\geq 0$. 3. A.3 The TCL’s cycling and temperature constraint are both simultaneously feasible. The sizing/design of the TCL is most likely to ensure that A.1 holds. With A.1 , we depart from discussing the cycling constraint until Section 5 since up to that point the mode state is assumed to evolve according to (6). Assumption A.2 states that when the TCL is on, the temperature does not increase and when the TCL is off the temperature does not decrease. All prior works focusing on cooling TCLs (e.g., air conditioners) implicitly make this assumption. Every result that is to follow is also valid for heating TCLs (e.g., a water heater or a heat pump) with a sign reversal. Like A.2, assumption A.3 is also implicit in any work that considers both the TCLs temperature and cycling constraint. ### 2.2 PDE model We now describe a PDE model of a TCL’s temperature with thermostat policy originally derived in [21]. Consider the following marginal pdfs $\mu_{\text{\footnotesize{on}}},\mu_{\text{\footnotesize{off}}}$: $\displaystyle\mu_{\text{\footnotesize{on}}}(\lambda,t)d\lambda$ $\displaystyle={\sf P}\left((\lambda<\theta(t)\leq\lambda+d\lambda),\ m(t)=\text{on}\right),$ (7) $\displaystyle\mu_{\text{\footnotesize{off}}}(\lambda,t)d\lambda$ $\displaystyle={\sf P}\left((\lambda<\theta(t)\leq\lambda+d\lambda),\ m(t)=\text{off}\right),$ (8) where ${\sf P}(\cdot)$ denotes probability, $\theta(t)$ evolves according to (2) and for now $m(t)$ evolves according to (6). It was shown in [21] that the densities $\mu_{\text{\footnotesize{on}}}$ and $\mu_{\text{\footnotesize{off}}}$ satisfy the Fokker-Planck equations, $\displaystyle\frac{\partial}{\partial t}\mu_{\text{\footnotesize{on}}}(\lambda,t)$ $\displaystyle=\frac{\sigma^{2}}{2}\nabla^{2}_{\lambda}\mu_{\text{\footnotesize{on}}}(\lambda,t)-\nabla_{\lambda}\Big{(}f_{\text{\footnotesize{on}}}(\lambda,t)\mu_{\text{\footnotesize{on}}}(\lambda,t)\Big{)}$ (9) $\displaystyle\frac{\partial}{\partial t}\mu_{\text{\footnotesize{off}}}(\lambda,t)$ $\displaystyle=\frac{\sigma^{2}}{2}\nabla^{2}_{\lambda}\mu_{\text{\footnotesize{off}}}(\lambda,t)-\nabla_{\lambda}\big{(}f_{\text{\footnotesize{off}}}(\lambda,t)\mu_{\text{\footnotesize{off}}}(\lambda,t)\big{)}$ (10) that are coupled through their boundary conditions [21]. The boundary conditions are listed in Appendix B.2. ###### Remark 1. The coupled equations (9)-(10) can be used to model either: (i) a _single_ TCL or (ii) a _collection_ of TCLs. For (i) the quantities (7)-(8) represent the _probability_ that a single TCLs temperature and on/off mode reside in the respective region. For (ii) the quantities (7)-(8) represent the _fraction_ of TCLs whose temperature and on/off mode reside in the respective region. How the equations (9)-(10) (specifically their discretized form) can be used to model an ensemble is discussed further in Section 5.2. Figure 2: The control volumes (CVs). The colors correspond to the colors found in Figure 3. The values in each CV represent the nodal temperature for the CV. The arrows describe the sign of the convection of the TCL through the CVs. The values are such that $N=m+q$. The terms involving $\alpha$ model rate of transfer between the corresponding CVs due to the thermostat policy, where $\alpha=\gamma+\frac{\sigma^{2}}{(\Delta\lambda)^{2}}$. The parameter $\gamma>0$ is a design parameter; see Remark 4.3. ## 3 Markov model from PDE Discretization We use the finite volume method (FVM) to discretize the PDEs (9) and (10). The discretization of (9) and (10) yields a finite dimensional probabilistic model for a single TCL (equation (16)). We discretize the PDEs (9) and (10) in a way that a control input for the BA can then be identified. More on this point will be discussed in Section 4, however the discretization here will play a role. ### 3.1 Spatial discretization The FVM bins the continuous temperature into $N$ control volumes (CV). The layout of the CVs is shown in Figure 2. The $N$ CVs for both the on and off mode state, as shown in Figure 2, are defined through the nodal temperature values ($\lambda_{\text{\footnotesize{on}}}$ and $\lambda_{\text{\footnotesize{off}}}$) and their boundaries ($\lambda^{+}_{\text{\footnotesize{on}}}$ and $\lambda^{+}_{\text{\footnotesize{off}}}$) and ($\lambda^{-}_{\text{\footnotesize{on}}}$ and $\lambda^{-}_{\text{\footnotesize{off}}}$): $\displaystyle\lambda_{\text{\footnotesize{on}}}=(\lambda^{i}_{\text{\footnotesize{on}}})_{i=1}^{N},\quad\lambda_{\text{\footnotesize{on}}}^{+}=\lambda_{\text{\footnotesize{on}}}+\frac{\Delta\lambda}{2},\quad\lambda_{\text{\footnotesize{on}}}^{-}=\lambda_{\text{\footnotesize{on}}}-\frac{\Delta\lambda}{2},$ $\displaystyle\lambda_{\text{\footnotesize{off}}}=(\lambda^{i}_{\text{\footnotesize{off}}})_{i=1}^{N},\quad\lambda_{\text{\footnotesize{off}}}^{+}=\lambda_{\text{\footnotesize{off}}}+\frac{\Delta\lambda}{2},\quad\lambda_{\text{\footnotesize{off}}}^{-}=\lambda_{\text{\footnotesize{off}}}-\frac{\Delta\lambda}{2},$ where $\Delta\lambda$ is the CV width. All intermediate values of $\lambda_{\text{\footnotesize{on}}}$ and $\lambda_{\text{\footnotesize{off}}}$ are separated from each other by $\Delta\lambda$. The values in $\lambda^{+}_{\text{\footnotesize{on}}}$ (respectively, $\lambda^{+}_{\text{\footnotesize{off}}}$) are the right edges of the CVs and the values $\lambda^{-}_{\text{\footnotesize{on}}}$ (respectively, $\lambda^{-}_{\text{\footnotesize{off}}}$) are the left edges of the CVs, for example, $\lambda^{1,-}_{\text{\footnotesize{off}}}=\lambda^{\text{low}}$. The quantities $\lambda^{\text{min}}$ and $\lambda^{\text{max}}$ specify the thermostat deadband, and are _different_ from the quantities $\lambda^{\text{high}}$ and $\lambda^{\text{low}}$ (see Figure 2). Figure 3: Sparsity pattern of the matrix $A(t)$ for $N=51$ CVs for both the on and off state. The colors correspond to the colors found in Figure 2. The steps taken to obtain the spatially discretized PDEs is detailed in Appendix B. To give an overview, the discretization is done in two parts: (i) for the internal CV’s (Appendix B.1) and (ii) for the boundary CV’s (Appendix B.2). We describe here the end result of the derivation in Appendix B. First, define the following quantities $\displaystyle\nu_{\text{\footnotesize{off}}}(\lambda^{i},t)$ $\displaystyle\triangleq\mu_{\text{\footnotesize{off}}}(\lambda^{i},t)\Delta\lambda,\quad\text{and}$ (11) $\displaystyle\nu_{\text{\footnotesize{on}}}(\lambda^{i},t)$ $\displaystyle\triangleq\mu_{\text{\footnotesize{on}}}(\lambda^{i},t)\Delta\lambda,$ (12) then construct the row vector, $\nu(t)=[\nu_{\text{\footnotesize{off}}}(t),\nu_{\text{\footnotesize{on}}}(t)]$. with $\displaystyle\nu_{\text{\footnotesize{off}}}(t)$ $\displaystyle\triangleq[\nu_{\text{\footnotesize{off}}}(\lambda^{1},t),\dots,\nu_{\text{\footnotesize{off}}}(\lambda^{N},t)],\quad\text{and}$ (13) $\displaystyle\nu_{\text{\footnotesize{on}}}(t)$ $\displaystyle\triangleq[\nu_{\text{\footnotesize{on}}}(\lambda^{1},t),\dots,\nu_{\text{\footnotesize{on}}}(\lambda^{N},t)].$ (14) By combining all the ordinary differential equations (ODEs) for the $\nu_{\text{\footnotesize{off}}}(\lambda^{i},t),\nu_{\text{\footnotesize{on}}}(\lambda^{i},t)$ for all the $i$’s, we obtain the linear time varying system $\displaystyle\frac{d}{dt}\nu(t)=\nu(t)A(t).$ (15) The sparsity pattern of $A(t)$ is shown in Figure 3. The system (15) is the spatially discretized version of the PDEs (9)-(10). The matrix $A(t)$ also satisfies the properties of a transition rate matrix, described in the following lemma. ###### Lemma 1. For all $t$, the matrix $A(t)$ is a transition rate matrix. That is, for all $t$ (i): $\displaystyle\quad A(t)\mathbb{1}=\mathbf{0}.$ (ii): $\displaystyle\quad\text{for all}\ i,\ A_{i,i}(t)\leq 0,\ \text{and}\ \text{for all}\ j\neq i\ A_{i,j}(t)\geq 0.$ ###### Proof. See Appendix A.1. ∎ ###### Remark 2. The choice of the FVM and how we discretize the convection and diffusion terms appearing in (9)-(10) is important for $A(t)$ to satisfy the conditions in Lemma 1. This issue is well known in the CFD literature, and also recognized in the related work [2]. If a finite difference method had been used with central differences for both diffusion and convection terms, the resulting $A(t)$ would require restrictive conditions on both $\sigma^{2}$ and $\Delta\lambda$ to satisfy the properties in Lemma 1 [27]. ### 3.2 Temporal discretization To temporally integrate the dynamics (15) we use a first order Euler approximation with time step $\Delta t>0$. Making the identifications $\nu_{k}\triangleq\nu(t_{k})$ and $A_{k}\triangleq A(t_{k})$ we have $\displaystyle\nu_{k+1}$ $\displaystyle=\nu_{k}P_{k},\quad\text{with}\quad P_{k}=I+\Delta tA_{k}.$ (16) In the continuous time setting elements of the vector $\nu(t)$ were referred to as, for example, $\nu_{\text{\footnotesize{on}}}(\lambda^{i},t)$. The counterpart to this, in the discrete time setting, is referring to elements of $\nu_{k}$ as, for example, $\nu_{\text{\footnotesize{on}}}[\lambda^{i},k]$. We further have the following. ###### Lemma 2. The matrix $P_{k}$ is a Markov transition probability matrix if $\displaystyle\forall\ i,\ \text{and}\ \forall\ k,\quad 0<\Delta t\leq\left|[A_{k}]_{i,i}\right|^{-1}.$ where $[A_{k}]_{i,i}$ is the $i^{th}$ diagonal element of the matrix $A_{k}$. ###### Proof. From Lemma 1 we have that $P_{k}\mathbb{1}=I\mathbb{1}+\Delta tA_{k}\mathbb{1}=\mathbb{1}$ since $A_{k}\mathbb{1}=0$. Also from Lemma 1, every element of $A_{k}$ is non-negative, save for the diagonal elements. Under the hypothesis on $A_{k}$, then every diagonal element of $I+\Delta tA_{k}$ will be in $[0,1]$. ∎ ###### Remark 3. The bound on the time step $\Delta t$ given in Lemma 2 is $O(\Delta\lambda)$, which follows from the PDE discretization; see Appendix B . Since $\Delta\lambda=\frac{\lambda^{\text{high}}-\lambda^{\text{min}}}{N}$, as the temperature resolution $\Delta\lambda$ becomes finer the time resolution $\Delta t$ must also become finer at the same rate. See also Remark 4.3 for a related comment. ## 4 Discrete space model of a TCL: structure and grid friendly policies Recall that the dynamics (16) derived in the previous section was for the thermostat policy. We now delve into the structure of these dynamics so to introduce a BA control input. We first formalize a discrete state space for the dynamics (16). We will then show that the transition matrix in (16) can be written as $P_{k}=\Phi G_{k}$ where $\Phi$ depends on the thermostat policy and $G_{k}$ on the TCL temperature dynamics and weather. The isolation of the policy then indicates how a BA could introduce grid friendly policies in place of the thermostat policy $\Phi$. ### 4.1 Discrete state space When the conditions of Lemma 2 are met $P_{k}$ is a transition matrix and hence each $\nu_{k}$ is a marginal pmf if $\nu_{0}$ is a pmf. The structure of this marginal is given from (7) for the on state (a similar interpretation holds for the off state) as, $\displaystyle\nu_{\text{\footnotesize{on}}}[\lambda^{i},k]$ $\displaystyle={\sf P}\left(\theta(t_{k})\in\text{CV}(i),\ m(t_{k})=\text{on}\right),$ (17) where $\theta(t_{k})$ is the temperature. Now denote, $\theta_{k}\triangleq\theta(t_{k})$, $m_{k}\triangleq m(t_{k})$, and $\displaystyle I_{k}\triangleq\sum_{i=1}^{N}i\mathbf{I}_{\text{CV}(i)}(\theta_{k},m_{k}).$ (18) The quantity $I_{k}$ indicates which CV the TCLs temperature resides in at time $k$. It also is a function of $m_{k}$ since the CV index for the on mode is different from the index for the off mode. We then define the following discrete state space: $\displaystyle{\sf Z}\triangleq\\{m\in\\{\text{\footnotesize{on}},\text{\footnotesize{off}}\\},\ I\in\\{1,\dots,N\\}\\},$ (19) with cardinality $\left|{\sf Z}\right|=2N$. Using the newly defined quantity $I_{k}$ we rewrite the marginals $\nu_{\text{\footnotesize{on}}}[\lambda^{i},k]$ and $\nu_{\text{\footnotesize{off}}}[\lambda^{i},k]$ as functions on ${\sf Z}$, $\displaystyle\nu_{\text{\footnotesize{on}}}[\lambda^{i},k]$ $\displaystyle={\sf P}\left(I_{k}=i,\ m_{k}=\text{on}\right),\quad\text{and}$ (20) $\displaystyle\nu_{\text{\footnotesize{off}}}[\lambda^{i},k]$ $\displaystyle={\sf P}\left(I_{k}=i,\ m_{k}=\text{off}\right).$ (21) From the above, the matrix $P_{k}$ (with the conditions of Lemma 2 satisfied) is the transition matrix for the joint process $(I_{k},m_{k})$ on the state space ${\sf Z}$. The dynamic equation $\nu_{k+1}=\nu_{k}P_{k}$ is then a probabilistic model for a TCL with state space ${\sf Z}$ and operating under the thermostat policy. ### 4.2 Conditional independence in $P_{k}$ In the following, we refer to the values of $I_{k}$ with $i$ and $j$ and the values of $m_{k}$ with $u$ and $v$. We introduce the following notation to refer to the elements of the transition matrix $P_{k}$: $\displaystyle P_{k}((i,u),(j,v))\triangleq$ (22) $\displaystyle{\sf P}\Big{(}I_{k+1}=j,\ m_{k+1}=v\ \Big{|}\ I_{k}=i,\ m_{k}=u,\ \theta^{a}_{k}=w_{k}\Big{)}.$ Recall, the matrix $P_{k}$ is derived for the thermostat policy. We will now show that the matrix $P_{k}$ can be written as the product of two matrices. One depends only on the thermostat policy (control) and the other depends only on weather and TCL temperature dynamics. That is, we show that each entry of $P_{k}$ factors as $\displaystyle P_{k}((i,u),(j,v))=\phi^{\text{TS}}_{u}(v\ |\ i)P^{u}_{k}(i,j)$ (23) where, for each given values of $\theta^{a}_{k}$, $P^{u}_{k}(i,j)$ is a _controlled transition matrix_ on Z: $\displaystyle P^{u}_{k}(i,j)\triangleq{\sf P}\left(I_{k+1}=j\ |\ I_{k}=i,\ m_{k}=u,\ \theta^{a}_{k}=w_{k}\right)$ (24) and $\phi^{\text{TS}}_{u}(v\ |\ i)$ is an instance of a _randomized policy_ $\phi_{u}(v\ |\ i)$ on Z: $\displaystyle\phi_{u}(v\ |\ i)\triangleq{\sf P}\left(m_{k+1}=v\ |\ I_{k}=i,\ m_{k}=u\right).$ (25) We show the factorization (23) through construction next. #### 4.2.1 Constructing the factorization The quantity $\phi^{\text{TS}}_{u}(v\ |\ i)$ in (25) is the thermostat policy on ${\sf Z}$, which is formally defined as follows. ###### Definition 1. The thermostat policy on ${\sf Z}$ is specified by the two vectors, $\phi^{\text{TS}}_{\text{\footnotesize{off}}},\phi^{\text{TS}}_{\text{\footnotesize{on}}}\in\mathbb{R}^{N}$, where $\phi^{\text{TS}}_{\text{\footnotesize{off}}}\triangleq\phi^{\text{TS}}_{\text{\footnotesize{off}}}(\text{\footnotesize{on}}\ |\ \cdot)=\mathbf{e}_{N}$, $\phi^{\text{TS}}_{\text{\footnotesize{on}}}\triangleq\phi^{\text{TS}}_{\text{\footnotesize{on}}}(\text{\footnotesize{off}}\ |\ \cdot)=\mathbf{e}_{1}$, and $\phi^{\text{TS}}_{\text{\footnotesize{off}}}(\text{\footnotesize{off}}\ |\ \cdot)\triangleq 1-\phi^{\text{TS}}_{\text{\footnotesize{off}}}$, $\phi^{\text{TS}}_{\text{\footnotesize{on}}}(\text{\footnotesize{on}}\ |\ \cdot)\triangleq 1-\phi^{\text{TS}}_{\text{\footnotesize{on}}}$. The quantity $P^{u}_{k}(i,j)$ in (24) represents the policy-free (open loop) evolution of the TCL on ${\sf Z}$. That is, it describes how the TCLs temperature evolves under a fixed mode. We define matrices with entries $P^{u}_{k}(i,j)$ next. ###### Definition 2. Let $P_{k}^{\text{\footnotesize{off}}},P_{k}^{\text{\footnotesize{on}}}\in\mathbb{R}^{N\times N}$ have $(i,j)$ entries given by, $\displaystyle P_{k}^{\text{\footnotesize{off}}}(i,j)$ $\displaystyle=P_{w_{k}}((i,\text{\footnotesize{off}}),(j,\text{\footnotesize{off}})),\quad i\neq N\ \text{and}\ j\neq N,$ $\displaystyle P_{k}^{\text{\footnotesize{on}}}(i,j)$ $\displaystyle=P_{w_{k}}((i,\text{\footnotesize{on}}),(j,\text{\footnotesize{on}})),\quad i\neq 1\ \text{and}\ j\neq 1,$ with $P_{k}^{\text{\footnotesize{off}}}(N,N)=1$ and $P_{k}^{\text{\footnotesize{on}}}(1,1)=1$. The quantities defined in Definition 1 and 2 correspond to entries of $P_{k}$. To construct the promised factorization, from these definitions, the idea is to construct its four sub-matrices that correspond to all possible combinations of $u,v\in\\{\text{\footnotesize{on}},\text{\footnotesize{off}}\\}$ (see Figure 3). For example, the $\text{\footnotesize{off}}-\text{\footnotesize{off}}$ quadrant of $P_{k}$ is given by the matrix product $\displaystyle\big{(}I-\text{diag}(\phi^{\text{TS}}_{\text{\footnotesize{off}}})\big{)}P^{\text{\footnotesize{off}}}_{k}.$ However, since the temperature associated with the $i^{th}$ CV for the on mode is not the same temperature associated with the $i^{th}$ CV for the off state (see Figure 2) it is _not_ true that the $\text{\footnotesize{off}}-\text{\footnotesize{on}}$ quadrant of $P_{k}$ is given as $\text{diag}(\phi^{\text{TS}}_{\text{\footnotesize{off}}})P^{\text{\footnotesize{off}}}_{k}$. The entries of the matrix $P^{\text{\footnotesize{off}}}_{k}$ need to be re- arranged so to correctly account for the difference in CV index between the on/off mode. We define such correctly re-arranged matrices next. ###### Definition 3. Let $I^{\text{\footnotesize{off}}}=\\{m,\dots,N\\}$, $I^{\text{\footnotesize{on}}}=\\{1,\dots,q\\}$, $m^{-}=m-1$, and $S_{k}^{\text{\footnotesize{off}}},S_{k}^{\text{\footnotesize{on}}}\in\mathbb{R}^{N\times N}$ with $(i,j)$ entries $\displaystyle S_{k}^{\text{\footnotesize{off}}}(i,j-m^{-})$ $\displaystyle=\begin{cases}P_{k}^{\text{\footnotesize{off}}}(i,j)&i,j\in I^{\text{\footnotesize{off}}}\\\ 0&\text{otherwise}.\end{cases}$ (26) $\displaystyle S_{k}^{\text{\footnotesize{on}}}(i,j+m^{-})$ $\displaystyle=\begin{cases}P_{k}^{\text{\footnotesize{on}}}(i,j)&i,j\in I^{\text{\footnotesize{on}}}\\\ 0&\text{otherwise}.\end{cases}$ (27) The above definition is based on the construction that $N=q+m$. The quantities in Definition 3 let us construct, e.g., the $\text{\footnotesize{off}}-\text{\footnotesize{on}}$ quadrant of $P_{k}$ as $\text{diag}(\phi_{\text{\footnotesize{off}}}^{\text{TS}})S_{k}^{\text{\footnotesize{off}}}$. The next result shows that $P_{k}=\Phi^{\text{\footnotesize{TS}}}G_{k}$ under certain conditions and for appropriate choices of the matrices $\Phi^{\text{\footnotesize{TS}}}$ and $G_{k}$. ###### Lemma 4.1. Let the time discretization period $\Delta t$ and the parameter $\alpha$ that appears as a design choice in discretizing the PDEs to ODEs be chosen to satisfy $\alpha=(\Delta t)^{-1}$. Let $\Phi^{\text{TS}}_{\text{\footnotesize{off}}}=\text{diag}(\Phi^{\text{TS}}_{\text{\footnotesize{off}}})$ and $\Phi^{\text{TS}}_{\text{\footnotesize{on}}}=\text{diag}(\Phi^{\text{TS}}_{\text{\footnotesize{on}}})$, and $\displaystyle\Phi^{\text{\footnotesize{TS}}}$ $\displaystyle\triangleq\begin{bmatrix}I-\Phi^{\text{TS}}_{\text{\footnotesize{off}}}&\Phi^{\text{TS}}_{\text{\footnotesize{off}}}&\mathbf{0}&\mathbf{0}\\\ \mathbf{0}&\mathbf{0}&\Phi^{\text{TS}}_{\text{\footnotesize{on}}}&I-\Phi^{\text{TS}}_{\text{\footnotesize{on}}}\end{bmatrix}\quad\text{and}$ (28) $\displaystyle G_{k}$ $\displaystyle\triangleq\begin{bmatrix}\mathbf{0}&S_{k}^{\text{\footnotesize{off}}}&\mathbf{0}&P_{k}^{\text{\footnotesize{on}}}\\\ P_{k}^{\text{\footnotesize{off}}}&\mathbf{0}&S_{k}^{\text{\footnotesize{on}}}&\mathbf{0}\end{bmatrix}^{T},$ (29) then $\displaystyle P_{k}=\Phi^{\text{\footnotesize{TS}}}G_{k}.$ (30) ###### Proof 4.2. See Appendix A.2. ###### Remark 4.3. The condition $\alpha=1/\Delta t$ can be satisfied as long as time and temperature discretization intervals are chosen to satisfy $\Delta t<(\Delta\lambda)^{2}/\sigma^{2}$. To understand how, recall that in the discretizing the PDE to the coupled ODEs, a design parameter $\gamma>0$ appears: some rate of density is transferred out of the control volume $\lambda^{N}_{\text{\footnotesize{off}}}$ and into the CV $\lambda^{q}_{\text{\footnotesize{on}}}$ (as depicted in Figure 2) due to thermostatic control. The rate of the density transfer is then given as $-\gamma\nu_{\text{\footnotesize{off}}}(\lambda^{N},t)$, where $\gamma>0$ is a modeling choice and a constant of appropriate units that describes the discharge rate. We then define $\alpha\triangleq D+\gamma$ where $D=\frac{\sigma^{2}}{(\Delta\lambda)^{2}}$. Recall that $\sigma^{2}$ is the variance in the Fokker-Planck equation (9)-(10) and $\Delta\lambda$ is the temperature discretization interval. Thus, as long as $1/\Delta t>D$, a positive $\gamma$ can be chosen while meeting the condition $\alpha=1/\Delta t$. The condition $1/\Delta t>D$ is equivalent to $\Delta t<(\Delta\lambda)^{2}/\sigma^{2}$. ###### Remark 4.4. The conditional independence factorization (30) has been a useful assumption in the design of algorithms in [3]. In the present it is a byproduct of our spatial and temporal discretization of the PDEs (9)-(10). There are other works [2, 1, 25] that develop Markov models for TCLs through discretization of PDEs. However, to our knowledge, our work is the first to uncover this factorization. Lemma 4.1 informs us how to define the dynamics of the marginals (20) under a different policy than the thermostat policy, which is described next. ### 4.3 BA control command $=$ policy In light of the previous section, an arbitrary randomized policy can replace the thermostat policy to control the state process on ${\sf Z}$. That is equivalent to replacing $\Phi^{\text{\footnotesize{TS}}}$ in (30) with a new matrix $\Phi$ that corresponds to a policy designed for grid support. From the viewpoint of the BA this randomized policy _is_ the control input that it must design and broadcast to a TCL. The TCL now implements this policy to make on/off decisions instead of using the thermostat policy. As we shall soon see, if the BA appropriately designs and sends the randomized policy to multiple TCLs it can achieve coordination of the TCLs for grid support. To distinguish from thermostat policy $\phi^{\text{TS}}_{\text{\footnotesize{off}}}$ and $\phi^{\text{TS}}_{\text{\footnotesize{on}}}$ in the prior section that only maintains temperature, we denote the newly introduced policies for providing grid support with the superscript ‘GS’. We require the policies, $\phi^{\text{GS}}_{\text{\footnotesize{on}}}$ and $\phi^{\text{GS}}_{\text{\footnotesize{off}}}$, to have the following structure $\displaystyle\phi^{\text{GS}}_{\text{\footnotesize{off}}}(\text{on}\ |\ j)=\begin{cases}\kappa^{\text{\footnotesize{on}}}_{j},&(m+1)\leq j\leq(N-1).\\\ 1,&j=N.\\\ 0,&\text{o.w.}\end{cases}$ (31) $\displaystyle\phi^{\text{GS}}_{\text{\footnotesize{on}}}(\text{off}\ |\ j)=\begin{cases}\kappa^{\text{\footnotesize{off}}}_{j},&2\leq j\leq(q-1).\\\ 1,&j=1.\\\ 0,&\text{o.w.}\end{cases}$ (32) with $\phi^{\text{GS}}_{\text{\footnotesize{off}}}(\text{off}\ |\ \cdot)=1-\phi^{\text{GS}}_{\text{\footnotesize{off}}}(\text{on}\ |\ \cdot)$ and $\phi^{\text{GS}}_{\text{\footnotesize{on}}}(\text{on}\ |\ \cdot)=1-\phi^{\text{GS}}_{\text{\footnotesize{on}}}(\text{off}\ |\ \cdot)$ and $\kappa^{\text{\footnotesize{on}}}_{j},\kappa^{\text{\footnotesize{off}}}_{j}\in[0,1]$ for all $j$. The policies could also be time varying, for example: $\kappa^{\text{\footnotesize{off}}}_{j}[k]$ and $\kappa^{\text{\footnotesize{on}}}_{j}[k]$. The dependence of the policies on time is denoted as $\phi^{\text{GS}}_{\text{\footnotesize{off}}}[k]$ and $\phi^{\text{GS}}_{\text{\footnotesize{on}}}[k]$. Designing the grid support control policies is then equivalent to choosing the values of $\kappa^{\text{\footnotesize{on}}}_{j}[k]$ and $\kappa^{\text{\footnotesize{off}}}_{j}[k]$ for all $j$ and $k$. We have required $\phi^{\text{GS}}_{\text{\footnotesize{off}}}(\text{on}\ |\ j)=0$ for $1\leq j\leq m$ since the temperatures corresponding to these indices are below the permitted deadband temperature, $\lambda^{\text{min}}$. Hence, turning on at these temperature does not make physical sense. The arguments for the zero elements in $\phi^{\text{GS}}_{\text{\footnotesize{on}}}$ are symmetric. ###### Remark 4.5. From the individual TCL’s perspective, implementing grid support randomized policies of the form (31)-(32) is straightforward: (i) the TCL measures its current temperature and on/off status, (ii) the TCL “bins” this temperature value according to (18) and (iii) the TCL flips a coin to decide its next on/off state according to the probabilities given in (31)-(32). Note that the thermostat policy is a special case of the grid support control policy, and both policies enforce the temperature constraint. ## 5 Proposed framework We are now in a position to present our unified framework for coordination of TCLs. We first expand the state of the model (16) so to incorporate cycling, following [19, 26]. We then shift the viewpoint from a single TCL to that of a collection of TCLs (recall Remark 1) to develop our control oriented aggregate model. Using this model we develop a method for designing both reference and policy through convex optimization. ### 5.1 Individual TCL model with cycling We now augment the model for a TCL’s temperature evolution with cycling dynamics. Recall the cycling constraint: as soon as a TCL switches its mode, the TCL becomes stuck in that mode for $\tau$ time instances. This constraint can be represented as the evolution of a state, specifically, a counter variable. First defining the binary variable $s_{k}$ as $s_{k}=1$ if the TCL is stuck in the current mode at time $k$ and $0$ if it is not stuck. The counter variable is defined as follows $\displaystyle{\sf{L}}_{k+1}\triangleq\begin{cases}{\sf{L}}_{k}+1,&s_{k}=1.\\\ 0,&s_{k}=0.\end{cases}$ (33) This variable denotes the time spent in the “stuck” mode ($s_{k}=1$). A TCL has flexibility to help the grid only when ${\sf{L}}_{k}=0$, which means it is not stuck in either the on or off mode. If ${\sf{L}}_{k}>0$, it is stuck in either the on or off mode, and switching the mode to help the grid will violate the cycling constraint. Recall, the discrete state space ${\sf Z}$ for a TCL included binned temperature and on/off mode. The space ${\sf Z}$, the policies $\phi^{\text{GS}}_{\text{\footnotesize{on}}}$ and $\phi^{\text{GS}}_{\text{\footnotesize{off}}}$, the marginal pmf $\nu_{k}$, and the transition matrix $P_{k}$ (and consequently its factors $\Phi$ and $G_{k}$) now all need to be expanded to be defined over a state space consisting of $(I_{k},m_{k},{\sf{L}}_{k})$. This expansion is described next. We denote this newly expanded state space as the set of values: ${\sf X}\triangleq$ $\displaystyle\Big{\\{}m\in\\{\text{\footnotesize{on}},\text{\footnotesize{off}}\\},\ I\in\\{1,\dots,N\\},\ {\sf{L}}\in\\{0,\dots,\tau\\}\Big{\\}},$ (34) with cardinality $|{\sf X}|=2N(\tau+1)$. The policies on the expanded state space are: $\displaystyle\phi^{\text{E}}_{\text{\footnotesize{off}}}$ $\displaystyle=\mathbf{I}_{\\{0\\}}({\sf{L}})\phi^{\text{GS}}_{\text{\footnotesize{off}}}+(1-\mathbf{I}_{\\{0\\}}({\sf{L}}))\phi^{\text{TS}}_{\text{\footnotesize{off}}},\quad\text{and}$ (35) $\displaystyle\phi^{\text{E}}_{\text{\footnotesize{on}}}$ $\displaystyle=\mathbf{I}_{\\{0\\}}({\sf{L}})\phi^{\text{GS}}_{\text{\footnotesize{on}}}+(1-\mathbf{I}_{\\{0\\}}({\sf{L}}))\phi^{\text{TS}}_{\text{\footnotesize{on}}}.$ To ensure that expanded policy (35) will enforce the cycling constraint, we impose the following restriction at the design stage: a TCL with ${\sf{L}}_{k}>0$ will only implement the thermostat policy, and a TCL with ${\sf{L}}_{k}=0$ will make on/off decisions based on the grid support policy. The construction in this way ensures a TCL will not violate its cycling and temperature constraints under the conditions in Assumption A.2 and A.3. Each entry of the expanded policy is denoted as $\phi^{\text{E}}_{\text{\footnotesize{off}}}(u\ |\ j,\ l)$ and $\phi^{\text{E}}_{\text{\footnotesize{on}}}(u\ |\ j,\ l)$. The expanded marginals are $\nu_{\text{\footnotesize{off}}}[\lambda^{j},l,k]$ and $\nu_{\text{\footnotesize{on}}}[\lambda^{j},l,k]$, and $\nu_{\text{\footnotesize{off}},l}$ (resp., $\nu_{\text{\footnotesize{on}},l}$) is shorthand for $\nu_{\text{\footnotesize{off}}}[\cdot,l,k]$ (resp., $\nu_{\text{\footnotesize{on}}}[\cdot,l,k]$). In vectorized form, the expanded marginal is $\nu^{\text{E}}=[\nu_{\text{\footnotesize{off}}}^{\text{E}},\nu_{\text{\footnotesize{on}}}^{\text{E}}]$ where $\nu_{\text{\footnotesize{off}}}^{\text{E}}=[\nu_{\text{\footnotesize{off}},0},\dots,\nu_{\text{\footnotesize{off}},\tau}]$ and $\nu_{\text{\footnotesize{on}}}^{\text{E}}=[\nu_{\text{\footnotesize{on}},0},\dots,\nu_{\text{\footnotesize{on}},\tau}]$. Define $\displaystyle G^{\text{E}}_{k}$ $\displaystyle\triangleq\begin{bmatrix}\mathbf{0}&D_{\tau}\otimes S_{k}^{\text{\footnotesize{on}}}&\mathbf{0}&C_{\tau}\otimes P_{k}^{\text{\footnotesize{off}}}\\\ C_{\tau}\otimes P_{k}^{\text{\footnotesize{on}}}&\mathbf{0}&D_{\tau}\otimes S_{k}^{\text{\footnotesize{off}}}&\mathbf{0}\end{bmatrix}^{T},$ (36) where $D_{\tau}\triangleq\mathbb{1}^{T}\otimes\mathbf{e}_{2}\in\mathbb{R}^{\tau+1\times\tau+1}$ and $\displaystyle C_{\tau}\triangleq\begin{bmatrix}1&0&\mathbf{0}_{\tau-1}^{T}\\\ \mathbf{0}_{\tau-1}&\mathbf{0}_{\tau-1}&I_{\tau-1}\\\ 1&0&\mathbf{0}_{\tau-1}^{T}\end{bmatrix}\in\mathbb{R}^{(\tau+1)\times(\tau+1)}.$ (37) We define the matrix $\Phi_{k}^{\text{E}}$ as having the same structure as (28), but with the expanded policies $\phi^{\text{E}}_{\text{\footnotesize{off}}}$ and $\phi^{\text{E}}_{\text{\footnotesize{on}}}$, i.e., $\displaystyle\Phi^{\text{E}}_{k}\triangleq\begin{bmatrix}I-\Phi^{\text{E}}_{\text{\footnotesize{off}}}[k]&\Phi^{\text{E}}_{\text{\footnotesize{off}}}[k]&\mathbf{0}&\mathbf{0}\\\ \mathbf{0}&\mathbf{0}&\Phi^{\text{E}}_{\text{\footnotesize{on}}}[k]&I-\Phi^{\text{E}}_{\text{\footnotesize{on}}}[k]\end{bmatrix},$ (38) where $\Phi^{\text{E}}_{\text{\footnotesize{off}}}[k]\triangleq\text{diag}(\phi^{\text{E}}_{\text{\footnotesize{off}}}[k])$ and $\Phi^{\text{E}}_{\text{\footnotesize{on}}}[k]\triangleq\text{diag}(\phi^{\text{E}}_{\text{\footnotesize{on}}}[k])$. The model of a TCL with cycling dynamics and grid support policy becomes $\displaystyle\nu^{\text{E}}_{k+1}=\nu^{\text{E}}_{k}\Phi_{k}^{\text{E}}G_{k}^{\text{E}}.$ (39) The structure of the transition matrix $\Phi_{k}^{\text{E}}G_{k}^{\text{E}}$ is shown in Figure 4. For comparison, the transition matrix with policy $\phi^{\text{GS}}$ and _without_ the cycle counter variable would simply be the four red shaded blocks appearing in their respective quadrant. In the expanded system, an on to off mode switch forces probability mass from the red shaded region ($l=0$ and $m=\text{\footnotesize{on}}$) to the green shaded region ($l=1$ and $m=\text{\footnotesize{off}}$). Mass must then transition through the chain of $\tau$ green blocks until it reaches the red block again, so to respect the cycling constraint. Figure 4: The sparsity pattern of the expanded transition matrix (the dots represent non-zero entries in the matrix) with $\tau=5$. Each shaded block is over the entire range of temperature values. ### 5.2 Aggregate model of a collection of TCLs We now transition from the viewpoint of a single TCL to that of a collection of ${\sf{N_{tcl}}}$ TCLs: $\ell=1,\dots,{\sf{N_{tcl}}}$. For example, $m_{k}^{\ell}$ and $I_{k}^{\ell}$ are the mode and binned temperature of the $\ell^{th}$ TCL at time $k$. Recall Remark 1, the model (39) also describes an entire collection of TCLs. For a single TCL, we view the state $\nu_{k}^{\text{E}}$ as a marginal but for a collection of TCLs we expect the marginal pmf $\nu_{k}^{\text{E}}$ to approximate the histogram $\displaystyle h_{k}[u,i,l]$ $\displaystyle\triangleq\frac{1}{{\sf{N_{tcl}}}}\sum_{\ell=1}^{{\sf{N_{tcl}}}}\Big{(}\mathbf{I}_{\\{i\\}}(I^{\ell}_{k})\mathbf{I}_{\\{u\\}}(m^{\ell}_{k})\mathbf{I}_{\\{l\\}}({\sf{L}}^{\ell}_{k})\Big{)},$ (40) for each state $(u,i,l)\in{\sf X}$ as ${\sf{N_{tcl}}}\rightarrow\infty$. In the same regard, we define $\displaystyle\gamma_{k}^{\text{E}}\triangleq\nu_{k}^{\text{E}}C^{\text{E}},\quad\text{where}\quad C^{\text{E}}\triangleq[\mathbf{0}^{T},P_{\text{\footnotesize{agg}}}\mathbb{1}^{T}]^{T},$ (41) where $P_{\text{\footnotesize{agg}}}$ is the maximum possible power of the collection, defined in (4). We expect $\gamma_{k}^{\text{E}}$ to approximate the total power consumption $y_{k}$ of the collection of ${\sf{N_{tcl}}}$ TCLs: $\displaystyle y_{k}$ $\displaystyle\triangleq P\sum_{\ell=1}^{{\sf{N_{tcl}}}}m_{k}^{\ell}.$ (42) which is the discrete-time equivalent of $y(t)$ defined in (5). That is, we expect $\gamma_{k}^{\text{E}}\approx y_{k}$ for large ${\sf{N_{tcl}}}$, based on a law of large numbers argument [5]. The _control oriented aggregate model of a TCL collection_ is the dynamics (39) together with the output (41): $\displaystyle\nu^{\text{E}}_{k+1}=\nu^{\text{E}}_{k}\Phi_{k}^{\text{E}}G_{k}^{\text{E}}.\qquad\text{and}\qquad\gamma_{k}^{\text{E}}=\nu_{k}^{\text{E}}C^{\text{E}}.$ (43) #### 5.2.1 Evaluating the aggregate model Before proceeding to policy design with our developed model (43), we first show that it is effective in modeling a population of TCLs. We do this by comparing the state of the model to (40) and (42) obtained from a simulation of ${\sf{N_{tcl}}}$= 50,000 air conditioning TCLs. The comparison results are shown in Figure 5 and Figure 6. The mode state of each TCL evolves according to a control policy, where the $\phi^{\text{GS}}_{\text{\footnotesize{off}}}$ and $\phi^{\text{GS}}_{\text{\footnotesize{on}}}$ portion are shown in Figure 6 (bottom). The policy is arbitrary, designed merely to be an example of a non- thermostat policy. This policy satisfies the structure in (31) and (32) so that both temperature and cycling constraints are satisfied at each TCL. The temperature evolution evolves according to (2). We see the state $\nu^{\text{E}}_{k}$ matches the histogram $h_{k}$ of the collection for the devices that are not stuck (Figure 5 (top)) and for the devices that are stuck (Figure 5 (bottom)). Additionally, the output of the aggregate model, $\gamma_{k}^{\text{E}}$, matches it’s empirical counterpart $y_{k}$ (shown in Figure 6 (top)). Figure 5: (Top): Histogram of the collection for the devices that are on and not stuck. (Bottom): Histogram of the collection for the devices that are on and are stuck. Figure 6: (Top): Comparison of the output of the expanded aggregate model $\gamma_{k}^{\text{E}}$ and the ensembles power consumption $y_{k}$. (Bottom): The policies $\phi^{\text{GS}}_{\text{\footnotesize{off}}}$ and $\phi^{\text{GS}}_{\text{\footnotesize{on}}}$ used for the numerical experiment in Section 5.2. ### 5.3 Grid support Policy design The goal of coordinating TCLs is to help the BA balance supply and demand of electricity in the grid. We denote $r^{\text{BA}}_{k}$ as the desired demand from all flexible loads and batteries that will reduce the imbalance to 0. It is unreasonable to expect any collection of TCLs to meet the entire desired demand $r^{\text{BA}}_{k}$ while maintaining their QoS. Only a portion of $r^{\text{BA}}_{k}$ can be supplied by TCLs, and we denote this portion by $r_{k}$. Determining $r_{k}$ becomes an optimal control problem due to the time coupling produced by the TCL dynamics. We consider a planning horizon of $T_{\text{plan}}$. To simultaneously design grid support control policies $\phi^{\text{GS}}_{\text{\footnotesize{off}}}[k]$ and $\phi^{\text{GS}}_{\text{\footnotesize{on}}}[k]$ and determine a suitable reference signal $r_{k}$ over $T_{\text{plan}}$ the BA solves the following optimization problem, $\displaystyle\eta^{*}=\min_{\nu^{\text{E}}_{k},\Phi^{\text{E}}_{k}}\ $ $\displaystyle\eta(\hat{\nu})=\sum_{k\in{\sf T}}\Big{(}r^{\text{BA}}_{k}-\gamma^{\text{E}}_{k}\Big{)}^{2}$ (44) s.t. $\displaystyle\nu^{\text{E}}_{k+1}=\nu^{\text{E}}_{k}\Phi_{k}^{\text{E}}G_{k}^{\text{E}},\quad\nu^{\text{E}}_{{\sf T}(0)}=\hat{\nu},$ (45) $\displaystyle\gamma^{\text{E}}_{k}=\nu_{k}^{\text{E}}C^{\text{E}},\quad\nu^{\text{E}}_{k}\in[0,1],\quad\Phi^{\text{E}}_{k}\in\varPhi.$ (46) The solution at time $k$ is denoted $r_{k}\triangleq\gamma_{k}^{\text{E},*}$, $\phi^{\text{GS},*}_{\text{\footnotesize{off}}}[k]$, and $\phi^{\text{GS},*}_{\text{\footnotesize{on}}}[k]$. We have ${\sf T}\triangleq\\{{\sf T}(0),\dots,{\sf T}(0)+T_{\text{plan}}-1\\}$ is the index set of times, ${\sf T}(0)$ denotes the initial time index, $\hat{\nu}$ is the initial condition, and $\nu_{k}^{\text{E}}\in[0,1]$ holds elementwise. The set $\varPhi$ collects all of the constraints on the policy. This includes the equality constraints set by the structural requirements in (31)-(32) and (35) as well as the structural requirement in (38). These constraints require certain elements of the policy to be either zero or one. The policy should also be a valid conditional pmf and its elements in $[0,1]$. Hence, the set $\varPhi$ is the following convex set $\displaystyle\varPhi\triangleq\Big{\\{}\Phi\in\mathbb{R}^{|{\sf X}|\times 2|{\sf X}|}_{[0,1]}\ \big{|}\ $ $\displaystyle\Phi\ \text{satisfies}~{}\eqref{eq:fullBAcontPol},\ \mathbb{1}=\Phi\mathbb{1},$ $\displaystyle\phi^{\text{GS}}_{\text{\footnotesize{off}}}\ \text{satisfies}~{}\eqref{eq:randPolOff2On},$ $\displaystyle\phi^{\text{GS}}_{\text{\footnotesize{on}}}\ \text{satisfies}~{}\eqref{eq:randPolOn2Off},\ \text{and}$ $\displaystyle\phi^{\text{E}}_{\text{\footnotesize{off}}}\ \text{and}\ \phi^{\text{E}}_{\text{\footnotesize{on}}}\ \text{satisfy}~{}\eqref{eq:expPolStruct}\Big{\\}}.$ (47) Where, e.g., $\mathbb{R}^{|{\sf X}|\times|{\sf X}|}_{[0,1]}$ is the set of $|{\sf X}|\times|{\sf X}|$ matrices with elements in $[0,1]$. ##### QoS + Solution of (44) 1. 1. The equality constraints in $\Phi$ are present to ensure the individual TCL’s QoS constraints: the structure (35) ensures the cycling constraint and the structure (31)-(32) ensures the temperature constraint. Recall that this structure guarantees QoS by requiring the policy to place zero probability on state transitions that would violate QoS. 2. 2. A solution to (44) yields, for $k\in{\sf T}$, two things: (i) the optimal randomized policies $\phi^{\text{GS},*}_{\text{\footnotesize{off}}}[k]$ and $\phi^{\text{GS},*}_{\text{\footnotesize{on}}}[k]$ and (ii) an optimal reference for the power demand of the TCL collection $r_{k}(=\gamma^{\text{E},*}_{k})$. The reference is optimal in the following sense: among all power demand signals the collection can track without requiring any TCL to violate its local QoS constraints in so doing, it is the closest to the BA’s desired demand $r^{\text{BA}}$ in 2-norm. The reference is also the predicted power consumption of the TCLs whilst using the policies $\phi^{\text{GS},*}_{\text{\footnotesize{off}}}[k]$ and $\phi^{\text{GS},*}_{\text{\footnotesize{on}}}[k]$. ###### Remark 5.1. Since the reference $r_{k}(=\gamma^{\text{E},*}_{k})$ from (44) is the best the TCLs can do to help the BA without any TCL having to violate its QoS, Problem (44) therefore also provides an answer to the “aggregate flexibility” question: how much can a collection of TCLs vary their demand while maintaining their local QoS constraints. This question has been investigated by many works [25, 13, 10, 15]. #### 5.3.1 Convex control synthesis The problem (44) is non-convex due to the product $\nu^{\text{E}}_{k}\Phi_{k}^{\text{E}}$ in the constraint. A well known convexification remedy for (44) is to consider optimizing over the marginal and joint distribution instead of the marginal and the policy [22, 2]. Using our identified structure from Section 4.2 we construct the following joint distribution (written in matrix form): $\displaystyle J_{k}=\text{diag}(\nu^{\text{E}}_{k})\Phi^{\text{E}}_{k}\in\mathbb{R}^{|{\sf X}|\times 2|{\sf X}|}.$ (48) By construction, we have that $\nu^{\text{E}}_{k+1}=\mathbb{1}^{T}J_{k}G^{\text{E}}_{k}$ and $(\nu^{\text{E}}_{k})^{T}=J_{k}\mathbb{1}$ since $\mathbb{1}^{T}\text{diag}(\nu^{\text{E}}_{k})=\nu^{\text{E}}_{k}$ and $\mathbb{1}=\Phi^{\text{E}}_{k}\mathbb{1}$. It is straightforward to convert the constraint set $\Phi^{\text{E}}_{k}\in\varPhi$ to the new decision variables. For the equality constraints in $\varPhi$ if we have that $\phi^{\text{E}}_{\text{\footnotesize{off}}}(u\ |\ j,l)=\kappa$, then in the decision variables $J_{k}$ and $\nu_{k}^{\text{E}}$ we will have a linear constraint of the form $\displaystyle{\sf P}\left(m_{k+1}=u,\ I_{k}=j,\ {\sf{L}}_{k}=l,\ m_{k}=\text{off}\right)$ $\displaystyle=\kappa\nu_{\text{\footnotesize{off}}}[\lambda^{j},l,k],$ (49) where the LHS of the above is some element in the matrix $J_{k}$. In addition to the above equality constraints, requiring both $J_{k}$ and $\nu_{k}^{\text{E}}$ to be within $[0,1]$ and the constraint $(\nu^{\text{E}}_{k})^{T}=J_{k}\mathbb{1}$ will allow one to reconstruct a policy $\Phi_{k}^{\text{E}}\in\varPhi$ from $J_{k}$ and $\nu_{k}^{\text{E}}$ (described shortly in Lemma 5.2). We denote the transcription of $\Phi^{\text{E}}_{k}\in\varPhi$ to the new variables as $(J_{k},\nu_{k}^{\text{E}})\in\bar{\varPhi}$. Optimizing over $J_{k}$ and $\nu_{k}^{\text{E}}$ yields the convex program: $\displaystyle\begin{split}\eta^{*}=&\min_{\nu^{\text{E}}_{k},J_{k}}\ \eta(\hat{\nu})=\sum_{k\in{\sf T}}\Big{(}r^{\text{BA}}_{k}-\gamma^{\text{E}}_{k}\Big{)}^{2}\\\ \text{s.t.}\quad&\nu^{\text{E}}_{k+1}=\mathbb{1}^{T}J_{k}G^{\text{E}}_{k},\quad\nu^{\text{E}}_{{\sf T}(0)}=\hat{\nu},\quad\gamma^{\text{E}}_{k}=\nu_{k}^{\text{E}}C^{\text{E}},\\\ &\nu^{\text{E}}_{k},J_{k}\in[0,1],\ (\nu^{\text{E}}_{k})^{T}=J_{k}\mathbb{1},\ (J_{k},\nu_{k}^{\text{E}})\in\bar{\varPhi}.\end{split}$ (50) Once the convex problem is solved, the grid support control policies need to be recovered from it by using the relation (48). If the matrix $\text{diag}(\nu_{k}^{\text{E}})$ is invertible, then the policy can be obtained trivially from inversion of $\text{diag}(\nu_{k}^{\text{E}})$. If $\text{diag}(\nu_{k}^{\text{E}})$ is not invertible, then slight care is required when reconstructing a policy from the solution of (50). We describe this in the following Lemma. ###### Lemma 5.2. Suppose for all $k\in{\sf T}$ that $\nu_{k}^{\text{E}}$ and $J_{k}$ satisfy the constraints in problem (50). Then, there exists matrices $H_{k}=H_{k}(\nu_{k}^{\text{E}})$ and $W_{k}=W_{k}(\nu_{k}^{\text{E}})$ so that for all $k\in{\sf T}$ the quantity $\Phi_{k}^{\text{E}}=H_{k}J_{k}+W_{k}$ satisfies (48) and $\Phi_{k}^{\text{E}}\in\varPhi$. ###### Proof 5.3. See Appendix A.3. Exact construction of $H_{k}$ and $W_{k}$ is given in the proof of Lemma 5.2. _Hence, the proof of Lemma 5.2 provides an algorithm for computing grid support control policies that are feasible for the problem (44) from the solutions of the convex problem (50)._ Further the two problems have a certain equivalence described here in the following Theorem. ###### Theorem 1. Denote $\eta^{*}_{\text{CVX}}$ the optimal cost for (50) and $\eta^{*}_{\text{NCVX}}$ the optimal cost for (44) we have that $\eta^{*}_{\text{CVX}}=\eta^{*}_{\text{NCVX}}$. ###### Proof 5.4. See Appendix A.4. This result, for a similar problem setup, is also reported in [2]. While we have no guarantee on the difference of the argument minimizers (and hence the policies obtained from both), Theorem (1) says that the policies will produce the same tracking performance. Further, from Lemma 5.2, the policies produced from either problem are guaranteed to ensure TCL QoS. #### 5.3.2 Computational considerations The dimension of the program (50) can be quite large, so that even though it is convex obtaining a solution requires care. We discuss now some practical considerations that we found necessary to consider when solving the problem (50). Due to the structure of $\Phi^{\text{E}}_{k}$, we do not need to declare every element in the matrix $J_{k}$ as a decision variable since many of these elements will be zero. For instance, we see that $\text{diag}(\nu^{\text{E}}_{k})\Phi^{\text{E}}_{k}$ is a block matrix, where further each matrix block is diagonal. We express this as: $\text{diag}(\nu^{\text{E}}_{k})\Phi^{\text{E}}_{k}=$ $\displaystyle\begin{bmatrix}B_{\text{\footnotesize{off}},\text{\footnotesize{off}}}[k]&B_{\text{\footnotesize{off}},\text{\footnotesize{on}}}[k]&\mathbf{0}&\mathbf{0}\\\ \mathbf{0}&\mathbf{0}&B_{\text{\footnotesize{on}},\text{\footnotesize{off}}}[k]&B_{\text{\footnotesize{on}},\text{\footnotesize{on}}}[k]\end{bmatrix}$ $\displaystyle\mathrel{\ensurestackMath{\stackunder[1pt]{=}{\scriptstyle\triangledown}}}\ \text{sparse}(J_{k})$ where, e.g., $B_{\text{\footnotesize{off}},\text{\footnotesize{off}}}[k]=\text{diag}(\nu^{\text{E}}_{\text{\footnotesize{off}}}[k])(I-\Phi^{\text{E}}_{\text{\footnotesize{off}}}[k])$. The other diagonal matrices appearing in (5.3.2) can be inferred by carrying out the matrix multiplication. If $J_{k}$ was declared directly as a decision variable the problem (50) would have $(8N^{2}+2N(\tau+1))T_{\text{plan}}$ primal variables, whereas the problem with $\text{sparse}(J_{k})$ as a decision variable only has $2NT_{\text{plan}}(\tau+3)$ primal variables. As an example, consider $N=12$, $T_{\text{plan}}=360$, and $\tau=5$, which are values used in numerical results reported later. The problem (50) without the structure exploited has $\approx 0.5$ million decision variables, but only $\approx 75,000$ when the structure is exploited. We also have found it helpful to include constraints of the form, $\displaystyle\phi^{\text{GS}}_{\text{\footnotesize{off}}}(\text{on}\ |\ j-1)\nu_{\text{\footnotesize{off}}}[\lambda^{j-1},0,k]\leq\phi^{\text{GS}}_{\text{\footnotesize{off}}}(\text{on}\ |\ j)\nu_{\text{\footnotesize{off}}}[\lambda^{j},0,k],$ (51) $\displaystyle\phi^{\text{GS}}_{\text{\footnotesize{on}}}(\text{off}\ |\ j+1)\nu_{\text{\footnotesize{on}}}[\lambda^{j+1},0,k]\leq\phi^{\text{GS}}_{\text{\footnotesize{on}}}(\text{off}\ |\ j)\nu_{\text{\footnotesize{on}}}[\lambda^{j},0,k],$ (52) so to suggest that the switching on (resp., switching off) probability increases as temperature increases (resp., decreases). Adding the constraints (51)-(52) to the problem (50) is straightforward as both the LHS and RHS of the inequalities are elements in the matrix $J_{k}$. Matlab implementation of (50) and the algorithm to extract the policies from $J_{k}$ (described in the proof of Lemma 5.2) is available at [8]. #### 5.3.3 Communication burden Once solved, the policies obtained from (50) need to be sent to each individual TCL. Many of the policy state values are constrained to either zero or one, which could be pre-programmed into each TCL. At each time index, $q-2$ (for the on to off policy) plus $N-m-1$ (for the off to on policy) numbers are not constrained and need to be sent from the BA to each TCL. Recall that the numbers $m$ and $q$ are temperature bin indices (see Figure 2) and $N$ is the number of temperature bins. For illustrative purposes, consider the values used in numerical experiments reported in the sequel: $N=12$ with $q=10$ and $m=2$ and a time discretization $\Delta t=1$ minute. Since $N=q+m$, then the BA has to broadcast $2(q-1)=18$ numbers every 1 minute to the TCLs. Each TCL receives the same 18 numbers. Communication from TCLs to the BA - about their temperature and on/off state - is needed at the beginning of every planning period so that the BA can determine the initial condition $\hat{\nu}$ in (50). The frequency of this feedback is a design choice. In our numerical simulations reported later, a planning horizon of 6 hours was used, and this feedback was necessary only once in six hours.More frequent loop closure may be needed for higher robustness to uncertainty in weather prediction etc., a topic outside the scope of this paper. ## 6 Numerical experiments Simulation involving coordination of ${\sf{N_{tcl}}}=20,000$ TCLs through our proposed framework is presented here. Recall the two parts of the coordination architecture shown in Figure 1: (i) planning and (ii) real time control. Planning refers to the solution of the problem (50) at the BA to compute the following two things for the planning period ${\sf T}$: 1. 1. $r_{k}$: the reference power consumption of the TCL collection, given the problem data $r^{\text{BA}}_{k}$. 2. 2. $\phi^{\text{GS},*}_{\text{\footnotesize{off}}}[k]$ and $\phi^{\text{GS},*}_{\text{\footnotesize{on}}}[k]$: grid support control policies for each TCL. This computation is performed at ${\sf T}(0)$. Real time control is then the implementation of the grid support policies by each TCL to make on/off decisions in real time. We imagine the BA broadcasts the policies $\phi^{\text{GS},*}_{\text{\footnotesize{off}}}[k]$ and $\phi^{\text{GS},*}_{\text{\footnotesize{on}}}[k]$ at each $k$, though it can also broadcast all the policies, for all $k\in{\sf T}$, at ${\sf T}(0)$ and not broadcast again until the beginning of the next planning horizon. The goal of the numerical simulations of real time control is to show the following. 1. 1. When each TCL uses the policies $\phi^{\text{GS},*}_{\text{\footnotesize{off}}}[k]$ and $\phi^{\text{GS},*}_{\text{\footnotesize{on}}}[k]$ to decide on/off actuation, the collection’s power demand indeed tracks $r_{k}$. 2. 2. Every TCL’s QoS constraints - both temperature and cycling - are satisfied at all times. Temperature of each TCL is computed in these simulations with the ODE model (1). Table 1: Simulation Parameters Par. | Unit | value | Par. | Unit | value ---|---|---|---|---|--- ${\sf{N_{tcl}}}$ | N/A | 2$\times 10^{4}$ | $\eta$ | $\frac{\text{kW-e}}{\text{kW-th.}}$ | $2.5$ $C$ | kWh$/^{\circ}$C | 1 | $P_{0}$ | kW | 5.5 $\lambda^{\text{min}}$ | ∘C | 20 | $\lambda^{\text{max}}$ | ∘C | $22$ $(\Delta t)\tau$ | Mins. | 5 | $P_{\text{\footnotesize{agg}}}$ | MW | 110 $R$ | ∘C$/$kW | 2 | $\Delta t$ | Mins. | 1 $q$ | N/A | 10 | $m$ | N/A | 2 $N$ | N/A | 12 | $T_{\text{plan}}$ | N/A | 360 ### 6.1 Planning The demand needed for demand-supply imbalance at the BA, $r^{\text{BA}}_{k}$, is chosen arbitrarily, and shown in Figure 7 (top). It is infeasible for the collection: sometimes negative and sometimes far higher than the maximum power demand of the collection. This is done to simulate a realistic scenario in which many sources of demand and generation, not just TCLs, are managed by the BA. The baseline demand trajectory is defined by the equation (4), which is approximately the power consumption for this collection of air conditioners under thermostat control. The ambient air temperature is time varying and is obtained from wunderground.com for a typical summer day in Gainesville, Florida, USA. The other parameters that affect the Markov model are shown in Table 1. Planning computations are done with Matlab and CVX [14] using a desktop Linux machine, with $N=12$, and for a six hour planning horizon with 1 minute discretization ($T_{\text{plan}}=360$). The problem (50) takes about a minute to solve. The quantity $r^{\text{BA}}_{k}$, the baseline power $\bar{P}_{k}$, and the reference signal $r_{k}$, obtained from solving (50), are shown in Figure 7 (top). The optimal reference for the collection, $r_{k}$, is as close to $r^{\text{BA}}_{k}$ as the dynamics of TCLs allows without violating their QoS constraints; recall Remark 5.1. Figure 7 (bottom) shows the two grid support control polices for one time instant. ### 6.2 Real time control The power consumption of the collection making on/off decisions according to the obtained policies is shown in Figure 8 (top). The figure shows that the TCLs are able to collectively track the reference signal $r_{k}$. We emphasize that the computational effort at each TCL is negligible. Recall Remark 4.5: once a TCL receives a grid support policy ($\approx$ 18 floating point numbers, see Section 5.3.3) it only has to measure its current state (temperature and on/off mode) and generate a uniformly distributed random number in $[0,1]$ to implement the policy. Verification of the grid support policies in ensuring QoS is shown in Figure 8. The bottom plots shows a histogram of the times between switches for 300 randomly chosen TCLs. The middle plot shows a histogram of temperature from 200 randomly chosen TCLs’ temperature trajectories. The histograms show that the policies designed with (50) indeed satisfy the QoS constraints, which is specified by the vertical lines in the figures. Some TCLs do escape the temperature deadband by a little bit, which is expected and occurs also in thermostatic control: the sensor must _first_ register a value outside the deadband in order decide to switch the on/off state. Figure 7: (Top): The quantity $r_{k}$ obtained from solving (50), the dashed horizontal lines represent all of the TCLs on (top line) and off (bottom line). (Bottom): Grid support control policies, obtained from solving (50), at one time instance. Figure 8: (Top): Reference tracking results for the TCLs under the influence of the grid support control policies obtained by solving (50). (Middle): Histogram of the 200 TCL’s temperature trajectories over the entire simulation horizon. (Bottom): Histogram of the time between switches over 3000 TCLs with the vertical line representing the minimum allowable time between switches. ## 7 Conclusion In this work we present a unified framework for the distributed control of TCLs. The framework enables: (i) reference planning for a collection of TCLs and (ii) design of a randomized control policy for the individual TCLs, so that both the BA’s requirement and consumers’ QoS are satisfied. The resulting framework is (i) scalable to an arbitrary number of loads and is implemented through _local_ feedback and minimal communication, (ii) able to guarantee both temperature and cycling constraints maintenance in each TCL, and (iii) based on convex optimization. Matlab/cvx implementation is publicly available [8]. There are several avenues for future work. The optimal control problem is solved in an open-loop fashion here. Feedback from TCLs is used only to compute an initial condition that is needed as problem data for the off- line planning problem. It is straightforward to close the loop between the TCL collection and the BA with greater frequency for robustness to uncertainty in weather forecast and TCL parameters. It will be of interest to identify scenarios where closing loop, say, by using Model Predictive Control, is (i) necessary, and (ii) at what frequency should information be communicated from the TCLs to the BA. Another avenue is to investigate how the problem (50) could be solved at each TCL, intermittently, instead of at the BA. Since the computational power of the processor at each TCL is lower than that of the processor at the BA, online distributed algorithms for convex optimization could play a role. The Fokker-Planck equations from [21] we used here are convenient for modeling TCL populations with a small deegree of heterogeneity. Distributed computation of optimal policies locally at each TCLs may help extend the method to a highly heterogeneous population of TCLs. ## References * [1] Saeid Bashash and Hosam K Fathy. Modeling and control of aggregate air conditioning loads for robust renewable power management. IEEE Transactions on Control Systems Technology, 21(4):1318–1327, 2013. * [2] Emilio Benenati, Marcello Colombino, and Emiliano Dall’Anese. A tractable formulation for multi-period linearized optimal power flow in presence of thermostatically controlled loads. In 2019 IEEE 58th Conference on Decision and Control (CDC), pages 4189–4194. IEEE, 2019. * [3] Ana Bušić and Sean Meyn. Distributed randomized control for demand dispatch. In IEEE conference on decision and control, pages 6964–6971, 2016\. * [4] D.S. Callaway and I.A. Hiskens. Achieving controllability of electric loads. Proceedings of the IEEE, 99(1):184–199, 2011. * [5] Yue Chen, Ana Bušić, and Sean P. Meyn. State estimation for the individual and the population in mean field control with application to demand dispatch. IEEE Transactions on Automatic Control, 62(3):1138–1149, 2017. * [6] Yue Chen, Md Umar Hashmi, Joel Mathias, Ana Bušić, and Sean Meyn. Distributed control design for balancing the grid using flexible loads. In IMA Volume on the Control of Energy Markets and Grids, pages 1–26. 2017. * [7] Chi-Min Chu and Tai-Lang Jong. A novel direct air-conditioning load control method. IEEE transactions on power systems, 23(3):1356–1363, 2008. * [8] Austin Coffman. Distributed control of TCLs through convex optimization. https://gitlab.com/austinrc925/distributed-control-of-tcls-through-convex-optimization/-/tree/main, 2021\. * [9] Austin Coffman, Ana Bušić, and Prabir Barooah. Virtual energy storage from TCLs using QoS preserving local randomized control. In 5th ACM International Conference on Systems for Built Environments (BuildSys), page 10, November 2018. * [10] Austin Coffman, Neil Cammardella, Prabir Barooah, and Sean P. Meyn. Aggregate capacity of TCLs with cycling constraints. ArXiV.org, October 2020. arXiv:1909.11497. * [11] Austin R. Coffman, Ana Bušić, and Prabir Barooah. Aggregate capacity for TCLs providing virtual energy storage with cycling constraints. In IEEE Conference on Decision and Control, December 2019. * [12] Austin R. Coffman, Ana Bušić, and Prabir Barooah. Control oriented modeling of TCLs. In American Control Conference, May 2021. accepted, extended version in arXiv:2009.12960. * [13] Austin R. Coffman, Zhong Guo, and Prabir Barooah. Characterizing capacity of flexible loads for providing grid support. IEEE Transactions on Power Systems, 36:2428 – 2437, May 2021\. * [14] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex programming, version 1.21. http://cvxr.com/cvx, February 2011. * [15] H. Hao, B. M. Sanandaji, K. Poolla, and T. L. Vincent. Aggregate flexibility of thermostatically controlled loads. IEEE Transactions on Power Systems, 30(1):189–198, Jan 2015. * [16] Emre C Kara, Mario Bergés, and Gabriela Hug. Impact of disturbances on modeling of thermostatically controlled loads for demand response. IEEE Transactions on Smart Grid, 6(5):2560–2568, 2015. * [17] Adil Khurram, Roland Malhamé, Luis Duffaut Espinosa, and Mads Almassalkhi. Identification of hot water end-use process of electric water heaters from energy measurements. Electric Power Systems Research, 189:106625, 2020. * [18] Zachary E. Lee, Qingxuan Sun, Zhao Ma, Jiangfeng Wang, Jason S. MacDonald, and K. Max Zhang. Providing Grid Services With Heat Pumps: A Review. ASME Journal of Engineering for Sustainable Buildings and Cities, 1(1), 01 2020. 011007\. * [19] M. Liu and Y. Shi. Model predictive control of aggregated heterogeneous second-order thermostatically controlled loads for ancillary services. IEEE Transactions on Power Systems, 31(3):1963–1971, May 2016. * [20] M. Liu, Y. Shi, and X. Liu. Distributed MPC of aggregated heterogeneous thermostatically controlled loads in smart grid. IEEE Transactions on Industrial Electronics, 63(2):1120–1129, 2016\. * [21] Roland Malhame and Chee-Yee Chong. Electric load model synthesis by diffusion approximation of a high-order hybrid-state stochastic system. IEEE Transactions on Automatic Control, 30(9):854–860, 1985. * [22] Alan S. Manne. Linear programming and sequential decisions. Management Science, 6(3):259–267, 1960. * [23] J. L. Mathieu, S. Koch, and D. S. Callaway. State estimation and control of electric loads to manage real-time energy imbalance. IEEE Transactions on Power Systems, 28:430–440, 2013. * [24] Md Salman Nazir and Ian Hiskens. Analysis of synchronization in load ensembles. Electric Power Systems Research, 190:106779, 2021. * [25] Dario Paccagnan, Maryam Kamgarpour, and John Lygeros. On the range of feasible power trajectories for a population of thermostatically controlled loads. In 2015 54th IEEE Conference on Decision and Control (CDC), pages 5883–5888, 2015. * [26] L. C. Totu, R. Wisniewski, and J. Leth. Demand response of a TCL population using switching-rate actuation. IEEE Transactions on Control Systems Technology, 25(5):1537–1551, 2017. * [27] Henk Kaarle Versteeg and Weeratunge Malalasekera. An introduction to computational fluid dynamics: the finite volume method. Pearson education, 2007. * [28] Wei Zhang, Jianming Lian, Chin-Yao Chang, and Karanjit Kalsi. Aggregated modeling and control of air conditioning loads for demand response. IEEE Transactions on Power Systems, 28(4):4655–4664, 2013. ## Appendix A Proofs ### A.1 Proof of Lemma 1 See Appendix B before reading this proof. Property (ii) is a consequence of the upwind difference scheme used. We see that for the internal CVs we have off CVs: $\displaystyle\quad-\Big{(}F^{i,+}_{\text{\footnotesize{off}}}+D\Big{)}$ (A.61) on CVs: $\displaystyle\quad\Big{(}F^{i,-}_{\text{\footnotesize{on}}}-D\Big{)}.$ (A.62) From Assumption A.2 we have that $F^{i,-}_{\text{\footnotesize{on}}}\leq 0$ and $F^{i,-}_{\text{\footnotesize{off}}}\geq 0$ so that both of the above terms are negative. The upwind scheme is what ensured appropriate sign was added to the terms $F^{i,-}_{\text{\footnotesize{on}}}$ and $F^{i,-}_{\text{\footnotesize{off}}}$ so that the above coefficients are negative. Similar arguments can be applied for the off diagonal terms of the internal CVs and the boundary CVs. To show property (i) we consider solely an internal CV for the off state as the arguments for all other CVs are identical in structure. Note that showing $A(t)\mathbb{1}=\mathbf{0}$ is equivalent to $\mathbb{1}^{T}\mathcal{A}(t)=\mathbf{0}^{T}$. Hence we need to show, for an arbitrary $i$ that all coefficients acting on $\nu_{\text{\footnotesize{off}}}(\lambda^{i},t)$ sum to $0$. We collect the coefficients corresponding to $\nu_{\text{\footnotesize{off}}}(\lambda^{i},t)$: $\displaystyle\text{From CV($i$)}:$ $\displaystyle\quad-F^{i,+}_{\text{\footnotesize{off}}}(t)-D.$ $\displaystyle\text{From CV($i-1$)}:$ $\displaystyle\quad\frac{D}{2}.\quad$ $\displaystyle\text{From CV($i+1$)}:$ $\displaystyle\quad\frac{D}{2}+F^{i+1,-}_{\text{\footnotesize{off}}}(t).\quad$ We then require the sum of these coefficients to be zero for all $t$ and any index $i$ for the internal off CVs, adding yields $\displaystyle F^{i+1,-}_{\text{\footnotesize{off}}}(t)-F^{i,+}_{\text{\footnotesize{off}}}(t)=\frac{f_{\text{\footnotesize{off}}}(\lambda^{i+1,-},t)-f_{\text{\footnotesize{off}}}(\lambda^{i,+},t)}{\Delta\lambda}=0$ since by construction $\lambda^{i+1,-}=\lambda^{i,+}$ for the off CV’s. This procedure can be repeated for $\nu_{\text{\footnotesize{off}}}(\lambda^{i},t)$ with $i\in\\{1,m,N\\}$, i.e., the boundary CVs in the off state and all of the on CVs in a similar fashion. ### A.2 Proof of Lemma 3 If $\alpha=(\Delta t)^{-1}$, the diagonal elements of $A_{k}$ with $\alpha$ in them will go to zero and the non diagonal elements will go to 1. These non- diagonal elements with value $1$ are the red dots in Figure 3 and encapsulate the thermostat control law. Thus the construction of $\Phi^{\text{\footnotesize{TS}}}$ with the canonical basis vectors. Now, multiplying out the matrix we have, $\displaystyle\Phi^{\text{\footnotesize{TS}}}G_{k}=\begin{bmatrix}\big{(}I-\Phi^{\text{\footnotesize{TS}}}_{\text{\footnotesize{off}}}\big{)}P_{k}^{\text{\footnotesize{off}}}&\Phi^{\text{\footnotesize{TS}}}_{\text{\footnotesize{off}}}S_{k}^{\text{\footnotesize{off}}}\\\ \Phi^{\text{\footnotesize{TS}}}_{\text{\footnotesize{on}}}S_{k}^{\text{\footnotesize{on}}}&\big{(}I-\Phi^{\text{\footnotesize{TS}}}_{\text{\footnotesize{on}}}\big{)}P_{k}^{\text{\footnotesize{on}}}\end{bmatrix}$ (A.63) where $\big{(}I-\Phi^{\text{\footnotesize{TS}}}_{\text{\footnotesize{off}}}\big{)}P_{k}^{\text{\footnotesize{off}}}$ (respectively, $\big{(}I-\Phi^{\text{\footnotesize{TS}}}_{\text{\footnotesize{on}}}\big{)}P_{k}^{\text{\footnotesize{on}}}$) is the matrix $P_{k}^{\text{\footnotesize{off}}}$ (respectively, $P_{k}^{\text{\footnotesize{on}}}$) but with the last (respectively, first) row zeroed out. The exact opposite statement is true for $\Phi^{\text{\footnotesize{TS}}}_{\text{\footnotesize{off}}}P_{k}^{\text{\footnotesize{on}}}$ and $\Phi^{\text{\footnotesize{TS}}}_{\text{\footnotesize{on}}}P_{k}^{\text{\footnotesize{off}}}$. Hence, by definition of the matrices in $G_{k}$ we have $P_{k}=\Phi^{\text{\footnotesize{TS}}}G_{k}$ where each non-zero element holds the interpretation (23). ### A.3 Proof of Lemma 4 We define the following transformation for $l\in\\{0,\dots,\tau\\}$ and $j\in\\{1,\dots,N\\}$ as $\displaystyle T(j,l)=lN+j$ (A.64) that maps the integers $j$ and $l$ that label the state values to the absolute index of either of the vectors $\nu^{\text{E}}_{\text{\footnotesize{off}}}$ and $\nu^{\text{E}}_{\text{\footnotesize{on}}}$. Now consider the following two sets $\displaystyle\mathcal{W}_{\text{\footnotesize{off}}}$ $\displaystyle\triangleq\Big{\\{}(u,j,l)\in{\sf X}\ \Big{|}\ \phi^{\text{E}}_{\text{\footnotesize{off}}}(u\ |\ j,\ l)=\beta_{\text{\footnotesize{off}}}(u,j,l)\Big{\\}}$ (A.65) $\displaystyle\mathcal{W}_{\text{\footnotesize{on}}}$ $\displaystyle\triangleq\Big{\\{}(u,j,l)\in{\sf X}\ \Big{|}\ \phi^{\text{E}}_{\text{\footnotesize{on}}}(u\ |\ j,\ l)=\beta_{\text{\footnotesize{on}}}(u,j,l)\Big{\\}}.$ (A.66) The values $\beta_{\text{\footnotesize{off}}}$ and $\beta_{\text{\footnotesize{on}}}$ are chosen to ensure the structural requirements in (31)-(32) and (35). For example, for $l=1$ and $u=\text{on}$ we have that $\beta_{\text{\footnotesize{off}}}(\text{\footnotesize{on}},\cdot,1)=\phi^{\text{TS}}_{\text{\footnotesize{off}}}(\text{\footnotesize{on}}|\cdot)$ (and hence $\beta_{\text{\footnotesize{off}}}(\text{\footnotesize{off}},\cdot,1)=1-\phi^{\text{TS}}_{\text{\footnotesize{off}}}(\text{\footnotesize{on}}|\cdot)$) so to enforce the structural requirement in (35). Define for each $k\in{\sf T}$ , $u,v\in\\{\text{\footnotesize{on}},\text{\footnotesize{off}}\\}$, $j\in\\{1,\dots,N\\}$, and $l\in\\{0,\dots,\tau\\}$ the following vectors $\displaystyle h^{u}_{k}[T(j,l)]$ $\displaystyle\triangleq\begin{cases}(\nu_{u}^{\text{E}}[\lambda^{j},l,k])^{-1}&\text{if}\ \nu_{u}^{\text{E}}[\lambda^{j},l,k]>0.\\\ 0&\text{otherwise}.\end{cases}$ $\displaystyle w^{u,v}_{k}[T(j,l)]$ $\displaystyle\triangleq\begin{cases}\beta_{v}(u,j,l)&\text{if}\ (u,j,l)\in\mathcal{W}_{v}\ \text{and}\\\ &\nu_{v}^{\text{E}}[\lambda^{j},l,k]=0.\\\ 0.5&\text{if}\ (u,j,l)\notin\mathcal{W}_{v}\ \text{and}\\\ &\nu_{v}^{\text{E}}[\lambda^{j},l,k]=0.\\\ 0&\text{otherwise}.\end{cases}$ where $T(\cdot,\cdot)$ is defined in (A.64), $\mathcal{W}_{\text{\footnotesize{off}}}$ in (A.65), and $\mathcal{W}_{\text{\footnotesize{on}}}$ in (A.66). Let $W^{u,v}_{k}=\text{diag}(w^{u,v}_{k})$ and $H^{u,v}_{k}=\text{diag}(h^{u,v}_{k})$, and construct the following matrices $\displaystyle H_{k}$ $\displaystyle=\begin{bmatrix}H_{k}^{\text{\footnotesize{off}}}&\mathbf{0}\\\ \mathbf{0}&H_{k}^{\text{\footnotesize{on}}}\end{bmatrix},\quad\text{and}$ (A.67) $\displaystyle W_{k}$ $\displaystyle=\begin{bmatrix}W^{\text{\footnotesize{off}},\text{\footnotesize{off}}}_{k}&W^{\text{\footnotesize{off}},\text{\footnotesize{on}}}_{k}&\mathbf{0}&\mathbf{0}\\\ \mathbf{0}&\mathbf{0}&W^{\text{\footnotesize{on}},\text{\footnotesize{off}}}_{k}&W^{\text{\footnotesize{on}},\text{\footnotesize{on}}}_{k}\end{bmatrix}.$ (A.68) We first show that $\Phi_{k}^{\text{E}}=H_{k}J_{k}+W_{k}$ satisfies (48). Note that $\text{diag}(\nu_{k}^{\text{E}})W_{k}=\mathbf{0}$ since by construction if the $i^{th}$ row of $W_{k}$ has a non zero entry then the $i^{th}$ diagonal entry of $\text{diag}(\nu_{k}^{\text{E}})$ is zero. The product $\text{diag}(\nu_{k}^{\text{E}})H_{k}$ is a diagonal matrix with with entries of either zero or one. The zero entries also correspond to the zero entries of $\nu_{k}^{\text{E}}$. In this case, the respective entry in $J_{k}$ is also zero so that $\text{diag}(\nu_{k}^{\text{E}})H_{k}J_{k}=J_{k}$ as desired. We now show that $\phi_{k}^{\text{E}}=H_{k}J_{k}+W_{k}\in\varPhi$. First consider an arbitrary state indexed by $(\text{\footnotesize{off}},j,l)$ at time $k$, if the corresponding value in $\nu_{\text{\footnotesize{off}}}[\lambda^{j},l,k]>0$ then the two policy values are defined as $\displaystyle\frac{{\sf P}\left(m_{k+1}=\text{on},\ I_{k}=j,\ {\sf{L}}_{k}=l,\ m_{k}=\text{off}\right)}{\nu_{\text{\footnotesize{off}}}[\lambda^{j},l,k]}$ (A.69) $\displaystyle\frac{{\sf P}\left(m_{k+1}=\text{off},\ I_{k}=j,\ {\sf{L}}_{k}=l,\ m_{k}=\text{off}\right)}{\nu_{\text{\footnotesize{off}}}[\lambda^{j},l,k]}.$ (A.70) If either of the above values are fixed in the constraint set $\Phi$, then the constraint (49) will ensure this. Further, since we have that $\nu_{k}^{\text{E}},J_{k}\in[0,1]$ and that $(\nu_{k}^{\text{E}})^{T}=J_{k}\mathbb{1}$ this ensures that the above policy values are within $[0,1]$ and sum to 1. The above argument is valid for any pair of state values such that the corresponding value of $\nu_{k}^{\text{E}}$ is non-zero. If the corresponding value of $\nu_{\text{\footnotesize{off}}}[\lambda^{j},l,k]=0$ and the policy (conditioned on this state value) has a constraint, the first if case in the definition of $w^{u,v}_{k}$ ensures this constraint. Further, the constraint values must also be chosen to ensure the respective policy values are in $[0,1]$ and sum to one. Lastly, if $\nu_{\text{\footnotesize{off}}}[\lambda^{j},l,k]=0$ and there is no constraint for the policy conditioned on this state value the second if case in the definition of $w^{u,v}_{k}$ ensures the policy value sums to 1 and the respective elements are in $[0,1]$. Thus $\Phi_{k}^{\text{E}}\in\varPhi$ for all $k\in{\sf T}$. ### A.4 Proof of Theorem 1 The proof structure is similar to the one in [2]. The idea is to exploit the fact that: (i) $\nu_{k}^{\text{E}}$ is a decision variable for both optimization problems (50) and (44) and (ii) the objective function is the same for both problems and solely a function of the marginal $\nu_{k}^{\text{E}}$. We rewrite these problem compactly below, $\displaystyle\eta^{*}_{\text{CVX}}$ $\displaystyle=\min_{(\nu^{\text{E}},J)\in X}\eta(\nu^{\text{E}}),$ (A.71) $\displaystyle\eta^{*}_{\text{NCVX}}$ $\displaystyle=\min_{(\nu^{\text{E}},\Phi^{\text{E}})\in Y}\eta(\nu^{\text{E}}),$ (A.72) where the sets $X$ and $Y$ collect all of the relevant constraints for the problems. The variables $\nu^{\text{E}}$, $\Phi^{\text{E}}$, and $J$ are concatenated over the considered finite time horizon and hence are not sub- scripted by $k$. We proceed by showing that $\eta^{*}_{\text{CVX}}\leq\eta^{*}_{\text{NCVX}}$ and $\eta^{*}_{\text{NCVX}}\leq\eta^{*}_{\text{CVX}}$ to give the desired result. #### A.4.1 $\eta^{*}_{\text{CVX}}\leq\eta^{*}_{\text{NCVX}}$ Pick any argument minimizer that achieves value $\eta^{*}_{\text{NCVX}}$ and denote the pair as $(\nu^{\text{E}}_{\text{NCVX}},\Phi^{\text{E}}_{\text{NCVX}})$. Trivially construct $J$ through the relation (48) so that this constructed $J$ and $\nu^{\text{E}}_{\text{NCVX}}$ (that is optimal for (44)) are also feasible for (50), i.e., $(\nu^{\text{E}}_{\text{NCVX}},J)\in X$. This is since $\mathbb{1}^{T}\text{diag}(\nu^{\text{E}}_{k})=\nu^{\text{E}}_{k}$ and $\mathbb{1}=\Phi^{\text{E}}_{k}\mathbb{1}$. Hence we have that $\displaystyle\eta^{*}_{\text{CVX}}$ $\displaystyle=\min_{(\nu^{\text{E}},J)\in X}\eta(\nu^{\text{E}})\leq\eta(\nu^{\text{E}}_{\text{NCVX}})=\eta^{*}_{\text{NCVX}}$ (A.73) since by definition $\eta^{*}_{\text{CVX}}$ is the minimum value over the set of feasible solutions. #### A.4.2 $\eta^{*}_{\text{NCVX}}\leq\eta^{*}_{\text{CVX}}$ We take a pair $(\nu^{\text{E}}_{\text{CVX}},J_{\text{CVX}})$ that achieve optimal cost $\eta^{*}_{\text{CVX}}$ and construct a feasible solution for (44), denoted $(\eta^{\text{E}}_{\text{NCVX}},\Phi^{\text{E}}_{\text{NCVX}})$, as follows (for each $k$) $\displaystyle\Phi^{\text{E}}_{k,\text{NCVX}}$ $\displaystyle=H_{k}J_{k}+W_{k},\quad\text{and}$ (A.74) $\displaystyle\nu^{\text{E}}_{k,\text{NCVX}}$ $\displaystyle=\nu^{\text{E}}_{k,\text{CVX}}.$ (A.75) Where $H_{k}$ and $W_{k}$ are defined in Lemma 5.2. This constructed solution is then feasible for (44) as the constraint $\Phi^{\text{E}}_{\text{NCVX}}\in\varPhi$ is part of the result in Lemma 5.2 and $\displaystyle\nu^{\text{E}}_{k,\text{NCVX}}\Phi^{\text{E}}_{k,\text{NCVX}}G_{k}^{\text{E}}$ $\displaystyle=\nu^{\text{E}}_{k,\text{NCVX}}\big{(}H_{k}J_{k}+W_{k}\big{)}G_{k}^{\text{E}}$ (A.76) $\displaystyle=\mathbb{1}^{T}J_{k}G_{k}^{\text{E}}=\nu^{\text{E}}_{k+1,\text{NCVX}}.$ (A.77) The fact that $\nu^{\text{E}}_{k,\text{NCVX}}\big{(}H_{k}J_{k}+W_{k}\big{)}=\mathbb{1}^{T}J_{k}$ is since $\nu^{\text{E}}_{k,\text{NCVX}}W_{k}=\mathbf{0}$ and $\nu^{\text{E}}_{k,\text{NCVX}}H_{k}J_{k}=\mathbb{1}^{T}J_{k}$. The matrix $W_{k}$ only has non zero entries for row indices where the index of the row vector $\nu^{\text{E}}_{k,\text{NCVX}}$ is zero so that the resulting product is the zero vector. The product $\nu^{\text{E}}_{k,\text{NCVX}}H_{k}$ is a vector of ones and zeros, specifically, if the $i^{th}$ element of this vector is zero then the entire $i^{th}$ column of the matrix $J_{k}$ will be the zero vector. Thus the equivalence between $\nu^{\text{E}}_{k,\text{NCVX}}\big{(}H_{k}J_{k}+W_{k}\big{)}$ and $\mathbb{1}^{T}J_{k}$. Since the constructed solution is feasible we have that $\displaystyle\eta^{*}_{\text{NCVX}}$ $\displaystyle=\min_{(\nu^{\text{E}},\Phi^{\text{E}})\in Y}\eta(\nu^{\text{E}})\leq\eta(\nu^{\text{E}}_{\text{CVX}})=\eta^{*}_{\text{CVX}}$ (A.78) since by definition $\eta^{*}_{\text{NCVX}}$ is the minimum value over the set of feasible solutions. ## Appendix B PDE discretization We denote the $i^{th}$ CV as CV($i$) and further adopt the following notational simplifications, $\displaystyle\mu_{\text{\footnotesize{off}}}(\lambda^{i},t)\triangleq\mu_{\text{\footnotesize{off}}}(\lambda^{i}_{\text{\footnotesize{off}}},t),\quad\text{and}\quad\mu_{\text{\footnotesize{on}}}(\lambda^{i},t)\triangleq\mu_{\text{\footnotesize{on}}}(\lambda^{i}_{\text{\footnotesize{on}}},t).$ Highlighted red in Figure 2 are the two control volumes to assist in enforcing boundary conditions that coincide with the thermostat policy (6). This is discussed further in Appendix B.2 when the boundary conditions CVs are discretized. ### B.1 Internal CV’s Consider the RHS of the pde (9) integrated over CV($i$): $\displaystyle\int_{\text{CV(i)}}\bigg{(}\frac{\sigma^{2}}{2}\frac{\partial^{2}}{\partial\lambda^{2}}\big{(}\mu_{\text{\footnotesize{on}}}(\lambda,t)\big{)}-\frac{\partial}{\partial\lambda}\big{(}f_{\text{\footnotesize{on}}}(\lambda,t)\mu_{\text{\footnotesize{on}}}(\lambda,t)\big{)}\bigg{)}d\lambda$ $\displaystyle=\bigg{(}\frac{\sigma^{2}}{2}\frac{\partial}{\partial\lambda}\mu_{\text{\footnotesize{on}}}(\lambda,t)-f_{\text{\footnotesize{on}}}(\lambda,t)\mu_{\text{\footnotesize{on}}}(\lambda,t)\bigg{)}\bigg{|}_{\lambda^{i,-}}^{\lambda^{i,+}},$ (B.79) where equality is by the divergence theorem [27]. Note, the points $\lambda^{i,-}$ and $\lambda^{i,+}$ are not control volume variables, but rather the boundaries of a single control volume. Hence, quantities in (B.79) need to be approximated in terms of the nodal points of the neighboring control volumes. The approximations for the partial derivative are, $\displaystyle\frac{\partial}{\partial\lambda}\mu_{\text{\footnotesize{on}}}(\lambda^{i,+},t)$ $\displaystyle\approx\frac{\mu_{\text{\footnotesize{on}}}(\lambda^{i+1},t)-\mu_{\text{\footnotesize{on}}}(\lambda^{i},t)}{\Delta\lambda},\quad\text{and}$ (B.80) $\displaystyle\frac{\partial}{\partial\lambda}\mu_{\text{\footnotesize{on}}}(\lambda^{i,-},t)$ $\displaystyle\approx\frac{\mu_{\text{\footnotesize{on}}}(\lambda^{i},t)-\mu_{\text{\footnotesize{on}}}(\lambda^{i-1},t)}{\Delta\lambda}.$ (B.81) For the integrated convective term, we use the so-called upwind scheme [27]. This scheme elects the FVM equivalent of a forward or backward difference based on the sign of the convective velocity $f_{\text{on}}(\lambda,t)$. By assumption A.2, $f_{\text{\footnotesize{on}}}(\lambda,t)\leq 0$ and the upwind scheme prescribes: $\displaystyle f_{\text{\footnotesize{on}}}(\lambda^{i,-},t)\mu_{\text{\footnotesize{on}}}(\lambda^{i,-},t)$ $\displaystyle=f_{\text{\footnotesize{on}}}(\lambda^{i,-},t)\mu_{\text{\footnotesize{on}}}(\lambda^{i},t),\quad\text{and}$ $\displaystyle f_{\text{\footnotesize{on}}}(\lambda^{i,+},t)\mu_{\text{\footnotesize{on}}}(\lambda^{i,+},t)$ $\displaystyle=f_{\text{\footnotesize{on}}}(\lambda^{i,+},t)\mu_{\text{\footnotesize{on}}}(\lambda^{i+1},t).$ (B.82) When the TCL is off $f_{\text{\footnotesize{off}}}(\lambda,t)\geq 0$ (also by Assumption A.2) the upwind scheme prescribes: $\displaystyle f_{\text{\footnotesize{off}}}(\lambda^{i,-},t)\mu_{\text{\footnotesize{off}}}(\lambda^{i,-},t)$ $\displaystyle=f_{\text{\footnotesize{off}}}(\lambda^{i,-},t)\mu_{\text{\footnotesize{off}}}(\lambda^{i-1},t),\ \text{and}$ $\displaystyle f_{\text{\footnotesize{off}}}(\lambda^{i,+},t)\mu_{\text{\footnotesize{off}}}(\lambda^{i,+},t)$ $\displaystyle=f_{\text{\footnotesize{off}}}(\lambda^{i,+},t)\mu_{\text{\footnotesize{off}}}(\lambda^{i},t).$ (B.83) Now returning to the discretization of the PDE (9) over an arbitrary internal CV. We approximate the LHS of (9) integrated over the control volume as, $\displaystyle\int_{\text{CV(i)}}\frac{\partial}{\partial t}\mu_{\text{\footnotesize{on}}}(\lambda,t)d\lambda\approx\frac{d}{dt}\mu_{\text{\footnotesize{on}}}(\lambda^{i},t)\Delta\lambda=\frac{d}{dt}\nu_{\text{\footnotesize{on}}}(\lambda^{i},t),$ where we have defined $\displaystyle\nu_{\text{\footnotesize{on}}}(\lambda^{i},t)\triangleq\mu_{\text{\footnotesize{on}}}(\lambda^{i},t)\Delta\lambda.$ (B.84) Now, denote the following $\displaystyle D\triangleq\frac{\sigma^{2}}{(\Delta\lambda)^{2}},\quad\text{and}\quad F_{\text{\footnotesize{on}}}^{i}(t)\triangleq\frac{f_{\text{\footnotesize{on}}}(\lambda^{i},t)}{\Delta\lambda},$ (B.85) where the quantities $F_{\text{\footnotesize{off}}}^{i}(t)$, $F_{\text{\footnotesize{on}}}^{i,+}(t)/F_{\text{\footnotesize{off}}}^{i,+}(t)$, and $F_{\text{\footnotesize{on}}}^{i,-}(t)/F_{\text{\footnotesize{off}}}^{i,-}(t)$ are defined similarly to $F_{\text{\footnotesize{on}}}^{i}(t)$, e.g., $F_{\text{\footnotesize{off}}}^{i,+}(t)\triangleq f_{\text{\footnotesize{off}}}(\lambda^{i,+},t)/\Delta\lambda$. Now equating the approximation of the RHS (9) with the approximation of the LHS of (9) we have, $\displaystyle\frac{d}{dt}\nu_{\text{\footnotesize{on}}}(\lambda^{i},t)$ $\displaystyle=\Big{(}F_{\text{\footnotesize{on}}}^{i,-}(t)-D\Big{)}\nu_{\text{\footnotesize{on}}}(\lambda^{i},t)+\frac{D}{2}\nu_{\text{\footnotesize{on}}}(\lambda^{i-1},t)$ $\displaystyle+\Big{(}\frac{D}{2}-F_{\text{\footnotesize{on}}}^{i,+}(t)\Big{)}\nu_{\text{\footnotesize{on}}}(\lambda^{i+1},t).$ (B.86) The spatial discretization for the PDE (10) is similar and yields, $\displaystyle\frac{d}{dt}\nu_{\text{\footnotesize{off}}}(\lambda^{i},t)$ $\displaystyle=\frac{D}{2}\nu_{\text{\footnotesize{off}}}(\lambda^{i+1},t)-\Big{(}F_{\text{\footnotesize{off}}}^{i,+}(t)+D\Big{)}\nu_{\text{\footnotesize{off}}}(\lambda^{i},t)$ $\displaystyle+\Big{(}\frac{D}{2}+F_{\text{\footnotesize{off}}}^{i,-}(t)\Big{)}\nu_{\text{\footnotesize{off}}}(\lambda^{i-1},t),$ (B.87) where $\nu_{\text{\footnotesize{off}}}(\lambda^{i},t)\triangleq\mu_{\text{\footnotesize{off}}}(\lambda^{i},t)\Delta\lambda.$ ### B.2 Boundary CV’s The boundary CVs are the CVs associated with the nodal values: $\lambda_{\text{\footnotesize{on}}}^{1}$, $\lambda_{\text{\footnotesize{on}}}^{q}$, $\lambda_{\text{\footnotesize{on}}}^{N}$, $\lambda_{\text{\footnotesize{off}}}^{1}$, $\lambda_{\text{\footnotesize{off}}}^{m}$, and $\lambda_{\text{\footnotesize{off}}}^{N}$. The superscript, for example the integer $q$ in $\lambda_{\text{\footnotesize{on}}}^{q}$ represents the CV index. All boundary CVs can be seen in Figure 2. Discretization of the boundary CVs requires care for atleast two reasons. First, this is typically where one introduces the BCs of the PDE into the numerical approximation. Secondly, on finite domains the endpoints present challenges, for example, there is no variable $\mu_{\text{\footnotesize{on}}}(\lambda^{N+1},t)$ for computation of the derivative values for node $\lambda^{N}_{\text{\footnotesize{on}}}$. The BC’s for the coupled PDEs (9)-(10) are [21]: Absorbing Boundaries: $\displaystyle\qquad\qquad\qquad\mu_{\text{\footnotesize{on}}}(\lambda^{\text{min}},t)=\mu_{\text{\footnotesize{off}}}(\lambda^{\text{max}},t)=0.$ (B.88) Conditions at Infinity: $\displaystyle\qquad\qquad\qquad\mu_{\text{\footnotesize{on}}}(+\infty,t)=\mu_{\text{\footnotesize{off}}}(-\infty,t)=0.$ (B.89) Conservation of Probability: $\displaystyle\frac{\partial}{\partial\lambda}\bigg{[}\mu_{\text{\footnotesize{on}}}(\lambda^{q,-},t)-\mu_{\text{\footnotesize{on}}}(\lambda^{q-1,+},t)-\mu_{\text{\footnotesize{off}}}(\lambda^{N-1,+},t)\bigg{]}=0.$ (B.90) $\displaystyle\frac{\partial}{\partial\lambda}\bigg{[}\mu_{\text{\footnotesize{off}}}(\lambda^{m,+},t)-\mu_{\text{\footnotesize{on}}}(\lambda^{2,-},t)-\mu_{\text{\footnotesize{off}}}(\lambda^{m+1,-},t)\bigg{]}=0.$ (B.91) Continuity: $\displaystyle\qquad\qquad\qquad\mu_{\text{\footnotesize{on}}}(\lambda^{q,-},t)=\mu_{\text{\footnotesize{on}}}(\lambda^{q-1,+},t).$ (B.92) $\displaystyle\qquad\qquad\qquad\mu_{\text{\footnotesize{off}}}(\lambda^{m,+},t)=\mu_{\text{\footnotesize{off}}}(\lambda^{m+1,-},t).$ (B.93) As we will see, implementation of some of the above conditions will require a bit of care. However, some are quite trivial to enforce. For example, by default, the continuity conditions (B.92) and (B.93) are satisfied due to our choice of CV structure, since, for example, for any $i$ we have $\lambda_{\text{\footnotesize{off}}}^{i,-}=\lambda_{\text{\footnotesize{off}}}^{i-1,+}$ and $\lambda_{\text{\footnotesize{off}}}^{i,+}=\lambda_{\text{\footnotesize{off}}}^{i+1,-}$. Now focusing on the conditions at infinity BC (B.89), we enforce instead the following conditions: $\displaystyle\frac{\partial}{\partial\lambda}\mu_{\text{\footnotesize{off}}}(\lambda^{1,-},t)=0,\quad\text{and}\quad\frac{\partial}{\partial\lambda}\mu_{\text{\footnotesize{on}}}(\lambda^{N,+},t)=0.$ (B.94) Our computational domain cannot extend to infinity, where the BC (B.89) is required to hold, but the temperature values $\lambda_{\text{\footnotesize{off}}}^{1}$ and $\lambda_{\text{\footnotesize{on}}}^{N}$ are quite far away from the deadband and so the density here will be near zero. Now, consider the spatial discretization of the CVs associated with the BC at infinity. First considering the CV associated with the temperature $\lambda^{1}_{\text{\footnotesize{off}}}$, we have that the differential equation is $\displaystyle\frac{d}{dt}\nu_{\text{\footnotesize{off}}}(\lambda^{1},t)$ $\displaystyle=\Big{(}-F_{\text{\footnotesize{off}}}^{1,+}(t)-\frac{D}{2}\Big{)}\nu_{\text{\footnotesize{off}}}(\lambda^{1},t)$ (B.95) $\displaystyle+\Big{(}\frac{D}{2}+F_{\text{\footnotesize{off}}}^{2,-}(t)\Big{)}\nu_{\text{\footnotesize{off}}}(\lambda^{2},t).$ Considering the CV associated with the temperature $\lambda^{N}_{\text{on}}$, we have $\displaystyle\frac{d}{dt}\nu_{\text{\footnotesize{on}}}(\lambda^{N},t)$ $\displaystyle=\Big{(}F_{\text{\footnotesize{on}}}^{N,+}(t)-\frac{D}{2}\Big{)}\nu_{\text{\footnotesize{on}}}(\lambda^{N},t)$ (B.96) $\displaystyle+\Big{(}\frac{D}{2}-F_{\text{\footnotesize{on}}}^{N,+}(t)\Big{)}\nu_{\text{\footnotesize{on}}}(\lambda^{N-1},t).$ In the above we make the assumption that $\nu_{\text{\footnotesize{off}}}(\lambda^{1,-}-\Delta\lambda,t)=0$ and $\nu_{\text{\footnotesize{on}}}(\lambda^{N,+}+\Delta\lambda,t)=0$. Now focus on the absorbing boundary (B.88) and conservation of probability (B.90)-(B.91) boundary conditions. These BCs have the following meaning. The condition (B.88) clamps the density at the end of the deadband to zero. BC (B.90) reads: the net-flux across the temperature value $\lambda^{q}_{\text{\footnotesize{on}}}$ is equal to the flux of density going from off to on. In order to enforce both (B.90) and (B.91) we will model the flux of density due to the thermostat control policy as a source/sink. Before doing this, we mention some issues with enforcing the BC (B.88). A TCL’s temperature trajectory will not satisfy the BC (B.88) since to switch its mode the TCL’s temperature sensor will have to register a value outside the deadband. Therfore, we introduce two additional CV’s associated with the temperatures $\lambda^{1}_{\text{\footnotesize{on}}}$ and $\lambda^{N}_{\text{\footnotesize{off}}}$, which are shown in red in Figure 2. We then transfer the BC (B.88) to one on the added CVs, which becomes: $\displaystyle\mu_{\text{\footnotesize{on}}}(\lambda^{1,-},t)=\mu_{\text{\footnotesize{off}}}(\lambda^{N,+},t)$ $\displaystyle=0.$ (B.97) As mentioned, to enforce the conservation of probability BC we use a source/sink type argument, which we also enforce on the added CVs. To see what we mean by source/sink argument, consider the following: some rate of density is transferred out of the CV $\lambda^{N}_{\text{\footnotesize{off}}}$ and into the CV $\lambda^{q}_{\text{\footnotesize{on}}}$ (as depicted in Figure 2) due to thermostatic control. We model the sink as simply $-\nu_{\text{\footnotesize{off}}}(\lambda^{N},t)$. The rate of the sink is then given as $-\gamma\nu_{\text{\footnotesize{off}}}(\lambda^{N},t)$, where $\gamma>0$ is a modeling choice and a constant of appropriate units that describes the discharge rate. We shortly give insight on how to select a value for $\gamma$. Now discretizing the CV corresponding to the nodal value $\lambda^{N}_{\text{\footnotesize{off}}}$ subject to the BC (B.97) and the sink $-\nu_{\text{\footnotesize{off}}}(\lambda^{N},t)$ we obtain, $\displaystyle\frac{d}{dt}\nu_{\text{\footnotesize{off}}}(\lambda^{N},t)$ $\displaystyle=\Big{(}\frac{D}{2}+F_{\text{\footnotesize{off}}}^{N,-}(t)\Big{)}\nu_{\text{\footnotesize{off}}}(\lambda^{N-1},t)$ (B.98) $\displaystyle-\alpha\nu_{\text{\footnotesize{off}}}(\lambda^{N},t),$ where $\alpha\triangleq\big{(}\gamma+D\big{)}$. In obtaining the above, we have made the reasonable assumption that $\nu_{\text{\footnotesize{off}}}(\lambda^{N,+}+\Delta\lambda,t)=0$. The quantity $\alpha\nu_{\text{\footnotesize{off}}}(\lambda^{N},t)$ represents the rate of change of density from the CV $\lambda^{N}_{\text{\footnotesize{off}}}$ to the CV $\lambda^{q}_{\text{\footnotesize{on}}}$, as depicted in Figure 2. Consequently, to conserve probability, we must add this quantity as a source to the ode for the CV $\lambda^{q}_{\text{\footnotesize{on}}}$, i.e., $\displaystyle\frac{d}{dt}\nu_{\text{\footnotesize{on}}}(\lambda^{q},t)$ $\displaystyle=\dots+\alpha\nu_{\text{\footnotesize{off}}}(\lambda^{N},t).$ (B.99) The dots in equation (B.99) represent the portion of the dynamics for the standard internal CV (i.e., the RHS of (B.86)) for the temperature node $\lambda^{q}_{\text{\footnotesize{on}}}$. A similar argument is used for the BC (B.91) with the CV’s $\lambda^{1}_{\text{\footnotesize{on}}}$ and $\lambda^{m}_{\text{\footnotesize{off}}}$, and the corresponding differential equations are, $\displaystyle\frac{d}{dt}\nu_{\text{\footnotesize{on}}}(\lambda^{1},t)$ $\displaystyle=\Big{(}\frac{D}{2}-F_{\text{\footnotesize{on}}}^{1,+}(t)\Big{)}\nu_{\text{\footnotesize{on}}}(\lambda^{2},t)-\alpha\nu_{\text{\footnotesize{on}}}(\lambda^{1},t),$ (B.100) $\displaystyle\frac{d}{dt}\nu_{\text{\footnotesize{off}}}(\lambda^{m},t)$ $\displaystyle=\dots+\alpha\nu_{\text{\footnotesize{on}}}(\lambda^{1},t).$ (B.101) To better understand the role of $\gamma$ consider the following example. Electing $\gamma$ in the above so that $\alpha=(\Delta t)^{-1}$, where $\Delta t$ is a time increment, has the following interpretation: all mass starting in state $\nu_{\text{\footnotesize{off}}}(\lambda^{N},\cdot)$ at time $t$ is transferred out by time $t+\Delta t$ into the state $\nu_{\text{\footnotesize{on}}}(\lambda^{q},\cdot)$. #### B.2.1 Additional conditions Two additional conditions are enforced, namely that once mass is transferred to the nodes $\lambda^{N}_{\text{\footnotesize{off}}}$ or $\lambda^{1}_{\text{\footnotesize{on}}}$ it cannot “travel backwards.” For example, mass is transferred from $\lambda^{N}_{\text{\footnotesize{off}}}$ entirely to the corresponding on temperature bin and no mass is transferred backwards to $\lambda^{N-1}_{\text{\footnotesize{off}}}$. This corresponds to setting: (i) the coefficient on $\nu_{\text{\footnotesize{off}}}(\lambda^{N},t)$ in the ode for $\nu_{\text{\footnotesize{off}}}(\lambda^{N-1},t)$ to zero and (ii) the coefficient on $\nu_{\text{\footnotesize{on}}}(\lambda^{1},t)$ in the ode for $\nu_{\text{\footnotesize{on}}}(\lambda^{2},t)$ to zero. ### B.3 Overall system Now, combining the odes–(B.86) and (B.87) for all of the internal CVs and (B.95), (B.96), (B.98), (B.99), (B.100), (B.101) for the BC CVs–we obtain the linear time varying system, $\displaystyle\frac{d}{dt}\nu(t)=\nu(t)A(t).$ (B.102)
# Impact of surface anisotropy on the spin-wave dynamics in thin ferromagnetic film Krzysztof Szulc<EMAIL_ADDRESS>Institute of Spintronics and Quantum Information, Faculty of Physics, Adam Mickiewicz University, Poznań, Uniwersytetu Poznańskiego 2, 61-614 Poznań, Poland Julia Kharlan Institute of Spintronics and Quantum Information, Faculty of Physics, Adam Mickiewicz University, Poznań, Uniwersytetu Poznańskiego 2, 61-614 Poznań, Poland Institute of Magnetism, National Academy of Sciences of Ukraine, 36b Vernadskogo Boulevard, 03142 Kyiv, Ukraine Pavlo Bondarenko Institute of Magnetism, National Academy of Sciences of Ukraine, 36b Vernadskogo Boulevard, 03142 Kyiv, Ukraine Elena V. Tartakovskaya Institute of Spintronics and Quantum Information, Faculty of Physics, Adam Mickiewicz University, Poznań, Uniwersytetu Poznańskiego 2, 61-614 Poznań, Poland Institute of Magnetism, National Academy of Sciences of Ukraine, 36b Vernadskogo Boulevard, 03142 Kyiv, Ukraine Maciej Krawczyk Institute of Spintronics and Quantum Information, Faculty of Physics, Adam Mickiewicz University, Poznań, Uniwersytetu Poznańskiego 2, 61-614 Poznań, Poland ###### Abstract The spin-wave dynamics in the thin CoFeB film in Damon-Eshbach geometry are studied in three cases of boundary conditions—free boundary conditions, symmetrical surface anisotropy, and one-sided surface anisotropy. The analytical model created by Wolfram and De Wames was extended to include perpendicular surface anisotropy in boundary conditions. Its comparison with numerical simulations demonstrate perfect agreement between the approaches. The analysis of the dispersion relation indicates that the presence of surface anisotropy increases the avoided crossing size between Damon-Eshbach mode and perpendicular standing modes. Additionally, asymmetrical one-sided surface anisotropy induces nonreciprocity in the dispersion relation. In-depth analysis of the avoided crossing size is conducted for systems with different boundary conditions, different thicknesses, surface anisotropy constant values, and external magnetic fields. It shows the significant role of the strength of surface localization of Damon-Eshbach mode and the symmetry of perpendicular standing modes in the avoided crossing broadening. Interestingly, for specific set of parameters the interaction between the particular modes can be suppressed, resulting in a mode crossing. Such a crossing, which occurs only on one side of the dispersion relation in a one- sided surface anisotropy system, can be utilized in nonreciprocal devices. ††preprint: APS/123-QED ## I Introduction In recent years, spin waves (SWs), which are collective, harmonic oscillations of spins that propagate within magnetic materials, have received increased attention due to their potential to transport and process information with the reduction of Joule heating and energy dissipation [1]. One of the interesting properties of propagating SWs in thin magnetic films in Damon-Eshbach (DE) geometry [2] is the hybridization between the fundamental SW mode and perpendicular standing SW (PSSW) modes [3, 4, 5, 6, 7, 8, 9]. This may result in the formation of avoided crossings (ACs), which can be a crucial physical characteristic for the development of magnonic devices such as filters and phase shifters. However, the control of the dynamic magnetic properties is a fundamental problem for the implementation of these devices. It has been demonstrated that surface anisotropy significantly impacts the dispersion relation and the AC size between propagating SW mode and PSSW modes [10]. Another studies have shown that surface anisotropy can be controlled by the voltage applied across the ferromagnetic-metal/insulator heterostructures due to the charge accumulation at the interface [11, 12, 13] or across insulator/ferromagnet/insulator multilayer due to the dielectric polarization influence on the interface [14]. Therefore, it can be concluded that hybridization between fundamental SW mode and higher-order PSSW modes could be controlled by electric field. However, there has been no systematic study on the influence of surface anisotropy on the hybridization between SW modes in the ferromagnetic film. In general, there are two alternative approaches which can be used for the analytical evaluation of dipole-exchange SW spectrum including interaction between fundamental SW mode and PSSW modes. One approach, proposed by Wolfram and De Wames [15, 16], involves solving a sixth-order differential equation derived from Maxwell’s equations along with equations of the magnetization motion. The extension of Damon and Eshbach’s theory for pure dipolar SWs by including exchange interactions provides evidence that, as a result of exchange, the surface and bulk modes mix. This theoretical approach was used for explanation of the first experiments on magnon branch repulsion in thin ferromagnetic films with in-plane magnetization [17, 18] and in thin single- crystal disks of yttrium-iron garnet [19]. Much later, researchers applied the same method to characterize SWs in infinitely long cylindrical wires with magnetization along the wire [20, 21]. However, it turned out that the Wolfram and De Wames approach is not suitable for a broad range of sample geometries and magnetic moment directions. In fact, its effectiveness is limited to cases of unbroken symmetry in infinite films, as well as in infinite wires with a magnetic moment along the wire axis, as previously noted. For more general cases, Kalinikos and Slavin proposed an alternative approach for mixed exchange boundary conditions in thin films and the arbitrary direction of external magnetic field and magnetic moment relative to the film plane [22, 23]. The first step of this method is to solve Maxwell’s equations separately in the magnetostatic approximation [24]. Then, the dynamical scalar potential obtained in the form of the tensorial magnetostatic Green’s functions [25] is inserted into the equations of motion for the magnetic moment (linearized Landau-Lifshitz equations), and the resulting integro-differential equation is solved through perturbation theory. This method has resolved most theoretical issues of spin dynamics in laterally confined magnetic elements under different magnetic field configurations. It has been previously applied to describe SW dynamics in isolated magnetic stripes [26] as well as rectangular [27, 28], cylindrical [29], and triangular [30] magnetic dots. A notable benefit of the Kalinikos and Slavin method is that it utilizes a simple analytical formula to achieve good agreement between theory and experiment for thin, circular nanoelements with perpendicularly-magnetized states, such as rings [31] and dots [32]. In more complex cases with broken cylindrical symmetry, it is necessary to consider a greater number of perturbation theory terms (i.e., the interaction of SW modes) [33, 34]. However, the applicability of this theory to any case of nanostructures and geometry of applied fields is not in question. The method of Wolfram and De Wames turned out to be somewhat forgotten, which forced Harms and Duine [35] to ”rediscover” this ansatz since in some cases it provides a more direct path to the result. A comprehensive review of the two mentioned approaches with an analysis of their applicability for various cases of the direction of the external field and magnetization in a ferromagnetic film is given by Arias [36]. The potential drawbacks of the Kalinikos-Slavin method were identified, including possible inaccuracies of the results obtained in the region of hybridization of SW modes, as well as the complexity of describing the interaction of surface and bulk modes. The theoretical approach proposed in [36] is based on the method developed by Wolfram and De Wames and provides strict solutions to the problem. It is important to note that the hybridization of SWs was only examined in the case of mixed symmetrical boundary conditions. In this paper, we conduct a systematical analysis of the impact of surface anisotropy on the SW hybridization, which was presented in [10]. We are confronted with a choice between the two methods described above for calculating the dynamics of SWs should be chosen. Following the conclusions of Arias [36], the Wolfram and De Wames method not only leads to the goal more efficiently in this case, despite the asymmetry of the boundary conditions, but also provides a rigorous solution. This is in contrast to the Kalinikos- Slavin perturbation theory which requires a significant number of iterations and provides only an approximate solution. Therefore, we compared the dispersion relations of SWs in an DE geometry using symmetrical and asymmetrical boundary conditions via the extended Wolfram and De Wames approach. The results of analytical calculations perfectly matched the numerical simulations on the example of CoFeB thin film. We provide an in- detail analysis of the dispersion relations, SW mode profiles, and the effect of material parameters on the SW coupling in the frame of AC size. ## II Methods ### II.1 Investigated system Figure 1: (a) A general schematic of the system and coordinate system. (b-d) Schematics of the boundary conditions investigated in the manuscript: (b) free boundary conditions, (c) symmetrical surface anisotropy, and (d) one-sided surface anisotropy. The system under investigation is presented in Fig. 1(a). It is a thin CoFeB film of thickness $L$ magnetized in-plane in $y$-direction by the external magnetic field $H_{0}$. We consider the DE geometry, i.e., the SWs propagating along the $x$-direction, perpendicular to the external field $H_{0}$. The $z$-axis corresponds to the direction perpendicular to the film plane, where the surfaces of the film are located at $z=\pm L/2$. The following parameters were used for CoFeB: magnetization saturation $M_{\mathrm{S}}=$1335\text{\,}\mathrm{k}\mathrm{A}\mathrm{/}\mathrm{m}$$, exchange stiffness $A_{\mathrm{ex}}=$15\text{\,}\mathrm{p}\mathrm{J}\mathrm{/}\mathrm{m}$$, and gyromagnetic ratio $\gamma=$30\text{\,}\mathrm{G}\mathrm{Hz}\mathrm{/}\mathrm{T}$$. In this study, we consider three cases of boundary conditions: free boundary conditions (FBC) where the surface anisotropy is absent in the system [Fig. 1(b)]; symmetrical surface anisotropy (SSA), i.e., the surface anisotropy of equal strength is present on both boundaries of the film [Fig. 1(c)]; one- sided surface anisotropy (OSA) where the bottom surface has non-zero surface anisotropy while the top surface is described with FBC [Fig. 1(d)]. ### II.2 Analytical model We use the approach proposed by Wolfram and De Wames [15, 16] to calculate the dispersion relation in DE geometry in dipole-exchange regime and extend it to include the perpendicular surface anisotropy introduced by Rado and Weertman [37]. The magnetic free energy of the system can be presented as $F=\int\left(-\mu_{0}\mathbf{H}_{0}\cdot\mathbf{M}+\frac{A_{\mathrm{ex}}}{M_{\mathrm{S}}^{2}}\left(\nabla\mathbf{M}\right)^{2}-\frac{1}{2}\mu_{0}\mathbf{H}_{\mathrm{d}}\cdot\mathbf{M}\right)\mathrm{d}V,$ (1) where there are three terms in the integral—the first term represents the Zeeman energy, the second term represents the exchange energy, and the third term represents the magnetostatic energy, $\mathbf{M}$ is the magnetization vector, $\mu_{0}$ is the vacuum permeability, $\mathbf{H}_{\mathrm{d}}$ is the demagnetizing field. The dynamics of the magnetic system are described with Landau-Lifshitz equation $\frac{\partial\mathbf{M}}{\partial t}=-|\gamma|\mu_{0}\mathbf{M}\times\mathbf{H}_{\mathrm{eff}},$ (2) where $\mathbf{H}_{\mathrm{eff}}=-\frac{1}{\mu_{0}}\frac{\delta F}{\delta\mathbf{M}}$ is the effective magnetic field. The demagnetizing field $\mathbf{H}_{\mathrm{d}}$ is derived from the Maxwell equations in magnetostatic approximation: $\nabla\times\mathbf{H}_{\mathrm{d}}=0,\,\,\,\,\,\nabla\cdot\mathbf{B}=0,$ (3) where $\mathbf{B}=\mu_{0}(\mathbf{H}_{\mathrm{d}}+\mathbf{M})$ is the magnetic induction. Equation (3) enables the introduction of magnetic scalar potential $\varphi$, which satisfies the formula $\mathbf{H}_{\mathrm{d}}=-\nabla\varphi$. As a result, the magnetostatic Maxwell equations are replaced with a single equation for the magnetic scalar potential $\Delta\varphi=\nabla\cdot\mathbf{M}.$ (4) Thanks to the uniform magnetization, the Landau-Lifshitz equation can be easily linearized. We assume that the static $y$-component of the magnetization remains constant and is equal to the saturation magnetization $M_{\mathrm{S}}$, while the dynamic component $\mathbf{m}=(m_{x},m_{z})$, which is much smaller than the static component $M_{y}$ ($|\mathbf{m}|\ll M_{\mathrm{S}}$), precesses in the $xz$-plane. Therefore, $\mathbf{M}(x,y,z,t)=M_{\mathrm{S}}\hat{y}+\mathbf{m}(x,z)e^{i\omega t}$, where $\omega=2\pi f$ is the angular frequency and $f$ is the frequency. After linearization, the SW dynamics are described with a set of three coupled equations: $i\omega m_{x}=\gamma\mu_{0}\left(H_{0}-\frac{2A_{\mathrm{ex}}}{\mu_{0}M_{\mathrm{S}}}\Delta\right)m_{z}+M_{\mathrm{S}}\partial_{z}\varphi,$ (5) $-i\omega m_{z}=\gamma\mu_{0}\left(H_{0}-\frac{2A_{\mathrm{ex}}}{\mu_{0}M_{\mathrm{S}}}\Delta\right)m_{x}+M_{\mathrm{S}}\partial_{x}\varphi,$ (6) $\Delta\varphi-\partial_{x}m_{x}-\partial_{z}m_{z}=0.$ (7) The solutions to Eqs. (5)-(7) take the form of plane waves. Two wave vectors can be defined due to the system’s symmetry: in-plane wave vector $k$ (in the $x$-direction) and out-of-plane wave vector $q$ (in the $z$-direction), as shown in Fig. 1(a). As a result, we have $(m_{x},m_{z},\varphi)\propto(m_{x0},m_{z0},\varphi_{0})e^{ikx}e^{iqz}$. The system in the $x$-direction is infinite, therefore the wave vector $k$ can only have real values for the solution to be physical. On the other hand, the wave vector $q$ may take on complex values. For simplicity, we introduce the following dimensionless parameters: $\Omega=\frac{\omega}{\gamma\mu_{0}M_{\mathrm{S}}}$, $\theta=\Omega_{H}+\lambda^{2}(k^{2}+q^{2})$, $\Omega_{H}=\frac{H_{0}}{M_{\mathrm{S}}}$, and $\lambda^{2}=\frac{2A_{\mathrm{ex}}}{\mu_{0}M_{\mathrm{S}}^{2}}$. After substituting the plane-wave solution into Eqs. (5)-(7) and expressing them in the matrix form, we obtain $\begin{pmatrix}i\Omega&\theta&iq\\\ \theta&-i\Omega&ik\\\ ik&iq&k^{2}+q^{2}\end{pmatrix}\begin{pmatrix}m_{x0}\\\ m_{z0}\\\ \varphi_{0}\end{pmatrix}=0.$ (8) The condition that the determinant of the 3x3 matrix in Eq. (8) is equal to zero leads to the following formula: $(k^{2}+q^{2})(\Omega^{2}-\theta^{2}-\theta)=0.$ (9) As $\theta=\theta(q^{2})$, Eq. (9) is a third-degree function with respect to $q^{2}$. Two roots, $q=\pm ik$, are obtained by setting the first bracket to zero whereas four roots, $q=\pm q_{1}$ and $q=\pm iq_{2}$ where $q_{1},q_{2}\in\mathbb{R}$, are obtained by setting the second bracket to zero. From the zeroing of the second bracket in Eq. (9), we can also derive the formula for the dimensionless frequency $\Omega=\sqrt{\theta(\theta+1)}.$ (10) Let $\theta(q=q_{1})=\theta_{1}$ and $\theta(q=q_{2})=\theta_{2}$. Since $q_{1}$ and $q_{2}$ correspond to the same frequency, $\Omega=\sqrt{\theta_{1}(\theta_{1}+1)}=\sqrt{\theta_{2}(\theta_{2}+1)}$, and therefore, $\theta_{2}=-(\theta_{1}+1)$. From this formula we can obtain the connection between wave vectors $k$, $q_{1}$, and $q_{2}$, which is the following: $q_{2}=\pm\sqrt{2k^{2}+q_{1}^{2}+\frac{2\Omega_{H}+1}{\lambda^{2}}}.$ (11) We can interpret the solutions obtained for the out-of-plane wave vector $q$ as follows. Since our solution is a plane wave, wave vector $q_{1}$ will give a volume contribution of the sinusoidal character to the mode profile while wave vectors $k$ and $q_{2}$ denote exponentially-decaying modes localized on the surfaces. Since the wave vector $k$ represents also the propagating in- plane wave vector, this solution has a character of a DE mode. Next, knowing that $\Omega_{H}\geq 0$, we can derive from Eq. (11) that $|q_{2}|\geq 1/\lambda$, indicating that $q_{2}$ has a character of a surface exchange mode. The solution of Eq. (8) can be represented by a vector $\begin{pmatrix}m_{x0}\\\ m_{z0}\\\ \varphi_{0}\end{pmatrix}=\begin{pmatrix}ik\theta-q\Omega\\\ iq\theta+k\Omega\\\ \Omega^{2}-\theta^{2}\end{pmatrix}C,$ (12) where $C$ is an arbitrary constant. The general solution for the full vector $(m_{x},m_{z},\varphi)$ is a superposition of six terms, one for each solution of the wave vector $q$ $\begin{pmatrix}m_{x}\\\ m_{z}\\\ \varphi\end{pmatrix}=\left[\begin{pmatrix}X_{1}\\\ Z_{1}\\\ F_{1}\end{pmatrix}C_{1}e^{iq_{1}z}+\begin{pmatrix}X_{2}\\\ Z_{2}\\\ F_{2}\end{pmatrix}C_{2}e^{-iq_{1}z}+\begin{pmatrix}X_{3}\\\ Z_{3}\\\ F_{3}\end{pmatrix}C_{3}e^{kz}+\begin{pmatrix}X_{4}\\\ Z_{4}\\\ F_{4}\end{pmatrix}C_{4}e^{-kz}+\begin{pmatrix}X_{5}\\\ Z_{5}\\\ F_{5}\end{pmatrix}C_{5}e^{q_{2}z}+\begin{pmatrix}X_{6}\\\ Z_{6}\\\ F_{6}\end{pmatrix}C_{6}e^{-q_{2}z}\right]e^{ikx}$ (13) where $X_{1}=ik\theta_{1}-q_{1}\Omega$, $X_{2}=ik\theta_{1}+q_{1}\Omega$, $X_{3}=X_{4}=ik$, $X_{5}=ik\theta_{2}-iq_{2}\Omega$, $X_{6}=ik\theta_{2}+iq_{2}\Omega$, $Z_{1}=k\Omega+iq_{1}\theta_{1}$, $Z_{2}=k\Omega-iq_{1}\theta_{1}$, $Z_{3}=-k$, $Z_{4}=k$, $Z_{5}=k\Omega- q_{2}\theta_{2}$, $Z_{6}=k\Omega+q_{2}\theta_{2}$, $F_{1}=F_{2}=\Omega^{2}-\theta_{1}^{2}$, $F_{3}=-(\Omega+\Omega_{H})$, $F_{4}=\Omega-\Omega_{H}$, $F_{5}=F_{6}=\Omega^{2}-\theta_{2}^{2}$, as it follows from Eq. (12). As the system under consideration is an infinite film, boundary conditions must be applied on top and bottom surfaces. Our goal was to extend the model derived by Wolfram and De Wames to include the presence of the perpendicular surface anisotropy. It requires the extension of exchange boundary condition by adding the term depending on the surface anisotropy [37] $\left\\{\begin{aligned} \partial_{z}m_{x}&=0|_{z=\pm L/2}\\\ \partial_{z}m_{z}\mp\sigma_{\mathrm{t(b)}}m_{z}&=0|_{z=\pm L/2}\end{aligned}\right.$ (14) where $\sigma_{\mathrm{t(b)}}=K_{\mathrm{s}}^{\mathrm{t(b)}}/A_{\mathrm{ex}}$ and $K_{\mathrm{s}}^{\mathrm{t(b)}}$ is surface anisotropy constant for the top (bottom) surface. Since the equation for the magnetic scalar potential [Eq. (7)] outside of the film gives $\Delta\varphi_{\mathrm{out}}=0$ and, subsequently, $-\varphi_{0}(k^{2}+q^{2})e^{ikx}e^{iqz}=0$, the asymptotic solutions outside the film for the magnetic scalar potential are given by expression $\varphi_{\mathrm{out}}=\begin{cases}C_{7}e^{ikx}e^{-|k|z}&\text{for }z\geq L/2,\\\ C_{8}e^{ikx}e^{|k|z}&\text{for }z\leq-L/2.\end{cases}$ (15) As the tangential components of the demagnetizing field $\mathbf{H}_{\mathrm{d}}$ are continuous across the surfaces of the film, the magnetic scalar potential must also be continuous. Additionally, the normal component of $\mathbf{B}$ must also be continuous. Therefore, this results in the effective magnetostatic boundary conditions: $\varphi=\varphi_{\mathrm{out}},$ (16) $B_{z}=B_{z}^{\mathrm{out}},$ (17) where $\varphi$ and $B_{z}$ are magnetic scalar potential and magnetic induction in the magnetic material, and $\varphi_{\mathrm{out}}$ and $B_{z}^{\mathrm{out}}$ – out of the magnetic material, respectively. Then, Eq. (17) can be rewritten in terms of scalar potential as $\partial_{z}\varphi-m_{z}=\partial_{z}\varphi_{\mathrm{out}}.$ (18) The complete set of boundary conditions in Eqs. (14), (16), and (18) evaluated for the SW modes in Eq. (13) leads to the following degeneracy matrix $\bm{A}$: $\begin{split}\bm{A}=\left(\begin{matrix}iq_{1}X_{1}e^{iq_{1}\frac{L}{2}}&-iq_{1}X_{2}e^{-iq_{1}\frac{L}{2}}&kX_{3}e^{k\frac{L}{2}}\\\ iq_{1}X_{1}e^{-iq_{1}\frac{L}{2}}&-iq_{1}X_{2}e^{iq_{1}\frac{L}{2}}&kX_{3}e^{-k\frac{L}{2}}\\\ (iq_{1}-\sigma_{\mathrm{t}})Z_{1}e^{iq_{1}\frac{L}{2}}&(-iq_{1}-\sigma_{\mathrm{t}})Z_{2}e^{-iq_{1}\frac{L}{2}}&(k-\sigma_{\mathrm{t}})Z_{3}e^{k\frac{L}{2}}\\\ (iq_{1}+\sigma_{\mathrm{b}})Z_{1}e^{-iq_{1}\frac{L}{2}}&(-iq_{1}+\sigma_{\mathrm{b}})Z_{2}e^{iq_{1}\frac{L}{2}}&(k+\sigma_{\mathrm{b}})Z_{3}e^{-k\frac{L}{2}}\\\ [(iq_{1}+|k|)F_{1}-Z_{1}]e^{iq_{1}\frac{L}{2}}&[(-iq_{1}+|k|)F_{2}-Z_{2}]e^{-iq_{1}\frac{L}{2}}&[(k+|k|)F_{3}-Z_{3}]e^{k\frac{L}{2}}\\\ [(iq_{1}-|k|)F_{1}-Z_{1}]e^{-iq_{1}\frac{L}{2}}&[(-iq_{1}-|k|)F_{2}-Z_{2}]e^{iq_{1}\frac{L}{2}}&[(k-|k|)F_{3}-Z_{3}]e^{-k\frac{L}{2}}\end{matrix}\right|\\\ \left|\begin{matrix}-kX_{4}e^{-k\frac{L}{2}}&q_{2}X_{5}e^{q_{2}\frac{L}{2}}&-q_{2}X_{6}e^{-q_{2}\frac{L}{2}}\\\ -kX_{4}e^{k\frac{L}{2}}&q_{2}X_{5}e^{-q_{2}\frac{L}{2}}&-q_{2}X_{6}e^{q_{2}\frac{L}{2}}\\\ (-k-\sigma_{\mathrm{t}})Z_{4}e^{-k\frac{L}{2}}&(q_{2}-\sigma_{\mathrm{t}})Z_{5}e^{q_{2}\frac{L}{2}}&(-q_{2}-\sigma_{\mathrm{t}})Z_{6}e^{-q_{2}\frac{L}{2}}\\\ (-k+\sigma_{\mathrm{b}})Z_{4}e^{k\frac{L}{2}}&(q_{2}+\sigma_{\mathrm{b}})Z_{5}e^{-q_{2}\frac{L}{2}}&(-q_{2}+\sigma_{\mathrm{b}})Z_{6}e^{q_{2}\frac{L}{2}}\\\ [(-k+|k|)F_{4}-Z_{4}]e^{-k\frac{L}{2}}&[(q_{2}+|k|)F_{5}-Z_{5}]e^{q_{2}\frac{L}{2}}&[(-q_{2}+|k|)F_{6}-Z_{6}]e^{-q_{2}\frac{L}{2}}\\\ [(-k-|k|)F_{4}-Z_{4}]e^{k\frac{L}{2}}&[(q_{2}-|k|)F_{5}-Z_{5}]e^{-q_{2}\frac{L}{2}}&[(-q_{2}-|k|)F_{6}-Z_{6}]e^{q_{2}\frac{L}{2}}\end{matrix}\right).\end{split}$ (19) The condition $\det{\bm{A}}=0$ allows obtaining the solutions of wave vector $q$ and, subsequently, the resonance frequencies as a function of wave vector $k$. The eigenvectors of matrix $\bm{A}$ provide the coefficients $C_{i}$ in Eq. (13). Compared to the approach suggested by Kalinikos et al. [23], the solution mentioned above is precise within the examined geometry. Calculating multiple integrals for components of a demagnetizing tensor and expanding dynamical magnetization components into a series is not required to obtain coupled modes, which simplifies analytical calculations and significantly reduces computation time. ### II.3 Numerical simulations The Landau-Lifshitz equation in the linear approximation [Eqs. (5),(6)] and the magnetostatic Maxwell equation-based formula for the magnetic scalar potential [Eq. (7)] along with the boundary conditions for perpendicular surface anisotropy [Eq. (14)] and magnetostatic potential [Eq. (15)] were solved numerically using finite-element method simulations in COMSOL Multiphysics [10]. The problem was solved in 1D geometry with reduced $x$\- and $y$-directions. Eqs. (5)-(7) were modified accordingly to introduce the terms coming from the implementation of plane-wave solution representing the propagation of SWs in $x$-direction $(m_{x},m_{z},\varphi)=(m_{x0},m_{z0},\varphi_{0})e^{ikx}$. The dispersion relations were calculated using eigenfrequency study. ## III Results and discussion ### III.1 Dispersion relation analysis Figure 2: (a-d) Dispersion relations of six lowest modes of a 100 nm-thick CoFeB film with (a) FBC, (b) SSA with $K_{\mathrm{s}}^{\mathrm{t}}=K_{\mathrm{s}}^{\mathrm{b}}=$-700\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$, and (c,d) OSA with $K_{\mathrm{s}}^{\mathrm{t}}=0$ and $K_{\mathrm{s}}^{\mathrm{b}}=$-1500\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$ for (c) negative and (d) positive wave vectors in the external magnetic field $\mu_{0}H_{0}=$50\text{\,}\mathrm{mT}$$. The plots present the comparison between the analytical model (orange lines) and numerical simulations (blue lines). Avoided crossings (ACs) are marked with labels. (e) The frequency difference between neighboring modes in FBC system. In plots (a-e) wave vector $k$ on the $x$-axis is presented in the logarithmic scale. (f-i) Close-up on the ACs: (f) AC1, (g) AC2, (h) AC3, and (i) AC4. The plot axis are showing the wave vector and frequency values relative to the AC position calculated from Eqs. (20) and (24), respectively. Plots present numerical simulations results only which are in agreement with analytical results. First, we study the effect of the surface anisotropy on the dispersion relation. We chose the thickness of the CoFeB film $L=$100\text{\,}\mathrm{nm}$$ and external magnetic field $\mu_{0}H_{0}=$50\text{\,}\mathrm{mT}$$. We show the dispersion relation of the six lowest modes for three cases—free boundary conditions (FBC), i.e., $K_{\mathrm{s}}^{\mathrm{t}}=K_{\mathrm{s}}^{\mathrm{b}}=0$ [Fig. 2(a)]; symmetrical surface anisotropy (SSA) with $K_{\mathrm{s}}^{\mathrm{t}}=K_{\mathrm{s}}^{\mathrm{b}}=$-700\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$ [Fig. 2(b)]; one-sided surface anisotropy (OSA) with $K_{\mathrm{s}}^{\mathrm{t}}=0$ and $K_{\mathrm{s}}^{\mathrm{b}}=$-1500\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$ separately for negative [Fig. 2(c)] and positive [Fig. 2(d)] wave vector $k$. Values of surface anisotropy are comparable to the values presented in literature [38]. The dispersion relation calculated with the analytical model is shown as a dashed orange line while the numerical simulation results are shown with dashed blue line. Figs. 2(a-d) demonstrate the perfect agreement between these two methods, yielding identical outcomes. The nature of dispersions is characteristic of the system in DE geometry. Each plot consists of one branch with a significant slope in the center of the investigated range of wave vector $k$, displaying a DE surface mode character, and the remaining five are flat branches representing PSSW modes. All the modes start to increase significantly in frequency at about $10^{7}$ rad/m as a result of the increasing contribution of the exchange interaction to the SW energy. Positions of PSSW modes at $k\approx 0$ are determined by the wave vector $q_{1}\approx n\pi/L$ ($n=1,2,3...$ is the PSSW mode number). In the presence of negative surface anisotropy, the value of $q_{1}>n\pi/L$ for corresponding PSSW modes (the reverse happens for positive surface anisotropy). The increase of the frequency of the DE mode correlates with the increase of its wave vector $q_{1}$ with the increase of $k$. However, $q_{1}$ begins to decrease at some point, leading to $q_{1}\approx n\pi/L$ for very large wave vectors $k$. Similarly as for the case of $k\approx 0$, for very large $k$ in the presence of negative surface anisotropy, $q_{1}>n\pi/L$ (the reverse happens for positive surface anisotropy). Detailed explanation of the correlation between wave vectors $k$ and $q_{1}$ is provided in Appendix A. The DE mode increases in frequency and intersects with the three lowest PSSW modes, leading to the emergence of ACs. These ACs are labeled in Figs. 2(a-d) with the abbreviation AC and a number indicating their sequence, beginning with the lowest. The discussion of ACs requires a precise definition of where AC occurs. Neglecting the atomic distance limit, the theory provides infinite number of SW modes. Though it is hypothetically possible for AC to be present between all modes, it is apparent that the number of ACs is not infinite for the finite-thickness film. To denote the presence of AC, we establish two distinct criteria. The first is the local minimum criterion. If the function that represents the frequency difference between the neighboring modes $\Delta f_{mn}=f_{m}(k)-f_{n}(k)$ (20) (where $m,n$ is a mode number) has a local minimum $\Delta f_{\mathrm{AC}n}$, this minimum represents an AC (or simply crossing if $\Delta f_{\mathrm{AC}}=0$). In this way, we can define an AC for any boundary conditions and it allows multiple ACs if multiple local minima exist. The second is a frequency limit criterion. It could be clearly defined only for FBC. It says that an AC is present between the DE mode and $n$-th PSSW mode if $f_{k\to\infty}^{\mathrm{DE}}>f_{k=0}^{n}$ in case where $f_{k\to\infty}^{\mathrm{DE}}$ is calculated for $A_{\mathrm{ex}}=0$ [2], i.e. $f_{k\to\infty}^{\mathrm{DE}}=\frac{\mu_{0}\gamma}{2\pi}\left(H_{0}+\frac{M_{\mathrm{S}}}{2}\right)$ (21) and [39] $\displaystyle f_{k=0}^{n}=\frac{\mu_{0}\gamma}{2\pi}\left(\left(H_{0}+\frac{2A_{\mathrm{ex}}}{\mu_{0}M_{\mathrm{S}}}\left(\frac{n\pi}{L}\right)^{2}\right)\right.\times$ $\displaystyle\left.\times\left(H_{0}+M_{\mathrm{S}}+\frac{2A_{\mathrm{ex}}}{\mu_{0}M_{\mathrm{S}}}\left(\frac{n\pi}{L}\right)^{2}\right)\right)^{1/2}$ (22) This criterion is valid under the assumption that the contribution of the exchange interaction to the $k$ dependence of the frequency of DE mode and PSSW modes is identical. The AC position is determined by the minimum of Eq. (20). It means that the choice of criterion does not influence the value of the AC size. In this paper, we present the results based on the local minimum criterion because of its broader definition. However, we will also mention the frequency limit criterion and its impact on the results. To address AC occurrence accurately, the frequency difference between neighboring modes is presented as a function of wave vector $k$ in Fig. 2(e) for the case of FBC, for which the dispersion relation is shown in Fig. 2(a). In the range of small and large wave vectors, the distance between the modes is almost constant. The discrepancy between these ranges is due to the fact that in the limit of small wave vectors, the dispersion relation of the modes can be described by Eq. (III.1) [39], while in the large wave vector limit with the function $f_{n}=\frac{\mu_{0}\gamma}{2\pi}\left(H_{0}+M_{\mathrm{S}}+\frac{2A_{\mathrm{ex}}}{\mu_{0}M_{\mathrm{S}}}k^{2}+\frac{2A_{\mathrm{ex}}}{\mu_{0}M_{\mathrm{S}}}\left(\frac{n\pi}{L}\right)^{2}\right).$ (23) Table 1: AC size of AC1-AC5 for FBC, SSA, and OSA systems, which dispersion relations are shown in Figs. 2(a-d). System | AC1 (MHz) | AC2 (MHz) | AC3 (MHz) | AC4 (MHz) | AC5 (MHz) ---|---|---|---|---|--- FBC | $11.14$ | $6.04$ | $158.8$ | $1137.2$ | $6022.4$ SSA | $21.03$ | $231.5$ | $280.6$ | $1327.7$ | $5781.5$ OSA ($k-$) | $162.8$ | $254.8$ | $565.6$ | $1389.2$ | $5474.4$ OSA ($k+$) | $122.8$ | $175.7$ | $24.67$ | $1396.0$ | $6119.0$ In the mid-range, each curve shown in Fig. 2(e) has a local minimum corresponding to the AC, which is labeled and marked with an arrow. The first three ACs are relatively small, not exceeding a size of 200 MHz. The AC4, represented by a deep minimum, has a size of 1.14 GHz. On the other hand, AC5 has a very shallow minimum with a size of 6.02 GHz. Interestingly, it is not the global minimum, as according to Eq. (23) the distance between the modes can reach 5.99 GHz, which value is in agreement with the analytical model. However, according to the local minimum criterion, it is considered to be an AC. In case of the frequency limit criterion, only the first three minima can be identified as ACs. The AC4 does not meet this criterion as $f_{k=0}^{n=4}=$27.55\text{\,}\mathrm{GHz}$$ exceeds $f_{k\to\infty}^{\mathrm{DE}}=$26.66\text{\,}\mathrm{GHz}$$ slightly. After presenting the similarities between the systems, it is time to highlight the differences. Firstly, the symmetry of the system, specifically the boundary conditions, leads to the symmetry of the dispersion relation with respect to wave vector. Therefore, the FBC and SSA systems have symmetrical dispersions since $K_{\mathrm{s}}^{\mathrm{t}}=K_{\mathrm{s}}^{\mathrm{b}}$. In contrast, the OSA system has different surface anisotropy constants on the top and bottom surfaces, resulting in a frequency difference between negative and positive wave vectors. Additionally, the presence of the negative surface anisotropy causes a slight increase in the frequency of all modes. Comparing the results in Figs. 2(a) and (b), for $K_{\mathrm{s}}=$-700\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$ the increase does not pass 1 GHz. Conversely, for a positive surface anisotropy, a decrease in frequency would be noted. The most significant difference between the systems lies in the size of the ACs. Close-up plots are shown in Fig. 2(f-i) for AC1-AC4, respectively. They show the dispersion relation for the values of wave vector and frequency relative to the AC location ($k_{\mathrm{AC}}$,$f_{\mathrm{AC}}$), which is defined separately for each AC in the following way –- $k_{\mathrm{AC}_{n}}$ represents the wave vector of the local minimum of Eq. (20), while $f_{\mathrm{AC}_{n}}$ $f_{\mathrm{AC}_{n}}=\frac{f_{n}(k_{\mathrm{AC}_{n}})+f_{n+1}(k_{\mathrm{AC}_{n}})}{2}.$ (24) The values of the AC size for each AC type can be found in Table 1. AC1 [Fig. 2(f)] exhibits a negligible size for the FBC and SSA systems, but a more significant size of 162.8 MHz for negative and 122.8 MHz for positive wave vectors in the OSA system. As the dispersion relations for FBC and SSA are symmetrical, the AC sizes are always equal for both negative and positive wave vectors. The size of AC2 [Fig. 2(g)] remains small only in the FBC system, whereas it opens up in the SSA and OSA systems reaching the sizes larger than AC1. In the OSA system, there is a slight asymmetry between negative and positive wave vector range. In the case of AC3 [Fig. 2(h)], it opens up for all the considered cases. The most interesting case is present for OSA system. In the range of negative wave vectors, this AC is large, whereas in the range of positive wave vectors, it is very small, measuring only 24.67 MHz. AC4 is much larger than lower order ACs, having a size above 1 GHz, however, the size is very similar in all of the systems [Fig. 2(i)]. AC5 presents a similar case, with its size being even larger, measuring above 5 GHz. ### III.2 Mode profiles Figure 3: Distribution across the film thickness of a dynamic magnetization components $m_{x}$ (solid lines) and $m_{z}$ (dashed lines) for (a) a first mode for wave vector $k=0$, (b) a first mode for wave vector $k=$-5\text{\times}{10}^{8}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$, (c) a third mode for wave vector $k=$-2.5\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$, and (d) a third mode for wave vector $k=$2.5\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$. Mode profiles are presented for system with FBC (blue lines), SSA (orange lines), and OSA (green lines). Plots present numerical simulations results only which are in agreement with analytical results. Figure 4: The AC size $\Delta f_{\mathrm{AC}}$ as a function of film thickness $L$ for the system with (a) FBC, (b) SSA, and OSA for (c) negative and (d) positive wave vector $k$. Odd- numbered ACs are shown with solid lines, even-numbered ACs with dashed lines. The $y$-axis is in the logarithmic scale. Plots present results of numerical simulations. The surface anisotropy has a significant impact on the dynamic magnetization distribution of SW modes, with mode profiles shown in Fig. 3. Firstly, we present the profile of the lowest frequency mode at $k=0$ in Fig. 3(a). Due to the low external field, the spin precession is strongly elliptical with the domination of in-plane $m_{x}$ component. In the case of FBC (blue lines), the mode is uniform throughout the thickness. The negative surface anisotropy leads to the reduction of the SW amplitude close to the film boundary. The mode is symmetrical for SSA, while for OSA it becomes asymmetrical. Interestingly, although the surface anisotropy affects directly only the out- of-plane $m_{z}$ component, the in-plane $m_{x}$ component is also impacted. However, in the dipole-dominated low-wave vector regime, the effect of surface anisotropy is generally small. The impact on the PSSW modes (not shown here) is even smaller. However, the anisotropy has a substantial effect on the mode profiles in the exchange-dominated large-wave vector region, as evidenced in Fig. 3(b) for the lowest frequency mode at $k=$-5\text{\times}{10}^{8}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$. In both the SSA and OSA cases, the mode amplitude is significantly lower near the boundary with surface anisotropy in comparison to the FBC case. Interestingly, in this case, the $m_{z}$ component exceeds the $m_{x}$ component, and the precession is close to circular. In Figs. 3(c,d), profiles of the third lowest mode are shown at $k$ between AC2 and AC3 for the negative [$k=$-2.5\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$, Fig. 3(c)] and positive [$k=$2.5\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$, Fig. 3(d)] wave vectors. The mode has a character of a DE mode, although the first and second term of Eq. (13) connected with wave vector $q_{1}$ also have a significant impact on the mode shape, which results in the sinusoidal character of these profiles. Their contribution is enhanced when the surface anisotropy is present. The $m_{x}$ component is larger than $m_{z}$ component, but the precession is less elliptical than when $k=0$. For both FBC and SSA cases, where the boundary conditions are identical on both surfaces, the mode profiles for opposite wave vectors are their mirror images. However, this is not true for OSA as the mode profiles differ between negative and positive wave vectors. For negative wave vectors [Fig. 3(c)], the contribution from first and second terms in Eq. (13) are significantly stronger for both $m_{x}$ and $m_{z}$ components. ### III.3 Analysis of thickness dependence In the next step, we present a detailed analysis of the impact of the surface anisotropy on AC formation. Firstly, we study the effect of the film thickness $L$ on the AC size $\Delta f_{\mathrm{AC}}$ in four cases—FBC [Fig. 4(a)], SSA [Fig. 4(b)], and OSA for both negative [Fig. 4(c)] and positive [Fig. 4(d)] wave vector $k$. In general, the increase of film thickness results in an increase in the number of ACs. This phenomenon is well-explained by the frequency limit criterion. The thickness has no impact on the maximum DE frequency $f_{k\to\infty}^{\mathrm{DE}}$ [Eq. (21)]. In contrast, the formula for the PSSW frequency $f_{k=0}^{n}$ [Eq. (III.1)] includes thickness in the denominator; thus, an increase of thickness results in a decrease of frequency. This allows for a higher number of PSSW modes to satisfy the frequency limit criterion, resulting in more ACs. Another relevant effect is that the AC size decreases with an increase of thickness. Figure 4 shows that the rate of the AC size decrease depends on the boundary conditions and the parity of the AC number. In the FBC system [Fig. 4(a)], the AC size decreases rapidly, but much faster for even-numbered ACs than for odd- numbered ACs. In the case of SSA [Fig. 4(b)], the rate of decrease for odd- numbered ACs is slightly smaller, but for even-numbered ACs the change is significant; in this case, the decrease is much smaller compared to the odd- numbered ACs. In the OSA system [Figs. 4(c,d)], the rate of decrease is similar across all ACs and comparable to the even-numbered ACs in the SSA system. This effect, which depends on parity, originates from the symmetry of modes and boundary conditions. Due to the dominant contribution of the $k$-dependent term in the magnetization profile shape, the DE mode has a symmetry closer to the odd-numbered PSSW modes, which are connected with the odd-numbered ACs. In the case of FBC, there is no additional source of symmetry breaking and, therefore, odd-numbered ACs are larger. On the other hand, SSA causes a symmetric disturbance of all modes, primarily affecting the dynamic magnetization amplitude in close proximity to the surface. Odd- numbered PSSW modes have opposite amplitude on the opposite boundaries, therefore, the effect of the anisotropy on the mode symmetry cancels out. On the other hand, both DE mode and even-numbered PSSW modes exhibit the same amplitude on the opposite surfaces, so the anisotropy breaks the symmetry of these modes and, as an effect, these modes induce larger ACs. In the case of OSA, the asymmetry of the anisotropy generates the asymmetry in the mode profiles, leading to large ACs in all cases. An explanation based on a simplified model of mode profiles is provided in Appendix B. The final effect is present only in the OSA system in the positive wave vector $k$ range [Fig. 4(d)]. It is the presence of a local minimum of AC size with a change of thickness. Interestingly, this effect only occurs for odd-numbered ACs. Upon analyzing this effect, one may question whether this local minimum reaches zero, or in other words, whether exists such a critical film thickness for which AC does not occur, i.e., the mode crossing is present. Obviously, the numerical study of the AC size can not provide a definite answer while we were not able to derive it from the analytical model. Nevertheless, the mode profiles analysis can resolve this issue. Figure 5: (a) The AC1 size $\Delta f_{\mathrm{AC1}}$ as a function of film thickness $L$ for the system with OSA for positive wave vector $k$—the close- up of Fig. 4(d) to the AC1 local minimum in small thickness range. Inset plots show the magnetization profiles of first (orange line) and second (green line) mode at $k_{\mathrm{AC1}}$ for the thickness of 39 nm (on the left) and 46 nm (on the right). (b) Dispersion relation of three lowest modes for the system with OSA for positive wave vector $k$ for the thickness of 42.4 nm. Inset in the bottom-right corner: the close-up to the AC1 with marking of three wave vectors – $k_{\mathrm{L}}=$4.853\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$, $k_{\mathrm{AC}}=$5.033\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$, and $k_{\mathrm{R}}=$5.2\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$. Inset at top: magnetization profiles of first (orange line) and second (green line) mode at $k_{\mathrm{L}}$ (left), $k_{\mathrm{AC}}$ (center), and $k_{\mathrm{R}}$ (right). Plots present results of numerical simulations. The close-up to the local minimum of the AC1 [Fig. 4(d), solid blue line] is shown in Fig. 5(a). In this case, the step in simulation was 0.2 nm. The minimum value of $\Delta f_{\mathrm{AC1}}=1.33$ MHz was obtained for a thickness of 42.4 nm. We can take a look on the magnetization profiles for the first and second modes at wave vector $k_{\mathrm{AC}}$ [the inset plots in Fig. 5(a)] for the thickness smaller (39 nm, left plot) and larger (46 nm, right plot) than the thickness of the AC1 size minimum. In both cases, the mode profiles are very similar, indicating the superposition of the DE and first PSSW mode. However, the most important fact is to notice that the modes are interchanged. For $L=$39\text{\,}\mathrm{nm}$$, the lower frequency mode (orange line) has higher amplitude at the bottom of the film, while for $L=$46\text{\,}\mathrm{nm}$$, higher amplitude is at the top of the film. The higher frequency mode (green line) demonstrates the opposite trend. The detailed analysis of the profiles indicates that this behavior is connected with each local minimum of $\Delta f_{\mathrm{AC}}$, i.e. the mode profiles at $k_{\mathrm{AC}}$ interchange. According to our analysis, it indicates the existence of a critical thickness value $L_{\mathrm{C}}$ where a crossing occurs instead of AC, indicating the absence of a gap between the first and second mode. This observation suggests the possible occurrence of an accidental degeneracy in the system [40, 41], meaning that there are two solutions with the same values of wave vector and frequency. Figure 6: (a) Dispersion relation of the lowest six modes of the system with SSA for $K_{\mathrm{s}}=$2500\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$. (b) Frequency difference between the neighboring modes of the system with SSA presented in (a). The $x$-axis is in the logarithmic scale. (c-e) The AC size $\Delta f_{\mathrm{AC}}$ as a function of the surface anisotropy $K_{\mathrm{s}}$ for the system with (c) SSA and OSA for (d) negative and (e) positive wave vector $k$. Odd-numbered ACs are shown with solid lines, even- numbered ACs with dashed lines, and second-order ACs with dotted lines. The $y$-axis is in the logarithmic scale. Plots present results of numerical simulations. Another observation concerns the system with the lowest value of $\Delta f_{\mathrm{AC1}}$ found for the thickness of 42.4 nm. Its dispersion relation is shown in Fig. 5(b). The AC1 is not visible in the full dispersion. Interestingly, the AC1 is still too small to be visible even after a close-up of the AC1 vicinity (inset plot in the lower right corner). To study the mode profiles in the vicinity of AC1, we chose three wave vectors: $k_{\mathrm{AC}}=$5.033\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$, $k_{\mathrm{L}}=$4.853\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$, and $k_{\mathrm{R}}=$5.2\text{\times}{10}^{6}\text{\,}\mathrm{rad}\mathrm{/}\mathrm{m}$$. The mode profiles are shown in the inset plot at the top part of Fig. 5(b). For $k_{\mathrm{AC}}$ (middle plot), the profiles are similar to the case of $L=$39\text{\,}\mathrm{nm}$$. It suggests, that the critical value of the thickness $L_{\mathrm{C}}>$42.4\text{\,}\mathrm{nm}$$. In the case of $k_{\mathrm{L}}$ (left plot), the profile of the first mode (orange line) has a character of the DE mode with a small amplitude reduction at the bottom due to the surface anisotropy. The second mode (green line) has the character of the first PSSW mode. This mode has a slightly larger amplitude at the bottom than at the top. The modes at $k_{\mathrm{R}}$ (right plot) have the same character as the modes at $k_{\mathrm{L}}$, but their order is reversed. It clearly shows that far from the AC (where $f_{2}-f_{1}\gg\Delta f_{\mathrm{AC1}}$), the modes have the same character on both sides of the AC, as if the interaction between them is negligible. It is worth noting that this interchange is not so clear in the case where $\Delta f_{\mathrm{AC}}$ is relatively large. In this case, the intermixing of the effects of the wave vector dependence and the short distance between the ACs relative to their size leads to a significant change in the mode profiles. ### III.4 Analysis of surface anisotropy constant dependence The analysis presented above was done for the case where the surface anisotropy constant $K_{\mathrm{s}}$ has a negative value, resulting in the partial pinning condition for the out-of-plane dynamic component of the magnetization. Now we can look at the case where $K_{\mathrm{s}}$ is positive, so the magnetization amplitude close to the surface is enhanced. The dispersion relation for the system with SSA for $K_{\mathrm{s}}^{\mathrm{t}}=K_{\mathrm{s}}^{\mathrm{b}}=$2500\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$ is shown in Fig. 6(a). The small $k$ range is comparable to the case of negative $K_{\mathrm{s}}$. However, for about $k=10^{7}$ rad/m, the DE mode reaches the local maximum at about 23 GHz and obtains negative group velocity. This effect is analogous to the effect of the volume perpendicular magnetic anisotropy [42]. On its way, the DE mode produces additional ACs, which did not occur in the case of FBC and negative $K_{\mathrm{s}}$. The frequency difference between the adjacent modes [Fig. 6(b)] shows that additional ACs are present for the first, second, and third PSSW modes. These ACs are marked with the letter ’x’. Also, an AC5 is present. However, it is not related to the AC5 occurring for negative anisotropy, therefore, it is also marked with ’x’. Figure 7: The AC size $\Delta f_{\mathrm{AC}}$ as a function of the external magnetic field $B_{0}$ for the system with (a) FBC, (b) SSA, and OSA for (c) negative and (d) positive wave vector $k$. Odd-numbered ACs are shown with solid lines, even-numbered ACs with dashed lines. The $y$-axis is in the logarithmic scale. Plots present results of numerical simulations. Next, we study the AC size as a function of the surface anisotropy constant $K_{\mathrm{s}}$ for the case of SSA [Fig. 6(c)] and OSA for negative [Fig. 6(d)] and positive [Fig. 6(e)] wave vector $k$. We calculated it numerically in the range from $-3000\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$ to $3000\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$ with a step of $100\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$. Almost all curves have a minimum similar to the one present in Fig. 4(d). A detailed analysis of the mode profiles agrees with the previous observation—in each case, the mode profiles at $k_{\mathrm{AC}}$ interchange, so we expect that for a critical value of $K_{\mathrm{s}}$ a crossing between modes should occur. The position of the minimum depends on the AC parity. AC2 has the smallest size at $K_{\mathrm{s}}=0$ (however, we expect the critical value to be very low, i.e., $|K_{\mathrm{s}}^{\mathrm{critical}}|<$50\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$). The odd-numbered ACs (AC1 and AC3) have the smallest size for positive $K_{\mathrm{s}}$ in the system with SSA and OSA at negative $k$, while for the system with OSA at positive $k$, the smallest value occurs for negative $K_{\mathrm{s}}$. Interestingly, for the system with OSA, $K_{\mathrm{s}}^{\mathrm{critical}}$ of the same AC is different for positive and negative wave vector range, which means that we can get a situation where the AC is present only on one side of the dispersion relation, while on the opposite side a crossing will be present. In general, the AC tends to have larger size for positive surface anisotropy than for negative surface anisotropy of the same value. In addition, we can see that for a wide range of positive surface anisotropy, additional ACs (marked with the letter ’x’) occur in all systems. Their source lies in the negative slope range of the dispersion relation, as discussed above. Their minima also follow the rule of the mode profile interchange, so we should expect these minima to go to zero as well. ### III.5 Analysis of external magnetic field dependence Finally, the effect of the external magnetic field $B_{0}$ on the AC size is shown in Fig. 7. The ACs have been calculated in the field range between 10 and 500 mT with a step of 10 mT. Almost all ACs are increasing with the increase of the external field. This observation correlates with the fact that $k_{\mathrm{AC}}$ also increases with the increase of external field. Then, the $k$-dependent terms in the DE mode profile [Eq. (13)] give a stronger contribution, and the profile asymmetry is increased, resulting in a stronger interaction with PSSW modes and a larger AC size. The most remarkable example is AC4 in the FBC system, which increases by a factor of 4.06 in the investigated field range. On the other hand, AC5 in the SSA system increases by only 3% in the same range. Interestingly, in the OSA system, the local minimum for the AC3 occurs in the positive wave vector range. The lowest detected value is 0.94 MHz at 280 mT. This minimum also has the source in the mode profile interchange at $k_{\mathrm{AC}}$, indicating the closing of the AC gap. In the direction of lower fields, the local maximum is present for 100 mT with the AC size of 27.4 MHz, while for higher fields it increases up to 120 MHz in the upper limit of the study of 500 mT. The results show that the external field provides a simple way to control the AC size, which is the easiest source of control from the experimental point of view. ## IV Conclusions In this article, we provide a comprehensive investigation of the SW dynamics in the ferromagnetic film in the DE geometry in the presence of surface anisotropy with the use of analytical model and numerical simulations. We compare three different cases: free boundary conditions, symmetrical surface anisotropy, and one-sided surface anisotropy. We show that the surface anisotropy significantly increases the size of the AC between DE and PSSW modes. In the case of OSA, the mirror symmetry breaking leads to the asymmetrical dispersion relation with respect to the wave vector $k$, which particularly affects the AC size. The surface anisotropy also has a strong influence on the shape of the mode profiles. We have studied in detail the impact of various parameters (i.e., film thickness, surface anisotropy constant, and external magnetic field) on the AC size. In general, the ACs shrink with the increase of film thickness or the decrease of the external magnetic field. Also, the parity of the AC has a strong influence on the AC size. For large positive surface anisotropy constant, the mode of DE character has a non-monotonic dispersion relation, which leads to the appearance of additional ACs for large wave vectors. In most cases, the increase in anisotropy leads to the increase in AC size. Interestingly, we found that under certain conditions, the AC can close and turn into a crossing. This phenomenon, known as accidental degeneracy, occurs for some particular ACs when the value of the surface anisotropy constant, the layer thickness, or the external magnetic field is changed. In the system with SSA, it occurs for both negative and positive wave vector, while in the system with OSA only on one side of the dispersion relation for a given set of parameters. The transition through the accidental degeneracy point in any parameter space is always associated with the exchange of the order of the mode profiles in the AC region. It is worth to note that the results shown in the paper are calculated for typical material parameters of CoFeB but the presented effects are universal and should also occur for different materials. The presence of surface anisotropy in magnetic thin films is ubiquitous. It is often considered a detrimental feature, but it can also be an essential property. The ability to control the anisotropy by voltage as well as to control its effects by an external magnetic field gives an additional advantage. Moreover, surface anisotropy of different strength on opposite surfaces provides a simple way to induce the nonreciprocity in the structure. We believe that surface anisotropy can be exploited in magnonic devices where asymmetrical transmission or the possibility to control the propagation of the SW is a fundamental property. ###### Acknowledgements. K.S. and M.K. acknowledge the financial support from National Science Centre, Poland, grants no. UMO-2020/39/I/ST3/02413 and UMO-2021/41/N/ST3/04478. The research leading to these results has received funding from the Norwegian Financial Mechanism 2014-2021 project no. UMO-2020/37/K/ST3/02450. K.S. acknowledges the financial support from the Foundation for Polish Science (FNP). The dataset for this manuscript is available on https://doi.org/10.5281/zenodo.8382924. ## Appendix A Relation between wave vectors $k$ and $q_{1}$ Figure 8: Wave vector $\tilde{q}_{1}=q_{1}L$ as a function of wave vector $k$ of six lowest modes of a CoFeB film of thickness $L=$100\text{\,}\mathrm{nm}$$ with (a) FBC, (b) SSA with $K_{\mathrm{s}}^{\mathrm{t}}=K_{\mathrm{s}}^{\mathrm{b}}=$-700\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$, and (c,d) OSA with $K_{\mathrm{s}}^{\mathrm{t}}=0$ and $K_{\mathrm{s}}^{\mathrm{b}}=$-1500\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$ for (c) negative and (d) positive wave vector $k$ in the external magnetic field $\mu_{0}H_{0}=$50\text{\,}\mathrm{mT}$$. Horizontal dashed black lines represent the values $q_{1}=n\pi/L$ for $n=1,2,3...$ Plots present analytical results. Figure 8 shows the wave vector $\tilde{q}_{1}=q_{1}L$ as a function of wave vector $k$ for three cases studied in the manuscript—FBC [Fig. 8(a)], SSA with $K_{\mathrm{s}}^{\mathrm{t}}=K_{\mathrm{s}}^{\mathrm{b}}=$-700\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$ [Fig. 8(b)], and OSA with $K_{\mathrm{s}}^{\mathrm{t}}=0$ and $K_{\mathrm{s}}^{\mathrm{b}}=$-1500\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{J}\mathrm{/}\mathrm{m}^{2}$$ for negative [Fig. 8(c)] and positive [Fig. 8(d)] wave vector $k$. In the low wave vector range (up to about $10^{7}$ rad/m), the plots are very similar to the dispersion relations shown in Figs. 2(a-d), including the presence of the gaps between the modes. For the case of FBC, the PSSW modes are placed exactly at $q_{1}=n\pi/L$, while for DE mode the value of $q_{1}$ is increasing and produces ACs with PSSW modes exactly as in the dispersion relation. Nevertheless, the large value of $q_{1}$ for DE mode is not decisive for the shape of the mode profile since the coefficients in Eq. (13) associated with $q_{1}$ give smaller contribution that those associated with $k$ (however, this contribution is not negligible). For the case of SSA and OSA, values of $q_{1}$ of PSSW modes are larger than $n\pi/L$. It is clear that a larger value of $q_{1}$ results in larger frequency of PSSW modes according to Eq. (10). In the large $k$ range for the case of FBC, the values of $q_{1}$ go back to $n\pi/L$, but this time for $n$ starting from 0. To achieve this feat, all modes in the range of the DE mode are decreasing in the value of $q_{1}$ in the dipole-exchange regime of the wave vector $k$. In the case of SSA and OSA, the values of $q_{1}$ are also larger than $n\pi/L$ but the difference is much larger than in the small $k$ range. ## Appendix B Toy model of interaction between modes Figure 9: The absolute value of overlapping integral $I$ as a function of film thickness $L$ for the system with (a) FBC, (b) SSA, and OSA for (c) negative and (d) positive wave vector $k$. Odd-numbered ACs are shown with solid lines, even-numbered ACs with dashed lines. The $y$-axis is in the logarithmic scale. The interaction between the modes can be explained using of a simplified model of mode profiles. In the case of FBC at $k=0$, the modes form a basis of cosine functions $m_{n}=A_{n}\cos{\left(n\pi\left(z-\frac{L}{2}\right)\right)},$ (25) where $m_{0}$ represents the DE mode and $m_{n>0}$ represents $n$-th order PSSW modes. $A_{n}$ is the normalization constant which assure that $\int_{-L/2}^{L/2}m_{n}^{2}\mathrm{d}z=1$. Assume that in the regime of small $k$, the PSSW modes remain unchanged with the change of the wave vector $k$, so their profiles are represented by Eq. (25). On the other hand, the DE mode is described by the function $m_{0}(k)=A_{0}e^{kz}.$ (26) In the presence of negative surface anisotropy, the PSSW modes are “squeezed” to satisfy the boundary conditions. Due to this effect, their profiles are modified such that $q_{1}=(n+p_{n})\pi/L$, where $p_{n}$ is the relative shift of wave vector. In the case of SSA, the mode profile is modified in the following way: $m_{n}=A_{n}\cos{\left((n+p_{n})\pi\left(z-\frac{n}{n+p_{n}}\frac{L}{2}\right)\right)}.$ (27) In the case of OSA, the mode profile of PSSW modes is represented by the function: $m_{n}=A_{n}\cos{\left((n+p_{n})\pi\left(z-\frac{L}{2}\right)\right)}.$ (28) We assume that the change in DE mode due to surface anisotropy is negligible. Basing on the results in Fig. 8, we assume that $p_{n}$ has a constant value of 0.03 for all PSSW modes and all thicknesses. Strength of the interaction between the modes is described by the overlapping integral $I_{ij}=\int_{-L/2}^{L/2}m_{i}m_{j}\mathrm{d}z.$ (29) Figure 9 shows the overlapping integral between the DE mode and the $n$-th order PSSW mode at $k_{\mathrm{AC}_{n}}$ as a function of layer thickness $L$. The model qualitatively reproduces the behavior shown in Fig. 4 showing that the overlapping integral is connected with the AC size. Firstly, there is an identical dependence on the PSSW mode parity. In the case of FBC [Fig. 9(a)], the overlapping integral has a larger value for the function representing odd- numbered PSSW modes than for even-numbered PSSW modes. In the case of SSA [Fig. 9(b)], the overlapping integral for even-numbered PSSW modes grows over the integral for odd-numbered PSSW modes which only increases slightly compared to the FBC case. In the case of OSA for negative wave vectors $k$ [Fig. 9(c)], the value of the overlapping integral is similar for all modes. In the case of positive wave vectors $k$ [Fig. 9(d)], we have successfully reproduced the presence of the minima for odd-numbered PSSW modes shown in Fig. 4(d). The positions of the minima—at 48, 104, 154, and 204 nm for the first, third, fifth, and seventh PSSW modes, respectively, is in good agreement with the positions of the minima in Fig. 4(d) (42, 98, 148, and 200 nm, respectively). ## References * Chumak _et al._ [2015] A. V. Chumak, V. I. Vasyuchka, A. A. Serga, and B. Hillebrands, Magnon spintronics, Nat. Phys. 11, 453 (2015). * Damon and Eshbach [1961] R. Damon and J. Eshbach, Magnetostatic modes of a ferromagnet slab, J. Phys. Chem. Solids 19, 308 (1961). * Grünberg _et al._ [1982] P. Grünberg, M. G. Cottam, W. Vach, C. Mayr, and R. E. Camley, Brillouin scattering of light by spin waves in thin ferromagnetic films (invited), J. Appl. Phys. 53, 2078 (1982). * Zhang _et al._ [2021] Z. Zhang, H. Yang, Z. Wang, Y. Cao, and P. Yan, Strong coupling of quantized spin waves in ferromagnetic bilayers, Phys. Rev. B 103, 104420 (2021). * Gladii _et al._ [2016] O. Gladii, M. Haidar, Y. Henry, M. Kostylev, and M. Bailleul, Frequency nonreciprocity of surface spin wave in permalloy thin films, Phys. Rev. B 93, 054430 (2016). * Tacchi _et al._ [2019] S. Tacchi, R. Silvani, G. Carlotti, M. Marangolo, M. Eddrief, A. Rettori, and M. G. Pini, Strongly hybridized dipole-exchange spin waves in thin Fe-N ferromagnetic films, Phys. Rev. B 100, 104406 (2019). * Wang _et al._ [2022] H. Wang, W. He, R. Yuan, Y. Wang, J. Wang, Y. Zhang, I. Medlej, J. Chen, G. Yu, X. Han, J.-P. Ansermet, and H. Yu, Hybridized propagating spin waves in a CoFeB/IrMn bilayer, Phys. Rev. B 106, 064410 (2022). * Dreyer _et al._ [2021] R. Dreyer, N. Liebing, E. R. J. Edwards, A. Müller, and G. Woltersdorf, Spin-wave localization and guiding by magnon band structure engineering in yttrium iron garnet, Phys. Rev. Mater. 5, 064411 (2021). * Song _et al._ [2021] W. Song, X. Wang, C. Jia, X. Wang, C. Jiang, D. Xue, and G. Chai, Nonreciprocal emergence of hybridized magnons in magnetic thin films, Phys. Rev. B 104, 014402 (2021). * Vaňatka _et al._ [2021] M. Vaňatka, K. Szulc, O. Wojewoda, C. Dubs, A. V. Chumak, M. Krawczyk, O. V. Dobrovolskiy, J. W. Kłos, and M. Urbánek, Spin-wave dispersion measurement by variable-gap propagating spin-wave spectroscopy, Phys. Rev. Appl. 16, 054033 (2021). * Rana and Otani [2019] B. Rana and Y. Otani, Towards magnonic devices based on voltage-controlled magnetic anisotropy, Commun. Phys. 2, 90 (2019). * Wang _et al._ [2017] Q. Wang, A. V. Chumak, L. Jin, H. Zhang, B. Hillebrands, and Z. Zhong, Voltage-controlled nanoscale reconfigurable magnonic crystal, Phys. Rev. B 95, 134433 (2017). * Rana and Otani [2018] B. Rana and Y. Otani, Voltage-controlled reconfigurable spin-wave nanochannels and logic devices, Phys. Rev. Appl. 9, 014033 (2018). * Ibrahim _et al._ [2016] F. Ibrahim, H. X. Yang, A. Hallal, B. Dieny, and M. Chshiev, Anatomy of electric field control of perpendicular magnetic anisotropy at Fe/MgO interfaces, Phys. Rev. B 93, 014429 (2016). * De Wames and Wolfram [2003] R. E. De Wames and T. Wolfram, Dipole‐Exchange Spin Waves in Ferromagnetic Films, J. Appl. Phys. 41, 987 (2003). * Wolfram [2003] T. Wolfram, Magnetostatic Surface Waves in Layered Magnetic Structures, J. Appl. Phys. 41, 4748 (2003). * Camley and Grimsditch [1980] R. E. Camley and M. Grimsditch, Brillouin scattering from magnons in ferromagnetic thin films of polycrystalline iron, Phys. Rev. B 22, 5420 (1980). * Kabos _et al._ [1984] P. Kabos, W. D. Wilber, C. E. Patton, and P. Grünberg, Brillouin light scattering study of magnon branch crossover in thin iron films, Phys. Rev. B 29, 6396 (1984). * Henry _et al._ [1972] R. Henry, S. D. Brown, P. E. Wigen, and P. J. Besser, Magnetoexchange branch repulsion in thin single-crystal disks of yttrium iron garnet, Phys. Rev. Lett. 28, 1272 (1972). * Arias and Mills [2001] R. Arias and D. L. Mills, Theory of spin excitations and the microwave response of cylindrical ferromagnetic nanowires, Phys. Rev. B 63, 134439 (2001). * Rychły _et al._ [2018] J. Rychły, V. S. Tkachenko, J. W. Kłos, A. Kuchko, and M. Krawczyk, Spin wave modes in a cylindrical nanowire in crossover dipolar-exchange regime, J. Phys. D: Appl. Phys. 52, 075003 (2018). * Kalinikos and Slavin [1986] B. A. Kalinikos and A. N. Slavin, Theory of dipole-exchange spin wave spectrum for ferromagnetic films with mixed exchange boundary conditions, J. Phys. C: Solid State Phys. 19, 7013 (1986). * Kalinikos _et al._ [1990] B. A. Kalinikos, M. P. Kostylev, N. V. Kozhus, and A. N. Slavin, The dipole-exchange spin wave spectrum for anisotropic ferromagnetic films with mixed exchange boundary conditions, J. Phys. Condens. Matter 2, 9861 (1990). * Akhiezer _et al._ [1968] A. Akhiezer, V. Bar’yakhtar, and S. Peletminskii, _Spin waves_ (North-Holland Pub. Co., Amsterdam, 1968). * Guslienko and Slavin [2011] K. Y. Guslienko and A. N. Slavin, Magnetostatic Green’s functions for the description of spin waves in finite rectangular magnetic dots and stripes, J. Magn. Magn. Mater. 323, 2418 (2011). * Guslienko _et al._ [2002] K. Y. Guslienko, S. O. Demokritov, B. Hillebrands, and A. N. Slavin, Effective dipolar boundary conditions for dynamic magnetization in thin magnetic stripes, Phys. Rev. B 66, 132402 (2002). * Gubbiotti _et al._ [2004] G. Gubbiotti, M. Conti, G. Carlotti, P. Candeloro, E. D. Fabrizio, K. Y. Guslienko, A. Andre, C. Bayer, and A. N. Slavin, Magnetic field dependence of quantized and localized spin wave modes in thin rectangular magnetic dots, J. Phys. Condens. Matter 16, 7709 (2004). * Bayer _et al._ [2005] C. Bayer, J. Jorzick, B. Hillebrands, S. O. Demokritov, R. Kouba, R. Bozinoski, A. N. Slavin, K. Y. Guslienko, D. V. Berkov, N. L. Gorn, and M. P. Kostylev, Spin-wave excitations in finite rectangular elements of ${\mathrm{Ni}}_{80}{\mathrm{Fe}}_{20}$, Phys. Rev. B 72, 064427 (2005). * Guslienko and Slavin [2000] K. Y. Guslienko and A. N. Slavin, Spin-waves in cylindrical magnetic dot arrays with in-plane magnetization, J. Appl. Phys. 87, 6337 (2000). * Kharlan _et al._ [2019] J. Kharlan, P. Bondarenko, M. Krawczyk, O. Salyuk, E. Tartakovskaya, A. Trzaskowska, and V. Golub, Standing spin waves in perpendicularly magnetized triangular dots, Phys. Rev. B 100, 184416 (2019). * Zhou _et al._ [2021] X. Zhou, E. V. Tartakovskaya, G. N. Kakazei, and A. O. Adeyeye, Engineering spin wave spectra in thick ${\mathrm{Ni}}_{80}{\mathrm{Fe}}_{20}$ rings by using competition between exchange and dipolar fields, Phys. Rev. B 104, 214402 (2021). * Kakazei _et al._ [2004] G. N. Kakazei, P. E. Wigen, K. Y. Guslienko, V. Novosad, A. N. Slavin, V. O. Golub, N. A. Lesnik, and Y. Otani, Spin-wave spectra of perpendicularly magnetized circular submicron dot arrays, Appl. Phys. Lett. 85, 443 (2004). * Tartakovskaya _et al._ [2016] E. V. Tartakovskaya, M. Pardavi-Horvath, and R. D. McMichael, Spin-wave localization in tangentially magnetized films, Phys. Rev. B 93, 214436 (2016). * Tartakovskaya [2005] E. V. Tartakovskaya, Quantized spin-wave modes in long cylindrical ferromagnetic nanowires in a transverse external magnetic field, Phys. Rev. B 71, 180404 (2005). * Harms and Duine [2022] J. Harms and R. Duine, Theory of the dipole-exchange spin wave spectrum in ferromagnetic films with in-plane magnetization revisited, J. Magn. Magn. Mater. 557, 169426 (2022). * Arias [2016] R. E. Arias, Spin-wave modes of ferromagnetic films, Phys. Rev. B 94, 134408 (2016). * Rado and Weertman [1959] G. Rado and J. Weertman, Spin-wave resonance in a ferromagnetic metal, J. Phys. Chem. Solids 11, 315 (1959). * Johnson _et al._ [1996] M. T. Johnson, P. J. H. Bloemen, F. J. A. den Broeder, and J. J. de Vries, Magnetic anisotropy in metallic multilayers, Rep. Prog. Phys. 59, 1409 (1996). * Gurevich and Melkov [1996] A. G. Gurevich and G. A. Melkov, _Magnetization Oscillations and Waves_ (CRC Press, London, 1996). * Herring [1937] C. Herring, Accidental degeneracy in the energy bands of crystals, Phys. Rev. 52, 365 (1937). * Huang _et al._ [2011] X. Huang, Y. Lai, Z. H. Hang, H. Zheng, and C. T. Chan, Dirac cones induced by accidental degeneracy in photonic crystals and zero-refractive-index materials, Nat. Mater. 10, 582 (2011). * Banerjee _et al._ [2017] C. Banerjee, P. Gruszecki, J. W. Klos, O. Hellwig, M. Krawczyk, and A. Barman, Magnonic band structure in a Co/Pd stripe domain system investigated by Brillouin light scattering and micromagnetic simulations, Phys. Rev. B 96, 024421 (2017).
# Constant Chemical Potential-Quantum Mechanical-Molecular Dynamics simulations of the Graphene-electrolyte double layer Nicodemo Di Pasquale Corresponding author<EMAIL_ADDRESS>Department of Chemical Engineering, Brunel University London, Uxbridge, UB8 3PH, United Kingdom Aaron R. Finney Department of Chemical Engineering, University College London, London, WC1E 7JE, United Kingdom Joshua Elliott Department of Chemical Engineering, University of Manchester, Manchester, M13 9PL, United Kingdom Diamond Light Source, Harwell Science and Innovation Park, Didcot, Oxfordshire OX11 8UQ, United Kingdom Paola Carbone Department of Chemical Engineering, University of Manchester, Manchester, M13 9PL, United Kingdom Matteo Salvalaglio Department of Chemical Engineering, University College London, London, WC1E 7JE, United Kingdom (October 2021) ###### Abstract We present the coupling of two frameworks—the pseudo-open boundary simulation method known as constant potential Molecular Dynamics simulations (C$\mu$MD), combined with QMMD calculations—to describe the properties of graphene electrodes in contact with electrolytes. The resulting C$\mu$QMMD model was then applied to three ionic solutions (LiCl, NaCl and KCl in water) at bulk solution concentrations ranging from 0.5 M up to 6 M in contact with a charged graphene electrode. The new approach we are describing here provides a simulation protocol to control the concentration of the electrolyte solutions while including the effects of a fully polarizable electrode surface. Thanks to this coupling, we are able to accurately model both the electrode and solution side of the double layer and provide a thorough analysis of the properties of electrolytes at charged interfaces, such as the screening ability of the electrolyte and the electrostatic potential profile. We also report the calculation of the integral electrochemical double layer capacitance in the whole range of concentrations analysed for each ionic species, while the QM simulations provide access to the differential and integral quantum capacitance. We highlight how subtle features, such as the adsorption of potassium at the interface or the tendency of the ions to form clusters, emerge from our simulations, contribute to explaining the ability of graphene to store charge and suggest implications for desalination. ## 1 Introduction Interest in graphene-based devices has grown in recent years, thanks of the versatility and physical characteristics of this new material, in particular for applications in which it is in contact with an electrolyte solution. Use of nanoporous graphene as a membrane for water desalination [1, 2] is one important example. The presence of pores of equal size to the electrolytes allows the selective passage of water through the membrane. Combined with the atomic scale thickness of graphene, this can lead to the creation of desalination membranes with higher performances than common polymer-based ones [3]. Another promising technologically relevant applications is the use of graphene electrodes in electrochemical double layer (super)capacitor (EDLC) devices[4, 5, 6]. In fact, graphene [7, 8, 9, 10], porous activated carbon [11] and carbon nanotube [12, 13] electrodes potentially have relatively high charge storage capacity and a favourable specific energy to power ratio, due to rapid charge-discharge cycling [8] controlled by changes of an applied potential, together with lifetimes that can reach millions of cycles [11]. Typically, charge storage at carbonaceous electrodes is a non-faradaic process, where mobile ionic species accumulate at the interface between the electrode and the liquid phase. An important class of systems of this kind, which has gained lots of attention recently, is represented by cheap and easy- to-prepare aqueous-based electrolytes in contact with a graphene electrode [6]. Carbon-based EDLCs with aqueous-based electrolytes do not generally suffer from electrochemical degradation, can be non-toxic, and provide an attractive alternative solution at the problem of energy storage compared with traditional battery devices. Combined with a longer lifetime and high power density,[14] these energy storage systems could be increasingly applied to power small electronic devices and for acceleration and breaking in electrical vehicles [5]. Several experimental works were undertaken to understand the physicochemical properties of neutral and charged graphene interfaces in contact with electrolyte solutions and the nature of these systems charge storage capacity [15, 16, 17]. However, the delicate balance between hydration-free energy and surface effects, which regulate the physisorption of ionic species at surfaces, resulted in conflicting experimental findings (see [18] for a more detailed account). For instance, there are reports both supporting the conclusion that the capacitance of graphene films is ion-independent [16], as well as contrasting observations suggesting that basal capacitance is instead ion-specific (with, for example, a greater propensity for Na+ and K+ adsorption over Li+ adsorption at negatively charged electrodes in the case of group I cations)[17]. Atomic-scale defects in the graphitic surface, its topography, dimensionality and chemical modifications are difficult to control and have non-negligible effects in experimental measurements. As an example, mechanical cutting produces structural defects known as “dangling bonds” which modifies the measured capacitance of the sample [15, 19]. In this respect, a model of the graphene interface and its interactions with an electrolyte solution can exclude all the spurious effects coming from uncontrolled defects and chemical modification of the surface. Molecular modelling and simulations can help to improve understanding of the mechanisms involved in such complex systems and guide the interpretation of experimental results. Many key features of supercapacitive devices are underpinned by the properties of the electrochemical double layer, and their responses to the charging of the electrode. Gouy-Chapman theory [20, 21] describes the double layer as a diffuse charged layer in the solution that compensates an applied surface charge on the electrode. Modifications to this model include the adsorption of counter-ions at the surface in the so-called Stern layer [22]. The development of a mean-field theory based on the Poisson-Boltzmann lattice-gas model [23] has shown that features not present in the Gouy-Chapman theory, such as steric effects, ion correlations, and preferential adsorption [24, 25, 26] need to be accounted for in order to correctly describe the interactions between the ions and the electrode. Mechanistic insight for these kinds of effects and how they control charge storage can be gained by atomistic simulations of the graphene/electrolyte interface; these also enable the evaluation of ensemble properties, such as the free energy of adsorption of the ions at the interface [27]. Furthermore, simulations can establish the effect of solution concentration on ion accumulation at the electrode, their interfacial structure, and dynamical properties. In order to compare simulations with a macroscopic system, this adsorption should ideally be modelled in the presence of bulk electroneutral solution with fixed composition to ensure a constant driving force for the adsorption at a charged surface. This can be obtained for example as shown in Finney et al. [28], where the authors performed MD simulations using constant chemical potential MD simulations, C$\mu$MD [29], which mimics open-boundary conditions. With C$\mu$MD, the authors simulated NaCl(aq) with concentrations spanning $\sim 0.1-10$ M at graphite surfaces. Their results indicate that the interface charge screening behaviour is a function of bulk solution concentration, with a transition (at $\sim 1$M) from diffuse charge screening, qualitatively consistent with the picture from simple mean field models, to a complex multi-layered structuring that systematically either over or under screens the surface potential. The multiple charged layers result from ion finite-size effects, over-compensation of the surface charge by oppositely charged ions closest to the surface, and non-idealities in solution, i.e., when the hypothesis of non-interaction between oppositely charged ions breaks down for large ions concentrations [30]. This last effect also has consequences on the conductance of the ions, which deviates from the prediction of the Nerst-Einstein equations [31]. Together with a constant driving force for ion adsorption from the bulk, another important effect to consider in the description of such systems is the the polarisation of the electrode exerted by the adsorbing electrolytes [18]. Classical simulations typically model the non-bonded interactions between atoms within the electrolyte and atoms belonging to the interface using additive pairwise potentials such as the Lennard-Jones potential and Coulomb interactions between fixed point atom charges. Polarisation can be introduced using e.g., oscillating charge models, or by fitting short-range potentials to binding energies obtained from ab initio methods [27, 32, 33]. However, these models may not accurately capture the complex many-body effect associated with charge polarization at the electrode-solution interface. Another way to include polarisation in classical MD simulations is through the constant potential method developed in [34]. This constant potential method has been successfully deployed to describe the properties of the electrochemical double layer of aqueous electrolytes and ionic liquids in contact with metal electrodes such as Au and Cu. Despite its successes, one of the key approximations of the constant potential method is that the electrode is fully metallic and can perfectly screen charges, which is not the case for (semimetallic) graphene [18]. On the other hand, a full Quantum Mechanical (QM) treatment of the interactions between the electrolyte and the substrate is still unfeasible, due to the length (tens of nm) and time (hundreds of ns) scales required for modeling the effect of the aqueous electrolytes. However, while the full QM model of the electrode/electrolyte system is out of reach, QM calculations can be used to compute a set of atomic partial charges on the electrode in the presence of the electrostatic potential arising from the position of the electrolyte atoms. This is exactly the spirit of our QMMD scheme, where QM calculations are coupled to MD simulations at fixed intervals of time integration. As such, the surface atom partial charges within the classical force field are updated on the fly. In a more recent development Machine Learning models have recently proven to be a viable option in tuning the surface polarization if the scope of the system becomes too large for QM simulations. This is achieved by replacing the QM calculations with a Neural Network (NN) model trained to reproduce results from a wide range of QM calculations with varying distributions of electrolytes in solution. The NN acts as a polarizable-like force field, combining fast classical MD simulations with more accurate QM calculations of the interface polarization [35]. This present work leverages the QMMD framework introduced in [18] and the C$\mu$MD introduced in [29, 28]. The approach simultaneously captures surface polarization and concentration effects that can modify the structure and composition of the electrochemical double layer. We use the resulting C$\mu$QMMD protocol to examine interfaces between aqueous alkali chloride solutions at different concentrations with a graphene electrode surface, elucidating complex interfacial structure, dynamics, and electrochemical properties. This paper is organized as follows: we first provide a brief overview of the QMMD and C$\mu$MD protocols, pointing to the relevant literature for the interested reader; we present the systems to which we apply the C$\mu$QMMD framework: a charged graphene electrode in contact with three different electrolyte solution, NaCl(aq), LiCl(aq), KCl(aq) at different concentrations. We derive the electrical properties of the interface in terms of the screening factor and electrical potential and calculate the total integral capacitance of this system by deriving the quantum and electrical double layer capacitance. Finally, we discuss the effects of complex solute speciation on the performance of graphene-electrolyte devices and draw some conclusions regarding this new proposed simulation scheme. ## 2 Computational Models In order to capture the dynamic polarization of a charged graphene surface in response to the evolving configuration of an electrolyte at a prescribed concentration, we coupled the classical C$\mu$MD simulation to the electronic structure theory calculations at regular time intervals. We will give a more detailed account of both models (C$\mu$MD and QMMD) in the following sections, while here we will only discuss their coupling. A sketch of the sequence of the operations involved is given in Figure 1. All the operations shown in Figure 1 are obtained through an in-house python wrapper. During the MD time integration obtained with GROMACS 2018.4 MD package [36], ion positions are passed to the Plumed software (v. 2.7) [37] patched with GROMACS, to compute the C$\mu$MD forces (see section 2.1 for more details). After the evolution of the atom positions, the final configuration of the electrolyte is extracted to compute the electrostatic potential. In turn, this latter quantity is used as input for the QM calculations obtained with the Dftb+ software package [38]. From the QM results, the distribution of the charges on the graphene is extracted (see section 2.2 for more details) and used as input for the new iteration of the loop. Figure 1: The computational workflow adopted in this work highlighting the two “black boxes” (the MD software and the QM software) in the blue squares and the operations included in the python wrapper (red squares). ### 2.1 C$\mu$MD Model Figure 2: Example configuration from a C$\mu$QMMD simulation of KCl(aq) in contact with graphene in this work projected onto simulation $x,z$ dimensions. K+, Cl-, O of water and C of graphene are shown by the pink, cyan, red and grey spheres. The blue lines highlight the C$\mu$QMMD control and reservoir regions, which also indicate the simulation cell boundaries. An extended vacuum region, around 8 nm in $z$, is truncated in the image. The graphene electrode we considered is located at $z=0$ and is in contact with an electrolyte slab of thickness 8 nm. A further 8 nm of vacuum separates the system from its periodically repeating images. The electrolyte phase is divided into three regions: the first region starts at the graphene electrode up to a distance of 4 nm. The second one is the control region, which is used to control the concentrations. The third region is the reservoir region which provides the reservoir of ions to adjust the concentration of the electrolytes in the other regions. Figure 2 provides an example of the set-up adopted in this work, where we highlighted the different C$\mu$MD simulation cell regions. The control of the concentration of the ions in solution is obtained by applying a force at the edge of the reservoir region according to a continuous function of the form, $F_{i}^{\mu}(z)=k_{i}(n_{i}^{\mathrm{CR}}-n_{i}^{0})\left[\frac{1}{4\omega}\left(1+\mathrm{cosh}\left(\frac{z-z_{F}}{\omega}\right)\right)^{-1}\right].$ (1) Here, $\omega$ was set to 0.2 nm, and represents the width of the force region (between the control and reservoir regions highlighted by the blue lines in Figure 2) while $k$ was $2\times 10^{4}$ kJ mol-1 nm-1, giving the correct densities in the bulk (see [28] for a discussion on these parameters). $n^{0}$ is the target ion number density, while $n^{\mathrm{CR}}$ is the density calculated instantaneously during time integration in the control region. Finally, $z_{F}$ is the position in $z$ where the C$\mu$MD forces are applied. In our simulations this is set to 5.5 nm beyond the graphene surface. Using this approach, the densities of cations and anions are constrained in the control region to maintain target concentrations of 0.5, 2.0, 3.0, 4.0, 4.4 and 6 M. At each MD time-step ion positions are passed to Plumed in order to compute the C$\mu$MD forces only acting on those ions in the region of $z_{F}$. No external forces are applied to the ions outside of this region, and any local change in the ion density at the interface results from the physical interactions between graphene and the solution. ### 2.2 QMMD Model The generality of electronic structure theory and its ability to reproduce the electronic charge density distribution in semiconductors, metals, and semimetals implies that that the QMMD approach can describe both long- and short-ranged redistribution of the surface charge induced by the presence of the electrolyte. Within each iteration (see Figure 1) of our scheme, the fully classical system is taken as input for a quantum mechanical calculation. The simulation box is partitioned into surface atoms whose electronic structure is explicitly treated, and electrolyte atoms that are converted into a set of point charges. The point charges take the values of the partial charges contained in the classical force field and form the background electrostatic potential during the computation of the electron structure. Upon derivation of the electronic structure, partitioning of the charge density via Mulliken population analysis yields the set of surface atom partial charges, which are then passed to the classical force field. Finally, a short MD trajectory on the order of several picoseconds can then be carried out (in the presence of the quantum mechanically polarized surface) to generate the electrolyte configuration for the following iteration. In our simulations we employ a coupling between QM and MD calculations of 5 ps. We previously found for this class of systems that 5 ps represents a good compromise in terms of computational accuracy of the computed charges (0.004 $e$) vs computing time when compared with a QMMD simulation where the charges were updated at every MD time step [18]. In practice, in order to describe the electronic structure of solid- electrolyte interfaces on the length scales required, we leverage the self- consistent charge Density Functional Tight-Binding (SCC-DFTB) [39] approach, which is an approximation to Kohn-Sham Density Functional Theory. The empirical description in our Dftb+ calculations of the interactions between the C atoms in the surface are described by the mio-1-1 parameter set. The SCC charge threshold and Fermi temperature have been set to $1\times 10^{-2}$ Hartree and 300 K, respectively. Whereas, on first inspection, these criteria can be considered loose and should not be adopted for the calculation of the total electronic energy, rigorous testing in our previous works [18, 40] found that they provide a sufficiently accurate description of the surface charge distribution with respect to fully converged simulations, at a fraction of the computational cost. Finally, to compute the partial charges passed to the graphene force field at each MD step, we perform a Mulliken population analysis [41], which gives reasonable results for this class of systems[18, 40]. ### 2.3 Simulations Details In our simulations, we consider a graphene electrode composed of 336 carbon atoms in contact with aqueous electrolyte solutions. We investigated three electrolyte systems, NaCl, KCl and LiCl at concentrations ranging from 0.5 M to 6 M. However, due to the solubility limits of the KCl(aq) [42, 43], we limit the investigated concentrations to 4.4 M for the KCl system. Our simulations are carried out at constant surface charge, which makes it difficult to draw comparisons across different electrodes since the potential applied is not necessarily constant. As such, when we compute the capacitance, we use the potential drop of the neutral electrode as a reference. This approach has been applied previously to compare the properties of the electrochemical double layer for different electrolytes [44]. Each operating condition was therefore repeated for two different total charges of the electrode: a charged graphene layer with a constant charge on the surface [45] $\sigma$ of -0.449 e nm-2 (-0.0719 C/$m^{2}$) and a neutral one ($\sigma=0$). Structural analyses of the solutions are carried out using PLUMED by post- processing the simulation trajectories. The first-shell coordination numbers for cations with anions ($N_{\mathrm{X-Cl}}$) and cations with water oxygen atoms ($N_{\mathrm{X-Ow}}$) were computed using a continuous switching function: $N=\frac{1}{M}\sum_{i}^{M}\mathbf{e}^{\left(\frac{-(r-d_{0})^{2}}{2r_{0}^{2}}\right)}$ (2) where $r$ are distances between atoms, $r_{0}$ was 0.01 nm, and $d_{0}$ was chosen such that the function goes smoothly from one to zero at the position of the first minimum in radial distribution functions for the cations with anions and water oxygen atoms. This ensured that a conservative definition of first-shell coordination was adopted in the analyses. Coordination numbers were evaluated in 1.3 nm regions in $z$ closest to the graphene surface and 3.5 nm from the surface, representing the double layer and bulk solution regions, respectively. The first coordination sphere distributions for ions were used to construct a graph of ion-ion contacts using the NetworkX Python library [46]. This allowed us to identify and compute the size of the ion clusters formed. Ion clusters at the interface and within the bulk were identified by sampling the regions defined for computing the coordination numbers. Molecular dynamics calculations in the NVT ensemble are carried out using GROMACS [47, 48], version 2018.4. The leapfrog algorithm with a timestep of 1 fs was used to integrate the equations of motion at a constant temperature of 298.15 K, controlled with the Nosé-Hoover thermostat, with a relaxation time of 0.1 ps. Long-range electrostatic interactions were treated using the particle-mesh Ewald approach, with a cut-off of 1.4 nm. Non-bonded interactions were computed using a Lennard-Jones 12-6 potential, truncated smoothly at 1.0 nm using a switch function starting at a distance of 0.99 nm. In all simulations, graphene carbon atoms were frozen, and water was modelled using the SPC/E model [49] with the SETTLE algorithm used to maintain rigid molecule geometries [50]. This choice is compatible with the Werder water- graphene parameters that reproduce the experimentally measured graphene/water contact angle [7, 51]. Ion force field parameters (for K+, Li+, Na+, Cl+), also compatible with the SPC/E model, are taken from the work of Joung and Cheatham [52]. In order to prevent water molecules and ions from escaping the solution into the vacuum space, we added a fixed wall above the reservoir, interacting with water molecules and ions only through a short-range Lennard- Jones potential. We equilibrated each system for 20 ns followed by 130 ns production runs to collect data for subsequent analyses of the steady-state structure of the interface. In all analyses discussed below, mean values and standard deviations (error bars) are obtained via averaging performed using 5 ns windows. ## 3 Results and Discussion Thanks to the simulation protocol implemented, electroneutral solutions with fixed ion concentrations can be maintained in the Control Region in Figure 2, representing bulk solutions in equilibrium with the electrode-solution interfaces. This allows us to compare the behaviour of different electrolytes while controlling the electrolyte background concentration. ### 3.1 Density Profiles We start this section by reporting in Figure 3 the concentration of the different ionic species in solution as a function of the $z$-coordinate, corresponding to the simulation cell direction orthogonal to the surface of negatively charged graphene electrodes. As expected, these profiles show preferential adsorption of cations at the electrode surfaces. For Na+ and Li+, a sharp density peak is observed at a distance of $0.5$ nm from graphene, followed by a second, less pronounced peak at $0.75$ nm. At the highest concentrations, a third cation peak emerges around $1.15$ nm, which is more pronounced for Li+. In contrast, in the case of K+, a small peak at $0.3$ nm is followed by a much larger and relatively diffuse density peak at $0.6$ nm. This is due to specific adsorption of the larger cation at the carbon surface, a small number of which partially dehydrate to directly coordinate to carbon. The difference in the $z$-density profiles for the different systems is less notable when considering Cl- with respect to cations. At the lowest bulk concentrations, there is a monotonically increasing density which reaches bulk values around 1.5 nm from the graphene interface. As the concentration rises, further density peaks are observed close to the carbon substrate, determined by the emergence of a multi-layered electrical double-layer structure, consistent with previously reported results [28, 18]. In such double-layer configurations, adjacent solution layers, rich in cations or anions, arise at the interface due to ion crowding (as in the case of the cations that are attracted towards the negatively charged surface of the electrode) and ion correlation (the localized positive excess charge in the closest layers to the electrode, in turn, attracts the anions). The results reported in Figure 3 are consistent with those of [40] with NaCl(aq) and LiCl(aq) systems displaying, qualitatively, the same solution side double layer structure. The case of KCl(aq) differs somewhat, with the same position of the two first two peaks for Figure 3(e) in both cases, but a different intensity compared with [40]. In turn, this intensity difference can be due to the different classical force fields used for water, carbon and ions as well as the use of scaled ionic charges not considered here. Besides these rather minor differences, the results presented here seem to be robust with respect to the chosen classical model. However, other results in the literature (see [53]) show clear qualitative differences (in particular for the KCl(aq) system where no adsorption is observed), most likely due to the lack of dynamic polarization considered for the graphene electrodes. (a) Na+ (b) K+ (c) Li+ (d) Cl- (NaCl) (e) Cl- (KCl) (f) Cl- (LiCl) Figure 3: Molar (M) density of the cations (top row) and the corresponding anions (bottom row) for the three systems considered in this work. Black, green, magenta, red, orange, blue curves correspond to bulk solution concentrations 0.5, 2, 3, 4, 4.4 and 6 M, respectively. ### 3.2 Electrical Double Layer Properties In this section we will derive and analyze the electrical properties of the electrode-electrolyte systems considered in this work. #### Electrode Charge Screeninig We begin by considering the screening factor [28] $f$ defined as: $f(z)=-\int_{0}^{z}\frac{\rho_{ions}(z^{\prime})}{\sigma}\mbox{d}z^{\prime}$ (3) where $\sigma$ is the superficial charge of the electrode interface and $\rho_{ions}(z)$ is the density charge of ions only, which is considered a function of just the $z$ coordinate, i.e., it is averaged over the $x$ and $y$ coordinates. The screening factor represents the extent to which the electrolyte phase electrically screens the charged interface. When $f$ converges to a value of one, the charge on the electrode is entirely shielded by the electrolyte. By considering only the ions in the calculations of $f$, we can compare their screening potential to predictions of simple mean field models. The integration shown in Equation 3 and Equation 5 is performed numerically. Data are first smoothed by applying the Savitzky-Golay [54] finite impulse response smoothing filter of order 3 with a window width of 5 points, implemented in Matlab. The smoothed curves obtained are then integrated using the trapezoidal rule. Error bars are computed by error propagation through the integration procedure. The screening factors for all systems are reported in Figure 4. When the concentration of the ions is below 1 M, an under-screening near the interface can be observed. $f$ increases smoothly to a value of one at around $z=2$ nm. This is qualitatively consistent with the predictions of Gouy-Chapman’s theory. For higher concentrations, however, $f$ transitions to over-screening at relatively small values of the $z$ coordinate. The over-screening, highlighted by the first peak at $z\approx 0.6$ nm reported in Figures 4(a), 4(b) and 4(c), depends both on the particular ion and the bulk concentration. In particular, the LiCl system has the strongest over-screening effect on the electrode across the entire concentration range considered. Over-screening is a well-known effect for ionic liquids [25] and is usually not considered important in the electrolyte solutions, as this was only apparent at relatively high concentrations [18, 28, 53]. The fact that over-screening appears for higher concentration of the solute, in turn, can be linked directly to the structuring of the ions near the interface observed in Figure 3. With the increase in concentration, the density of the cations closest to the electrode increases with respect to their value in the solution bulk (see Figure 3). The excess charge associated with this ion accumulation is balanced in adjacent solution layers until the average bulk density is reached [14]. This description is consistent with our observations, where lithium and sodium show a high degree of structuring near the interface relative to potassium (i.e., multiple ion density peaks are observed, accompanied by a significant over-screening effect). In contrast, potassium, with the lowest degree of structuring near the interface, shows the smallest over-screening among the three ion solutions considered. Moreover, for potassium, we observe a variation in the slope of the screening factor when $z\approx 0.5$ nm, which increases (becoming more pronounced) as a function of concentration. This additional feature in the screening factor, absent in NaCl and LiCl, can be explained by the direct coordination of the K+ (i.e., through the first coordination sphere) to carbon atoms (as also observed in [40]), as opposed to the behaviour of the cations in LiCl(aq) and NaCl(aq) systems (see the first peak at $\approx$0.35 nm in Figure 3(b) with respect to the first peak at $\approx$0.5 nm in Figures 3(c) and 3(a)). (a) KCl (b) LiCl (c) NaCl Figure 4: Screening factor as defined in Equation 3 for the three systems considered using ion solution charge densities only. We included only a subset of the concentrations for clarity, and the results for all the concentrations are reported in the SI (see Fig.S3 of the SI). #### Electrode Polarisation The coordination of the K+ with the carbon atoms on the graphene electrode is shown qualitatively in Figure 3(b) for the lowest (0.5 M) and the highest concentration (4.4 M) considered here. The plots in Figure 5 represents a single snapshot in the 150 ns long simulation with the highest number of potassium cations in direct contact with the interface (i.e., at a distance of $0.26$ nm from the interface). As expected, the number of K+ in direct contact with the interface increases as the bulk concentration of the cations increases, consistent with the observation in Figure 4 for the short-distance (from the electrode) behaviour of the screening factor, which increases with concentration. The accumulation of K+ in the nearby region to the negative electrode (see Figure 3(b)) results in an increased non-uniformity of the partial charge distribution on the electrode, with higher negative charges located on the carbons closer to the coordinated K+. This, in turn, demonstrates how polarisation effects are important to be considered in systems where direct coordination of electrolytes to the electrode may occur. (a) 0.5 M (b) 4.4 M Figure 5: Representative plot of the computed Mulliken charges on the graphene sheet charged with 4 and in contact with kCl solutions at different concentrations. Circled X’s mark the coordinates of K ions directly adsorbed on the surface. (a) KCl (b) LiCl (c) NaCl Figure 6: Electrostatic potential as defined in Equation 5 for the three systems considered. We included only a subset of the concentrations for clarity. We included the results for all the concentrations in the SI (see Fig.S3 of the SI). #### Electrical Potential in the Double Layer We calculated the electrical field $E(z)$ and the electrical potential, $\psi(z)$ in the direction orthogonal to the interface using the Poisson equation: $-\frac{\mbox{d}^{2}\psi(z)}{\mbox{d}z^{2}}=\frac{\mbox{d}E(z)}{\mbox{d}z}=\frac{\rho(z)}{\epsilon(z)}$ (4) where $\rho(z)$ is the charge density calculated for all atoms on the perpendicular axis and we defined $\epsilon(z)=\epsilon_{r}(z)\epsilon_{0}$, the product of the permittivity in vacuum $\epsilon_{0}$ and relative permittivity $\epsilon_{r}$. It was reported that this latter quantity could be a function of the distance from the electrode [55], a function of the concentration of the electrolyte [16], or possibly both. Given such uncertainties, we consider a constant relative permittivity equal to one in this work. The electrical potential, $\psi$(z), is obtained from Equation 4 by integrating twice with respect to the $z$-coordinate: $\psi(z)=-\int_{0}^{z}\int_{0}^{z^{\prime}}\frac{\rho(\zeta)}{\epsilon(\zeta)}\mbox{d}\zeta\mbox{d}z^{\prime}$ (5) The two integration constants in Equation 5 are chosen to set the electrostatic field and potential equal to zero in the bulk, which amounts to considering the bulk as the reference for the calculation of the electrostatic potential. The results of Equation 5 are reported in Figure 6 for a selection of concentrations (see Figure S.3 of the SM for the entire range of concentrations). In stark contrast to the exponential behaviour predicted by models based on the Gouy-Chapman double layer theory which treats the solvent medium as a continuum with known dielectric, atom/molecule finite-size effects give rise to an undulating $\psi(z)$ function in the interfacial region at all concentrations and in all systems. When calculating the charge distribution, we include all solution species, including water molecules partial charges. Hence, it is unsurprising that the structuring of ions and water molecules at the interface gives rise to a significant departure from the predictions of simple mean field models. Indeed, these finite size effects are a well- reported feature of electrode-electrolyte systems [56, 57]. From a relatively large negative value of the potential at the electrode, the (partial) charges of ions and water give rise to fluctuations that attenuate at larger values of $z$, where the bulk solution behaviour is recovered. Generally, increasing the bulk solution concentration increase the amplitude of $\psi(z)$ fluctuations. Furthermore, it is evident from Figure 6(b) and Figure 6(c) that the crowding of ions in the double-layer increases with concentration as the positions of peaks and minima in $z$ shift to lower values, a feature also observed by Finney et al. [28] with graphite and which was related to changes in the screening factor. This concentration dependence is less apparent in the case of KCl(aq), where the value of $\psi(z)$ at the first maximum is less susceptible to changes in the concentration as opposed to NaCl(aq) and LiCl(aq). #### Electrical Double Layer Capacitance The total capacitance $C_{TOT}$ in these kinds of systems is usually considered as composed of three independent components combined in series: the Electrochemical Double-Layer Capacitance (EDLC), $C_{EDL}$, and the quantum capacitance (or the space charge capacitance)($C_{Q}$), depending on the spatial distribution of the charges on the graphene [40]. The total capacitance is then given by $\frac{1}{C_{TOT}}=\frac{1}{C_{EDL}}+\frac{1}{C_{Q}}$ (6) From Figure 6 we can easily derive the potential drop, $\Delta\psi$, across the interface as*** A more precise notation for the potential drop across the interface would have been $\Delta\Delta\psi=\Delta\psi^{-}-\Delta\psi_{ref}$. [40] $\Delta\psi=\Delta\psi^{-}-\Delta\psi_{ref}$ where $\Delta\psi^{-}$ and $\Delta\psi_{ref}$ represent the potential drop at the interface with respect to the bulk for the charged and neutral electrodes, respectively. As a reference for the calculation of the potential drop, we use the potential at the interface in a neutral electrode with all other conditions unchanged. We report the calculation of the potential across the system for a neutral electrode in the SI (see Figure S.4 of the SI) along with the potential drop at the interface ($\Delta\psi_{ref}$) (see Table S.2 of the SI). With this definition of the potential drop, the EDLC can be obtained, as $C_{EDL}=\frac{\sigma}{\Delta\psi}$ (7) The quantum capacitance instead is obtained by calculating the differential quantum capacitance $C_{Q}^{diff}$ according to [40]: $C_{Q}^{diff}(\psi)=\frac{e^{2}}{4k_{B}T}\int_{-\infty}^{\infty}\left[D(E)\mbox{sech}^{2}\left(E+\psi\right)\right]\mbox{d}E$ (8) Where $e$ is the electron charge, $E$ is the energy relative to the Fermi level, $D(E)$ is the density of states at a given energy, $k_{B}$ is the Boltzmann constant, and $T$ is the temperature. By integrating the differential quantum capacitance with respect to the potential $\psi$ up to the potential drop $\Delta\psi$ calculated for each system, we obtain the integral quantum capacitance $C_{Q}$: $C_{Q}=\frac{1}{\Delta\psi}\int_{0}^{\Delta\psi}C_{Q}^{diff}(\psi)\mbox{d}\psi$ (9) For more detailed information about the calculation of the quantum capacitance we refer the reader to our previous work [40]. concentration | $\Delta\psi$ | $C_{EDL}$ | $C_{Q}$ | $C_{TOT}$ ---|---|---|---|--- | KCl 0.5 | -1.03 | 6.95 | 10.56 | 4.19 2.0 | -1.01 | 7.10 | 10.31 | 4.20 3.0 | -1.00 | 7.16 | 10.23 | 4.21 4.0 | -0.984 | 7.28 | 9.81 | 4.18 4.4 | -0.988 | 7.25 | 10.07 | 4.22 | LiCl 0.5 | -1.05 | 6.82 | 10.78 | 4.18 2.0 | -1.01 | 7.09 | 10.31 | 4.20 3.0 | -1.00 | 7.16 | 10.21 | 4.21 4.0 | -1.00 | 7.16 | 10.21 | 4.21 4.4 | -1.09 | 6.57 | 11.25 | 4.15 6.0 | -1.02 | 7.02 | 10.44 | 4.20 | NaCl 0.5 | -1.05 | 6.82 | 10.78 | 4.18 2.0 | -1.02 | 7.02 | 10.44 | 4.20 3.0 | -0.997 | 7.18 | 10.17 | 4.21 4.0 | -0.996 | 7.19 | 10.15 | 4.21 4.4 | -0.986 | 7.26 | 10.05 | 4.22 6.0 | -1.00 | 7.16 | 10.21 | 4.21 Table 1: Electrostatic potential drop ($\Delta\psi$) across the interface (in V), Electrochemical double layer capacitance $C_{EDL}$, Quantum Capacitance $C_{Q}$, and total capacitance $C_{TOT}$ (in $\mu$F cm-2) for each concentration considered (in M). The results for $C_{Q}$, $C_{EDL}$, and $C_{TOT}$ for all of the systems considered are reported in Table 1. The data show that the total capacitance is practically constant across all the concentration range and for all solution types. The largest variation in $C_{TOT}$ we obtained among all the systems is $\approx$2% (between the LiCl(aq) and KCl(aq) at 4.4 M). This result contrasts with the different behaviour of the three cations in solution and near the electrode interfaces, as highlighted in the discussion of the number density of ionic species at the interface (see Figure 3) their screening effect on the charge of the electrode (Figure 4), and as further discussed in the following section in relation to their clustering properties. An important point we want to highlight here is that such differences in the behaviour of the cation in solution can be correctly captured through the use of a simulation protocol that combines the pseudo-open boundary condition, i.e., C$\mu$MD to maintain constant composition electroneutral bulk solutions beyond the double layer, and the quantum mechanical description for the distribution of partial charges of the electrode. However, while the capacitance is a critical parameter for this kind of system’s applications as supercapacitors, we showed here that the physics of the interfaces between graphene electrodes and electrolytes is much richer than the one captured by such quantity. ### 3.3 Ion Association An often overlooked effect in systems in alkali chloride solutions is the tendency for ions to associate, forming clusters. In particular, even simple salt solutions exhibit significant non-ideal behaviour at high concentrations. Recent experiments [58] and simulations [59] have shown that extended liquid- like clusters exist in bulk NaCl(aq) at high concentrations and the extent of these ionic networks is promoted in the double layer at carbon surfaces [28]. Since the effectiveness of the graphene-electrolyte devices often depends on the ability to ‘build up the double layer’ (i.e., accumulate ions from the bulk solution in the interfacial region), the structure and mobility of ion species can be essential to this. #### Ion Clusters To identify and characterise ion associates in the simulations in this work, pairwise RDFs were computed (see Fig. S.1 of the Supplementary Material, (SM)), and the first minima in these informed truncation distances ($r_{c}$) for first-sphere ion-ion coordination. $r_{c}=0.29$, 0.34 and 0.39 nm for Li– Na– and K–Cl, respectively, reflecting the different sizes of the cations. Clusters were identified as fully connected networks in the graph of adjacent ion-ion connections according to this geometric criteria, regardless of their total charge or lifetime. Figure 7 provides the average first-sphere coordination number between cations and O of water (see Figure 7(a)) as well as cations and anions for all systems, calculated using Equation 2. (a) Cation/Water (b) Cation/Anion Figure 7: Coordination number for the different systems at the different concentrations. The results shown in Figure 7 indicate no significant surface effect on the coordination of cations with water or chloride when ions in the interface ($0<z<2.5$ nm) and bulk ($2.5<z<4.5$ nm) regions were investigated. There is a slight increase in the mean cation-anion coordination, and a concomitant decrease in cation-water coordination, at the interface compared to the bulk; however, this difference is within the margin of error. Generally, the effect of increasing concentration is to increase the number of cation-anion contacts, particularly for KCl(aq), where the coordination number is more than double that of the other systems for all concentrations (and with Li–Cl coordination being negligible even at 6 M). From the largest to smallest variation in the coordination number we can write $\mbox{K}^{+}\rightarrow\mbox{Na}^{+}\rightarrow\mbox{Li}^{+}$. This trend follows the decrease of the ion radius and it is likely due to the stronger binding of water in the solvation spheres of smaller cations. Furthermore, the average cation-water coordination number is unchanging with a concentration within the margin of error. In simulations of NaCl(aq) in contact with graphite, [28] the substrate was found to increase cation-anion correlations in the double layer with respect to the bulk, particularly beyond $5$ M. It is important to note that different models (due to the different system) were used and also that system size likely plays a role in the extent that clusters can grow (both in e.g., the system-size dependence of the availability of ions to form associates and the extent to which finite-size and percolating clusters may form in effectively confined canonical systems.) The change in coordination for different salts is reflected in the cluster size probability distributions presented in Figure 8 for the case of 4.4 M (we report the results for the entire range of concentrations in Figure S.5 of the SM). There is a clear difference in the extent to which clusters can grow, with lithium forming clusters containing at most four ions and potassium forming much larger networks containing as many as 35 ions. Even at the highest concentrations, the majority of the Li+ are dispersed in solution, fully solvated in their first shell. A snapshot of a configuration obtained during the simulation of KCl at 4.4 M is shown in Figure 8. Although the most probable clusters contain only a few ions (for clusters composed of five ion units, we obtained a relative frequency of 0.01), larger species do contribute to the charge storage capacity and must be considered. What we observe is a stronger tendency of the potassium to associate into large aggregates—albeit ones which are highly dynamic on the timescales of the simulations—compared to sodium or lithium. Since the KCl(aq) system shows the formation of large aggregates of ions, it is interesting to study the relative frequency of the charge of these aggregates. In Figure 9 we plot the 2-dimensional histogram showing the relative frequencies of the charge vs. the cluster size for the KCl(aq) system. The histogram is skewed towards positive charges, with the appearance of clusters containing an excess of positive charge as large as +7e, although the majority of the clusters are neutral. Figure 8: On the left: Histogram of the relative frequency of the cluster of different sizes for the concentration of 4.4 M. In the inset, the same quantity for the 0.5 M case. On the right: an example of a cluster composed of 26 ions for the KCl system at 4.4 M. Figure 9: 2-dimensional histogram (charge VS dimension of the clusters) for the KCl(aq) system at the largest concentration considered (4.4 M). #### Ion Mobilities As well as a high capacity to store charge, an optimal charge storage device must also be a good electrical conductor. In this regard, we might expect the conductivity of solutions to decrease when clusters are present. Indeed, this can be perceived as a relative decrease in the activity of charge carriers due to increasingly non-ideal solutions. To test this, we calculated the conductivity of bulk NaCl(aq) solutions with concentrations ranging from 1–10 M from the ion diffusion coefficients calculated by Finney and Salvalaglio in finite size systems and in the dilute limit [59, 60]. To determine conductivity we use the Nernst-Einstein equation: $\sigma_{NE}=\frac{e^{2}}{Vk_{\mathrm{B}}T}(N_{+}z_{+}^{2}D_{+}+N_{-}z_{-}^{2}D_{-})$ (10) where $e$, $V$, $k_{\mathrm{B}}$ and $T$ are the elementary charge, simulation cell volume, Boltzmann’s constant and temperature, respectively. $N$ and $D$ are the total number of ions and diffusion coefficients for ions with charge indicated by the subscript. Furthermore, we assume that, given the highly dynamic nature of the clusters observed in solution, the valency of ionic species, $z$, is equal to one. Figure 10: Solution conductivities, $\sigma_{NE}$, of bulk NaCl(aq) solutions calculated for a range of concentrations. To this aim, the Nernst-Einstein equation was adopted where ion diffusion coefficients were determined from simulations at finite concentration, $D_{ion}$ (blue), or from a single simulation at the dilute limit, $D_{ion}^{0}$ (red). Dashed lines are a guide for the eye, while error bars indicate uncertainties in the conductivities associated with the calculated $D$ value from Refs. [59] and [60]. Figure 10 provides the solution conductivities for NaCl(aq) where either the diffusion of ions in finite concentration simulations was used ($D_{ion}$) or the diffusion of ions in the dilute limit ($D_{ion}^{0}$) was considered. For the latter, ions are assumed to be completely, dispersed as association, even beyond the second solvation sphere, did not occur in simulations at the dilute limit. For the estimate of $D_{ion}^{0}$, Finney and Salvalaglio [60] performed extended simulations of a single cation and anion in 4,000 water molecules; here, $D_{+}^{0}=1.223\pm 0.005\times 10^{-5}$ cm2 s-1 and $D_{-}^{0}=1.282\pm 0.008\times 10^{-5}$ cm2 s-1. In all cases, diffusion coefficients were corrected to account for simulation finite size effects [61]. Unsurprisingly, a linear correlation in $\sigma_{NE}$ as a function of concentration is found when a constant $D_{ion}^{0}$ is used for the diffusion of ions, independent of concentration. This is inaccurate at relatively high concentrations, given the simulation and experimental observations of ion-ion correlations[59, 58]. When accounting for the non-idealities in the solution and the formation of clusters explicitly in the diffusion of ions, we find that the solution conductivity reaches an upper limit between 4 and 5 M. At the lowest concentrations (1–2 M), the conductivity from finite concentration and dilute simulation data agree, and the simulation predictions match well with experimental measurements [62]. A crossover in the conductivity behaviour from the ‘pseudo-ideal’ to non-ideal regime occurs between 2 and 3 M. Therefore, over a wide concentration range up to the salt solubility, non-idealities will affect the performance of electrical devices; depending upon the chosen application, electrolytes should be chosen to minimize these effects. ## 4 Conclusions In this work, we presented an extended set of simulations describing the interface between three different electrolyte solutions - (KCl(aq), LiCl(aq), and NaCl(aq)) - in contact with the surface of a negatively charged graphene electrode. To investigate these systems, we combined QM/MD and C$\mu$MD methodologies into a new simulation framework. QM/MD models of the graphene electrode in contact with an electrolyte enabled the explicit coupling of the electrode polarizability with the instantaneous configuration of the electrolyte. The latter was maintained in equilibrium with a liquid phase at constant bulk concentration thanks to the C$\mu$MD model, which mimics open- boundary conditions. We performed a thorough analysis of the interaction of the ions with the electrode by showing the different behaviour of the three cations in the double layer, focusing on K+, which, according to our results, is able to directly adsorb at the electrode surface at shorter distances compared to Li+ and Na+, modifying the screening effect of the solution. Calculations of the integral capacitance indicated no concentration dependence or specific ion effects, with a total capacitance of around 4.2 µF cm-2 across all systems. However, the lack of variation in capacitance hides the rich electrolyte solution behaviour, particularly for the ions close to the electrode. We showed, for example, that large KCl clusters emerge in solution, which might be important when considering properties associated with ion mobility and charge transfer. Our results indicate that accurate models of the interface - able to account for the position-dependent non-ideality of electrolyte solutions - better capture the configurational and dynamical details underpinning the electrochemical behavior of interfaces at the atomistic level, and that is often overshadowed by the calculation of aggregated quantities such as the integral capacitance. We plan to extend our calculations to include a range of charged electrodes, both positive and negative, and further investigate ion dynamics in solution. ## 5 Acknowledgements We acknowledge the support provided by the IT Services use of the Computational Shared Facility (CSF) and at the University of Manchester. NDP, JDE and PC thank the European Union’s Horizon 2020 research and innovation programme project VIMMP under Grant Agreement No. 760907. ARF and MS acknowledge funding from an EPSRC Programme Grant (Grant EP/R018820/1), which funds the Crystallisation in the Real World consortium. ## References * Cohen-Tanugi and Grossman [2012] David Cohen-Tanugi and Jeffrey C Grossman. Water desalination across nanoporous graphene. _Nano letters_ , 12(7):3602–3608, 2012. * Heiranian et al. [2021] Mohammad Heiranian, Yechan Noh, and Narayana R Aluru. Dynamic and weak electric double layers in ultrathin nanopores. _The Journal of Chemical Physics_ , 154(13):134703, 2021. * Surwade et al. [2015] Sumedh P Surwade, Sergei N Smirnov, Ivan V Vlassiouk, Raymond R Unocic, Gabriel M Veith, Sheng Dai, and Shannon M Mahurin. Water desalination using nanoporous single-layer graphene. _Nature nanotechnology_ , 10(5):459–464, 2015\. * Simon and Gogotsi [2008] Patrice Simon and Yury Gogotsi. Materials for electrochemical capacitors. In _Nanoscience and technology: a collection of reviews from Nature journals_ , pages 320–329. World Scientific, 2008. * Wang et al. [2021] Yifan Wang, Lin Zhang, Haoqing Hou, Wenhui Xu, Gaigai Duan, Shuijian He, Kunming Liu, and Shaohua Jiang. Recent progress in carbon-based materials for supercapacitor electrodes: a review. _Journal of Materials Science_ , 56(1):173–200, 2021. * Elliott et al. [2022a] J. Elliott, A. A. Papaderakis, R. Dryfe, and P. Carbone. The electrochemical double layer at the graphene/aqueous electrolyte interface: what we can learn from simulations, experiments, and theory. _Journal of Materials Chemistry C_ , 2022a. * Wang et al. [2009] Yan Wang, Zhiqiang Shi, Yi Huang, Yanfeng Ma, Chengyang Wang, Mingming Chen, and Yongsheng Chen. Supercapacitor devices based on graphene materials. _Journal of Physical Chemistry C_ , 113(30):13103–13107, 2009. * Liu et al. [2010] Chenguang Liu, Zhenning Yu, David Neff, Aruna Zhamu, and Bor Z Jang. Graphene-based supercapacitor with an ultrahigh energy density. _Nano letters_ , 10(12):4863–4868, 2010. * Yu et al. [2010] Aiping Yu, Isaac Roes, Aaron Davies, and Zhongwei Chen. Ultrathin, transparent, and flexible graphene films for supercapacitor application. _Applied physics letters_ , 96(25):253105, 2010\. * Zhang et al. [2010] Li Li Zhang, Rui Zhou, and XS Zhao. Graphene-based materials as supercapacitor electrodes. _Journal of Materials Chemistry_ , 20(29):5983–5992, 2010. * Zhu et al. [2011] Y. Zhu, S. Murali, M. D. Stoller, K. J. Ganesh, W. Cai, P. J. Ferreira, A. Pirkle, R. M. Wallace, K. A. Cychosz, M. Thommes, D. Su, E. A. Stach, and R. S. Ruoff. Carbon-based supercapacitors produced by activation of graphene. _science_ , 332(6037):1537–1541, 2011. * An et al. [2001] Kay Hyeok An, Won Seok Kim, Young Soo Park, Young Chul Choi, Seung Mi Lee, Dong Chul Chung, Dong Jae Bae, Seong Chu Lim, and Young Hee Lee. Supercapacitors using single-walled carbon nanotube electrodes. _Advanced Materials_ , 13(7):497–500, 2001. * Yang et al. [2019] Zhoufei Yang, Jiarui Tian, Zefang Yin, Chaojie Cui, Weizhong Qian, and Fei Wei. Carbon nanotube-and graphene-based nanomaterials and applications in high-voltage supercapacitor: A review. _Carbon_ , 141:467–480, 2019. * Merlet et al. [2013] C. Merlet, B. Rotenberg, P. A. Madden, and M. Salanne. Computer simulations of ionic liquids at electrochemical interfaces. _Physical Chemistry Chemical Physics_ , 15(38):15781–15792, 2013. * Iamprasertkun et al. [2019] P. Iamprasertkun, W. Hirunpinyopas, A. Keerthi, B. Wang, B. Radha, M. A. Bissett, and R. A. W. Dryfe. Capacitance of basal plane and edge-oriented highly ordered pyrolytic graphite: specific ion effects. _The Journal of Physical Chemistry Letters_ , 10(3):617–623, 2019. * Yang et al. [2017] H. Yang, J. Yang, Z. Bo, X. Chen, X. Shuai, J. Kong, J. Yan, and K. Cen. Kinetic-dominated charging mechanism within representative aqueous electrolyte-based electric double-layer capacitors. _The Journal of Physical Chemistry Letters_ , 8(15):3703–3710, 2017. * Qu et al. [2008] QT Qu, B Wang, LC Yang, Y Shi, S Tian, and YP Wu. Study on electrochemical performance of activated carbon in aqueous li2so4, na2so4 and k2so4 electrolytes. _Electrochemistry Communications_ , 10(10):1652–1655, 2008. * Elliott et al. [2020] J. D. Elliott, A. Troisi, and P. Carbone. A QM/MD coupling method to model the ion-induced polarization of graphene. _J. Chem. Theory Comput._ , 16(8):5253–5263, 2020\. * Wang et al. [2008] Xinran Wang, Scott M Tabakman, and Hongjie Dai. Atomic layer deposition of metal oxides on pristine and functionalized graphene. _Journal of the American Chemical Society_ , 130(26):8152–8153, 2008. * Gouy [1910] MJJPTA Gouy. Sur la constitution de la charge électrique à la surface d’un électrolyte. _J. Phys. Theor. Appl._ , 9(1):457–468, 1910\. * Chapman [1913] D. L. Chapman. LI. A contribution to the theory of electrocapillarity. _The London, Edinburgh, and Dublin philosophical magazine and journal of science_ , 25(148):475–481, 1913. * Stern [1924] Otto Stern. Zur theorie der elektrolytischen doppelschicht. _Zeitschrift für Elektrochemie und angewandte physikalische Chemie_ , 30(21-22):508–516, 1924. * Popović and Šiber [2013] Marko Popović and Antonio Šiber. Lattice-gas poisson-boltzmann approach for sterically asymmetric electrolytes. _Physical Review E_ , 88(2):022302, 2013. * Borukhov et al. [1997] Itamar Borukhov, David Andelman, and Henri Orland. Steric effects in electrolytes: A modified poisson-boltzmann equation. _Physical review letters_ , 79(3):435, 1997. * Fedorov and Kornyshev [2008] M. V. Fedorov and A. A. Kornyshev. Ionic liquid near a charged wall: Structure and capacitance of electrical double layer. _The Journal of Physical Chemistry B_ , 112(38):11868–11872, aug 2008. doi: 10.1021/jp803440q. * Howard et al. [2010] J. J. Howard, J. S. Perkyns, and B. M. Pettitt. The behavior of ions near a charged wall-dependence on ion size, concentration, and surface charge. _The journal of physical chemistry B_ , 114(18):6074–6083, 2010. * Misra and Blankschtein [2021a] Rahul Prasanna Misra and Daniel Blankschtein. Ion adsorption at solid/water interfaces: Establishing the coupled nature of ion–solid and water–solid interactions. _The Journal of Physical Chemistry C_ , 125(4):2666–2679, 2021a. * Finney et al. [2021] Aaron R Finney, Ian J McPherson, Patrick R Unwin, and Matteo Salvalaglio. Electrochemistry, ion adsorption and dynamics in the double layer: a study of nacl (aq) on graphite. _Chemical science_ , 12(33):11166–11180, 2021\. * Perego et al. [2015] C. Perego, M. Salvalaglio, and M. Parrinello. Molecular dynamics simulations of solutions at constant chemical potential. _The Journal of Chemical Physics_ , 142(14):144113, apr 2015. doi: 10.1063/1.4917200. * Doblhoff-Dier and Koper [2021] K. Doblhoff-Dier and M. T. M. Koper. Modeling the Gouy–Chapman diffuse capacitance with attractive ion–surface interaction. _The Journal of Physical Chemistry C_ , 125(30):16664–16673, 2021. * France-Lanord and Grossman [2019] Arthur France-Lanord and Jeffrey C Grossman. Correlations from ion pairing and the nernst-einstein equation. _Physical review letters_ , 122(13):136001, 2019\. * Misra and Blankschtein [2021b] Rahul Prasanna Misra and Daniel Blankschtein. Uncovering a Universal Molecular Mechanism of Salt Ion Adsorption at Solid/Water Interfaces. _Langmuir_ , 37(2):722–733, January 2021b. * Williams et al. [2017] C. D. Williams, J. Dix, A. Troisi, and P. Carbone. Effective Polarization in Pairwise Potentials at the Graphene–Electrolyte Interface. _J. Phys. Chem. Lett._ , 8(3):703, February 2017\. * Siepmann and Sprik [1995] J Ilja Siepmann and Michiel Sprik. Influence of surface topology and electrostatic potential on water/electrode systems. _The Journal of chemical physics_ , 102(1):511, 1995. * Di Pasquale and Davidchack [2020] N. Di Pasquale and R. L. Davidchack. Shuttleworth Equation: A Molecular Simulations Perspective. _Journal of Chemical Physics_ , 153:154705, 2020. * Abraham et al. [2015] M. J. Abraham, T. Murtola, R. Schulz, S. Páll, J. C. Smith, B. Hess, and E. Lindahl. GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. _SoftwareX_ , 1:19–25, 2015. * Tribello et al. [2014] G. A. Tribello, M. Bonomi, D. Branduardi, C. Camilloni, and G. Bussi. Plumed 2: New feathers for an old bird. _Computer physics communications_ , 185(2):604–613, 2014. * Hourahine et al. [2020] B. Hourahine, B. Aradi, V. Blum, F. Bonafé, A. Buccheri, C. Camacho, C. Cevallos, M. Y. Deshaye, T. Dumitrică, A. Dominguez, S. Ehlert, M. Elstner, T. van der Heide, J. Hermann, S. Irle, J. J. Kranz, C. Köhler, T. Kowalczyk, T. Kubař, I. S. Lee, V. Lutsker, R. J. Maurer, S. K. Min, I. Mitchell, C. Negre, T. A. Niehaus, A. M. N. Niklasson, A. J. Page, A. Pecchia, G. Penazzi, M. P. Persson, J. Řezáč, C. G. Sánchez, M. Sternberg, M. Stöhr, F. Stuckenberg, A. Tkatchenko, V. W.-z. Yu, and T. Frauenheim. DFTB$+$, a software package for efficient approximate density functional theory based atomistic simulations. _The Journal of Chemical Physics_ , 152(12):124101, 2020. doi: 10.1063/1.5143190. * Elstner et al. [1998] Marcus Elstner, Dirk Porezag, G Jungnickel, J Elsner, M Haugk, Th Frauenheim, Sandor Suhai, and Gotthard Seifert. Self-consistent-charge density-functional tight-binding method for simulations of complex materials properties. _Physical Review B_ , 58(11):7260, 1998. * Elliott et al. [2022b] J. D. Elliott, M. Chiricotto, A. Troisi, and P. Carbone. Do specific ion effects influence the physical chemistry of aqueous graphene-based supercapacitors? perspectives from multiscale qmmd simulations. _arXiv_ , 2022b. doi: 10.48550/ARXIV.2203.02469. * Mulliken [1955] Robert S Mulliken. Electronic population analysis on lcao–mo molecular wave functions. i. _The Journal of Chemical Physics_ , 23(10):1833–1840, 1955. * Haynes et al. [2016] William M Haynes, David R Lide, and Thomas J Bruno. _CRC handbook of chemistry and physics_. CRC press, 2016. * Zeron et al. [2019] I. M. Zeron, J. L. F. Abascal, and C. Vega. A force field of Li+, Na+, K$+$, Mg2+, Ca2+, Cl-, and SO${}_{4}^{2-}$ in aqueous solution based on the TIP4P/2005 water model and scaled charges for the ions. _The Journal of chemical physics_ , 151(13):134504, 2019. * Ho and Striolo [2013] Tuan A Ho and Alberto Striolo. Capacitance enhancement via electrode patterning. _The Journal of Chemical Physics_ , 139(20):204708, 2013. * Xu et al. [2020] Kui Xu, Hui Shao, Zifeng Lin, Céline Merlet, Guang Feng, Jixin Zhu, and Patrice Simon. Computational insights into charge storage mechanisms of supercapacitors. _Energy & environmental materials_, 3(3):235–246, 2020. * Hagberg et al. [2008] Aric Hagberg, Pieter Swart, and Daniel S Chult. Exploring network structure, dynamics, and function using networkx. Technical report, Los Alamos National Lab.(LANL), Los Alamos, NM (United States), 2008. * Berendsen et al. [1995] H. J. C. Berendsen, D. van der Spoel, and R. van Drunen. Gromacs: a message-passing parallel molecular dynamics implementation. _Computer physics communications_ , 91(1-3):43–56, 1995. * van der Spoel et al. [2005] D. van der Spoel, E. Lindhal, B. Hess, G. Groenhof, A. E. Mark, and H. J. C. Berendsen. Gromacs: Fast, flexible and free. _J. Comput. Chem._ , 26(16):1701–1718, 2005. * Berendsen et al. [1987] H. J. C. Berendsen, J. R. Grigera, and T. P. Straatsma. The missing term in effective pair potentials. _Journal of Physical Chemistry_ , 91(24):6269–6271, 1987. * Miyamoto and Kollman [1992] S. Miyamoto and P. A. Kollman. SETTLE: An analytical version of the SHAKE and RATTLE algorithm for rigid water molecules. _J. Comput. Chem_ , 13:952–962, 1992. * Huang et al. [2012] Yi Huang, Jiajie Liang, and Yongsheng Chen. An overview of the applications of graphene-based materials in supercapacitors. _small_ , 8(12):1805–1834, 2012. * Joung and Cheatham [2008] I. S. Joung and T. E. III Cheatham. Determination of alkali and halide monovalent ion parameters for use in explicitly solvated biomolecular simulations. _The journal of physical chemistry B_ , 112(30):9020–9041, 2008. * Dočkal et al. [2022] J. Dočkal, M. Lísal, and F. Moučka. Molecular dynamics of the interfacial solution structure of alkali-halide electrolytes at graphene electrodes. _Journal of Molecular Liquids_ , 353:118776, 2022. * Savitzky and Golay [1964] Abraham Savitzky and Marcel JE Golay. Smoothing and differentiation of data by simplified least squares procedures. _Analytical chemistry_ , 36(8):1627–1639, 1964\. * Olivieri et al. [2021] Jean-François Olivieri, James T Hynes, and Damien Laage. Confined water’s dielectric constant reduction is due to the surrounding low dielectric media and not to interfacial molecular ordering. _The Journal of Physical Chemistry Letters_ , 12(17):4319–4326, 2021. * Kornyshev [2007] Alexei A Kornyshev. Double-layer in ionic liquids: paradigm change? _The Journal of Physical Chemistry B_ , 111(20):5545–5557, 2007. * Snook and van Megen [1981] Ian Snook and William van Megen. Finite ion size effects in the electrical double layer—a monte carlo study. _The Journal of Chemical Physics_ , 75(8):4104–4106, 1981. * Hwang et al. [2021] Hyerim Hwang, Yong Chan Cho, Sooheyong Lee, Yun-Hee Lee, Seongheun Kim, Yongjae Kim, Wonhyuk Jo, Patrick Duchstein, Dirk Zahn, and Geun Woo Lee. Hydration breaking and chemical ordering in a levitated NaCl solution droplet beyond the metastable zone width limit: evidence for the early stage of two-step nucleation. _Chemical Science_ , 12(1):179–187, 2021. doi: 10.1039/d0sc04817h. URL https://doi.org/10.1039%2Fd0sc04817h. * Finney and Salvalaglio [2022] Aaron R Finney and Matteo Salvalaglio. Multiple pathways in nacl homogeneous crystal nucleation. _Faraday Discussions_ , 2022. * Finney and Salvalaglio [2021] A. R. Finney and M. Salvalaglio. Bridging the gap between mesoscopic and molecular models of solid/liquid interfaces out-of-equilibrium. _arXiv preprint arXiv:2109.00568_ , 2021. * Yeh and Hummer [2004] I. Yeh and G. Hummer. System-Size Dependence of Diffusion Coefficients and Viscosities from Molecular Dynamics Simulations with Periodic Boundary Conditions. _The Journal of Physical Chemistry B_ , 108(40):15873–15879, October 2004. * Widodo et al. [2018] C. S. Widodo, H. Sela, and D. R. Santosa. The effect of nacl concentration on the ionic nacl solutions electrical impedance value using electrochemical impedance spectroscopy methods. In _AIP Conference Proceedings_ , volume 2021, page 050003. AIP Publishing LLC, 2018.
# Variational Loop Vertex Expansion Vasily<EMAIL_ADDRESS> Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France ###### Abstract Loop Vertex Expansion (LVE) was developed to construct QFT models with local and non-local interactions. Using LVE, one can prove the analyticity in the finite cardioid-like domain in the complex plain of the coupling constant of the free energies and cumulants of various vector, matrix, or tensor-type models. Here, applying the idea of choosing the initial approximation depending on the coupling constant, we construct the analytic continuation of the free energy of the quartic matrix model beyond the standard LVE cardioid over the branch cut and for arbitrary large couplings. ## 1 Introduction The present work stems from the constructive field theory method of the Loop Vertex Expansion (LVE) and ideas of the variational perturbation theory. The concept of LVE was first introduced in [1] as a constructive approach for quartic matrix models aimed to provide bounds that are uniform in the size of the matrix. In its original form, LVE combines an intermediate field representation with replica fields and a forest formula [2, 3] to express the free energy of the theory through a convergent sum over trees. Unlike conventional constructive methods, this loop vertex expansion does not rely on cluster expansions and does not entail conditions related to small/large field considerations. Like Feynman’s perturbative expansion, the LVE provides a straightforward way to calculate connected quantities. In this method, the theory’s partition function is represented as a sum over forests, and its logarithm is essentially the same sum but constrained to connected forests – trees. This property arises from the fact that the amplitudes factorize over the connected components of the forest. The functional integrands associated with each forest or tree exhibit absolute and _uniform_ convergence for all field values. Together with the non-proliferation of trees (in comparison to Feynman diagrams) this leads to the convergence of the LVE in the ’pacman’-like or cardioid-like domains, see Fig. 1. Figure 1: The left-hand side represents the typical LVE ’pacman’-type domain of analyticity with the radius $r_{\alpha}$ shrinking with decreasing of the angle $\alpha$. On the right, we present the cardioid domain, which can be understood as a union of ’pacman’-type domains. The convergence of the LVE implies the Borel summability of the standard perturbation series [4], and the LVE directly computes the Borel sum. Essentially, the loop vertex expansion performs an _explicit repacking_ of infinitely many subsets of Feynman amplitude components, leading to a convergent expansion rather than a divergent one [5]. In the context of combinatorial field theories involving matrices and tensors [6, 7], appropriately rescaled to possess a non-trivial $N\to\infty$ limit [8, 9, 10, 11], the Borel summability achieved by the LVE is _uniform_ with respect to the model’s size $N$, [1, 12, 13, 14]. The LVE method is extendable to ordinary field theories with cutoffs, as discussed in [15]. An adapted multiscale version, known as MLVE [16], incorporates renormalization techniques [17, 18, 19, 20, 21], although it’s worth noting that models developed so far are limited to the superrenormalizable type. The MLVE is particularly effective in resumming renormalized series for non-local field theories of matrix or tensorial types. Originally developed for the quartic interactions, the LVE method was then generalized to higher-order interactions under the name of the Loop Vertex Representation [22, 23, 24]. The main principle of the Variational Perturbation Theory (VPT) is the utilization of the initial approximation chosen depending on the coupling constant and/or on the order of the perturbative expansion, or optimized depending on these parameters. Such ideas of the VPT were applied under different names: Variational Perturbation Theory [25, 26], Optimized Perturbation Theory [27], Convergent Perturbation Theory [28], Delta Expansion [29], etc., – with the different levels of rigor to a various range of physical problems. The applications of VPT include: energy levels of quantum anharmonic oscillator [30, 31, 29]; critical indices in scalar quantum field theories [32]; optimization of the QCD perturbative computations [27, 33]; computations in lattice models with real and scalar actions [34, 35, 36]; coefficients of the $\frac{1}{D}$ expansion in the vector model [37]. VPT was used for constructing strong coupling expansions [38]. It was shown that VPT (Delta Expansion) should be applicable in cases where the standard perturbation theory is non-Borel summable [29]. It is worth noting that the proof of the convergence of the VPT (Delta Expansion) for the double-well case of the anharmonic oscillator is based on the analyticity of its energy levels [39] in the area of the coupling constant Riemann surface, which is larger than the one required for the Borel summability in the single-well potential case. In the current work, we unite the LVE method with the VPT idea of the initial approximation depending on the coupling constant. Taking the quartic matrix model as an example, we construct LVE with the modified initial approximation depending on the coupling constant and prove the convergence of the corresponding series for the free energy of the model for arbitrary large coupling constants with the arguments $-\frac{3\pi}{2}<\phi<\frac{3\pi}{2}$. The latter result is not optimal and might be improved up to $-2\pi+\epsilon<\phi<2\pi-\epsilon$, where $\epsilon\geq\epsilon_{0}$, $\epsilon_{0}>0$. This and the extension of the current results to cumulants and a constructive version of the $1/N$ expansion are additional outcomes of the method, left for further exploration. We also aim to investigate connections with the resurgence theory in the spirit of the recently carried out analysis of the vector model [40]. ## 2 Statement of the result To illustrate the main advantages of the Loop Vertex Expansion modified by the initial approximation depending on the coupling, we study a quartic matrix model. The partition function of this model is defined by $\displaystyle{\cal Z}[\lambda,N]=\frac{1}{Z_{0}}\int dM\exp\Big{\\{}-\operatorname{Tr}(MM^{\dagger})-\frac{\lambda}{2N}\operatorname{Tr}([MM^{\dagger}]^{2})\Big{\\}}\,,$ (1) where $M$ are complex $N\times N$ matrices, and $\displaystyle Z_{0}=\int dM\exp\Big{\\{}-\operatorname{Tr}(MM^{\dagger}))\Big{\\}}\,.$ (2) The measure $dM$ is given by $dM=\pi^{N}\prod_{1\leq i,j\leq N}d\text{Re}(M_{ij})d\text{Im}(M_{ij})\,.$ (3) The core object of our studies here is the free energy of the model defined, as $\displaystyle F[\lambda,N]=-\frac{1}{N^{2}}\log{\cal Z}[\lambda,N]\,.$ (4) As was mentioned in the introduction the standard LVE tools allow one to prove the analyticity of the free energy (4) or of the cumulants of the model in the ’pacman’-like or cardioid-like domains, as in Fig. 1. The sharpest result regarding the model (1) was obtained by LVE in [12], among other results it was shown there that the free energy (4) is analytic in the cardioid domain of the coupling constant $\lambda$, ${\cal C}=\Big{\\{}\lambda\in\mathbb{C}\,\Big{|}\arg\lambda=\phi\,,4|\lambda|<\cos^{2}\Big{(}{\frac{\phi}{2}}\Big{)}\Big{\\}}\,.$ (5) The main result of the current paper, obtained by merging the ideas of the variational perturbation theory and loop vertex expansion is given by the following theorem. ###### Theorem 1. For any $\lambda\in{\cal X}$, where ${\cal X}={\cal C}\cup{\cal Y}$, and ${\cal Y}$ is defined as a subset of the Riemann surface of $\sqrt{\lambda}$ with $\lambda\neq 0$ and $|\phi|<\frac{3\pi}{2}$, the free energy $F[\lambda,N]$ of the model (1), is analytic in $\lambda$ uniformly in $N$. ## 3 New initial approximation and intermediate field representation Hereafter we assume that $\lambda\neq 0$. The case of $\lambda=0$ is trivial and the vicinity of $\lambda=0$ can be easily treated with the standard LVE tools [12]. As a first step of our program, we change the Gaussian part of the action which we treat as an unperturbed measure, in other words, we shift the initial approximation. Then, introducing the intermediate field representation, we obtain, $\displaystyle{\cal Z}[\lambda,N]=\frac{1}{Z_{0}}\int dM\int dA\,\exp\Big{\\{}-\frac{1}{2}\operatorname{Tr}(A^{2})-a\operatorname{Tr}(MM^{\dagger})$ $\displaystyle+\mathrm{i}\operatorname{Tr}\Big{(}A\big{[}\sqrt{\frac{\lambda}{N}}MM^{\dagger}+\frac{(1-a)\sqrt{N}}{\sqrt{\lambda}}\mathbb{1}\big{]}\Big{)}+\operatorname{Tr}\big{[}\frac{(1-a)^{2}N}{2\lambda}\mathbb{1}\big{]}\Big{\\}}\,,$ (6) where $Re\,a>0$, $A$ field is realised by the $N\times N$ Hermitian matrices, and the integral over $A$ is assumed to be normalized. The integral over the initial degrees of freedom, matrices $M$ and $M^{\dagger}$, is Gaussian, so it can be evaluated. The corresponding covariance is given by $\big{(}a-\mathrm{i}\sqrt{\frac{\lambda}{N}}A\big{)}\otimes 1$, and using that $\displaystyle\det\left[\Big{(}a-\mathrm{i}\sqrt{\frac{\lambda}{N}}A\Big{)}\otimes 1\right]=a^{N}\exp\bigg{\\{}N\operatorname{Tr}\log\Big{(}1-\mathrm{i}\sqrt{\frac{\lambda}{a^{2}N}}A\Big{)}\bigg{\\}}\,,$ (7) we obtain $\displaystyle{\cal Z}[\lambda,N]$ $\displaystyle=$ $\displaystyle\frac{e^{\frac{N^{2}(1-a)^{2}}{2\lambda}}a^{N}}{\widetilde{Z}_{0}}\int dA\,\exp\bigg{\\{}-\frac{1}{2}\operatorname{Tr}(A^{2})$ (8) $\displaystyle-$ $\displaystyle N\operatorname{Tr}\log\left(1-\mathrm{i}\sqrt{\frac{\lambda}{a^{2}N}}A\right)-\mathrm{i}\operatorname{Tr}\bigg{(}A\frac{(1-a)\sqrt{N}}{\sqrt{\lambda}}\bigg{)}\bigg{\\}}\,.$ Then, we define the non-polynomial interaction, as $\displaystyle{\cal S}[\lambda,N,a](A)=\operatorname{Tr}\log\left(1-\mathrm{i}\sqrt{\frac{\lambda}{a^{2}N}}A\right)+\frac{\mathrm{i}}{\sqrt{N}}\operatorname{Tr}\bigg{(}A\frac{(1-a)}{\sqrt{\lambda}}\bigg{)}\,,$ (9) and the normalization, as $\displaystyle K[\lambda,N,a]=\frac{e^{\frac{N^{2}(1-a)^{2}}{2\lambda}}a^{N}}{\widetilde{Z}_{0}}\,.$ (10) Thus, the partition function can be written in the following form $\displaystyle{\cal Z}[\lambda,N]=K[\lambda,N,a]\int dA\,\exp\bigg{\\{}-\frac{1}{2}\operatorname{Tr}(A^{2})-N{\cal S}[\lambda,N,a](A)\bigg{\\}}\,.$ (11) Let us study the analytic properties of the partition function. We first have the following bound. ###### Lemma 1. Let $\frac{\lambda}{a^{2}}=\rho\mathrm{e}^{\mathrm{i}\theta}$ with $\rho>0$, then: $\Big{\|}\Big{(}1-\mathrm{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A\Big{)}^{-1}\Big{\|}\leq\frac{1}{\cos\frac{\theta}{2}}\,,$ (12) where $\Big{\|}\cdot\Big{\|}$ stands for the operator norm. ###### Proof. The first step is to rewrite the resolvent, as an integral, $\Big{(}1-\mathrm{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A\Big{)}^{-1}=\frac{a\sqrt{N}}{\sqrt{\lambda}}\int_{0}^{\infty}d\alpha\,\exp\Big{\\{}-\alpha\frac{a\sqrt{N}}{\sqrt{\lambda}}+\alpha\mathrm{i}A\Big{\\}}\,.$ (13) Using the latter representation, we immediately arrive at the bound for the operator norm, $\displaystyle\Big{\|}\Big{(}1-\mathrm{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A\Big{)}^{-1}\Big{\|}$ $\displaystyle\leq$ $\displaystyle\frac{|a|\sqrt{N}}{|\sqrt{\lambda}|}\int_{0}^{\infty}\exp\Big{\\{}-\alpha\text{Re}\big{(}\frac{a\sqrt{N}}{\sqrt{\lambda}}\big{)}\Big{\\}}\Big{\|}\exp\Big{\\{}\mathrm{i}\alpha A\Big{\\}}\Big{\|}$ (14) $\displaystyle=$ $\displaystyle\frac{1}{\cos\frac{\theta}{2}}\,.$ ∎ Hereafter, we write the parameter $a$ describing the initial approximation in the following form, $\displaystyle a=x|\sqrt{\lambda}|e^{\mathrm{i}\psi}\,,\qquad-\frac{\pi}{2}<\psi<\frac{\pi}{2}\,,\qquad x>0\,.$ (15) Writing (11), as ${\cal Z}[\lambda,N]=K[\lambda,N,a]\int dA\,\frac{\exp\bigg{\\{}-\frac{1}{2}\operatorname{Tr}(A^{2})-\mathrm{i}\operatorname{Tr}\bigg{(}A\frac{(1-a)\sqrt{N}}{\sqrt{\lambda}}\bigg{)}\bigg{\\}}}{\Big{[}\det\Big{(}1-\mathrm{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A\Big{)}\Big{]}^{N}}\,,$ (16) and using lemma 1, one can show that this integral is convergent for $\theta\in(-\pi,\pi)$, or equivalently for $\lambda/a^{2}\in{\mathbb{C}}-{\mathbb{R}}^{-}$. Since the integrand is analytic for $\theta\in(-\pi,\pi)$, and $\theta=\phi-2\psi$, we have the following result. ###### Proposition 1. ${\cal Z}[N,\lambda]$ is analytic in $\lambda$ on the Riemann surface of the square root with $-2\pi<\phi<2\pi$. Note, that since ${\cal Z}[N,\lambda]$ may be zero for certain values of $\lambda$, its analyticity does not imply the analyticity of the free energy. ## 4 The Loop Vertex Expansion Using the representation (11) for the partition function, we expand the interaction part of the exponent into the Taylor series, which is convergent if $\lambda/a^{2}\in{\mathbb{C}}-{\mathbb{R}}^{-}$, $\displaystyle{\cal Z}=K[\lambda,N,a]\sum_{n=0}^{\infty}\frac{(-1)^{n}}{n!}\int d\mu(A)\,\bigg{[}N{\cal S}[\lambda,N,a](A)\bigg{]}^{n}\,,$ (17) where $d\mu(A)=dA\exp\\{-\frac{1}{2}\operatorname{Tr}(A^{2})\\}$ is the normalized Gaussian measure for the Hermitian matrices $A$ (and we dropped the arguments $\lambda$ and $N$ of ${\cal Z}$, and ${\cal S}[\lambda,N,a](A)$ we will write ${\cal S}(A)$ in order to simplify the notation). Applying the replica trick, we replace (at each order $n$) the integral over a single matrix $A$ by the integral over a $n$-copies of $N\times N$ Hermitian matrices $A=(A_{i})_{1\leq i\leq n}$. The measure of the Gaussian integral over the replicated matrices $A$ is normalized with a degenerated covariance $C_{ij}=1$. For any real positive symmetric matrix $C_{ij}$ the Gaussian integral obeys $\int d\mu_{C}(A)\,A_{i|ab}A_{j|cd}=C_{ij}\,\delta_{ad}\delta_{bc}\,,~{}~{}~{}\int d\mu_{C}(A)=1\,,$ (18) where $A_{i|ab}$ corresponds to the matrix element of $A_{i}$ in the row $a$ and column $b$. The degenerated covariance in the Gaussian integral is equivalent to insertion $n-1$ Dirac $\delta$-functions $\delta(A_{1}-A_{2})\cdots\delta(A_{n-1}-A_{n})$. In Feynman diagrams, the uniform covariance connects the various replicas together (with the appropriate weights). Then, the partition function is given by $\displaystyle{\cal Z}=K[\lambda,N,a]\sum_{n=0}^{\infty}\frac{(-1)^{n}}{n!}\int d\mu_{C}(A)\prod_{i=1}^{n}\bigg{[}N{\cal S}(A_{i})\bigg{]}\,.$ (19) To take the logarithm of the partition function, we are going to convert the latter expansion into the sum over forests by applying the Bridges-Kennedy- Abdessalam-Rivasseau [2, 3]. The first step to do it is to replace the covariance $C_{ij}=1$ by $C_{ij}(x)=x_{ij}$, $x_{ij}=x_{ji}$ (which at the end should be evaluated at $x_{ij}=1$) for $i\neq j$ and $C_{ii}(x)=1$. Then, $\displaystyle{\cal Z}$ $\displaystyle=$ $\displaystyle K[\lambda,N,a]\sum_{F\,\text{labeled forest}}\frac{(-1)^{n}}{n!}\int_{0}^{1}\prod_{(i,j)\in F}dt_{ij}\,\,\left(\prod_{(i,j)\in F}\frac{\partial}{\partial x_{ij}}\right)$ $\displaystyle\times$ $\displaystyle\bigg{\\{}\int d\mu_{C(x)}(A)\prod_{i=1}^{n}\bigg{[}N{\cal S}(A_{i})\bigg{]}\bigg{\\}}\bigg{|}_{x_{ij}=v^{F}_{ij}}\,,$ where $n$ is the number of vertices of the forest $F$, $i$ and $j$ label the forest’s vertices, and there is a weakening parameter $t_{ij}$ per each edge $(i,j)$ of the forest, and $v^{F}_{ij}=\left\\{\begin{array}[]{ccl}\inf_{(k,l)\in{P}_{i\leftrightarrow j}^{{F}}}t_{kl}&\text{if}&{P}_{i\leftrightarrow j}^{{F}}\,\text{exists}\\\ 0&\text{if}&{P}_{i\leftrightarrow j}^{{F}}\,\text{does not exist}\end{array}\right.\,,$ (21) ${P}_{i\leftrightarrow j}^{{F}}$ is the unique path in the forest $F$ joining $i$ and $j$ (the infimum is taken to be $1$ if $i=j$). Applying the following lemma, [12], we can take the logarithm. ###### Lemma 2. Let ${\cal W}(T)$ be the weight of a tree $T$, not depending on the labels of the tree vertices and the weight of a forest ${\cal W}(F)$ is defined to be the product of the weights of its trees. Then, in the formal series sense, we have $\log\sum_{F\text{ labeled forests}}\frac{{\cal W}(F)}{|V(F)|!}=\sum_{T\text{ labeled trees}}\frac{{\cal W}(T)}{|V(T)|!}\,,$ (22) where $|V(F)|$ and $|V(T)|$ are the number of vertices in $F$ and $T$ correspondingly. Since the differentiation with respect to $x_{ij}$ and the Gaussian integration factor over the trees in the forest $F$, we obtain $\displaystyle\log{\cal Z}$ $\displaystyle=$ $\displaystyle\log K[\lambda,N,a]+\sum_{T\,\text{labeled trees}}\frac{(-1)^{n}}{n!}\int_{0}^{1}\prod_{(i,j)\in T}dt_{ij}\,\left(\prod_{(i,j)\in T}\frac{\partial}{\partial x_{ij}}\right)$ (23) $\displaystyle\times$ $\displaystyle\bigg{\\{}\int d\mu_{C(x)}(A)\prod_{i=1}^{n}\bigg{[}N{\cal S}(A_{i})\bigg{]}\bigg{\\}}\bigg{|}_{v^{T}_{ij}}\,,$ (24) $\displaystyle v^{T}_{ij}$ $\displaystyle=$ $\displaystyle\inf_{(k,l)\in{P}_{i\leftrightarrow j}^{T}}t_{kl}\,.$ (25) where ${P}_{i\leftrightarrow j}^{T}$ stands for the unique path joining vertices $i$ and $j$ in the tree $T$. Expressing the Gaussian integral as a differential operator, $\displaystyle\int d\mu_{C(x)}(A)F(A)=\left[e^{\frac{1}{2}\sum_{i,j}x_{ij}\operatorname{Tr}\left[\frac{\partial}{\partial A_{i}}\frac{\partial}{\partial A_{j}}\right]}F(A)\right]_{A_{i}=0}\,,$ (26) we see that $\frac{\partial}{\partial x_{ij}}\bigg{(}\int d\mu_{C(x)}(A)F(A)\bigg{)}=\frac{1}{2}\int d\mu_{C(x)}(A)\,\operatorname{Tr}\left[\frac{\partial}{\partial A_{i}}\frac{\partial}{\partial A_{j}}\right]F(A)\,.$ (27) The latter differential operator acts on $i$ and $j$ vertices and connects them by an edge. The first derivative of the loop vertex (non-polynomial action ${\cal S}(A_{i})$) is given by $\displaystyle\frac{\partial}{\partial A_{i|cd}}\bigg{[}\operatorname{Tr}\log\left(1-\mathrm{i}\sqrt{\frac{\lambda}{a^{2}N}}A\right)+\frac{\mathrm{i}}{\sqrt{N}}\operatorname{Tr}\bigg{(}A\frac{(1-a)}{\sqrt{\lambda}}\bigg{)}\bigg{]}=$ $\displaystyle\frac{\sqrt{\lambda}}{a\sqrt{N}}\bigg{(}1-\text{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A_{i}\Big{)}_{cd}^{-1}+\mathrm{i}\frac{(1-a)}{\sqrt{\lambda}\sqrt{N}}\mathbb{1}_{cd}\,.$ (28) Only the first term of (28) is relevant for applying all further derivatives. Therefore all other derivatives can be computed using the following recursive relation, $\frac{\partial}{\partial A_{i|ab}}\frac{\sqrt{\lambda}}{a\sqrt{N}}\Big{(}1-\text{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A_{i}\Big{)}_{cd}^{-1}=\mathrm{i}\frac{\lambda}{a^{2}N}\Big{(}1-\text{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A_{i}\Big{)}_{ca}^{-1}\Big{(}1-\text{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A_{i}\Big{)}_{bd}^{-1}\,.$ (29) Let $V(T)$ be the set of all vertices of the tree, and $E(T)$ be the set of edges of the tree. We observe that in each tree $T$ with $|V(T)|>2$ there are two types of vertices: internal vertices and leaves. If the vertex is differentiated only once, it is a leaf, and it brings the contribution of the form of (28). We say that leaf vertices have only one corner. Multiple derivatives acting on the same vertex (corresponding to multiple edges hooked to it) can act on either of the corners of the vertex, splitting it into two corners and eliminating the constant term of (28), if it was present, see Fig. 2. Figure 2: From left to right: a vertex without corners – corresponds to a trivial tree, a vertex with only one corner – a leaf, vertex with two corners. Each derivative brings an additional factor of $\frac{1}{\sqrt{N}}$, and consequently, each edge of any tree comes with the factor $\frac{1}{N}$. In the following we attribute factors $\frac{1}{N}$ to the edges, and define a corner operator as ${\cal C}=\begin{cases}\frac{\sqrt{\lambda}}{a}\Big{(}1-\text{i}\sqrt{\frac{\lambda}{a^{2}N}}A_{i}\Big{)}_{cd}^{-1}+\mathrm{i}\frac{(1-a)}{\sqrt{\lambda}}\mathbb{1}_{cd}\,,\text{ only one corner in the vertex}\\\ ~{}\\\ \frac{\sqrt{\lambda}}{a}\Big{(}1-\text{i}\sqrt{\frac{\lambda}{a^{2}N}}A_{i}\Big{)}_{cd}^{-1}\,.\text{ if there are more corners}\end{cases}\,,$ (30) Therefore, the logarithm of ${\cal Z}$ can be expressed, as $\displaystyle\log{\cal Z}=$ $\displaystyle\log K[\lambda,N,a]+\sum_{T\,\atop\text{LVE tree}}{\cal A}_{T}[\lambda,N]\,,$ (31) $\displaystyle{\cal A}_{T}[\lambda,N]=$ $\displaystyle\frac{N^{|V(T)|-|E(T)|}}{|V(T)|!}\int_{0}^{1}\prod_{e\in E(T)}dt_{e}\,$ (32) $\displaystyle\times\int d\mu_{C_{T}}(A)\,\operatorname{Tr}\Big{[}\mathop{\overrightarrow{\prod}}\limits_{c\in\partial T\,\text{corner}}{\cal C}_{c}(i_{c})\Big{]}\,,$ (33) where $i_{c}$ labels the vertex, to which the corner $c$ is attached, the covariance $C_{T}$ is $(C_{T})_{ij}=\inf_{(k,l)\in P^{T}_{i\leftrightarrow j}}t_{kl}$ (34) and the infimum is taken to be $1$ if $i=j$. ## 5 Bounds The expansion (33) contains three types of contributions, which should be considered separately: a trivial tree with only one vertex, a tree with two vertices, and all other trees. In the following we treat these cases step by step. ### 5.1 Trivial tree As always in the LVE formalism there is a trivial tree with only one vertex, see Fig. 2. It requires a special treatment. The contribution of the trivial tree is given by $\displaystyle{\cal A}_{T_{1}}=\int d\mu_{C(x)}(A)\bigg{[}N\operatorname{Tr}\log\left(1-\mathrm{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A\right)+\mathrm{i}\operatorname{Tr}\bigg{(}A\frac{(1-a)\sqrt{N}}{\sqrt{\lambda}}\bigg{)}\bigg{]}\,.$ (35) The second term gives zero after the integration and the first can be bounded by integrating by parts as $\displaystyle{\cal A}_{T_{1}}$ $\displaystyle=$ $\displaystyle N\int dA\,e^{-\frac{1}{2}\operatorname{Tr}\big{[}A^{2}\big{]}}\operatorname{Tr}_{i=j}\log(\mathbb{1}-\text{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A_{ij})$ (36) $\displaystyle=$ $\displaystyle N\int dA\,e^{-\frac{1}{2}\operatorname{Tr}\big{[}A^{2}\big{]}}\operatorname{Tr}_{i=j}\int_{0}^{1}dt\,\frac{-\text{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}tA_{ik}}{(\mathbb{1}-\text{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A)_{kj}}$ $\displaystyle=$ $\displaystyle\int d\mu(A)\,\operatorname{Tr}_{i=j}\int_{0}^{1}dt\,\frac{-\frac{\lambda}{a^{2}}t}{(\mathbb{1}-\text{i}\frac{\sqrt{\lambda}}{a\sqrt{N}}A)^{2}_{ij}}$ Since the integrand is analytic for $\theta\in(-\pi,\pi)$, and $\theta=\phi-2\psi$, employing (15), we arrive at the following conclusion. ###### Lemma 3. The amplitude of the trivial tree, ${\cal A}_{T_{1}}$, is analytic in $\lambda$ and bounded on the Riemann surface of the square root with $-2\pi<\phi<2\pi$. ### 5.2 Two-vertex tree To derive the bound for the tree with two vertices, see Fig. 3, we start with a general bound for the corner operators that are also useful for all other trees. Using the lemma 1, we obtain $\|{\cal C}\|\leq\begin{cases}\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}+\Bigg{|}\frac{(1-a)}{\sqrt{\lambda}}\Bigg{|}\,,\text{ if there is only one corner in the vertex}\\\ ~{}\\\ \Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}\,,\text{ if there are more corners}\end{cases}\,.$ (37) Figure 3: Two-vertex tree. Remembering the representation (15), for sufficiently large $x$, we can simplify the one-corner bound in (37), requiring that $\displaystyle\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}+\Bigg{|}\frac{(1-a)}{\sqrt{\lambda}}\Bigg{|}\leq\Bigg{|}\frac{3(1-a)}{2\sqrt{\lambda}}\Bigg{|}\,.$ (38) It is not hard to see that the latter inequality is valid when $\displaystyle x\geq x_{1}\,,\qquad x_{1}=\frac{\cos\frac{\theta}{2}+\sqrt{\cos^{2}\frac{\theta}{2}+8|\lambda|\cos\frac{\theta}{2}}}{2|\sqrt{\lambda}|\cos\frac{\theta}{2}}\,.$ (39) Note, as $-\pi<\theta<\pi$, we have that $\cos\frac{\theta}{2}>0$. In the two-vertex tree, both vertices have only one corner, therefore, using (38), taking into account that there are $2$ vertices, $1$ edge, and one factor $N$ coming from the trace, we can bound the amplitude of this tree as $\displaystyle|{\cal A}_{T_{2}}|\leq\frac{9N^{2}}{8}\int d\mu_{C(x)}(A)\Bigg{|}\frac{(1-a)}{\sqrt{\lambda}}\Bigg{|}^{2}\,.$ (40) Due to the analyticity of the integrand for $\theta\in(-\pi,\pi)$ and the latter bound, we arrive at the following. ###### Lemma 4. The amplitude of the trivial tree, ${\cal A}_{T_{2}}$, is analytic in $\lambda$ and bounded on the Riemann surface of the square root with $\lambda\neq 0$, $-2\pi<\phi<2\pi$. ### 5.3 Other trees Obviously, the bound for the leaves is larger than for the internal vertices. If we use it for all vertices, we can bound the amplitude of each tree, as $\displaystyle\bigg{|}\operatorname{Tr}\Big{[}\mathop{\overrightarrow{\prod}}\limits_{c\in\partial T\,\text{corner}}{\cal C}_{c}(i_{c})\Big{]}\bigg{|}$ $\displaystyle\leq N\Bigg{(}\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}+\Bigg{|}\frac{(1-a)}{\sqrt{\lambda}}\Bigg{|}\Bigg{)}^{2|E(T)|}\,.$ (41) However, the best results given by this bound are achieved when one takes $a=1$, which corresponds to classic LVE results. It happens since it is impossible to make simultaneously small both terms: $\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}\text{ and }\Bigg{|}\frac{(1-a)}{\sqrt{\lambda}}\Bigg{|}$. To overcome this difficulty, we observe that: * • We can always make the term $\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}$ as small as needed, by varying the ratio $\frac{\sqrt{\lambda}}{a}$ (considering large $x$). * • The factor $\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}+\Bigg{|}\frac{(1-a)}{\sqrt{\lambda}}\Bigg{|}$, for which hereafter we will use the bound (38), comes only from the leave vertices, and in a general tree there are not so many leaves, see Fig. 4. Figure 4: 1) A tree with a maximal amount of leaves at the given order of the loop vertex expansion. 2) A tree with a minimal amount of leaves at the given order of the LVE. 3) Representation of an average LVE tree with not so many leaves. Being inspired by the latter observations, we first establish the bound for the corner operators of a general tree. ###### Lemma 5. If a tree has $n_{l}$ leaves, $n_{i}$ internal vertices, and $n_{l}+n_{i}=|V(T)|>2$, its amplitude given by a trace of the product of the corner operators is bounded by $\displaystyle\bigg{|}\operatorname{Tr}\Big{[}\mathop{\overrightarrow{\prod}}\limits_{c\in\partial T_{(n_{i},n_{l})}\,\text{corner}}{\cal C}_{c}(i_{c})\Big{]}\bigg{|}$ $\displaystyle\leq N\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}^{2n_{i}-2}\Bigg{|}\frac{3(1-a)}{2a\cos\frac{\theta}{2}}\Bigg{|}^{n_{l}}\,.$ (42) ###### Proof. We prove (42) by induction. For the base, it is enough to check (42) for the tree with $n_{i}=1$ and $n_{l}=2$. According to (38), two leaves bring the factor $\displaystyle\Bigg{|}\frac{3(1-a)}{2\sqrt{\lambda}}\Bigg{|}^{2}\,,$ (43) and two corner operators of the internal vertex give $\displaystyle\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}^{2}\,.$ (44) All together, taking into account the factor $N$ coming from the trace, we have exactly (42) for $n_{i}=1$ and $n_{l}=2$. Then, for the induction, we assume that (42) is valid for all trees with $n_{v}$ vertices, $n_{l}+n_{i}=n_{v}$, and prove that it is then valid for all trees with $(n_{v}+1)$ vertices. The trees with $(n_{v}+1)$ vertices can be obtained from the trees with $n_{v}$ vertices by increasing the number of leaves or number of internal vertices. In the first case, see Fig. 5, the bound (42) gets an extra factor $\displaystyle\Bigg{|}\frac{3(1-a)}{2\sqrt{\lambda}}\Bigg{|}\,,$ (45) from the leaf, and an additional factor $\displaystyle\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}$ (46) from the new corner of the internal vertex, where the new leaf is attached. Altogether, this gives (42) with the number of internal $n_{i}$ and number of leaves $(n_{l}+1)$. Figure 5: Possible ways to add a leaf to a tree. The second case, see Fig. 6, when the number of vertices is augmented by increasing the number of the internal vertices, can be interpreted as attaching a new leaf to the already existing leaf – since the only way to obtain an additional internal vertex is to convert a leaf into it. The contribution of the original leaf will be preserved – by shifting to the contribution of the new leaf. In addition, there will be a contribution from two corner operators of the new internal vertex, the same as (44). All together this gives (42) with the number of internal $(n_{i}+1)$ and number of leaves $n_{l}$. Figure 6: Possible ways to add an internal vertex to a tree. ∎ To prove the convergence of the series (33), we split the sum over the LVE trees as $\displaystyle\sum_{T\,\atop\text{LVE tree}}=\sum_{2<|V(T)|<60}+\sum_{T_{\geq}}+\sum_{T_{<}}\,,$ (47) where the first sum runs over all LVE trees with $2<|V(T)|<60$ number of vertices, $T_{\geq}$ denotes the LVE trees which $\alpha|V(T)|$ leaves or more and $|V(T)|\geq 60$ vertices, and $T_{<}$ stands for the LVE trees with less than $\alpha|V(T)|$ leaves and $|V(T)|\geq 60$ vertices. Let us now bound each sum of (47) separately. We start with the following lemma. ###### Lemma 6. For any $x$, satisfying $\displaystyle x\geq x_{2}\,,\qquad x_{2}=\frac{|\sqrt{\lambda}|+1}{|\sqrt{\lambda}|}\,,$ (48) it is easy to see that $\displaystyle\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}\leq\Bigg{|}\frac{(1-a)}{a\cos\frac{\theta}{2}}\Bigg{|}\leq\frac{\sqrt{2}}{\cos\frac{\theta}{2}}\,,$ (49) and obviously $\displaystyle\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}\leq\Bigg{|}\frac{3(1-a)}{2a\cos\frac{\theta}{2}}\Bigg{|}\leq\frac{3\sqrt{2}}{2\cos\frac{\theta}{2}}\,.$ (50) ###### Proof. Recall that $\sqrt{\lambda}=e^{\text{i}\phi/2}|\sqrt{\lambda}|$, $\lambda\neq 0$, and that according to the representation (15), $a=x|\sqrt{\lambda}|e^{\mathrm{i}\psi}$, $x>0$, $-\frac{\pi}{2}<\psi<\frac{\pi}{2}$. Then, the first inequality in (50) transforms to $\displaystyle\frac{1}{x}\leq\Bigg{|}\frac{1-|\sqrt{\lambda}|xe^{\mathrm{i}\psi}}{|\sqrt{\lambda}|x}\Bigg{|}\,,$ (51) and it is enough to satisfy it for $\psi=0$. Simplifying the absolute value operation for $x\geq\frac{1}{|\sqrt{\lambda}|}$, we arrive at (48). The second inequality in (50) rewrites, as $\displaystyle\Bigg{|}\frac{(1-|\sqrt{\lambda}|xe^{\mathrm{i}\psi})}{|\sqrt{\lambda}|x\cos\frac{\theta}{2}}\Bigg{|}\leq\frac{\sqrt{2}}{\cos\frac{\theta}{2}}\,.$ (52) Using the upper bound for the left-hand side of (52), we find that it is enough to satisfy $\displaystyle\Bigg{|}\frac{\sqrt{1+|\lambda|x^{2}}}{|\sqrt{\lambda}|x}\Bigg{|}\leq\sqrt{2}\,,$ (53) what can be achieved for $x\geq\frac{1}{|\sqrt{\lambda}|}$. ∎ ###### Lemma 7. For $-\pi<\theta<\pi$, the sum of absolute values of trees amplitudes from the finite set of trees in (47) is bounded by a constant, $\displaystyle\sum_{2<|V(T)|<30}|{\cal A}_{T}|<const\,.$ (54) ###### Proof. Follows from lemmas 5 and 6. ∎ Now we can estimate how many trees have more than $\alpha|V(T)|$ leaves (we will be interested in $1>\alpha>1/2$). ###### Lemma 8. The number of trees with $|V(T)|$ vertices and $\alpha|V(T)|$ or more leaves with $1>\alpha>1/2$, $|T_{\geq}|$ is bounded by $\displaystyle|T_{\geq}|$ $\displaystyle\leq$ $\displaystyle(|V(T)|-\left\lceil\alpha|V(T)|\right\rceil+1)\big{(}|V(T)|!\big{)}$ (55) $\displaystyle\times$ $\displaystyle 2^{|V(T)|-3}e^{\left\lceil\alpha|V(T)|\right\rceil}\bigg{(}\frac{1-\alpha}{\alpha}\bigg{)}^{\left\lceil\alpha|V(T)|\right\rceil}\,.$ ###### Proof. The number of labeled trees with $n$ vertices and $k$ leaves is given by $\displaystyle{\cal N}(n,k)=\frac{n!}{k!}S(n-2,n-k)\,,$ (56) where $S(n-2,n-k)$ is the Stirling number of the second type. We can bound ${\cal N}(n,k)$ using an upper bound for the Stirling’s numbers of the second type, $\displaystyle S(n-2,n-k)\leq\binom{n-3}{n-k-1}(n-k)^{k}\leq 2^{n-3}(n-k)^{k}\,,$ (57) and a lower bound for the Gamma function $\displaystyle k!\geq e^{-k}(k+1)^{k}\,.$ (58) Thus, we have $\displaystyle{\cal N}(n,k)\leq n!\frac{e^{k}}{k^{k}}2^{n-3}(n-k)^{k}\,.$ (59) Applying it to $n=|V(T)|$, $k=\left\lceil\alpha|V(T)|\right\rceil$ – an integer number larger or equal to $\alpha|V(T)|$, we obtain $\displaystyle{\cal N}(|V(T)|,\left\lceil\alpha|V(T)|\right\rceil)\leq\big{(}|V(T)|!\big{)}2^{|V(T)|-3}e^{\left\lceil\alpha|V(T)|\right\rceil}\bigg{(}\frac{1-\alpha}{\alpha}\bigg{)}^{\left\lceil\alpha|V(T)|\right\rceil}\,,$ (60) where we have used an obvious relation $\displaystyle\frac{|V(T)|-\left\lceil\alpha|V(T)|\right\rceil}{\left\lceil\alpha|V(T)|\right\rceil}\leq\frac{1-\alpha}{\alpha}\,.$ (61) If we take $\frac{e}{1+e}<\alpha<1$, then $e\frac{1-\alpha}{\alpha}<1$, and it decreases with increasing of $\alpha$ leading to decreasing of the bound (60). Therefore for large values of $\alpha$ (for larger amounts of leaves), we will have fewer trees. Consequently, we can bound the number of trees that have $\alpha|V(T)|$ leaves or more just by multiplying the (60) by $(|V(T)|-\left\lceil\alpha|V(T)|\right\rceil+1)$. ∎ Now we are ready to bound the sum of amplitudes of all trees $T_{\geq}$ with $\alpha|V(T)|$ leaves or more. ###### Lemma 9. For $-\frac{\pi}{2}<\theta<\frac{\pi}{2}$ and $\alpha=\frac{59}{60}$, the sum of absolute values of amplitudes of the trees $T_{\geq}$ is smaller than the sum of an absolutely convergent series, $\displaystyle\sum_{T_{\geq}}\big{|}{\cal A}_{T_{\geq}}\big{|}\leq\frac{N^{2}e}{472}\sum_{v=60}^{\infty}(\frac{1}{60}v+1)\,\Bigg{(}\frac{7}{8}\Bigg{)}^{v}\,.$ (62) ###### Proof. First of all, we chose $x\geq x_{3}$, $x_{3}=\max\\{x_{1},x_{2}\\}$. Then, according to lemmas 5 and 6, the contributions of the corner operators of each tree from $T_{\geq}$ can be bounded, as $\displaystyle\bigg{|}\operatorname{Tr}\Big{[}\mathop{\overrightarrow{\prod}}\limits_{c\in\partial T_{\geq}\,\text{corner}}{\cal C}_{c}(i_{c})\Big{]}\bigg{|}\leq N\Bigg{(}\frac{3\sqrt{2}}{2\cos\frac{\theta}{2}}\Bigg{)}^{2n_{i}+n_{l}-2}\leq N\Bigg{(}\frac{3\sqrt{2}}{2\cos\frac{\theta}{2}}\Bigg{)}^{2|V(T)|}\,.$ (63) Taking into account the number of trees in $T_{\geq}$ at each order of LVE by lemma 8, we obtain $\displaystyle\sum_{T_{\geq}}\big{|}{\cal A}_{T_{\geq}}\big{|}$ $\displaystyle\leq$ $\displaystyle\frac{N^{2}}{8}\sum_{v=60}^{\infty}(v-\left\lceil\alpha v\right\rceil+1)\,2^{v}e^{\left\lceil\alpha v\right\rceil}\bigg{|}\frac{1-\alpha}{\alpha}\bigg{|}^{\left\lceil\alpha v\right\rceil}\Bigg{(}\frac{3\sqrt{2}}{2\cos\frac{\theta}{2}}\Bigg{)}^{2v}$ (64) $\displaystyle\leq$ $\displaystyle\frac{N^{2}}{8}\sum_{v=60}^{\infty}(v-\alpha v+1)\,e^{\alpha v+1}\bigg{|}\frac{1-\alpha}{\alpha}\bigg{|}^{\alpha v+1}\Bigg{(}\frac{3}{\cos\frac{\theta}{2}}\Bigg{)}^{2v}\,.$ Now let us take $\alpha$ in such a way that $\displaystyle\Bigg{(}\frac{3}{\cos\frac{\theta}{2}}\Bigg{)}^{2}\,e^{\alpha}\Bigg{|}\frac{1-\alpha}{\alpha}\Bigg{|}^{\alpha}<1\,.$ (65) It is easy to see that, for any value of the $\cos\frac{\theta}{2}$, one can choose such $\alpha$’s, sufficiently close to $\alpha=1$, that the inequality (65) is satisfied. However, to handle our bounds we need to fix a certain constant value of $\alpha$. For this, we need to additionally restrict values of $\theta$. For simplicity, hereafter we consider $-\frac{\pi}{2}<\theta<\frac{\pi}{2}$. Then, it is enough to find $\alpha$’s satisfying $\displaystyle 18\,e^{\alpha}\Bigg{|}\frac{1-\alpha}{\alpha}\Bigg{|}^{\alpha}<1\,.$ (66) If, for instance, we take $\alpha=\frac{59}{60}$, the left-hand side of (66) $\approx 0.873<\frac{7}{8}$ (note that $\frac{59}{60}>\frac{e}{1+e}$). This completes the proof. ∎ ###### Lemma 10. For $-\frac{\pi}{2}<\theta<\frac{\pi}{2}$ and $\alpha=\frac{59}{60}$, the sum of absolute values of amplitudes of the trees $T_{<}$ is smaller than the sum of an absolutely convergent series, $\displaystyle\sum_{T_{<}}\big{|}{\cal A}_{T_{<}}\big{|}$ $\displaystyle\leq$ $\displaystyle N^{2}\frac{2(x_{4}\cos\frac{\theta}{2})^{5}\sqrt{\lambda}}{3(x_{4}\sqrt{\lambda}-1)}\sum_{v=60}^{\infty}\frac{1}{v^{2}}\Bigg{(}\frac{1}{2}\Bigg{)}^{v}\,,$ (67) where $\displaystyle x_{4}=\max\\{x3,\frac{2^{30}e^{30}}{\cos\frac{\theta}{2}}\Big{(}\frac{3\sqrt{2}}{2\cos\frac{\theta}{2}}\Big{)}^{59/30}\\}+1\,.$ (68) ###### Proof. Recall that, for $x\geq x_{3}$, $\displaystyle\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}<\Bigg{|}\frac{(1-a)}{a\cos\frac{\theta}{2}}\Bigg{|}<\frac{\sqrt{2}}{\cos\frac{\theta}{2}}\,.$ (69) Taking this into account, we conclude that the corner operators of trees with less than $\alpha|V(T)|$ leaves, with $\alpha=\frac{59}{60}$, are bounded by the bounds for corner operators of trees from $T_{<}$ with the maximal possible amount of leaves $(\left\lceil\frac{59}{60}|V(T)|\right\rceil-1)$. Note that for any $x>\max\\{x_{3},\frac{2^{30}e^{30}}{\cos\frac{\theta}{2}}\Big{(}\frac{3\sqrt{2}}{2\cos\frac{\theta}{2}}\Big{)}^{59/30}\\}$, we have $\displaystyle e\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}^{\frac{1}{30}}\bigg{(}\frac{3\sqrt{2}}{2\cos\frac{\theta}{2}}\bigg{)}^{\frac{59}{60}}<\frac{1}{2}\,.$ (70) Therefore, fixing $x_{4}$ as in (68), we obtain $\displaystyle\bigg{|}\operatorname{Tr}\Big{[}\mathop{\overrightarrow{\prod}}\limits_{c\in\partial T_{n_{l},n_{i}}\,\text{corner}}{\cal C}_{c}(i_{c})\Big{]}\bigg{|}\leq$ $\displaystyle N\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}^{2|V(T)|-2\left\lceil\frac{59}{60}|V(T)|\right\rceil-2}\Bigg{|}\frac{3(1-a)}{2a\cos\frac{\theta}{2}}\Bigg{|}^{\left\lceil\frac{59}{60}|V(T)|\right\rceil-1}\leq$ $\displaystyle N\Bigg{|}\frac{\sqrt{\lambda}}{a}\frac{1}{\cos\frac{\theta}{2}}\Bigg{|}^{\frac{1}{30}|V(T)|-4}\Bigg{|}\frac{3\sqrt{2}}{2\cos\frac{\theta}{2}}\Bigg{|}^{\frac{59}{60}|V(T)|}\leq$ $\displaystyle\frac{2(x_{4}\cos\frac{\theta}{2})^{5}\sqrt{\lambda}}{3(x_{4}\sqrt{\lambda}-1)}\frac{N}{e^{|V(T)|}}\Big{(}\frac{1}{2}\Big{)}^{|V(T)|}\,.$ (71) The number of all LVE trees with $n$ vertices is given by $n^{n-2}$, this amount can be employed as an upper bound for the number of trees $T_{<}$. Using the lower bound for the factorial, we obtain $\displaystyle|T_{<}|\leq\frac{n^{n-2}}{n!}\leq\frac{n^{n-2}}{\Big{(}\frac{n+1}{e}\Big{)}^{n}}\leq\frac{e^{n}}{n^{2}}\,.$ (72) Combining (71) with (72), and (33), we obtain (67). ∎ Equations (33), (47), lemmas 3, 4, 7, 9, and 10 together with the theorem 1 from [12] give the proof of the theorem 1 of the current paper. ## References * [1] V. Rivasseau, “Constructive Matrix Theory,” JHEP 0709 (2007) 008, arXiv:0706.1224 [hep-th]. * [2] D. Brydges and T. Kennedy, Mayer expansions and the Hamilton-Jacobi equation, Journal of Statistical Physics, 48, 19 (1987). * [3] A. Abdesselam and V. Rivasseau, “Trees, forests and jungles: A botanical garden for cluster expansions,” arXiv:hep-th/9409094. * [4] A. D. Sokal, “An Improvement Of Watson’s Theorem On Borel Summability,” J. Math. Phys. 21, 261 (1980). * [5] V. Rivasseau and Z. Wang, “How to Resum Feynman Graphs,” Annales Henri Poincaré 15, no. 11, 2069 (2014), arXiv:1304.5913 [math-ph]. * [6] R. Gurau and J. P. Ryan, “Colored Tensor Models - a review,” SIGMA 8, 020 (2012), arXiv:1109.4812 [hep-th]. * [7] R. Gurau, “Random Tensors”, Oxford University Press (2016). * [8] G. ’t Hooft, ”A planar diagram theory for strong interactions.” Nucl. Phys. B 72, 461 (1974). * [9] R. Gurau, “The 1/N expansion of colored tensor models,” Annales Henri Poincaré 12, 829 (2011), arXiv:1011.2726 [gr-qc]. * [10] R. Gurau and V. Rivasseau, “The 1/N expansion of colored tensor models in arbitrary dimension,” Europhys. Lett. 95, 50004 (2011), arXiv:1101.4182 [gr-qc]. * [11] R. Gurau, “The complete 1/N expansion of colored tensor models in arbitrary dimension,” Annales Henri Poincaré 13, 399 (2012), arXiv:1102.5759 [gr-qc]. * [12] Gurau, Razvan G., and Thomas Krajewski. ”Analyticity results for the cumulants in a random matrix model.” Annales de l’Institut Henri Poincare D 2, no. 2 (2015): 169-228. * [13] R. Gurau, “The 1/N Expansion of Tensor Models Beyond Perturbation Theory,” Commun. Math. Phys. 330, 973 (2014), arXiv:1304.2666 [math-ph]. * [14] T. Delepouve, R. Gurau and V. Rivasseau, “Universality and Borel Summability of Arbitrary Quartic Tensor Models,” arXiv:1403.0170 [hep-th]. * [15] J. Magnen and V. Rivasseau, “Constructive $\phi^{4}$ field theory without tears,” Annales Henri Poincaré 9 (2008) 403, arXiv:0706.2457 [math-ph]. * [16] R. Gurau and V. Rivasseau, “The Multiscale Loop Vertex Expansion,” Annales Henri Poincaré 16, no. 8, 1869 (2015), arXiv:1312.7226 [math-ph]. * [17] T. Delepouve and V. Rivasseau, “Constructive Tensor Field Theory: The $T^{4}_{3}$ Model,” arXiv:1412.5091 [math-ph]. * [18] V. Lahoche, “Constructive Tensorial Group Field Theory II: The $U(1)-T^{4}_{4}$ Model,” arXiv:1510.05051 [hep-th]. * [19] V. Rivasseau and F. Vignes-Tourneret, “Constructive tensor field theory: The $T^{4}_{4}$ model,” arXiv:1703.06510 [math-ph]. * [20] V. Rivasseau, “Constructive Tensor Field Theory,” SIGMA 12, 085 (2016), arXiv:1603.07312 [math-ph]. * [21] V. Rivasseau and Z. Wang, “Corrected loop vertex expansion for $\Phi_{2}^{4}$ theory,” J. Math. Phys. 56, no. 6, 062301 (2015), arXiv:1406.7428 [math-ph]. * [22] V. Rivasseau, “Loop Vertex Expansion for Higher Order Interactions,” arXiv:1702.07602 [math-ph]. * [23] Krajewski, Thomas, Vincent Rivasseau, and Vasily Sazonov. ”Constructive matrix theory for higher-order interaction.” In Annales Henri Poincare, vol. 20, pp. 3997-4032. Springer International Publishing, 2019. * [24] Krajewski, Thomas, Vincent Rivasseau, and Vasily Sazonov. ”Constructive matrix theory for higher-order interaction II: Hermitian and real symmetric cases.” In Annales Henri Poincare, vol. 23, no. 10, pp. 3431-3452. Cham: Springer International Publishing, 2022. * [25] Feynman, R.P. and Kleinert, H., 1986. ”Effective classical partition functions,” Physical Review A, 34(6), p.5080. * [26] Kleinert, H., and W. Janke. ”Convergence behavior of variational perturbation expansion - A method for locating Bender-Wu singularities,” Physics Letters A 206, no. 5-6 (1995): 283-289. * [27] Stevenson, Paul M. ”Optimized perturbation theory.” Physical Review D 23, no. 12 (1981): 2916. * [28] Shaverdyan, B. S., and A. G. Ushveridze. ”Convergent perturbation theory for the scalar $\phi^{2p}$ field theories; the Gell-Mann-Low function.” Physics Letters B 123, no. 5 (1983): 316-318. * [29] Guida, Riccardo, Kenichi Konishi, and Hiroshi Suzuki. ”Improved convergence proof of the delta expansion and order dependent mappings.” Annals of Physics 249, no. 1 (1996): 109-145. * [30] Caswell, William E. ”Accurate energy levels for the anharmonic oscillator and a summable series for the double-well potential in perturbation theory.” Annals of Physics 123, no. 1 (1979): 153-184. * [31] Halliday, I. G., and P. Suranyi. ”Anharmonic oscillator: A new approach.” Physical Review D 21, no. 6 (1980): 1529. * [32] Kleinert, Hagen. ”Strong-coupling behavior of $\varphi^{4}$ theories and critical exponents.” Physical Review D 57, no. 4 (1998): 2264. * [33] Stevenson, Paul M. ”Optimization and the ultimate convergence of QCD perturbation theory.” Nuclear Physics B 231, no. 1 (1984): 65-90. * [34] Ivanov, Aleksandr., and Vasily Sazonov. ”Convergent series for lattice models with polynomial interactions.” Nuclear Physics B 914 (2017): 43-61. * [35] Ivanov, Aleksandr, and Vasily Sazonov. ”Infinite lattice models by an expansion with a non-Gaussian initial approximation.” Physics Letters B 796 (2019): 52-58. * [36] Sazonov, Vasily. ”Convergent series for polynomial lattice models with complex actions.” Modern Physics Letters A 34, no. 30 (2019): 1950243. * [37] Brandt, Sebastian F., and Axel Pelster. ”Large-D expansion from variational perturbation theory.” Journal of Mathematical Physics 46, no. 11 (2005). * [38] Janke, W. ”Variational Perturbation Theory: a Powerful Method for Deriving Strong-Coupling Expansions.” In Fluctuating Paths And Fields: Festschrift Dedicated to Hagen Kleinert on the Occasion of His 60th Birthday, pp. 301-314. 2001. * [39] Simon, Barry, and A. Dicke. ”Coupling constant analyticity for the anharmonic oscillator.” Annals of Physics 58, no. 1 (1970): 76-136. * [40] Benedetti, Dario, Razvan Gurau, Hannes Keppler, and Davide Lettera. ”The small-$N$ series in the zero-dimensional $O(N)$ model: constructive expansions and transseries.” arXiv preprint arXiv:2210.14776 (2022).
††thanks: These authors contributed equally to this work††thanks: These authors contributed equally to this work # Ionization-induced Long-lasting Orientation of Symmetric-top Molecules Long Xu AMOS and Department of Chemical and Biological Physics, The Weizmann Institute of Science, Rehovot 7610001, Israel Ilia Tutunnikov AMOS and Department of Chemical and Biological Physics, The Weizmann Institute of Science, Rehovot 7610001, Israel Yehiam Prior<EMAIL_ADDRESS>AMOS and Department of Chemical and Biological Physics, The Weizmann Institute of Science, Rehovot 7610001, Israel Ilya Sh. Averbukh <EMAIL_ADDRESS>AMOS and Department of Chemical and Biological Physics, The Weizmann Institute of Science, Rehovot 7610001, Israel ###### Abstract We theoretically consider the phenomenon of field-free long-lasting orientation of symmetric-top molecules ionized by two-color laser pulses. The anisotropic ionization produces a significant long-lasting orientation of the surviving neutral molecules. The degree of orientation increases with both the pulse intensity and, counterintuitively, with the rotational temperature. The orientation may be enhanced even further by using multiple delayed two-color pulses. The long-lasting orientation may be probed by even harmonic generation or by Coulomb-explosion-based methods. The effect may enable the study of relaxation processes in dense molecular gases, and may be useful for molecular guiding and trapping by inhomogeneous fields. _Introduction_.—Field-free oriented molecules are essential in many studies, such as ultrafast dynamic imaging, molecular tomography, and electron diffraction, to name just a few. A much more comprehensive list of applications may be found in a recent review by Koch et al. Koch _et al._ (2019). Naturally, for practical applications, a sizable degree of orientation is beneficial. One of the tools for inducing the molecular orientation is a non-resonant two-color laser pulse consisting of the fundamental wave (FW) and its second harmonic (SH). Using such fields, two different orientation mechanisms have been identified and studied theoretically and experimentally Kanai and Sakai (2001); De _et al._ (2009); Oda _et al._ (2010); Spanner _et al._ (2012); Frumker _et al._ (2012a); Znakovskaya _et al._ (2014); Kraus _et al._ (2015). The first orientation mechanism, which is dominant at low to moderate (non-ionizing) intensities, relies on the interaction of the external fields with the molecular hyperpolarizability, which results in asymmetric torques that orient the molecules along the polarization direction of the SH field Kanai and Sakai (2001); De _et al._ (2009); Oda _et al._ (2010); Lin _et al._ (2018); Xu _et al._ (2021a). At high (ionizing) intensities, the dominant orientation mechanism Spanner _et al._ (2012) is different – probability of ionization depends on the molecular orientation with respect to the polarization direction of the asymmetric electric field of the two-color pulse Spanner _et al._ (2012); Frumker _et al._ (2012a); Znakovskaya _et al._ (2014). As a result, immediately after the pulse, the angular distribution of the surviving neutral molecules is asymmetric and has a non-zero orientation on average. Note that for linear molecules, this orientation disappears shortly after the excitation, but it periodically reemerges due to the phenomenon of rotational quantum revivals Sh. Averbukh and Perelman (1989); Robinett (2004). Here, we theoretically investigate the ionization-induced orientation of _symmetric-top molecules_ excited by intense two-color femtosecond laser pulses. We demonstrate that in addition to the transient post-pulse orientation, and unlike linear molecules, there also exists a significant long-lasting orientation in these molecules. Long-lasting means that the orientation exists not only at the revival times but between the revivals too. In other words, the orientation signal has a non-zero baseline. Within the idealized model of non-interacting rigid rotors used here, this orientation lasts indefinitely. In practice, however, it will eventually be suppressed by additional physical effects, e.g., by intermolecular collisions in gas cell experiments. Related effects of long-lasting orientation have been recently investigated in chiral Milner _et al._ (2019); Tutunnikov _et al._ (2020, 2021); Xu _et al._ (2021a) and other non-linear Xu _et al._ (2020, 2021b, 2021a) molecules excited by non-ionizing THz and laser pulses. In what follows, we present our numerical analysis, outline our results on significant long-lasting orientation, and discuss its dependence on intensity and temperature. We conclude with a discussion of the experimental feasibility of observing the predicted effect. _Numerical methods_.—In our analysis, we simulate the rotational dynamics of symmetric-top molecules within the rigid rotor approximation both classically and quantum-mechanically. The symmetric-top molecules are excited by a two- color laser pulse, consisting of the co-linearly polarized and phase-locked FW field and its SH. The electric field is described by $\mathbf{E}(t)=E_{0}f(t)[\cos(\omega t)+\varepsilon\cos(2\omega t+\phi_{0})]\mathbf{e}_{Z},$ (1) where $E_{0}$ and $\omega$ are the peak amplitude and the carrier frequency of the FW, respectively. The laser pulse envelope is defined by $f(t)=\exp[-2\ln 2\,(t^{2}/\sigma^{2})]$, where $\sigma$ is the full width at half maximum (FWHM) of the pulse intensity profile, and $\mathbf{e}_{Z}$ is a unit vector along the laboratory $Z$ axis. Here, we set $\varepsilon=1$ and $\phi_{0}=0$. The Hamiltonian describing molecular rotation driven by a two-color laser pulse is given by $H(t)=H_{r}+H_{\mathrm{int}}$, where $H_{r}$ is the rotational kinetic energy Hamiltonian and the interaction Hamiltonian is given by $H_{\mathrm{int}}=V_{\mathrm{pol}}+V_{\mathrm{hyp}}+V_{\mathrm{ion}}.$ (2) The field-polarizability and field-hyperpolarizability interaction terms are defined as Buckingham (2007) $V_{\mathrm{pol}}=-\frac{1}{2}\sum_{i,j}\alpha_{ij}E_{i}E_{j},V_{\mathrm{hyp}}=-\frac{1}{6}\sum_{i,j,k}\beta_{ijk}E_{i}E_{j}E_{k},\\!\\!$ (3) where $E_{i}$, $\alpha_{ij}$, and $\beta_{ijk}$ are the components of the electric field vector, polarizability tensor, and hyperpolarizability tensor, respectively. Figure 1: (a) $\mathrm{CH_{3}F}$ molecule. Atoms are color-coded: black, carbon; gray, hydrogen; green, fluorine. $\theta$ is the angle between the molecular $a$ axis and the laboratory $Z$ axis, and $\chi$ represents the rotation angle about the $a$ axis. (b) Structure factor $G(\theta,\chi)$ [see Eqs. (7) and (8)] determining the angle dependence of the ionization rate. The ionization depletion term $V_{\mathrm{ion}}$ is sensitive to molecular orientation. For linear molecules, e.g., $\mathrm{HCl}$ Akagi _et al._ (2009) and $\mathrm{CO}$ Li _et al._ (2011); Wu _et al._ (2012); Spanner _et al._ (2012), the ionization rate depends on the angle between the molecular axis and the polarization direction. For symmetric-top molecules, e.g., $\mathrm{CH_{3}F}$ and $\mathrm{CH_{3}Br}$, belonging to the $C_{3v}$ point group, the ionization rate depends on two angles, $\theta$ and $\chi$ (see Fig. 1) Kraus _et al._ (2015). In this work we consider $\mathrm{CH_{3}F}$ as our example symmetric-top molecule. The ionization process is modeled using a complex absorbing potential $V_{\mathrm{ion}}=-\frac{i}{2}\Gamma(\theta,\chi,t),$ (4) where the ionization rate $\Gamma(\theta,\chi,t)$ is defined as Kraus _et al._ (2015) (within the weak-field asymptotic theory) $\Gamma(\theta,\chi,t)=\begin{cases}W(t)|G(\theta,\chi)|^{2},&E(t)>0,\\\ W(t)|G(\pi-\theta,\pi+\chi)|^{2},&E(t)<0.\end{cases}$ (5) Here, $E(t)=\mathbf{e}_{Z}\cdot\mathbf{E}(t)$, $W(t)$ is the field factor, and $G(\theta,\chi)$ is the structure factor. The field factor is given by Kraus _et al._ (2015) $W(t)=\frac{\kappa}{2}\left(\frac{4\kappa^{2}}{|E(t)|}\right)^{2/\kappa-1}\exp\left[-\frac{2\kappa^{3}}{3|E(t)|}\right],$ (6) where $\kappa=\sqrt{2I_{p}}$, and $I_{p}$ is the field-free energy of the highest occupied molecular orbital (HOMO). We use the following model for the structure factor $G(\theta,\chi)=\left[\sin(\theta)+\frac{3}{2}\sin(2\theta)\right]G_{1}(\chi),$ (7) where $G_{1}(\chi)=\begin{cases}\,\,\,\,\sqrt{1+\sin(3\chi)},&0\leq\chi<7\pi/6,\\\ -\sqrt{1+\sin(3\chi)},&7\pi/6\leq\chi<11\pi/6,\\\ \,\,\,\,\sqrt{1+\sin(3\chi)},&11\pi/6\leq\chi<2\pi.\end{cases}\\!\\!\\!$ (8) $G(\theta,\chi)$ defined in Eqs. (7) and (8) closely approximates the structure factor of field-dressed HOMO of $\mathrm{CH_{3}F}$ with the largest dipole moment Kraus _et al._ (2015) (the orbital from which the strong field ionization preferentially occurs). The definition in Eq. (5) accounts for the oscillations of the laser electric field along the $Z$ axis. We quantify the degree of orientation of surviving neutral molecules using the thermally averaged quantum expectation value of $\cos(\theta)$, $\braket{\cos(\theta)}$. Further details on the quantum simulations can be found in Xu _et al._ (2021b); Tutunnikov _et al._ (also see Sec. I of the Supplemental Material Sup ). In classical simulations, we use the Monte Carlo approach to simulate the behavior of a classical ensemble consisting of $N=10^{7}$ sample molecules. A detailed description can be found in Tutunnikov _et al._ (2019); Xu _et al._ (2021b) (also see Sec. I of the Supplemental Material Sup ). Following the ionization depletion, the classical degree of orientation is given by $\displaystyle\braket{\cos(\theta)}(t)=\sum\limits_{n=1}^{N}\rho(\theta_{n},\chi_{n},t)\cos(\theta_{n}),$ (9) where the relative weight (non-ionized fraction) of the $n$-th molecule is $\displaystyle\rho(\theta_{n},\chi_{n},t)=N_{\mathrm{neu}}^{-1}\exp\left[-\int_{0}^{t}\Gamma(\theta_{n},\chi_{n},t^{\prime})\,dt^{\prime}\right],$ (10) and the total number of surviving neutral molecules is $\displaystyle N_{\mathrm{neu}}=\sum\limits_{n=1}^{N}\exp\left[-\int_{0}^{t}\Gamma(\theta_{n},\chi_{n},t^{\prime})\,dt^{\prime}\right],$ (11) Here, $\theta_{n}$ and $\chi_{n}$ are the time-dependent angles of the $n$-th molecule. The population of surviving neutral molecules is defined as $N_{\mathrm{neu}}/N$. _Results_.—Molecular parameters of $\mathrm{CH_{3}F}$ are provided in Sec. II of the Supplemental Material Sup . Figure 2 shows the calculated, classically and quantum-mechanically, time-dependent orientation factor of neutral molecules following a single two-color pulse applied at $t=0$. The parameters used in this calculation are: the rotational temperature is $T=300\,\mathrm{K}$, the laser wavelengths are 800 nm (FW) and 400 nm (SH), the peak intensity is $7\times 10^{13}\,\mathrm{W/cm^{2}}$, and $\sigma=20\,\mathrm{fs}$, see Eq. (1). The three panels of Fig. 2 show the orientation factor obtained for various combinations of interaction terms [see Eq. (2)]. All cases include the field- polarizability interaction, $V_{\mathrm{pol}}\propto\cos^{2}(\theta)$ which is a symmetric function of $\theta$ (about $\theta=\pi/2$). A torque-kick by such a potential results in molecular alignment only (for review, see Stapelfeldt and Seideman (2003)). The two other terms, $V_{\mathrm{hyp}}$ and $V_{\mathrm{ion}}$ are asymmetric functions of $\theta$ and thus induce molecular orientation. All three panels depict the immediate response of $\braket{\cos(\theta)}$ to the laser excitation near $t=0$. This transient orientation effect is similar to the one studied in linear molecules excited by two-color laser pulses De _et al._ (2009); Oda _et al._ (2010); Wu and Zeng (2010); Mun and Sakai (2018); Mun _et al._ (2019); Mun and Kim (2020); Mellado-Alcedo _et al._ (2020); Wang and Henriksen (2020). At room temperature and field parameters used here, the transient molecular orientation induced by the field-hyperpolarizability interaction alone [Fig. 2(a)] is negligible compared to the orientation resulting from the ionization depletion [Fig. 2(b)]. Accordingly, the curves in Fig. 2(b) and Fig. 2(c) [in which all the interaction terms are included, see Eq. (2)] are almost indistinguishable. These results are consistent with previous results reported for linear molecules Spanner _et al._ (2012); Znakovskaya _et al._ (2014), where it was shown that at high (ionizing) intensities, the orientation mechanism of ionization depletion dominates. Figure 2: Time-dependent orientation factor at $T=300\,\mathrm{K}$ for different orientation mechanisms calculated classically and quantum- mechanically. Here the field intensity is $I_{0}=7\times 10^{13}\,\mathrm{W/cm^{2}}$ and the pulse duration is $\sigma=20\,\mathrm{fs}$. About 40% of neutral molecules survive after the ionization. In this work, we focus on the long-term orientation existing in symmetric-top molecules. Figures 2(b) and 2(c) show that under the stated conditions and for these molecules, following the transient orientation the degree of orientation doesn’t return to zero but persists at a constant value till the first revival and beyond. _This effect of ionization-induced long-lasting orientation doesn’t exist in linear molecules, and it is the main result of this Letter_. Note that the quantum and classical results agree well on the short time scale, during the initial transient response, and predict the same degree of long-lasting orientation on the long time scale, suggesting that the long- lasting orientation stems from a classical origin. Long-lasting orientation has been previously observed in chiral Milner _et al._ (2019); Tutunnikov _et al._ (2020, 2021); Xu _et al._ (2021a) and studied in other non-linear molecules Xu _et al._ (2020, 2021b, 2021a) excited by non-ionizing THz and laser pulses. In these cases, the orientation mechanisms, including the interactions with the polarizability and hyperpolarizability, were considered. Here, the ionization depletion mechanism of ionization depletion gives rise to unprecedented degrees of long-lasting orientation at room temperature. The degree of long-lasting orientation ($\approx-0.033$), as shown in Figs. 2(b) and 2(c), is an order of magnitude higher than values reported in previous studies. _Mechanism_.—Next, we discuss the mechanism behind this large ionization- induced long-lasting orientation. Under field-free condition, the symmetry axis (dipole) of symmetric-top molecules precesses around the (conserved, space-fixed) vector of angular momentum, whereas linear molecules rotate in a plane perpendicular to the angular momentum Landau and Lifshitz (1976). It is this precession that is the source of the long-lasting orientation in symmetric-top molecules. The degree of long-lasting orientation (with respect to the $Z$ axis) of a single symmetric-top molecule is given by the combination of three quantities $L_{a}L_{Z}/L^{2}$ Xu _et al._ (2020, 2021b) (see also Sec. III of the Supplemental Material Sup ), where $L_{a}$ and $L_{Z}$ are the projections of the angular momentum along the molecular symmetry axis [the molecular $a$ axis in Fig. 1(a)] and the laboratory $Z$ axis, respectively, and $L$ is the magnitude of the angular momentum. Note that in the presence of $Z$-polarized pulses, $L_{a}$ and $L_{Z}$ are conserved quantities. Initially, before the laser pulse, the molecules are isotropically distributed in space, namely there is an equal number of molecules having positive (along the $Z$ axis) and negative (against the $Z$ axis) long-lasting orientations. Therefore, the ensemble-averaged long-lasting orientation $\braket{L_{a}L_{Z}/L^{2}}$ vanishes Xu _et al._ (2021b). The co-linearly polarized two-color laser pulse preferentially ionizes molecules oriented more or less along its polarization axis (with $\theta<\pi/2$) [see Fig. 1(b)]. Thus, this selective ionization depletion breaks the symmetry of the molecular ensemble, generating a non-zero long-lasting orientation of the surviving (not ionized) neutral molecules, as shown in Figs. 2(b) and 2(c). For symmetric-top molecules, within the model used here, the long-lasting orientation is permanent. In experiments, however, this orientation will be gradually destroyed by other physical effects such as intermolecular collisions, centrifugal distortion, and radiation emission caused by rapidly rotating molecular permanent dipole moments. Furthermore, while the centrifugal distortion is known to lead to the decay of the revivals’ peaks due to the dephasing of the rotational states Damari _et al._ (2016); Babilotte _et al._ (2016), the average dipole remains almost unchanged (see Xu _et al._ (2020)). The radiative emission gradually decreases the rotational energy Damari _et al._ (2016); Babilotte _et al._ (2016), but the relative energy loss during a single revival is negligible for a rarefied molecular gas. _Intensity dependence_.—One of the ways to enhance the degree of the ionization-induced long-lasting orientation is by increasing the laser pulse energy – peak intensity and/or pulse duration, or investing the higher energy in several pulses. Figure 3 shows the classically calculated long-lasting orientation factor, $\braket{\cos(\theta)}_{p}$, and the population of surviving neutral molecules as functions of the laser intensity for single or multiple delayed pulses. The value of $\braket{\cos(\theta)}_{p}$ is taken after the pulse(s), when the orientation factor reaches a constant value. As expected, with the increasing input energy, the long-lasting orientation factor (in absolute value) initially grows [for $I_{0}<7\times 10^{13}\,\mathrm{W/cm^{2}}$, see Fig. 3(a)], while the population of surviving neutral molecules decreases monotonically [see Fig. 3(b)]. After reaching its maximum value, the long-lasting orientation decreases with increasing laser intensity. The reason for this decrease is the structure factor $G(\theta,\chi)$ shown in Fig. 1(b). Due to the anisotropy of the ionization process, molecules at $\theta\approx 0,\,0.6\pi,\,\pi$ have a relatively low ionization rate. Therefore, when the ionization saturates at high laser intensities, only these molecules survive, but these molecules have a relatively low contribution to the long-lasting orientation. The combination of increased ionization yield, but counterproductive selectivity of low- contributing molecules limits the usefulness of increasing the laser intensity beyond a certain point. Figure 3: Classically calculated (a) long-lasting orientation and (b) population of neutral molecules after the pulse(s) as functions of the laser intensity. The time delay between pulses is 0.5 ps. Here $T=300\,\mathrm{K}$ and $\sigma=20\,\mathrm{fs}$. Both orientation mechanisms are taken into account. A way to avoid the limits imposed on high intensities is to use multiple pulses instead of a single strong pulse. Consider the application of several delayed two-color pulses. Figure 3(a) shows that the maximum long-lasting orientation is about $-0.033$ for one pulse, $-0.069$ for two pulses, and $-0.107$ for three pulses. Here, the time delay between each pulse is set to 0.5 ps, a relatively long time delay which allows the molecules to rotate between the pulses. This way, after each pulse additional neutral molecules rotate to an orientation more favorable for ionization by the next pulse. This approach overcomes the ionization saturation limit that exists in the case of a single pulse, and is similar to the use of multiple pulses for achieving enhanced alignment while avoiding ionization Leibscher _et al._ (2003, 2004). Naturally, adding more pulses results in a progressively lower population of surviving neutral molecules. Nevertheless, for a fixed population, a higher long-lasting orientation is achieved by applying multiple delayed pulses. For example, for $N_{\mathrm{neu}}/N\approx 40\%$, the long-lasting orientation is about $-0.033$ for one pulse, $-0.049$ for two pulses, and $-0.054$ for three pulses. _This observation is an additional main result of this Letter._ Optimization of the time delay may allow further enhancement of the long- lasting orientation. _Temperature dependence_.—Next, we consider the temperature dependence of the long-lasting orientation as depicted in Fig. 4. At $T=0$, the long-lasting orientation is zero for all three curves shown in Fig. 4(a), as a result of $L_{a}=L_{Z}=0$ (these are conserved quantities). There is an optimal temperature ($T<1\,\mathrm{K}$ in this example) at which the long-lasting orientation induced by the field-hyperpolarizability interaction (together with the field-polarizability interaction) reaches the maximum ($\approx 0.006$). Above the optimal temperature, the long-lasting orientation decays with increasing temperature Xu _et al._ (2021b). In sharp contrast, the ionization-induced long-lasting orientation factor (in absolute value) increases monotonically with the temperature, which is another principal finding of this work. Figure 4: Classically calculated (a) long-lasting orientation factor and (b) population of neutral molecules as functions of temperature for different orientation mechanisms. The laser parameters are the same as in Fig. 2: $I_{0}=7\times 10^{13}\,\mathrm{W/cm^{2}}$, $\sigma=20\,\mathrm{fs}$. The population for the orientation mechanism of hyperpolarizability remains 1. As mentioned above, the long-lasting orientation is given by $L_{a}L_{Z}/L_{f}^{2}$, where $L_{f}=|\mathbf{L}_{f}|$ is the magnitude of the angular momentum after the pulse. $L_{Z}$ and $L_{a}$ are conserved in the case of excitation by linearly polarized laser pulses and thus are functions of temperature only. The short two-color pulse has a two-fold effect: (i) it (almost) instantaneously ionizes the molecules at particular angles (the effect described by $V_{\mathrm{ion}}$), and (ii) it changes the molecular angular momentum, $\mathbf{L}_{f}=\mathbf{L}_{i}+\delta\mathbf{L}$, where $\mathbf{L}_{i}$ is the initial angular momentum, and $\delta\mathbf{L}$ is the change caused by $V_{\mathrm{pol}}$. Around the room temperature (in the examples considered here), the effect of $V_{\mathrm{pol}}$ becomes negligible, i.e., $L_{a}L_{Z}/L_{f}^{2}\approx L_{a}L_{Z}/L_{i}^{2}$, and the degree of long-lasting orientation reaches a constant value determined by $V_{\mathrm{ion}}$. At lower temperatures, $V_{\mathrm{pol}}$ has a detrimental effect by increasing the total angular momentum, such that $L_{a}L_{Z}/L_{f}^{2}<L_{a}L_{Z}/L_{i}^{2}$ effectively lowers the long- lasting orientation. At temperatures above room temperature, the ionization can no longer be considered instantaneous because the molecules rotate fast enough to change their orientation during the pulse. Accordingly, more molecules get ionized, as seen from the populations in Fig. 4(b). The higher the ionization yield, the higher asymmetry of the molecular ensemble, manifesting in the higher long-lasting orientation. _Conclusions_.—We have theoretically demonstrated a sizable ionization-induced long-lasting orientation of symmetric-top molecules excited by two-color laser pulses. The mechanism leading to this observation is the selective ionization of the polar molecules at particular angles with respect to the laser’s electric field, and the ability of symmetric-top molecules to precess around the fixed angular momentum vector. We show that using a proper sequence of delayed two-color pulses allows for enhancing the long-lasting orientation of neutral molecules without drastic depletion of their population. Due to the required precession, the enhanced orientation favors high temperature (room temperature and above in our examples). The long-lasting orientation may be measured with the help of Coulomb explosion Znakovskaya _et al._ (2014), or by observing second (or higher order) harmonic generation in the gas phase, which is sensitive to the lack of inversion symmetry Frumker _et al._ (2012b, c); Kraus _et al._ (2012). Long-lasting orientation can pave the way to the study of relaxation processes in dense molecular gases that are not otherwise accessible to direct probing (see analogous applications of persistent alignment Hartmann and Boulet (2012); Vieillard _et al._ (2013)). In addition, molecular focusing, guiding, and trapping by inhomogeneous fields rely on molecular orientation Gershnabel and Sh. Averbukh (2011a, b); Küpper _et al._ (2012), and their observation may be facilitated by this long-lasting orientation effect. ###### Acknowledgements. L.X. is a recipient of the Sir Charles Clore Postdoctoral Fellowship. This research was made possible in part by the historic generosity of the Harold Perlman Family. ## References * Koch _et al._ (2019) C. P. Koch, M. Lemeshko, and D. Sugny, Quantum control of molecular rotation, Rev. Mod. Phys. 91, 035005 (2019). * Kanai and Sakai (2001) T. Kanai and H. Sakai, Numerical simulations of molecular orientation using strong, nonresonant, two-color laser fields, J. Chem. Phys 115, 5492 (2001). * De _et al._ (2009) S. De, I. Znakovskaya, D. Ray, F. Anis, N. G. Johnson, I. A. Bocharova, M. Magrakvelidze, B. D. Esry, C. L. Cocke, I. V. Litvinyuk, and M. F. Kling, Field-Free Orientation of CO Molecules by Femtosecond Two-Color Laser Fields, Phys. Rev. Lett. 103, 153002 (2009). * Oda _et al._ (2010) K. Oda, M. Hita, S. Minemoto, and H. Sakai, All-optical molecular orientation, Phys. Rev. Lett. 104, 213901 (2010). * Spanner _et al._ (2012) M. Spanner, S. Patchkovskii, E. Frumker, and P. Corkum, Mechanisms of two-color laser-induced field-free molecular orientation, Phys. Rev. Lett. 109, 113001 (2012). * Frumker _et al._ (2012a) E. Frumker, C. T. Hebeisen, N. Kajumba, J. B. Bertrand, H. J. Wörner, M. Spanner, D. M. Villeneuve, A. Naumov, and P. B. Corkum, Oriented rotational wave-packet dynamics studies via high harmonic generation, Phys. Rev. Lett. 109, 113901 (2012a). * Znakovskaya _et al._ (2014) I. Znakovskaya, M. Spanner, S. De, H. Li, D. Ray, P. Corkum, I. V. Litvinyuk, C. L. Cocke, and M. F. Kling, Transition between mechanisms of laser-induced field-free molecular orientation, Phys. Rev. Lett. 112, 113005 (2014). * Kraus _et al._ (2015) P. M. Kraus, O. I. Tolstikhin, D. Baykusheva, A. Rupenyan, J. Schneider, C. Z. Bisgaard, T. Morishita, F. Jensen, L. B. Madsen, and H. J. Wörner, Observation of laser-induced electronic structure in oriented polyatomic molecules, Nat. Commun. 6, 7039 (2015). * Lin _et al._ (2018) K. Lin, I. Tutunnikov, J. Qiang, J. Ma, Q. Song, Q. Ji, W. Zhang, H. Li, F. Sun, X. Gong, H. Li, P. Lu, H. Zeng, Y. Prior, I. Sh. Averbukh, and J. Wu, All-optical field-free three-dimensional orientation of asymmetric-top molecules, Nat. Commun. 9, 5134 (2018). * Xu _et al._ (2021a) L. Xu, I. Tutunnikov, Y. Prior, and I. Sh. Averbukh, Three dimensional orientation of small polyatomic molecules excited by two-color femtosecond pulses, J. Phys. B 54, 164003 (2021a). * Sh. Averbukh and Perelman (1989) I. Sh. Averbukh and N. F. Perelman, Fractional revivals: Universality in the long-term evolution of quantum wave packets beyond the correspondence principle dynamics, Phys. Lett. A 139, 449 (1989). * Robinett (2004) R. W. Robinett, Quantum wave packet revivals, Phys. Rep 392, 1 (2004). * Milner _et al._ (2019) A. A. Milner, J. A. M. Fordyce, I. MacPhail-Bartley, W. Wasserman, V. Milner, I. Tutunnikov, and I. Sh. Averbukh, Controlled Enantioselective Orientation of Chiral Molecules with an Optical Centrifuge, Phys. Rev. Lett. 122, 223201 (2019). * Tutunnikov _et al._ (2020) I. Tutunnikov, J. Floß, E. Gershnabel, P. Brumer, I. Sh. Averbukh, A. A. Milner, and V. Milner, Observation of persistent orientation of chiral molecules by a laser field with twisted polarization, Phys. Rev. A 101, 021403(R) (2020). * Tutunnikov _et al._ (2021) I. Tutunnikov, L. Xu, R. W. Field, K. A. Nelson, Y. Prior, and I. Sh. Averbukh, Enantioselective orientation of chiral molecules induced by terahertz pulses with twisted polarization, Phys. Rev. Res. 3, 013249 (2021). * Xu _et al._ (2020) L. Xu, I. Tutunnikov, E. Gershnabel, Y. Prior, and I. Sh. Averbukh, Long-lasting molecular orientation induced by a single terahertz pulse, Phys. Rev. Lett. 125, 013201 (2020). * Xu _et al._ (2021b) L. Xu, I. Tutunnikov, Y. Prior, and I. Sh. Averbukh, Long-lasting orientation of symmetric-top molecules excited by two-color femtosecond pulses, Front. Phys. 9, 689635 (2021b). * Buckingham (2007) A. D. Buckingham, Permanent and induced molecular moments and long-range intermolecular forces, in _Advances in Chemical Physics_ (John Wiley & Sons, New York, 2007) pp. 107–142. * Akagi _et al._ (2009) H. Akagi, T. Otobe, A. Staudte, A. Shiner, F. Turner, R. Dörner, D. M. Villeneuve, and P. B. Corkum, Laser tunnel ionization from multiple orbitals in HCl, Science 325, 1364 (2009). * Li _et al._ (2011) H. Li, D. Ray, S. De, I. Znakovskaya, W. Cao, G. Laurent, Z. Wang, M. F. Kling, A. T. Le, and C. L. Cocke, Orientation dependence of the ionization of CO and NO in an intense femtosecond two-color laser field, Phys. Rev. A 84, 043429 (2011). * Wu _et al._ (2012) J. Wu, L. P. H. Schmidt, M. Kunitski, M. Meckel, S. Voss, H. Sann, H. Kim, T. Jahnke, A. Czasch, and R. Dörner, Multiorbital tunneling ionization of the CO molecule, Phys. Rev. Lett. 108, 183001 (2012). * (22) I. Tutunnikov, L. Xu, Y. Prior, and I. Sh. Averbukh, Echo-enhanced molecular orientation at high temperatures, arXiv:2207.08274 . * (23) See Supplemental Material at URL for the further details on the quantum and classical simulations (Sec. I), for the molecular properties of methyl fluoride (Sec. II), as well as for the derivation of the long-lasting orientation factor in the classical model (Sec. III), which includes Refs. Buckingham (2007); Tolstikhin _et al._ (2011); Kraus _et al._ (2015); Kallush and Fleischer (2015); McDowell (1990); Sidje (1998); Goldstein _et al._ (2002); Kuipers (1999); Coutsias and Romero (2004); Tutunnikov _et al._ (2019); Xu _et al._ (2021b); Johnson (2019); Chong (1992); Bieri _et al._ (1981). * Tolstikhin _et al._ (2011) O. I. Tolstikhin, T. Morishita, and L. B. Madsen, Theory of tunneling ionization of molecules: Weak-field asymptotics including dipole effects, Phys. Rev. A 84, 053423 (2011). * Kallush and Fleischer (2015) S. Kallush and S. Fleischer, Orientation dynamics of asymmetric rotors using random phase wave functions, Phys. Rev. A 91, 063420 (2015). * McDowell (1990) R. S. McDowell, Rotational partition functions for symmetric-top molecules, J. Chem. Phys. 93, 2801 (1990). * Sidje (1998) R. B. Sidje, Expokit: A software package for computing matrix exponentials, ACM Trans. Math. Softw. 24, 130 (1998). * Goldstein _et al._ (2002) H. Goldstein, C. Poole, and J. Safko, _Classical Mechanics_ (Addison Wesley, San Francisco, CA, 2002). * Kuipers (1999) J. B. Kuipers, _Quaternions and Rotation Sequences: A Primer with Applications to Orbits, Aerospace and Virtual Reality_ (Princeton University Press, Princeton, N.J., 1999). * Coutsias and Romero (2004) E. A. Coutsias and L. Romero, The quaternions with an application to rigid body dynamics, Sandia Technical Report, SAND2004-0153 (2004). * Johnson (2019) R. D. Johnson, _NIST Computational chemistry comparison and benchmark database, Release 20_ , Tech. Rep. (2019). * Chong (1992) D. P. Chong, Theoretical calculations of dipole moments, polarizabilities, and hyperpolarizabilities of HF, OCS, $\mathrm{O_{3}}$, $\mathrm{CH_{3}F}$, and $\mathrm{CH_{3}Cl}$ by local density approximation, J. Chin. Chem. Soc. 39, 375 (1992). * Bieri _et al._ (1981) G. Bieri, L. Åsbrink, and W. Von Niessen, 30.4-nm He(II) photoelectron spectra of organic molecules: Part IV. fluoro-compounds (C, H, F), J. Electron Spectrosc. Relat. Phenom. 23, 281 (1981). * Tutunnikov _et al._ (2019) I. Tutunnikov, J. Floß, E. Gershnabel, P. Brumer, and I. Sh. Averbukh, Laser-induced persistent orientation of chiral molecules, Phys. Rev. A 100, 043406 (2019). * Stapelfeldt and Seideman (2003) H. Stapelfeldt and T. Seideman, Colloquium: Aligning molecules with strong laser pulses, Rev. Mod. Phys. 75, 543 (2003). * Wu and Zeng (2010) J. Wu and H. Zeng, Field-free molecular orientation control by two ultrashort dual-color laser pulses, Phys. Rev. A 81, 053401 (2010). * Mun and Sakai (2018) J. H. Mun and H. Sakai, Improving molecular orientation by optimizing relative delay and intensities of two-color laser pulses, Phys. Rev. A 98, 013404 (2018). * Mun _et al._ (2019) J. H. Mun, H. Sakai, and R. González-Férez, Orientation of linear molecules in two-color laser fields with perpendicularly crossed polarizations, Phys. Rev. A 99, 053424 (2019). * Mun and Kim (2020) J. H. Mun and D. E. Kim, Field-free molecular orientation by delay- and polarization-optimized two fs pulses, Sci. Rep 10, 18875 (2020). * Mellado-Alcedo _et al._ (2020) D. Mellado-Alcedo, N. R. Quintero, and R. González-Férez, Linear polar molecule in a two-color cw laser field: A symmetry analysis, Phys. Rev. A 102, 023110 (2020). * Wang and Henriksen (2020) S. Wang and N. E. Henriksen, Optimal field-free molecular orientation with nonresonant two-color adiabatic-turn-on and sudden-turn-off laser pulses, Phys. Rev. A 102, 063120 (2020). * Landau and Lifshitz (1976) L. Landau and E. Lifshitz, _Mechanics_, 3rd ed. (Butterworth-Heinemann, Oxford, 1976). * Damari _et al._ (2016) R. Damari, S. Kallush, and S. Fleischer, Rotational control of asymmetric molecules: Dipole- versus polarizability-driven rotational dynamics, Phys. Rev. Lett. 117, 103001 (2016). * Babilotte _et al._ (2016) P. Babilotte, K. Hamraoui, F. Billard, E. Hertz, B. Lavorel, O. Faucher, and D. Sugny, Observation of the field-free orientation of a symmetric-top molecule by terahertz laser pulses at high temperature, Phys. Rev. A 94, 043403 (2016). * Leibscher _et al._ (2003) M. Leibscher, I. Sh. Averbukh, and H. Rabitz, Molecular alignment by trains of short laser pulses, Phys. Rev. Lett. 90, 213001 (2003). * Leibscher _et al._ (2004) M. Leibscher, I. Sh. Averbukh, and H. Rabitz, Enhanced molecular alignment by short laser pulses, Phys. Rev. A 69, 013402 (2004). * Frumker _et al._ (2012b) E. Frumker, C. T. Hebeisen, N. Kajumba, J. B. Bertrand, H. J. Wörner, M. Spanner, D. M. Villeneuve, A. Naumov, and P. B. Corkum, Oriented rotational wave-packet dynamics studies via high harmonic generation, Phys. Rev. Lett. 109, 113901 (2012b). * Frumker _et al._ (2012c) E. Frumker, N. Kajumba, J. B. Bertrand, H. J. Wörner, C. T. Hebeisen, P. Hockett, M. Spanner, S. Patchkovskii, G. G. Paulus, D. M. Villeneuve, A. Naumov, and P. B. Corkum, Probing polar molecules with high harmonic spectroscopy, Phys. Rev. Lett. 109, 233904 (2012c). * Kraus _et al._ (2012) P. M. Kraus, A. Rupenyan, and H. J. Wörner, High-harmonic spectroscopy of oriented OCS molecules: Emission of even and odd harmonics, Phys. Rev. Lett. 109, 233903 (2012). * Hartmann and Boulet (2012) J.-M. Hartmann and C. Boulet, Quantum and classical approaches for rotational relaxation and nonresonant laser alignment of linear molecules: A comparison for co2 gas in the nonadiabatic regime, J. Chem. Phys. 136, 184302 (2012). * Vieillard _et al._ (2013) T. Vieillard, F. Chaussard, F. Billard, D. Sugny, O. Faucher, S. Ivanov, J.-M. Hartmann, C. Boulet, and B. Lavorel, Field-free molecular alignment for probing collisional relaxation dynamics, Phys. Rev. A 87, 023409 (2013). * Gershnabel and Sh. Averbukh (2011a) E. Gershnabel and I. Sh. Averbukh, Electric deflection of rotating molecules, J. Chem. Phys. 134, 054304 (2011a). * Gershnabel and Sh. Averbukh (2011b) E. Gershnabel and I. Sh. Averbukh, Deflection of rotating symmetric top molecules by inhomogeneous fields, J. Chem. Phys. 135, 084307 (2011b). * Küpper _et al._ (2012) J. Küpper, F. Filsinger, G. Meijer, and H. Stapelfeldt, Manipulating the motion of complex molecules: Deflection, focusing, and deceleration of molecular beams for quantum-state and conformer selection, in _Methods in Physical Chemistry_ (John Wiley & Sons, Ltd, 2012) Chap. 1, pp. 1–28.
# Generalized Relation Learning with Semantic Correlation Awareness for Link Prediction Yao Zhang1, Xu Zhang1, Jun Wang2, Hongru Liang1, Wenqiang Lei3111Corresponding authors., Zhe Sun4, Adam Jatowt5, Zhenglu Yang111footnotemark: 1 ###### Abstract Developing link prediction models to automatically complete knowledge graphs has recently been the focus of significant research interest. The current methods for the link prediction task have two natural problems: 1) the relation distributions in KGs are usually unbalanced, and 2) there are many unseen relations that occur in practical situations. These two problems limit the training effectiveness and practical applications of the existing link prediction models. We advocate a holistic understanding of KGs and we propose in this work a unified Generalized Relation Learning framework GRL to address the above two problems, which can be plugged into existing link prediction models. GRL conducts a generalized relation learning, which is aware of semantic correlations between relations that serve as a bridge to connect semantically similar relations. After training with GRL, the closeness of semantically similar relations in vector space and the discrimination of dissimilar relations are improved. We perform comprehensive experiments on six benchmarks to demonstrate the superior capability of GRL in the link prediction task. In particular, GRL is found to enhance the existing link prediction models making them insensitive to unbalanced relation distributions and capable of learning unseen relations. ## Introduction Figure 1: (a) The unbalanced relation distribution in the FB15K-237 dataset where relations are sorted according to their frequency. (b) Lots of unseen relations. Three film-related relations are respectively categorized into the many-shot class, few-shot class, and zero-shot class as marked. Knowledge graphs (KGs), representing facts in semantic graph structures, have been applied to multiple artificial intelligence tasks, e.g., recommendation (Lei et al. 2020a; Wang et al. 2021), dialogue generation (Moon et al. 2019; Lei et al. 2020b), and question answering (Christmann et al. 2019; Zhu et al. 2021). In KGs, facts are formed as triples, (head entity, relation, tail entity), where the head entity is linked to the tail entity via the relation. New knowledge emerges continuously, and hence the issue of incompleteness of KGs has triggered wide research interests in link prediction task, which requires predicting the missing links in KGs (Seyed and David 2018). The mainstream link prediction models (Bordes et al. 2013; Dettmers et al. 2018) learn the embeddings of entities and relations, and then use a score function to estimate the validity of triples. However, we believe using the embedding learning for mainstream link prediction models results in two key problems: * $\bullet$ Unbalanced relation distribution. As shown in Figure 1, the relation distribution in an off-the-shelf KG learning resource (i.e., FB15K-237 (Toutanova and Chen 2015)) is quite unbalanced. For example, the frequencies of the two relations /film/film/language and /film/film/edited differ greatly. Mainstream link prediction models assume enough training instances for all relations and pay less attention to few-shot relations, disregarding the fact that few-shot relation learning may influence the learning performance to a high degree. * $\bullet$ Existence of unseen relations. Real-world KGs tend to be open and evolve quickly, and accordingly there is a large number of zero-shot relations unseen in the off-the-shelf learning resources, for example, the relation /film/film/sequel in Figure 1. The unseen relations are beyond the capacity of mainstream link prediction models, as there are no training instances to learn their embeddings. This problem may restrict the use of these models in downstream tasks. Recently, some efforts have been conducted on addressing the above problems. Xiong et al. (2018), Shi and Weninger (2018), and Chen et al. (2019a) adopted the meta-learning or metric-based approaches to train on limited training samples and perform fast learning on new few-shot relations. These studies show promise in few-shot relation learning, however they have difficulty in tackling unbalanced relation distributions, which is mainly attributed to the excessive time cost required for training numerous relations. More recently, Chen et al. (2019b), Qin et al. (2020) predicted the unseen relations by extracting information from textual descriptions. They were able to successfully complete the unseen relation prediction task. However, these models are not appropriate for link prediction task, since textual descriptions tend to be noisy and also cannot build a bridge between seen and unseen relations. In general, an ideal link prediction model should be able to jointly learn many-, few-, and zero- shot relations. Regarding the joint relation learning, we noticed that semantic correlations, which denote the similarities of relations in semantics, can serve as a bridge to connect the learning of many-, few-, and zero- shot relations. Take Figure 1 as an instance. The many-shot relation “/film/film/language”, few-shot relation “/film/film/edited”, and zero-shot relation “/film/film/sequel” are all related to “film”. Based on the assumption that semantically similar relations should be located near each other in embedding space (Yang et al. 2015), it makes sense to exploit semantic correlations, such as the one in the above-mentioned example, to accomplish the joint relation learning. Inspired by this, we propose a Generalized Relation Learning framework (abbreviated to GRL) based on learning semantic correlations. GRL can be plugged into a mainstream link prediction model to make it (1) insensitive to unbalanced relation distributions and (2) capable of learning zero-shot relations. Specifically, GRL is plugged into a link prediction model after the embedding learning stage. To optimize the relation embedding, GRL extracts rich semantic correlations through an attention mechanism, fuses different relations, and minimizes the classification-aware loss to enable these implicitly embedded semantic correlations in the relation embeddings. Then, the closeness of semantically similar relations in vector space and the discrimination of dissimilar relations can be improved. In this way, few-shot relations can learn knowledge from the semantically similar many-shot relations; for zero- shot relations, their most semantically similar relation can also be predicted. In our experiments, we improve two base models (DistMult (Yang et al. 2015) and ConvE (Dettmers et al. 2018)) by incorporating the proposed GRL framework on all relation classes, i.e., many, few, and zero-shot relations. Our work is an important step towards a holistic understanding of KGs and a generalized solution of relation learning for the link prediction task. Our contributions are as follows: * $\bullet$ We carefully consider two key problems of the embedding learning used by mainstream link prediction models and we highlight the necessity of jointly learning many-, few-, and zero- shot relations. * $\bullet$ We propose GRL framework by leveraging the rich semantic correlations between relations to make the link prediction models insensitive to unbalanced relation distributions and capable of learning zero-shot relations. * $\bullet$ We perform experiments on six benchmarks to evaluate the link prediction capability of GRL, and show that GRL lets the base link prediction models perform well across many-, few-, and zero- shot relations. ## Related Work Since KGs are populated by automatic text processing they are often incomplete, and it is usually infeasible to manually add to them all the relevant facts. Hence, many researches approached the task of predicting missing links in KGs. Mainstream link prediction models widely use embedding-based methods to map entities and relations into continuous low-dimensional vector space and use a score function to predict whether the triples are valid. They can be broadly classified as translational based (Bordes et al. 2013; Wang et al. 2014; Lin et al. 2015; Ji et al. 2016), multiplicative based (Nickel, Tresp, and Kriegel 2011; Yang et al. 2015; Trouillon et al. 2016), and neural network based (Dettmers et al. 2018; Schlichtkrull et al. 2018). These models are based on the implicit assumption that all relations are distributed within the dataset in a balanced way. Hence, they perform poorly in few-shot relation learning scenarios because these models neglect the imbalanced distributions, as well as they cannot properly handle zero-shot relations due to keeping only the knowledge of existing relations and not learning information on unseen relations. Few-shot relation learning models attempt to adopt the meta-learning (Chen et al. 2019a; Lv et al. 2019) and metric-based (Xiong et al. 2018; Wang et al. 2019) methods to learn knowledge from only a few samples. However, the few- shot learning models are computationally expensive because they need to spend extra time retraining on each few-shot relation (meta-learning), or need to compare the few-shot relations one by one (metric-based). In practice, the many-shot and few-shot scenarios are not explicitly distinguished. Zero-shot relation learning models aim to learn relations that are unseen in the training set. Researchers have proposed several models to deal with zero-shot relations by leveraging information from textual descriptions (Chen et al. 2019b; Qin et al. 2020). They perform well on predicting the zero-shot relations, but are not appropriate in the link prediction task because textual descriptions could be noisy and a bridge connecting seen and unseen relations could be missing. Figure 2: The illustration of GRL, which consists of the intuitive explanation (a), the base model (b) and the detailed architecture (c). The base model denotes the mainstream link prediction model. GRL is plugged after the embedding component of the base model and contains three components: attention, fusion, and classifier. In this work, we focus on jointly learning many-, few-, and zero- shot relations without requiring extra textual knowledge. Recently, some computer vision works (Ye et al. 2019; Shi et al. 2019) have attempted to approach the generalized image classification. Nonetheless, they are not designed for coping with graph structures, e.g., KGs. We leverage in this work the rich semantic correlations between relations as a bridge to connect the learning of many-, few-, and zero- shot relations. Zhang et al. (2019) integrated the rich semantic correlations between specific hierarchical relations into relation extraction. That method performs well only on hierarchical relations, as well as it predicts relations from text, hence it does not cope with the link prediction task. ## Method Figure 2 provides the illustration of the proposed framework GRL. The figure consists of three parts: the intuitive explanation of GRL in Figure 2 (a), base model shown in Figure 2 (b), and the detailed architecture in Figure 2 (c). The intuitive explanation of GRL is shown to utilize the semantic correlations between many-shot and few-shot relations so that the relation embedding learning can benefit from semantically similar relations. We devise three modules, i.e., Attention, Fusion and Classifier, to embed and fuse the rich semantic correlations among many-shot and few-shot relations in the training phase; and to select the most similar relation embedding for zero-shot relations in the testing phase. In this way, GRL can improve the performance on all relation classes, i.e., many, few, and zero-shot relations. The base model denotes the existing mainstream link prediction model consisting of an embedding component and a score function component. GRL can be plugged between the embedding and the score function components to make it (1) insensitive to imbalanced relation distributions and (2) capable of detecting zero-shot relations. Before delving into the model description, we first formally represent a KG as a collection of triples $\mathcal{T}=\left\\{(e_{h},r,e_{t})|e_{h}\in\mathcal{E},e_{t}\in\mathcal{E},r\in\mathcal{R}\right\\}$, where $\mathcal{E}$ and $\mathcal{R}$ are the entity and relation sets, respectively. Each directed link in KG represents a triple (i.e., $e_{h}$ and $e_{t}$ are represented as the nodes and $r$ as the labeled edge between them). The link prediction task is to predict whether a given triple $(e_{h},r,e_{t})$ is valid or not. In particular, for the zero-shot relations, we need to emphasize that we mainly focus on predicting the validity of the triple with a zero-shot relation, rather than predicting the zero-shot relations, i.e., the relation prediction task (Chen et al. 2019b; Qin et al. 2020). However, GRL has also the ability to predict the most semantically similar relation of a given zero-shot relation through learning from the many- and few-shot relations, not from the text description. ### Base Model We select a mainstream link prediction model as the base model and apply GRL to it. The base model can be seen as multi-layer neural network consisting of an embedding component and a score function component. For the base link prediction model, given an input triple $(e_{h},r,e_{t})$, the embedding component maps the head and tail entities $(e_{h},e_{t})$ and the relation $r$ to their distributed embedding representations $(\bm{e}_{h},\bm{r},\bm{e}_{t})$ through the entity and relation embedding layers, respectively. After the embedding representations are obtained, the score function component is adopted to calculate the likelihood of $(\bm{e}_{h},\bm{r},\bm{e}_{t})$ being a valid fact. The following binary cross entropy loss is used to train model parameters: $\mathcal{L}_{s}=-\frac{1}{N}\sum_{i=1}^{N}(t_{i}\log p(s_{i})+(1-t_{i})\log(1-p(s_{i}))),$ (1) where $s_{i}$ is the score of the $i$-th input triple, $t_{i}$ is the ground truth label, $t_{i}$ is 1 if the input relation is valid and 0 otherwise, and $N$ is the number of input triples. ### GRL Framework The loss used by mainstream link prediction models is score-oriented and lacks an in-depth exploration of rich semantic correlations in KGs. We propose the GRL framework to learn appropriate representations for relations by embedding semantic correlations into classification-aware optimization. GRL contains three specific modules: 1) Attention Module, which builds the knowledge-aware attention distribution and the relational knowledge vector. The aim of this module is to extract the semantic correlations and the degree of these correlations. 2) Fusion Module, which fuses the relational knowledge vector with the joint vector obtained from the attention module. This module realizes the fusion of different relations, according to semantic correlations. 3) Classifier Module, which calculates the classification-aware loss to implicitly enable the rich semantic correlations embedded in the embeddings. Thanks to it, both the compactness of semantically similar relations and discrimination of dissimilar relations can be enhanced. The following is a detailed introduction to each module. Attention Module Joint Block. The classification-aware loss is calculated by the relation classification results based on the head and tail entities from the given triple $(e_{h},r,e_{t})$. Inspired by (Qin et al. 2020), the joint vector of the head and tail entities has the ability to represent the potential relation between them. The head and tail entities representations (i.e., $\bm{e}_{h}$ and $\bm{e}_{t}$) are jointed together at the joint block for which we adopt three different alternatives: $\bm{j}=\left\\{\begin{array}[]{ll}\bm{e}_{h}-\bm{e}_{t},&sub\\\ \bm{e}_{h}\otimes\bm{e}_{t},&multiply\\\ W_{1}[\bm{e}_{h};\bm{e}_{t}]+b_{1},&concat\end{array},\right.$ (2) where $\otimes$ denotes the element-wise multiplication operator, and $W_{j}$ and $b_{j}$ are the learnable parameters. Relation Memory Block. Using a memory block to store class information is widely used in image classification (Snell, Swersky, and Zemel 2017; Karlinsky et al. 2019; Liu et al. 2019). Following these studies, we design a relation memory block to store all relation information by sharing parameters with the relation embedding layer as $\bm{M}=\left\\{\bm{r}_{1},\bm{r}_{2},...,\bm{r}_{K-1},\bm{r}_{K}\right\\},$ (3) where $M\in\mathbb{R}^{K\times dim}$, $K$ is the number of relation classes. As the training progresses, the relation embedding layer and relation memory block are updated synchronously. Relational Knowledge. To realize the classification-aware optimization objective, we extract useful relational knowledge from the relation memory block to enrich the joint vector. The semantic correlation degree between different relations may vary; thus, we adopt the attention mechanism to customize specific relational knowledge for each joint vector. Concretely, the relational knowledge vector $\bm{rk}$ is computed as a weighted sum of each relation representation in the relation memory block $M$, i.e., $\bm{rk}=\alpha_{sim}\bm{M}$, where $\alpha_{sim}\in\mathbb{R}^{K}$ represents the knowledge-aware attention distribution. Attention Distribution The knowledge-aware attention distribution $\alpha_{sim}$ describes the similarity between the joint vector and each relation representation in the relation memory block. We estimate $\alpha_{sim}$ as $\alpha_{sim}=softmax(\bm{j}\bm{M}^{\top}),$ (4) where $softmax$ is the activation function, and $\bm{M}^{\top}$ represents the transposed matrix of $\bm{M}$. Note that the attention value of the ground- truth relation is masked with 0. Fusion Module In this module, the joint vector and relational knowledge vector are fused. Intuitively, the proportion of fusion is different for each joint vector. Inspired by the pointer-generator network (See, Liu, and Manning 2017) that facilitates copying words from the source text during new words generation, we propose a soft switch, that is, the fusion probability $p_{f}\in[0,1]$, to adaptively adjust the fusion proportion between the joint vector and relational knowledge vector. The fusion probability $p_{f}$ is estimated according to the joint vector as $p_{f}=sigmoid(FC(\bm{j})),$ where $FC$ is the fully connected neural network, and $sigmoid$ is the activation function. Finally, we obtain the following fusion vector $\bm{f}$ over the joint vector $\bm{j}$ and relational knowledge vector $\bm{rk}$ as $\bm{f}=(1-p_{f})\bm{j}+p_{f}\bm{rk}.$ (5) Classifier Module Classification-aware Loss. The fusion vector $\bm{f}$ is mapped to a class probability through the classifier block as $D\sim softmax(\bm{f}^{\top}W_{c}),$ (6) where $W_{c}\in\mathbb{R}^{dim\times K}$ is the classification weight matrix, and $softmax$ is the activation function. Given the ground truth relation $r_{i}$ from the $i$-th input $(e_{h_{i}},r_{i},e_{t_{i}})$, we adopt cross entropy to assess the classification-aware loss as $\mathcal{L}_{c}=-\frac{1}{N}\sum_{i=1}^{N}\log p(r_{i}|(e_{h_{i}},e_{t_{i}})),$ (7) where $p(r_{i}|(e_{h_{i}},e_{t_{i}}))\in D_{i}$ is the probability of the ground truth relation $r_{i}$. | $\mathcal{\left|E\right|}$ | $\mathcal{\left|R\right|}$ | Train | Valid | Test ---|---|---|---|---|--- YAGO3-10 | 123k | 37 | 1M | 5k | 5k FB15K-237 | 15k | 237 | 273k | 18k | 20k NELL-995 | 75k | 200 | 150k | 543 | 4k Kinship | 104 | 25 | 9k | 1k | 1k WN18 | 41k | 18 | 141k | 5k | 5k NELL-ONE | 69k | 358 | 190k | 1k | 2k Table 1: Statistics of datasets. $\mathcal{\left|E\right|}$ and $\mathcal{\left|R\right|}$ represent the cardinalities of the entity and relation sets. | YAGO3-10 | FB15K-237 | NELL-995 | Kinship | WN18 ---|---|---|---|---|--- | MRR | HITS@N | MRR | HITS@N | MRR | HITS@N | MRR | HITS@N | MRR | HITS@N | | @10 | @1 | | @10 | @1 | | @10 | @1 | | @10 | @1 | | @10 | @1 ComplEx | 36.0 | 55.0 | 26.0 | 24.7 | 42.8 | 15.8 | 48.2 | 60.6 | 39.9 | 82.3 | 97.1 | 73.3 | 94.1 | 94.7 | 93.6 R-GCN | - | - | - | 24.8 | 41.7 | 15.3 | 12.0 | 18.8 | 8.2 | 10.9 | 23.9 | 3.0 | 81.4 | 96.4 | 69.7 ConvKB | - | - | - | 28.9 | 47.1 | 19.8 | 43.0 | 54.5 | 37.0 | 61.4 | 95.3 | 43.6 | - | - | - D4-STE | 47.2 | 64.3 | 38.1 | 32.0 | 50.2 | 23.0 | - | - | - | - | - | - | 94.6 | 95.2 | 94.2 D4-Gumbel | 38.8 | 57.3 | 29.4 | 30.0 | 49.6 | 20.4 | - | - | - | - | - | - | 94.6 | 95.4 | 94.2 DistMult | 34.0 | 54.0 | 24.0 | 24.1 | 41.9 | 15.5 | 48.5 | 61.0 | 40.1 | 51.6 | 86.7 | 36.7 | 82.2 | 93.6 | 72.8 +GRL | 41.2 | 59.9 | 31.1 | 25.8 | 43.9 | 16.9 | 54.3 | 64.6 | 47.6 | 52.2 | 86.4 | 37.3 | 86.1 | 95.2 | 79.2 ($\pm$ sd) | (0.3) | (1.0) | (0.1) | (0.2) | (0.3) | (0.1) | (0.2) | (0.3) | (0.3) | (0.2) | (0.8) | (0.2) | (1.0) | (0.4) | (1.1) ConvE | 52.0 | 66.0 | 45.0 | 31.6 | 49.1 | 23.9 | 49.1 | 61.3 | 40.3 | 83.3 | 98.1 | 73.8 | 94.2 | 95.5 | 93.5 +GRL | 55.4 | 69.0 | 47.4 | 32.6 | 50.2 | 24.8 | 49.4 | 60.6 | 41.5 | 83.4 | 97.8 | 74.5 | 94.8 | 95.7 | 94.4 ($\pm$ sd) | (1.0) | (0.1) | (0.1) | (0.3) | (0.2) | (0.3) | (0.2) | (0.3) | (0.3) | (0.2) | (0.5) | (0.5) | (0.1) | (0.4) | (0.0) Table 2: Link prediction results (mean $\pm$ sd) of the compared models (%): the best results are marked in bold (pairwise t-test at 5% significance level). | YAGO3-10 | NELL-995 ---|---|--- | Many | Few | All | Many | Few | All DistMult | 38.1 | 26.7 | 34.0 | 52.6 | 41.9 | 48.5 DistMult+GRL | 44.8 | 34.2 | 41.2 | 57.3 | 48.8 | 54.3 (Increment) | ($\uparrow$6.7) | ($\uparrow$7.5) | ($\uparrow$7.2) | ($\uparrow$4.7) | ($\uparrow$6.9) | ($\uparrow$5.8) ConvE | 57.9 | 20.0 | 52.4 | 52.0 | 42.2 | 49.1 ConvE+GRL | 59.4 | 24.6 | 55.4 | 52.4 | 43.9 | 49.4 (Increment) | ($\uparrow$1.5) | ($\uparrow$4.6) | ($\uparrow$3.0) | ($\uparrow$0.4) | ($\uparrow$1.7) | ($\uparrow$0.3) Table 3: Link prediction results with the increment (%) on many-shot and few- shot sub-groups, and entire test set. Most Similar Relation. Existing mainstream link prediction models have achieved impressive performance, yet they can only learn the patterns observed in the closed datasets, thereby limiting their scalability for handling the rapidly evolving KGs. Specifically, when a zero-shot relation $r_{z}$ (i.e., one not existing in the training set) occurs between an entity pair $(e_{h},e_{t})$, it is almost impossible for the existing models to distinguish whether this new triple $(e_{h},r_{u},e_{t})$ is valid or not. All $r_{z}$ will be then identified as an ‘unknown‘ vector $\bm{u}$ by the embedding component, and the newly constructed triple representation $(\bm{e}_{h},\bm{u},\bm{e}_{t})$ will receive a low score. To alleviate this defect, GRL selects the most semantically similar relation for replacing to enhance the learning ability of base model on zero-shot relations. We argue that the relation which corresponds to the maximum similarity in $\alpha_{sim}$ reflects the semantic relation of two entities in the best way. Therefore, we use the vector of the most similar relation $\bm{r}_{ms}$ to replace the vector $\bm{u}$ and evaluate the newly constructed triple representation $(\bm{e}_{h},\bm{r}_{ms},\bm{e}_{t})$. ### Learning Scheme We follow the definition of score-aware loss in existing base models and propose a classification-aware loss to train the model. The overall optimization follows the joint learning paradigm that is defined as a weighted combination of constituent losses as $\mathcal{L}=\mathcal{L}_{s}+\lambda\mathcal{L}_{c},$ where $\lambda$ is a hyper-parameter to balance the importance between the score-aware loss and classification-aware loss for optimization. ## Experiments and Results ### Datasets We select two categories of datasets to comprehensively evaluate GRL as follows, whose statistical descriptions are shown in Table 1: * $\bullet$ Imbalanced datasets: YAGO3-10 (Mahdisoltani, Biega, and Suchanek 2015), FB15K-237 (Toutanova and Chen 2015), NELL-995 (Xiong, Hoang, and Wang 2017), Kinship (Lin, Socher, and Xiong 2018), and WN18 (Bordes et al. 2013). These datasets contain both many-shot and few-shot relations. * $\bullet$ Few-shot dataset: NELL-ONE (Xiong et al. 2018), which is specially constructed for the few-shot learning task in KG. The relations with less than 500 but more than 50 training triples are selected as testing data. ### Baselines We adopt two embedding-based models, DistMult (Yang et al. 2015) and ConvE (Dettmers et al. 2018), as the base models of our proposed modules, and compare the two enhanced models with the following popular relation prediction models: RESCAL (Nickel, Tresp, and Kriegel 2011), TransE (Bordes et al. 2013), DistMult (Yang et al. 2015), ComplEx (Trouillon et al. 2016), ConvE (Dettmers et al. 2018), ConvKB (Nguyen et al. 2018), D4-STE, D4-Bumbel (Xu and Li 2019), and TuckER (Balazevic, Allen, and Hospedales 2019). Besides the above general models, we test two additional models, GMatching (Xiong et al. 2018) and CogKR (Du et al. 2019), which are designed specifically for the few-shot relation learning. ### Experimental Configuration We implement the base models and our proposed two modules in PyTorch (Paszke et al. 2017) . Throughout the experiments, we optimize the hyperparameters in a grid search setting for the best mean reciprocal rank (MRR) on the validation set. We use Adam to optimize all the parameters with initial learning rate at 0.003. The dimensions of entity and relation embeddings are both set to 200. The loss weight $\lambda$ is set to 0.1. According to the frequency of relations, we take the top 20% and bottom 80% of relations as many-shot and few-shot relation classes, respectively. The experimental results of our model are averaged across three training repetitions, and standard deviations (sd) are also reported. ### Experiment I: Link Prediction #### Setting We follow the evaluation protocol of (Dettmers et al. 2018): each input $(e_{h},r,e_{t})$ is converted to two queries, that is, tail query $(e_{h},r,?)$ and head query $(?,r,e_{t})$; then, the ranks of correct entities are recorded among all entities for each query, excluding other correct entities that were observed in any of the train/valid/test sets for the same query. We use the filtered HITS@1, 5, 10, and MRR as evaluation metrics. #### Results Table 2 records the results on five imbalanced datasets, which reflect the general performance of the compared models in solving the link prediction task. It shows that two base models (DistMult and ConvE) are generally improved by incorporating the proposed GRL framework. That is, GRL improves DistMult by an average of 3.84% and improves ConvE by an average of 1.08% under the MRR evaluation. Especially, the enhanced model ConvE+GRL generally outperforms the compared models on YAGO3-10, FB15K-237, Kinship, and WN18, and the enhanced model DistMult+GRL also performs well on NELL-995. We also evaluate the performance of GRL in learning many-shot and few-shot relations and show the MRR results of DistMult, DistMult+GRL, ConvE, and ConvE+GRL on YAGO3-10 and NELL-995 (c.f. Table 3). The results indicate that GRL achieves consistent improvements on both “many-shot” and “few-shot” sub-groups. We assume this may be because handling many-shot relations can be improved thanks to useful implicit information from few-shot relations, even though there are already numerous training samples for many-shot relations. From this aspect, it is sensible for the mainstream link prediction models to rely on GRL regarding the imbalanced relation issue. | MRR | HITS@N ---|---|--- | | @10 | @5 | @1 TransE† | 9.3 | 19.2 | 14.1 | 4.3 GMatching† | 18.8 | 30.5 | 24.3 | 13.3 CogKR∗ | 25.6 | 35.3 | 31.4 | 20.5 DistMult† | 10.2 | 17.7 | 12.6 | 6.6 DistMult+GRL | 14.4 | 23.0 | 18.2 | 9.8 ($\pm$ sd) | (2.0) | (2.1) | (1.9) | (2.3) ConvE∗ | 17.0 | 30.6 | 23.0 | 10.5 ConvE+GRL | 25.6 | 38.9 | 33.6 | 18.8 ($\pm$ sd) | (2.3) | (3.7) | (3.1) | (2.1) Table 4: Few-shot relation learning results (mean $\pm$ sd) on NELL-ONE dataset (%): the results marked by ‘$\dagger$’ or ‘$\ast$’ are taken from (Xiong et al. 2018; Du et al. 2019). ### Experiment II: Few-shot Relation Learning #### Setting To further evaluate the performance of GRL in the few-shot relation learning case, which is tricky for a link prediction model, especially, when relations are very insufficient, we test approaches on the NELL-ONE dataset wherein each test relation has only one instance in the training set. We follow the evaluation protocol and metrics of (Xiong et al. 2018). #### Results Table 4 shows that GRL consistently improves both of the base models by average 4.2% and 8.6% MRR scores. Especially for ConvE, incorporating GRL helps it outperform the other approaches on three metrics. CogKR, a path learning based model, performs best under HITS@1. The reason might be that the testing query is easy to be completed by finding KG paths on the few-shot relation datasets, such as NELL-ONE. Although there is only one training instance for each testing query, GRL can effectively embed the few-shot relations by learning from the semantically similar relations in the many-shot class. ### Experiment III: Zero-shot Relation Learning #### Setting To evaluate the performance on zero-shot relations of GRL, we construct a testing set containing 500 triples whose relations are unseen in the training phase. The testing triples are randomly sampled from the FB15K dataset (Bordes et al. 2013), and the training set is FB15K-237 to ensure the authenticity of the triples. We adopt the fundamental testing protocol that quantitatively determines the scores of triples with zero-shot relations. Most of existing zero-shot relation studies have to depend on textual descriptions, while the zero-shot learning addressed in this work does not require this information. Therefore, we select the GMatching model (Xiong et al. 2018) for comparison, which can predict similar relations by learning a matching metric without any additional information. We use the classical method TransE (Bordes et al. 2013) to learn the relation embeddings in the FB15K dataset and calculate the similarity between the zero-shot relation and the predicted relation. Figure 3: Zero-shot relation learning results: (a) the average score of the testing triples, and (b) the average similarity between the zero-shot relation with the predicted relation. #### Results Figure 3 (a) demonstrates results of the average score of the testing triples with zero-shot relations. Note that we use the fusion vector as the zero-shot relations embedding. We can see that two base models (DistMult and ConvE) cannot get a good average score because all zero-shot relations will be identified as an ‘unknown’ relation. When GRL is plugged, two enhanced models (DistMult+GRL and ConvE+GRL) are both boosted in learning positive relations, proving that the GRL framework can effectively improve the validation capabilities on triples with zero-shot relations of the base models. Figure 3 (b) shows the performance on predicting zero-shot relations. We can see that the base models perform worse due to their superficial way of embedding zero- shot relations as mentioned before. When equipping with GRL, the enhanced models perform better than GMatching, indicating that learning from the semantic correlations between unseen relations and seen relations provides a comparably good way as learning from neighbor information. | | YAGO3-10 | NELL-ONE ---|---|---|--- (1) | ConvE | 52.0 | 17.0 (2) | ConvE+GRL($p_{f}=0$) | 52.6 | 23.3 (3) | ConvE+GRL($p_{f}=0.5$) | 53.9 | 24.7 (4) | ConvE+GRL($p_{f}=1$) | 52.2 | 20.3 (5) | ConvE+Direct | 51.2 | 10.5 (6) | ConvE+GRL | 55.4 | 25.6 Table 5: Ablation Study. ## Further Analysis of GRL ### Ablation Study Study of Fusion Probability To assess the effect of the fusion vector, we make a comparison on three variants from the fusion probability perspective based on ConvE, see Table 5 (2)-(4). The three variants are the followings: only using the joint vector (i.e., $p_{f}=0$), only using the relational knowledge vector (i.e., $p_{f}=1$), and using the joint and relational knowledge vectors with an equal weight (i.e., $p_{f}=0.5$). Compared with three variants, fusing the joint and relational knowledge vectors (i.e., ConvE+GRL) performs best, which suggests the semantic correlations in the relational knowledge vectors can help the base model learn more appropriate representations of relations and thus boost the general performance. Moreover, the adaptive fusion probability can improve the flexibility of the fusion operator. Direct Fusion vs. GRL We test now a direct fusion method that fuses the relational knowledge vector with the relation representation as the updated relation representation without considering the classification-aware loss. Table 5 (5) shows the MRR performance of ConvE when enhanced by the direct method. Rich semantic correlations in KGs cannot be adequately learned by the direct method because it simply leverage the superficial semantic correlations, rather than embedding them into relation vectors. Moreover, the direct method will make embedding learning more confusing especially for the few-shot relation data such as NELL-ONE. ### Case Study Visualization of Knowledge-aware Attention GRL is able to make the base model fully learn semantic correlations between relations. To verify this, we display the attention distribution for the base model (ConvE) and enhanced model (ConvE+GRL) on FB15K-237 in Figure 4, and show the average attention distribution of 237 relation classes where each row represents a type of relation. The base model learns little about semantic correlations between relations, while the enhanced model (ConvE+GRL) can perfectly capture the semantic correlations. The attention distribution of few-shot relations is more discrete than many-shot relations due to insufficient training data. Figure 4: Case study: knowledge-aware attention cases with a heat map. Figure 5: Case study: t-SNE visualization of relation embeddings in FB15K-237 (better view in color). The semantically similar relations get closer after plugging GRL. Visualization of Relation Embedding In addition, we also show in Figure 5 the t-SNE (Maaten and Hinton 2008) plot of all relations on FB15K-237 in embedding space. To provide more insights we highlight the relations associated with “film”. The Stars and Triangles represent the many-shot and few-shot relations, respectively. We can see that the many-shot and few-shot relations are more compact in the case of the enhanced model than the base model . ## Conclusion and Future Work In this work, we study two natural problems in the link prediction task: 1) unbalanced relation distribution, and 2) unseen relations. To address them, we focus on generalized relation learning and propose a framework, GRL, that uses semantic correlations among relations as a bridge to connect semantically similar relations. Through extensive experiments on six datasets, we demonstrate the effectiveness of GRL, providing a comprehensive insight into the generalized relation learning of KGs. There are a few loose ends for further investigation. We will consider combining the external text information and the semantic knowledge of KGs to facilitate the relation learning. We will also try to deploy GRL to downstream applications that involve generalized relation learning scenarios to gain more insights. ## Acknowledgments This work was supported in part by the Ministry of Education of Humanities and Social Science project under grant 16YJC790123 and the Natural Science Foundation of Shandong Province under grant ZR2019MA049. ## References * Balazevic, Allen, and Hospedales (2019) Balazevic, I.; Allen, C.; and Hospedales, T. 2019. TuckER: Tensor Factorization for Knowledge Graph Completion. In _Proceedings of EMNLP-IJCNLP_ , 5184–5193. * Bordes et al. (2013) Bordes, A.; Usunier, N.; Garcia-Durán, A.; Weston, J.; and Yakhnenko, O. 2013. Translating Embeddings for Modeling Multi-relational Data. In _Proceedings of NeurIPS_ , 2787–2795. * Chen et al. (2019a) Chen, M.; Zhang, W.; Zhang, W.; Chen, Q.; and Chen, H. 2019a. Meta Relational Learning for Few-Shot Link Prediction in Knowledge Graphs. In _Proceedings of EMNLP-IJCNLP_ , 4216–4225. * Chen et al. (2019b) Chen, W.; Zhu, H.; Han, X.; Liu, Z.; and Sun, M. 2019b. Quantifying Similarity between Relations with Fact Distribution. In _Proceedings of ACL_ , 2882–2894. * Christmann et al. (2019) Christmann, P.; Saha Roy, R.; Abujabal, A.; Singh, J.; and Weikum, G. 2019. Look before you Hop: Conversational Question Answering over Knowledge Graphs Using Judicious Context Expansion. In _Proceedings of CIKM_ , 729–738. * Dettmers et al. (2018) Dettmers, T.; Minervini, P.; Stenetorp, P.; and Riedel, S. 2018. Convolutional 2D Knowledge Graph Embeddings. In _Proceedings of AAAI_ , 1811–1818. * Du et al. (2019) Du, Z.; Zhou, C.; Ding, M.; Yang, H.; and Tang, J. 2019. Cognitive Knowledge Graph Reasoning for One-shot Relational Learning. _arXiv preprint arXiv:1906.05489_ . * Ji et al. (2016) Ji, G.; Liu, K.; He, S.; and Zhao, J. 2016. Knowledge Graph Completion with Adaptive Sparse Transfer Matrix. In _Proceedings of AAAI_ , 985–991. * Karlinsky et al. (2019) Karlinsky, L.; Shtok, J.; Harary, S.; Schwartz, E.; Aides, A.; Feris, R.; Giryes, R.; and Bronstein, A. M. 2019. RepMet: Representative-Based Metric Learning for Classification and Few-Shot Object Detection. In _Proceedings of CVPR_ , 5197–5206. * Lei et al. (2020a) Lei, W.; He, X.; de Rijke, M.; and Chua, T.-S. 2020a. Conversational Recommendation: Formulation, Methods, and Evaluation. In _Proceedings of SIGIR_ , 2425–2428. * Lei et al. (2020b) Lei, W.; Zhang, G.; He, X.; Miao, Y.; Wang, X.; Chen, L.; and Chua, T.-S. 2020b. Interactive Path Reasoning on Graph for Conversational Recommendation. In _Proceedings of KDD_ , 2073–2083. * Lin, Socher, and Xiong (2018) Lin, X. V.; Socher, R.; and Xiong, C. 2018. Multi-Hop Knowledge Graph Reasoning with Reward Shaping. In _Proceedings of EMNLP_ , 3243–3253. * Lin et al. (2015) Lin, Y.; Liu, Z.; Sun, M.; Liu, Y.; and Zhu, X. 2015. Learning Entity and Relation Embeddings for Knowledge Graph Completion. In _Proceedings of AAAI_ , 2181–2187. * Liu et al. (2019) Liu, Z.; Miao, Z.; Zhan, X.; Wang, J.; Gong, B.; and Yu, S. X. 2019. Large-Scale Long-Tailed Recognition in an Open World. In _Proceedings of CVPR_ , 2537–2546. * Lv et al. (2019) Lv, X.; Gu, Y.; Han, X.; Hou, L.; Li, J.; and Liu, Z. 2019. Adapting Meta Knowledge Graph Information for Multi-Hop Reasoning over Few-Shot Relations. In _Proceedings of EMNLP-IJCNLP_ , 3374–3379. * Maaten and Hinton (2008) Maaten, L. v. d.; and Hinton, G. 2008. Visualizing Data Using T-SNE. _Journal of Machine Learning Research_ 2579–2605. * Mahdisoltani, Biega, and Suchanek (2015) Mahdisoltani, F.; Biega, J.; and Suchanek, F. 2015. YAGO3: A Knowledge Base from Multilingual Wikipedias. In _Proceedings of CIDR_. * Moon et al. (2019) Moon, S.; Shah, P.; Kumar, A.; and Subba, R. 2019. OpenDialKG: Explainable Conversational Reasoning with Attention-based Walks over Knowledge Graphs. In _Proceedings of ACL_ , 845–854. * Nguyen et al. (2018) Nguyen, D. Q.; Nguyen, T. D.; Nguyen, D. Q.; and Phung, D. 2018. A Novel Embedding Model for Knowledge Base Completion Based on Convolutional Neural Network. In _Proceedings of NAACL_ , 327–333. * Nickel, Tresp, and Kriegel (2011) Nickel, M.; Tresp, V.; and Kriegel, H.-P. 2011. A Three-Way Model for Collective Learning on Multi-Relational Data. In _Proceedings of ICML_ , 809–816. * Paszke et al. (2017) Paszke, A.; Gross, S.; Chintala, S.; Chanan, G.; Yang, E.; DeVito, Z.; Lin, Z.; Desmaison, A.; Antiga, L.; and Lerer, A. 2017. Automatic Differentiation in PyTorch. In _Proceedings of NeurIPS Workshop_. * Qin et al. (2020) Qin, P.; Wang, X.; Chen, W.; Zhang, C.; Xu, W.; and Wang, W. Y. 2020. Generative Adversarial Zero-Shot Relational Learning for Knowledge Graphs. In _Proceedings of AAAI_ , 8673–8680. * Schlichtkrull et al. (2018) Schlichtkrull, M.; Kipf, T. N.; Bloem, P.; Van Den Berg, R.; Titov, I.; and Welling, M. 2018. Modeling Relational Data with Graph Convolutional Networks. In _Proceedings of ESWC_ , 593–607. * See, Liu, and Manning (2017) See, A.; Liu, P. J.; and Manning, C. D. 2017. Get To The Point: Summarization with Pointer-Generator Networks. In _Proceedings of ACL_ , 1073–1083. * Seyed and David (2018) Seyed, M. K.; and David, P. 2018. SimplE Embedding for Link Prediction in Knowledge Graphs. In _Proceedings of NeurIPS_ , 4284–4295. * Shi and Weninger (2018) Shi, B.; and Weninger, T. 2018. Open-World Knowledge Graph Completion. In _Proceedings of AAAI_ , 1957–1964. * Shi et al. (2019) Shi, X.; Salewski, L.; Schiegg, M.; Akata, Z.; and Welling, M. 2019. Relational Generalized Few-Shot Learning. _arXiv preprint arXiv:1907.09557_ . * Snell, Swersky, and Zemel (2017) Snell, J.; Swersky, K.; and Zemel, R. 2017. Prototypical Networks for Few-shot Learning. In _Proceedings of NeurIPS_ , 4077–4087. * Toutanova and Chen (2015) Toutanova, K.; and Chen, D. 2015. Observed Versus Latent Features for Knowledge Base and Text Inference. In _Proceedings of CVSC_ , 57–66. * Trouillon et al. (2016) Trouillon, T.; Welbl, J.; Riedel, S.; Gaussier, É.; and Bouchard, G. 2016. Complex Embeddings for Simple Link Prediction. In _Proceedings of ICML_ , 2071–2080. * Wang et al. (2021) Wang, X.; Huang, T.; Wang, D.; Yuan, Y.; and Chua, T. S. 2021. Learning Intents behind Interactions with Knowledge Graph for Recommendation . * Wang et al. (2019) Wang, Z.; Lai, K.; Li, P.; Bing, L.; and Lam, W. 2019. Tackling Long-Tailed Relations and Uncommon Entities in Knowledge Graph Completion. In _Proceedings of EMNLP-IJCNLP_ , 250–260. * Wang et al. (2014) Wang, Z.; Zhang, J.; Feng, J.; and Chen, Z. 2014. Knowledge Graph Embedding by Translating on Hyperplanes. In _Proceedings of AAAI_ , 1112–1119. * Xiong, Hoang, and Wang (2017) Xiong, W.; Hoang, T.; and Wang, W. Y. 2017. DeepPath: A Reinforcement Learning Method for Knowledge Graph Reasoning. In _Proceedings of EMNLP_ , 564–573. * Xiong et al. (2018) Xiong, W.; Yu, M.; Chang, S.; Guo, X.; and Wang, W. Y. 2018. One-Shot Relational Learning for Knowledge Graphs. In _Proceedings of EMNLP_ , 1980–1990. * Xu and Li (2019) Xu, C.; and Li, R. 2019. Relation Embedding with Dihedral Group in Knowledge Graph. In _Proceedings of ACL_ , 263–272. * Yang et al. (2015) Yang, B.; Yih, W.-t.; He, X.; Gao, J.; and Deng, L. 2015. Embedding Entities and Relations for Learning and Inference in Knowledge Bases. In _Proceedings of ICLR_. * Ye et al. (2019) Ye, H.-J.; Hu, H.; Zhan, D.-C.; and Sha, F. 2019. Learning Classifier Synthesis for Generalized Few-Shot Learning. _arXiv preprint arXiv:1906.02944_ . * Zhang et al. (2019) Zhang, N.; Deng, S.; Sun, Z.; Wang, G.; Chen, X.; Zhang, W.; and Chen, H. 2019. Long-tail Relation Extraction via Knowledge Graph Embeddings and Graph Convolution Networks. In _Proceedings of NAACL_ , 3016–3025. * Zhu et al. (2021) Zhu, F.; Lei, W.; Wang, C.; Zheng, J.; Poria, S.; and Chua, T.-S. 2021. Retrieving and Reading: A Comprehensive Survey on Open-domain Question Answering.
# An Overview of the State of the Art for Practical Quantum Key Distribution Daniel D. Moskovich Center for Quantum Information Science and Technology, Ben-Gurion University of the Negev, Israel (Date: 9 April, 2015) ###### Abstract. This is an overview of the state of the art for quantum key distribution (QKD) as of March 2015. It is written by a non-expert for non-experts. Additions and corrections are welcome. ## 1\. Introduction The goal of this overview to concisely summarize, in a way that is accessible to a non-expert, where practical Quantum Key Distribution (QKD) stands now in early 2015 and what seem to be promising directions for the near future to the best of the author’s knowledge and understanding. We begin with a general overview of what QKD is, followed by a discussion of the major practical QKD players at the moment, a discussion of protocols, and a discussion of photon sources, transmission, and detection. We conclude with a section on attacks against QKD. ## 2\. What is quantum key distribution? Quantum key distribution (QKD) uses principles of quantum information theory to ensure secure communication (Weisner, 1983; Bennett & Brassard, 1984; Ekert, 1991). Its goal is for two parties called (A)lice and (B)ob to share a secret key made up of $0$ s and $1$ s which they will later use to encrypt and decrypt communications between them. The information used to compose the key is carried between Alice and Bob on qubits (two state quantum systems). The _SECOCQ White Paper_ convincingly argued that QKD is a form of _trusted courier_ (Alice hands a message to somebody she trusts, who carries that to Bob), so that it is useful in contexts in which a trusted courier would be useful (SECOQC, 2007). With idealized hardware and with perfect accuracy, the advantages of QKD over classical trusted courier methods would be: * • Mathematically-proven security against all classical and quantum attacks. * • Alice and Bob can detect any active attempt by an eavesdropper (E)ve to eavesdrop on the key distribution process. The _Black Paper of Quantum Cryptography_ convincingly argued that real-life QKD security is less than perfect, so that each different QKD setup should be carefully and individually studied to assure its security (Scarani & Kurtsiefer, 2014). All security threats discovered so far have been redeemable, and we have no reason to believe that any given QKD setup cannot be made perfectly secure against all known attacks in principle. Real world QKD has become a focus of interest for industrial players, for governments, and for security agencies. ## 3\. Fundamental challenges A number of fundamental challenges to the widescale use of QKD have been identified (Pritchard & Till, 2010). 1. (1) Limited transmission rate and range. Both the range and the maximal bit-rate of QKD are low compared to classical communications. It is considered technologically impossible, for instance, to transmit a polarized photon reliably over more than 400km of fiber, although quantum repeaters will allow for longer range QKD. 2. (2) QKD protocols are fundamentally point-to-point, and do not integrate with packet-based protocols such as those used on the internet. 3. (3) QKD requires expensive special-purpose hardware such as single-photon sources and detectors. Such hardware is difficult to upgrade and to maintain. 4. (4) QKD addresses only one aspect of the security problem. For example authentication and integrity are not covered and must be handled classically. 5. (5) There is nothing fundamentally wrong with existing classical cryptographic techniques. Even if some classical ciphers (e.g. RSA) may be cracked using quantum algorithms at some unspecified point in the future, other classical ciphers are being developed that would be immune to quantum attacks. 6. (6) Because it is a new technology, there are potential discovered and undiscovered vulnerabilities in practical QKD systems. Indeed, several proposed conditions for unconditional security of practical implementations were found wanting and had to be revised, e.g. because key-length is finite while many security proofs assume infinite-length keys (Inamori, Lütkenhaus & Mayers, 2007; Tomamichel et al., 2012), or because the quantum effect of _locking_ might allow unexpectedly large information leak during the error- correction and privacy amplification steps (König et al., 2007; Iwakoshi & Hirota, 2014; Yuen, 2013; Portmann & Renner, 2014). Commercial QKD systems have been successfully hacked (Section 12.4). Because QKD is unconditionally secure while the security of any quantum protocol implementation is a probabilistic quantitative matter, these vulnerabilities will in principle never be fatal flaws. But each new vulnerability might require re-tuning of parameters and modifications to technological implementation. ## 4\. Advantages of QKD 1. (1) QKD provides the possibility to establish a secret key in a way that is provable secure against eavesdropping. Moreover, QKD can be composed with other encryptions, so as to provide an additional layer of security for an already secure message. For example a message encrypted using an RSA public key may be once again encrypted using a quantum key. To intercept the message, an attacker would have to break both the quantum key and the classical key. 2. (2) Eavesdropping can be detected, following which countermeasures may be adopted. This capability distinguished QKD from among all encryption methods. 3. (3) Expertise and knowledge gained in QKD research will, in large part, be useful for developing future technologies in future manifestations of the coming quantum revolution predicted by Michael Berry when he said, “It is easy to predict that in the twenty-first century, it will be quantum mechanics that influences all our lives.” Berry (1998). ## 5\. World QKD projects ### 5.1. Large scale networks Microwave apparatus used in quantum experiments. Retrieved from http://www.foxnews.com/tech/2013/05/08/quantum-network-secretly-running- for-2-years/ In the last 10 years, a number of multi-user QKD networks have been constructed. All use relay between trusted nodes and optical switching. The first of these was the $10$–node DARPA Quantum Network which has been operating since 2004 (Elliot et al., 2005). It uses active optical switching (i.e. an electrically powered switching device similar to a router) to distribute the key between pairs of nodes. It is being developed by BBN Technologies, Harvard University, Boston University and QinetiQ, with the support of the US Defense Advanced Research Projects Agency (DARPA). The SECOQC (Secure Communication Based on Quantum Cryptography) Quantum Network is an EU project which integrated several different QKD systems into one quantum backbone (QBB) network, developing a cross-platform interface (http://www.secoqc.net/). This provided impetus for the European Telecommunications Standards Institute (ETSI) to launch an industry specification group for QKD (ISG-QKD)in order to create universally accepted QKD standards (ETSI, 2015). The Swiss Quantum Network and the Durban Network are testing long-term QKD operation in field environments (http://swissquantum.idquantique.com and (Mirza & Petruccione, 2010)). Transparent network implementation of QKD using only beam splitters, which facilitate secure communication without requiring clients to be reconfigured, have been demonstrated by several groups (Telecordia Technologies, Universidad Politécnica de Madrid and Telefónica Investigación y Desarrollo, and two teams from the University of Science and Technology of China). The Tokyo QKD network used a central Key Management Service (KMS) and newer technologies to increase its speed to the point of transmitting a QKD-secured live teleconference between two nodes (Sasaki et al., 2011). This is suitable for government or municipal settings in which one central body controls the flow of information. Mitsubishi combined this system with an application for secure telephony to demonstrate QKD-secured mobile telephony (Mitsubishi, 2015). Finally, Los Alamos National Laboratory (LANL) runs a hub-and-spokes one-to-many quantum network (Hughes et al., 2013). The LANL photon generator has been miniaturized to around the size of a house key. China is currently constructing a 1200-mile line between Beijing and Shanghai as part of a proposed $20$-node QKD network which it aims to complete in 2016. Its current network, the Hefei-Chaohu-Wuhu wide area QKD network, is the largest in the world (Wang et al., 2014). To overcome the need to use trusted nodes, where one compromised node could impact the security of the entire network, there has been work aimed at using techniques of classical multiple access optical communication in the quantum context (Sarwar Pasha & Bala Ram, 2014). Such technologies have been applied for one part of the DARPA network, and also for an experimental three-node network at NIST (see http://www.nist.gov/itl/quantum/threeusernetwork.cfm). Taking the above technologies into account, the Engineering Science and Research Council (EPSRC) estimated in their 2014 report that hand-held QKD systems should be commercially available “with sufficient investment and encouragement” within 4–7 years, and that long-range highly-connected quantum networks should become available within 10–25 years (EPSRC, 2014). ### 5.2. University Centres There are a growing number of university centers in the world which specialize in quantum communication. We list a few of the most active. The world’s foremost dedicated quantum communications center is the Group of Applied Physics (GAP) at Geneva University (http://www.unige.ch/gap/quantum/) and their commercial spinoff company Id Quantique (http://www.idquantique.com/). They have developed what is today the world’s best single photon detector (Korzh _et al._ , 2014) with which they have achieved the current world record distance for QKD through fiber (Korzh _et al._ , 2015). They also produce and sell photon detectors and random number generators using patented technologies. The Centre for Quantum Technologies (CQT) in Singapore, directed by Ekert who developed the E91 protocol, specializes in quantum hacking http://www.quantumlah.org/. They have developed several successful attacks, which have taken advantage both of side-channels (e.g. (Lamas-Linares & Kurtsiefer, 2007)) and of erroneous parameters in security proofs (e.g. (Gerhardt et al., 2011)). The Institute of Quantum Computing (IQC) in Waterloo also has a research group working on QKD. It is directed by Norbert Lütkenhaus, who previously worked at MagiQ to develop practical QKD. Vadim Makarov of that group discovered some successful side-channel attacks against QKD (e.g. (Makarov _et al._ , 2006; Makarov, 2009)). The Key Laboratory of Quantum Information is China’s leading quantum information center, which is creating the world’s longest and most sophisticated QKD networks http://en.physics.ustc.edu.cn/research_9/Quantum/201107/t20110728_116550.html. Artist’s conception of quantum key distribution over free space ground to satellite quantum information links. Retrieved from http://www.esa.int/ESA ### 5.3. Commercial companies A number of commercial companies sell QKD systems and related devices. MagiQ Technologies in the US sells the QPN-8505, a QKD system which combines BB84 QKD with classical 3DES and AES encryption (http://www.magiqtech.com/). It works using decoy-state optimized BB84, with a secure key rate of 256Hz over 100km, or 140km with decoy states. In Europe the leading QKD company is Id Quantique, whose flagship product is the Clavis2, a pure QKD system (http://www.idquantique.com/). Clavis2 implements both BB84 and SARG04, with secure key-rates of around 1KHz on a 25km fiber. SeQureNet is a Paris-based company that produces QKD parts and that specialized is continuous variable (CV) QKD (http://sequrenet.com). Quintessence Labs in Australia provides true random number generators (http://www.quintessencelabs.com/). ## 6\. Protocols ### 6.1. BB84 The most widely used QKD protocol, which was also the world’s first QKD protocol, was developed by Bennett and Brassard in 1984, and is called BB84. It is typically divided into three layers: The physical layer in which the quantum communication is carried out,the key-extraction in which the actual key is extracted from the qubits that Alice sent to Bob, and the key- application layer where the secret key is used to encode a communication such as a telephone or a video conversation (Bennett et al., 1992; Gisin et al., 2002). In the physical layer, Alice sends random photons, $1$ with 50% probability and $0$ with 50% probability, either in the so-called $X$ basis or in the so- called $Z$ basis, each with 50% probability. Bob measures each bit he receives in a random basis, either in the $X$ basis with 50% probability, or in the $Y$ basis with 50% probability. This is the hardware-intensive portion of the protocol, for which good random-number generators, single-photon sources, and single-photon detectors are required. In the key-extraction layer, BB84 becomes classical. The first classical sublevel is called _sifting_. Alice and Bob both reveal which bases they used over a public channel. They then discard the bits which they measured in different bases. The second sublevel is called _authentication_. In it, Alice and Bob compare some of their sifted bits over the public channel to determine whether eavesdropping has occurred. If the bits they compare are more different than can be accounted by from random noise, then they can guess that Eve has eavesdropped, and adopt countermeasures. The reason that they can make this deduction is that Eve’s direct attack, intercept–resend REF, would involve measuring some of the bits sent by Alice, and sending them on to Bob. But since Eve does not know which basis was originally used by Alice, she will choose the wrong basis with 50% probability, and if she chooses the wrong basis then she will send the wrong qubit to Bob, and that incorrect qubit will survive into the sifted key with 50% probability. The third sublevel is called _error correction_. In it, Alice and Bob apply classical error-correcting algorithms to remedy the effect of random errors caused by channel noise and by the fact that equipment is non-ideal. The fourth and final level is called _privacy amplification_ , in which Alice and Bob apply classical cryptography algorithms to minimize the effect on the final key of any under-the-radar eavesdropping by Eve. In other words, security of a QKD key is always a quantitative affair because of non-ideal equipment and channel noise, so some non-trivial information might have been picked-up by Eve without being detected in the authentication phase. But the amount of leaked information is guaranteed to be below a certain threshold, and privacy amplification can negate the knowledge about the final key which that partial information imparts. The BB84 protocol. Figure retrieved from http://swissquantum.idquantique.com/?Key-Sifting. ### 6.2. Modified BB84 protocols The best known modification of BB84 is SARG04, which adapts it for use with attenuated laser pulses (Scarani et al., 2004). SARG04 is more robust than BB84 against so-called ‘coherent attacks’, but unfortunately it performs worse against certain ‘incoherent attacks’ (Branciard et al., 2005). Lo, Chau, and Ardehali presents a modification of BB84 which essentially doubles its efficiency (Lo _et al._ , 2005b). The key differences are that significantly different probabilities are assigned to the two bases so that few bits are discarded, and that key extraction is performed separately for data in each of the bases. The Cambridge–Toshiba team further improved efficiency and included decoy states, developing a new protocol called T12 (Lucamarini _et al._ , 2013). The authors prove it to be unconditionally secure. As of February 2015, this is the protocol with which the highest ranges and secure key rates have been obtained (Korzh _et al._ , 2015). Decoy state QKD comes to solve the problem that the secure key rate of a quantum key from a coherent source scales like the square of the transmittance (the proportion of photons that make it through from Alice to Bob) of the medium, and thus a secure key becomes too long to be practical when it must be transmitted for long distances. For decoy-state QKD, the key length scales like the transmittance. Three-source decoy-state QKD was what was used in (Lucamarini _et al._ , 2013). Decoy state QKD works also with non-coherent sources such as PDC sources (Ma & Lo, 2008). Additionally, there has been work on measuring-device independent (MDI) QKD, in which Alice and Bob independently prepare phase randomized coherent pulses in one of the four BB84 states (with decoy states) and send these to an untrusted third party, Charlie. Charlie then performs Bell state measurements (BSM), and announces to Alice and Bob over a public channel the successful BSM events. Alice and Bob can get a sifted key by dropping events where they sent pulses in different bases (Wang, 2013). This has been implemented and gives good key rates in the laboratory (Tang et al., 2014). A further improvement has been examined, using four-source decoy states (Jiang et al., 2015). There has been recent work to modify the BB84 protocol to deal with higher bit error rates on the sifted key, in order to distribute quantum keys for longer distances without using repeaters in a way that is compatible with optical amplification (Hughes & Norholdt, 2014). ### 6.3. Continuous Variable (CV) QKD Continuous-variable (CV) QKD protocols employ continuous or discrete modulations of the quadratures of an electromagnetic field. CV-QKD setups rely on a coherent detection between the quantum signal and a classical reference signal, and their implementation requires only standard telecom components. They are compatible with wavelength division multiplexing, which greatly eases their deployment into telecommunication networks. They should be easier to integrate on silicon photonics chips (Jouguet et al., 2013; Kumar, Qin, & Alléaume, 2014). The bottleneck for CV-QKD is a classical cryptography problem, that of error- correction. For a long time, the range of CV-QKD was limited to 25-30km. New error-correcting codes have improved this range to 80km (Jouguet et al., 2012). SeQureNet’s Cygnus module for CV-QKD features this range (http://sequrenet.com/products.html). Currently, CV-QKD keyrates are competitive with DV-QKD keyrates up to about 30km. But it may be more difficult to increase ranges for CV-QKD than for DV-QKD because security proofs for CV-QKD are penalized heavily for finite size effects. An additional concern is that CV-QKD is a newer technology, and therefore has different vulnerabilities, some of which may be unmapped. Several potential vulnerabilities have been identified and addressed in (Jouguet, Kunz-Jacques, & Diamanti, 2013; Huang et al., 2014; Kunz-Jacques & Jouguet, 2015). Considering the above, CV-QKD should be considered a promising future technology for medium-range QKD. The state of the art for CV-QKD is surveyed in (Jouguet et al., 2014). ### 6.4. Entanglement-based protocols There are a number of protocols involving entangled pairs of photons, chief among these being E91 (Ekert, 1991). In E91, Alice and Bob each have half of an entangled state (EPR pair, or singlet). The working concept of this scheme is that there is nothing for Eve to intercept, as the qubit state manifests only after a measurement has been made. If Eve attempts an intercept-resend attack, her measurement will break the entanglement between the photons. The protocol proceeds as follows. Alice and Bob each choose one of two different bases to measure, with 50% probability of choosing one basis and 50% probability of choosing the other. After having performed their measurements, they disclose which bases they used over a public channel. If the results of measurements which were made in different bases violate Bell inequalities, then the state is still entangled and there has been no eavesdropping. Despite being theoretically more secure than BB84 and its variants (fewer side-channels and thus fewer bits required for a secure key), entanglement- based protocols are not currently considered to be practical for long-range large-scale systems because of the difficulty of controlling entangled pairs caused by decoherence (Scarani & Kurtsiefer, 2014). ### 6.5. Counterfactual QKD Tae-Gon Noh has demonstrated that a QKD can be achieved ostensibly without sending the key through the quantum channel (Noh, 2009). The quantum principle in play is that the possibility of sending a photon can be detected even if the photon is seemingly not actually sent. Counterfactual QKD has been demonstrated experimentally in the laboratory (Liu et al., 2011). ## 7\. Real-time key extraction Key generation bandwidth in a pure CPU-based implementation has been shown to saturate at rates of around 1MHz (Restelli et al., 2009). High speed QKD networks routinely exceed this data rate— for example, the NIST system generates sifted keys at around 2MHz and has a maximal capacity above 30MHz. For secure real-time practical applications, GHz data rates are anticipated. In order to shift the bottleneck from the classical computation layer back to the physical layer where it should be, hardware-based implementations have become necessary. Sifting is computationally straightforward, and is relatively easy to perform at high speed. There is good privacy-amplification software which can work directly on a CPU-based system (Zhang et al., 2014), and the Wegman–Carter strongly universal hashing method, as used by e.g. Id Quantique, is also good. It is the error correction step which is complicated and which sets a hard upper limit on the secure key rate. The Cascade error-correction algorithm, developed for QKD in (Brassard & Salvail, 1994), is the fastest at current data rates, and is best implemented in a Field Programmable Gate Array (FPGA) because it requires many simple but different logical bit-level operations. When the NIST QKD network began to exceed data rates of 1MHz, implementation of the Cascade algorithm was moved to hardware (Mink et al., 2006). The maximal throughput they were able to achieve was 12MHz in theory, but they were not able to approach that limit in a practical system due to timing jitter in their photon detectors (Mink & Bienfang, 2013). The Wuhu metropolitan area QKD network uses a similar FPGA- based system (Zhang et al., 2012). For next-generation real-time error correction as data rates push towards the GHz mark, the Low-Density Parity-Check (LDPC) algorithm is expected to replace the Cascade algorithm for error correction (Elkouss et al., 2009). The LDPC algorithm requires 20 to 30 bytes of memory per bit of data being corrected, as opposed to 1 or 2 bytes for the Cascade algorithm. On an FPGA, LDPC implementation rates of up to 607MHz have been reported (Mhaske et al., 2015). The current fastest implementation of the LDPC algorithm runs on a GPU-based system (Falcão et al., 2009) and has been tested for QKD (Martinez-Mateo et al., 2013; Dixon & Sato, 2014). For even faster rates, LDPC performance of 47 GHz has been reported for a custom chip implementation but not in the context of QKD (Zhang et al., 2009). ## 8\. Hardware: Photon sources A common method of encoding qubits is the use of polarized photons (less common methods include time-bin encoding (Marcikic et al., 2002) and frequency encoding (Zhu et al., 2011)). To preclude photon number splitting attacks, each qubit should be sent on a single photon. The ideal single-photon source would send a single photon 100% of the time whenever the user wishes (“on demand”), would send multiple photons 0% of the time, and the photons it sends would be indistinguishable. If a photon cannot be sent 100% of the time on demand, the detector might have to be left on for a longer time, increasing ‘dark count’ (detection of photons which were not sent to it) and thus increasing noise. If the source were to send multiple photons, then Eve would be able to intercept one photon and transmit the remaining photons to Bob, executing a _photon number splitting attack_. And if photons were distinguishable, then interception one photon could give non-trivial information about another photon. Photon sources are classified as _deterministic_ versus _probabilistic_. A deterministic single photon source emits a single photon on demand, whereas a probabilistic source might emit more than one photon, and its photon emission timing not be entirely on demand. One should note, however, that even the most ‘deterministic’ photon source might in practice exhibit probabilistic behaviour because, for example, photons might get lost during emission with some probability (“ _extraction loss_ ”). A common measure for the efficiency of a single-photon source is the _$2$ nd order correlation function_ $g^{(2)}$. If $g^{(2)}=1$, this means that the number of photons emitted by the source follows a Poisson distribution, which is the distribution one would expect from completely random and uncorrelated emissions. It is usually assumed that $g^{(2)}=1$ for attenuated lasers, although, as pointed out by the European Telecom Standards Institute, standard number GS QKD 003 Section 6.4.1 (ETSI, 2015), experiment hasn’t always born this out and perhaps further study is necessary e.g. when the diode is biased close to the lasing threshold (Dixon _et al._ , 2009). The $g^{(2)}<1$ situation is referred to as _photon antibunching_. In this case the probability of emitting one photon relative to the probability of emitting multiple photons is higher than in a Poisson process. The ideal state is $g^{(2)}=0$, which means that we get a single photon $100\%$ of the time. The most common single photon sources are attenuated lasers, in which a laser beam is sent through a powerful attenuator which weakens it to the point that the probability of emitting one photon is greater than the probability of emitting multiple photons. Attenuated lasers are relatively cheap, convenient, and robust. When higher performance (lower $g^{(2)}$) is desired, the most common single- photon sources make use of parametric down-conversion (PDC). This type of source is not on-demand, but it probabilistically produces a pair of photons, one of which can be used as a _heralding photon_ to instruct the detector to activate. This is a major advantage in QKD where it is important to minimize detector dark-count. The heralding photon could also be used as an entangled pair with the first photon, although here PDC makes it difficult in general to obtain the desired wavelength and phase-matching for the pair (Eisaman _et al._ , 2011). A promising future technology is the use of nitrogen vacancy (NV) color centers in diamond as single photon sources. An NV center is a defect in a diamond lattice which occurs when a nitrogen atom is substituted for a carbon atom, leaving a vacancy next to it. As single photon sources, NV centers are on-demand and exhibit low $g^{(2)}$. The current challenges are that they are not identical, although some tunability has been demonstrated (Tamarat et al., 2006), and that the ‘shelving level’ reduces their efficiency. There are several proposed approaches to solving these problems (e.g. (Babinec _et al._ , 2010)); but it is the promise of a single photon coupled with a long-lived spin qubit (the vacancy itself is an excellent qubit) that makes NV centers especially promising. Note also that must be thousands of optical defects in solids which could potentially be used for single-photon generation; only two of these have so far been seriously studied in this context (Santori, Fattal, & Yamamoto, 2010). An NV center used as a single photon source. Retrieved from http://xqp.physik.uni-muenchen.de/research/single_photon/index.html There are many other single-photon sources, including single atoms, ions, and molecules, ensembles, quantum dots, nanowires, four-wave mixing, and mesoscopic quantum wells, but these do not currently seem as suitable for QKD as the sources discussed above. ## 9\. Transmission Quantum key distribution can be performed through fiber, through free space, or (experimentally) bounced off a satellite. The principles of sending photons through fiber and through free space are the same, but fiber provides a channel in which the amount of noise can be determined and even controlled to some extent, whereas the amount noise in a free-space channel is unknown (although sometimes one may try to estimate it as in e.g. (Gabay & Arnon, 2005)) and typically is changing. Frequencies used to send photons are typically around 800nm for free space and around 1550nm for fibers. Experimental QKD usually uses dark fibers with no other signals passing through it, but real-world applications will typically involve sending messages through bright fibers which are carrying other signals. Scattering effects in bright fibers will raise the BER of Alice’s transmissions, and will cause more of Alice’s photons to ‘get lost on the way’. Despite this, by smartly time-filtering QKD photon and other communication photons, in 1992 a team from Toshiba was able to obtain a secure bit-rate of 507KHz over a 95km bright fiber, several factors of 10 over what had been achieved previously (Patel et al., 2012). The greatest distance positive key rates have been experimentally obtained through fiber is 307km (Korzh _et al._ , 2015) and through free space is 144km between two Canary Islands (Ursin _et al._ , 2007). The problem with free space transmission is atmospheric turbulence— random fluctuations in the refractive index of air. One potential solution is to bounce the polarized photons off satellites. The distance to the International Space Station is 400km, but the atmospheric thickness is an order of magnitude smaller than the Canary Islands experiment. In 2014 a team from the University of Padua bounced photons off four satellites to show feasibility (Vallone _et al._ , 2014), and China claims to have done so as well, and aims to have a dedicated QKD satellite in orbit by 2016 (Yikra, 2014). When such technologies become practically viable, they will significantly increase free space QKD ranges. ## 10\. Hardware: Photon detection For Bob to receive qubits from Alice in the form of single photon polarizations, Bob needs to have a good single photon detector. The main technological bottleneck in the development of practical and secure QKD systems for short to medium distances is the development of good single photon detectors. Thus, in the last few years, any improvement to single photon detection technology has immediately led to improved QKD capabilities. Our main reference for this section are (Eisaman _et al._ , 2011) and (Hadfield, 2009). Single crystal diamond nanowires for photon detection. Retrieved from http://www.osa-opn.org/opn/media/Images/photocontests/gallery09_36.jpg An ideal single photon detector for QKD should have 100% _detection efficiency_ (every photon sent to the detector should be successfully detected), 0% _dark count_ (the detector should not detect photons which were not sent to it), no _dead time_ (the recovery time for the detector after it has detected a photon until it had detected another photon), and no _timing jitter_ (the time between the photon’s arrival and its registration by the detector). Additionally, an ideal detector would have complete _photon number resolution_ , meaning that it would be able to count the number of photons it had received. It would also be _asynchronous_ , meaning that it need not know the arrival times of photons in advance. Low detection capacities and high dark counts create noise in the communication channel, reducing its capacity. A low capacity channel is vulnerable to an intercept-resent attack, because it is difficult to detect eavesdropping in the presence of random noise (Section 12.1). High dead time reduces the channel bit rate, and creates a vulnerability to faked-state attacks and to time-shift attacks (Makarov _et al._ , 2006; Burenkov _et al._ , 2010; Makarov, 2009). Timing jitter can lead to a leak of secret key information (Lamas-Linares & Kurtsiefer, 2007). Poor or nonexistent photon number resolution creates a vulnerability to photon number splitting attacks. A single photon detector typically works by converting a photon into a charge carrier which in turn triggers an avalanche process in a physical system which is held very close to a critical state, leading to a macroscopic current pulse. Superconducting nanowire single photon detectors (SNSPD) are the best single photon detectors known currently. They were first developed in 2001. They have high detection efficiency (ten times better than the best semiconductor detectors), low dark count, low dead time, and low time jitter. They are also fully asynchronous. Their drawback is that they require cryogenic temperatures ($<3K$) to operate, which makes them bulky and expensive, and has limited their uses outside the lab. There is work being done, however, on variations of SNSPD that can function at temperatures of over $20K$, making them a promising future technology for military and government applications (Wang, Miki, & Fujiwara, 2009). Single photon avalanche detectors (SPADs) are the single photon detectors which are most currently used in practice. They are cheap and compact, with high detection efficiency and low time jitter. They are also fully asynchronous. The challenge in building good SPADs has been _afterpulsing_ , which is the phenomenon of a spontaneous dark count occurring shortly after a photon detection. If we wait until afterpulsing ends before reactivating the SPAD, then we increase the dead time. There has recently been dramatic progress in SPAD design. In 2013, the University of Geneva Applied Physics team developed an InGaAs negative feedback avalanche diode (NFAD) single photon detector whose performance rivals many SNSPD systems, but which operates at temperatures of approximately 150–220 K (as opposed to $<3K$ for SNSPD systems) (Korzh _et al._ , 2014). Using these InGaAs NFADs, the same team were able in February 2015 to demonstrate provably secure QKD transmission over 307km of optical fiber, which is the current record (Korzh _et al._ , 2015). ## 11\. Auxiliary systems ### 11.1. Random number generators A QKD system is only as good as its random number generator. If Eve can predict Alice and Bob’s random choices, they she can read the entire key. The entire selling point of Quintessence Technologies is their random number generators. CPU-based random number generators are trusted for many classical cryptography tasks, and are implemented in most operating systems. The numbers they produce are not truly random, however, and therefore they are usually referred to as _pseudorandom number generators_. When higher speeds are required and when stronger random numbers are needed, hardware-based implementations are preferred. These come in two flavours— they either use filtered random physical processes within the FPGA as random number seeds (Tsoi et-al., 2007; Kwok & Lam, 2006), or they use less random seeds and strong permutations (Alimohammad et al., 2008; Cheung et al., 2007; Xiang & Benkrid, 2009). Currently, both alternatives are considered cryptographically equivalent. In QKD, in order to physically guarantee unconditional security, quantum effects are desired for use as true random number generators (TRNG). A quick and dirty way to do this, for Alice at least, is to send an unpolarized single photon through a beam splitter— if it comes out one end then count that as a zero, and if it comes out of the other end count it as a 1. A more sophisticated version of this scheme which eliminates this bias is marketed by Id Quantique, and another by Quintessence Technologies, and reaches rates of 16 KHz. Looking into the future, an experimental idea with great promise is to use quantum vacuum fluctuations for high bandwidth truly random number generation of up to 100GHz (Jofre et al., 2011). We note that TRNGs arise as commercial spinoffs of QKD projects. ### 11.2. Memories and repeaters The concept of a quantum repeater. Retrieved from http://www.uqcc.org/research/index.html To extend the range of QKD beyond a few hundred kilometers, quantum repeaters will be necessary, which in turn will require quantum memories. A quantum memory absorbs a photon, stores its quantum states for as long as possible, and releases it on demand. A key feature is that it does not break entanglement. The primary candidates for practical quantum memories for QKD in the near future are Raman gas based quantum memories (Simon et al., 2010) and quantum memories using NV centers (de Riedmatten & Afzelius, 2015). The advantages of the former include its greater capacity, while the advantages of the latter include that it is solid-state and that it allows longer storage times. It is still unclear which of these approaches will be best. It is still unclear which repeater technology will be best, although the first quantum repeaters which outperform direct transmission will probably be based on atomic ensembles, linear optics, and photon counting (Sangouard et al., 2011). Quantum memories and repeaters are expected to become a commercial technology within 10-15 years. ## 12\. Attacks While the protocols of QKD operating under certain conditions are unconditionally secure, practical implementations have been successfully attacked. While none of these attacks is fatal to the QKD concept— effective countermeasures to each attack have been devised— it is generally agreed that the security of each setup should be the object of a dedicated study whose goal is to find and patch up all vulnerabilities (Scarani & Kurtsiefer, 2014). ### 12.1. Intercept-resend The simplest and most direct attack against BB84 and its relatives is for Eve to intercept a photon sent by Alice to Bob, to measure that photon, to prepare her own photon encoding the bit which she measured, and to send that photon off to Bob. Because Eve doesn’t know in which basis Alice’s photon was sent, she’ll measure in the wrong basis approximately half of the time and she will send a photon in the wrong basis to Bob approximately half of the time. An intercept-resend attack thus introduces a bit error rate (BER) of around 25%, although weaknesses in certain practical systems allow modified versions of this attack to introduce BERs of 19.7% (Xi, Qi, & Lo, 2010). Because acceptably BERs in commercial systems are around 8%, practical QKD is indeed secure against pure intercept-resent attacks, which are caught during the error-correction key-establishment phase— if the error rate is too high then Eve has been there. ### 12.2. Photon Number Splitting (PNS) Due to hardware limitations, most photon sources used in QKD are not true single photon sources in the sense that there is a non-negligible probability that they will generate multiple photons to transmit a single qubit. If Alice sends two or more identical photons to Bob, then Eve can split off one photon and send the remaining photons through. Eve stores the qubit she has learnt in quantum memory until Alice has revealed her encoding basis. Then Eve measures her photons in the correct basis and gains information about the key. A successful photon number splitting attack requires sophisticated equipment— Eve must be able to count photons and to split off just one to quantum memory while sending others through. Moreover, various countermeasures have been developed. Better single-photon sources and modifications of the BB84 protocol, such as for instance SARG04, make successful PNS attacks much more difficult to carry out. Another solution is to use decoy states, in which photons are randomly sent at different intensities. The security of decoy state QKD against PNS attacks was proven in (Lo _et al._ , 2005a). Because a successful photon number splitting attack is much more difficult to carry out against decoy-state QKD, we can use attenuated lasers as photon sources when transmitting keys, increasing secure key-rates (Yuan, 2007). The current state-of-the-art for decoy-state QKD is 320MHz over a 200km fiber, yielding a 15Hz secure key rate (Liu _et al._ , 2010). ### 12.3. Timing attacks When different light sources are used for beams in different polarizations, and/or different detectors are used to make different measurements, it may be possible to ‘listen in’ to which bit was sent or to which bases was used without actually intercepting a photon. Such side-channels were evident already in the first implementations QKD, as noted by Brassard (Brassard, 2005): > The funny thing is that, while our theory had been serious, our prototype > was mostly a joke. Indeed, the largest piece in the prototype was the power > supply needed to feed in the order of one thousand volts to Pockels cells, > used to turn photon polarization. But power supplies make noise, and not the > same noise for the different voltages needed for different polarizations. > So, we could literally hear the photons as they flew, and zeroes and ones > made different noises. Thus, our prototype was unconditionally secure > against any eavesdropper who happened to be deaf ! :-) It is therefore critical to the security of the QKD system that different light-sources and detectors be indistinguishable to Eve. One particular vulnerability is that different light sources and detectors may not be perfectly synchronized, so that Eve can figure out which detector clicked, for example, by examining the time signature publicly announced by Bob in order to determine which photon he detected of the photons sent by Alice. Such an attack could read-off $\geq 25\%$ of the key for a detector mismatch of $0.5$ nanoseconds, an amount that could easily go unnoticed (Lamas-Linares & Kurtsiefer, 2007). An attempt to carry out such an attack against a commercial system was unsuccessful because of several practical difficulties (Zhao et al., 2008). ### 12.4. Trojan attacks The logo of the Quantum Hacking Lab. Retrieved from http://www.vad1.com/lab/ In a Trojan attack, Eve shines bright light at either Alice or Bob, determining which base was used by analyzing the reflection. The Trojan attack has successfully read off complete keys both in the lab (Gerhardt et al., 2011) and also of commercial QKD systems, QPN-5505 from MagiQ Technologies and Clavis2 of IP Quantique (Lydersen et al., 2010). This has been the most powerful and the best-performing hack on QKD so far. Although these specific attacks can be protected against, Trojan attacks using pulses of different wavelengths may still be able to hack complete keys, and we still do not know the full scope of the vulnerability of practical QKD systems to Trojan attacks (Jain et al., 2014, 2015). ### 12.5. Other side-channel attacks Many other attacks against QKD systems have been investigated, and new vulnerabilities are periodically discovered. Some of these attacks (denial of service, man-in-the-middle,…) can be carried out against classical systems as well, and the vulnerability of QKD to these attacks is identical to the vulnerability of any classical protocol. Other attacks which take advantage of a weakness in an auxiliary system— e.g. a randomization attacks, in which the random bases are successfully computed by Eve because the random number generator is faulty— can be counteracted by using better hardware. Of course each side-channel attack must be investigated and ruled out for each setup. ## 13\. Conclusion Quantum Key Distribution (QKD) is a modern form of trusted courier, which in principle allows Alice to communicate a message to Bob with complete confidence that the message will not be eavesdropped on during transmission. Real-life QKD security, however, is a quantitative issue, and each setup should be individually studied to ensure its security. QKD is currently a focus of interest for many private, governmental, and military groups all over the world. Current state of the art setups still use the first QKD protocol, BB84, and its variants. The Cascade protocol is still the fastest for error correction, but LDPC is expected to overtake it as key rates rise. In both cases, hardware implementation using FPGA’s is the current state of the art and is likely to remain so for the next decade at least. Qubits are typically transmitted as polarized photons. Decoy-state QKD using attenuated laser pulses are the current state of the art photon sources, despite not being true single photon sources. NV centers are a promising future technology. Polarized photons can be transmitted through fiber, through air (free space), or bounced off satellites. There are various attempts to send polarized photons via bright fibers through which other messages are travelling, but the key rates being obtained are still quite low. Photon detectors are the main technological bottleneck for practical QKD. The current state of the art are InGaAs NFAD’s. A promising future technology are SNSPD’s, which currently require cryogenic temperatures to operate, but future SNSPD’s be able to operate at above 20K. Memories and repeaters, which are thought to be required for QKD at ranges over around 400km, are still in the experimental stage, and it is too early to say which technology will be best. Security of QKD is a well-studied field, and there have been numerous attempts to attack QKD implementations both using standard attacks and also using side- channel attacks. Only one of these attacks, a Trojan attack, has successfully stolen a secret key, and the vulnerability it highlights can be plugged. QKD of course has the same vulnerability to classical attacks such as denial-of- service and man-in-the-middle as classical implementations. QKD is an exciting emerging technology which is beginning to enter the marketplace. We expect it to be successful in its own right, and also to serve as a stepping stone towards greater and higher goals in quantum communication and computation. ## 14\. Acknowledgement The author wishes to thank Judith Kupferman for useful comments, and wishes to thank Paul Jouguet at Takehisa Iwakoshi for corrections and references. ## References * Alimohammad et al. (2008) Alimohammad, A., Fard, S.F., Cockburn, B.F., & Schlegel, C. 2008 A Compact and Accurate Gaussian Variate Generator. _IEEE T. VLSI Syst._ , 16(5), 517–527. * Babinec _et al._ (2010) Babinec, T.M., Hausmann, B.J., Khan, M., Zhang, Y., Maze, J.R., Hemmer, P.R., & Lončar, M. 2010 A diamond nanowire single-photon source. _Nat. Nanotechnol._ , 5(3), 195–199. * Bennett & Brassard (1984) Bennett, C.H. & Brassard, G. 1984 Quantum cryptography: Public key distribution and coin tossing. In _Proceedings of IEEE International Conference on Computers, Systems and Signal Processing_ , volume 175, page 8. http://researcher.watson.ibm.com/researcher/files/us-bennetc/BB84highest.pdf * Bennett et al. (1992) Bennett, C.H., Bessette, F., Brassard, G., Salvail, L., & Smolin, J. 1992 Experimental quantum cryptography. _J. of Cryptography_ , 5(1) 3–28. * Berry (1998) Berry, M. 1998 Foreword, in _Introduction to Quantum Computation and Information_ eds. Lo, H.-K., Popescu S., & Spiller, T. World Scientific, Singapore, hardcover 1998, paperback 2001. * Branciard et al. (2005) Branciard, C., Gisin, N., Kraus, B., & Scarani, V. 2005 Security of two quantum cryptography protocols using the same four qubit states. _Phys. Rev. A_ , 72(3), 032301. arXiv:quant-ph/0505035 * Brassard & Salvail (1994) Brassard, G., & Salvail, L. 1994 Secret-key reconciliation by public discussion. In _Advances in Cryptology EUROCRYPT93_ , Springer Berlin Heidelberg, 410–423. * Brassard (2005) Brassard, G. 2005 Brief history of quantum cryptography: A personal perspective. In _Theory and Practice in Information-Theoretic Security, 2005_. IEEE Information Theory Workshop, 19–23. arXiv:quant-ph/0604072 * Burenkov _et al._ (2010) Burenkov, V., Qi, B., Fortescue, B., & Lo H.-K. 2010 Security of high speed quantum key distribution with finite detector dead time. _Quantum Inf. Comput._ , 14(3-4), 217–235. arXiv:1005.0272 * Cheung et al. (2007) Cheung, R.C.C., Lee, D.-U., Luk, W., & Villasenor, J.D. 2007 Hardware Generation of Arbitrary Random Number Distributions From Uniform Distributions Via the Inversion Method. _IEEE T. VLSI Syst._ , 15(8), 952–962. * Comandar _et al._ (2014) Comandar, L.C., Fröhlich, B., Lucamarini, M., Patel, K.A., Sharpe, A.W., Dynes, J.F., Yuan, Z.L., Penty, R.V., & Shields, A.J. 2014 Room temperature single-photon detectors for high bit rate quantum key distribution. _Appl. Phys. Lett._ , 104, 021101. arXiv:1402.2210 * de Riedmatten & Afzelius (2015) de Riedmatten, H., & Afzelius, M. 2015 Quantum Light Storage in Solid State Atomic Ensembles. Preprint. arXiv:1502.00307 * Dixon _et al._ (2009) Dixon, A.R., Dynes, J.F., Yuan, Z.L., Sharpe, A.W., Bennett, A.J., & Shields, A.J. 2009 Ultrashort dead time of photon-counting InGaAs avalanche photodiodes _Appl. Phys. Lett._ , 94(23), 231113. arXiv:0905.2931 * Dixon & Sato (2014) Dixon, A.R., & Sato, H. 2014 High speed and adaptable error correction for megabit/s rate quantum key distribution. _Scientific Reports_ , 4. * Eisaman _et al._ (2011) Eisaman, M.D., Fan, J., Migdall, A., & Polyakov, S.V. 2011 Single-photon sources and detectors. _Rev. Sci. Instrum._ , 82(7), 071101. * Ekert (1991) Ekert, A.K. 1991 Quantum cryptography based on Bell’s theorem. _Phys. Rev. Lett._ 67(6), 661. * Elkouss et al. (2009) Elkouss, D., Leverrier, A., Alléaume, R., & Boutros, J.J. 2009 Efficient reconciliation protocol for discrete-variable quantum key distribution. In _IEEE International Symposium on Information Theory, 2009_ , ISIT 2009, 1879–1883). * Elliot et al. (2005) Elliott, C., Colvin, A., Pearson, D., Pikalo, O., Schlafer, J., & Yeh, H. 2005 Current status of the DARPA quantum network. In _Defense and Security_ , International Society for Optics and Photonics, 138–149. arXiv:quant-ph/0503058 * EPSRC (2014) UK Quantum Technology Landscape, 2014. Retrieved from http://www.epsrc.ac.uk/newsevents/pubs/dstl-uk-quantum-technology-landscape-2014/ * ETSI (2015) European Telecom Standards Institute, Standards for Quantum Key Distribution. Retrieved from http://www.etsi.org/technologies-clusters/technologies/quantum-key-distribution * Falcão et al. (2009) Falcão, G., Silva, V., & Sousa, L. 2009 How GPUs can outperform ASICs for fast LDPC decoding. In _Proceedings of the 23rd international conference on Supercomputing_ , ACM, 390–399. * Gabay & Arnon (2005) Gabay, M., & Arnon, S. 2005 Effect of turbulence on a quantum-key distribution scheme based on transformation from the polarization to the time domain: laboratory experiment. _Opt. Engineering_ , 44(4), 045002. * Gerhardt et al. (2011) Gerhardt, I., Liu, Q., Lamas-Linares, A., Skaar, J., Kurtsiefer, C., & Makarov, V. 2011 Full-field implementation of a perfect eavesdropper on a quantum cryptography system. _Nat. Commun._ , 2, 349. arXiv:1011.0105 * Gisin et al. (2002) Gisin, N., Ribordy, G., Tittel, W., & Zbinden, H. 2002 Quantum cryptography. _Rev.Mod.Phys._ , 74, 145–190. * Hadfield (2009) Hadfield, R.H. 2009 Single-photon detectors for optical quantum information applications. _Nat. Photonics_ , 3(12), 696–705. * Huang et al. (2014) Huang, J.Z., Kunz-Jacques, S., Jouguet, P., Weedbrook, C., Yin, Z.Q., Wang, S., Chen, W., Guo, G.C., & Han, Z.F. (2014). Quantum hacking on quantum key distribution using homodyne detection. _Phys. Rev. A_ , 89(3), 032304. arXiv:1402.6921 * Hughes et al. (2013) Hughes, R.J., Nordholt, J.E., McCabe, K.P., Newell, R.T., Peterson, C.G., & Somma, R.D. 2013 Network-Centric Quantum Communications. In _Frontiers in Optics_ , Optical Society of America. arXiv:1305.0305 * Hughes & Norholdt (2014) Hughes, R.J., & Nordholt, J.E. 2014 Long-range quantum cryptography: Amplified quantum key distribution (AQKD). Preprint. arXiv:1406.6990 * Inamori, Lütkenhaus & Mayers (2007) Inamori, H., Lütkenhaus, N., & Mayers, D. 2007 Unconditional security of practical quantum key distribution. _Eur. Phys. J. D._ , 41(3), 599–627. arXiv:quant-ph/0107017 * Iwakoshi & Hirota (2014) Iwakoshi, T., & Hirota, O. 2014 Misinterpretation of statistical distance in security of quantum key distribution shown by simulation. In _SPIE Security + Defence_ , International Society for Optics and Photonics, 92540L–92540L. * Jain et al. (2014) Jain, N., Anisimova, E., Khan, I., Makarov, V., Marquardt, Ch., & Leuchs, G. 2014 Trojan-horse attacks threaten the security of practical quantum cryptography. _New J. Phys._ , 16(12), 123030. arXiv:1406.5813 * Jain et al. (2015) Jain, N., Stiller, B., Khan, I., Makarov, V., Marquardt, Ch., & Leuchs, G. 2015 Risk analysis of Trojan-horse attacks on practical quantum key distribution systems. _IEEE J. Sel. Top. Quantum Electron._ 21, 6600710. arXiv:1408.0492 * Jiang et al. (2015) Jiang, H., Gao, M., Wang, H., Li, H., & Ma, Z. 2015 Four-intensity Decoy-state Quantum Key Distribution with enhanced resistance against statistical fluctuation. Preprint. arXiv:1502.02249 * Jofre et al. (2011) Jofre, M., Curty, M., Steinlechner, F., Anzolin, G., Torres, J.P., Mitchell, M.W. & Pruneri, V. 2011 True random numbers from amplified quantum vacuum. _Opt. Express_ , 19(21), 20665–20672. * Jouguet et al. (2012) Jouguet, P., Kunz-Jacques, S., Leverrier, A., Grangier, P., & Diamanti, E. 2013 Experimental demonstration of long-distance continuous-variable quantum key distribution. _Nat. Photonics_ , 7(5), 378–381. arXiv:1210.6216 * Jouguet, Kunz-Jacques, & Diamanti (2013) Jouguet, P., Kunz-Jacques, S., & Diamanti, E. 2013 Preventing calibration attacks on the local oscillator in continuous-variable quantum key distribution. _Phys. Rev. A_ , 87(6), 062313. arXiv:1304.7024 * Jouguet et al. (2013) Jouguet, P., Kunz-Jacques, S., Kumar, R., Qin, H., Gabet, R., Diamanti, E., & Alléaume, R. 2013 Experimental demonstration of the coexistence of continuous-variable quantum key distribution with an intense DWDM classical channel. In _Proc. 3rd Ann. Conf. Quantum Cryptography_. * Jouguet et al. (2014) Jouguet, P., Elkouss, D., & Kunz–Jacques, S. 2014 High-bit-rate continuous-variable quantum key distribution. _Phys. Rev. A_ 90, 042329. * König et al. (2007) König, R., Renner, R., Bariska, A., & Maurer, U. 2007 Small accessible quantum information does not imply security. _Phys. Rev. Lett._ , 98(14), 140502. arXiv:quant-ph/0512021 * Korzh _et al._ (2014) Korzh, B., Walenta, N., Lunghi, T., Gisin, N., & Zbinden, H. 2014 Free-running InGaAs single photon detector with 1 dark count per second at 10% efficiency. _Appl. Phys. Lett._ , 104, 081108. arXiv:1312.2636 * Korzh _et al._ (2015) Korzh, B., Lim, C.C.W., Houlmann, R., Gisin, N., Li M.J., Nolan, D., Sanguinetti, B., Thew, R., & Zbiden, H. 2015 Provably secure and practical quantum key distribution over 307km of optical fibre. _Nat. Photonics_. arXiv:1407.7427 * Kumar, Qin, & Alléaume (2014) Kumar, R., Qin, H., & Alléaume, R. 2014 Coexistence of continuous variable QKD with intense DWDM classical channels. Preprint. arXiv:1412.1403 * Kunz-Jacques & Jouguet (2015) Kunz-Jacques, S., & Jouguet, P. 2015 Robust Shot Noise Measurement for Continuous Variable Quantum Key Distribution. _Phys. Rev. A_ 91, 022307. arXiv:1406.7554 * Kwok & Lam (2006) Kwok, S.H.M., & Lam, E.Y. 2006 FPGA–based High-speed True Random Number Generator for Cryptographic Applications. _TENCON 2006. 2006 IEEE Region 10 Conference_ , 1–4 & 14–17. * Lamas-Linares & Kurtsiefer (2007) Lamas-Linares, A., & Kurtsiefer, C. 2011 Breaking a quantum key distribution system through a timing side channel. _Opt. Express_ , 15, 9388. arXiv:0704.3297 * Liu _et al._ (2010) Liu, Y., Chen, T.-Y., Wang, J., Cai, W.-Q., Wan, X., Chen, L.-K., Wang, J.-H., Liu, S.-B. Liang, H., Yang, L., Peng, C.-Z. Chen, K., Chen, Z.-B. & Pan J.-W. 2010 Decoy–state quantum key distribution with polarized photons over 200 km. _Opt. Express_ , 18 (2010) 8587–8594. arXiv:0908.4063 * Liu et al. (2011) Liu, Y., Ju, L., Liang, X.L., Tang, S.B., Tu, G.L.S., Zhou, L., Peng, C.-Z., Chen, K. Chen, T.-Y. Chen, Z.-B. & Pan, J.W. 2012 Experimental demonstration of counterfactual quantum communication. _Phys. Rev. Lett._ , 109(3), 030501. arXiv:1107.5754 * Lo _et al._ (2005a) Lo, H.K., Ma, X., & Chen, K. 2005 Decoy state quantum key distribution. _Phys. Rev. Lett._ , 94(23), 230504. arXiv:quant-ph/0411004 * Lo _et al._ (2005b) Lo, H.-K., Chau, H.F. & Ardehali, M. 2005 Efficient quantum key distribution scheme and proof of its unconditional security. _J. of Crypt._ , 18, 133. arXiv:quant-ph/0011056 * Lucamarini _et al._ (2013) Lucamarini, M., Patel, K.A., Dynes, J.F., Fröhlich, B., Sharpe, A.W., Dixon, A.R., Yuan, Z.L., Penty, R.V., & Shields, A. J. 2013 Efficient decoy-state quantum key distribution with quantified security. _Opt. Express_ , 21(21), 24550–24565. http://arxiv.org/abs/1310.0240 * Lydersen et al. (2010) Lydersen, L., Wiechers, C., Wittmann, C., Elser, D., Skaar, J., & Makarov, V. 2010 Hacking commercial quantum cryptography systems by tailored bright illumination. _Nat. Photonics_ , 4(10), 686–689. arXiv:1008.4593 * Ma & Lo (2008) Ma, X., & Lo, H.K. 2008 Quantum key distribution with triggering parametric down-conversion sources. _New J. Phys._ , 10(7), 073018. * Makarov _et al._ (2006) Makarov, V., Anisimov, A. & Skaar, J. 2006 Effects of detector efficiency mismatch on security of quantum cryptosystems. _Phys. Rev. A_ , 74, 022313; Erratum _Phys. Rev. A_ , 78, 019905. arXiv:quant-ph/0511032 * Makarov (2009) Makarov, V. 2009 Controlling passively quenched single photon detectors by bright light. _New J. Phys._ , 11(6), 065003. arXiv:quant-ph/0707.3987 * Marcikic et al. (2002) Marcikic, I., De Riedmatten, H., Tittel, W., Scarani, V., Zbinden, H., & Gisin, N. 2002 Time-bin entangled qubits for quantum communication created by femtosecond pulses. _Phys. Rev. A_ , 66(6), 062308. arXiv:quant-ph/0205144 * Martinez-Mateo et al. (2013) Martinez-Mateo, J., Elkouss, D., & Martin, V. 2013 Key Reconciliation for High Performance Quantum Key Distribution. _Scientific Reports_ , 3. http://www.nature.com/srep/2013/130402/srep01576/full/srep01576.html * Mhaske et al. (2015) Mhaske, S., Kee, H., Ly, T., Aziz, A., & Spasojevic, P. 2015 High-throughput FPGA-based QC-LDPC decoder architecture. Preprint. arXiv:quant-ph/1503.02986. * Mink et al. (2006) Mink, A., Tang, X., Ma, L., Nakassis, T., Hershman, B., Bienfang, J.C., Su, D., Boisvert, R., Clark, C.W., & Williams, C. J. 2006 High speed quantum key distribution system supports one-time pad encryption of real-time video. In _Defense and Security Symposium_ , International Society for Optics and Photonics, 62440M. * Mink & Bienfang (2013) Mink, A., & Bienfang, J. 2013 QKD on a Board Limited by Detector Rates in a Free-Space Environment. In _ICQNM 2013, The Seventh International Conference on Quantum, Nano and Micro Technologies_ , 28–33. * Mirza & Petruccione (2010) Mirza, A., & Petruccione, F. 2010 Realizing long-term quantum cryptography. _JOSA B_ , 27(6), A185–A188. * Mitsubishi (2015) Quantum Cryptography: Realizing next-generation security that cannot be broken. [R& D Highlights] 2015. Retrieved from http://www.mitsubishielectric.com/company/rd/research/highlights/communications/quantum.html * Nazarathy et al. (2006) Nazarathy, M., Tselniker, I., Regev, Y., Orenstein, M., & Katz, M. 2006 Integrated–optical realizations of quantum key distribution over maximally unbiased bases. _IEEE J. Sel. Top. Quant._ , 12(4), 897–913. * Noh (2009) Noh, T.-G. 2009 Counterfactual quantum cryptography. _Phys. Rev. Lett._ , 103(23), 230501. arXiv:0809.3979 * Patel et al. (2012) Patel, K.A., Dynes, J.F., Choi, I., Sharpe, A.W., Dixon, A.R., Yuan, Z.L., Penty, R.V. & Shields, A.J. 2012 Coexistence of high-bit-rate quantum key distribution and data on optical fiber. _Phys. Rev. X_ , 2(4), 041010. arXiv:1212.0033 * Portmann & Renner (2014) Portmann, C., & Renner, R. 2014 Cryptographic security of quantum key distribution. Preprint. arXiv:1409.3525 * Pritchard & Till (2010) Pritchard, J. & Till, S. 2014 UK Quantum Technology Landscape 2014. Engineering and Physical Sciences Research Council Report DSTL/PUB75620 Retrieved from http://www.epsrc.ac.uk/newsevents/pubs/dstl-uk-quantum-technology-landscape-2014/ * Restelli et al. (2009) Restelli, A., Bienfang, J.C., Mink, A., & Clark, C.W. 2009 Quantum key distribution at GHz transmission rates. In _SPIE OPTO: Integrated Optoelectronic Devices_ , International Society for Optics and Photonics, 72360L. * Sangouard et al. (2011) Sangouard, N., Simon, C., De Riedmatten, H., & Gisin, N. 2011 Quantum repeaters based on atomic ensembles and linear optics. _Rev. Mod. Phys._ , 83(1), 33. http://www.arxiv.org/abs/quant-ph/0906.2699 * Santori, Fattal, & Yamamoto (2010) Santori, C., Fattal, D., & Yamamoto, Y. 2010 _Single-photon Devices and Applications_. John Wiley & Sons. * Sarwar Pasha & Bala Ram (2014) Sarwar Pasha M.D. & Bala Ram, A. 2014 Key pre-distribution using quantum key channel— A survey. _International Journal of Computer Science and Mobile Applications_ , 2(3) 85–101. * Sasaki et al. (2011) Sasaki, M., Fujiwara, M., Ishizuka, H., Klaus, W., Wakui, K., Takeoka, M., Tanaka, A., Yoshino, K., Nambu, Y., Takahashi, S., Tajima, A., Tomita, A., Domeki, T., Hasegawa, T., Sakai, Y., Kobayashi, H., Asai, T., Shimizu, K., Tokura, T., Tsurumaru, T., Matsui, M., Honjo, T., Tamaki, K., Takusue, H., Tokura, Y., Dynes, J.F., Dixon, A.R., Sharpe, A.W., Yuan, Z.-L., Shields, A.J., Uchikoga, S., Legré, M., Robyr, S., Trinkler, P., Monat, L., Page, J.-B., Ribordy, G., Poppe, A., Allacher, A., Maurhart, O., Länger, T., Peev, M., & Zeilinger, A. 2011 Field test of quantum key distribution in the Tokyo QKD Network. _Opt. Express_ , 19(11), 10387–10409. arXiv:1103.3566 * Scarani et al. (2004) Scarani, V., Acín, A., Ribordy, G. & Gisin, N. 2004 Quantum Cryptography Protocols Robust against Photon Number Splitting Attacks for Weak Laser Pulse Implementations. _Phys. Rev. Lett._ 92, 057901. * Scarani & Kurtsiefer (2014) Scarani, V., & Kurtsiefer, C. 2014 The black paper of quantum cryptography: Real implementation problems. _Theor. Comput. Sci._ , 560, 27–32. arXiv:0906.4547 * SECOQC (2007) Alléaume, R., Bouda, J., Branciard, C., Debuisschert, T., Dianati, M., Gisin, N., Godfrey, M., Grangier, P., Länger, T., Leverrier, A., Lütkenhaus, N., Painchault, P., Peev, M., Poppe, A., Pornin, T., Rarity, J., Renner, R., Ribordy, G., Riguidel, M., Salvail, L., Shields, A., Weinfurter, H., & Zeilinger, A. 2007 SECOQC White Paper on Quantum Key Distribution and Cryptography. Preprint. arXiv:quant-ph/0701168 * Simon et al. (2010) Simon, C., Afzelius, M., Appel, J., Boyer de La Giroday, A., Dewhurst, S.J., Gisin, N., Hu., C.-Y., Jelezko, F., Kröhl, S., Müller, J.H., Nunn, J., Polzik, E., Rarity, J., de Riedmatten, H., Rosenfeld, W., Shields, A.J., Sköld, N., Stevenson, R.M., Thew, R., Walmsley, I., Weber, M., Weinfurter, H., Wrachtrup, J., & Young, R.J. 2010 Quantum memories. _Eur. Phys. J. D._ , 58(1), 1–22. arXiv:1003.1107 * Tamarat et al. (2006) Tamarat, Ph., Gaebel, T., Rabeau, J.R., Khan, M., Greentree, A.D, Wilson, H., Hollenberg, L.C.L., Prawer, S., Hemmer, P., Jelezko, F., & Wrachtrup J. 2006 Stark shift control of single optical centers in diamond. _Phys. Rev. Lett._ , 97(8), 083002. arXiv:quant-ph/0607170 * Tang et al. (2014) Tang, Z., Liao, Z., Xu, F., Qi, B., Qian, L., & Lo, H.K. 2014 Experimental demonstration of polarization encoding measurement-device-independent quantum key distribution. _Phys. Rev. Lett._ , 112(19), 190503. http://2013.qcrypt.net/contributions/Tang-abstract.pdf * Tomamichel et al. (2012) Tomamichel, M., Lim, C.C.W., Gisin, N., & Renner, R. 2012 Tight finite-key analysis for quantum cryptography. _Nat. Commun._ , 3, 634. * Tsoi et-al. (2007) Tsoi, K.H., Leung, K.H., & Leong, P.H.W. 2007 High performance physical random number generator. _IET Comput. Digit. Tech._ , 1(4), 349–352. * Ursin _et al._ (2007) Ursin, R., Tiefenbacher, F., Schmitt-Manderbach, T., Weier, H., Scheidl, T., Lindenthal, M., Blauensteiner, B., Jennewein, T., Perdigues, J., Trojek, P., Ömer, B., Fürst, M., Meyenburg, M., Rarity, J., Sodnik, Z., Barbieri, C., Weinfurter, H.& Zeilinger, A. (2007) Entanglement-based quantum communication over 144 km. _Nature Phys._ , 3(7), 481–486. * Vallone _et al._ (2014) Vallone, G., Bacco, D., Dequal, D., Galarin, S., Luceri, V., Bianco, G., & Villoresi, P. 2014 Experimental Satellite Quantum Communications Preprint. arXiv:quant-ph/1406.4051 * Wang, Miki, & Fujiwara (2009) Wang, Z., Miki, S., & Fujiwara, M. 2009 Superconducting nanowire single-photon detectors for quantum information and communications. _IEEE J. Sel. Top. Quantum Electron._ , 15(6), 1741–1747. * Wang (2013) Wang, X.B. 2013 Three-intensity decoy-state method for device-independent quantum key distribution with basis-dependent errors. _Phys. Rev. A_ , 87(1), 012320. arXiv:1308.5677 * Wang et al. (2014) Wang, S., Chen, W., Yin, Z.-Q., Li, H.-W., He, D.-Y., Li, Y.-H., Wang, D., Chen, H., Han, Y.-G., Huang, J.-Z., Guo, J.-F., Hao, P.-L., Li, M., Zhang, C.-M., Liu, D., Liang, W.-Y., Miao, C.-H., Wu, P., Guo, G.-C., & Han, Z.F. 2014 Field and long-term demonstration of a wide area quantum key distribution network. _Opt. Express_ , 22(18), 21739–21756. arXiv:1409.1568 * Weisner (1983) Wiesner, S. 1981 Conjugate coding. _SIGACT News_ , 15(1), 78–88. Original manuscript written circa 1969. * Xi, Qi, & Lo (2010) Xu, F., Qi, B., & Lo, H.K. 2010 Experimental demonstration of phase-remapping attack in a practical quantum key distribution system. _New J. Phys._ , 12(11), 113026. arXiv:1005.2376 * Xiang & Benkrid (2009) Xiang, T., & Benkrid, K. 2009 Mersenne Twister Random Number Generation on FPGA, CPU and GPU. _NASA/ESA Conference on Adaptive Hardware and Systems_ , 460–464. * Yikra (2014) Yikra,B. 2014 Researchers bounce polarized photons off satellites to show feasibility of space based quantum communications. _Phys.org_. Retrieved from http://phys.org/news/2014-06-polarized-photons-satellites-feasibility-space.html * Yuan (2007) Yuan, Z.L., Sharpe, A.W., Shields, A.J. 2007 Unconditionally secure one-way quantum key distribution using decoy pulses. _Appl. Phys. Lett._ , 90(1), 011118. arXiv:quant-ph/0610015 * Yuen (2013) Yuen, H.P. 2013 Essential elements lacking in security proofs for quantum key distribution. In _Proceedings of SPIE- The International Society for Optical Engineering_. arXiv:quant-ph/1310.0842 * Zhang et al. (2009) Zhang, Z., Anantharam, V., Wainwright, M.J., & Nikolic, B. 2009 A 47 Gb/s LDPC decoder with improved low error rate performance. In _2009 Symposium on VLSI Circuits_ , IEEE, 286–287. * Zhang et al. (2012) Zhang, H.F., Wang, J., Cui, K., Luo, C.L., Lin, S.Z., Zhou, L., Liang, H., Chen, T.-Y., Chen, K., & Pan, J.W. 2012 A real-time QKD system based on FPGA. _J. Lightwave Technol._ , 30(20), 3226–3234. arXiv:quant-ph/1301.2383 * Zhang et al. (2014) Zhang, C.-M., Li M., Huang J.-Z., Treeviriyanupab, P., Li, H.-W., Li, F.-Y., Wang, C., Yin, Z.-Q., Chen, W., Sripimanwat, K., & Han, Z.-F., 2013 Fast implementation of length-adaptive privacy amplification in quantum key distribution. _Chinese Phys. B_ , 23(9), 090310. * Zhao et al. (2008) Zhao, Y., Fung, C.H.F., Qi, B., Chen, C., & Lo, H.K. 2008 Quantum hacking: Experimental demonstration of time-shift attack against practical quantum-key-distribution systems. _Phys. Rev. A_ , 78(4), 042333. arXiv:0704.3253 * Zhu et al. (2011) Zhu, C.-H., Pei, C.-X., Quan, D.-X., Gao, J.-L., Chen, N., Yi, Y.-H., 2010 A new quantum key distribution scheme based on frequency and time coding. _Chinese Phys. Lett._ , 27(9), 090301. arXiv:1003.2472
# PP-MeT: a Real-world Personalized Prompt based Meeting Transcription System ###### Abstract Speaker-attributed automatic speech recognition (SA-ASR) improves the accuracy and applicability of multi-speaker ASR systems in real-world scenarios by assigning speaker labels to transcribed texts. However, SA-ASR poses unique challenges due to factors such as speaker overlap, speaker variability, background noise, and reverberation. In this study, we propose PP-MeT system, a real-world personalized prompt based meeting transcription system, which consists of a clustering system, target-speaker voice activity detection (TS- VAD), and TS-ASR. Specifically, we utilize target-speaker embedding as a prompt in TS-VAD and TS-ASR modules in our proposed system. In constrast with previous system, we fully leverage pre-trained models for system initialization, thereby bestowing our approach with heightened generalizability and precision. Experiments on M2MeT2.0 Challenge dataset show that our system achieves a cp-CER of 11.27% on the test set, ranking first in both fixed and open training conditions. Index Terms— SA-ASR, TS-VAD, TS-ASR, personalized prompt, M2MeT2.0 Challenge ## 1 Introduction The rapid advancements in deep learning have led to remarkable strides in automatic speech recognition (ASR), substantially enhancing its overall performance. Despite these achievements, ASR systems continue to face challenges in real-world far-field scenarios, such as meetings or home parties, where background noise, unavoidable reverberation, and overlapping speech from multiple speakers can significantly degrade their performance. In order to develop a robust ASR system in such challenging acoustic environments, numerous research studies have concentrated on multi-channel multi-party speech recognition and diarization within dinner party scenarios [1, 2]. The objective of the M2MeT2.0 challenge [3, 4] is to address the ASR task in multi-party meetings, which involves providing precise transcriptions and identifying the corresponding speakers. To advance the practical application of current multi-speaker speech recognition systems, the M2MET 2.0 Challenge evaluates the task of Speaker-attributed ASR (SA-ASR). Additionally, the challenge includes two sub-tracks: fixed training condition track and open training condition track. Speaker-attributed ASR (SA-ASR) poses several challenges due to the complexity of accurately attributing speech to specific speakers. The SA-ASR task improves the accuracy and applicability of multi- speaker ASR systems in real-world scenarios by assigning speaker labels to transcribed texts. Unlike traditional ASR systems that transcribe speech without considering speaker identities, SA-ASR goes a step further by associating each recognized word or phrase with the corresponding speaker. SA-ASR faces unique challenges due to factors like speaker overlap, speaker variability, background noise, and reverberation. Overcoming these challenges involves developing advanced algorithms and techniques for speaker diarization, speech separation, and speaker recognition to accurately attribute spoken words to their respective speakers. The development of SA-ASR systems has the potential to improve the performance and usability of speech recognition in scenarios where multiple speakers are present, enabling applications that require speaker-specific information and analysis. In this study, we present the PP-MeT system, a personalized-prompt based meeting transcription system designed to address the ASR task in multi-party meetings. Our approach comprises three essential components: a clustering system, target-speaker voice activity detection (TS-VAD), and target-speaker ASR (TS-ASR). To enhance the system’s performance and applicability, we integrate target-speaker embeddings as prompts within the TS-VAD and TS-ASR modules. Leveraging pre-trained models during system initialization further empowers our approach, granting it superior generalizability and precision. In experiments conducted on the M2MeT2.0 dataset, our integrated PP-MeT system achieves a concatenated minimum permutation character error rate (cp-CER) of only 11.27% on the test set, achieving the top position in both fixed and open training conditions. We also release our inference system with pre-trained models at website111https://github.com/XimalayaEverestIntelligentLab/M2MET2.0. The rest of this paper is organized as follows. In Section 2, we detail the architecture of the PP-MeT system. Datasets and experimental setup are described in Section 3. Section 4 presents the experimental results of M2MeT2.0 Challenge test set and our ablation study. Finally, we conclude in Section 5. ## 2 Proposed System Description The overview of our proposed PP-MeT system for the M2MeT2.0 Challenge is shown in Figure 1. Fig. 1: The overview of our proposed PP-MeT system. ### 2.1 Speaker Embedding System As M2MeT2.0 encourage the participants to use pre-trained models, we use two pre-trained models222https://github.com/wenet-e2e/wespeaker/blob/master/docs/pretrained.md from Wespeaker toolkit [5, 6]. One is Resnet34 from the CN-Celeb example, and another is ResNet34-LM, which is obtained by further training ResNet34 with a large-margin technique. We also train a ResNet34 model with Speechbrain toolkit333https://github.com/speechbrain/speechbrain/tree/develop to introduce diversity to our speaker embedding model. We will refer to these three speaker embedding models as SV-1/2/3 and the corresponding personalized prompt as Prompt-1/2/3 for simplicity. ### 2.2 Clustering System Before proceeding to TS-VAD and TS-ASR systems, we need to estimate the number of speakers and initialize personalized prompts using clustering algorithm. First, we extract voice speech segments based on VAD results for each session. Then we split each segment into subsegments using a fixed 3s window size and 1.5s window-shift. After that, we use speaker embedding model to extract embedding for each subsegment. Finally, we feed the L2-normalized embedding into the clustering algorithm and obtain the speaker number for each session, as well as the label for each subsegment. We use DOVER-lap toolkit [7]444https://github.com/desh2608/dover-lap to merge clustering results from different channel and speaker embedding models. We compare Auto-tuning Spectral Clustering with Normalized Eigen Gap [8](NME-SC) with Agglomerative Hierarchical Clustering(AHC) algorithm. As NME-SC outperforms AHC by a large margin, we use NME-SC algorithm results to initialize personalized prompts. After obtaining the clustering system result, for each speaker, we extract speech that contains only the targeted speaker as personalized speech. Then we repeat the speaker embedding extraction steps over the personalized speech and use the mean-pooled L2-normalized speaker embedding as the personalized prompt. ### 2.3 TS-VAD System As the clustering system can not handle overlap speech, it results in a high miss error in multi-party meeting scenarios. To further reduce DER, we use TS- VAD system to give a more accurate estimate of speaker labels. We use ResNet34 model as backbone for our TS-VAD system, which is the same as that of speaker embedding model. First, we extract the pooling layer input as frame-level speaker embedding. Then we do a stats-pooling with 3-second stride to extract the frame level mean and std feature, and concatenate it with the original frame-level speaker embedding. We do mean-pooling and attention- pooling for frame-level speaker embedding and personalized prompts, respectively. After that, we use a conformer decoder layer to explore the relationship between the frame-level speaker embedding and personalized prompt. We feed the frame-level speaker embedding feature as conformer decoder input, and each personalized prompt as decoder memory. Finally, we concatenate the conformer decoder layer output and use a BiLSTM layer to explore the relationship among each speaker. The BiLSTM output is fed into a fully- connect(FC) layer with a sigmoid activation function to generate the final TS- VAD probability[9, 10]. The detailed TS-VAD model structure is shown in Figure 2. Fig. 2: TS-VAD model structure ### 2.4 TS-ASR system Far-Field ASR poses a greater challenge compared to ASR of speech captured by a close-proximity microphone due to the degraded quality of the signal. To address this, we endeavor to engage in speech enhancement. In our practical pursuit, there exist two pivotal components. Firstly, we employ a sophisticated dereverberation method based on weighted prediction error (WPE) [11] to mitigate the effects of late reverberation. In the challenge, we utilize an accelerated GPU-version of WPE, incorporating the following parameters: taps=12, delay=2, iterations=3. Secondly, in order to further attenuate late reverberation and minimize noise interference, the weighted delay-and-sum acoustic beamforming (BeamformIt) method [12] is employed. As M2MeT2.0 requires participants to give transcription for each speaker, we upgrade the traditional ASR model into TS-ASR system with personalized prompt module, which enables it to yield different transcription given different personalized prompt [13, 14]. We feed the personalized prompt into a FC layer, and do Hadamard product with the output from the first layer of both asr encoder and decoder. As our TS-ASR model makes little modification to the traditional ASR model, we can easily adapt a pre-trained ASR model into a TS- ASR model. We use Unfied-Conformer [15] model pretrained on wenetspeech555https://github.com/wenet-e2e/wenet/blob/main/docs/pretrained_models.md from [16] as the TS-ASR model backbone. The detailed TS-ASR model structure is shown in Figure 3. Fig. 3: TS-ASR model structure ## 3 Experimental Setup ### 3.1 Datasets The original M2MeT1.0 dataset [3] contains 118.75 hours of speech data in total. The dataset is divided into 104.75 hours for training, 4 hours for development (denoted as Dev 1.0), and 10 hours as test set (denoted as Test 1.0) for scoring and ranking in M2MeT1.0 Challenge. Test 1.0 is used as development set in M2MeT2.0 Challenge. M2MeT2.0 uses a new 10 hours dataset (denote as Test 2.0) as test set. AISHELL4 [17] is a real-recorded Mandarin speech dataset collected by 8-channel circular microphone array for speech processing in a conference scenario. This dataset consists of 211 recorded meeting sessions, each containing 4 to 8 speakers, with a total length of 120 hours, aiming to bridge the advanced research on multi-speaker processing and the practical application scenarios. CN-Celeb [18] is a large-scale speaker recognition dataset collected ‘in the wild’. This dataset contains more than 130, 000 utterances from 1, 000 Chinese celebrities, and covers 11 different genres in the real world. Both M2MeT and AISHELL4 datasets are far-field multi-channel datasets, while the CN-Celeb dataset is a near-field dataset. Figure 4 shows the data preparation. By Oracle VAD, the non-overlap speech of each speaker is obtained from both near-field and far-field data. Then the personalized prompt is extracted. The M2MeT dataset is processed according to the given prior information into continuous voice speech. All far-field multi-channel datasets are pre-processed to generate enhanced data by WPE and BF. The original far- field 8-channel data and the enhanced data compose the speech of each speaker, which is used in the next training process. Fig. 4: Data preparation before training. The data flow of each training process is shown in Figure 5. The near-field data is processed into 3-second uniform segments and used in speaker embedding training. In TS-VAD model training, the continuous voice speech and non- overlap speech with online augmentation are processed into 16-second uniform segments, and the target-speaker embedding is used as a prompt. Moreover, the speech segment of each speaker and the personalized prompt are used in TS-ASR model training. Fig. 5: Data flow in each training process. ### 3.2 System Setup For all systems, we use 80-dimension log-mel filter bank (Fbank) feature as input. The Fbank feature is extracted using a 25ms window length and 10ms window shift. #### 3.2.1 Speaker embedding system we use CN-Celeb data [18] to train our speaker embedding model and split each utterance into 3s uniform length segments. When iterating over all segments, we introduce diversity by randomly offsetting the start frame of the segments from -1.5s to 1.5s. All these three speaker embedding models are trained using AAM softmax loss [19] and generate 256 dimension speaker embedding as output. We use a cyclical learning rate policy to dynamically adjust the lr for 16 epochs. #### 3.2.2 TS-VAD system we use M2MeT2.0 training data and Aishell-4 data for training. For each session, first, we extract and combine all voiced speech as our real training data. Then, for each speaker, we extract and combine speech that contains only the target speaker as personalized speech. Finally, we initialize Prompt-1/2/3 using personalized speech. If the number of speaker is less than 4, we pad Prompt-1/2/3 using zero vectors. During training, we split the real training data into 16s segments and iterate over each segment. We also do an online data simulation by choosing personalized speech from random speakers to fill up the voiced region of real data[20]. It is important that the randomly chosen speakers are from the same session, in case the model learns background noise feature of each session, rather than the essential difference of each speaker. We train three TS-VAD models based on SV-1/2/3 and Prompt-1/2/3. For all TS- VAD models, we use 2 layers, 256-dimension input, 512-dimension hidden dimension, and 8 heads for the conformer decoder setup. We use 2 layers, 1024 dimension input, and 512 hidden dimensions for BiLSTM setup. TS-VAD training consists three key stages. In stage 1, we copy the pre-trained speaker embedding parameter into the TS-VAD model, freeze the backbone part and train the model using real and simulated data until convergence with 1e-3 lr. In stage 2, we train the whole model using real and simulated data until convergence with 1e-4 lr. In stage 3, we finetune the whole model only using real data with 1e-5 lr. We choose the model with the lowest DER on Test 1.0 for decoding. During TS-VAD decoding, we initialize Prompt-1/2/3 from clustering system. We can iterate over the TS-VAD system by re-initialize Prompt-1/2/3 using TS-VAD system outputs. #### 3.2.3 TS-ASR system We use the WeNet toolkit and its pre-trained Unified-Conformer model on WeNetSpeech as backbone. Since M2MeT2.0 and Aishell-4 training data comprise multiple channels, on one hand, we directly feed the model with raw mean- pooled data, on the other hand, we feed the model with enhanced single-channel data. Additionally, we incorporate speed augmentation techniques during the training process. It is imperative to note that when the audio speed is altered, the corresponding personalized prompt for that particular speed variation should be rendered. We also train three TS-ASR systems based on Prompt-1/2/3. For all TS-ASR models, we use 12-layer conformer encoder with 512 dimension output, 2048 dimension linear units, and 8 attention heads. We use a 3-layer bi-transformer decoder with 2048 dimension linear units and 8 attention heads. For the personalized prompt module, we feed the 256 dimension personalized prompt into a FC layer, project it into a 512 dimension vector and do a Hadamard product with the first layer output of both encoder and decoder. TS-ASR training also consists three key stages. In stage 1, we freeze the Unified-Conformer backbone, and only train the personalized prompt module using raw data and enhanced data. In stage 2, we train the whole model with 1e-4 lr. In stage 3, we finetune the whole model with 1e-5 lr using enhanced data. ## 4 Experimental Results ### 4.1 Results on M2MeT2.0 Challenge M2MeT2.0 challenge uses concatenated minimum permutation character error rate (cp-CER) as the evaluation metric. It computes the minimum CER given all speaker permutations, which requires the system to give the correct transcription and speaker label. The calculation of cp-CER is divided into three steps. First, recognition results and reference transcriptions belonging to the same speaker are concatenated on the timeline in a session. Second, the character error rate (CER) of all permutations of speakers is calculated. Finally, the lowest CER is selected as the cp-CER. Table 1 presents the cp-CER results of the official baseline and each competition system. Our system achieves 15.05%, 16.84%, and 11.27% cp-CER on Dev 1.0, Test 1.0, and Test 2.0, respectively. Notice that cp-CER on Dev 1.0 and Test 1.0 is achieved using oracle diarization result. We can observe that our PP-MeT model gives better results over the official baseline and achieve up to 30.28% absolute cp-CER improvement due to the enhanced dataset and advanced model architectures. achieving first place in the challenge. System | Dev 1.0 | Test 1.0 | Test 2.0 ---|---|---|--- PP-MeT (Rank 1st) | 15.05 | 16.84 | 11.27 Rank 2nd Team | – | – | 18.64 Rank 3rd Team | – | – | 22.83 Rank 4th Team | – | – | 23.51 Rank 5th Team | – | – | 24.82 Official Baseline | 47.4 | 52.57 | 41.55 Table 1: The cp-CER (%)$\downarrow$ results of each competition system on the M2MeT Dev 1.0, Test 1.0, and Test 2.0. ### 4.2 Ablation Study We conduct a detailed ablation study to better understand the contribution of cp-CER from each system, and the significance of pre-trained models. Clustering Method | SV Model | Channel 1-8 DER(%) $\downarrow$ | DER (%)_Channel $\downarrow$ | DER (%)_Model $\downarrow$ ---|---|---|---|--- SC | SV-1 | 16.87 | 16.21 | 16.89 | 17.49 | 18.61 | 18.00 | 16.09 | 18.85 | 16.40 | 15.22 SV-2 | 15.96 | 16.22 | 16.41 | 16.39 | 16.86 | 16.57 | 16.55 | 17.99 | 15.22 SV-3 | 17.26 | 16.18 | 17.16 | 16.97 | 17.07 | 16.55 | 16.52 | 17.11 | 15.75 AHC | SV-1 | 28.22 | 27.11 | 26.92 | 25.79 | 25.90 | 24.52 | 26.61 | 26.25 | 22.95 | 22.43 SV-2 | 26.48 | 23.85 | 26.13 | 26.70 | 25.28 | 25.61 | 24.58 | 26.65 | 22.43 SV-3 | 25.90 | 26.74 | 29.33 | 27.19 | 27.51 | 27.02 | 26.09 | 25.87 | 22.94 Table 2: DER Results for each clustering system on Test 1.0 #### 4.2.1 Clustering System As clustering system gives the estimate of speaker number and rough speaker label, its performance determines the superior limit of the whole PP-MeT system. In Table 2, we study the impact of different speaker embedding models and clustering algorithms in clustering systems. SV-1/2/3 achieves 7.13%, 6.49%, and 7.06% on CN-Celeb dev trials, respectively. The threshold for AHC clustering is tuned on Dev 1.0. Results show that given each model and channel, NME-SC outperforms AHC significantly. DOVER-lap makes the clustering result more stable by leveraging clustering results from different channels and models. As the accuracy of speaker embedding directly affects the quality of speaker embedding, DER relates to speaker embedding performance evidently. The lowest DER is achieved by SV-2, which also achieves the lowest EER on CN-Celeb trials. #### 4.2.2 TS-VAD system In Table 3, we study the impact of pre-trained speaker embedding model and different model architectures in TS-VAD system. Results show that pre-trained model contributed heavily to the performance of TS-VAD system. If TS-VAD model backbone parameter is randomly initialized, it only achieves 13.28% DER on Test 1.0, which is only slightly better than that of clustering system. Also, TS-VAD model backbone should match that of the personalized prompt. If we initialize the TS-VAD model backbone parameter using pre-trained ECAPA-TDNN speaker embedding model and train with Prompt-1. It achieves 7.68% DER, which is much worse than its counterpart using matched speaker embedding model and prompt. The above results demonstrate the importance of pre-trained models in TS-VAD system, and using matched speaker embedding model for initialization and personalized prompt makes it easier to explore the relationship between frame-level speaker embedding and personalized prompt. We can also observe that the DER drops moderately if we iterate the TS-VAD system by refining Prompt-1/2/3 using TS-VAD system output. Initialization Model Parameter | Personalized Prompt | DER (%)_Iter0 $\downarrow$ | DER (%)_Model $\downarrow$ | DER (%)_Iter1 $\downarrow$ | DER (%)_Model $\downarrow$ ---|---|---|---|---|--- SV-1 | Prompt-1 | 5.22 | 3.19 | 4.87 | 2.99 SV-2 | Prompt-2 | 4.52 | 4.25 SV-3 | Prompt-3 | 4.02 | – | 3.64 Random | Prompt-1 | 13.28 | – ECAPA | Prompt-1 | 7.68 | – Table 3: DER Results for each TS-VAD model on Test 1.0 #### 4.2.3 TS-ASR System In Table 4, we study the impact of pre-trained models and personalized prompts in TS-ASR system. First, we try to finetune the pre-trained unified-conformer ASR model directly without any structure modification. Results show that the pre-trained ASR model achieves 32.63% and 35.89% cp-CER on Dev 1.0 and Test 1.0. After finetuning the model on M2MET2.0 and Aishell-4 data, the cp-CER drops to 22.55% and 26.43%, respectively. However, the cp-CER improvement is largely due to the model performance on nonoverlap speech. It fails to decrease further because the traditional ASR model cannot handle overlap speech. Then, we try to train the TS-ASR model from scratch with Prompt-1. However, the TS-ASR model with a unified-conformer backbone fails to converge. This demonstrates the necessity of pre-trained ASR model backbone in our TS-ASR system. Finally, we train three TS-ASR models based on the Prompt-1/2/3. cp-CER on TS- ASR model with pretrained ASR model backbone and Prompt-1/2/3 drops dramatically both on Dev 1.0 and Test 1.0. The result shows that pre-trained ASR model with Prompt-2 achieves the lowest cp-CER, which means that the performance of pre-trained speaker embedding model also affects the performance of TS-ASR on overlapped speeches. We also try to finetune the TS-ASR model further using LF-MMI with k2 toolkit666https://github.com/k2-fsa/k2, and introducing LM information by decoding with HLG. However, the cp-CER fails to drop on both Dev 1.0 and Test 1.0. This is due to the fact that in multi-party meeting scenario, the transcription from each session is highly irrelevant. External LM information can not help to decrease cp-CER. In Table 4, Test 1.0 cp-CER is calculated using segments and prompts from TS- VAD system. The gap between cp-CER of Test 1.0 and Test 1.0 means the degradation introduced by TS-VAD system, which is approximately 2%. We obtain final results by leveraging each system results using SCTK rover toolkit777https://github.com/usnistgov/SCTK. Personalized Prompt | Dev 1.0 | Test 1.0 | Test 1.0 ---|---|---|--- – | 22.55 | 26.43 | – Prompt-1 | 15.35 | 17.20 | 19.45 Prompt-2 | 15.13 | 17.08 | 19.06 Prompt-3 | 15.28 | 17.16 | 19.06 Rover | 15.05 | 16.84 | 18.92 Table 4: cp-CER (%) results for each TS-ASR model on Dev 1.0, Test 1.0 and Test 1.0. ## 5 Conclusion In this paper, we present our PP-MET system for the Multi-channel Multi-party Meeting Transcription Challenge 2.0 (M2MeT2.0) to address the ASR task in a multi-party meeting scenario. Compared with the other conventional systems, we incorporate target-speaker embedding as a personalized prompt in both TS-VAD and TS-ASR stage. Moreover, to further enhance the system’s robustness and reduce the training cost, pre-trained models are used in our system’s initialization, enabling fast adaptation across all modules. Experimental results shows proposed system outperforms conventional systems by a large margin. In future work, we plan to explore the potential of expanding personalized prompts on the time axis. Additionally, we aim to enhance the TS-ASR model by jointly training the speaker embedding module with the ASR backbone, further improving its performance. ## References * [1] Jon Barker, Shinji Watanabe, Emmanuel Vincent, and Jan Trmal, “The fifth’chime’speech separation and recognition challenge: dataset, task and baselines,” arXiv preprint arXiv:1803.10609, 2018. * [2] Shinji Watanabe, Michael Mandel, Jon Barker, Emmanuel Vincent, Ashish Arora, Xuankai Chang, Sanjeev Khudanpur, Vimal Manohar, Daniel Povey, Desh Raj, et al., “Chime-6 challenge: Tackling multispeaker speech recognition for unsegmented recordings,” arXiv preprint arXiv:2004.09249, 2020. * [3] Fan Yu, Shiliang Zhang, Yihui Fu, Lei Xie, Siqi Zheng, Zhihao Du, Weilong Huang, Pengcheng Guo, Zhijie Yan, Bin Ma, et al., “M2met: The icassp 2022 multi-channel multi-party meeting transcription challenge,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6167–6171. * [4] Yuhao Liang, Mohan Shi, Fan Yu, Yangze Li, Shiliang Zhang, Zhihao Du, Qian Chen, Lei Xie, Yanmin Qian, Jian Wu, Zhuo Chen, Kong Aik Lee, Zhijie Yan, and Hui Bu, “The second multi-channel multi-party meeting transcription challenge (m2met) 2.0): A benchmark for speaker-attributed asr,” in arxiv preprint arxiv: 2309.13573, 2023. * [5] Hongji Wang, Chengdong Liang, Shuai Wang, Zhengyang Chen, Binbin Zhang, Xu Xiang, Yanlei Deng, and Yanmin Qian, “Wespeaker: A research and production oriented speaker embedding learning toolkit,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5. * [6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. * [7] Desh Raj, Leibny Paola Garcia-Perera, Zili Huang, Shinji Watanabe, Daniel Povey, Andreas Stolcke, and Sanjeev Khudanpur, “Dover-lap: A method for combining overlap-aware diarization outputs,” in 2021 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2021, pp. 881–888. * [8] Tae Jin Park, Kyu J Han, Manoj Kumar, and Shrikanth Narayanan, “Auto-tuning spectral clustering for speaker diarization using normalized maximum eigengap,” IEEE Signal Processing Letters, vol. 27, pp. 381–385, 2019. * [9] Weiqing Wang, Xiaoyi Qin, and Ming Li, “Cross-channel attention-based target speaker voice activity detection: Experimental results for the m2met challenge,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 9171–9175. * [10] Ivan Medennikov, Maxim Korenevsky, Tatiana Prisyach, Yuri Khokhlov, Mariya Korenevskaya, Ivan Sorokin, Tatiana Timofeeva, Anton Mitrofanov, Andrei Andrusenko, Ivan Podluzhny, et al., “Target-speaker voice activity detection: a novel approach for multi-speaker diarization in a dinner party scenario,” arXiv preprint arXiv:2005.07272, 2020. * [11] Tomohiro Nakatani, Takuya Yoshioka, Keisuke Kinoshita, Masato Miyoshi, and Biing-Hwang Juang, “Speech dereverberation based on variance-normalized delayed linear prediction,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 18, no. 7, pp. 1717–1731, 2010. * [12] Xavier Anguera, Chuck Wooters, and Javier Hernando, “Acoustic beamforming for speaker diarization of meetings,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 15, no. 7, pp. 2011–2022, 2007. * [13] Takafumi Moriya, Hiroshi Sato, Tsubasa Ochiai, Marc Delcroix, and Takahiro Shinozaki, “Streaming target-speaker asr with neural transducer,” Proc. Interspeech 2022, pp. 2673–2677, 2022. * [14] Hiroshi Sato, Tsubasa Ochiai, Marc Delcroix, Keisuke Kinoshita, Takafumi Moriya, and Naoyuki Kamo, “Should we always separate?: Switching between enhanced and observed signals for overlapping speech recognition,” arXiv preprint arXiv:2106.00949, 2021. * [15] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al., “Conformer: Convolution-augmented transformer for speech recognition,” arXiv preprint arXiv:2005.08100, 2020. * [16] Binbin Zhang, Di Wu, Zhendong Peng, Xingchen Song, Zhuoyuan Yao, Hang Lv, Lei Xie, Chao Yang, Fuping Pan, and Jianwei Niu, “Wenet 2.0: More productive end-to-end speech recognition toolkit,” arXiv preprint arXiv:2203.15455, 2022. * [17] Yihui Fu, Luyao Cheng, Shubo Lv, Yukai Jv, Yuxiang Kong, Zhuo Chen, Yanxin Hu, Lei Xie, Jian Wu, Hui Bu, et al., “Aishell-4: An open source dataset for speech enhancement, separation, recognition and speaker diarization in conference scenario,” arXiv preprint arXiv:2104.03603, 2021. * [18] Yue Fan, JW Kang, LT Li, KC Li, HL Chen, ST Cheng, PY Zhang, ZY Zhou, YQ Cai, and Dong Wang, “Cn-celeb: a challenging chinese speaker recognition dataset,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 7604–7608. * [19] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4690–4699. * [20] Weiqing Wang, Danwei Cai, Qingjian Lin, Lin Yang, Junjie Wang, Jin Wang, and Ming Li, “The dku-dukeece-lenovo system for the diarization task of the 2021 voxceleb speaker recognition challenge,” arXiv preprint arXiv:2109.02002, 2021.
# Self-Supervised Visual Place Recognition by Mining Temporal and Feature Neighborhoods Chao Chen1, Xinhao Liu1, Xuchu Xu1, Yiming Li1, Li Ding2, Ruoyu Wang1, and Chen Feng✉,1 ✉ Corresponding author.1Chao Chen, Xinhao Liu, Xuchu Xu, Yiming Li, Ruoyu Wang, and Chen Feng are with New York University, Brooklyn, NY 11201, USA <EMAIL_ADDRESS>Ding is with University of Rochester, Rochester, NY 14627, USA<EMAIL_ADDRESS> ###### Abstract Visual place recognition (VPR) using deep networks has achieved state-of-the- art performance. However, most of them require a training set with ground truth sensor poses to obtain positive and negative samples of each observation’s spatial neighborhood for supervised learning. When such information is unavailable, temporal neighborhoods from a sequentially collected data stream could be exploited for self-supervised training, although we find its performance suboptimal. Inspired by noisy label learning, we propose a novel self-supervised framework named TF-VPR that uses temporal neighborhoods and learnable feature neighborhoods to discover unknown spatial neighborhoods. Our method follows an iterative training paradigm which alternates between: (1) representation learning with data augmentation, (2) positive set expansion to include the current feature space neighbors, and (3) positive set contraction via geometric verification. We conduct comprehensive experiments on both simulated and real datasets, with either RGB images or point clouds as inputs. The results show that our method outperforms our baselines in recall rate, robustness, and heading diversity, a novel metric we propose for VPR. Our code and datasets can be found at https://ai4ce.github.io/TF-VPR/. ## I Introduction Visual place recognition (VPR), which aims to identify previously visited places based on current visual observation, is a well-known problem in computer vision and plays a crucial role in autonomous robots [1]. Meanwhile, VPR is closely related to re-localization [2], loop closure detection [3], and image retrieval [4]. Despite all the efforts, VPR remains a difficult task due to various challenges such as perceptual aliasing and view direction differences [5, 6]. Classic VPR methods based on hand-crafted feature matching do not require supervised learning, but are less robust to the challenges mentioned above [7, 8, 9]. Therefore, learning-based methods have been proposed to learn local or global feature descriptors [10] by place classification [11] or contrastive-like similarity learning [12]. Some works also use a short image sequence instead of a single image to mitigate the issue of perceptual aliasing [13, 14, 15]. So far, most learning-based VPR methods are supervised, focusing on either learning better feature representations or designing robust matching strategies. They assume that accurate positions (and sometimes orientations) are available in their training set, for obtaining the positive and negative neighbors of each visual observation [12, 10], or defining place categories [11]. However, this information could be onerous to obtain due to GPS errors and its indoor unavailability or SLAM robustness challenges, especially at a large scale. Considering human’s extraordinary VPR ability that does not seem to need ground truth pose information for its training, we ask the following question: is it possible to relax such an assumption and design a learning- based VPR approach without pose-dependent supervision? (a) A partial sensory stream & 3 types of neighbors (red/blue/green circles). (b) A spatial neighborhood’s iterative update. Figure 1: Our idea is based on the interconnections between the temporal, spatial, and feature neighborhoods in sensory data: a query’s spatial neighborhood expands from its temporal to feature neighbors, then contracts to exclude wrong neighbors, iterated in training until such neighborhoods’ convergence. To achieve this goal, our main idea is to leverage fixed temporal neighborhoods and learnable feature neighborhoods to discover the unknown spatial neighborhoods (which require ground truth pose to compute), leading to a self-supervised VPR method shown in Fig. 1. We are inspired by research work utilizing sensory streams (RGB videos or point cloud sequences) to obtain the positive and negative neighbors in the temporal domain such as [16]. However, we find that VPR learned from temporal cues alone will miss spatial neighboring places with large viewpoint differences, because temporal neighbors tend to share similar viewpoints. This is suboptimal in applications such as visual navigation or loop closure for SLAM. To automatically discover the true spatial neighbors with diverse viewpoints, we propose a novel iterative learning strategy inspired by noisy label training such as bootstrapping [17]. More specifically, we exploit the temporal information for label initialization as shown in Fig. 1(a). We further augment the labels by simulating observations perceived from different view directions. Afterward, a feature representation will be learned based on the current labels. Finally, as shown in Fig. 1(b), we re-label the dataset using the learned feature space by: (1) adding feature-space neighbors as tentative positive labels and (2) rejecting false positives via geometric verification, in order to further refine the feature representation. Note that the above steps are iteratively conducted until convergence. To evaluate our self-supervised VPR with Temporal and Feature neighborhood interactions (TF-VPR), we simulate datasets with different input modalities, and develop a real-world dataset NYU-VPR-360 including two scenes with 38,426 GPS-georeferenced images in total. All the datasets are sequentially-collected sensory streams. Meanwhile, we develop a novel metric to measure the heading diversity of the retrieval results. In summary, our contributions are: 1. 1. We propose a novel self-supervised VPR solution termed TF-VPR that eliminates pose-dependent supervision by mining temporal and feature neighborhoods. 2. 2. We propose a new evaluation metric to assess the heading diversity of VPR retrieval results. 3. 3. We conduct comprehensive experiments in both simulation and real-world to demonstrate the advantages of our solution compared with other baseline methods. Our codes and data will be released with this paper. ## II Related Work Visual place recognition. Visual place recognition (VPR) is the problem of identifying a previously visited place based on visual information [5]. Existing VPR methods mainly lie in two categories: (1) traditional VPR techniques using hand-crafted features [18, 19, 20, 21, 22, 23], and (2) state-of-the-art VPR techniques using deep learning [12, 24, 25, 26, 27]. Among those, NetVLAD [12] is a seminal deep-learning-based VPR framework, followed by various research extensions such as (1) learning powerful feature representation [28, 29], (2) designing robust matching strategies [13, 14, 15], and (3) investigating different input modalities in VPR [30, 31, 32, 33]. However, most methods are supervised by pose-dependent data [24, 25, 12, 26]. To relax such a constraint, several attempts have been made [34, 35, 36, 37], yet failed to handle diverse re-visiting viewpoints. The most relevant work to our method is the semi-parametric topological memory (SPTM) [16], which utilizes temporal positives and negatives to train a binary classification network for adding edges for topological mapping, similar to VPR. However, since temporal neighbors tend to have very similar viewpoints, SPTM still struggles to recognize revisits of the same place from different viewpoints. To the best of our knowledge, no research has addressed self-supervised VPR that can recognize places observed from various viewpoints as either 2D images or 3D point clouds. Noisy label learning. Noisy labels become a problem as training data size increases, resulting in degraded performance [38]. To mitigate the issues of noisy labels, several learning attempts [39] have been made from directions including latent variable optimization [40], loss function design [41], and pseudo-label-based self-training [42, 17, 43, 44]. Among all pseudo-label methods, label refurbishment was firstly introduced by Bootstrapping [17]. Later on, another method addressed this problem using a self-training-based approach with an iterative workflow [43] which was used as a baseline architecture in their more recent framework confidence regularized [44]. Our method’s iterative training paradigm takes inspiration from [43, 44]. Contrastive learning. It is a self-supervised learning technique to find feature representations that differentiate similar data pairs from dissimilar ones without labels. Data augmentation is often used, and the learning objective is to decrease feature distances between the original and augmented images (positive samples), while increasing those distances between different images (negative samples) [45, 46, 47, 48]. In VPR, NetVLAD [30] uses the triplet loss which is similar to contrastive learning, yet relies on ground truth pose to define positive/negative samples. Building on NetVLAD, we further adopt data augmentation to synthesize observations captured by different viewpoints in order to learn more robust features for VPR. VPR evaluation. There are several evaluation metrics for visual place recognition, e.g., the popular AUC-PR [1] provides a good overview of precision and recall performance but is less indicative in the cases when ground truth match could take multiple values. Recall Rate@N, as used in [49, 50, 12], is designed to address such cases that the correct retrieval may be in the top-N results, and multiple correct query matches are neither penalized nor rewarded. However, existing VPR metrics rarely evaluate the viewpoint diversity of the retrieved results [5], which is important in downstream applications such as SLAM. In this work, we develop such a metric to fill this gap. It assesses a VPR model’s capacity to recognize places revisited from different directions. ## III Methodology Figure 2: Overview of TF-VPR. Labeling, training, expansion, and contraction are four major steps in our approach. Labels can be refurbished by iteratively learning feature representations (training), adding feature neighborhoods (expansion), and removing false positives by geometric verification (contraction). Initial labels are generated by temporal adjacency. We focus on a robot that is collecting a sensory stream of surround-view visual observations while navigating in a certain area. Our goal is to enable the robot to achieve visual place recognition (VPR) in the same spatial area where the data stream is collected, without relying on any frame-wise pose information. To this end, we utilize a learnable neural network $f_{\theta}$ parameterized by $\theta$ to map each visual observation to discriminative feature space for VPR. However, it is non-trivial to train such a neural network without ground-truth labels. In this work, we propose a novel iterative learning paradigm based on the following intuition: for the $i$-th query observation ${\bf{\mathbf{q}}}_{i}$, its spatial positive set $\mathcal{P}_{{\bf{\mathbf{q}}}_{i}}$ could be inferred by its temporal positive set $\mathcal{P}^{t}_{{\bf{\mathbf{q}}}_{i}}$ together with its tentative feature-space positive set $\mathcal{P}^{f}_{{\bf{\mathbf{q}}}_{i}}$. Our objective: auto-labeling with self-supervised learning. Different from existing supervised VPR methods that need to address the generalizability of learned models, we aim to solve VPR only for a certain spatial area where the data has been collected. Given an observation sequence, we want to automatically label each data frame’s spatial topology without ground truth poses for supervision. Similar to DeepMapping [51], we do not expect the learned $f_{\theta}$ to generalize its VPR ability either to other areas or to the same area but under different times/seasons/weather conditions than what has been covered in the training dataset. Note that this is non-trivial and useful in real-world, because (1) ground truth pose information is not always easy to obtain, and (2) achieving our objective would enable auto-labeling of large-scale datasets without relying on ground truth pose information, which can be further used to provide training dataset for supervised VPR methods. ### III-A Overall framework Given a sensory stream $\\{{\bf{\mathbf{o}}}_{i}\\}_{i=1}^{I}$ which can cover the entire spatial area (each observation denoted by ${\bf{\mathbf{o}}}_{i}$ is either an RGB image or a LiDAR point cloud), we aim to discover the spatial connectivity for these observations without any spatial information. Specifically, we propose a self-supervised learning paradigm that utilizes the interconnection of spatial, temporal, and feature neighborhoods to iteratively refine the noisy pseudo-labels as well as the feature representations, as shown in Fig. 2. The final output is the discovered spatial connectivity for the given observation sequence, as well as the learned neural network $f_{\theta}$ which can map each observation ${\bf{\mathbf{o}}}_{i}$ to a discriminative feature space that supports VPR. Note that the optimized $f_{\theta}$ can also discover the spatial connectivity for a new observation sequence collected in the same spatial area: for each query observation ${\bf{\mathbf{q}}}_{i}$, a set of nearest feature neighbors $\mathcal{P}^{f}_{{\bf{\mathbf{q}}}_{i}}=\\{{\bf{p}}_{j}\\}_{j\neq i}$ which can approximate the spatial neighbors is retrieved based on the Euclidean distance in the learned feature space: $d\left({\bf{\mathbf{q}}}_{i},{\bf{p}}_{j}\right)=||f_{\theta}\left({\bf{\mathbf{q}}}_{i}\right)-f_{\theta}\left({\bf{p}}_{j}\right)||$. The four sub-modules are detailed next. ### III-B Label initialization Temporal adjacency between any two data frames generally implies their spatial adjacency, but not vice versa [16]. Therefore, we could utilize temporal adjacency to generate noisy labels for each observation. Note that the term noisy means that such labels may include false negatives, e.g., it would overlook positive labels for observations that are spatially adjacent yet temporally distant. In this work, the labels are initialized solely with the knowledge of temporal adjacency: given a query ${\bf{q}}_{i}$, the temporal positive set is $\mathcal{P}^{t}_{{\bf{\mathbf{q}}}_{i}}=\\{{\bf{p}}_{j}\\}_{|i-j|<n}$, and the temporal negative set is $\mathcal{N}^{t}_{{\bf{\mathbf{q}}}_{i}}=\\{{\bf{n}}_{v}\\}_{|i-v|>u\times n}$, where $i,j,v$ denote the frame indices of the observations, $n$ controls the size of temporal positives, $u$ controls the range of temporal negatives. After initialization, the feature will be refined iteratively by inferring spatial neighborhoods according to the known temporal neighborhoods and current feature neighborhoods. ### III-C Training In order to learn a more robust feature representation for recognizing places viewed from diverse directions, we augment the training set by simulating various sensor headings, i.e., an observed panoramic image/point cloud is randomly rolled/rotated around the sensor’s vertical axis. Meanwhile, the pseudo ground-truth for the $e$-th training epoch includes spatial positives denoted by $\mathcal{P}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$ and spatial negatives denoted by $\mathcal{N}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$, both for the $i$-th query ${\bf{q}}_{i}$. Note that they are initialized by $\mathcal{P}^{t}_{{\bf{\mathbf{q}}}_{i}}$ and $\mathcal{N}^{t}_{{\bf{\mathbf{q}}}_{i}}$ for $e=1$ and will be iteratively updated. For the training objective, we employ weakly supervised ranking loss denoted by $L_{\theta}$ for the $i$-th training tuple $({\bf{q}}_{i},\\{{\bf{p}}_{j}\\},\\{{\bf{n}}_{v}\\})$, following [12]: $L_{\theta}=\sum_{i}\sum_{v}l\left(\min_{j}d\left({\bf{q}}_{i},{\bf{p}}_{j}\right)+m-d\left({\bf{q}}_{i},{\bf{n}}_{v}\right)\right),$ (1) where ${\bf{p}}_{j}\in\mathcal{P}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$ and ${\bf{n}}_{v}\ \in\mathcal{N}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$, $d$ is the Euclidean distance in feature space, $l$ is the hinge loss: $l(x)=\max(x,0)$, and $m$ defines the margin (set to $0.2$ in this paper). ### III-D Expansion At the initial stage (epoch $e=1$), the current positive label set $\mathcal{P}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$ has a limited size since we only incorporate the temporal positives. Actually, there could be a number of spatial positives in $\mathcal{N}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$ which are temporally distant yet spatially adjacent to the query. To improve the labels, we need to expand our positive set $\mathcal{P}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$ by including current feature space neighbors. Specifically, once the $f_{\theta}$ is trained for one epoch with current labels, the $K$ nearest neighbors (KNN) could be retrieved as the feature neighbors. Dynamic KNN. However, it is hard to use the same $K$ for different query observations, because places like intersections naturally have larger spatial neighborhoods than hallways. Thus, we need a frame-specific threshold. To design such a mechanism, we observe the following hypothesis: if an observation is closer to the query than at least one of the query’s temporal neighbors in the feature space, then such an observation is more likely to be a ground truth spatial neighbor of the query. Therefore, we firstly select $K$ nearest feature neighbors as the feature neighborhood candidates $\\{{\bf{\mathbf{o}}}_{k}\\}_{k=1}^{K}$. Then we utilize the maximum feature distance between the query and its temporal neighbors as a frame-specific threshold $\tau_{i}$ to determine which candidates should be included in the expanded positive set: $\tau_{i}=\max\limits_{{\bf{p}}_{j}\in\mathcal{P}^{t}_{{\bf{\mathbf{q}}}_{i}}}d\left({\bf{q}}_{i},{\bf{p}}_{j}\right),$ (2) where $\mathcal{P}^{t}_{{\bf{\mathbf{q}}}_{i}}$ is the temporal neighborhood set for query ${\bf{q}}_{i}$, $i$ and $j$ are the frame indices. Finally, any candidate ${\bf{o}}_{k}$ (selected from KNN) with a smaller feature distance to the query than the threshold $\tau_{i}$ will form the feature neighborhood set: $\mathcal{P}^{f}_{{\bf{\mathbf{q}}}_{i}}=\\{{\bf{\mathbf{o}}}_{k}\\}_{d\left({\bf{q}}_{i},{\bf{\mathbf{o}}}_{k}\right)<\tau_{i}}$. ### III-E Contraction To avoid potential false positives caused by noisy feature space during learning, we employ geometric verification to check the validity of feature neighborhoods $\mathcal{P}^{f}_{{\bf{\mathbf{q}}}_{i}}$ before using it to update $\mathcal{P}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$. We adopt different distance measures (independent of the neural network) based on input modality: the number of matching points for images [52] and the Chamfer distance after ICP for point clouds [53]. The verification strategy is similar to the one for KNN feature neighborhoods verification: we consider the maximum distance between the query and its temporal neighborhoods as the threshold. Then any candidate with a distance value within the threshold will pass the verification. The verified feature neighborhoods denoted by $\hat{\mathcal{P}}^{f}_{{\bf{\mathbf{q}}}_{i}}$ are permanently added into the positive set: $\mathcal{P}^{(e+1)}_{{\bf{q}}_{i}}=\mathcal{P}^{(e)}_{{\bf{q}}_{i}}\cup\hat{\mathcal{P}}^{f}_{{\bf{q}}_{i}},$ (3) where $e$ is the number of learning epochs. Note that $\mathcal{N}_{{\bf{\mathbf{q}}}_{i}}^{(e)}$ will also be updated similarly. Afterward, the model will be trained with the updated labels until convergence. ## IV Experiments We test TF-VPR in three different kinds of environments: simulated 2D point clouds [51], simulated RGB images [54], and real-world RGB images. Our codebase uses PyTorch [55] with network parameters optimized using Adam [56]. The learning rate is tuned to 0.001. We compare TF-VPR with both supervised and self-supervised baseline methods. ### IV-A Evaluation metrics Recall rate (Recall@N) is the ratio of successful retrievals to a total number of queries, where a successful retrieval means at least one of the top-N retrieved results is also the ground-truth spatial neighbor of the query. Ground truth can be obtained by K-D tree search on geographical location $(x,y,z)$ within a certain radius $R$. Note that when computing this metric, we need to exclude the temporal neighbors of a query from its top-$N$ retrievals. This is because in our auto-labeling setup, the temporal-based methods can easily overfit the temporal neighborhood, leading to uninformative evaluation with recall rates that are always close to $100\%$ if the temporal neighbors are kept in ground truth. Heading diversity (HD) measures the diversity of sensor headings of the true positives (w.r.t. the query’s heading) among the top-$|GT|$ retrievals. $|GT|$ is the size of ground-truth set that can be obtained based on a specific radius $R$ as described in Recall@N. First, we believe that those positives with headings different than the query are more valuable in downstream applications. Thus, we evenly divide the 360∘ range of headings into 8 angular bins in our setup, ignoring the first and last bins because they contain positives with similar headings w.r.t. the query. In this case, the $m$-th bin covers the heading difference range ${\mathcal{Q}_{m}}$ as: $\mathcal{Q}_{m}=[m\times 45^{\circ},(m+1)\times 45^{\circ}],\,m\in[1,2,3,4,5,6].$ (4) Then, we define HD for a query $\bf{q}$ as the bin coverage ratio between the true positives and the ground truth: $\text{HD(\bf{q})}=\frac{\sum_{m\in{[1...6]}}\mathds{1}(\exists\bf{x}\in\tilde{\mathcal{P}}^{|GT|}_{{\bf{\mathbf{q}}}}\land(\theta_{\bf{q}}-\theta_{\bf{x}})\in\mathcal{Q}_{m})}{\varepsilon+\sum_{m\in{[1...6]}}\mathds{1}(\exists\bf{y}\in\mathcal{P}^{|GT|}_{{\bf{\mathbf{q}}}}\land(\theta_{\bf{q}}-\theta_{\bf{y}})\in\mathcal{Q}_{m})},$ (5) where $\theta_{\bf{q}}$ and $\theta_{\bf{x}}$ are respectively the heading of the query and a frame ${\bf{x}}$, $\tilde{\mathcal{P}}^{|GT|}_{{\bf{\mathbf{q}}}}$ is the set of true positives in the top-$|GT|$ retrievals, $\mathcal{P}^{|GT|}_{{\bf{\mathbf{q}}}}$ is the ground truth positive set, $\varepsilon$ is an arbitrarily small positive quantity to avoid zero division error, and $\mathds{1}(\cdot)$ is the indicator function. See Fig. 3 for an example. Finally, we report the averaged HD for all queries. Figure 3: Heading diversity illustration. The angle represents the heading difference between the query and the evaluated positives. HD represents how many angular bins are covered by true positives vs. that by the ground truth. The figure gives an example of how to calculate HD. Excluding the first and last bins, $\mathcal{P}^{f}_{{\bf{\mathbf{q}}}_{i}}$ contains 5 retrieved non- temporal positives, 4 of which are true positives, and they fall into 3 different bins, while ground truth covers 5 bins, so HD is $3/5$ (not $4/5$). ### IV-B Experiments on simulated point cloud dataset Dataset. We generate the 2D point cloud dataset using the tool provided in [51]. Specifically, we create a large environment as a $1024\times 1024$ binary image in which black and white pixels respectively represent the occupied and free-space locations in the 2D environment. We then manually generate a set of trajectories in this environment, each of which contains a sequence of $2048$ poses. At each pose, a point cloud scan is simulated by finding the intersection points between 2D LiDAR rays and occupied space in the environment. The simulated 2D point cloud dataset contains $3$ different environments and $18$ trajectories in each environment. Each scan contains $256$ points. Baseline methods. We compare TF-VPR with the following baselines: (1) PointNetVLAD [30] trained with pose-based supervision, (2) PointNetVLAD trained with temporal pseudo labels as in [16] (SPTM), (3) SPTM with data augmentation (SPTM+A), (4) SPTM with feature-neighbor KNN expansion (SPTM+F(K)) , and (5) SPTM+A+F(K). Note that the contraction step is not used for this simulated toy dataset, because we found the expansions are all correct in this case. Implementation details. The network architecture for the SPTM baseline is the same as PointNetVLAD. We select $n=5$ and $u=2$ in Section III-B. The value of $n$ depends on the sampling rate of the sensor. Additionally, similar to the training in [30], we randomly select only $2$ positives and $18$ negatives to speed up the loss computation. The output feature dimension is $512$. Based on the sampling rate of the sensor, for each query, we exclude its closest $10$ temporal neighbors from the top-N retrievals as explained in IV-A. Figure 4: Qualitative VPR results on 2D point cloud dataset. The first row shows the results of SPTM [16] and the second row shows the results of TF-VPR. The first column is the query point cloud and column 2-5 are the top 4 retrievals. Green and red respectively indicate true and false positives. Figure 5: Recall@10 and HD with respect to training epoch on 2D point cloud dataset. F denotes method with feature neighborhoods, A denotes method with data augmentation, (K) denotes method with KNN during expansion, (D) denotes method with dynamic KNN in expansion. Figure 6: Per-frame performance visualization over ground truth trajectories on point cloud (top), simulated RGB (middle), and NYU-VPR-360 (bottom) datasets. The left block shows the Recall@10 and the right block shows the HD metrics. Each query’s metrics at epoch 30 are color-coded. Data augmentation. Fig. 5 shows a significant improvement in both recall@10 and HD after augmentation is added to SPTM. SPTM fails to retrieve the true neighbors from different headings, leading to insufficient retrievals. Differently, SPTM+A uses augmented positives in the triplet loss, leading to a network that is insensitive to the orientation of the point cloud inputs. Thus, in Fig. 4, with augmentation, TF-VPR demonstrates its capability to retrieve positives from diverse directions. Feature-neighbor expansion. SPTM+F(K) has slightly poorer performance than the original SPTM as shown in Fig. 5, because only adding positives from the same direction does not provide extra valuable training data. In comparison, SPTM with data augmentation and feature space neighborhood (SPTM+A+F(K)) helps feature-neighbor expansion more effectively discover true neighbors from different headings, thereby improving performance. Dynamic KNN. We further study the effectiveness of dynamic KNN compared to fixed KNN in TF-VPR, i.e., SPTM+A+F(D) vs. SPTM+A+F(K). The difference is negligible on the simulated point cloud dataset. However, the difference would be larger on NYU-VPR-360 dataset. Performance difference between TF-VPR and other baselines. In Table I, among all of these methods, TF-VPR shows a comparable result of recall rate and heading diversity to PointNetVLAD, and outperforms SPTM by a large margin. Moreover, Fig. 6 visualizes the per-frame retrieval quality over the datasets’ ground truth trajectories, and SPTM clearly performs the worst with more failures. TABLE I: Quantitative results on 2D simulated point cloud data. We employ best recall rate (R) and heading diversity (HD). Method / Metric | | R@10 | R@5 | R@1 | HD ---|---|---|---|---|--- SPTM [16] | | 93.36 | 91.06 | 79.10 | 4.25 TF-VPR (Ours) | | 99.71 | 99.66 | 98.73 | 93.62 PointNetVLAD [30] | | 100.00 | 100.00 | 100.00 | 89.38 ### IV-C Experiments on photorealistic RGB dataset TABLE II: Quantitative results on photorealistic RGB data. We employ best recall rate (R) and heading diversity (HD). We report results on three rooms in habitat-sim (Goffs, Micanopy, and Spotswood). Scene | Goffs | Micanopy | Spotswood ---|---|---|--- Metric | R@10 | R@5 | R@1 | HD | R@10 | R@5 | R@1 | HD | R@10 | R@5 | R@1 | HD SPTM [16] | 95.52 | 94.69 | 90.15 | 55.28 | 95.13 | 94.32 | 90.06 | 56.37 | 96.33 | 95.65 | 92.16 | 54.41 SPTM (epoch 30) [16] | 94.50 | 93.28 | 88.03 | 49.61 | 92.89 | 90.20 | 78.27 | 50.47 | 93.53 | 91.23 | 81.52 | 49.69 PCL [57] | 49.27 | 42.86 | 27.00 | 0.35 | 60.36 | 54.18 | 37.73 | 0.19 | 57.36 | 51.66 | 35.33 | 0.03 VLAD [9, 8] | 42.48 | 30.69 | 12.54 | 0.07 | 49.23 | 37.82 | 17.36 | 0.00 | 53.52 | 41.23 | 16.53 | 0.12 TF-VPR (Ours) | 96.31 | 96.00 | 93.62 | 67.67 | 95.59 | 95.21 | 93.42 | 65.97 | 96.91 | 96.57 | 95.35 | 69.39 TF-VPR (Ours) (epoch 30) | 96.15 | 95.82 | 93.31 | 66.54 | 95.46 | 94.93 | 92.82 | 63.94 | 96.66 | 96.32 | 94.66 | 65.59 NetVLAD [12] | 99.63 | 99.16 | 96.36 | 60.48 | 99.48 | 99.26 | 97.38 | 63.38 | 99.56 | 99.17 | 96.60 | 64.64 Figure 7: Recall@10 and HD vs. training epoch on the photorealistic dataset (Goffs). V denotes method with geometric verification. Other abbreviations follow Fig. 5. Figure 8: Qualitative VPR results on the photorealistic RGB dataset. The upper part shows an example where TF-VPR outperforms SPTM [16], and the lower part depicts a challenging example in which none of the methods retrieve the data correctly. Green and red indicates true and false non- temporal spatial positives, respectively. Although the frames in the first row appear to be positives, their positions are still far from the query. Due to the high frame density in the habitat-sim dataset, frames can only be considered positives if they are no more than $20$cm apart from the query. Dataset. TF-VPR has also been tested via habitat-sim [54] simulator on the Gibson photorealistic RGB dataset [58], which provides panoramic RGB images for a variety of indoor scans. We capture RGB images by a panoramic camera mounted on a robot moving randomly in the virtual environment. We captured a total of $33,679$ RGB images in three Gibson rooms. Each image is downsampled to $256\times 64$ pixels. In contrast to other datasets, this simulated RGB dataset contains a large number of revisits of places from both similar and different directions, which is useful for testing recall rate and heading diversity for VPR. Baseline methods. The following baselines are evaluated: (1) NetVLAD [12] trained with pose-based supervision, (2) NetVLAD trained with temporal pseudo labels as in [16] (SPTM), (3) prototypical contrastive learning (PCL) [57] as another self-supervised VPR method used in visual navigation [37], (4) VLAD [9, 8] as a classic non-deep-learning VPR, and the previous ablation study baselines. Implementation details. To speed up computation, the output dimension is set to $512$. Our implementation is based on the code from NetVLAD. The method for converting NetVLAD to SPTM is similar to Section IV-B. Similarly, we also choose $n=5$ and $u=2$ in Section III-B, and exclude for each query’s closest $30$ temporal neighbors from the top-N retrievals as explained in IV-A. Considering new baselines involved, PCL is implemented as the default setting. We need to tune the total number of clusters. In this experiment, we tried to set the total number of clutter to $200$, $500$, and $1000$ and select the best result. For the VLAD baseline, we use the classic 128-dimensional SIFT features, and a cluster size of $32$. The raw VLAD descriptor dimension of $32\times 128$ is further reduced to $512$ by PCA. Importance of contraction. Geometric verification becomes more important for RGB data. The generated positive candidate is not as trustworthy as in the toy dataset in Section IV-B. As a result, including geometric verification helps maintain or even increase accuracy. From Fig. 7, we can observe that both metrics in all methods other than TF-VPR decline with respect to training epochs, because they overfit temporal neighbors as depicted in the first stage in Fig. 1(b). And they tend to miss true spatial neighbors. To prevent this, the feature space neighborhood together with the verification discovers more reliable positives in $\hat{\mathcal{P}}^{f}_{{\bf{\mathbf{q}}}_{i}}$ to be added into $\mathcal{P}_{{\bf{\mathbf{q}}}_{i}}$, as depicted in the third stage of Fig. 1(b). The stability of the model performance over training epochs is critical for self-supervised VPR in reality because we would not know when to stop training without ground truth poses. Ablation study on habitat-sim. As shown in Fig. 7 and Table II, TF-VPR outperforms all unsupervised baselines, and approaches the supervised NetVLAD performance in both recall rate and heading diversity (HD). Particularly, HD is improved by a large margin, particularly for scenes with a high number of revisits from different directions. On average, TF-VPR improves recall rate by $1\%$ and heading diversity by $10\%$. Fig. 6 visualizes the retrieval quality on the simulated RGB dataset. SPTM can make reasonably accurate estimates in recall@10 and HD. TF-VPR outperforms SPTM, but the improvement is not significant. Especially, the improvement of HD in the RGB dataset is not comparable to that in the point cloud dataset, because RGB images appear to be less resistant to environmental changes such as lighting and seasons. The qualitative results are in Fig. 8. PCL, VLAD, and conventional visual SLAM system. Table II shows the poor performance of PCL and VLAD. We believe that PCL might not be a good VPR solution, because most contrastive learning methods do form several disjoint clusters for each category, which is not suitable to represent a continuous feature in vision-based SLAM problems. Similarly, VLAD performs poorly when the input resolution is low and the output dimension is small. Moreover, we test the conventional visual SLAM system, like OpenVSLAM [59]. However, OpenVSLAM easily loses track of odometry. A total of $17.75\%$ of the frames are lost during the tracking process. OpenVSLAM builds a disjoint topology graph while tracking odometry, and the recall rate is $54.98\%$ versus $96.23\%$ for TF-VPR. As a result of the poor performance, we do not use OpenVSLAM as a benchmark. ### IV-D Experiments on NYU-VPR-360 dataset Figure 9: Recall@10 and HD vs. training epoch on NYU-VPR-360 (scene 1). Abbreviations follow Fig. 7. TABLE III: Quantitative results on NYU-VPR-360 dataset. We employ best recall rate (R) and heading diversity (HD). Two scenes have been reported (Scene1, Scene2). Scene | Scene 1 | Scene 2 ---|---|--- Metric | R@10 | R@5 | R@1 | HD | R@10 | R@5 | R@1 | HD SPTM [16] | 69.02 | 60.08 | 48.27 | 79.49 | 65.97 | 62.52 | 53.25 | 50.79 SPTM (epoch 30) [16] | 60.16 | 50.27 | 39.34 | 75.46 | 58.12 | 54.06 | 43.92 | 46.39 PCL [57] | 64.75 | 57.30 | 40.99 | 0.04 | 62.80 | 58.42 | 46.96 | 13.46 VLAD [9, 8] | 37.31 | 27.89 | 10.53 | 0.01 | 21.23 | 16.92 | 8.78 | 0.01 TF-VPR (Ours) | 71.94 | 63.89 | 52.56 | 83.34 | 66.16 | 62.05 | 52.49 | 51.19 TF-VPR (Ours) (epoch 30) | 70.97 | 62.61 | 50.99 | 80.79 | 64.18 | 60.09 | 51.44 | 49.33 NetVLAD [12] | 100.00 | 100.00 | 100.00 | 86.38 | 100.00 | 100.00 | 99.99 | 54.45 NYU-VPR-360 dataset. There are several VPR datasets with panoramic images [60, 50], yet few of them have repeated visits to the same place from a variety of angles. To show the ability of our method in retrieving images of different headings, we proposed the NYU-VPR-360 dataset captured by Gopro MAX (a dual- lens 360∘ camera with GPS recording), which is composed of sequentially collected panoramic RGB images of street views in New York City. The GoPro camera was mounted on the top of the driving vehicle. We utilize the GPS readings of the camera to provide the ground truth of spatial neighborhoods. Note that we select the panoramic images from the whole video to make them synchronized with GPS. The dataset is composed of two driving trajectories, covering an area of approximately $80,000m^{2}$. There are over $15,000$ images of $3840\times 1920$ pixels in the dataset with their corresponding locations for each scene. Most junctions have at least two types of actions with different driving directions, with the exception of a few intersections that are for traffic reasons. Baseline methods. Following IV-C, we use SPTM [16], NetVLAD [12], VLAD [9, 8], and PCL [57] as baselines. Implementation details. The images are resized to $128\times 64$ pixels. We set $n=10$ and $u=5$ as described in Section III-B. For scene 1 and scene 2, based on different sampling rates of the sensors, for each query, we respectively exclude its closest $30$ and $100$ temporal neighbors from the top-N retrievals as explained in IV-A. Comparison with baselines. The edge of our method is more distinguishable on NYU-VPR-360 dataset as proven by the qualitative results in Fig. 10. As shown in Fig. 6, there is an improvement in the recall rate using TF-VPR . Furthermore, as shown in Table III, TF-VPR surpasses other baselines in recall rate and heading diversity in scene 1 by about $2\%$ and $4\%$ respectively. Furthermore, performance in scene 2 does not improve because there are no spatial positives from different headings in scene 2. More importantly, the performance gap between TF-VPR and other baselines becomes larger over epochs. As shown in Fig. 9, the recall@10 of TF-VPR outperforms the baselines by approximately $8\%$-$10\%$ at epoch $30$ as discussed in IV-C. Dynamic KNN mechanism. The distinction between the K-nearest-neighbor specification and the dynamic KNN described in Section III-D is sharper in our real-world experiment. This proves that the dynamic KNN is a better mechanism for selecting feature neighborhoods because its flexibility allows the model to find an adequate number of neighbors for each location in the dataset. Figure 10: Qualitative VPR results on NYU-VPR-360 dataset. The upper part shows an example where TF-VPR outperforms SPTM, and the lower part shows a challenging example in the dataset where only the top-1 retrieval of TF-VPR and VLAD is correct. Green and red respectively indicate true and false positives . ## V Conclusion We propose TF-VPR as a self-supervised auto-labeling method for determining the unknown spatial neighbors from the fixed temporal neighbors and learnable feature neighbors. Extensive experiments show that TF-VPR not only improves the recall rate over the existing method but also can retrieve spatial positives with more diverse viewpoints on various datasets. TF-VPR enables easier use of VPR in real-world robotics and computer vision applications, and can be applied to most existing deep-learning-based VPR methods. Acknowledgment. This work is supported by NSF grants under CMMI-1932187, CNS-2121391, and EEC-2036870. ## References * [1] S. Lowry, _et al._ , “Visual place recognition: A survey,” _T-RO_ , 2015\. * [2] E. Brachmann and C. Rother, “Visual camera re-localization from RGB and RGB-D images using DSAC,” _TPAMI_ , 2021. * [3] A. Angeli, S. Doncieux, J.-A. Meyer, and D. Filliat, “Real-time visual loop-closure detection,” in _2008 ICRA_ , 2008. * [4] F. Radenović, A. Iscen, G. Tolias, Y. Avrithis, and O. Chum, “Revisiting oxford and paris: Large-scale image retrieval benchmarking,” in _CVPR_ , 2018\. * [5] M. Zaffar, _et al._ , “VPR-Bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change,” _IJCV_ , 2021. * [6] D. Sheng, _et al._ , “NYU-VPR: Long-term visual place recognition benchmark with view direction and data anonymization influences,” in _IROS_ , 2021. * [7] D. G. Lowe, “Object recognition from local scale-invariant features,” in _ICCV_ , 1999. * [8] H. Jégou, M. Douze, C. Schmid, and P. Pérez, “Aggregating local descriptors into a compact image representation,” in _CVPR_ , 2010. * [9] R. Arandjelovic and A. Zisserman, “All about vlad,” in _CVPR_ , June 2013\. * [10] P.-E. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk, “From coarse to fine: Robust hierarchical localization at large scale,” in _CVPR_ , 2019. * [11] Z. Chen, _et al._ , “Deep learning features at scale for visual place recognition,” in _ICRA_ , 2017. * [12] R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “Netvlad: Cnn architecture for weakly supervised place recognition,” in _CVPR_ , June 2016\. * [13] S. Garg, M. Vankadari, and M. Milford, “Seqmatchnet: Contrastive learning with sequence matching for place recognition & relocalization,” in _CoRL_. PMLR, 2022. * [14] A. Forechi, A. F. De Souza, C. Badue, and T. Oliveira-Santos, “Sequential appearance-based global localization using an ensemble of knn-dtw classifiers,” in _IJCNN_ , 2016. * [15] O. Vysotska and C. Stachniss, “Lazy data association for image sequences matching under substantial appearance changes,” _RA-L_ , 2015. * [16] N. Savinov, A. Dosovitskiy, and V. Koltun, “Semi-parametric topological memory for navigation,” in _ICLR_ , 2018. * [17] S. E. Reed, H. Lee, D. Anguelov, C. Szegedy, D. Erhan, and A. Rabinovich, “Training deep neural networks on noisy labels with bootstrapping,” in _ICLR (Workshop)_ , 2015. * [18] M. Cummins and P. Newman, “FAB-MAP: Probabilistic localization and mapping in the space of appearance,” _IJRR_ , 2008. * [19] N. Cummins, Mark and Paul, “Appearance-only SLAM at large scale with FAB-MAP 2.0,” _IJRR_ , 2011. * [20] M. J. Milford, G. F. Wyeth, and D. Prasser, “Ratslam: a hippocampal model for simultaneous localization and mapping,” in _ICRA_ , 2004. * [21] D. Galvez-Lopez and J. D. Tardos, “Real-time loop detection with bags of binary words,” in _IROS_ , 2011. * [22] M. J. Milford and G. F. Wyeth, “Seqslam: Visual route-based navigation for sunny summer days and stormy winter nights,” in _ICRA_ , 2012. * [23] G. Costante, T. A. Ciarfuglia, P. Valigi, and E. Ricci, “A transfer learning approach for multi-cue semantic place recognition,” in _IROS_ , 2013. * [24] N. Sünderhauf, _et al._ , “Place recognition with convnet landmarks: Viewpoint-robust, condition-robust, training-free,” in _Robotics: Science and Systems XI_. Robotics: Science and Systems Conference, 2015. * [25] M. Lopez-Antequera, R. Gomez-Ojeda, N. Petkov, and J. Gonzalez-Jimenez, “Appearance-invariant place recognition by discriminatively training a convolutional neural network,” _Pattern Recognition Letters_ , 2017. * [26] T. Naseer, G. L. Oliveira, T. Brox, and W. Burgard, “Semantics-aware visual localization under challenging perceptual conditions,” in _ICRA_ , 2017. * [27] S. Garg, N. Suenderhauf, and M. Milford, “Lost? appearance-invariant place recognition for opposite viewpoints using visual semantics,” _arXiv preprint arXiv:1804.05526_ , June 2018. * [28] C. Choy, J. Park, and V. Koltun, “Fully convolutional geometric features,” in _ICCV_ , 2019. * [29] H. Deng, T. Birdal, and S. Ilic, “Ppfnet: Global context aware local features for robust 3D point matching,” in _CVPR_ , 2018. * [30] M. A. Uy and G. H. Lee, “PointNetVLAD: Deep point cloud based retrieval for large-scale place recognition,” in _CVPR_ , 2018. * [31] Z. Liu, _et al._ , “Lpd-net: 3d point cloud learning for large-scale place recognition and environment analysis,” in _ICCV_ , 2019. * [32] X. Chen, _et al._ , “Overlapnet: Loop closing for lidar-based slam,” in _Robotics: Science and Systems_ , 2020. * [33] K. Cai, B. Wang, and C. X. Lu, “Autoplace: Robust place recognition with low-cost single-chip automotive radar,” in _ICRA_ , 2022. * [34] S. Lowry and M. J. Milford, “Supervised and unsupervised linear learning techniques for visual place recognition in changing environments,” _T-RO_ , 2016. * [35] N. Merrill and G. Huang, “Lightweight unsupervised deep loop closure,” in _Robotics: Science and Systems_ , 2018. * [36] X. Gao and T. Zhang, “Unsupervised learning to detect loops using deep neural networks for visual SLAM system,” _Autonomous robots_ , 2017. * [37] O. Kwon, N. Kim, Y. Choi, H. Yoo, J. Park, and S. Oh, “Visual graph memory with unsupervised representation for visual navigation,” in _ICCV_ , 2021\. * [38] N. Natarajan, I. S. Dhillon, P. K. Ravikumar, and A. Tewari, “Learning with noisy labels,” _NeurIPS_ , 2013. * [39] H. Song, M. Kim, D. Park, Y. Shin, and J.-G. Lee, “Learning from noisy labels with deep neural networks: A survey,” _TNNLS_ , 2022. * [40] Z. Yu, _et al._ , “Simultaneous edge alignment and learning,” in _ECCV_ , 2018. * [41] G. Patrini, A. Rozza, A. Krishna Menon, R. Nock, and L. Qu, “Making deep neural networks robust to label noise: A loss correction approach,” in _CVPR_ , 2017. * [42] D.-H. Lee _et al._ , “Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks,” in _ICML_ , 2013\. * [43] Y. Zou, Z. Yu, B. Kumar, and J. Wang, “Unsupervised domain adaptation for semantic segmentation via class-balanced self-training,” in _ECCV_ , 2018\. * [44] Y. Zou, Z. Yu, X. Liu, B. Kumar, and J. Wang, “Confidence regularized self-training,” in _ICCV_ , 2019. * [45] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in _CVPR_ , 2020. * [46] A. v. d. Oord, Y. Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” _arXiv preprint arXiv:1807.03748_ , 2018\. * [47] Y. Tian, D. Krishnan, and P. Isola, “Contrastive multiview coding,” in _ECCV_. Springer, 2020. * [48] Y. Tian, C. Sun, B. Poole, D. Krishnan, C. Schmid, and P. Isola, “What makes for good views for contrastive learning?” _NeurIPS_ , 2020. * [49] R. Arandjelović and A. Zisserman, “Visual vocabulary with a semantic twist,” in _ACCV_. Springer, 2014\. * [50] A. Torii, R. Arandjelovic, J. Sivic, M. Okutomi, and T. Pajdla, “24/7 place recognition by view synthesis,” in _CVPR_ , 2015. * [51] L. Ding and C. Feng, “Deepmapping: Unsupervised map estimation from multiple point clouds,” in _CVPR_ , 2019. * [52] P. C. Ng and S. Henikoff, “Sift: Predicting amino acid changes that affect protein function,” _Nucleic acids research_ , 2003. * [53] P. Achlioptas, O. Diamanti, I. Mitliagkas, and L. Guibas, “Learning representations and generative models for 3d point clouds,” in _ICML_. PMLR, 2018. * [54] M. Savva, _et al._ , “Habitat: A platform for embodied AI research,” in _ICCV_ , 2019. * [55] A. Paszke, _et al._ , “Automatic differentiation in pytorch,” _arXiv preprint arXiv:2010.07922_ , 2017. * [56] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in _ICLR_ , 2015. * [57] J. Li, P. Zhou, C. Xiong, and S. C. Hoi, “Prototypical contrastive learning of unsupervised representations,” in _ICLR_ , 2021. * [58] F. Xia, A. R. Zamir, Z.-Y. He, A. Sax, J. Malik, and S. Savarese, “Gibson env: real-world perception for embodied agents,” in _CVPR_ , 2018. * [59] S. Sumikura, M. Shibuya, and K. Sakurada, “Openvslam: A versatile visual slam framework,” in _ACM Multimedia_ , 2019. * [60] A. Torii, J. Sivic, T. Pajdla, and M. Okutomi, “Visual place recognition with repetitive structures,” in _CVPR_ , 2013.
# FWD: Real-time Novel View Synthesis with Forward Warping and Depth Ang Cao Chris Rockwell Justin Johnson University of Michigan Ann Arbor <EMAIL_ADDRESS> ###### Abstract Novel view synthesis (NVS) is a challenging task requiring systems to generate photorealistic images of scenes from new viewpoints, where both quality and speed are important for applications. Previous image-based rendering (IBR) methods are fast, but have poor quality when input views are sparse. Recent Neural Radiance Fields (NeRF) and generalizable variants give impressive results but are not real-time. In our paper, we propose a generalizable NVS method with sparse inputs, called FWD , which gives high-quality synthesis in real-time. With explicit depth and differentiable rendering, it achieves competitive results to the SOTA methods with 130-1000$\times$ speedup and better perceptual quality. If available, we can seamlessly integrate sensor depth during either training or inference to improve image quality while retaining real-time speed. With the growing prevalence of depths sensors, we hope that methods making use of depth will become increasingly useful. ## 1 Introduction Given several posed images, _novel view synthesis_ (NVS) aims to generate photorealistic images depicting the scene from unseen viewpoints. This long- standing task has applications in graphics, VR/AR, bringing life to still images. It requires a deep visual understanding of geometry and semantics, making it appealing to test visual understanding. Figure 1: Real-time Novel View Synthesis. We present a real-time and generalizable method to synthesize images from sparse inputs. NeRF variants model the scene via an MLP, which is queried millions of times during rendering and leads to low speeds. Our method utilizes explicit depths and point cloud renderers for fast rendering, inspired by SynSin [82]. The model is trained end-to-end with a novel fusion transformer to give high-quality results, where regressed depths and features are optimized for synthesis. Early work on NVS focused on _image-based rendering_ (IBR), where models generate target views from a set of input images. Light field [39] or proxy geometry (like mesh surfaces) [12, 24, 61, 62] are typically constructed from posed inputs, and target views are synthesized by resampling or blending warped inputs. Requiring dense input images, these methods are limited by 3D reconstruction quality, and can perform poorly with _sparse_ input images. Recently, Neural Radiance Fields (NeRF) [48] have become the leading methods for NVS, using MLPs to represent the 5D radiance field of the scene _implicitly_. The color and density of each sampling point are queried from the network and aggregated by volumetric rendering to get the pixel color. With dense sampling points and _differentiable renderer_ , explicit geometry isn’t needed, and densities optimized for synthesis quality are learned. Despite impressive results, they are not _generalizable_ , requiring MLP fitting for each scene with dense inputs. Also, they are extremely _slow_ because of tremendous MLP query times for a single image. Generalizable NeRF variants like PixelNeRF [89], IBRNet [78] and MVSNeRF [9] emerge very recently, synthesizing novel views of unseen scenes without per- scene optimization by modeling an MLP conditioned on sparse inputs. However, they still query the MLP millions of times, leading to slow speeds. Albeit the progress of accelerating NeRF with per-scene optimization [88, 27, 18], fast and generalizable NeRF variants are still under-explored. In this paper, we target a _generalizable_ NVS method with _sparse_ inputs, refraining dense view collections. Both _real-time speed_ and _high-quality_ synthesis are expected, allowing interactive applications. Classical IBR methods are fast but require dense input views for good results. Generalizable NeRF variants show excellent quality without per-scene optimization but require intense computations, leading to slow speeds. Our method, termed FWD, achieves this target by Forward Warping features based on Depths. Our key insight is that explicitly representing the _depth_ of each input pixel allows us to apply _forward warping_ to each input view using a differentiable point cloud renderer. This avoids the expensive volumetric sampling used in NeRF-like methods, enabling real-time speed while maintaining high image quality. This idea is deeply inspired by the success of SynSin [82], which employs a differentiable point cloud renderer for single image NVS. Our paper extends SynSin to multiple inputs settings and explores effective and efficient methods to fuse multi-view information. Like prior NVS methods, our approach can be trained with RGB data only, but it can be progressively enhanced if noisy sensor depth data is available during training or inference. Depth sensors are becoming more prevalent in consumer devices such as the iPhone 13 Pro and the LG G8 ThinQ, making RGB-D data more accessible than ever. For this reason, we believe that methods making use of RGB-D will become increasingly useful over time. Our method estimates depths for each input view to build a point cloud of latent features, then synthesizes novel views via a point cloud renderer. To alleviate the inconsistencies between observations from various viewpoints, we introduce a view-dependent feature MLP into point clouds to model view- dependent effects. We also propose a novel Transformer-based fusion module to effectively combine features from multiple inputs. A refinement module is employed to inpaint missing regions and further improve synthesis quality. The whole model is trained end-to-end to minimize photometric and perceptual losses, learning depth and features optimized for synthesis quality. Our design possesses several advantages compared with existing methods. First, it gives both high-quality and high-speed synthesis. Using explicit point clouds enables real-time rendering. In the meanwhile, differentiable renderer and end-to-end training empower high-quality synthesis results. Also, compared to NeRF-like methods, which cannot synthesize whole images during training because of intensive computations, our method could easily utilize perceptual loss and refinement module, which noticeably improves the visual quality of synthesis. Moreover, our model can seamlessly integrate sensor depths to further improve synthesis quality. Experimental results support these analyses. We evaluate our method on the ShapeNet and DTU datasets, comparing it with representative NeRF-variants and IBR methods. It outperforms existing methods, considering speed and quality jointly: compared to IBR methods we improve both speed and quality; compared to recent NeRF-based methods we achieve competitive quality at realtime speeds (130-1000$\times$ speedup). A user study demonstrates that our method gives the most perceptually pleasing results among all methods. The code is available at https://github.com/Caoang327/fwd_code. ## 2 Related Work Novel view synthesis is a long-standing problem in computer vision, allowing for the generation of novel views given several scene images. A variety of 3D representations (both implicit and explicit) have been used for NVS, including depth and multi-plane images [74, 94, 72, 57, 7, 67], voxels [69, 21], meshes [61, 23, 28, 62], point clouds [82, 40, 64] and neural scene representations [65, 41, 19, 34, 47, 55, 48]. In this work, we use point clouds as our 3D representations for computational and memory efficiency. Image-based Rendering. IBR synthesizes novel views from a set of reference images by weighted blending [15, 39, 20, 24, 57, 61, 12, 62]. They generally estimate proxy geometry from dense captured images for synthesis. For instance, Riegler et al. [61] uses multi-view stereo [66, 87, 77, 77, 45, 29] to produce scene mesh surface and warps source view images to target views based on proxy geometry. Despite promising results in some cases, they are essentially limited by the quality of 3D reconstructions, where dense inputs (tens to hundreds) with large overlap and reasonable baselines are necessary for decent results. These methods estimate geometry as an intermediate task not directly optimized for image quality. In contrast we input sparse views and learn depth jointly to optimize for synthesis quality. Neural Scene Representations. Recent work uses implicit scene representations for view synthesis [65, 41, 19, 34, 47, 55]. Given many views, neural radiance fields (NeRF) show impressive results [48, 92, 46, 56, 81], but require expensive per-scene optimization. Recent methods [78, 89, 75, 9, 31] generalize NeRF without per-scene optimization by learning a shared prior, with sparse inputs. However these methods require expensive ray sampling and therefore are very slow. In contrast, we achieve significant speedups using explicit representations. Some concurrent work accelerates NeRF by reformulating the computation [18], using precomputation [88, 27], or adding view dependence to explicit 3D representations [41, 83, 2, 8, 49]; unlike ours, these all require dense input views and per-scene optimization. Utilizing RGB-D in NVS. The growing availability of annotated depth maps [13, 5, 10, 1, 71, 68] facilitates depth utilization in NVS [54, 40, 26], which serves as extra supervision or input to networks. Our method utilizes explicit depths as 3D representations, allowing using sensor depths as additional inputs for better quality. Given the increasing popularity of depth sensors, integrating sensor depths is a promising direction for real-world applications. Figure 2: System Overview. Given a sparse set of images, we construct a point cloud $\mathcal{P}_{i}$ for each image $I_{i}$ using Feature Network $f$, View-Dependent Feature MLP $\psi$, and Depth Network $d$. Besides images, $d$ takes MVS estimated depths or sensor depths as inputs and regresses refined depths. Per-pixel features $F^{\prime}_{i}$ are regressed by $f$ and $\psi$ based on images and relative view changes. A differentiable point cloud renderer $\pi$ is employed to project and render point clouds to target views. We use Transformer $T$ to fuse rendered results from arbitrary number inputs and apply refinement module $R$ for final results. The model is trained with photometric loss and content loss. Depth has been used in neural scene representations for speedups [51, 73], spaser inputs [16] and dynamic scenes[84]. However, these works still require per-scene optimization. Utilizing RGB-D inputs to accelerate generalizable NeRF like [89, 78] is still an open problem. Differentiable Rendering and Refinement. We use advances in differentiable rendering [42, 35, 11, 52, 43] to learn 3D end-to-end. Learned geometries rely heavily on rendering and refinement [90, 86, 3, 79] to quickly synthesize realistic results. Refinement has improved dramatically owing to generative modeling [38, 36, 91, 95] and rendering frameworks [60, 32, 50, 30]. Instead of aggregating information across viewpoints before rendering [44], we render viewpoints separately and fuse using a Transformer [76, 17, 4], enabling attention across input views. ## 3 Method Given a sparse set of input images $\\{I_{i}\\}_{i=1}^{N}$ and corresponding camera poses $\\{R_{i},T_{i}\\}$, our goal is to synthesize a novel view with camera pose $\\{R_{t},T_{t}\\}$ fast and effectively. The depths $\\{D^{sen}_{i}\\}$ of $I_{i}$ captured from sensors are optionally available, which are generally incomplete and noisy. The insight of our method is that using explicit depths and forward warping enables real-time rendering speed and tremendous accelerations. Meanwhile, to alleviate quality degradations caused by inaccurate depth estimations, a differentiable renderer and well-designed fusion & refinement modules are employed, encouraging the model to learn geometry and features optimized for synthesis quality. As illustrated in Figure 2, with estimated depths, input view $I_{i}$ is converted to a 3D point cloud $\mathcal{P}_{i}$ containing geometries and view-dependent semantics of the view. A differentiable neural point cloud renderer $\pi$ is used to project point clouds to target viewpoints. Rather than directly aggregating point clouds across views before rendering, we propose a Transformer-based module $T$ fusing rendered results at target view. Finally, a refinement module $R$ is employed to generate final outputs. The whole model is trained end-to-end with photometric and perceptual loss. ### 3.1 Point Cloud Construction We use point clouds to represent scenes due to their efficiency, compact memory usage, and scalability to complex scenes. For input view $I_{i}$, point cloud $\mathcal{P}_{i}$ is constructed by estimating depth $D_{i}$ and feature vectors $F^{\prime}_{i}$ for each pixel in the input image, then projecting the feature vectors into 3D space using known camera intrinsics. The depth $D_{i}$ is estimated by a _depth network_ $d$; features $F^{\prime}_{i}$ are computed by a _spatial feature encoder_ $f$ and _view-dependent MLP_ $\psi$. Spatial Feature Encoder $f$. Scene semantics of input view $I_{i}$ are mapped to per-pixel feature vectors $F_{i}$ by spatial feature encoder $f$. Each feature vector in $F_{i}$ is 61-dimensions and is concatenated with RGB channels, which is 64 dimensions in total. $f$ is built on BigGAN architecture [3]. Depth Network $d$. Estimating depth from a single image has scaling/shifting ambiguity, losing valuable multi-view cues and leading to inconsistent estimations across views. Applying multi-view stereo algorithms (MVS) [66, 87, 77, 85] solely on sparse inputs is challenging because of limited overlap and huge baselines between input views, leading to inaccurate and low-confidence estimations. Therefore, we employ a hybrid design cascading a U-Net after the MVS module. The U-Net takes image $I_{i}$ and estimated depths from the MVS module as inputs, refining depths with multi-view stereo cues and image cues. PatchmatchNet [77] is utilized as the MVS module, which is fast and lightweight. Depth Estimation with sensor depths. As stated, U-Net receives an initial depth estimation from the MVS module and outputs a refined depth used to build the point cloud. If sensor depth $D^{sen}_{i}$ is available, it is directly input to the U-Net as the initial depth estimations. In this setting, U-Net servers as completion and refinement module taking $D^{sen}_{i}$ and $I_{i}$ as inputs, since $D^{sen}_{i}$ is usually noisy and incomplete. During training, loss $L_{s}$ is employed to encourage the U-Net output to match the sensor depth. $\displaystyle\mathcal{L}_{s}=\|M_{i}\odot D_{i}-M_{i}\odot D^{sen}_{i}\|$ (1) where $M_{i}$ is a binary mask indicating valid sensor depths. View-Dependent Feature MLP $\psi$. The appearance of the point could vary across views because of lighting and view direction, causing inconsistency between multiple views. Therefore, we propose to insert view direction changes into scene semantics to model this view-dependent effects. An MLP $\psi$ is designed to compute view-dependent features $F^{\prime}_{i}$ by taking $F_{i}$ and relative view changes $\Delta v$ from input to target view as inputs. For each point in the cloud, $\Delta v$ is calculated based on normalized view directions $v_{i}$ and $v_{t}$, from the point to camera centers of input view $i$ and target view $t$. The relative view direction change is calculated as: $\displaystyle\Delta v=[(v_{i}-v_{t})/\|v_{i}-v_{t}\|,v_{i}\cdot v_{t}],v_{i},v_{t}\in\mathbb{R}^{3}.$ (2) and the view-dependent feature $F^{\prime}_{i}$ is: $\displaystyle F^{\prime}_{i}=\psi(F_{i},\delta(\Delta v))$ (3) where $\delta$ is a two-layer MLP mapping $\Delta v$ to a 32-dimensions vector and $\psi$ is also a two-layer MLP. ### 3.2 Point Cloud Renderer To observe the constructed point cloud $\mathcal{P}_{i}$ at target views, we employ a neural point cloud renderer $\pi$. $\mathcal{P}_{i}$ is first transformed to target view coordinates based on camera poses and then rendered by $\pi$. The rendered feature maps $\tilde{F}_{i}$ share the same dimension as feature $F^{\prime}_{i}$ at each pixel. With explicit geometry transformation, our rendered results are geometrically consistent and correct across views. We use the _differentiable_ renderer design of [82], which splats 3D points to the image plane and gets pixel values by blending point features. The blending weights are computed based on _z_ -buffer depths and distances between pixel and point centers. It is implemented using Pytorch3D [60]. This fully differentiable renderer allows our model to be trained end-to-end, where photometric and perceptual loss gradients can be propagated to points’ position and features. In this way, the model learns to estimate depths and features optimized for synthesis quality, leading to superior quality. We show the effectiveness of it in experiments. ### 3.3 Fusion and Refinement Unlike SynSin [82] using a single image for NVS, fusing multi-view inputs is required in our method. A naïve fusion transforms each point cloud to target view and aggregates them into a large one for rendering. Despite high efficiency, it is vulnerable to inaccurate depths since points with wrong depths may occlude points from other views, leading to degraded results. Methods like PointNet [58] may be feasible to apply on the aggregated point cloud for refinement, but they are not efficient with significant point numbers. Instead, we render each point cloud individually at target viewpoints and fuse rendered results by a fusion Transformer $T$. A refinement module $R$ is used to inpaint missing regions, decode feature maps and improve synthesis quality. Figure 3: Fusion Transformer. We use a lightweight transformer $T$ to fuse the features from $N$ input views on each pixel. We use a learnable token to query the fusion results. Fusion Transformer $T$. Given an arbitrary number of rendered feature maps $\\{\tilde{F}_{i}\\}$, fusion should be effective, fast, and permutation invariant. Inspired by the success of Transformers, we propose a pixel-wise Transformer $T$ for fusion, detailed in Figure 3. At each pixel, $T$ inputs rendered feature vectors and queries fused results using a learnable “token”. Applied on features, $T$ utilizes semantics for fusion. Rendered results may lose geometry cues for fusion when rendered from 3D to 2D. For instance, depths may reveal occlusion relationships across views, and relative view changes from input to target views relate to each input’s importance for fusion. Therefore, we also explored to use geometry features as position encoding while not helpful. Refinement Module $R$. Built with 8 ResNet [22] blocks, $R$ decodes fused feature maps $\tilde{F}$ to RGB images $\tilde{I}$ at target views. It inpaints regions invisible for inputs in a semantically and geometrically meaningful manner. Also, it corrects local errors caused by inaccurate depths and improves perceptual quality based on semantics contained by feature maps, leading to coherent and high-quality synthesis. ### 3.4 Training and Implementation Details Our model is trained end-to-end with photometric $\mathcal{L}_{l_{2}}$ and perceptual $\mathcal{L}_{c}$ losses between generated and ground-truth target images. The whole loss function is: $\displaystyle\mathcal{L}=\lambda_{l_{2}}\mathcal{L}_{l_{2}}+\lambda_{c}\mathcal{L}_{c}$ (4) where $\lambda_{l_{2}}=5.0,\lambda_{c}=1.0$. The model is trained end-to-end on 4 2080Ti GPUs for 3 days, using Adam [37] with learning rate $10^{-4}$ and $\beta_{1}{=}0.9,\beta_{2}{=}0.999$. When sensors depths are available as inputs, $\mathcal{L}_{s}$ is used with $\lambda_{s}=5.0$. Table 1: Model variants settings. We predefine three model variants with different settings. FWD utilizes a pre-trained MVS module, in which way it gets access to depths during training. Name | Test Depth | Train Depth | Depth Network | MVS Module | Losses ---|---|---|---|---|--- FWD-U | | | MVS + U-Net | Random ini. | $\mathcal{L}_{l_{2}}+\mathcal{L}_{c}$ FWD | | ✓ | MVS + U-Net | Pre-trained | $\mathcal{L}_{l_{2}}+\mathcal{L}_{c}$ FWD-D | ✓ | ✓ | RGB-D + U-Net | - | $\mathcal{L}_{l_{2}}+\mathcal{L}_{c}+\mathcal{L}_{s}$ ## 4 Experiments The goal of our paper is _real-time_ and _generalizable_ novel view synthesis with _sparse_ inputs, which can optionally use sensor depths. To this end, our experiments aim to identify the speed and quality at which our method can synthesize novel images and explore the advantage of explicit depths. We evaluate our methods on ShapeNet [6] and DTU [33] datasets, comparing results with the SOTA methods and alternative approaches. Experiments take place with held-out test scenes and no per-scene optimization. We conduct ablations to validate the effectiveness of designs. Metrics. We conduct A/B test to measure the visual quality, in which workers select the image most similar to the ground truth from competing methods. Automatic image quality metrics including PSNR, SSIM [80] and LPIPS [93] are also reported, and we find LPIPS best reflects the image quality as perceived by humans. Frames per second (FPS) during rendering is measured on the same platform (single 2080Ti GPU with 4 CPU cores). All evaluations are conducted using the same protocol (same inputs and outputs). Model Variants. Three models are evaluated with various accessibility to depths for training and test, as defined in Table 1. FWD utilizes a pretrained PatchmatchNet [77] as the MVS module for depth estimations, which is also updated during end-to-end training with photometric and perceptual loss. FWD-U learns depth estimations in an Unsupervised manner, sharing the same model and settings as FWD while PatchmatchNet is randomly initialized without any pretraining. FWD-D takes sensor depths as additional inputs during both training and inference. It doesn’t use any MVS module since sensor depths provide abundant geometry cues. For pretraining PatchmatchNet, we train it following typical MVS settings and using the same data splitting as NVS. Input | PixelNeRF | FWD-U | GT ---|---|---|--- | | | | | | | | | | | | | | | | | | | | | | | | Figure 4: Qualitative results of category-agnostic NVS on ShapeNet. We test the capacity of our model by training it across 13 categories of ShapeNet with single view input, and compare with PixelNeRF [89]. No gt depths are available during inference and training. Our results have better visual quality and details. Table 2: Category-agnostic NVS on ShapeNet. Quantitative results for category-agnostic view-synthesis are presented. | 1-view | 2-view ---|---|--- model | PSNR | SSIM | LPIPS | FPS | PSNR | SSIM | LPIPS | FPS DVR [53] | 22.70 | 0.860 | 0.130 | 1.5 | - | - | - | - SRN [70] | 23.28 | 0.849 | 0.139 | 24 | - | - | - | - PixelNeRF | 26.80 | 0.910 | 0.108 | 1.2 | 28.88 | 0.936 | 0.076 | 1.1 FWD-U | 26.66 | 0.911 | 0.055 | 364 | 28.43 | 0.931 | 0.043 | 336 Input: 3 views of held-out scene | Novel views ---|--- FWD-D | | | | | | | | FWD | | | | | | | | FWD-U | | | | | | | | Figure 5: View synthesis results from FWD. We show the view synthesis results with 3 input views on DTU dataset from FWD-D (row. 1), FWD (row. 2) and FWD-U (row. 3). Our methods synthesize high-quality and geometrically correct novel views in real time. ### 4.1 ShapeNet Benchmarks We first evaluate our model for category-agnostic synthesis task on ShapeNet [6]. Following the setting of [89], we train and evaluate a single model on 13 ShapeNet [6] categories. Each instance contains 24 fixed views of 64 $\times$ 64 resolution. During training, one random view is selected as input and the rests are served as target views. For testing, we synthesize all the other views from a fixed informative view. The model is finetuned with two random input views for 2-view experiments. We find that U-Net is sufficient for good results on this dataset without the MVS module. Qualitative comparisons to PixelNeRF are shown in Figure 4, where FWD-U gets noticeably superior results. Our synthesized results are more realistic and closely matching to target views, while PixelNeRF’s results tend to be blurry. We observe the same trend in the DTU benchmark and evaluate the visual quality quantitatively there. We show quantitative results in Table 2, adding SRN [70] and DVR [53] as other baselines. Our method outperforms others significantly for LPIPS, indicating a much better perceptual quality, as corroborated by qualitative results. PixelNeRF has a slightly better PSNR while its results are blurry. Most importantly, FWD-U runs at a speed of over 300 FPS, which is 300$\times$ faster than PixelNeRF. ### 4.2 DTU MVS Benchmarks We also evaluate our model on DTU MVS dataset [33], which is a real scene dataset consisting of 103 scenes. Each scene contains one or multiple objects placed on a table, while images and incomplete depths are collected by the camera and structured light scanner mounted on an industrial robot arm. Corresponding camera poses are provided. As stated in [89], this dataset is challenging since it consists of complex real scenes without apparent semantic similarities across scenes. Also, images are taken under varying lighting conditions with distinct color inconsistencies between views. Moreover, with only under 100 scenes available for training, it is prone to overfitting in training. We follow the same training and evaluation pipelines as PixelNeRF [89] for all methods to give a fair comparison. The data consists of 88 training and 15 test scenes, between which there are no shared or highly similar scenes. Images are down-sampled to a resolution of 300 $\times$ 400\. For training, three input views are randomly sampled, with the rest as target views. For inference, we choose three fixed informative input views and synthesize other views of the scene. Input Image | PixelNeRF | IBRNet | FWD-U | FWD | Blending+R | FWD-D | Target View ---|---|---|---|---|---|---|--- | | | | | | | | | | | | | | | | | | | | | Figure 6: Qualitative Comparison. We compare synthesis results from different methods with 3 input views (one of them shown in figure). Our methods give geometrically consistent and visually appealing results, while other results suffering shaking artifacts at some views. Unlike other methods, FWD-D and Blending+R get access sensor depths as inputs during inference. Figure 7: User study on DTU. We conduct a user study by asking subjects to select the results most similar to the ground truth. The numbers indicate the percentage of preference. Methods are grouped based whether using depths during test. We also report FWD vs. FWD-D showing the advantages of sensor depths. Baselines. We evaluate a set of representatives of generalizable NeRF and IBR methods in two different scenarios: with RGB or RGB-D available as inputs during inference. PixelNeRF [89], IBRNet [78] and MVSNeRF [9] are the SOTA generalizable NeRF variants, taking RGB as inputs. We use the official PixelNeRF model trained on DTU MVS and carefully retrain IBRNet and MVSNeRF with the same 3-input-view settings. PixelNeRF-DS is also included as reported in [16], which is PixelNeRF supervised with depths. Please note that our settings are very different from evaluations used in original papers of IBRNet and MVSNeRF. A series of IBR methods are also evaluated. Since COLMAP [66] fails to give reasonable outputs with sparse input images, methods using COLMAP like FVS [61], DeepBlending [25] cannot estimate scene geometry in this setting. For these methods, we use depths captured by sensors as estimated depths, which should give upper-bound performance of these methods. To better cope with missing regions, we add our refinement model to DeepBlending [25] and retrain it on DTU dataset, termed Blending-R. For fairness, we evaluate all methods using the same protocol, distinct from some of their original settings. Although we try our best to adopt these methods, our reported results may still not perfectly reflect their true capacity. Qualitative Results. Synthesis results are shown in Figure 5, where high- quality and geometrically correct novel views are synthesized in real-time (over 35 FPS) under significant viewpoint changes. Our refinement module faithfully inpaints invisible regions; also, synthesized images have good shadows, light reflections, and varying appearances across views, showing the efficacy of view-dependent MLP. With sensor depths, results can be further improved. We show comparisons to baselines in Figure 6. Our methods provide noticeably better results than baselines across different depth settings. For models without depths in test, IBRNet and PixelNeRF give blurry results in areas of high detail such as the buildings in the top row, while our FWD-U and FWD give more realistic and sharper images. With sensor depths in test, baseline Blending-R produces more cogent outputs, but still struggles to distinguish objects from the background, such as in the middle row, while FWD-D gives faithfully synthesis and clear boundaries. Quantitative Results. We evaluate synthesis quality quantitatively by user study following a standard A/B paradigm. Workers choose the closest to a ground truth image between competing methods, and are monitored using a qualifier and sentinel examples. All views in the test set (690 in total) are evaluated, and each view is judged by three workers. In Figure 7, user study results support qualitative observations. Among all baselines with and without test depths , users choose our method as more closely matching ground truth images than others most of the time. FWD-U is selected over PixelNeRF in 65.6% of examples, and 77.8% compared to IBRNet. Also, over 90% workers prefer FWD-D to FWD , showing advantage of using sensor depths. We show automated view synthesis metrics and speeds in Table 3. Across all depth availability settings, our method is competitive with the SOTA baselines while significantly faster. FWD-D runs in real-time and gives substantially better image quality than others. FWD has competitive metrics to PixelNeRF-DS while 1000$\times$ faster. Notably, NeRF variants such as PixelNeRF, IBRNet, MVSNeRF, and PixelNeRF-DS are at least two orders of magnitude slower. The exception to highly competitive performance is weaker PSNR and SSIM of our unsupervised FWD-U against PixelNeRF and IBRNet. However, FWD-U has noticeably better perceptual quality with the best LPIPS, and human raters prefer it to other methods in A/B tests. The visual quality in figure 6 also illustrates the disparity between comparisons using PSNR and LPIPS. Meanwhile, FWD-U is above $1000\times$ faster than PixelNeRF and above $100\times$ faster than IBRNet. Depth estimations, rendering and CNN would introduce tiny pixel shiftings, which harm the PSNR of our method. NeRF-like methods are trained to optimize L2 loss for each pixel independently, leading to blur results. Among all methods without test depths, FWD has the best results. Although it uses a pretrained MVS module, we think this comparison is still reasonable since pretrained depth module is easy to get. Also, training depths can be easily calculated from training images since they are dense. Baseline comparisons also show that IBR methods are fast, but do not give images that are competitive with our method. Our method outperforms them in both perceptual quality and standard metrics, showing the efficacy of proposed methods. Note that Blending+R doesn’t support variable number inputs and our refinement module improves its results significantly. We also compare FWD-U with SynSin [82] which only receives a single input image, showing the benefits of using multi-view inputs in NVS. ### 4.3 Ablations and Analysis We evaluate the effectiveness of our designs and study depth in more detail through ablation experiments. Effects of Fusion Transformer. We design a model without Transformer, which concatenates point clouds across views into a bigger one for later rendering and refinement. Its results in FWD-U settings are shown in Figure 8. The ablated version is vulnerable to inaccurate depths learned in unsupervised manner and synthesizes “ghost objects” since points with bad depths occlude other views’ points. We repeat the same ablation in FWD-D settings, shown in Table 4, which settings give much better depth estimations with sensor depths. The ablated model has notably worse results for all metrics, indicating that the proposed method is not only powerful to tackle inaccurate depth estimations, but also fuse semantic features effectively. Table 3: Quantitative comparison on DTU real images. We compare our method with representatives of generalizable NeRF variants and IBR methods for image quality and rendering speed. Our method achieves significantly better speed- quality tradeoff, indicating the effectiveness and efficiency of our design. $\dagger$ Unlike other methods, SynSin receives only one image as input. Test | Train | Model | PSNR$\uparrow$ | SSIM$\uparrow$ | LPIPS$\downarrow$ | FPS$\uparrow$ ---|---|---|---|---|---|--- RGB | RGB | PixelNeRF [89] | 19.24 | 0.687 | 0.399 | 0.03 IBRNet [78] | 18.86 | 0.695 | 0.387 | 0.27 MVSNeRF [9] | 17.13 | 0.611 | 0.444 | 0.34 SynSin [82] $\dagger$ | 15.66 | 0.564 | 0.388 | 51.8 FWD-U | 17.42 | 0.598 | 0.341 | 35.4 RGB | RGB-D | PixelNeRF-DS [16] | 19.87 | 0.710 | 0.370 | 0.03 FWD | 20.15 | 0.721 | 0.259 | 35.4 RGB-D | RGB-D | Blending-R [25] | 16.98 | 0.661 | 0.351 | 41.8 FVS [61] | 15.92 | 0.733 | 0.267 | 9.70 FWD-D | 21.98 | 0.791 | 0.208 | 43.2 Effects of View Dependent MLP. For ablation, we remove the view-dependent feature MLP and report its results in Table 4. Removing this module reduces model’s ability to produce view-dependent appearances, leading to worse performance for all metrics. We show more results in Supp. Depth Analysis and Ablations. We visualize depths in Figure 9. Estimating depths from sparse inputs is challenging and gives less accurate results because of the huge baselines between inputs. We show estimated depths from PatchmatchNet here, filtered based on the confidence scores. Therefore, refinement is essential in our design to propagate multi-view geometry cues to the whole image. Our end-to-end model learns it by synthesis losses. We ablate the depth network in Table 5 and report depth error $\delta_{3cm}$, which is the percentage of estimated depths within 3 cm of sensor depths. MVS module is critical (row 2) to give geometrically consistent depths. U-Net further refines depths and improves the synthesis quality (row 3). PatchmatchNet has its own shallow refinement layer, already giving decent refinements. Learning unsupervised MVS and NVS jointly from scratch is challenging (row 4), and training depth network without supervision [14] first may give a good initialization for further jointly training. | | | ---|---|---|--- | | | Input View | FWD-U | w/o Transformer | Target View Figure 8: Ablation on Fusion Transformer. We show results for FWD-U with and without Transformed-based fusion. Table 4: Ablation Studies. We show the effectiveness of Transformer Fusion and View-dependent MLP by ablation study on FWD-D. These designs improve synthesize quality noticeably while maintaining real-time rendering speed. Model | PSNR | SSIM | LPIPS | FPS ---|---|---|---|--- Full model | 21.98 | 0.791 | 0.208 | 43.2 w/o Transformer | 20.95 | 0.748 | 0.241 | 48.4 w/o View dependence | 21.16 | 0.769 | 0.212 | 44.0 ## 5 Conclusion We propose a real-time and generalizable method for NVS with sparse inputs by using explicit depths. This method inherits the core idea of SynSin while extending it to multi-view input settings, which is more challenging. Our experiments show that estimating depths can give impressive results with a real-time speed, outperforming existing methods. Moreover, the proposed method could utilize sensor depths seamlessly and improve synthesis quality significantly. With the increasing availability of mobile depth sensors, we believe our method has exciting real-world 3D applications. We acknowledge there could be the potential for the technology to be used for negative purposes by nefarious actors, like synthesizing fake images for cheating. There are also challenges and limitations yet to be explored. 1) Although using explicit depths gives tremendous speedups, it potentially inflicts depth reliance on our model. We designed a hybrid depth regressor to improve the quality of depth by combining MVS and single image depth estimations. We also employed an effective fusion and refinement module to reduce the degrades caused by inaccurate depths. Despite these designs, the depth estimator may still work poorly in some challenging settings (like very wide camera baselines), and it would influence the synthesis results. Exploring other depth estimation methods like MiDaS [59] could be an interesting direction for future work. 2) The potential capacity of our method is not fully explored. Like SynSin [82], our model (depth/feature network and refinement module especially) is suitable and beneficial from large-scale training data, while the DTU MVS dataset is not big enough and easy to overfit during training. Evaluating our method on large-scale datasets like Hypersim [63] would potentially reveal more advantages of our model, which dataset is very challenging for NeRF-like methods. 3) Although our method gives more visually appealing results, our PSNR and SSIM are lower than NeRF-like methods. We hypothesize that our refinement module is not perfectly trained to decode RGB colors from feature vectors because of limited training data. Also, tiny misalignments caused during the rendering process may also harm the PSNR, although it is not perceptually visible. | | | | | ---|---|---|---|---|--- | | | | | input image | sensor depths | FWD-D | filtered MVS | FWD | FWD-U Figure 9: Depth visualizations. We visualize the normalized inverse depths involved in our method. Sensor depths are incomplete because of hardware limitations and MVS estimated depths are inaccurate, where many predictions have low confidence. This demonstrates the necessity of depth completion and refinement. Table 5: Depths network ablation and error. We ablate depth network and compute $\delta_{3cm}$ as error, which is the percentage of predicted depths within 3 cm of sensor depths. Test | Train | Model | PSNR | SSIM | LPIPS | $\delta_{3cm}$ ---|---|---|---|---|---|--- RGB | RGB-D | FWD | 20.15 | 0.721 | 0.259 | 79.07 RGB | RGB-D | -w/o MVS | 16.69 | 0.594 | 0.357 | 61.62 RGB | RGB-D | -w/o U-Net | 19.10 | 0.702 | 0.285 | 73.62 RGB | RGB | FWD-U | 17.42 | 0.598 | 0.341 | 54.27 Acknowledgement. Toyota Research Institute provided funds to support this work. We thank Dandan Shan, Hao Ouyang, Jiaxin Xie, Linyi Jin, Shengyi Qian for helpful discussions. ## References * [1] Henrik Aanæs, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 120(2):153–168, 2016. * [2] Alex Yu and Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks, 2021. * [3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. * [4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213–229. Springer, 2020. * [5] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. * [6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. Technical Report arXiv:1512.03012, 2015. * [7] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3):1–12, 2013. * [8] Anpei Chen, Zexiang Xu, Andreas Geiger, , Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields, 2022. * [9] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124–14133, 2021. * [10] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. In NeurIPS, volume 29, 2016. * [11] Wenzheng Chen, Jun Gao, Huan Ling, Edward J Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. 2019\. * [12] Inchang Choi, Orazio Gallo, Alejandro Troccoli, Min H Kim, and Jan Kautz. Extreme view synthesis. In ICCV, pages 7781–7790, 2019. * [13] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017\. * [14] Yuchao Dai, Zhidong Zhu, Zhibo Rao, and Bo Li. Mvs2: Deep unsupervised multi-view stereo with multi-view symmetry. In 2019 International Conference on 3D Vision (3DV), pages 1–8, 2019. * [15] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 11–20, 1996. * [16] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. * [17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. * [18] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346–14355, 2021. * [19] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857–4866, 2020. * [20] Steven J Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43–54, 1996. * [21] Pengsheng Guo, Miguel Angel Bautista, Alex Colburn, Liang Yang, Daniel Ulbricht, Joshua M Susskind, and Qi Shan. Fast and Explicit Neural View Synthesis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3791–3800, 2022. * [22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. * [23] Peter Hedman and Johannes Kopf. Instant 3d photography. TOG, 37(4):1–12, 2018. * [24] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (TOG), 37(6):1–15, 2018. * [25] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. 37(6):257:1–257:15, 2018. * [26] Peter Hedman, Tobias Ritschel, George Drettakis, and Gabriel Brostow. Scalable inside-out image-based rendering. ToG, 35(6):1–11, 2016. * [27] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875–5884, 2021. * [28] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12528–12537, 2021. * [29] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. Conference on Computer Vision and Pattern Recognition (CVPR), 2018\. * [30] Rui Huang, Wanyue Zhang, Abhijit Kundu, Caroline Pantofaru, David A Ross, Thomas Funkhouser, and Alireza Fathi. An lstm approach to temporal 3d object detection in lidar point clouds. arXiv preprint arXiv:2007.12392, 2020. * [31] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885–5894, 2021. * [32] Krishna Murthy Jatavallabhula, Edward Smith, Jean-Francois Lafleche, Clement Fuji Tsang, Artem Rozantsev, Wenzheng Chen, Tommy Xiang, Rev Lebaredian, and Sanja Fidler. Kaolin: A pytorch library for accelerating 3d deep learning research. arXiv preprint arXiv:1911.05063, 2019. * [33] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406–413. IEEE, 2014. * [34] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6001–6010, 2020. * [35] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1251–1261, 2020. * [36] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8110–8119, 2020. * [37] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2015\. * [38] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, and W. Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, pages 105–114, 2017. * [39] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31–42, 1996. * [40] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458–14467, 2021. * [41] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651–15663, 2020. * [42] Shichen Liu, Weikai Chen, Tianye Li, and Hao Li. Soft rasterizer: Differentiable rendering for unsupervised single-view mesh reconstruction. ICCV, 2019. * [43] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. Dist: Rendering deep implicit signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2019–2028, 2020. * [44] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. * [45] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (TOG), 39(4):71–1, 2020. * [46] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. * [47] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460–4470, 2019. * [48] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. * [49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv:2201.05989, Jan. 2022. * [50] Mahyar Najibi, Guangda Lai, Abhijit Kundu, Zhichao Lu, Vivek Rathod, Thomas Funkhouser, Caroline Pantofaru, David Ross, Larry S Davis, and Alireza Fathi. Dops: learning to detect 3d objects and predict their 3d shapes. In CVPR, pages 11913–11922, 2020. * [51] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H Mueller, Chakravarty R Alla Chaitanya, Anton Kaplanyan, and Markus Steinberger. DONeRF: Towards real-time rendering of compact neural radiance fields using depth oracle networks. In Computer Graphics Forum, volume 40, pages 45–59. Wiley Online Library, 2021. * [52] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504–3515, 2020. * [53] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. * [54] David Novotny, Ben Graham, and Jeremy Reizenstein. Perspectivenet: A scene-consistent image generator for new view synthesis in real indoor environments. Advances in Neural Information Processing Systems, 32:7601–7612, 2019. * [55] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165–174, 2019. * [56] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. * [57] Eric Penner and Li Zhang. Soft 3d reconstruction for view synthesis. ToG, 36(6):1–11, 2017. * [58] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017. * [59] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. * [60] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. * [61] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, pages 623–640. Springer, 2020. * [62] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. * [63] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10912–10922, 2021. * [64] Chris Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In ICCV, 2021. * [65] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356–14366, 2021. * [66] Johannes Lutz Schönberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. * [67] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028–8038, 2020. * [68] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, pages 746–760, 2012. * [69] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, pages 2437–2446, 2019. * [70] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. * [71] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, pages 567–576, 2015. * [72] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snavely. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, pages 175–184, 2019. * [73] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. * [74] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Multi-view 3d models from single images with a convolutional network. In European Conference on Computer Vision, pages 322–337. Springer, 2016. * [75] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182–15192, 2021. * [76] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017\. * [77] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14194–14203, 2021. * [78] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2021. * [79] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In CVPR, pages 8798–8807, 2018. * [80] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. * [81] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf–: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. * [82] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467–7477, 2020. * [83] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. Nex: Real-time view synthesis with neural basis expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8534–8543, 2021. * [84] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421–9431, 2021. * [85] Jiaxin Xie, Chenyang Lei, Zhuwen Li, Li Erran Li, and Qifeng Chen. Video depth estimation by fusing flow-to-depth proposals. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10100–10107, 2020. * [86] Chao Yang, Xin Lu, Zhe Lin, Eli Shechtman, Oliver Wang, and Hao Li. High-resolution image inpainting using multi-scale neural patch synthesis. In CVPR, pages 6721–6729, 2017. * [87] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European Conference on Computer Vision (ECCV), pages 767–783, 2018. * [88] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In arXiv, 2021. * [89] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578–4587, 2021. * [90] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, pages 5505–5514, 2018. * [91] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, pages 7354–7363. PMLR, 2019. * [92] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. * [93] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586–595, 2018. * [94] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. * [95] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017.
11institutetext: Department of Mathematics, University of Milan, via Saldini 50, 10133 Milan Italy <EMAIL_ADDRESS> <EMAIL_ADDRESS> <EMAIL_ADDRESS> # A set–valued framework for birth–and–growth process Giacomo Aletti Enea G. Bongiorno Vincenzo Capasso ###### Abstract We propose a set–valued framework for the well–posedness of birth–and–growth process. Our birth–and–growth model is rigorously defined as a suitable combination, involving Minkowski sum and Aumann integral, of two very general set–valued processes representing nucleation and growth respectively. The simplicity of the used geometrical approach leads us to avoid problems arising by an analytical definition of the front growth such as boundary regularities. In this framework, growth is generally anisotropic and, according to a mesoscale point of view, it is not local, i.e. for a fixed time instant, growth is the same at each space point. ## Introduction Nucleation and growth processes arise in several natural and technological applications (cf. [8, 7] and the references therein) such as, for example, solidification and phase–transition of materials, semiconductor crystal growth, biomineralization, and DNA replication (cf., e.g., [15]). A _birth–and–growth process_ is a RaCS family given by $\Theta_{t}=\bigcup_{n:T_{n}\leq t}\Theta_{T_{n}}^{t}(X_{n})$, for $t\in\mathbb{R}_{+}$, where $\Theta^{t}_{T_{n}}\left(X_{n}\right)$ is the RaCS obtained as the evolution up to time $t>T_{n}$ of the germ born at (random) time $T_{n}$ in (random) location $X_{n}$, according to some growth model. An analytical approach is often used to model birth–and–growth process, in particular it is assumed that the growth is driven according to a non–negative normal velocity, i.e. for every instant $t$, a border point $x\in\partial\Theta_{t}$ “grows” along the outward normal unit (e.g. [3, 22, 13, 4, 6, 5, 11]). Thus, growth is pointwise isotropic; i.e. given a point belonging $\partial\Theta_{t}$, the growth rate is independently from outward normal direction. Note that, the existence of the outward normal vector imposes a regularity condition on $\partial\Theta_{t}$ and also on the nucleation process (it cannot be a point process). This paper is an attempt to offer an original alternative approach based on a purely geometric stochastic point of view, in order to avoid regularity assumptions describing birth–and–growth process. In particular, Minkowski sum (already employed in [19] to describe self–similar growth for a single convex germ) and Aumann integral are used here to derive a mathematical model of such process. This model, that emphasizes the geometric growth without regularity assumptions on $\partial\Theta_{t}$, is rigorously defined as a suitable combination of two very general set–valued processes representing nucleation $\left\\{B_{t}\right\\}_{t\in[t_{0},T]}$ and growth $\left\\{G_{t}\right\\}_{t\in[t_{0},T]}$ respectively $\begin{array}[]{rl}\Theta_{t}=&\left(\Theta_{t_{0}}\oplus\int_{t_{0}}^{t}G_{s}ds\right)\cup\bigcup_{s\in[t_{0},t]}dB_{s}\\\ d\Theta_{t}=&\oplus G_{t}dt\cup dB_{t}\qquad\textrm{ or }\qquad\Theta_{t+dt}=(\Theta_{t}\oplus G_{t}dt)\cup dB_{t}.\end{array}$ Roughly speaking, increment $d\Theta_{t}$, during an infinitesimal time interval $dt$, is an enlargement due to an infinitesimal Minkowski addend $G_{t}dt$ followed by the union with the infinitesimal nucleation $dB_{t}$. As a consequence of Minkowski sum definition, for every instant $t$, each point $x\in\Theta_{t}$ (and then each point $x\in\partial\Theta_{t}$) grows up by $G_{t}dt$ and no regularity border assumptions are required. Then we deal with _not–local_ growth; i.e. growth is the same Minkowski addend for every $x\in\Theta_{t}$. Nevertheless, under mesoscale hypothesis we can only consider constant growth region as described, for example, in [6]. On the other hand, growth is anisotropic whenever $G_{t}$ is not a ball. The aim of this paper is to ensure the well–posedness of such a model and, hence, to show that above “integral” and “differential” notations are meaningful. In view of well–posedness, in [1], the authors show how the model leads to different and significant statistical results. The article is organized as follows. Section 0.1 contains some assumptions about (random) closed sets and their basilar properties. Model assumptions are collected in Section 0.2 and integrability properties of growth process are studied in Section 0.3. For the sake of simplicity, we present, in Section 0.4, main results of the paper (that imply well-posedness of the model), whilst correspondent proofs are in Section 0.4.1. At the last, Section 0.5 proposes a discrete time point of view, also justifying integral and differential notations. ## 0.1 Preliminary results Let $\mathbb{N}$, $\mathbb{Z}$, $\mathbb{R}$, $\mathbb{R}_{+}$ be the sets of all non–negative integer, integer, real and non–negative real numbers respectively. Let $\mathfrak{X}$, ${\mathfrak{X}^{*}}$, $B^{*}_{1}$ be a Banach space, its dual space and the unit ball of the dual space centered in the origin respectively. We shall consider $\begin{array}[]{llcll}\mathfrak{P}^{\,0}(\mathfrak{X})&=\textrm{ the family of all subsets of }\mathfrak{X},&&\mathfrak{P}(\mathfrak{X})&=\mathfrak{P}^{\,0}(\mathfrak{X})\setminus\\{\emptyset\\}\\\ \mathbb{F}^{\,0}(\mathfrak{X})&=\textrm{ the family of all closed subsets of }\mathfrak{X},&&\mathbb{F}(\mathfrak{X})&=\mathbb{F}^{\,0}(\mathfrak{X})\setminus\\{\emptyset\\}.\end{array}$ The suffixes $c$ and $b$ denote convexity and boundedness properties respectively (e.g. $\mathbb{F}^{\,0}_{bc}(\mathfrak{X})$ denotes the family of all closed, bounded and convex subsets of $\mathfrak{X}$). For all $A,B\in\mathfrak{P}^{\,0}(\mathfrak{X})$ and $\alpha\in\mathbb{R}_{+}$, let us define $\begin{array}[]{rll}A+B=&\left\\{a+b:a\in A,\ b\in B\right\\}=\bigcup_{b\in B}b+A,&\textrm{(Minkowski Sum)}\\\ \alpha\cdot A=&\alpha A=\left\\{\alpha a:a\in A\right\\},&\textrm{(Scalar Product)}\end{array}$ By definition, $\forall A\in\mathfrak{P}^{\,0}(\mathfrak{X})$, $\alpha\in\mathbb{R}_{+}$, we have $\emptyset+A=\emptyset=\alpha\emptyset$. It is well known that $+$ is a commutative and associative operation with a neutral element but $(\mathfrak{P}(\mathfrak{X}),+)$ is not a group (cf. [20]). The following relations are useful in the sequel (see [21]): for all $\forall A,B,C\in\mathfrak{P}(\mathfrak{X})$ $\begin{array}[]{c}(A\cup B)+C=(A+C)\cup(B+C)\\\ \textrm{if }B\subseteq C,\quad A+B\subseteq A+C\end{array}$ In the following, we shall work with closed sets. In general, if $A,B\in\mathbb{F}^{\,0}(\mathfrak{X})$ then $A+B$ does not belong to $\mathbb{F}^{\,0}(\mathfrak{X})$ (e.g., in $\mathfrak{X}=\mathbb{R}$ let $A=\left\\{n+1/n:n>1\right\\}$ and $B=\mathbb{Z}$, then $\left\\{1/n=\left(n+1/n\right)+(-n)\right\\}\subset A+B$ and $1/n\downarrow 0$, but $0\not\in A+B$). In view of this fact, we define $A\oplus B=\overline{A+B}$ where $\overline{(\cdot)}$ denotes the closure in $\mathfrak{X}$. For any $A,B\in\mathbb{F}(\mathfrak{X})$ the _Hausdorff distance_ (or _metric_) is defined by $\delta_{H}(A,B)=\max\left\\{\sup_{a\in A}\inf_{b\in B}\left\|a-b\right\|_{\mathfrak{X}},\sup_{b\in B}\inf_{a\in A}\left\|a-b\right\|_{\mathfrak{X}}\right\\}.$ For all $(x^{*},A)\in B^{*}_{1}\times\mathbb{F}(\mathfrak{X})$, the _support function_ is defined by $s(x^{*},A)=\sup_{a\in A}x^{*}(a)$. It can be proved (cf. [14, 2]) that for each $A,B\in\mathbb{F}_{bc}(\mathfrak{X})$, $\delta_{H}(A,B)=\sup\left\\{\left|s(x^{*},A)-s(x^{*},B)\right|:x^{*}\in B^{*}_{1}\right\\}.$ (1) Let $(\Omega,\mathfrak{F})$ be a measurable space with $\mathfrak{F}$ complete with respect to some $\sigma$-finite measure, let $X:\Omega\to\mathfrak{P}^{\,0}(\mathfrak{X})$ be a set–valued map, and $\displaystyle D(X)$ $\displaystyle=\left\\{\omega\in\Omega:X(\omega)\neq\emptyset\right\\}$ be the domain of $X$ $\displaystyle X^{-1}(A)$ $\displaystyle=\left\\{\omega\in\Omega:X(\omega)\cap A\neq\emptyset\right\\},\quad A\subset\mathfrak{X},$ be the inverse image of $X$ Roughly speaking, $X^{-1}(A)$ is the set of all $\omega$ such that $X(\omega)$ hits set $A$. Different definitions of measurability for set–valued functions are developed over the years by several authors (cf. [17, 10, 2, 16] and reference therein). Here, $X$ is _measurable_ if, for each $O$, open subset of $\mathfrak{X}$, $X^{-1}(O)\in\mathfrak{F}$. ###### Proposition 0.1.1 _(See[17]) _ $X:\Omega\to\mathfrak{P}^{\,0}(\mathfrak{X})$ is a measurable set–valued map if and only if $D(X)\in\mathfrak{F}$, and $\omega\mapsto d(x,X(\omega))$ is a measurable function of $\omega\in D(X)$ for each $x\in\mathfrak{X}$. From now on, $\mathcal{U}[\Omega,\mathfrak{F},\mu;\mathbb{F}(\mathfrak{X})]$ ($=\mathcal{U}[\Omega;\mathbb{F}(\mathfrak{X})]$ if the measure $\mu$ is clear) denotes the family of $\mathbb{F}(\mathfrak{X})$–valued measurable maps (analogous notation holds whenever $\mathbb{F}(\mathfrak{X})$ is replaced by another family of subsets of $\mathfrak{X}$). Let $(\Omega,\mathfrak{F},\mathbb{P})$ be a complete probability space and let $X\in\mathcal{U}[\Omega,\mathfrak{F},\mathbb{P};\mathbb{F}(\mathfrak{X})]$, then $X$ is a RaCS. It can be proved (see [18]) that, if $X,X_{1},X_{2}$ are RaCS and if $\xi$ is a measurable real–valued function, then $X_{1}\oplus X_{2}$, $X_{1}\ominus X_{2}$, $\xi X$ and $(\textrm{Int}\ X)^{C}$ are RaCS. Moreover, if $\left\\{X_{n}\right\\}_{n\in\mathbb{N}}$ is a sequence of RaCS then $X=\overline{\bigcup_{n\in\mathbb{N}}X_{n}}$ is so. Let $(\Omega,\mathfrak{F},\mu)$ be a finite measure space (although most of the results are valid for $\sigma$-finite measures space). The _Aumann integral_ of $X\in\mathcal{U}[\Omega,\mathfrak{F},\mu;\mathbb{F}(\mathfrak{X})]$ is defined by $\int_{\Omega}Xd\mu=\left\\{\int_{\Omega}xd\mu:x\in S_{X}\right\\},$ where $S_{X}=\left\\{x\in L^{1}[\Omega;\mathfrak{X}]:x\in X\ \mu-\textrm{a.e.}\right\\}$ and $\int_{\Omega}xd\mu$ is the usual Bochner integral in $L^{1}[\Omega;\mathfrak{X}]$. Moreover, $\int_{A}Xd\mu=\left\\{\int_{A}xd\mu:x\in S_{X}\right\\}$ for $A\in\mathfrak{F}$. If $\mu$ is a probability measure, we denote the Aumann integral by $\mathbb{E}{X}=\int_{\Omega}Xd\mu$. Let $X\in\mathcal{U}[\Omega,\mathfrak{F},\mu;\mathbb{F}(\mathfrak{X})]$, it is _integrably bounded_ , and we shall write $X\in L^{1}[\Omega,\mathfrak{F},\mu;\mathbb{F}(\mathfrak{X})]=L^{1}[\Omega;\mathbb{F}(\mathfrak{X})]$, if $\left\|X\right\|_{h}\in L^{1}[\Omega,\mathfrak{F},\mu;\mathbb{R}]$. ## 0.2 Model assumptions Let us consider $\begin{array}[]{rl}\Theta_{t}=&\left(\Theta_{t_{0}}\oplus\int_{t_{0}}^{t}G_{s}ds\right)\cup\bigcup_{s\in[t_{0},t]}dB_{s}\\\ d\Theta_{t}=&\oplus G_{t}dt\cup dB_{t}\qquad\textrm{ or }\qquad\Theta_{t+dt}=(\Theta_{t}\oplus G_{t}dt)\cup dB_{t}.\end{array}$ (2) In fact, above equation is not a definition since, for example, problems arise handling non–countable union of (random) closed sets. The well–posedness of (2) and hence the existence of such a process are the main purpose of this paper. From now on, let us consider the following assumptions. * ​​​​​​​​​​(A-0) * - $(\mathfrak{X},\left\|\cdot\right\|_{\mathfrak{X}})$ is a reflexive Banach space with separable dual space $({\mathfrak{X}^{*}},\left\|\cdot\right\|_{{\mathfrak{X}^{*}}})$, (then, $\mathfrak{X}$ is separable too, see [12, Lemma II.3.16 p. 65]). * - $[t_{0},T]\subset\mathbb{R}$ is the _time observation interval_ (or _time interval_), * - $\left(\Omega,\mathfrak{F},\left\\{\mathfrak{F}_{t}\right\\}_{t\in[t_{0},T]},\mathbb{P}\right)$ is a filtered probability space, where the filtration $\left\\{\mathfrak{F}_{t}\right\\}_{t\in[t_{0},T]}$ is assumed to have the usual properties. (_Nucleation Process_). $B=\left\\{B(\omega,t)=B_{t}:\omega\in\Omega,\ t\in[t_{0},T]\right\\}$ is a process with non–empty closed values, i.e. $B:\Omega\times[t_{0},T]\to\mathbb{F}(\mathfrak{X})$ such that * ​​​​​​​​​​(A-1) $B(\cdot,t)\in\mathcal{U}[\Omega,\mathfrak{F}_{t},\mathbb{P};\mathbb{F}(\mathfrak{X})]$, for every $t\in[t_{0},T]$, i.e. $B_{t}$ is an _adapted_ (to $\left\\{\mathfrak{F}_{t}\right\\}_{t\in[t_{0},T]}$) process. * ​​​​​​​​​​(A-2) $B_{t}$ is increasing: for every $t,s\in[t_{0},T]$ with $s<t$, $B_{s}\subseteq B_{t}$. (_Growth Process_). $G=\left\\{G_{t}=G(\omega,t):\omega\in\Omega,\ t\in[t_{0},T]\right\\}$ is a process with non–empty closed values, i.e. $G:\Omega\times[t_{0},T]\to\mathbb{F}(\mathfrak{X})$ such that * ​​​​​​​​​​(A-3) for every $\omega\in\Omega$ and $t\in[t_{0},T]$, $0\in G(\omega,t)$. * ​​​​​​​​​​(A-4) for every $\omega\in\Omega$ and $t\in[t_{0},T]$, $G(\omega,t)$ is convex, i.e. $G:\Omega\times[t_{0},T]\to\mathbb{F}_{c}(\mathfrak{X})$. * ​​​​​​​​​​(A-5) there exists $K\in\mathbb{F}_{b}(\mathfrak{X})$ such that $G(\omega,t)\subseteq K$ for every $t\in[t_{0},T]$ and $\omega\in\Omega$. As a consequence, $G(\omega,t)\in\mathbb{F}_{b}(\mathfrak{X})$ and $\left\|G(\omega,t)\right\|_{h}\leq\left\|K\right\|_{h}$, $\forall(\omega,t)\in\Omega\times[t_{0},T]$. In order to establish the well–posedness of integral $\int_{t_{0}}^{t}G_{s}ds$ in (2), let us consider a suitable hypothesis of measurability for $G$ (analogously to what is). A $\mathbb{F}(\mathfrak{X})$–valued process $G=\left\\{G_{t}\right\\}_{t\in[t_{0},T]}$ has _left continuous trajectories_ on $[t_{0},T]$ if, for every $\overline{t}\in[t_{0},T]$ with $t<\overline{t}$, $\lim_{t\to\overline{t}}\delta_{H}\left(G(\omega,t),G(\omega,\overline{t})\right)=0,\qquad\textrm{a.s.}$ The $\sigma$-algebra on $\Omega\times[t_{0},T]$ generated by the processes $\left\\{G_{t}\right\\}_{t\in[t_{0},T]}$ with left continuous trajectories on $[t_{0},T]$, is called the _previsible_ (or _predictable_) $\sigma$-algebra and it is denoted by $\mathcal{P}$. ###### Proposition 0.2.1 __The previsible $\sigma$-algebra is also generated by the collection of random sets $A\times t_{0}$ where $A\in\mathfrak{F}_{t_{0}}$ and $A\times(s,t]$ where $A\in\mathfrak{F}_{s}$ and $(s,t]\subset[t_{0},T]$. Proof. Let the $\sigma$-algebra generated by the above collection of sets be denoted by $\mathcal{P}^{\prime}$. We shall show $\mathcal{P}=\mathcal{P}^{\prime}$. Let $G$ be a left continuous process and let $\alpha=(T-t_{0})$, consider for $n\in\mathbb{N}$ $G_{n}(\omega,t)=\left\\{\begin{array}[]{ll}G(\omega,t_{0}),&t=t_{0}\\\ \\\ G\left(\omega,t_{0}+\frac{k\alpha}{2^{n}}\right),&\begin{array}[]{c}\left(t_{0}+\frac{k\alpha}{2^{n}}\right)<t\leq\left(t_{0}+\frac{(k+1)\alpha}{2^{n}}\right)\\\ k\in\left\\{0,\ldots,(2^{n}-1)\right\\}\end{array}\end{array}\right.$ It is clear that $G_{n}$ is $\mathcal{P}^{\prime}$-measurable, since $G$ is adapted. As $G$ is left continuous, the above sequence of left-continuous processes converges pointwise (with respect to $\delta_{H}$) to $G$ when $n$ tends to infinity, so $G$ is $\mathcal{P}^{\prime}$-measurable, thus $\mathcal{P}\subseteq\mathcal{P}^{\prime}$. Conversely consider $A\times(s,t]\in\mathcal{P}^{\prime}$ with $(s,t]\subset[t_{0},T]$ and $A\in\mathfrak{F}_{s}$. Let $b\in\mathfrak{X}\setminus\\{0\\}$ and $G$ be the process $G(\omega,v)=\left\\{\begin{array}[]{ll}b,&v\in(s,t],\ \omega\in A\\\ 0,&\textrm{otherwise}\end{array}\right.$ this function is adapted and left continuous, hence $\mathcal{P}^{\prime}\subseteq\mathcal{P}$. $\blacksquare$ Then let us consider the following assumption. * ​​​​​​​​​​(A-6) $G$ is $\mathcal{P}$-measurable. ## 0.3 Growth process properties Theorem 0.3.2 is the main result in this section. It shows that $\omega\mapsto\int_{a}^{b}G(\omega,\tau)d\tau$ is a RaCS with non–empty bounded convex values. This is the first step in order to obtain well–posedness of (2). ###### Proposition 0.3.1 __Suppose ​​​​​​​​​​(A-3), …, ​​​​​​​​​​(A-6), and let $\mu_{\lambda}$ be the Lebesgue measure on $[t_{0},T]$, then * • $G(\omega,\cdot)\in\mathcal{U}\left[[t_{0},T],\mathcal{B}_{[t_{0},T]},\mu_{\lambda};\mathbb{F}_{bc}(\mathfrak{X})\right]$ for every $\omega\in\Omega$. * • $G(\cdot,t)\in\mathcal{U}[\Omega,\widetilde{\mathfrak{F}}_{t^{-}},\mathbb{P};\mathbb{F}_{bc}(\mathfrak{X})]$ for each $t\in[t_{0},T]$, where $\widetilde{\mathfrak{F}}_{t^{-}}$ is the so called _history $\sigma$-algebra_ i.e. $\widetilde{\mathfrak{F}}_{t^{-}}=\sigma\left(\mathfrak{F}_{s}:0\leq s<t\right)\subseteq\mathfrak{F}$. * • $G\in\ L^{1}[[t_{0},T],\mathcal{B}_{[t_{0},T]},\mu_{\lambda};\mathbb{F}_{bc}(\mathfrak{X})]\cap L^{1}[\Omega,\mathfrak{F},\mathbb{P};\mathbb{F}_{bc}(\mathfrak{X})]$ Proof. Assumptions ​​​​​​​​​​(A-3) and ​​​​​​​​​​(A-4) imply that $G$ is non–empty and convex. Measurability and integrability properties are consequence of ​​​​​​​​​​(A-6) and ​​​​​​​​​​(A-5) respectively. $\blacksquare$ ###### Theorem 0.3.2 __Suppose ​​​​​​​​​​(A-3), …, ​​​​​​​​​​(A-6). For every $a,b\in[t_{0},T]$, the integral $\int_{a}^{b}G(\omega,\tau)d\tau$ is non–empty and the set–valued map $\begin{array}[]{rccl}G_{a,b}:&\Omega&\to&\mathfrak{P}(\mathfrak{X})\\\ &\omega&\mapsto&\int_{a}^{b}G(\omega,\tau)d\tau\end{array}$ is measurable. Moreover, $G_{a,b}$ is a non–empty, bounded convex RaCS. In order to prove Theorem 0.3.2, consider following properties for real processes. A real–valued process $X=\left\\{X_{t}\right\\}_{t\in[t_{0},T]}$ is _predictable_ with respect to filtration $\left\\{\mathfrak{F}_{t}\right\\}_{t\in\mathbb{R}_{+}}$, if it is measurable with respect to the _predictable $\sigma$-algebra_ $\mathcal{P}_{\mathbb{R}}$, i.e. the $\sigma$-algebra generated by the collection of random sets $A\times\left\\{0\right\\}$ where $A\in\mathfrak{F}_{0}$ and $A\times(s,t]$ where $A\in\mathfrak{F}_{s}$. ###### Proposition 0.3.3 _(See[9, Propositions 2.30, 2.32 and 2.41]) _Let $X=\left\\{X_{t}\right\\}_{t\in[t_{0},T]}$ be a predictable real–valued process, then $X$ is $(\mathfrak{F}\otimes\mathcal{B}_{[t_{0},T]},\mathcal{B}_{\mathbb{R}})$-measurable. Further, for every $\omega\in\Omega$, the trajectory $X(\omega,\cdot):[t_{0},T]\to\mathbb{R}$ is $(\mathcal{B}_{[t_{0},T]},\mathcal{B}_{\mathbb{R}})$-measurable . ###### Lemma 0.3.4 __Let $x^{*}$ be an element of the unit ball in the dual space $B^{*}_{1}$, then $G\mapsto s(x^{*},G)$ is a measurable map. Proof. By definition $s(x^{*},G)=\sup\left\\{x^{*}(g):g\in G\right\\}$. Since $\mathfrak{X}$ is separable ​​​​​​​​​​(A-0), there exists $\left\\{g_{n}\right\\}_{n\in\mathbb{N}}\subset G$ such that $G=\overline{\left\\{g_{n}\right\\}}$. Then, for every $x^{*}\in B^{*}_{1}$ we have $s(x^{*},G)=\sup_{g\in G}x^{*}(g)=\sup_{n\in\mathbb{N}}x^{*}(g_{n}).$ Since $x^{*}$ is a continuous map then, $s(x^{*},\cdot)$ is measurable. $\blacksquare$ Proof of Theorem 0.3.2. At first, we prove that $G_{a,b}$ is a measurable map. From Proposition 0.3.1, integral $G_{a,b}=\int_{a}^{b}G(\omega,\tau)d\tau$ is well defined for all $\omega\in\Omega$. Assumption ​​​​​​​​​​(A-3) implies $0\in G_{a,b}(\omega)\neq\emptyset$ for every $\omega\in\Omega$. Hence, the domain of $G_{a,b}$ is the whole $\Omega$ for all $a,b\in[t_{0},T]$ $D\left(G_{a,b}\right)=\left\\{\omega\in\Omega:G_{a,b}\neq\emptyset\right\\}=\Omega\in\mathfrak{F}.$ Thus, by Proposition 0.1.1 and for a fixed couple $a,b\in[t_{0},T]$, $G_{a,b}$ is (weakly) measurable if and only if, for every $x\in\mathfrak{X}$, the map $\omega\mapsto d\left(x,\int_{a}^{b}G(\omega,\tau)d\tau\right)=\delta_{H}\left(x,\int_{a}^{b}G(\omega,\tau)d\tau\right)$ (3) is measurable. Equation (1) guarantees that (3) is measurable if and only if, for every $x\in\mathfrak{X}$, the map $\omega\mapsto\sup_{x^{*}\in B^{*}_{1}}\left|s(x^{*},x)-s\left(x^{*},\int_{a}^{b}G(\omega,\tau)d\tau\right)\right|$ is measurable. The above expression can be computed on a countable family dense in $B^{*}_{1}$ (note that such family exists since ${\mathfrak{X}^{*}}$ is assumed separable ​​​​​​​​​​(A-0)): $\omega\mapsto\sup_{n\in\mathbb{N}}\left|s(x^{*}_{i},x)-s\left(x^{*}_{i},\int_{a}^{b}G(\omega,\tau)d\tau\right)\right|.$ It can be proved ([18, Theorem 2.1.12 p. 46]) that $s\left(x^{*},\int_{a}^{b}G(\omega,\tau)d\tau\right)=\int_{a}^{b}s\left(x^{*},G(\omega,\tau)\right)d\tau,\qquad\forall x^{*}\in B^{*}_{1}$ and therefore, since $s(x^{*}_{i},x)$ is a constant, $G_{a,b}$ is measurable if, for every $x^{*}\in\left\\{x_{i}^{*}\right\\}_{i\in\mathbb{N}}$, the following map $\begin{array}[]{ccl}(\Omega,\mathfrak{F})&\to&(\mathbb{R},\mathcal{B}_{\mathbb{R}})\\\ \omega&\mapsto&\int_{a}^{b}s\left(x^{*},G(\omega,\tau)\right)d\tau\end{array}$ (4) is measurable. Note that $s(x^{*},G(\cdot,\cdot))$, as a map from $\Omega\times[t_{0},T]$ to $\mathbb{R}$, is predictable since it is the composition of a predictable map ​​​​​​​​​​(A-6) with a measurable one (see Lemma 0.3.4): $\begin{array}[]{rccccc}s\left(x^{*},G(\cdot,\cdot)\right):&(\Omega\times[t_{0},T],\mathcal{P})&\to&(\mathbb{F}(\mathfrak{X}),\sigma_{f})&\to&(\mathbb{R},\mathcal{B}_{\mathbb{R}})\\\ &(\omega,t)&\mapsto&G(\omega,t)&\mapsto&s\left(x^{*},G(\omega,t)\right)\end{array}$ thus, by Proposition 0.3.3, it is a $\mathcal{P}$-measurable map and hence (4) is a measurable map. In view of the first part, it remains to prove that $G_{a,b}$ is a bounded convex set for a.e. $\omega\in\Omega$. Since $\mathfrak{X}$ is reflexive ​​​​​​​​​​(A-0), by Proposition 0.3.1 we have that $G_{a,b}$ is closed ([18, Theorem 2.2.3]). Further, $G_{a,b}$ is also convex (see [18, Theorem 2.1.5 and Corollary 2.1.6]). To conclude the proof, it is sufficient to show that $G_{a,b}$ is included in a bounded set: $\displaystyle\int_{a}^{b}G(\omega,\tau)d\tau$ $\displaystyle=$ $\displaystyle\left\\{\int_{a}^{b}g(\omega,\tau)d\tau:g(\omega,\cdot)\in G(\omega,\cdot)\subseteq K\right\\}$ $\displaystyle\subseteq$ $\displaystyle\left\\{\int_{a}^{b}kd\tau:k\in K\right\\}=\left\\{(b-a)k:k\in K\right\\}=(b-a)K.$ $\blacksquare$ ## 0.4 Geometric Random Process For the sake of simplicity, let us present the main results which proofs will be given in Section 0.4.1. Let us assume conditions from ​​​​​​​​​​(A-0) to ​​​​​​​​​​(A-6). For every $t\in[t_{0},T]\subset\mathbb{R}$, $n\in\mathbb{N}$ and $\Pi=\left(t_{i}\right)_{i=0}^{n}$ partition of $[t_{0},t]$, let us define $\displaystyle s_{\Pi}(t)=$ $\displaystyle\left(B_{t_{0}}\oplus\int_{t_{0}}^{t}G(\tau)d\tau\right)\cup\bigcup_{i=1}^{n}\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{t}G(\tau)d\tau\right)$ (5) $\displaystyle S_{\Pi}(t)=$ $\displaystyle\left(B_{t_{0}}\oplus\int_{t_{0}}^{t}G(\tau)d\tau\right)\cup\bigcup_{i=1}^{n}\left(\Delta B_{t_{i}}\oplus\int_{t_{i-1}}^{t}G(\tau)d\tau\right)$ (6) where $\Delta B_{t_{i}}=B_{t_{i}}\setminus B_{t_{i-1}}^{o}$ ($B_{t_{i-1}}^{o}$ denotes the interior set of $B_{t_{i-1}}$) and where the integral is in the Aumann sense with respect to the Lebesgue measure $d\tau=d\mu_{\lambda}$. We write $s_{\Pi}$ and $S_{\Pi}$ instead of $s_{\Pi}(t)$ and $S_{\Pi}(t)$ when the dependence on $t$ is clear. Proposition 0.4.1 guarantees that both $s_{\Pi}$ and $S_{\Pi}$ are well defined RaCS, further, Proposition 0.4.3 shows $s_{\Pi}\subseteq S_{\Pi}$ as a consequence of different time intervals integration: if the time interval integration of $G$ increases then the integral of $G$ does not decrease with respect to set-inclusion (Lemma 0.4.2). Proposition 0.4.4 means that $\left\\{s_{\Pi}\right\\}$ ($\left\\{S_{\Pi}\right\\}$) increases (decreases) whenever a refinement of $\Pi$ is considered. At the same time, Proposition 0.4.5 implies that $s_{\Pi}$ and $S_{\Pi}$ become closer each other (in the Hausdorff distance sense) when partition $\Pi$ becomes finer. The “limit” is independent on the choice of the refinement as consequence of Proposition 0.4.6. Corollary 0.4.7 means that, given any $\left\\{\Pi_{j}\right\\}_{j\in\mathbb{N}}$ refinement sequence of $[t_{0},t]$, the random closed sets $s_{\Pi_{j}}$ and $S_{\Pi_{j}}$ play the same role that lower sums and upper sums have in classical analysis when we define the Riemann integral. In fact, if $\Theta_{t}$ denotes their limit value (see (7)), $s_{\Pi_{j}}$ and $S_{\Pi_{j}}$ are a lower and an upper approximation of $\Theta_{t}$ respectively. Note that, as a consequence of monotonicity of $s_{\Pi_{j}}$ and $S_{\Pi_{j}}$, we avoid problems that may arise considering uncountable unions in integral expression in (2). ###### Proposition 0.4.1 __Let $\Pi$ be a partition of $[t_{0},t]$. Both $s_{\Pi}$ and $S_{\Pi}$, defined in (5) and (6), are RaCS. ###### Lemma 0.4.2 __Let $X\in L^{1}[I,\mathfrak{F},\mu_{\lambda};\mathbb{F}(\mathfrak{X})]$, where $I$ is a bounded interval of $\mathbb{R}$, such that $0\in X$ $\mu_{\lambda}$-almost everywhere on $I$ and let $I_{1},I_{2}$ be two other intervals of $\mathbb{R}$ with $I_{1}\subset I_{2}\subset I$. Then $\int_{I_{1}}X(\tau)d\tau\subseteq\int_{I_{2}}X(\tau)d\tau.$ ###### Proposition 0.4.3 __Let $\Pi$ be a partition of $[t_{0},t]$. Then $s_{\Pi}\subseteq S_{\Pi}$ almost surely. ###### Proposition 0.4.4 __Let $\Pi$ and $\Pi^{\prime}$ be two partitions of $[t_{0},t]$ such that $\Pi^{\prime}$ is a refinement of $\Pi$. Then, almost surely, $s_{\Pi}\subseteq s_{\Pi^{\prime}}$ and $S_{\Pi^{\prime}}\subseteq S_{\Pi}$. ###### Proposition 0.4.5 __Let $\left\\{\Pi_{j}\right\\}_{j\in\mathbb{N}}$ be a refinement sequence of $[t_{0},t]$ (i.e. $\left|\Pi_{j}\right|\to 0$ if $j\to\infty$). Then, almost surely, $\lim_{j\to\infty}\delta_{H}\left(s_{\Pi_{j}},S_{\Pi_{j}}\right)=0$. ###### Proposition 0.4.6 __Let $\left\\{\Pi_{j}\right\\}_{j\in\mathbb{N}}$ and $\left\\{\Pi_{l}^{\prime}\right\\}_{l\in\mathbb{N}}$ be two distinct refinement sequences of $[t_{0},t]$, then, almost surely, $\lim_{\textrm{\tiny$\begin{array}[]{c}j\rightarrow\infty\\\ l\rightarrow\infty\end{array}$}}\delta_{H}\left(s_{\Pi_{j}},s_{\Pi^{\prime}_{l}}\right)=0\qquad\textrm{and}\qquad\lim_{\textrm{\tiny$\begin{array}[]{c}j\rightarrow\infty\\\ l\rightarrow\infty\end{array}$}}\delta_{H}\left(S_{\Pi_{j}},S_{\Pi^{\prime}_{l}}\right)=0.$ ###### Corollary 0.4.7 __For every $\left\\{\Pi_{j}\right\\}_{j\in\mathbb{N}}$ refinement sequence of $[t_{0},t]$, the following limits exist $\overline{\left(\bigcup_{j\in\mathbb{N}}s_{\Pi_{j}}\right)},\ \overline{\left(\lim_{j\rightarrow\infty}s_{\Pi_{j}}\right)},\ \lim_{j\rightarrow\infty}S_{\Pi_{j}},\ \bigcap_{j\in\mathbb{N}}S_{\Pi_{j}},$ (7) and they are equals almost surely. The convergences is taken with respect to the Hausdorff distance. We are now ready to define the continuous time stochastic process. ###### Definition 0.4.8 __Assume ​​​​​​​​​​(A-0), …, ​​​​​​​​​​(A-6). For every $t\in[t_{0},T]$, let $\left\\{\Pi_{j}\right\\}_{j\in\mathbb{N}}$ be a refinement sequence of the time interval $[t_{0},t]$ and let $\Theta_{t}$ be the RaCS defined by $\overline{\left(\bigcup_{j\in\mathbb{N}}s_{\Pi_{j}}(t)\right)}=\overline{\left(\lim_{j\rightarrow\infty}s_{\Pi_{j}}(t)\right)}=\Theta_{t}=\lim_{j\rightarrow\infty}S_{\Pi_{j}}(t)=\bigcap_{j\in\mathbb{N}}S_{\Pi_{j}}(t),$ then, the family $\Theta=\left\\{\Theta_{t}:t\in[t_{0},T]\right\\}$ is called _geometric random process G-RaP_ (on $[t_{0},T]$). ###### Theorem 0.4.9 __Let $\Theta$ be a G-RaP on $[t_{0},T]$, then $\Theta$ is a non-decreasing process with respect to the set inclusion, i.e. $\mathbb{P}\left(\Theta_{s}\subseteq\Theta_{t},\ \forall t_{0}\leq s<t\leq T\right)=1.$ Moreover, $\Theta$ is adapted with respect to filtration $\left\\{\mathfrak{F}_{t}\right\\}_{t\in[t_{0},T]}$. ###### Remark 0.4.10 __We want to point out that, assumptions we considered on $\left\\{B_{t}\right\\}$ and $\left\\{G_{t}\right\\}$ are so general, that a wide family of classical random sets and evolution processes can be described (for example, Boolean model is a birth–and–growth process with “null growth”). ### 0.4.1 Proofs of Propositions in Section 0.4 Proof of Proposition 0.4.1. For every $i\in\left\\{0,\ldots,n\right\\}$, $\int_{t_{i-1}}^{t}G(\tau)d\tau$ is a RaCS (Theorem 0.3.2). Thus, measurability Assumption ​​​​​​​​​​(A-1) on $B$ guarantees that, for every $t_{i}\in\Pi$, $B_{t_{i}}$, $\Delta B_{t_{i}}$, $\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{t}G(\tau)d\tau\right)$, and hence $s_{\Pi}$ and $S_{\Pi}$ are RaCS. $\blacksquare$ Proof of Lemma 0.4.2. Let $y\in\left(\int_{I_{1}}X(\tau)d\tau\right)$, then there exists $x\in S_{X}$, for which $y=\left(\int_{I_{1}}x(\tau)d\tau\right)$. Let us define on $I_{2}(\supset I_{1})$ $x^{\prime}(\tau)=\left\\{\begin{array}[]{ll}x(\tau),&\tau\in I_{1}\\\ 0,&\tau\in I_{2}\setminus I_{1}\end{array}\right.$ then $x^{\prime}\in S_{X}$ and $y=\left(\int_{I_{2}}x^{\prime}(\tau)d\tau\right)\in\left(\int_{I_{2}}X(\tau)d\tau\right)$. $\blacksquare$ Proof of Proposition 0.4.3. Thesis is a consequence of Lemma 0.4.2 and Minkowski addition properties, in fact $\left(\int_{t_{i-1}}^{t}G(\tau)d\tau\right)\subseteq\left(\int_{t_{i}}^{t}G(\tau)d\tau\right)$ implies $s_{\Pi}\subseteq S_{\Pi}$. $\blacksquare$ Proof of Proposition 0.4.4. Let $\Pi^{\prime}$ be a refinement of partition $\Pi$ of $[t_{0},t]$, i.e. $\Pi\subset\Pi^{\prime}$. We prove that $s_{\Pi}\subseteq s_{\Pi^{\prime}}$ ($S_{\Pi^{\prime}}\subseteq S_{\Pi}$ is analogous). It is sufficient to show the thesis only for $\Pi^{\prime}=\Pi\cup\left\\{\overline{t}\right\\}$ where $\Pi=\left\\{t_{0},\ldots,t_{n}\right\\}$ with $t_{0}<\ldots<t_{n}=t$ and $\overline{t}\in(t_{0},t)$. Let $i\in\left\\{0,\ldots,(n-1)\right\\}$ be such that $t_{i}\leq\overline{t}\leq t_{i+1}$ then $\displaystyle s_{\Pi}$ $\displaystyle=$ $\displaystyle\left(B_{t_{0}}\oplus\int_{t_{0}}^{t}G(\tau)d\tau\right)\cup\bigcup_{\textrm{\tiny$\begin{array}[]{c}j=1\\\ j\neq i+1\end{array}$}}^{n}\left(\Delta B_{t_{j}}\oplus\int_{t_{j}}^{t}G(\tau)d\tau\right)\cup$ $\displaystyle\left[\left(B_{t_{i+1}}\setminus B_{t_{i}}^{o}\right)\oplus\int_{t_{i+1}}^{t}G(\tau)d\tau\right]$ and $\displaystyle s_{\Pi^{\prime}}$ $\displaystyle=$ $\displaystyle\left(B_{t_{0}}\oplus\int_{t_{0}}^{t}G(\tau)d\tau\right)\cup\bigcup_{\textrm{\tiny$\begin{array}[]{c}j=1\\\ j\neq i+1\end{array}$}}^{n}\left(\Delta B_{t_{j}}\oplus\int_{t_{j}}^{t}G(\tau)d\tau\right)\cup$ $\displaystyle\left[\left(B_{\overline{t}}\setminus B_{t_{i}}^{o}\right)\oplus\int_{\overline{t}}^{t}G(\tau)d\tau\right]\cup\left[\left(B_{t_{i+1}}\setminus B_{\overline{t}}^{o}\right)\oplus\int_{t_{i+1}}^{t}G(\tau)d\tau\right]$ Definitely, in order to prove that $s_{\Pi}\subseteq s_{\Pi^{\prime}}$ we have to prove that $\displaystyle\left\\{\left[\left(B_{\overline{t}}\setminus B_{t_{i}}^{o}\right)\oplus\int_{\overline{t}}^{t}G(\tau)d\tau\right]\cup\left[\left(B_{t_{i+1}}\setminus B_{\overline{t}}^{o}\right)\oplus\int_{t_{i+1}}^{t}G(\tau)d\tau\right]\right\\}$ $\displaystyle\supseteq\left[\left(B_{t_{i+1}}\setminus B_{t_{i}}^{o}\right)\oplus\int_{t_{i+1}}^{t}G(\tau)d\tau\right].$ This inclusion is a consequence of $\left(\int_{\overline{t}}^{t}G(\tau)d\tau\right)\supseteq\left(\int_{t_{i+1}}^{t}G(\tau)d\tau\right)$ (Lemma 0.4.2) and of the Minkowski distribution property. $\blacksquare$ Proof of Proposition 0.4.5. Let $\Pi_{j}=\left(t_{i}\right)_{i=0}^{n}$ be the $j$-partition of the refinement sequence $\left\\{\Pi_{j}\right\\}_{j\in\mathbb{N}}$, then $\delta_{H}\left(s_{\Pi_{j}},S_{\Pi_{j}}\right)=\max\left\\{\sup_{x\in s_{\Pi_{j}}}d(x,S_{\Pi_{j}}),\sup_{y\in S_{\Pi_{j}}}d(y,s_{\Pi_{j}})\right\\}$ where $d(x,S_{\Pi_{j}})=\inf_{y\in S_{\Pi_{j}}}\left\|x-y\right\|_{\mathfrak{X}}$. By Proposition 0.4.3, $s_{\Pi_{j}}\subseteq S_{\Pi_{j}}$ then $\sup_{x\in s_{\Pi_{j}}}d(x,S_{\Pi_{j}})=0$ and hence we have to prove that, whenever $j\to\infty$ (i.e. $\left|\Pi_{j}\right|\to 0$), $\delta_{H}\left(s_{\Pi_{j}},S_{\Pi_{j}}\right)=\sup_{y\in S_{\Pi_{j}}}d(y,s_{\Pi_{j}})=\sup_{y\in S_{\Pi_{j}}}\inf_{x\in s_{\Pi_{j}}}\left\|x-y\right\|_{\mathfrak{X}}\longrightarrow 0.$ For every $\omega\in\Omega$, let $y$ be any element of $S_{\Pi_{j}}(\omega)$, then we distinguish two cases: (1) if $y\in\left(B_{t_{0}}(\omega)\oplus\int_{t_{0}}^{t}G(\omega,\tau)d\tau\right)$, then it is also an element of $s_{\Pi_{j}}(\omega)$, and hence $d\left(s_{\Pi_{j}}(\omega),y\right)=0$. (2) if $y\not\in\left(B_{t_{0}}(\omega)\oplus\int_{t_{0}}^{t}G(\omega,\tau)d\tau\right)$, then there exist $j\in\left\\{1,\ldots,n\right\\}$ such that $y\in\left(\Delta B_{t_{j}}(\omega)\oplus\int_{t_{j-1}}^{t}G(\omega,\tau)d\tau\right).$ By definition of $\oplus$, for every $\omega\in\Omega$, there exist $\left\\{y_{m}\right\\}_{m\in\mathbb{N}}\subseteq\left(\Delta B_{t_{j}}(\omega)+\int_{t_{j-1}}^{t}G(\omega,\tau)d\tau\right),$ such that $\lim_{m\to\infty}y_{m}=y$. Then, for every $\omega\in\Omega$, there exist $h_{m}\in\Delta B_{t_{j}}(\omega)$ and $g_{m}\in\left(\int_{t_{j-1}}^{t}G(\omega,\tau)d\tau\right)$ such that $y_{m}=(h_{m}+g_{m})$ and hence $y=\lim_{m\to\infty}(h_{m}+g_{m})=\lim_{m\to\infty}y_{m}$ where the convergence is in the Banach norm, then let $\overline{m}\in\mathbb{N}$ be such that $\left\|y-y_{m}\right\|_{\mathfrak{X}}<\left|\Pi_{j}\right|$, for every $m>\overline{m}$. Note that, for every $\omega\in\Omega$ and $m\in\mathbb{N}$, by Aumann integral definition, there exists a selection $\widehat{g_{m}}(\cdot)$ of $G(\omega,\cdot)$ (i.e. $\widehat{g_{m}}(t)\in G(\omega,t)$ $\mu_{\lambda}$-a.e.) such that $g_{m}=\int_{t_{j-1}}^{t}\widehat{g_{m}}(\tau)d\tau\qquad\textrm{ and }\qquad y_{m}={h_{m}+\int_{t_{j-1}}^{t}\widehat{g_{m}}(\tau)d\tau}.$ For every $\omega\in\Omega$, let us consider $x_{m}=h_{m}+\int_{t_{j}}^{t}\widehat{g_{m}}(\tau)d\tau$ then $x_{m}\in s_{\Pi_{j}}(\omega)$ for all $m\in\mathbb{N}$. Moreover, the following chain of inequalities hold, for all $m>\overline{m}$ and $\omega\in\Omega$, $\displaystyle\inf_{x^{\prime}\in s_{\Pi_{j}}}\left\|x^{\prime}-y\right\|_{\mathfrak{X}}$ $\displaystyle\leq$ $\displaystyle\left\|x_{m}-y\right\|_{\mathfrak{X}}\leq\left\|x_{m}-y_{m}\right\|_{\mathfrak{X}}+\left\|y_{m}-y\right\|_{\mathfrak{X}}$ $\displaystyle\leq$ $\displaystyle\left\|\int_{t_{j-1}}^{t_{j}}\widehat{g_{m}}(\tau)d\tau\right\|_{\mathfrak{X}}+\left|\Pi_{j}\right|\leq\int_{t_{j-1}}^{t_{j}}\left\|\widehat{g_{m}}(\tau)\right\|_{\mathfrak{X}}d\tau+\left|\Pi_{j}\right|$ $\displaystyle\leq$ $\displaystyle\int_{t_{j-1}}^{t_{j}}\left\|G(\tau)\right\|_{h}d\tau+\left|\Pi_{j}\right|\leq\left|t_{j}-t_{j-1}\right|\left\|K\right\|_{h}+\left|\Pi_{j}\right|$ $\displaystyle\leq$ $\displaystyle\left|\Pi_{j}\right|\left(\left\|K\right\|_{h}+1\right)\stackrel{{\scriptstyle j\to\infty}}{{\longrightarrow}}0$ since $\left\|K\right\|_{h}$ is a positive constant. By the arbitrariness of $y\in S_{\Pi_{j}}(\omega)$ we obtain the thesis. $\blacksquare$ Proof of Proposition 0.4.6. Let $\Pi_{j}$ and $\Pi_{l}^{\prime}$ be two partitions of the two distinct refinement sequences $\left\\{\Pi_{j}\right\\}_{j\in\mathbb{N}}$ and $\left\\{\Pi_{l}^{\prime}\right\\}_{l\in\mathbb{N}}$ of $[t_{0},t]$. Let $\Pi^{\prime\prime}=\Pi_{j}\cup\Pi_{l}^{\prime}$ be the refinement of both $\Pi_{j}$ and $\Pi_{l}^{\prime}$. Then Proposition 0.4.4 and Proposition 0.4.3 imply that $s_{\Pi_{j}}\subseteq s_{\Pi^{\prime\prime}}\subseteq S_{\Pi^{\prime\prime}}\subseteq S_{\Pi_{l}^{\prime}}$. Therefore $s_{\Pi_{j}}\subseteq S_{\Pi_{l}^{\prime}}$ for every $j,l\in\mathbb{N}$. Then $\overline{\left(\bigcup_{j\in\mathbb{N}}s_{\Pi_{j}}\right)}\subseteq\bigcap_{l\in\mathbb{N}}S_{\Pi_{l}^{\prime}}.$ Analogously $\overline{\left(\bigcup_{l\in\mathbb{N}}s_{\Pi_{l}^{\prime}}\right)}\subseteq\bigcap_{j\in\mathbb{N}}S_{\Pi_{j}}.$ Proposition 0.4.5 concludes the proof. $\blacksquare$ In order to prove Theorem 0.4.9, let us consider the following Lemma. ###### Lemma 0.4.11 __Let $s,t\in[t_{0},T]$ with $t_{0}<s<t$ and let $\Pi^{s}$ and $\Pi^{t}$ be two partition of $[t_{0},s]$ and $[t_{0},t]$ respectively, such that $\Pi^{s}\subset\Pi^{t}$. Then $s_{\Pi^{s}}(s)\subseteq s_{\Pi^{t}}(t)\qquad\textrm{ and }\qquad S_{\Pi^{s}}(s)\subseteq S_{\Pi^{t}}(t).$ Proof. The proofs of the two inclusions are similar. Let us prove that $s_{\Pi^{s}}(s)\subseteq s_{\Pi^{t}}(t)$. Since $\Pi^{s}\subset\Pi^{t}$, then $\Pi^{s}=\left(t_{i}\right)_{i=0}^{n}$ and $\Pi^{t}=\Pi^{s}\cup\left(t_{i}\right)_{i=n+1}^{n+m}$ with $m\in\mathbb{N}$. By Lemma 0.4.2, we have that $\displaystyle s_{\Pi^{s}}(s)$ $\displaystyle=$ $\displaystyle\left(B_{t_{0}}\oplus\int_{t_{0}}^{s}G(\tau)d\tau\right)\cup\bigcup_{i=1}^{n}\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{s}G(\tau)d\tau\right)$ $\displaystyle\subseteq$ $\displaystyle\left(B_{t_{0}}\oplus\int_{t_{0}}^{t}G(\tau)d\tau\right)\cup\bigcup_{i=1}^{n}\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{t}G(\tau)d\tau\right)$ $\displaystyle\subseteq$ $\displaystyle\left(B_{t_{0}}\oplus\int_{t_{0}}^{t}G(\tau)d\tau\right)\cup\bigcup_{i=1}^{n}\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{t}G(\tau)d\tau\right)$ $\displaystyle\cup\bigcup_{i=n+1}^{n+m}\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{t}G(\tau)d\tau\right)$ i.e. $s_{\Pi^{s}}(s)\subseteq s_{\Pi^{t}}(t)$. $\blacksquare$ Proof of Theorem 0.4.9. For every $s,t\in[t_{0},T]$ with $s<t$, let $\left\\{\Pi^{s}_{i}\right\\}_{i\in\mathbb{N}}$ and $\left\\{\Pi^{t}_{i}\right\\}_{i\in\mathbb{N}}$ be two refinement sequences of $[t_{0},s]$ and $[t_{0},t]$ respectively, such that $\Pi^{s}_{i}\subset\Pi^{t}_{i}$ for every $i\in\mathbb{N}$. Then, by Lemma 0.4.11, $S_{\Pi^{s}_{i}}\subseteq S_{\Pi^{t}_{i}}$. Now, as $i$ tends to infinity, we obtain $\Theta_{s}=\bigcap_{i\rightarrow\infty}S_{\Pi^{s}_{i}}\subseteq\bigcap_{i\rightarrow\infty}S_{\Pi^{t}_{i}}=\Theta_{t}.$ For the second part, note that Theorem 0.3.2 still holds replacing $\mathfrak{F}_{t}$ instead of $\mathfrak{F}$, so that for every $s\in[t_{0},T]$, the family $\left\\{\int_{s}^{t}G(\omega,\tau)d\tau\right\\}_{t\in[s,T]}$ is an adapted process to the filtration $\left\\{\mathfrak{F}_{t}\right\\}_{t\in[t_{0},T]}$. This fact together with Assumption ​​​​​​​​​​(A-1) guarantees that $\left\\{S_{\Pi}\right\\}_{t\in[s,T]}$ is adapted for every partition $\Pi$ of $[s,T]$ and hence $\Theta$ is adapted too. $\blacksquare$ ## 0.5 Discrete time case and infinitesimal notations Let us consider $\Theta_{s}$ and $\Theta_{t}$ with $s<t$. Let $\left\\{\Pi^{s}_{j}\right\\}_{j\in\mathbb{N}}$ and $\left\\{\Pi^{t}_{j}\right\\}_{j\in\mathbb{N}}$ be two refinement sequences of $[t_{0},s]$ and $[t_{0},t]$ respectively, such that $\Pi^{s}_{j}\subset\Pi^{t}_{j}$ for every $j\in\mathbb{N}$ (i.e. $\Pi_{j}^{s}=\left(t_{i}\right)_{i=0}^{n}$ and $\Pi_{j}^{t}=\Pi_{j}^{s}\cup\left(t_{i}\right)_{i=n+1}^{n+m}$ with $n,m\in\mathbb{N}$). It is easy to compute $s_{\Pi^{t}_{j}}=\left(s_{\Pi^{s}_{j}}\oplus\int_{s}^{t}G(\tau)d\tau\right)\cup\bigcup_{i=n+1}^{n+m}\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{t}G(\tau)d\tau\right).$ Then, by Definition 0.4.8, whenever $\left|\Pi^{t}_{j}\right|\to 0$, we obtain $\Theta_{t}=\left(\Theta_{s}\oplus\int_{s}^{t}G(\tau)d\tau\right)\cup\lim_{\left|\Pi^{t}_{j}\right|\rightarrow 0}\bigcup_{i=n+1}^{n+m}\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{t}G(\tau)d\tau\right).$ (10) The following notations $G_{k}=\int_{s}^{t}G(\tau)d\tau\qquad\textrm{and}\qquad B_{k}=\lim_{\left|\Pi^{t}_{j}\right|\rightarrow 0}\bigcup_{i=n+1}^{n+m}\left(\Delta B_{t_{i}}\oplus\int_{t_{i}}^{t}G(\tau)d\tau\right)$ lead us to the set-valued discrete time stochastic process $\Theta_{k}=\left\\{\begin{array}[]{ll}(\Theta_{k-1}\oplus G_{k})\cup B_{k},&k\geq 1,\\\ B_{0},&k=0.\end{array}\right.$ In view of this, we are able to justify infinitesimal notations introduced in (2). In particular, from Equation (10), whenever $\left|\Pi^{t}_{j}\right|\rightarrow 0$, we obtain $\Theta_{t}=\left(B_{t_{0}}\oplus\int_{t_{0}}^{t}G(\tau)d\tau\right)\cup\bigcup_{s=t_{0}}^{t}\left(dB_{s}\oplus\int_{s}^{t}G(\tau)d\tau\right),\qquad t\in[t_{0},T].$ Moreover, with a little abuse of this infinitesimal notation, we get two differential formulations $d\Theta_{t}=\oplus G_{t}dt\cup dB_{t}\qquad\textrm{and}\qquad\Theta_{t+dt}=(\Theta_{t}\oplus G_{t}dt)\cup dB_{t}.$ ## References * [1] G. Aletti, E. G. Bongiorno, and V. Capasso. Statistical aspects of set–valued continuous time stochastic processes. (submitted). * [2] J. Aubin and H. Frankowska. Set–valued Analysis, volume 2 of Systems & Control: Foundations & Applications. Birkhäuser Boston Inc., 1990. * [3] G. Barles, H. M. Soner, and P. E. Souganidis. Front propagation and phase field theory. SIAM J. Control Optim., 31(2):439–469, 1993. * [4] M. Burger. Growth of multiple crystals in polymer melts. European J. Appl. Math., 15(3):347–363, 2004. * [5] M. Burger, V. Capasso, and A. Micheletti. An extension of the Kolmogorov–Avrami formula to inhomogeneous birth–and–growth processes. In Math Everywhere (G. Aletti et al., Eds). Springer, Berlin, 63–76, 2007. * [6] M. Burger, V. Capasso, and L. Pizzocchero. Mesoscale averaging of nucleation and growth models. Multiscale Model. Simul., 5(2):564–592 (electronic), 2006. * [7] V. Capasso, editor. Mathematical Modelling for Polymer Processing. Polymerization, Crystallization, Manufacturing. Mathematics in Industry, Vol. 2, Springer–Verlag, Berlin, 2003. * [8] V. Capasso. On the stochastic geometry of growth. In Morphogenesis and Pattern Formation in Biological Systems (Sekimura, T. et al. Eds). Springer, Tokyo, 45–58, 2003. * [9] V. Capasso and D. Bakstein. An Introduction to Continuous–Time Stochastic Processes. Modeling and Simulation in Science, Engineering and Technology. Birkhäuser Boston Inc., 2005. * [10] C. Castaing and M. Valadier. Convex Analysis and Measurable Multifunctions. Lecture Notes in Mathematics, Vol. 580, Springer–Verlag, Berlin, 1977\. * [11] S. N. Chiu. Johnson–Mehl tessellations: asymptotics and inferences. In Probability, finance and insurance, pages 136–149. World Sci. Publ., River Edge, NJ, 2004. * [12] N. Dunford and J. T. Schwartz. Linear Operators. Part I. Wiley Classics Library. John Wiley & Sons Inc., New York, 1988. * [13] H. J. Frost and C. V. Thompson. The effect of nucleation conditions on the topology and geometry of two–dimensional grain structures. Acta Metallurgica, 35:529–540, 1987. * [14] E. Giné, M. G. Hahn, and J. Zinn. Limit theorems for random sets: an application of probability in Banach space results. In Probability in Banach Spaces, IV (Oberwolfach, 1982). Lecture Notes in Mathematics, Vol. 990, 112–135, Springer, Berlin, 1983. * [15] J. Herrick, S. Jun, J. Bechhoefer, and A. Bensimon. Kinetic model of DNA replication in eukaryotic organisms. J.Mol.Biol., 320:741–750, 2002. * [16] F. Hiai and H. Umegaki. Integrals, conditional expectations, and martingales of multivalued functions. J. Multivariate Anal., 7(1):149–182, 1977. * [17] C. J. Himmelberg. Measurable relations. Fund. Math., 87:53–72, 1975. * [18] S. Li, Y. Ogura, and V. Kreinovich. Limit Theorems and Applications of Set–Valued and Fuzzy Set–Valued Random Variables. Vol. 43 of Theory and Decision Library. Series B: Mathematical and Statistical Methods. Kluwer Academic Publishers Group, Dordrecht, 2002. * [19] A. Micheletti, S. Patti, and E. Villa. Crystal growth simulations: a new mathematical model based on the Minkowski sum of sets. In Industry Days 2003-2004 (D.Aquilano et al. Eds), volume 2 of The MIRIAM Project, pages 130–140. Esculapio, Bologna, 2005\. * [20] H. Rådström. An embedding theorem for spaces of convex sets. Proc. Amer. Math. Soc., 3:165–169, 1952. * [21] J. Serra. Image Analysis and Mathematical Morphology. Academic Press Inc. [Harcourt Brace Jovanovich Publishers], London, 1984\. * [22] Bo Su and Martin Burger. Global weak solutions of non-isothermal front propagation problem. Electron. Res. Announc. Amer. Math. Soc., 13:46–52 (electronic), 2007.
# Faster Relative Entropy Coding with Greedy Rejection Coding Gergely Flamich Department of Engineering University of Cambridge <EMAIL_ADDRESS>&Stratis Markou∗ Department of Engineering University of Cambridge <EMAIL_ADDRESS>&José Miguel Hernández Lobato Department of Engineering University of Cambridge <EMAIL_ADDRESS>Equal contribution. ###### Abstract Relative entropy coding (REC) algorithms encode a sample from a target distribution $Q$ using a proposal distribution $P$ using as few bits as possible. Unlike entropy coding, REC does not assume discrete distributions or require quantisation. As such, it can be naturally integrated into communication pipelines such as learnt compression and differentially private federated learning. Unfortunately, despite their practical benefits, REC algorithms have not seen widespread application, due to their prohibitively slow runtimes or restrictive assumptions. In this paper, we make progress towards addressing these issues. We introduce Greedy Rejection Coding (GRC), which generalises the rejection based-algorithm of Harsha et al. (2007) to arbitrary probability spaces and partitioning schemes. We first show that GRC terminates almost surely and returns unbiased samples from $Q$, after which we focus on two of its variants: GRCS and GRCD. We show that for continuous $Q$ and $P$ over $\mathbb{R}$ with unimodal density ratio $dQ/dP$, the expected runtime of GRCS is upper bounded by $\beta D_{\mathrm{KL}}[Q\|P]+\mathcal{O}(1)$ where $\beta\approx 4.82$, and its expected codelength is optimal. This makes GRCS the first REC algorithm with guaranteed optimal runtime for this class of distributions, up to the multiplicative constant $\beta$. This significantly improves upon the previous state-of-the-art method, A* coding (Flamich et al., 2022). Under the same assumptions, we experimentally observe and conjecture that the expected runtime and codelength of GRCD are upper bounded by $D_{\mathrm{KL}}[Q\|P]+\mathcal{O}(1)$. Finally, we evaluate GRC in a variational autoencoder-based compression pipeline on MNIST, and show that a modified ELBO and an index-compression method can further improve compression efficiency. ## 1 Introduction and motivation Over the past decade, the development of excellent deep generative models (DGMs) such as variational autoencoders (VAEs; Vahdat & Kautz, 2020; Child, 2020), normalising flows (Kingma et al., 2016) and diffusion models (Ho et al., 2020) demonstrated great promise in leveraging machine learning (ML) for data compression. Many recent learnt compression approaches have significantly outperformed the best classical hand-crafted codecs across a range of domains including, for example, lossless and lossy compression of images and video (Zhang et al., ; Mentzer et al., 2020, 2022). #### Transform coding. Most learnt compression algorithms are transform coding methods: they first map a datum to a latent variable using a learnt transform, and encode it using entropy coding (Ballé et al., 2020). Entropy coding assumes discrete variables while the latent variables in DGMs are typically continuous, so most transform coding methods quantize the latent variable prior to entropy coding. Unfortunately, quantization is a non-differentiable operation. Thus, state-of- the-art DGMs trained with gradient-based optimisation must resort to some continuous approximation to quantisation during training and switch to hard quantisation for compression. Previous works have argued that using quantisation within learnt compression is restrictive or otherwise harmful, and that a method which naturally interfaces with continuous latent variables is needed (Havasi et al., 2018; Flamich et al., 2020; Theis & Agustsson, 2021; Flamich et al., 2022). #### Relative entropy coding. In this paper, we study relative entropy coding (REC; Havasi et al., 2018; Flamich et al., 2020), an alternative to quantization and entropy coding. A REC algorithm uses a proposal distribution $P$, and a public source of randomness $S$, to produce a random code which represents a single sample from a target distribution $Q$. Thus REC does not assume discrete distributions and interfaces naturally with continuous variables. Remarkably, REC has fundamental advantages over quantization in lossy compression with realism constraints (Theis & Agustsson, ; Theis et al., 2022). More generally, it finds application across a range of settings including, for example, differentially private compression for federated learning (Shah et al., 2022). #### Limitations of existing REC algorithms. While algorithms for solving REC problems already exist, most of them suffer from limitations that render them impractical. These limitations fall into three categories: prohibitively long runtimes, overly restrictive assumptions, or additional coding overheads. In this work, we study and make progress towards addressing these limitations. #### General-purpose REC algorithms. On the one hand, some REC algorithms make very mild assumptions and are therefore applicable in a wide range of REC problems (Harsha et al., 2007; Li & El Gamal, 2018). Unfortunately, these algorithms have prohibitively long runtimes. This is perhaps unsurprising in light of a result by Agustsson & Theis (2020), who showed that without additional assumptions on $Q$ and $P$, the worst-case expected runtime of any general-purpose REC algorithm scales as $\smash{2^{D_{\mathrm{KL}}[Q\|P]}}$, which is impractically slow. There are also REC algorithms which accept a desired runtime as a user-specified parameter, at the expense of introducing bias in their samples (Havasi et al., 2018; Theis & Yosri, 2022). Unfortunately, in order to reduce this bias to acceptable levels, these algorithms require runtimes of an order of $\smash{2^{D_{\mathrm{KL}}[Q\|P]}}$, and are therefore also impractical. #### Faster algorithms with additional assumptions. On the other hand, there exist algorithms which make additional assumptions in order to achieve faster runtimes. For example, dithered quantisation (Ziv, 1985; Agustsson & Theis, 2020) achieves an expected runtime of $D_{\mathrm{KL}}[Q\|P]$, which is optimal since any REC algorithm has an expected runtime of at least $D_{\mathrm{KL}}[Q\|P]$. However, it requires both $Q$ and $P$ to be uniform distributions, which limits its applicability. Recently, Flamich et al. (2022) introduced A∗ coding, an algorithm based on A∗ sampling (Maddison et al., 2014) which, under assumptions satisfied in practice, achieves an expected runtime of $D_{\infty}[Q\|P]$. Unfortunately, this runtime is sub-optimal and is not always practically fast, since $D_{\infty}[Q\|P]$ can be arbitrarily large for fixed $D_{\mathrm{KL}}[Q\|P]$. Further, as discussed in Flamich et al. (2022) this runtime also comes at a cost of an additional, substantial, overhead in codelength, which limits the applicability of A∗ coding. AS∗ codingGlobal A∗ AD∗ codingGRCDGRCSGRCGSample partitioningGlobal partitioningDyadic partitioningBranch & bound search Rejection coding Figure 1: An illustration of the relations between the variants of GRC, introduced in this work, and the variants of A∗ coding. Algorithms in purple are introduced in this work. The algorithms of Harsha et al. (2007) and Li & El Gamal (2018) are equivalent to GRCG and Global A∗ coding respectively. #### Our contributions. In this work, we address some of these limitations. First, we propose greedy rejection coding (GRC), a REC algorithm based on rejection sampling. Then, inspired by A* coding (Flamich et al., 2022), we develop GRCS and GRCD, two variants of GRC that partition the sample space to dramatically speed up termination. Figure 1 illustrates the relations between GRC and its variants with existing algorithms. We analyze the correctness and the runtime of these algorithms and, in particular, prove that GRCS has an optimal codelength and order-optimal runtime on a wide class of one-dimensional problems. In more detail, our contributions are: * • We introduce Greedy Rejection Coding (GRC), which generalises the algorithm of Harsha et al. (2007) to arbitrary probability spaces and partitioning schemes. We prove that under mild conditions, GRC terminates almost surely and returns an unbiased sample from $Q$. * • We introduce GRCS and GRCD, two variants of GRC for continuous distributions over $\mathbb{R}$, which adaptively partition the sample space to dramatically improve their convergence, inspired by AS∗ and AD∗ coding (Flamich et al., 2022), respectively. * • We prove that whenever $dQ/dP$ is unimodal, the expected runtime and codelength of GRCS is $\mathcal{O}(D_{\mathrm{KL}}[Q\|P])$. This significantly improves upon the $\mathcal{O}(D_{\infty}[Q\|P])$ runtime of AS∗ coding, which is always larger than that of GRCS. This runtime is order-optimal, while making far milder assumptions than, for example, ditered quantization. * • We provide clear experimental evidence for and conjecture that whenever $dQ/dP$ is unimodal, the expected runtime and codelength of GRCD are $D_{\mathrm{KL}}[Q\|P]$. This also significantly improves over the $D_{\infty}[Q\|P]$ empirically observed runtime of AD∗ coding. * • We implement a compression pipeline with VAEs, using GRC to compress MNIST images. We propose a modified ELBO objective and show that this, together with a practical method for compressing the indices returned by GRC further improve compression efficiency. ## 2 Background and related work #### Relative entropy coding. First, we define REC algorithms. Definition 1 is stricter than the one given by Flamich et al. (2022), as it has a stronger condition on the the expected codelength of the algorithm. In this paper, all logarithms are base 2, and all divergences are measured in bits. ###### Definition 1 (REC algorithm). Let $(\mathcal{X},\Sigma)$ be a measurable space, let $\mathcal{R}$ be a set of pairs of distributions $(Q,P)$ over $(\mathcal{X},\Sigma)$ such that $D_{\mathrm{KL}}[Q\|P]<\infty$ and $\mathcal{P}$ be the set of all distributions $P$ such that $(Q,P)\in\mathcal{R}$ for some distribution $Q$. Let $S=(S_{1},S_{2},\dots)$ be a publicly available sequence of independent and fair coin tosses, with corresponding probability space $(\mathcal{S},\mathcal{F},\mathbb{P})$ and let $\mathcal{C}=\\{0,1\\}^{*}$ be the set of all finite binary sequences. A REC algorithm is a pair of functions $\mathtt{enc}:\mathcal{R}\times\mathcal{S}\to\mathcal{C}$ and $\mathtt{dec}:\mathcal{C}\times\mathcal{P}\times\mathcal{S}\to\mathcal{X}$, such that for each $(Q,P)\in\mathcal{R}$, the outputs of the encoder $C=\mathtt{enc}(Q,P,S)$ and the decoder $X=\mathtt{dec}(P,C,S)$ satisfy $X\sim Q\quad\text{and}\quad\mathbb{E}_{S}[|C|]=D_{\mathrm{KL}}[Q\|P]+\mathcal{O}(\log D_{\mathrm{KL}}[Q\|P]),$ (1) where $|C|$ is the length of the string $C$. We call $\mathtt{enc}$ the encoder and $\mathtt{dec}$ the decoder. In practice, $S$ is implemented with a pseudo-random number generator (PRNG) with a public seed. In the remainder of this section, we discuss relevant REC algorithms, building up to GRC in section 3. Figure 2: Example run of Harsha et al. (2007), for a pair of continuous $Q$ and $P$ over $[0,1]$. The green and red regions correspond to acceptance and rejection regions at each step. Here the algorithm rejects the first two samples and accepts the third one, terminating at the third step. #### Existing REC algorithms. While there are many REC algorithms already, they suffer from various issues limiting their applicability in practice. Our proposed algorithm, Greedy Rejection Coding (GRC), is based on and generalises the rejection-based algorithm of Harsha et al. (2007), by drawing inspiration from A∗ coding (Flamich et al., 2022). Specifically, A∗ coding can be viewed as a generalisation of an algorithm due to Li & El Gamal (2018). The former generalises the latter by introducing a partitioning scheme to speed up termination. In an analogous fashion, GRC generalises Harsha et al. (2007) by also introducing partitioning schemes, to speed up termination and achieve optimal runtimes. Here we discuss relevant algorithms, building up to GRC in section 3. #### REC with rejection sampling. Harsha et al. (2007) introduced a REC algorithm based on rejection sampling, which we generalise and extend in this work. While this algorithm was originally presented for discrete $Q$ and $P$, we will show that it can be generalised to arbitrary probability spaces. In this section, we present this generalised version and in section 3 we further extend it to arbitrary partitioning schemes (see definition 5). The generalisation to arbitrary probability spaces relies on the Radon-Nikodym derivative $dQ/dP$, which is guaranteed to exist since $Q\ll P$ by definition 1. When $Q$ and $P$ both have densities, $dQ/dP$ coincides with the density ratio. At each step, the algorithm draws a sample from $P$ and performs an accept- reject step, as illustrated in fig. 2. If it rejects the sample, it rules out part of $Q$ corresponding to the acceptance region, adjusts the proposal to account for the removed mass, and repeats until acceptance. More formally, define $T_{0}$ to be the zero-measure on $(\mathcal{X},\Sigma)$, and recursively for $d\in\mathbb{N}$, set: $\displaystyle T_{d+1}(S)$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}T_{d}(S)+A_{d+1}(S),$ $\displaystyle A_{d+1}(S)\stackrel{{\scriptstyle\text{def}}}{{=}}\int_{S}\alpha_{d+1}(x)\,dP(x),$ (2) $\displaystyle t_{d}(x)$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}\frac{dT_{d}}{dP}(x),$ $\displaystyle~{}\alpha_{d+1}(x)\stackrel{{\scriptstyle\text{def}}}{{=}}\min\left\\{\frac{dQ}{dP}(x)-t_{d}(x),(1-T_{d}(\mathcal{X}))\right\\},$ (3) $\displaystyle X_{d}\sim P,$ $\displaystyle~{}~{}U_{d}\sim\text{Uniform}(0,1)$ $\displaystyle~{}\beta_{d+1}(x)\stackrel{{\scriptstyle\text{def}}}{{=}}\frac{\alpha_{d+1}(x)}{1-T_{d}(\mathcal{X})},$ (4) for all $x\in\mathcal{X},S\in\Sigma$. The algorithm terminates at the first occurrence of $U_{d}\leq\beta_{d+1}(X_{d})$. The $T_{d}$ measure corresponds to the mass that has been ruled off up to and including the $d^{\text{th}}$ rejection: $T_{1}(\mathcal{X}),T_{2}(\mathcal{X})$ and $T_{3}(\mathcal{X})$ are the sums of the blue and green masses in the left, middle and right plots of fig. 2 respectively. The $A_{d}$ measure corresponds to the acceptance mass at the $d^{\text{th}}$ step: $A_{1}(\mathcal{X}),A_{2}(\mathcal{X})$ and $A_{3}(\mathcal{X})$ are the masses of the green regions in the left, middle and right plots of fig. 2 respectively. Lastly, $t_{d},\alpha_{d}$ are the Radon-Nikodym derivatives i.e., roughly speaking, the densities, of $T_{d},A_{d}$ with respect to $P$, and $\beta_{d+1}(X_{d})$ is the probability of accepting the sample $X_{d}$. Here, the encoder $\mathtt{enc}$ amounts to keeping count of the number of rejections that occur up to the first acceptance, setting $C$ equal to this count and returning $X$ and $C$. The decoder $\mathtt{dec}$ amounts to drawing $C+1$ samples from $P$, using the same seed as the encoder, and returning the last of these samples. While this algorithm is elegantly simple and achieves optimal codelengths, Flamich & Theis (2023) showed its expected runtime is $\smash{2^{D_{\infty}[Q\|P]}}$, where $D_{\infty}[Q\|P]=\sup_{x\in\mathcal{X}}\log(dQ/dP)(x)$ is the Rényi $\infty$-divergence. Unfortunately, this is prohibitively slow in most practical cases. #### REC with Poisson & Gumbel processes. Li & El Gamal (2018) introduced a REC algorithm based on Poisson processes, referred to as Poisson Functional Representation (PFR). PFR assumes that $dQ/dP$ is bounded above, and relies on the fact that (Kingman, 1992), if $T_{n}$ are the ordered arrival times of a homogeneous Poisson process on $\mathbb{R}^{+}$ and $X_{n}\sim P$, then $N\stackrel{{\scriptstyle\text{def}}}{{=}}\operatorname*{arg\,min}_{n\in\mathbb{N}}\left\\{T_{n}\frac{dP}{dQ}(X_{n})\right\\}\implies X_{N}\sim Q,$ (5) Therefore, PFR casts the REC problem into an optimisation, or search, problem, which can be solved in finite time almost surely. The PFR encoder draws pairs of samples $T_{n},X_{n}$, until it solves the search problem in eq. 5, and returns $X=X_{N},C=N-1$. The decoder can recover $X_{N}$ from $(P,C,S)$, by drawing $N$ samples from $P$, using the same random seed, and keeping the last sample. While, like the algorithm of Harsha et al. (2007), PFR is elegantly simple and achieves optimal codelengths, its expected runtime is also $2^{D_{\infty}[Q\|P]}$ (Maddison, 2016). #### Fast REC requires additional assumptions. These algorithms’ slow runtimes are perhaps unsurprising considering Agustsson & Theis’s result, which shows under the computational hardness assumption $\mathrm{RP}\neq\mathrm{NP}$ that without making additional assumptions on $Q$ and $P$, there is no REC algorithm whose expected runtime scales polynomially in $D_{\mathrm{KL}}[Q\|P]$. Therefore, in order achieve faster runtimes, a REC algorithm must make additional assumptions on $Q$ and $P$. #### A∗ coding. To this end, Flamich et al. (2022) proposed: (1) a set of appropriate assumptions which are satisfied by many deep latent variable models in practice and (2) a REC algorithm, referred to as A∗ coding, which leverages these assumptions to achieve a substantial speed-up over existing methods. In particular, A∗ coding generalizes PFR by introducing a partitioning scheme, which splits the sample space $\mathcal{X}$ in nested partitioning subsets, to speed up the solution of eq. 5. Drawing inspiration from this, our proposed algorithm generalises eqs. 2, 3 and 4 in an analogous manner (see fig. 1), introducing partitioning processes (definition 2) to speed up the algorithm’s termination. ###### Definition 2 (Partitioning process). A partitioning process is a process $Z:\mathbb{N}^{+}\to\Sigma$ such that $Z_{1}=\mathcal{X},~{}~{}Z_{2n}\cap Z_{2n+1}=\emptyset,~{}~{}Z_{2n}\cup Z_{2n+1}=Z_{n}.$ (6) In other words, a partitioning process $Z$ is a process indexed by the heap indices of an infinite binary tree, where the root node is $\mathcal{X}$ and any two children nodes $Z_{2n},Z_{2n+1}$ partition their parent node $Z_{n}$. In section 3 we present specific choices of partitioning processes which dramatically speed up GRC. #### Greedy Poisson Rejection Sampling. Contemporary to our work, Flamich (2023) introduces a rejection sampler based on Poisson processes, which can be used as a REC algorithm referred to as Greedy Poisson Rejection Sampling (GPRS). Similar to GRC and A* coding, GPRS partitions the sample space to speed up the convergence to the accepted sample. Furthermore, a variant of GPRS also achieves order-optimal runtime for one-dimensional distribution pairs with a unimodal density ratio. However, the construction of their method is significantly different from ours, relying entirely on Poisson processes. Moreover, GPRS requires numerically solving a certain ODE, while our method does not, making it potentially more favourable in practice. We believe establishing a closer connection between GPRS and GRC is a promising future research direction. ## 3 Greedy Rejection Coding #### Generalising Harsha et al. (2007). In this section we introduce Greedy Rejection Coding (GRC; definition 5), which generalises the algorithm of Harsha et al. (2007) in two ways. First, GRC can be used with distributions over arbitrary probability spaces. Therefore, it is applicable to arbitrary REC problems, including REC with continuous distributions. Second, similar to A∗ coding, GRC can be combined with arbitrary partitioning processes, allowing it to achieve optimal runtimes given additional assumptions on the REC problem, and an appropriate choice of partitioning process. (a) Sample & accept or reject (b) Partition & sample $b_{1}\in\\{{\color[rgb]{0.6875,0.4453125,0.69140625}\definecolor[named]{pgfstrokecolor}{rgb}{0.6875,0.4453125,0.69140625}\mathbf{0}},{\color[rgb]{0.96875,0.57421875,0.11328125}\definecolor[named]{pgfstrokecolor}{rgb}{0.96875,0.57421875,0.11328125}\mathbf{1}}\\}$ (c) Sample & accept or reject (d) Sample & accept or reject (e) Partition & sample $b_{1}\in\\{{\color[rgb]{0.6875,0.4453125,0.69140625}\definecolor[named]{pgfstrokecolor}{rgb}{0.6875,0.4453125,0.69140625}\mathbf{0}},{\color[rgb]{0.96875,0.57421875,0.11328125}\definecolor[named]{pgfstrokecolor}{rgb}{0.96875,0.57421875,0.11328125}\mathbf{1}}\\}$ (f) Sample & accept or reject Figure 3: Illustrations of the two variants of GRC considered in this work. (a) to (c) show GRC with the sample-splitting partitioning process (GRCS). (d) to (f) show GRC with the dyadic partition process (GRCD). GRC interleaves accept-reject steps with partitioning steps. In the former, it draws a sample and either accepts or rejects it. In the latter, it partitions the sample space and randomly chooses one of the partitions, ruling out large parts of the sample space and speeding up termination. ### 3.1 Algorithm definition #### Overview. Before specifying GRC, we summarise its operation with an accompanying illustration. On a high level, GRC interleaves accept-reject steps with partitioning steps, where the latter are determined by a partitioning process. Specifically, consider the example in figs. 3(d), 3(e) and 3(f), where $Q$ and $P$ are distributions over $\mathcal{X}=[0,1]$, and $Z$ is the partitioning process defined by $Z_{n}=[L,R]\implies Z_{2n}=[L,M),Z_{2n+1}=[M,R],\text{ where }M=(L+R)/2.$ (7) In each step $d=1,2,\dots$, GRC maintains a heap index $I_{d}$ of an infinite binary tree, and an active subset $S_{d}=Z_{I_{d}}\subseteq\mathcal{X}$ of the sample space, initialised as $I_{0}=1$ and $S_{1}=Z_{1}=\mathcal{X}$ respectively. #### Accept-reject step. In each step, GRC draws a sample from the restriction of $P$ to $S_{d}$, namely $P|_{S_{d}}/P(S_{d})$, and either accepts or rejects it. If the sample is accepted, the algorithm terminates. Otherwise, GRC performs a partitioning step as shown in fig. 3(d) #### Partitioning step. In each partitioning step, GRC partitions $S_{d}=Z_{I_{d}}$ into $Z_{2I_{d}}$ and $Z_{2I_{d}+1}$, as specified by the partitioning process $Z$. It then samples a Bernoulli random variable $b_{d}$, whose outcomes have probabilities proportional to the mass of $Q$ which has not been accounted for, up to and including step $d$, within the partitions $Z_{2I_{d}}$ and $Z_{2I_{d}+1}$ respectively. In fig. 3(e), these two masses correspond to the purple and orange areas, and the algorithm has sampled $b_{d}=1$. Last, GRC updates the heap index to $I_{d+1}=2I_{d}+b_{d}$ and the active subset to $S_{d+1}=Z_{I_{d+1}}$. GRC proceeds by interleaving accept-reject and partitioning steps until an acceptance occurs. #### Algorithm specification. The aforementioned algorithm can be formalised in terms of probability measures over arbitrary spaces and arbitrary partitioning processes. Above, algorithms 1 and 2 describe Harsha et al.’s rejection sampler and our generalisation of it, respectively. For the sake of keeping the exposition lightweight, we defer the formal measure-theoretic definition of GRC to the appendix (see definition 5 in section A.1), and refer to algorithm 2 as a working definition here. Algorithm 1 Harsha et al.’s rejection algorithm; equivalent to GRC with a global partition 1:Target $Q$, proposal $P$, space $\mathcal{X}$ 2:$d\leftarrow 0,T_{0}\leftarrow 0$ 3: 4:while True do 5: $X_{d+1}\sim P$ 6: $U_{d+1}\sim\text{Uniform}(0,1)$ 7: $\beta_{d+1}\leftarrow\texttt{AcceptProb}(Q,P,X_{d+1},T_{d})$ 8: if $U_{d+1}\leq\beta_{d+1}$ then 9: return $X_{d+1},d$ 10: end if 11: 12: 13: 14: $T_{d+1}\leftarrow\texttt{RuledOutMass}(Q,P,T_{d})$ 15: $d\leftarrow d+1$ 16:end while Algorithm 2 GRC with partition process $Z$; differences to Harsha et al.’s algorithm shown in green 1:Target $Q$, proposal $P$, space $\mathcal{X}$, partition $Z$ 2:$d\leftarrow 0,T_{0}\leftarrow 0$ 3:${\color[rgb]{0.234375,0.5,0.19140625}\definecolor[named]{pgfstrokecolor}{rgb}{0.234375,0.5,0.19140625}I_{0}\leftarrow 1,S_{1}\leftarrow\mathcal{X}}$ 4:while $\mathtt{True}$ do 5: $X_{I_{d}}\sim{\color[rgb]{0.234375,0.5,0.19140625}\definecolor[named]{pgfstrokecolor}{rgb}{0.234375,0.5,0.19140625}P|_{S_{d}}/P(S_{d})}$ 6: $U_{I_{d}}\sim\text{Uniform}(0,1)$ 7: $\beta_{I_{d}}\leftarrow\texttt{AcceptProb}(Q,P,X_{I_{d}},T_{d})$ 8: if $U_{I_{d}}\leq\beta_{d+1}$ or $d=D_{\max}$ then 9: return $X_{I_{d}},I_{d}$ 10: end if 11: $p\leftarrow\texttt{PartitionProb}(Q,P,T_{d},Z_{2d},Z_{2d+1})$ 12: $b_{d}\sim\text{Bernoulli}(p)$ 13: $I_{d+1}\leftarrow 2I_{d}+b_{d}$ and $S_{d+1}\leftarrow Z_{I_{d+1}}$ 14: $T_{d+1}\leftarrow\texttt{RuledOutMass}(Q,P,T_{d},{\color[rgb]{0.234375,0.5,0.19140625}\definecolor[named]{pgfstrokecolor}{rgb}{0.234375,0.5,0.19140625}S_{d+1}})$ 15: $d\leftarrow d+1$ 16:end while #### Comparison to Harsha et al. While algorithms 1 and 2 are similar, they differ in two notable ways. First, rather than drawing a sample from $P$, GRC draws a sample from the restriction of $P$ to an active subset $S_{d}=Z_{d}\subseteq\mathcal{X}$, namely $P|_{S_{d}}/P(S_{d})$. Second, GRC updates its active subset $S_{d}=Z_{d}$ at each step, setting it to one of the children of $Z_{d}$, namely either $Z_{2d}$ or $Z_{2d+1}$, by drawing $b_{d}\sim\text{Bernoulli}$, and setting $Z_{2d+b_{d}}$. This partitioning mechanism, which does not appear in algorithm 1, yields a different variant of GRC for each choice of partitioning process $Z$. In fact, as shown in Proposition 1 below, algorithm 1 is a special case of GRC with $S_{d}=\mathcal{X}$ for all $d$. See section A.2 for the proof. ###### Proposition 1 (Harsha et al. (2007) is a special case of GRC). Let $Z$ be the global partitioning process over $\Sigma$, defined as $Z_{1}=\mathcal{X},~{}~{}~{}Z_{2n}=Z_{n},~{}~{}~{}Z_{2n+1}=\emptyset,~{}~{}\text{ for all }~{}n=1,2,\dots.$ (8) Harsha et al. (2007) is equivalent to GRC using this $Z$ and setting $C=D^{*}$ instead of $C=I_{D^{*}}$. We refer to this algorithm as Global GRC, or GRCG for short. #### Partitioning processes and additional assumptions. While Proposition 1 shows that Harsha et al.’s algorithm is equivalent to GRC with a particular choice of $Z$, a range of other choices of $Z$ is possible, and this is where we can leverage additional structure. In particular, we show that when $Q$ and $P$ are continuous distributions over $\mathbb{R}$ with a unimodal density ratio $dQ/dP$, we can dramatically speed up GRC with an appropriate choice of $Z$. In particular, we will consider the sample- splitting and dyadic partitioning processes from Flamich et al. (2022), given in Definitions 3 and 4. ###### Definition 3 (Sample-splitting partitioning process). Let $\mathcal{X}=\mathbb{R}\cup\\{-\infty,\infty\\}$ and $P$ a continuous distribution. The sample-splitting partitioning process is defined as $Z_{n}=[a,b],a,b\in\mathcal{X}\implies Z_{2n}=[a,X_{n}],~{}~{}Z_{2n+1}=[X_{n},b],\text{ where }X_{n}\sim P|_{Z_{n}}/P(Z_{n}).$ In other words, in the sample-splitting process, $Z_{n}$ are intervals of $\mathbb{R}$, each of which is partitioned into sub-intervals $Z_{2n}$ and $Z_{2n+1}$ by splitting at the sample $X_{n}$ drawn from $P|_{Z_{n}}/P(Z_{n})$. We refer to GRC with the sample-splitting partitioning process as GRCS. ###### Definition 4 (Dyadic partitioning process). Let $\mathcal{X}=\mathbb{R}\cup\\{-\infty,\infty\\}$ and $P$ a continuous distribution. The dyadic partitioning process is defined as $Z_{n}=[a,b],a,b\in\mathcal{X}\implies Z_{2n}=[a,c],~{}~{}Z_{2n+1}=[c,b],\text{ such that }P(Z_{2n})=P(Z_{2n+1}).$ Similar to the sample-splitting process, in the dyadic process $Z_{n}$ are intervals of $\mathbb{R}$. However, in the dyadic process, $Z_{n}$ is partitioned into sub-intervals $Z_{2n}$ and $Z_{2n+1}$ such that $P(Z_{2n})=P(Z_{2n+1})$. We refer to GRC with the dyadic partitioning process as GRCD. #### GRC with a tunable codelength. Flamich et al. presented a depth-limited variant of AD∗ coding, DAD∗ coding, in which the codelength $|C|$ can be provided as a tunable input to the algorithm. Fixed-codelength REC algorithms are typically approximate because they introduce bias in their samples, but are nevertheless useful in certain contexts, such as for coding a group of random variables with the same fixed codelength. GRCD can be similarly modified to accept $|C|$ as an input, by limiting the maximum steps of the algorithm by $D_{\max}$ (see algorithm 2). Setting $D_{\max}=\infty$ in algorithm 2 corresponds to exact GRC, while setting $D_{\max}<\infty$ corresponds to depth-limited GRC. ### 3.2 Theoretical results #### Correctness of GRC. In theorem 1 we show that GRC terminates almost surely and produces unbiased samples from $Q$, given interchangeable mild assumptions on $Q,P$ and $Z$. Assumption 1 is the most general, since it holds for any $Q$ and $P$ over arbitrary probability spaces, and can be used to apply GRC to arbitrary coding settings. ###### Assumption 1. GRC has a finite ratio mode if $dQ/dP(x)<M$ for all $x\in\mathcal{X}$, for some $M\in\mathbb{R}$. Assumption 1 holds for GRCG, GRCS and GRCD, so long as $dQ/dP$ is bounded. While this assumption is very general, in some cases we may want to consider $Q,P$ with unbounded $dQ/dP$. To this end, we show that it can be replaced by alternative assumptions, such as assumptions 2 and 3. ###### Assumption 2. GRC is single-branch if for each $d$, $b_{d}=0$ or $b_{d}=1$ almost surely. GRC with the global partitioning process (eq. 8) satisfies assumption 2. In addition, if $Q$ and $P$ are distributions over $\mathbb{R}$ and $dQ/dP$ is unimodal, GRCS also satisfies assumption 2. ###### Assumption 3. Suppose $\mathcal{X}\subseteq\mathbb{R}^{N}$. GRC has nicely shrinking $Z$ if, almost surely, the following holds. For each $x\in\mathcal{X}$ which is in a nested sequence of partitions $x\in Z_{1}\supseteq\dots\supseteq Z_{k_{d}}\supseteq\dots$ with $P(Z_{k_{d}})\to 0$, there exist $\gamma,r_{1},r_{2},...\in\mathbb{R}_{>0}$ such that $r_{d}\to 0,~{}Z_{k_{d}}\subseteq B_{r_{d}}(x)\text{ and }P(Z_{k_{d}})\geq\gamma P(B_{r_{d}}(x)).$ (9) If $Q$ and $P$ are distributions over $\mathbb{R}$, GRCD satisfies assumption 3. Theorem 1 shows that if any of the above assumptions hold, then GRC terminates almost surely and yields unbiased samples from $Q$. We provide the proof of the theorem in appendix B. ###### Theorem 1 (Correctness of GRC). Suppose $Q,P$ and $Z$ satisfy any one of assumptions 1, 2 and 3. Then, algorithm 2 terminates with probability $1$, and its returned sample $X$ has law $X\sim Q$. #### Expected runtime and codelength of GRCS. Now we turn to the expected runtime and codelength of GRCS. Theorem 2 shows that the expected codelength of GRCS is optimal, while Theorem 3 establishes that its runtime is order-optimal. We present the proofs of the theorems in appendix C. ###### Theorem 2 (GRCS codelength). Let $Q$ and $P$ be continuous distributions over $\mathbb{R}$ such that $Q\ll P$ and with unimodal $dQ/dP$. Let $Z$ be the sample-splitting process, and $X$ its returned sample. Then, $\mathbb{H}[X|Z]\leq D_{\mathrm{KL}}[Q\|P]+2\log\left(D_{\mathrm{KL}}[Q\|P]+1\right)+\mathcal{O}(1).$ (10) ###### Theorem 3 (GRCS runtime). Let $Q$ and $P$ be continuous distributions over $\mathbb{R}$ such that $Q\ll P$ and with unimodal $dQ/dP$. Let $Z$ be the sample-splitting process and $D$ the number of steps the algorithm takes before accepting a sample. Then, for $\beta=2/\log(4/3)\approx 4.82$ we have $\mathbb{E}[D]\leq\beta~{}D_{\mathrm{KL}}[Q\|P]+\mathcal{O}(1)$ (11) #### Improving the codelength of GRCD. In Theorem 2 we state the bound for the REC setting, where we make no further assumptions on $Q$ and $P$. However, we can improve the bound if we consider the reverse channel coding (RCC) setting (Theis & Yosri, 2022). In RCC, we have a pair of correlated random random variables $X,Y\sim P_{X,Y}$. During one round of communication, the encoder receives $Y\sim P_{Y}$ and needs to encode a sample $X\sim P_{X|Y}$ from the posterior using $P_{X}$ as the proposal distribution. Thus, RCC can be thought of as the average-case version of REC, where the encoder sets $Q\leftarrow P_{X|Y}$ and $P\leftarrow P_{X}$. In this case, when the conditions of Theorem 2 hold for every $(P_{X|Y},P_{X})$ pair, in appendix C we show that the bound can be improved to $\mathbb{I}[X;Y]+2\log(\mathbb{I}[X;Y]+1)+\mathcal{O}(1)$, where $\mathbb{I}[X;Y]=\mathbb{E}_{Y\sim P_{Y}}\left[D_{\mathrm{KL}}[P_{X|Y}\|P_{Y}]\right]$ is the mutual information between $X$ and $Y$. #### GRCS runtime is order-optimal. Theorem 3 substantially improves upon the runtime of A∗ coding, which is the current fastest REC algorithm with similar assumptions. In particular, AS∗ coding has $\mathcal{O}(D_{\infty}[Q\|P])$ expected runtime, which can be arbitrarily larger than that of GRCS. Remarkably, the runtime of GRCS is optimal up to the multiplicative factor $\beta$. This term arises from the fact the sample-splitting process may occasionally rule out a small part of the sample space at a given step. ## 4 Experiments We conducted two sets of experiments: one on controlled synthetic REC problems to check the predictions of our theorems numerically, and another using VAEs trained on MNIST to study how the performance of GRC-based compression pipelines can be improved in practice. We conducted all our experiments under fair and reproducible conditions and make our source code public.111Source code to be published with the camera-ready version: https://github.com/source- code. ### 4.1 Synthetic Experiments #### Synthetic REC experiments. First, we compare GRCS and GRCD, against AS∗ and AD∗ coding, on a range of synthetic REC problems. We systematically vary distribution parameters to adjust the difficulty of the REC problems. Figure 4 shows the results of our synthetic experiments. Figure 4: Comparison between GRC and A∗ coding on synthetic REC problems with Gaussian $Q$ and $P$. Left: we fix $D_{\mathrm{KL}}[Q\|P]=3$ and vary $D_{\infty}[Q\|P]$, measuring the number of steps taken by each algorithm. Right: we fix $D_{\infty}[Q\|P]=D_{\mathrm{KL}}[Q\|P]+2$ and vary $D_{\mathrm{KL}}[Q\|P]$, plotting the codelengths produced by each algorithm. Reported codelengths do not include additional logarithmic overhead terms. Results are averaged over $4\times 10^{3}$ different random seeds for each datapoint. We have included error-bars in both plots but these are too small to see compared to the plot scales. #### Partitioning processes improve the runtime of GRC. First, we observe that, assuming that $dQ/dP$ is unimodal, introducing an appropriate partitioning process such as the sample-splitting or the dyadic process, dramatically speeds up GRC. In particular, fig. 4 shows that increasing the infinity divergence $D_{\infty}[Q\|P]$ (for a fixed $D_{\mathrm{KL}}[Q\|P]$) does not affect the runtimes of GRCS and GRCD, which remain constant and small. This is a remarkable speed-up over the exponential expected runtime of GRCG. #### GRC is faster than A∗ coding. Further, we observe that GRC significantly improves upon the runtime of A* coding, which is the fastest previously known algorithm with similar assumptions. In particular, Figure 4 shows that increasing the infinity divergence $D_{\infty}[Q\|P]$, while keeping the KL divergence $D_{\mathrm{KL}}[Q\|P]$ fixed, increases the runtime of both AS∗ and AD∗ coding, while the runtimes of GRCS and GRCD remain constant. More generally, for a fixed KL divergence, the infinity divergence can be arbitrarily large or even infinite. In such cases, A∗ coding would be impractically slow or even inapplicable, while GRCS and GRCD remain practically fast. #### GRCD improves on GRCS. In our experiments, we observe that the performance of GRCD (green in fig. 4) matches that of GRCS (blue in fig. 4) in terms of runtime and codelength. While in our experiments, GRCD does not yield an improvement over GRCS, we note the following behaviour. The sample-splitting process may occasionally rule out a only a small part of space, which can slow down convergence. In particular, in appendix C we show that on average, the sample-splitting process rules out $\nicefrac{{1}}{{2}}$ of the active sample space in the best case at each step, and $\nicefrac{{3}}{{4}}$ in the worst case. By contrast, the dyadic process always rules out $\nicefrac{{1}}{{2}}$ of the sample space, potentially speeding up termination. We conjecture that GRCD achieves an optimal expected runtime with $\beta=1$. ### 4.2 Compression with Variational Autoencoders #### Compressing images with VAEs and REC. One of the most promising applications of REC is in learnt compression. Here, we implement a proof-of-concept lossless neural compression pipeline using a VAE with a factorized Gaussian posterior on MNIST and take the architecture used by Townsend et al. (2018). To compress an image $Y$, we encode a latent sample $X$ from the VAE posterior $q(X\mid Y)$ by applying GRCD dimensionwise after which we encode the image $Y$ with entropy coding using the VAE’s conditional likelihood $p(Y\mid X)$ as the coding distribution. Unfortunately, in addition to the $D_{\mathrm{KL}}[q(X_{d}\mid Y)\|p(X_{d})]$ bits coding cost for latent dimension $d$, this incurs an overhead of ${\log(D_{\mathrm{KL}}[q(X_{d}\mid Y)\|p(X_{d})]+1)+\mathcal{O}(1)}$ bits, analogously to how a symbol code, like Huffman coding, incurs a constant overhead per symbol (MacKay, 2003). However, since $\log(1+x)\approx x$ when $x\approx 0$, the logarithmic overhead of GRC can become significant compared to the KL divergence. Hence, we now investigate two approaches to mitigate this issue. Training objective | # latent | Total BPP with $\zeta$ coding | Total BPP with $\delta$ coding | Neg. ELBO per pixel | Overhead BPP with $\delta$ coding ---|---|---|---|---|--- ELBO | 20 | $1.472\pm 0.004$ | $1.482\pm 0.004$ | $1.391\pm 0.004$ | $0.091\pm 0.000$ 50 | $1.511\pm 0.003$ | $1.530\pm 0.003$ | $1.357\pm 0.003$ | $0.172\pm 0.000$ 100 | $1.523\pm 0.003$ | $1.600\pm 0.003$ | $1.362\pm 0.003$ | $0.238\pm 0.000$ Modified ELBO | 20 | $1.470\pm 0.004$ | $1.478\pm 0.004$ | $1.393\pm 0.004$ | $0.085\pm 0.000$ 50 | $1.484\pm 0.003$ | $1.514\pm 0.003$ | $1.373\pm 0.003$ | $0.141\pm 0.000$ 100 | $1.485\pm 0.003$ | $1.579\pm 0.003$ | $1.373\pm 0.003$ | $0.205\pm 0.000$ Table 1: Lossless compression performance comparison on the MNIST test set of a small VAE with different latent space sizes, optimized using either the ELBO or the modified ELBO in eq. 12. We report the bits per pixel (BPP) attained using different coding methods, averaged over the 10,000 test images, along with the standard error, using GRCD. See section 4.2 for further details. #### Modified ELBO for REC. A principled approach to optimizing our neural compression pipeline is to minimize its expected codelength. For bits-back methods (Townsend et al., 2018, 2019), the negative ELBO indeed expresses their expected codelength, but in REC’s case, it does not take into account the additional dimensionwise logarithmic overhead we discussed above. Thus, we propose to minimize a modified negative ELBO to account for this (assuming that we have $D$ latent dimensions): $\displaystyle\underbrace{\mathbb{E}_{X\sim q(X|Y)}[-\log p(Y|X)]+D_{\mathrm{KL}}[q(X|Y)\|p(X)]}_{\text{Regular ELBO}}+\sum_{d=1}^{D}\underbrace{\log\left(D_{\mathrm{KL}}[q(X_{d}|Y)\|p(X_{d})]+1\right)}_{\text{Logarithmic overhead per dimension}}.$ (12) #### Coding the latent indices. As the final step during the encoding process, we need a prefix code to encode the heap indices $I_{d}$ returned by GRCD for each $d$. Without any further information, the best we can do is use Elias $\delta$ coding (Elias, 1975), which, assuming our conjecture on the expected runtime of GRCD holds, yields an expected codelength of $\mathbb{I}[Y;X]+2\log(\mathbb{I}[Y;X]+1)+\mathcal{O}(1)$. However, we can improve this if we can estimate $\mathbb{E}[\log I_{d}]$ for each $d$: it can be shown, that the maximum entropy distribution of a positive integer-valued random variable with under a constraint on the expectation on its logarithm is $\zeta(n|\lambda)\propto n^{-\lambda}$, with $\lambda^{-1}=\mathbb{E}[\log I_{d}]+1$. In this case, entropy coding $I_{d}$ using this $\zeta$ distribution yields improves the expected codelength to $\mathbb{I}[Y;X]+\log(\mathbb{I}[Y;X]+1)+\mathcal{O}(1)$. #### Experimental results. We trained our VAE with $L\in\\{20,50,100\\}$ latent dimensions optimized using the negative ELBO and its modified version in Equation 12, and experimented with encoding the heap indices of GRCD with both $\delta$ and $\zeta$ coding. We report the results of our in Table 1 on the MNIST test set in bits per pixel. In addition to the total coding cost, we report the negative ELBO per pixel, which is the fundamental lower bound on the compression efficiency of REC with each VAE. Finally, we report the logarithmic overhead due to $\delta$ coding. We find that both the modified ELBO and $\zeta$ coding prove beneficial, especially as the dimensionality of the latent space increases. This is expected, since the overhead is most significant for latent dimensions with small KLs, which becomes more likely as the dimension of the latent space grows. The improvements yielded by each of the two methods are significant, with $\zeta$ coding leading to a consistent $1-7\%$ gain compared to $\delta$ coding and the modified objective resulting in up to $2\%$ gain in coding performance. ## 5 Conclusion and Future Work #### Summary. In this work, we introduced Greedy Rejection Coding (GRC), a REC algorithm which generalises the rejection algorithm of Harsha et al. to arbitrary probability spaces and partitioning processes. We proved the correctness of our algorithm under mild assumptions, and introduced GRCS and GRCD, two variants of GRC. We showed that the runtimes of GRCS and GRCD significantly improve upon the runtime of A∗ coding, which can be arbitrarily larger. We evaluated our algorithms empirically, verifying our theory and conducted a proof-of-concept learnt compression experiment on MNIST using VAEs. We demonstrated that a principled modification to the ELBO and entropy coding GRCD’s indices using a $\zeta$ distribution can further improve compression efficiency. #### Limitations and Further work. One limitation of GRC is that, unlike A∗ coding, it requires us to be able to evaluate the CDF of $Q$. While in some settings this CDF may be intractable, this assumption is satisfied by most latent variable generative models, and is not restrictive in practice. However, one practical limitation of GRCS and GRCD, as well as AS∗ and AD∗ , is that they assume target-proposal pairs over $\mathbb{R}$. For multivariate distributions, we can decompose them into univariate conditionals and apply GRC dimensionwise, however this incurs an additional coding overhead per dimension, resulting in a non-negligible cost. Thus, an important direction is to investigate whether fast REC algorithms for multivariate distributions can be devised, to circumvent this challenge. ## References * Agustsson & Theis (2020) Eirikur Agustsson and Lucas Theis. Universally quantized neural compression. _Advances in Neural Information Processing Systems_ , 2020. * Ballé et al. (2020) Johannes Ballé, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. _IEEE Journal of Selected Topics in Signal Processing_ , 2020. * Child (2020) Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on images. _CoRR_ , abs/2011.10650, 2020. * Dudley (2018) Richard M Dudley. _Real analysis and probability_. CRC Press, 2018. * Dunford & Schwartz (1988) Nelson Dunford and Jacob T Schwartz. _Linear operators, part 1: general theory_ , volume 10. John Wiley & Sons, 1988. * Elias (1975) Peter Elias. Universal codeword sets and representations of the integers. _IEEE transactions on information theory_ , 1975. * Flamich (2023) Gergely Flamich. Greedy Poisson rejection sampling. _Advances in Neural Information Processing Systems_ , 2023. * Flamich & Theis (2023) Gergely Flamich and Lucas Theis. Adaptive greedy rejection sampling. _arXiv preprint arXiv:2304.10407_ , 2023. * Flamich et al. (2020) Gergely Flamich, Marton Havasi, and José Miguel Hernández-Lobato. Compressing images by encoding their latent representations with relative entropy coding. _Advances in Neural Information Processing Systems_ , 2020. * Flamich et al. (2022) Gergely Flamich, Stratis Markou, and Jose Miguel Hernandez-Lobato. Fast relative entropy coding with A* coding. In _Proceedings of the 39th International Conference on Machine Learning_ , Proceedings of Machine Learning Research. PMLR, 2022. * Harsha et al. (2007) Prahladh Harsha, Rahul Jain, David McAllester, and Jaikumar Radhakrishnan. The communication complexity of correlation. In _Twenty-Second Annual IEEE Conference on Computational Complexity (CCC’07)_. IEEE, 2007. * Havasi et al. (2018) Marton Havasi, Robert Peharz, and José Miguel Hernández-Lobato. Minimal random code learning: Getting bits back from compressed model parameters. In _International Conference on Learning Representations_ , 2018. * Ho et al. (2020) Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. _CoRR_ , 2020. * Kingma et al. (2016) Durk P Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling. Improved variational inference with inverse autoregressive flow. _Advances in neural information processing systems_ , 2016. * Kingman (1992) J.F.C. Kingman. _Poisson Processes_. Oxford Studies in Probability. Clarendon Press, 1992. * Li & El Gamal (2018) Cheuk Ting Li and Abbas El Gamal. Strong functional representation lemma and applications to coding theorems. _IEEE Transactions on Information Theory_ , 2018. * MacKay (2003) David JC MacKay. _Information theory, inference and learning algorithms_. Cambridge university press, 2003. * Maddison (2016) CA Maddison. Poisson process model for Monte Carlo. _Perturbation, Optimization, and Statistics_ , 2016. * Maddison et al. (2014) Chris J Maddison, Daniel Tarlow, and Tom Minka. A* sampling. _Advances in Neural Information Processing Systems_ , 2014. * Mentzer et al. (2020) Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. _Advances in Neural Information Processing Systems_ , 2020. * Mentzer et al. (2022) Fabian Mentzer, George Toderici, David Minnen, Sergi Caelles, Sung Jin Hwang, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. In _Advances in Neural Information Processing Systems_ , 2022. * Pérez-Cruz (2008) Fernando Pérez-Cruz. Kullback-Leibler divergence estimation of continuous distributions. In _2008 IEEE international symposium on information theory_ , 2008\. * Rudin (1986) Walter Rudin. _Real and Complex Analysis_. McGraw-Hill, 1986. * Shah et al. (2022) Abhin Shah, Wei-Ning Chen, Johannes Balle, Peter Kairouz, and Lucas Theis. Optimal compression of locally differentially private mechanisms. In _International Conference on Artificial Intelligence and Statistics_. PMLR, 2022. * Theis & Agustsson (2021) L. Theis and E. Agustsson. On the advantages of stochastic encoders. In _Neural Compression Workshop at ICLR_ , 2021. * (26) Lucas Theis and Eirikur Agustsson. On the advantages of stochastic encoders. In _Neural Compression: From Information Theory to Applications–Workshop@ ICLR 2021_. * Theis & Yosri (2022) Lucas Theis and Noureldin Yosri. Algorithms for the communication of samples. In _International Conference on Machine Learning_ , 2022. * Theis et al. (2022) Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. _arXiv preprint arXiv:2206.08889_ , 2022. * Townsend et al. (2018) James Townsend, Thomas Bird, and David Barber. Practical lossless compression with latent variables using bits back coding. In _International Conference on Learning Representations_ , 2018. * Townsend et al. (2019) James Townsend, Thomas Bird, Julius Kunze, and David Barber. Hilloc: lossless image compression with hierarchical latent variable models. In _International Conference on Learning Representations_ , 2019. * Vahdat & Kautz (2020) Arash Vahdat and Jan Kautz. Nvae: A deep hierarchical variational autoencoder. _Advances in neural information processing systems_ , 2020. * (32) Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. _CoRR_. * Ziv (1985) Jacob Ziv. On universal quantization. _IEEE Transactions on Information Theory_ , 1985. ## Appendix A Formal definition of Greedy Rejection Coding ### A.1 Formal definition Here we give a formal definition of GRC in terms of measures. We chose to omit this from the main text for the sake of exposition, and instead formally define GRC in definition 5 below. ###### Definition 5 (Greedy Rejection Coding). Let $Z$ be a partitioning process on $\Sigma$, and $I_{0}=1$, $S_{0}=Z_{I_{0}}$. Let $T_{0}(\cdot,S_{0})$ be the zero-measure on $(\mathcal{X},\Sigma)$. Then for $d=0,1,\dots$ define $\displaystyle t_{d}(x,S_{0:d})$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}\frac{dT_{d}(\cdot,S_{0:d})}{dP(\cdot)}(x),$ (13) $\displaystyle\alpha_{d+1}(x,S_{0:d})$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}\min\left\\{\frac{dQ}{dP}(x)-t_{d}(x,S_{0:d}),\frac{1-T_{d}(\mathcal{X},S_{0:d})}{P(S_{d})}\right\\}$ (14) $\displaystyle A_{d+1}(S,S_{0:d})$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}\int_{S}dP(x)~{}\alpha_{d+1}(x,S_{0:d}),$ (15) $\displaystyle\beta_{d+1}(x,S_{0:d})$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}\alpha_{d+1}(x,S_{0:d})~{}\frac{P(S_{d})}{1-T_{d}(\mathcal{X},S_{0:d})},$ (16) $\displaystyle X_{I_{d}}$ $\displaystyle\sim\frac{P|_{S_{d}}}{P(S_{d})},$ (17) $\displaystyle U_{I_{d}}$ $\displaystyle\sim\text{Uniform}(0,1),$ (18) $\displaystyle b_{d}$ $\displaystyle\sim\text{Bernoulli}\left(\frac{Q(Z_{2I_{d}+1})-T_{d}(Z_{2I_{d}+1},S_{0:d})-A_{d+1}(Z_{2I_{d}+1},S_{0:d})}{Q(S_{d})-T_{d}(S_{d},S_{0:d})-A_{d+1}(S_{d},S_{0:d})}\right),$ (19) $\displaystyle I_{d+1}$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}2I_{d}+b_{d},$ (20) $\displaystyle S_{d+1}$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}Z_{I_{d+1}},$ (21) $\displaystyle T_{d+1}(S,S_{0:d+1})$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}T_{d}(S\cap S_{d+1},S_{0:d})+A_{d+1}(S\cap S_{d+1},S_{0:d})+Q(S\cap S_{d+1}^{\prime}),$ (22) where $S\in\Sigma$ and $P|_{Z_{d}}$ denotes the restriction of the measure $P$ to the set $Z_{d}$. Generalised Greedy Rejection Coding (GRC) amounts to running this recursion, computing $\displaystyle D^{*}=\min\\{d\in\mathbb{N}:U_{I_{d}}\leq\beta_{d+1}(X_{I_{d}},S_{0:d})\\},$ (23) and returning $X=X_{I_{D^{*}}}$ and $C=I_{D^{*}}$. The functions AcceptProb and RuledOutMass in algorithm 2 correspond to calculating the quantities in eq. 16 and eq. 22. The function PartitionProb corresponds to computing the success probability of the Bernoulli coin toss in eq. 19. ### A.2 Harsha et al.’s algorithm is a special case of GRC Here we show that the algorithm of Harsha et al. is a special case of GRC which assumes discrete $P$ and $Q$ distributions and uses the global partitioning process, which we refer to as GRCG. Note that the original algorithm described by Harsha et al. assumes discrete $P$ and $Q$ distributions, whereas GRCG does not make this assumption. ###### Proposition 2 (Harsha et al. (2007) is a special case of GRC). Let $Z$ be the global partitioning process over $\Sigma$, defined as $Z_{1}=\mathcal{X},~{}~{}~{}Z_{2n}=Z_{n},~{}~{}~{}Z_{2n+1}=\emptyset,~{}~{}\text{ for all }~{}n=1,2,\dots.$ (24) Harsha et al. (2007) is equivalent to GRC using this $Z$ and setting $C=D^{*}$ instead of $C=I_{D^{*}}$. We refer to this variant of GRC as Global GRC, or GRCG for short. ###### Proof. With $Z$ defined as in eq. 24, we have $b_{d}\sim\text{Bernoulli}(0)$ by eq. 19, so $b_{d}=0$ almost surely. Therefore $S_{d}=\mathcal{X}$ for all $d\in\mathbb{N}^{+}$. From this, we have $T_{d+1}(S,S_{0:d})=T_{d}(S,S_{0:d})+A_{d}(S,S_{0:d})$ and also $P(S_{d})=P(\mathcal{X})=1$ for all $d\in\mathbb{N}^{+}$. Substituting these in the equations of definition 5, we recover eqs. 2, 3 and 4. Setting $C=D^{*}$ instead of $C=I_{D^{*}}$ makes the two algorithms identical. ∎ ## Appendix B Proof of correctness of GRC: Theorem 1 In this section we give a proof for the correctness of GRC. Before going into the proof, we outline our approach and the organisation of the proof. #### Proof outline. To prove theorem 1, we consider running GRC for a finite number of $d$ steps. We consider the measure $\tau_{d}:\Sigma\to[0,1]$, defined such that for any $S\in\Sigma$, the quantity $\tau_{d}(S)$ is equal to the probability that GRC terminates within $d$ steps and returns a sample $X\in S\subseteq\Sigma$. We then show that $\tau_{d}\to Q$ in total variation as $d\to\infty$, which proves theorem 1. #### Organisation of the proof. First, in section B.1 we introduce some preliminary definitions, assumptions and notation on partitioning processes, which we will use in later sections. Then, in B.2 we derive the $\tau_{d}$ measure, and prove some intermediate results about it. Specifically, proposition 3 shows that the measures $A_{d}$ and $T_{d}$ from the definition of GRC (definition 5) correspond to probabilities describing the termination of the algorithm, and lemma 1 uses these facts to derive the form of $\tau_{d}$ in terms of $A_{d}$. Then, lemma 2 shows that the measure $\tau_{d}$ is no larger than the measure $Q$ and lemma 3 shows that the limit of $\tau_{d}$ as $d\to\infty$ is also a measure. Lastly lemma 4 shows that $T_{d}$ and $\tau_{d}$ are equal on the active sets of the partition process followed within a run of GRC, and then lemma 5 uses that result to derive the subsets of the sample space on which $\tau_{d}$ is equal to $Q$ and $\tau$ is equal to $Q$. Then, in section B.3 we break down the proof of theorem 1 in four cases. First, we consider the probability $p_{d}$ that GRC terminates at step $d$, given that it has not terminated up to and including step $d-1$. Lemma 7 shows that if $p_{d}\not\to 0$, then $\tau_{d}\to Q$ in total variation. Then we consider the case $p_{d}\to 0$ and show that in this case, if any of assumptions 1, 2 or 3 hold, then again $\tau_{d}\to Q$ in total variation. Putting these results together proves theorem 1. ### B.1 Preliminary definitions, assumptions and notation For the sake of completeness, we restate relevant definitions and assumptions. Definition 6 restates our notation on the target $Q$ and proposal $P$ measures and assumption 4 emphasises our assumption that $Q\ll P$. Definition 7 restates the definition of partitioning processes. ###### Definition 6 (Target $Q$ and proposal $P$ distributions). Let $Q$ and $P$ be probability measures on a measurable space $(\mathcal{X},\Sigma)$. We refer to $Q$ and $P$ as the target and proposal measures respectively. ###### Assumption 4 ($Q\ll P$). We assume $Q$ is absolutely continuous w.r.t. $P$, that is $Q\ll P$. Under this assumption, the Radon-Nikodym derivative of $Q$ w.r.t. $P$ exists and is denoted as $dQ/dP:\mathcal{X}\to\mathbb{R}^{+}$. ###### Definition 7 (Partitioning process). A random process $Z:\mathbb{N}^{+}\to\Sigma$ which satisfies $Z_{1}=\mathcal{X},~{}~{}Z_{2n}\cap Z_{2n+1}=\emptyset,~{}~{}Z_{2n}\cup Z_{2n+1}=Z_{n}.$ (25) is called a partitioning process. That is, a partitioning process $Z$ is a random process indexed by the heap indices of an infinite binary tree, where the root node is $\mathcal{X}$ and any two children nodes $Z_{2n}$ and $Z_{2n+1}$ partition their parent node $Z_{n}$. Note that by definition, a partitioning process takes values which are measurable sets in $(\mathcal{X},\Sigma)$. Because GRC operates on an binary tree, we find it useful to define some appropriate notation. Definition 8 specifies the ancestors of a node in a binary tree. Notation 1 gives some useful indexing notation for denoting different elements of the partitioning process $Z$, as well as for denoting the branch of ancestors of an element in a partitioning process. ###### Definition 8 (Ancestors). We define the one-step ancestor function $A_{1}:2^{\mathbb{N}^{+}}\to 2^{\mathbb{N}^{+}}$ as $\displaystyle A_{1}(N)$ $\displaystyle=N\cup\\{n\in\mathbb{N}^{+}:n^{\prime}=2n\text{ or }n^{\prime}=2n+1,\text{ for some }n^{\prime}\in N\\},$ (26) and the ancestor function $A:2^{\mathbb{N}^{+}}\to 2^{\mathbb{N}^{+}}$ as $A(N)=\left\\{n\in\mathbb{N}^{+}:n\in A_{1}^{k}(\\{n^{\prime}\\})\text{ for some }n^{\prime}\in N,k\in\mathbb{N}^{+}\right\\}.$ (27) where $A_{1}^{k}$ denotes the composition of $A_{1}$ with itself $k$ times. Viewing $\mathbb{N}^{+}$ as the set of heap indices of an infinite binary tree, $A$ maps a set $N\subseteq\mathbb{N}$ of natural numbers (nodes) to the set of all elements of $N$ and their ancestors. ###### Notation 1 (Double indexing for $Z$, ancestor branch). Given a partitioning process $Z$, we use the notation $Z_{d,k}$, where $d=1,2,\dots$ and $k=1,\dots,2^{d-1}$ to denote the $k^{th}$ node at depth $d$, that is $Z_{d,k}:=Z_{2^{d-1}-1+k}.$ (28) We use the hat notation $\hat{Z}_{d,k}$ to denote the sequence of nodes consisting of $Z_{d,k}$ and all its ancestors $\hat{Z}_{d,k}:=(Z_{n}:n\in A(\\{2^{d-1}-1+k\\})),$ (29) and call $\hat{Z}_{d,k}$ the ancestor branch of $Z_{d,k}$. ###### Notation 2 ($\mathbb{P}$ measure). In definition 5, we defined $\mathbb{P}$ to be the measure associated with an infinite sequence of independent fair coin tosses over a measurable space $(\Omega,\mathcal{S})$. To avoid heavy notation, for the rest of the proof we will overload this symbol as follows: if $F$ is a random variable from $\Omega$ to some measurable space, we will abbreviate $\mathbb{P}\circ F^{-1}$ by simply $\mathbb{P}(F)$. ### B.2 Deriving the measure of samples returned by GRC For the remainder of the proof, we condition on a fixed partitioning process sample $Z$. For brevity, we omit this conditioning which, from here on is understood to be implied. Proposition 3 shows that the measures $A_{d}$ and $T_{d}$ correspond to the probabilities that GRC picks a particular branch of the binary tree and terminates at step $d$, or does not terminate up to and including step $d$, respectively. ###### Proposition 3 (Acceptance and rejection probabilities). Let $V_{d}$ be the event that GRC does not terminate up to and including step $d$ and $W_{d}$ be the event that it terminates at step $d$. Let $S_{0:d}=B_{0:d}$ denote the event that the sequence of the first $d$ bounds produced is $B_{0:d}$. Then $\displaystyle\mathbb{P}(V_{d},S_{0:d}=B_{0:d})$ $\displaystyle=1-T_{d}(\mathcal{X},B_{0:d}),$ $\displaystyle\text{ for }d=0,1,\dots,$ (30) $\displaystyle\mathbb{P}(W_{d+1},S_{0:d}=B_{0:d})$ $\displaystyle=A_{d+1}(\mathcal{X},B_{0:d}),$ $\displaystyle\text{ for }d=0,1,\dots.$ (31) ###### Proof. First we consider the probability that GRC terminates at step $k+1$ given that it has not terminated up to and including step $d$, that is the quantity $\mathbb{P}(W_{k+1}~{}|~{}V_{k},S_{0:k}=B_{0:k})$. By definition 5, this probability is given by integrating the acceptance probability $\beta_{k+1}(x,B_{0:k})$ over $x\in\mathcal{X}$, with respect to the measure $P|_{B_{k}}/P(B_{k})$, that is $\displaystyle\mathbb{P}(W_{k+1}~{}|~{}V_{k},S_{0:k}=B_{0:k})$ $\displaystyle=\int_{x\in B_{k}}dP(x)\frac{\beta_{k+1}(x,B_{0:k})}{P(B_{k})}$ (32) $\displaystyle=\int_{x\in\mathcal{X}}dP(x)\frac{\beta_{k+1}(x,B_{0:k})}{P(B_{k})}$ (33) $\displaystyle=\int_{x\in\mathcal{X}}dP(x)\frac{\alpha_{k+1}(x,B_{0:k})}{1-T_{k}(\mathcal{X},B_{0:k})}$ (34) $\displaystyle=\frac{A_{k+1}(\mathcal{X},B_{0:k})}{1-T_{k}(\mathcal{X},B_{0:k})},$ (35) Now, we show the result by induction on $d$, starting from the base case of $d=0$. Base case: For $d=0$, by the definition of GRC (definition 5) $S_{0}=Z_{I_{0}}=\mathcal{X}$, so $\displaystyle\mathbb{P}\left(V_{0},S_{0}=B_{0}\right)=1~{}\text{ and }~{}T_{0}(\mathcal{X},B_{0})=0,$ (36) which show the base case for eq. 30. Now, plugging in $k=0$ in eq. 35 we obtain $\mathbb{P}(W_{1},S_{0}=B_{0})=\mathbb{P}(W_{1}~{}|~{}V_{0},S_{0}=B_{0})=\frac{A_{1}(\mathcal{X},B_{0})}{1-T_{0}(\mathcal{X},B_{0})}=A_{1}(\mathcal{X},B_{0})$ (37) where we have used the fact that $T_{0}(\mathcal{X},B_{0})=0$, showing the base case for eq. 31. Inductive step: Suppose that for all $k=0,1,2,\dots,d$ it holds that $\mathbb{P}\left(V_{d},S_{0:k}=B_{0:k}\right)=1-T_{d}(\mathcal{X},B_{0:k})~{}~{}\text{ and }~{}~{}\mathbb{P}\left(W_{k+1},S_{0:k}=B_{0:k}\right)=A_{k+1}(\mathcal{X},B_{0:k}).$ (38) Setting $k=d$ in eq. 35, we obtain $\displaystyle\mathbb{P}(W^{\prime}_{d+1}~{}|~{}V_{d},S_{0:d}=B_{0:d})$ $\displaystyle=\frac{1-T_{d}(\mathcal{X},B_{0:d})-A_{d+1}(\mathcal{X},B_{0:d})}{1-T_{d}(\mathcal{X}.B_{0:d})},$ (39) and using the inductive hypothesis from eq. 38, we have $\displaystyle\mathbb{P}(V_{d+1},S_{0:d}=B_{0:d})$ $\displaystyle=\mathbb{P}(W_{d+1}^{\prime},V_{d},S_{0:d}=B_{0:d})=1-T_{d}(\mathcal{X},B_{0:d})-A_{d+1}(\mathcal{X},B_{0:d}).$ (40) Now, $B_{d}=Z_{n}$ for some $n\in\mathbb{N}^{+}$. Denote $B_{d}^{L}:=Z_{2n}$ and $B_{d}^{R}:=Z_{2n+1}$. Then, by the product rule $\displaystyle\mathbb{P}(V_{d+1},S_{0:d}$ $\displaystyle=B_{0:d},S_{d+1}=B^{R}_{d})=$ (41) $\displaystyle=\mathbb{P}(S_{d+1}=B^{R}_{d}~{}|~{}V_{d+1},S_{0:d}=B_{0:d})\mathbb{P}(V_{d+1},S_{0:d}=B_{0:d})$ (42) $\displaystyle=\frac{Q(B_{d}^{R})-T_{d}(B_{d}^{R},B_{0:d})-A_{d+1}(B_{d}^{R},B_{0:d})}{Q(B_{d})-T_{d}(B_{d},B_{0:d})-A_{d+1}(B_{d},B_{0:d})}\mathbb{P}(V_{d+1},S_{0:d}=B_{0:d})$ (43) $\displaystyle=\frac{Q(B_{d}^{R})-T_{d}(B_{d}^{R},B_{0:d})-A_{d+1}(B_{d}^{R},B_{0:d})}{\underbrace{Q(\mathcal{X})}_{=~{}1}-T_{d}(\mathcal{X},B_{0:d})-A_{d+1}(\mathcal{X},B_{0:d})}\mathbb{P}(V_{d+1},B_{0:d}=B_{0:d})$ (44) $\displaystyle=Q(B_{d}^{R})-T_{d}(B_{d}^{R},B_{0:d})-A_{d+1}(B_{d}^{R},B_{0:d})$ (45) $\displaystyle=1-T_{d+1}(\mathcal{X},B_{0:d+1})$ (46) where we have written $B_{0:d+1}=(B_{0},\dots,B_{d},B_{d}^{R})$. Above, to go from 41 to 42 we used the definition of conditional probability, to go from 42 to 43 we used the definition in 19, to go from 43 to 44 we used the fact that for $k=0,1,2,\dots,$ it holds that $\displaystyle Q(\mathcal{X})-T_{k}(\mathcal{X},B_{0:k})-A_{k+1}(\mathcal{X},B_{0:k})$ $\displaystyle=Q(B_{k})-T_{k}(B_{k},B_{0:k})-A_{k+1}(B_{k},B_{0:k})+$ $\displaystyle~{}~{}~{}~{}~{}~{}+Q(B_{k}^{\prime})-\underbrace{T_{k}(B_{k}^{\prime},B_{0:k})}_{=~{}Q(B_{k}^{\prime})}-\underbrace{A_{k+1}(B_{k}^{\prime},B_{0:k})}_{=~{}0}$ (47) $\displaystyle=Q(B_{k})-T_{d}(B_{k},B_{0:k})-A_{k+1}(B_{k},B_{0:k}),$ (48) from 44 to 45 we have used eq. 40, and lastly from 45 to 46 we have again used eq. 48. Equation 46 similarly holds if $B_{d+1}=B^{R}_{d}$ by $B_{d+1}=B^{L}_{d}$, so we arrive at $\displaystyle\mathbb{P}(V_{d+1},B_{0:d+1}$ $\displaystyle=B_{0:d+1})=1-T_{d+1}(\mathcal{X},B_{0:d+1}),$ (49) which shows the inductive step for eq. 30. Further, we have $\displaystyle\mathbb{P}(W_{d+2},B_{0:d+1}=B_{0:d+1})=\mathbb{P}(W_{d+2}~{}|~{}V_{d+1},B_{0:d+1}=B_{0:d+1})\mathbb{P}(V_{d+1},B_{0:d+1}=B_{0:d+1})$ (50) and also by setting $k=d+1$ in eq. 35 we have $\mathbb{P}(W_{d+2}~{}|~{}V_{d+1},B_{0:d+1}=B_{0:d+1})=\frac{A_{d+2}(\mathcal{X},B_{0:d+1})}{1-T_{d+1}(\mathcal{X},B_{0:d+1})}.$ (51) Combining eq. 49 and eq. 51 we arrive at $\displaystyle\mathbb{P}(W_{d+2},B_{0:d+1}=B_{0:d+1})=A_{d+2}(\mathcal{X},B_{0:d+1}),$ (52) which is the inductive step for eq. 31. Putting eqs. 49 and 52 together shows the result. ∎ We now turn to defining and deriving the form of the measure $\tau_{D}$. We will define $\tau_{D}$ to be the measure such that for any $S\in\Sigma$, the probability that GRC terminates up to and including step $D$ and returns a sample within $S$ is given by $\tau_{D}(S)$. We will also show that $\tau_{D}$ is non-increasing in $D$. ###### Lemma 1 (Density of samples generated by GRC). The probability that GRC terminates by step $D\geq 1$ and produces a sample in $S$ is given by the measure $\tau_{D}(S)=\sum_{d=1}^{D}\sum_{k=1}^{2^{d-1}}A_{d}(S,\hat{Z}_{d,k}),$ (53) where $\hat{Z}_{D,k}$ is the ancestor branch of $Z_{D,k}$ as defined in eq. 29. Further, $\tau_{D}$ is non-decreasing in $D$, that is if $n\leq m$, then $\tau_{n}(S)\leq\tau_{m}(S)$ for all $S\in\Sigma$. ###### Proof. Let $V_{d}$ be the event that GRC does not terminate up to and including step $d$ and let $W_{d}(S)$ be the event that GRC terminates at step $d$ and returns a sample in $S$. Then $\displaystyle\tau_{D}(S)$ $\displaystyle=\sum_{d=1}^{D}\mathbb{P}(W_{d}(S))$ (54) $\displaystyle=\sum_{d=1}^{D}\mathbb{P}(W_{d}(S),V_{d-1})$ (55) $\displaystyle=\sum_{d=1}^{D}\sum_{k=1}^{2^{d-1}}\mathbb{P}(W_{d}(S),V_{d-1},S_{0:d-1}=\hat{Z}_{d,k})$ (56) $\displaystyle=\sum_{d=1}^{D}\sum_{k=1}^{2^{d-1}}\mathbb{P}(W_{d}(S)~{}|~{}V_{d-1},S_{0:d-1}=\hat{Z}_{d,k})~{}\mathbb{P}(V_{d-1},S_{0:d-1}=\hat{Z}_{d,k}).$ (57) Further, the terms in the summand can be expressed as $\displaystyle\mathbb{P}(V_{d-1},S_{0:d-1}=\hat{Z}_{d,k})$ $\displaystyle=1-T_{d-1}(\mathcal{X},\hat{Z}_{d,k}),$ (58) $\displaystyle\mathbb{P}(W_{d}(S)~{}|~{}V_{d-1},S_{0:d-1}=\hat{Z}_{d,k})$ $\displaystyle=\int_{x\in S}dP(x)\frac{\beta_{d}(x,\hat{Z}_{d,k})}{P(Z_{d,k})}$ (59) $\displaystyle=\int_{x\in S}dP(x)\frac{\alpha_{d}(x,\hat{Z}_{d,k})}{1-T_{d-1}(\mathcal{X},\hat{Z}_{d,k})}$ (60) $\displaystyle=\frac{A_{d}(S,\hat{Z}_{d,k})}{1-T_{d-1}(\mathcal{X},\hat{Z}_{d,k})},$ (61) and substituting eqs. 58 and 61 into the sum in eq. 57, we obtain eq. 53. Further, since the inner summand is always non-negative, increasing $D$ adds more non-negative terms to the sum, so $\tau_{D}$ is also non-decreasing in $D$. ∎ Now we turn to proving a few results about the measure $\tau_{D}$. Lemma 2 shows that $\tau_{D}\leq Q$ for all $D$. This result implies that $||Q-\tau_{D}||_{TV}=Q(\mathcal{X})-\tau_{D}(\mathcal{X})$, which we will use later. ###### Lemma 2 ($Q-\tau_{D}$ is non-negative). Let $D\in\mathbb{N}^{+}$. Then $Q-\tau_{D}$ is a positive measure, that is $Q(S)-\tau_{D}(S)\geq 0\text{ for any }S\in\Sigma.$ (62) ###### Proof. Let $S\in\Sigma$ and write $\displaystyle Q(S)-\tau_{D}(S)$ $\displaystyle=\sum_{k=1}^{2^{D-1}}Q(S\cap Z_{D,k})-\tau_{D}(S\cap Z_{D,k})$ (63) $\displaystyle=\sum_{k=1}^{2^{D-1}}\left[Q(S\cap Z_{D,k})-\sum_{d=1}^{D}\sum_{k^{\prime}=1}^{2^{D-1}}A_{d}(S\cap Z_{D,k},\hat{Z}_{D,k^{\prime}})\right]$ (64) $\displaystyle=\sum_{k=1}^{2^{D-1}}\left[Q(S\cap Z_{D,k})-\sum_{d=1}^{D}A_{d}(S\cap Z_{D,k},\hat{Z}_{D,k})\right]$ (65) $\displaystyle=\sum_{k=1}^{2^{D-1}}\left[Q(S\cap Z_{D,k})-T_{D-1}(S\cap Z_{D,k},\hat{Z}_{D,k})-A_{D}(S\cap Z_{D,k},\hat{Z}_{D,k})\right]$ (66) We will show that the summand in eq. 66 is non-negative. From the definition in eq. 14 we have $\displaystyle\alpha_{D}(x,\hat{Z}_{D,k})$ $\displaystyle=\min\left\\{\frac{dQ}{dP}(x)-t_{D-1}(x,\hat{Z}_{D,k}),\frac{1-T_{D-1}(\mathcal{X},\hat{Z}_{D,k})}{P({Z_{D,k}})}\right\\}$ (67) $\displaystyle\leq\frac{dQ}{dP}(x)-t_{D-1}(x,\hat{Z}_{D,k})$ (68) and integrating both sides of eq. 68 over $S\cap Z_{D,k}$, we obtain $\displaystyle A_{D}(S\cap Z_{D,k},\hat{Z}_{D,k})\leq Q(S\cap Z_{D,k})-T_{D-1}(S\cap Z_{D,k},\hat{Z}_{D,k})$ (69) Putting this together with eq. 66 we arrive at $\displaystyle Q(S)-\tau_{D}(S)\geq 0,$ (70) which is the required result. ∎ Thus far we have derived the form of $\tau_{D}$, shown that it is non- decreasing in $D$ and that it is no greater than $Q$. As we are interested in the limiting behaviour of $\tau_{D}$, we next show that its limit, $\tau=\lim_{D\to\infty}\tau_{D}$, is also a measure. Further, it also holds that $\tau\leq Q$. ###### Lemma 3 (Measures $\tau_{D}$ converge to a measure $\tau\leq Q$). For each $S\in\Sigma$, $\tau_{D}(S)$ converges to a limit. Further, the function $\tau:\Sigma\to[0,1]$ defined as $\tau(S)=\lim_{D\to\infty}\tau_{D}(S)$ (71) is a measure on $(\mathcal{X},\Sigma)$ and $\tau(S)\leq Q(S)$ for all $S\in\Sigma$. ###### Proof. First, by lemma 1, $\tau_{D}(S)$ is non-decreasing in $D$, and bounded above by $Q(S)$ for all $S\in\Sigma$. Therefore, for each $S\in\Sigma$, $\tau_{D}(S)$ converges to some limit as $D\to\infty$. Define $\tau:\Sigma\to[0,1]$ as $\tau(S)=\lim_{D\to\infty}\tau_{D}(S),$ (72) and note that $\tau$ is a non-negative set function for which $\tau(\emptyset)=0$. By the Vitali-Hahn-Saks theorem (see Corollary 4, p. 160; Dunford & Schwartz, 1988), $\tau$ is also countably additive, so it is a measure. Also, by lemma 2, $\tau_{D}(S)\leq Q(S)$ for all $D\in\mathbb{N}^{+}$ and all $S\in\Sigma$, so $\tau(S)\leq Q(S)$ for all $S\in\Sigma$. ∎ ###### Definition 9 ($H_{d,k}$, $H_{d}$ and $H$). For $d=1,2,\dots$ and $k=1,\dots,2^{d-1}$, we define the sets $H_{d,k}$ as $H_{d,k}=\left\\{x\in Z_{d,k}~{}\Big{|}~{}\frac{dQ}{dP}(x)-t_{d-1}(x,\hat{Z}_{d,k})\geq\frac{1-T_{d-1}(\mathcal{X},\hat{Z}_{d,k})}{P(Z_{d,k})}\right\\}.$ (73) Also, define the sets $H_{d}$ and $H$ as $\displaystyle H_{d}=\bigcup_{k=1}^{2^{d-1}}H_{d,k}~{}\text{ and }~{}H=\bigcap_{d=1}^{\infty}H_{d}.$ (74) ###### Lemma 4 ($T_{D}(\cdot,\hat{Z}_{D+1,k})$ and $\tau_{D}$ agree in $Z_{D+1,k}$). Let $R\in\Sigma$. If $R\subseteq Z_{D+1,k}$, then $\tau_{D}(R)=T_{D}(R,\hat{Z}_{D+1,k}).$ (75) ###### Proof. Suppose $R\subseteq Z_{D+1,k}$. First, we have $\tau_{D}(R)=\sum_{d=1}^{D}\sum_{k^{\prime}=1}^{2^{d-1}}A_{d}(R,\hat{Z}_{d,k^{\prime}})=\sum_{d=1}^{D}A_{d}(R,(\hat{Z}_{D+1,k})_{1:d}).$ (76) From the definition of $T_{D}$ in eq. 22, we have $\displaystyle T_{D}(R,\hat{Z}_{D+1,k})$ $\displaystyle=T_{D-1}(R\cap Z_{D+1,k},(\hat{Z}_{D+1,k})_{1:D})+A_{D}(R\cap Z_{D+1,k},(\hat{Z}_{D+1,k})_{1:D})+$ (77) $\displaystyle\quad\quad+\underbrace{Q(R\cap Z_{D+1,k}^{\prime})}_{=~{}0}$ $\displaystyle=T_{D-1}(R\cap Z_{D+1,k},(\hat{Z}_{D+1,k})_{1:D})+A_{D}(R\cap Z_{D+1,k},(\hat{Z}_{D+1,k})_{1:D})$ (78) $\displaystyle=T_{D-1}(R,(\hat{Z}_{D+1,k})_{1:D})+A_{D}(R,(\hat{Z}_{D+1,k})_{1:D})$ (79) where we have used the assumption that $R\subseteq Z_{D+1,k}$. In a similar manner, applying eq. 79 recursively $D-1$ more times, we obtain $T_{D}(R,\hat{Z}_{D+1,k})=\sum_{d=1}^{D}A_{d}(R,(\hat{Z}_{D+1,k})_{1:d})=\tau_{D}(R).$ (80) which is the required result. ∎ ###### Lemma 5 (Equalities with $Q,\tau_{D}$ and $\tau$). The following two equalities hold $Q(\mathcal{X}\setminus H_{D})=\tau_{D}(\mathcal{X}\setminus H_{D})~{}\text{ and }~{}Q(\mathcal{X}\setminus H)=\tau(\mathcal{X}\setminus H).$ (81) ###### Proof. Let $R=Z_{D+1,k}\setminus H_{D,k}$. Then, by similar reasoning used to prove eq. 77, we have $\displaystyle T_{D}(R,\hat{Z}_{D+1,k})=T_{D-1}(R,(\hat{Z}_{D+1,k})_{1:D})+A_{D}(R,(\hat{Z}_{D+1,k})_{1:D})$ (82) Further, we also have $\displaystyle A_{D}(R,\hat{Z}_{D,k})$ $\displaystyle=\int_{R}dP(x)~{}\alpha_{D}(x,\hat{Z}_{D,k})$ (83) $\displaystyle=\int_{R}dP(x)~{}\min\left\\{\frac{dQ}{dP}(x)-t_{D-1}(x,\hat{Z}_{D,k}),\frac{1-T_{D-1}(\mathcal{X},\hat{Z}_{D,k})}{P(Z_{D,k})}\right\\}$ (84) $\displaystyle=\int_{R}dP(x)~{}\left(\frac{dQ}{dP}(x)-t_{D-1}(x,\hat{Z}_{D,k})\right)$ (85) $\displaystyle=Q(R)-T_{D-1}(R,\hat{Z}_{D,k})$ (86) where from eq. 84 to eq. 85 we have used the definition of $H_{D,k}$. Then, combining eqs. 82 and 86 and using lemma 4, we arrive at $Q(Z_{D+1,k}\setminus H_{D,k})=T_{D}(Z_{D+1,k}\setminus H_{D,k},\hat{Z}_{D+1,k})=\tau_{D}(Z_{D+1,k}\setminus H_{D,k}).$ (87) Now, using the equation above, we have that $\tau_{D}(\mathcal{X}\setminus H_{D})=\sum_{k=1}^{2^{D}}\tau_{D}(Z_{D+1,k}\setminus H_{D})=\sum_{k=1}^{2^{D}}Q(Z_{D+1,k}\setminus H_{D})=Q(\mathcal{X}\setminus H_{D}).$ (88) Now, using $\tau_{D}\leq\tau\leq Q$ and $\tau_{D}(\mathcal{X}\setminus H_{D})=Q(\mathcal{X}\setminus H_{D})$, we have that $\tau(\mathcal{X}\setminus H_{D})=Q(\mathcal{X}\setminus H_{D})$, which is the first part of the result we wanted to show. Taking limits, we obtain $Q(\mathcal{X}\setminus H)=\lim_{D\to\infty}Q(\mathcal{X}\setminus H_{D})=\lim_{D\to\infty}\tau(\mathcal{X}\setminus H_{D})=\tau(\mathcal{X}\setminus H),$ (89) which is the second part of the required result. ∎ ### B.3 Breaking down the proof of Theorem 1 in five cases In definition 10 we introduce the quantities $w_{d}=Q(\mathcal{X})-\tau_{d}(\mathcal{X})$ and $p_{d}=\mathbb{P}(W_{d}~{}|~{}V_{d-1})$. Then we break down the proof of theorem 1 in five cases. First, in lemma 7 we show that if $p_{d}\not\to 0$, then $w_{d}\to 0$. Second, in lemma 8 we show that if $P(H_{d})\to 0$, then $w_{d}\to 0$. In lemma 9 we show an intermediate result, used in the other three cases, which we consider in lemmas 10, 11 and 12. Specifically, in these three cases we show that if $p_{d}\to 0$ and $P(H_{d})\not\to 0$, and assumption 1, 2 or 3 hold respectively, we have $w_{d}\to 0$. Putting these results together shows theorem 1. ###### Definition 10 ($p_{d}$, $w_{d,k}$ and $w_{d}$). Define $p_{d}=\mathbb{P}(W_{d}~{}|~{}V_{d-1})$. Also define $w_{d,k}$ and $w_{d}$ as $\displaystyle w_{d,k}$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}Q(Z_{d,k})-\tau_{d}(Z_{d,k}),$ (90) $\displaystyle w_{d}$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}\sum_{k=1}^{2^{d-1}}w_{d,k}.$ (91) ###### Lemma 6 ($w_{d}$ non-increasing in $d$). The sequence $w_{d}$ is non-negative and non-increasing in $d$. ###### Proof. Since $\tau_{d}$ is non-decreasing in $d$ (from lemma 5) and $w_{d}=\sum_{k=1}^{2^{d-1}}Q(Z_{d,k})-\tau_{d}(Z_{d,k})=Q(\mathcal{X})-\tau_{d}(\mathcal{X}),$ (92) it follows that $w_{d}$ is a non-increasing and non-negative sequence. ∎ ###### Lemma 7 (Case 1). If $p_{d}\not\to 0$, then $w_{d}\to 0$. ###### Proof. Let $p_{d}=\mathbb{P}(W_{d}~{}|~{}V_{d-1})$ and suppose $p_{d}\not\to 0$. Then, there exists $\epsilon>0$ such that $p_{d}>\epsilon$ occurs infinitely often. Therefore, there exists an increasing sequence of integers $a_{d}\in\mathbb{N}$ such that $p_{a_{d}}>\epsilon$ for all $d\in\mathbb{N}$. Then $\displaystyle\tau_{a_{d}}(\mathcal{X})$ $\displaystyle=\mathbb{P}\left(\bigcup_{d=1}^{a_{d}}W_{d}\right)$ (93) $\displaystyle=1-\mathbb{P}\left(V_{a_{d}}\right),$ (94) $\displaystyle=1-\prod_{d=1}^{a_{d}}\mathbb{P}\left(V_{d}~{}|~{}V_{d-1}\right),$ (95) $\displaystyle=1-\prod_{d=1}^{a_{d}}(1-p_{d}),$ (96) $\displaystyle\geq 1-(1-\epsilon)^{d}\to 1\text{ as }d\to\infty.$ (97) Therefore, $\tau_{d}(\mathcal{X})\to 1$ as $d\to\infty$, which implies that $||Q-\tau_{d}||_{TV}\to 0$. ∎ ###### Lemma 8 (Case 2). If $P(H_{d})\to 0$, then $w_{d}\to 0$. ###### Proof. Suppose $P(H_{d})\to 0$. Since $Q\ll P$, we have $Q(H)=0$, and since $Q\geq\tau\geq 0$ (by lemma 3), we also have $\tau(H)=0$. Therefore $\displaystyle\lim_{d\to\infty}w_{d}$ $\displaystyle=\lim_{d\to\infty}||Q-\tau_{d}||_{TV}$ (98) $\displaystyle=Q(\mathcal{X})-\tau(\mathcal{X})$ (99) $\displaystyle=\underbrace{Q(\mathcal{X}\setminus H)-\tau(\mathcal{X}\setminus H)}_{=~{}0\text{ from lemma }\ref{lem:Qtau}}+\underbrace{Q(H)}_{=~{}0}-\underbrace{\tau(H)}_{=~{}0}$ (100) $\displaystyle=0$ (101) which is the required result. ∎ ###### Lemma 9 (An intermediate result). If $p_{d}\to 0$ and $w_{d}\not\to 0$ as $d\to\infty$, then $\sum_{k=1}^{2^{d-1}}\frac{P(H_{d,k})}{P(Z_{d,k})}~{}w_{d,k}\to 0\text{ as }d\to\infty.$ (102) ###### Proof. Suppose that $p_{d}=\mathbb{P}(W_{d}~{}|~{}V_{d-1})\to 0$ and $w_{d}\not\to 0$. Then $\displaystyle\mathbb{P}(W_{d}~{}|~{}V_{d-1})$ $\displaystyle\geq\mathbb{P}(W_{d}(H_{d})~{}|~{}V_{d-1})$ (103) $\displaystyle=\sum_{k=1}^{2^{d-1}}\mathbb{P}\left(W_{d}(H_{d,k})~{}|~{}V_{d-1}\right)$ (104) $\displaystyle=\sum_{k=1}^{2^{d-1}}\mathbb{P}\left(W_{d}(H_{d,k}),S_{0:d-1}=\hat{Z}_{d,k}~{}|~{}V_{d-1}\right)$ (105) $\displaystyle=\sum_{k=1}^{2^{d-1}}\mathbb{P}\left(W_{d}(H_{d,k})~{}|~{}V_{d-1},S_{0:d-1}=\hat{Z}_{d,k}\right)\mathbb{P}\left(S_{0:d-1}=\hat{Z}_{d,k}~{}|~{}V_{d-1}\right)$ (106) $\displaystyle=\sum_{k=1}^{2^{d-1}}\frac{P(H_{d,k})}{P(Z_{d,k})}\mathbb{P}\left(S_{0:d-1}=\hat{Z}_{d,k}~{}|~{}V_{d-1}\right)$ (107) $\displaystyle=\sum_{k=1}^{2^{d-1}}\frac{P(H_{d,k})}{P(Z_{d,k})}\frac{w_{d,k}}{w_{d}}\to 0.$ (108) In addition, if $w_{d}\not\to 0$, then since $0\leq w_{d}\leq 1$ we have $\sum_{k=1}^{2^{d-1}}\frac{P(H_{d,k})}{P(Z_{d,k})}w_{d,k}\to 0.$ (109) which is the required result. ∎ ###### Lemma 10 (Case 3). Suppose that $p_{d}\to 0$, $P(H_{d})\not\to 0$ and assumption 1 holds. Then $w_{d}\to 0$. ###### Proof. Suppose that $p_{d}\to 0$, $P(H_{d})\not\to 0$. Suppose also that assumption 1 holds, meaning there exists $M\in\mathbb{R}$ such that $dQ/dP(x)<M$ for all $x\in\mathcal{X}$. Then for any $S\in\Sigma$, we have $\frac{Q(S)-\tau(S)}{P(S)}\leq\frac{Q(S)}{P(S)}=\frac{\int_{S}\frac{dQ}{dP}dP}{P(S)}\leq M~{}\frac{\int_{S}dP}{P(S)}=M\implies\frac{Q(S)-\tau(S)}{M}\leq P(S).$ (110) Further, we have $\displaystyle\sum_{k=1}^{2^{d-1}}\frac{P(H_{d,k})}{P(Z_{d,k})}~{}w_{d,k}$ $\displaystyle\geq\sum_{k=1}^{2^{d-1}}\frac{P(H_{d,k})}{P(Z_{d,k})}~{}(Q(H_{d,k})-\tau(H_{d,k}))$ (111) $\displaystyle\geq\frac{1}{M}\sum_{k=1}^{2^{d-1}}\frac{(Q(H_{d,k})-\tau(H_{d,k}))^{2}}{P(Z_{d,k})}$ (112) $\displaystyle\geq\frac{1}{M}\sum_{k=1}^{2^{d-1}}\frac{(Q(H\cap H_{d,k})-\tau(H\cap H_{d,k}))^{2}}{P(Z_{d,k})}$ (113) $\displaystyle\geq\frac{1}{M}\sum_{k=1}^{2^{d-1}}\frac{\Delta_{d,k}^{2}}{P(Z_{d,k})}$ (114) $\displaystyle=\frac{1}{M}~{}\Phi_{d}$ (115) $\displaystyle\to 0,$ (116) where in the second inequality we have used eq. 110 and we have defined $\displaystyle\Delta_{d,k}$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}Q(H\cap H_{d,k})-\tau(H\cap H_{d,k}),$ (117) $\displaystyle\Phi_{d}$ $\displaystyle\stackrel{{\scriptstyle\text{def}}}{{=}}\sum_{k=1}^{2^{d-1}}\frac{\Delta_{d,k}^{2}}{P(Z_{d,k})}.$ (118) Now note that the sets $H\cap H_{d+1,2k}$ and $H\cap H_{d+1,2k+1}$ partition the set $H\cap H_{d,k}$. Therefore $\Delta_{d,k}=\Delta_{d+1,2k}+\Delta_{d+1,2k+1}.$ (119) By the definition of $\Phi_{d}$ in eq. 118, we can write $\Phi_{d+1}=\sum_{k=1}^{2^{d}}\frac{\Delta_{d,k}^{2}}{P(Z_{d+1,k})}=\sum_{k=1}^{2^{d-1}}\left[\frac{\Delta_{d+1,2k}^{2}}{P(Z_{d+1,2k})}+\frac{\Delta_{d+1,2k+1}^{2}}{P(Z_{d+1,2k+1})}\right],$ (120) where we have written the sum over $2^{d}$ terms as a sum over $2^{d-1}$ pairs of terms. We can rewrite the summand on the right hand side as $\displaystyle\frac{\Delta_{d+1,2k}^{2}}{P(Z_{d+1,2k})}+\frac{\Delta_{d+1,2k+1}^{2}}{P(Z_{d+1,2k+1})}$ $\displaystyle=\frac{\Delta_{d+1,2k}^{2}}{P(Z_{d+1,2k})}+\frac{(\Delta_{d,k}-\Delta_{d+1,2k})^{2}}{P(Z_{d+1,2k+1})}$ (121) $\displaystyle=\Delta_{d,k}^{2}\left[\frac{\rho^{2}}{P(Z_{d+1,2k-1})}+\frac{(1-\rho)^{2}}{P(Z_{d+1,2k})}\right]$ (122) $\displaystyle=\Delta_{d,k}^{2}~{}g(\rho)$ (123) where in eq. 121 we have used eq. 119, from eq. 121 to eq. 122 we defined the quantity $\rho=\Delta_{d+1,2k}/\Delta_{d,k}$, and from eq. 122 to eq. 123 we have defined $g:[0,1]\to\mathbb{R}$ as $g(r)\stackrel{{\scriptstyle\text{def}}}{{=}}\frac{r^{2}}{P(Z_{d+1,2k})}+\frac{(1-r)^{2}}{P(Z_{d+1,2k+1})}.$ (124) The first and second derivatives of $g$ are $\displaystyle\frac{dg}{dr}$ $\displaystyle=\frac{2r}{P(Z_{d+1,2k})}-\frac{2(1-r)}{P(Z_{d+1,2k+1})},$ (125) $\displaystyle\frac{d^{2}g}{dr^{2}}$ $\displaystyle=\frac{2}{P(Z_{d+1,2k})}+\frac{2}{P(Z_{d+1,2k+1})}>0,$ (126) so $g$ has a single stationary point that is a minimum, at $r=r_{\min}$, which is given by $\displaystyle r_{\min}:=\frac{P(Z_{d+1,2k})}{P(Z_{d+1,2k})+P(Z_{d+1,2k+1})}.$ (127) Plugging this back in $g$, we obtain $\displaystyle g(r_{\min})=\frac{1}{P(Z_{d+1,2k})+P(Z_{d+1,2k+1})}=\frac{1}{P(Z_{d,k})},$ (128) which implies that $\displaystyle\frac{\Delta_{d+1,2k}^{2}}{P(Z_{d+1,2k})}+\frac{\Delta_{d+1,2k+1}^{2}}{P(Z_{d+1,2k+1})}\geq\frac{\Delta_{d,k}^{2}}{P(Z_{d,k})}.$ (129) Therefore $\Phi_{d+1}=\sum_{k=1}^{2^{d}}\frac{\Delta_{d,k}^{2}}{P(Z_{d+1,k})}\geq\sum_{k=1}^{2^{d-1}}\frac{\Delta_{d,k}^{2}}{P(Z_{d,k})}=\Phi_{d},$ (130) but since $\Phi_{d}\to 0$, this is only possible if $\Phi_{d}=0$ for all $d$, including $d=1$, which would imply that $\Delta_{1,1}=Q(H\cap H_{1,1})-\tau(H\cap H_{1,1})=Q(H)-\tau(H)=0,$ (131) which, together with lemma 5, implies that $Q(\mathcal{X})-\tau(\mathcal{X})=Q(H)-\tau(H)=0,$ (132) and therefore $w_{d}=||Q-\tau_{d}||_{TV}\to 0$. ∎ ###### Lemma 11 (Case 4). Suppose that $p_{d}\to 0$, $P(H_{d})\not\to 0$ and assumption 3 holds. Then $w_{d}\to 0$. ###### Proof. Suppose that $p_{d}\to 0$, $P(H_{d})\not\to 0$. Suppose also that assumption that assumption 3 holds, meaning that for each $d$, we have $w_{d,k}>0$ for exactly one value of $k=k_{d}$, and $w_{d,k}=0$ for all other $k\neq k_{d}$. In this case, it holds that $H_{d,k}=\emptyset$ for all $k\neq k_{d}$ and $H_{d}=H_{d,k_{d}}$. Since $P(H_{d})\not\to 0$ and $P(H_{d})$ is a decreasing sequence, it converges to some positive constant. We also have $p_{d}\geq\sum_{k=1}^{2^{d-1}}\frac{P(H_{d,k})}{P(Z_{d,k})}~{}w_{d,k}=\frac{P(H_{d,k_{d}})}{P(Z_{d,k_{d}})}~{}w_{d,k_{d}}=\frac{P(H_{d,k_{d}})}{P(Z_{d,k_{d}})}~{}w_{d}\geq P(H_{d})~{}w_{d}\to 0,$ (133) which can only hold if $w_{d}\to 0$, arriving at the result. ∎ ###### Lemma 12 (Case 5). Suppose that $p_{d}\to 0$, $P(H_{d})\not\to 0$ and assumption 3 holds. Then $w_{d}\to 0$. ###### Proof. Suppose that $p_{d}\to 0$, $P(H_{d})\not\to 0$ and assumption 3 holds. Since each $x\in\mathcal{X}$ belongs to exactly one $Z_{d,k}$ we can define the function $B_{d}:\mathcal{X}\to\Sigma$ as $B_{d}(x)=Z_{d,k}\text{ such that }x\in Z_{d,k}.$ (134) Using this function we can write $p_{d}\geq\sum_{k=1}^{2^{d-1}}\frac{P(H_{d,k})}{P(Z_{d,k})}~{}w_{d,k}=\sum_{k=1}^{2^{d-1}}P(H_{d,k})~{}\frac{Q(Z_{d,k})-\tau_{d}(Z_{d,k})}{P(Z_{d,k})}=\int_{H_{d}}dP~{}\frac{Q(B_{d}(x))-\tau_{d}(B_{d}(x))}{P(B_{d}(x))}.$ Now, because the sets $H_{d}$ are measurable, their intersection $H:=\cap_{d=1}^{\infty}H_{d}$ is also measurable. We can therefore lower bound the integral above as follows $\displaystyle\int_{H_{d}}dP~{}\frac{Q(B_{d}(x))-\tau_{d}(B_{d}(x))}{P(B_{d}(x))}$ $\displaystyle\geq\int_{H}dP~{}\frac{Q(B_{d}(x))-\tau_{d}(B_{d}(x))}{P(B_{d}(x))}$ (135) $\displaystyle\geq\int_{H}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))},$ (136) where the first inequality holds as the integrand is non-negative and we are constraining the integration domain to $H\subseteq H_{d}$, and the second inequality holds because $\tau_{d}(S)\leq\tau(S)$ for any $S\in\Sigma$. Define $\mathcal{C}$ to be the set of all intersections of nested partitions, with non-zero mass under $P$ $\mathcal{C}=\left\\{\bigcap_{d=0}^{\infty}Z_{d,k_{d}}:P\left(\bigcap_{d=0}^{\infty}Z_{d,k_{d}}\right)>0,k_{0}=1,k_{d+1}=2k_{d}\text{ or }k_{d+1}=2k_{d}+1\right\\},$ (137) and note that all of its elements are pairwise disjoint. Each of the elements of $\mathcal{C}$ is a measurable set because it is a countable intersection of measurable sets. In addition, $\mathcal{C}$ is a countable set, which can be shown as follows. Define the sets $\mathcal{C}_{n}$ as $\mathcal{C}_{n}=\left\\{E\in\mathcal{C}:2^{-n-1}<P(E)\leq 2^{-n}\right\\}\text{ for }n=0,1,\dots$ (138) and note that their union equals $\mathcal{C}$. Further, note that each $\mathcal{C}_{n}$ must contain a finite number of elements. That is because if $\mathcal{C}_{n}$ contained an infinite number of elements, say $E_{1},E_{2},\dots\in\mathcal{C}_{n}$, then $\displaystyle P(\mathcal{X})\geq P\left(\bigcup_{k=1}^{\infty}E_{k}\right)=\sum_{k=1}^{\infty}P(E_{k})>\sum_{k=1}^{\infty}2^{-n-1}\to\infty,$ (139) where the first equality holds because $P$ is an additive measure and the $E_{n}$ terms are disjoint, and the second inequality follows because $E_{k}\in\mathcal{C}_{n}$ so $P(E_{k})>2^{-n-1}$. This results in a contradiction because $P(\mathcal{X})=1$, so each $\mathcal{C}_{n}$ must contain a finite number of terms. Therefore, $\mathcal{C}$ is a countable union of finite sets, which is also countable. This implies that the union of the elements of $\mathcal{C}$, namely $C=\cup_{C^{\prime}\in\mathcal{C}}C^{\prime}$ is a countable union of measurable sets and therefore also measurable. Since $C$ is measurable, $H\setminus C$ is also measurable and we can rewrite the integral in eq. 135 as $\displaystyle p_{d}$ $\displaystyle\geq\int_{H}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}$ (140) $\displaystyle=\int_{H\cap C}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}+\int_{H\setminus C}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}$ (141) $\displaystyle\to 0$ (142) Since both terms above are non-negative and their sum converges to $0$, the terms must also individually converge to $0$. Therefore, for the first term, we can write $\lim_{d\to\infty}\int_{H\cap C}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}=\liminf_{d\to\infty}\int_{H\cap C}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}=0.$ (143) Similarly to $B_{d}$ defined in eq. 134, let us define $B:C\to\Sigma$ as $B(x)=C^{\prime}\in\mathcal{C}\text{ such that }x\in C^{\prime}.$ (144) Applying Fatou’s lemma (4.3.3, p. 131; Dudley, 2018) to eq. 143, we obtain $\displaystyle\liminf_{d\to\infty}\int_{H\cap C}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}$ $\displaystyle\geq\int_{H\cap C}dP~{}\liminf_{d\to\infty}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}$ (145) $\displaystyle=\int_{H\cap C}dP~{}\frac{Q(B(x))-\tau(B(x))}{P(B(x))}$ (146) $\displaystyle=0,$ (147) where from eq. 145 to eq. 146 we have used the fact that $P(B_{d}(x))>0$ whenever $x\in C$ and also that $B_{1}(x)\supseteq B_{2}(x)\supseteq\dots$. Now we can re-write this integral as a sum, as follows. Let the elements of $\mathcal{C}$, which we earlier showed is countable, be $C_{1},C_{2},\dots$ and write $\displaystyle\int_{H\cap C}dP~{}\frac{Q(B(x))-\tau(B(x))}{P(B(x))}$ $\displaystyle=\sum_{n=1}^{\infty}\int_{H\cap C_{n}}dP~{}\frac{Q(B(x))-\tau(B(x))}{P(B(x))}$ (148) $\displaystyle=\sum_{n=1}^{\infty}\frac{P(H\cap C_{n})}{P(C_{n})}\left(Q(C_{n})-\tau(C_{n})\right)$ (149) $\displaystyle=0.$ (150) Now, from lemma 5, we have $\displaystyle\sum_{n=1}^{\infty}\frac{P(H\cap C_{n})}{P(C_{n})}\left(Q(C_{n})-\tau(C_{n})\right)=\sum_{n=1}^{\infty}\frac{P(H\cap C_{n})}{P(C_{n})}\left(Q(H\cap C_{n})-\tau(H\cap C_{n})\right)=0,$ (151) which in turn implies that for each $n=1,2,\dots$, we have either $Q(H\cap C_{n})-\tau(H\cap C_{n})=0$ or $P(H\cap C_{n})=0$. However, the latter case also implies $Q(H\cap C_{n})-\tau(H\cap C_{n})=0$ because $Q\ll P$, so $Q(H\cap C_{n})-\tau(H\cap C_{n})=0$ holds for all $n$. Therefore $\tau(H\cap C)=\sum_{n=1}^{\infty}\tau(H\cap C_{n})=\sum_{n=1}^{\infty}Q(H\cap C_{n})=Q(H\cap C).$ (152) Returning to the second term in the right hand of eq. 141, and again applying Fatou’s lemma $\displaystyle\liminf_{d\to\infty}\int_{H\setminus C}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}$ $\displaystyle\geq\int_{H\setminus C}dP~{}\liminf_{d\to\infty}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}.$ (153) Now, since $Z$ has the nice-shrinking property from assumption 3, we can apply a standard result from measure theory and integration Rudin (1986, given in Theorem 7.10, p. 140), to show that the following limit exists and the following equalities are satisfied $\displaystyle\lim_{d\to\infty}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}$ $\displaystyle=\lim_{d\to\infty}\frac{1}{P(B_{d}(x))}\int_{B_{d}}dP\left(\frac{dQ}{dP}(x)-\frac{d\tau}{dP}(x)\right)$ (154) $\displaystyle=\frac{dQ}{dP}(x)-\frac{d\tau}{dP}(x)$ (155) Inserting 155 into eq. 153, we obtain $\displaystyle\liminf_{d\to\infty}\int_{H\setminus C}dP~{}\frac{Q(B_{d}(x))-\tau(B_{d}(x))}{P(B_{d}(x))}\geq\int_{H\setminus C}dP~{}\left(\frac{dQ}{dP}(x)-\frac{d\tau}{dP}(x)\right)=0,$ (156) which in turn implies that $\frac{dQ}{dP}(x)-\frac{d\tau}{dP}(x)=0~{}~{}P\text{-almost-everywhere on }H\setminus C,$ (157) or equivalently that $Q(H\setminus C)=\tau(H\setminus C)$. Combining this with the fact that $Q(\mathcal{X}\setminus H)=\tau(\mathcal{X}\setminus H)$ and our earlier result that $Q(H\cap C)=\tau(H\cap C)$, we have $||Q-\tau||_{TV}=Q(\mathcal{X}\setminus H)-\tau(\mathcal{X}\setminus H)+Q(H\setminus C)-\tau(H\setminus C)+Q(H\cap C)-\tau(H\cap C)=0,$ which is equivalent to $w_{d}=||Q-\tau_{d}||_{TV}\to 0$, that is the required result. ∎ ###### Theorem (Correcness of GRC). If any one of the assumptions 1, 2 or 3 holds, then $||Q-\tau_{d}||_{TV}\to 0~{}\text{ as }~{}d\to\infty.$ (158) ###### Proof. If $p_{d}\to 0$, then $w_{d}\to 0$ by lemma 7. If $P(H_{d})\to 0$, then $w_{d}\to 0$ by lemma 8. Therefore suppose that $p_{d}\not\to 0$ and $P(H_{d})\not\to 0$. Then if any one of assumptions 1, 2 or 3 holds, we can conclude from lemma 10, 11 or 12 respectively, that $||Q-\tau_{d}||_{TV}\to 0$. ∎ ## Appendix C Optimality of GRCS Algorithm 3 GRCS with arthmetic coding for the heap index. 1:Target $Q$, proposal $P$ over $\mathbb{R}$ with unimodal density ratio $r=dQ/dP$ with mode $\mu$. 2:$d\leftarrow 0,T_{0}\leftarrow 0,L_{0}\leftarrow 0$ 3:$I_{0}\leftarrow 1,S_{1}\leftarrow\mathbb{R}$ 4:while $\mathtt{True}$ do 5: $X_{I_{d}}\sim P|_{S_{d}}/P(S_{d})$ 6: $U_{I_{d}}\sim\text{Uniform}(0,1)$ 7: $\beta_{I_{d}}\leftarrow\mathtt{clip}\left(P(S_{d})\cdot\frac{r(X_{I_{d}})-L_{d}}{1-T_{d}},0,1\right)$ $\triangleright$ $\mathtt{clip}(y,a,b)\stackrel{{\scriptstyle\mathit{def}}}{{=}}\max\\{\min\\{y,b\\},a\\}$ 8: if $U_{I_{d}}\leq\beta_{d+1}$ then 9: return $X_{I_{d}},I_{d}$ 10: end if 11: if $X_{I_{d}}>\mu$ then 12: $I_{d+1}\leftarrow 2I_{d}$ 13: $S_{d+1}\leftarrow S_{d}\cap(-\infty,X_{I_{d}})$ 14: else 15: $I_{d+1}\leftarrow 2I_{d}+1$ 16: $S_{d+1}\leftarrow S_{d}\cap(X_{I_{d}},\infty)$ 17: end if 18: $L_{d+1}\leftarrow L_{d}+T_{d}/P(S_{d})$ 19: $T_{d+1}\leftarrow\mathbb{P}_{Y\sim Q}[r(Y)\geq L_{d+1}]-L_{d+1}\cdot\mathbb{P}_{Y\sim P}[r(Y)\geq L_{d+1}]$ 20: $d\leftarrow d+1$ 21:end while In this section, we prove Theorems 2 and 3. We are only interested in continuous distributions over $\mathbb{R}$ with unimodal density ratio $dQ/dP$ for these theorems. Hence, we begin by specializing Algorithm 2 to this setting, shown in Algorithm 3. For simplicity, we also dispense with the abstraction of partitioning processes and show the bound update process directly. Furthermore, we also provide an explicit form for the AcceptProb and RuledOutMass functions. Before we move on to proving our proposed theorems, we first prove two useful results. First, we bound the negative log $P$-mass of the bounds with which Algorithm 3 terminates. ###### Lemma 13. Let $Q$ and $P$ be distributions over $\mathbb{R}$ with unimodal density ratio $r=dQ/dP$, given to Algorithm 3 as the target and proposal distribution as input, respectively. Let $d\geq 0$ and let $X_{1:d}\stackrel{{\scriptstyle\mathit{def}}}{{=}}X_{1},\ldots,X_{d}$ denote the samples simulated by Algorithm 3 up to step $d+1$, where for $d=0$ we define the empty list as $X_{1:0}=\emptyset$. Let $S_{d}$ denote the bounds at step $d+1$. Then, $\displaystyle-\sum_{j=0}^{d}A_{j+1}(\mathbb{R},S_{0:d})\cdot\log P(S_{j})\leq D_{\mathrm{KL}}[Q\|P]+\log e.$ (159) ###### Proof. For brevity, we will write $A_{d}=A_{d}(\mathbb{R},S_{0:d})$ and $T_{d}=T_{d}(\mathbb{R},S_{0:d})$. Furthermore, as in Algorithm 3, we define $\displaystyle L_{d}\stackrel{{\scriptstyle\mathit{def}}}{{=}}\sum_{j=0}^{d-1}\frac{1-T_{j}}{P(S_{j})}\quad\text{with}\quad L_{0}=0.$ (160) Note that $X_{1:d}$ is well-defined for all $d\geq 0$ since we could remove the return statement from the algorithm to simulate the bounds it would produce up to an arbitrary step $d$. Now, note that by Proposition 3 we have $\mathbb{P}[D=d\mid X_{1:d}]=A_{d+1}(\mathbb{R},S_{0:d})$. Now, fix $d\geq 0$ and bounds $S_{0:d}$, and let $x\in\mathbb{R}$ be such that $\alpha_{d+1}(x)>0$ which holds whenever $r(x)\geq L_{d}$. From this, for $d\geq 1$ we find $\displaystyle r(x)$ $\displaystyle\geq\sum_{j=0}^{d-1}\frac{1-T_{j}}{P(S_{j})}$ (161) $\displaystyle\geq\frac{1-T_{d-1}}{P(S_{d-1})},$ (162) where the second inequality follows from the fact that the $(1-T_{j})/P(S_{j})$ terms are all positive. taking logs, we get $\displaystyle\log r(x)-\log(1-T_{d-1})\geq-\log P(S_{d-1}).$ (163) Now, we consider the expectation of interest: $\displaystyle\sum_{j=0}^{d}-A_{j+1}\cdot\log P(S_{j})$ $\displaystyle=-\sum_{j=0}^{d}\int_{\mathbb{R}}\alpha_{j+1}(x)\log P(S_{j})\,dx$ (164) $\displaystyle\stackrel{{\scriptstyle\text{\lx@cref{creftype~refnum}{eq:log_bound_size_ineq}}}}{{\leq}}\sum_{j=0}^{d}\int_{\mathbb{R}}\alpha_{j+1}(x)(\log(r(x))-\log(1-T_{j}))\,dx$ (165) $\displaystyle\stackrel{{\scriptstyle(a)}}{{\leq}}\int_{\mathbb{R}}\sum_{j=0}^{\infty}\alpha_{j+1}(x)\log r(x)\,dx+\sum_{j=0}^{\infty}A_{j+1}\log\frac{1}{1-T_{j}}$ (166) $\displaystyle\stackrel{{\scriptstyle(b)}}{{=}}\int_{\mathbb{R}}q(x)\log r(x)\,dx+\sum_{j=0}^{\infty}(T_{j+1}-T_{j})\log\frac{1}{1-T_{j}}$ (167) $\displaystyle=D_{\mathrm{KL}}[Q\|P]+\sum_{j=0}^{\infty}(T_{j+1}-T_{j})\log\frac{1}{1-T_{j}}$ (168) $\displaystyle\stackrel{{\scriptstyle(c)}}{{\leq}}D_{\mathrm{KL}}[Q\|P]\cdot\log 2+\int_{0}^{1}\log\frac{1}{1-t}\,dt$ (169) $\displaystyle=D_{\mathrm{KL}}[Q\|P]+\log e.$ (170) Inequality (a) holds because all terms are positive. This is guaranteed by the fact that for $d\geq 1$, we have $L_{d}\geq 1$, hence $0\leq\log L_{d}\leq r(x)$ whenever Equation 163 holds. Equality (b) follows by the correctness of GRC (Theorem 1), which implies that for all $x\in\mathbb{R}$ we have $\sum_{j=0}^{\infty}\alpha_{d}(x)=q(x)$, and inequality (c) follows from the facts that $0\leq T_{d}\leq 1$ for all $d$ and that the summand in the second term forms a lower-Riemann sum approximation to $-\log(1-t)$. ∎ Second, we consider the contraction rate of the bounds $S_{0:d}$, considered by Algorithm 3. ###### Lemma 14. Let $Q$ and $P$ be distributions over $\mathbb{R}$ with unimodal density ratio $r=dQ/dP$, given to Algorithm 3 as the target and proposal distribution as input, respectively. Assume $P$ has CDF $F_{P}$ and the mode of $r$ is at $\mu$. Fix $d\geq 0$ and let $X_{1:d}$ be the samples considered by Algorithm 3 and $S_{d}$ the bounds at step $d+1$. Then, $\displaystyle\mathbb{E}_{X_{1:d}}[P(S_{d})]\leq\left(\frac{3}{4}\right)^{d}$ (171) ###### Proof. We prove the claim by induction. For $d=0$ the claim holds trivially, since $S_{0}=\mathbb{R}$, hence $P(S_{0})=1$. Assume now that the claim holds for $d=k-1$, and we prove the statement for $d=k$. By the law of iterated expectations, we have $\displaystyle\mathbb{E}_{X_{1:k}}[P(S_{k})]=\mathbb{E}_{X_{1:k-1}}[\mathbb{E}_{X_{k}\mid X_{1:k-1}}[P(S_{k})]].$ (172) Let us now examine the inner expectation. First, assume that $S_{k-1}=(a,b)$ for some real numbers $a<b$ and define $A=F_{P}(a),B=F_{P}(B),M=F_{P}(\mu)$ and $U=F_{P}(X_{k})$. Since $X_{k}\mid X_{1:k-1}\sim P|_{S_{k-1}}$, by the probability integral transform we have $U\sim\mathrm{Unif}(A,B)$, where $\mathrm{Unif}(A,B)$ denotes the uniform distribution on the interval $(A,B)$. The two possible intervals from which Algorithm 3 will choose are $(a,X_{k})$ and $(X_{k},b)$, whose measures are $P((a,X_{k}))=F_{P}(X_{k})-F_{P}(a)=U-A$ and similarly $P((X_{k},b))=B-U$. Then, $P(S_{k})\leq\max\\{U-A,B-U\\}$, from which we obtain the bound $\displaystyle\mathbb{E}_{X_{k}\mid X_{1:k-1}}[P(S_{k})]\leq\mathbb{E}_{U}[\max\\{U-A,B-U\\}]=\frac{3}{4}(B-A)=\frac{3}{4}P(S_{k-1}).$ (173) Plugging this into Equation 172, we get $\displaystyle\mathbb{E}_{X_{1:k}}[P(S_{k})]$ $\displaystyle\leq\frac{3}{4}\mathbb{E}_{X_{1:k-1}}\left[P(S_{k-1})\right]$ (174) $\displaystyle\leq\frac{3}{4}\cdot\left(\frac{3}{4}\right)^{k-1},$ (175) where the second inequality follows from the inductive hypothesis, which finishes the proof. ∎ The proof of Theorem 3: We prove our bound on the runtime of Algorithm 3 first, as this will be necessary for the proof of the bound on the codelength. First, let $D$ be the number of steps Algorithm 3 takes before it terminates minus $1$. Then, we will show that $\displaystyle\mathbb{E}[D]\leq\frac{1}{\log(4/3)}D_{\mathrm{KL}}[Q\|P]+4$ (176) We tackle this directly. Hence, let $\displaystyle\mathbb{E}_{D}[D]$ $\displaystyle=\lim_{d\to\infty}\mathbb{E}_{X_{1:j}}\left[\sum_{j=1}^{d}j\cdot A_{j+1}\right]$ (177) $\displaystyle=\lim_{d\to\infty}\mathbb{E}_{X_{1:j}}\left[\sum_{j=1}^{d}\frac{-j}{\log P(S_{j})}\cdot-A_{j+1}\log P(S_{j})\right]$ (178) $\displaystyle\leq\lim_{d\to\infty}\mathbb{E}_{X_{1:j}}\left[\max_{j\in[1:d]}\left\\{\frac{-j}{\log P(S_{j})}\right\\}\cdot\sum_{j=1}^{d}-A_{j+1}\log P(S_{j})\right]$ (179) $\displaystyle\stackrel{{\scriptstyle\text{\lx@cref{creftype~refnum}{lemma:negative_log_bound_size_bound}}}}{{\leq}}\left(D_{\mathrm{KL}}[Q\|P]+\log e\right)\cdot\lim_{d\to\infty}\mathbb{E}_{X_{1:j}}\left[\max_{j\in[1:d]}\left\\{\frac{-j}{\log P(S_{j})}\right\\}\right].$ (180) To finish the proof, we will now bound the term involving the limit. To do this, note, that for any finite collection of reals $F$, we have $\max_{x\in F}\\{x\\}=-\min_{x\in F}\\{-x\\}$, and that for a finite collection of real- valued random variables $\hat{F}$ we have $\mathbb{E}[\min_{{\mathbf{x}}\in\hat{F}}\\{{\mathbf{x}}\\}]\leq\min_{{\mathbf{x}}\in\hat{F}}\\{\mathbb{E}[{\mathbf{x}}]\\}$. Now, we have $\displaystyle\lim_{d\to\infty}\mathbb{E}_{X_{1:j}}\left[\max_{j\in[1:d]}\left\\{\frac{-j}{\log P(S_{j})}\right\\}\right]$ $\displaystyle=\lim_{d\to\infty}-\mathbb{E}_{X_{1:j}}\left[\min_{j\in[1:d]}\left\\{\frac{j}{\log P(S_{j})}\right\\}\right]$ (181) $\displaystyle\leq\lim_{d\to\infty}\left(-\min_{j\in[1:d]}\left\\{\mathbb{E}_{X_{1:j}}\left[\frac{j}{\log P(S_{j})}\right]\right\\}\right)$ (182) $\displaystyle\stackrel{{\scriptstyle\text{(a)}}}{{\leq}}\lim_{d\to\infty}\left(-\min_{j\in[1:d]}\left\\{\frac{j}{\log\mathbb{E}_{X_{1:j}}\left[P(S_{j})\right]}\right\\}\right)$ (183) $\displaystyle\stackrel{{\scriptstyle\text{\lx@cref{creftype~refnum}{lemma:bound_size_exp_bound}}}}{{\leq}}\lim_{d\to\infty}\left(-\min_{j\in[1:d]}\left\\{\frac{-j}{j\log(4/3)}\right\\}\right)$ (184) $\displaystyle=\lim_{d\to\infty}\left(\max_{j\in[1:d]}\left\\{\frac{1}{\log(4/3)}\right\\}\right)$ (185) $\displaystyle=\frac{1}{\log(4/3)}$ (186) Inequality (a) follows from Jensen’s inequality. Finally, plugging this back into the previous equation, we get $\displaystyle\mathbb{E}[D]\leq\frac{D_{\mathrm{KL}}[Q\|P]+\log e}{\log 4/3}\leq\frac{D_{\mathrm{KL}}[Q\|P]}{\log 4/3}+4$ (187) Proof of Theorem 2: For the codelength result, we need to encode the length of the search path and the search path itself. More formally, since the returned sample $X$ is a function of the partition process $Z$, the search path length $D$ and search path $S_{0:D}$, we have $\displaystyle\mathbb{H}[X\mid Z]\leq\mathbb{H}[D,S_{0:D}]=\mathbb{H}[D]+\mathbb{H}[S_{0:D}\mid D].$ (188) we can encode $D$ using Elias $\gamma$-coding, from which we get $\displaystyle\mathbb{H}[D]$ $\displaystyle\leq\mathbb{E}_{D}[2\log(D+1)]+1$ (189) $\displaystyle\leq 2\log(\mathbb{E}[D]+1)+1$ (190) $\displaystyle\leq 2\log\left(\frac{D_{\mathrm{KL}}[Q\|P]+\log e}{\log(4/3)}+1\right)+1$ (191) $\displaystyle\leq 2\log\left(D_{\mathrm{KL}}[Q\|P]+\log e+\log(4/3)\right)+1-2\log\left(\log(4/3)\right)$ (192) $\displaystyle\leq 2\log\left(D_{\mathrm{KL}}[Q\|P]+1\right)+1-2\log\left(\log(4/3)\right)+2\log(\log e+\log(4/3))$ (193) $\displaystyle\leq 2\log\left(D_{\mathrm{KL}}[Q\|P]+1\right)+6.$ (194) Given the search path length $D$, we can use arithmetic coding (AC) to encode the sequence of bounds $S_{0:D}$ using $-\log P(S_{D})+2$ bits (assuming infinite precision AC). Hence, we have that the average coding cost is upper bounded by $\displaystyle\mathbb{H}[S_{0:D}\mid D]\leq\mathbb{E}_{D}[-\log P(S_{D})]+2\stackrel{{\scriptstyle\text{\lx@cref{creftype~refnum}{lemma:negative_log_bound_size_bound}}}}{{\leq}}D_{\mathrm{KL}}[Q\|P]+5.$ (195) Putting everything together, we find $\displaystyle\mathbb{H}[D,S_{0:D}]\leq D_{\mathrm{KL}}[Q\|P]+2\log(D_{\mathrm{KL}}[Q\|P]+1)+11,$ (196) as required. ## Appendix D Additional experiments with depth-limited GRC In this section we show the results of some experiments comparing the approximation bias of depth limited GRCD, to that of depth limited AD∗ , following the setup of Flamich et al. (2022). Limiting the depth of each algorithm introduces bias in the resulting samples, as these are not guaranteed to be distributed from the target distribution $Q$, but rather from a different distribution $\smash{\hat{Q}}$. Figure 5 quantifies the effect of limiting the depth on the bias of the resulting samples. In our experiment we take $Q$ and $P$ to be Gaussian and we fix $D_{\mathrm{KL}}[Q\|P]=3$ (bits), and consider three different settings of $D_{\infty}[Q\|P]=5,7$ or $9$ (bits), corresponding to each of the panes in fig. 5. For each such setting, we set the depth limit of each of the two algorithms to $D_{\max}=D_{\mathrm{KL}}[Q\|P]+d$ bits, and refer to $d$ as the number of additional bits. We then vary the number of additional bits allowed for each algorithm, and estimate the bias of the resulting samples by evaluating the KL divergence between the empirical and the exact target distribution, that is $\smash{D_{\mathrm{KL}}[\hat{Q}\|Q]}$. To estimate this bias, we follow the method of Pérez-Cruz (2008). For each datapoint shown we draw 200 samples $X\sim\hat{Q}$ and use these to estimate $\smash{D_{\mathrm{KL}}[\hat{Q}\|Q]}$. We then repeat this for 10 different random seeds, reporting the mean bias and standard error in the bias, across these 10 seeds. Generally we find that the bias of GRCD is higher than that of AD∗ . This is likely because AD∗ is implicitly performing importance sampling over a set of $2^{D_{\max}+d}-1$ samples, and returning the one with the highest importance weight. By contrast, GRCD is running rejection sampling up to a maximum of $D_{\max}+d$ steps, returning its last sample if it has not terminated by its $(D_{\max}+d)^{\text{th}}$ step. While it might be possible to improve the bias of depth limited GRCD by considering an alternative way of choosing which sample to return, using for example an importance weighting criterion, we do not examine this here and leave this possibility for future work. Figure 5: Bias of depth-limited AD∗ and GRCD, as a function of the number of additional bit budget given to each algorithm. See text above for discussion.
# Optimal Moments on Redundancies in Noisy Parallel Computing Setup Sahasrajit Sarmasarkar<EMAIL_ADDRESS>Stanford UniversityUSA43017-6221 and Harish Pillai<EMAIL_ADDRESS>Indian Institute of Technology, BombayIndia ###### Abstract. We consider the problem of job assignment where a master server aims to compute some tasks and is provided a few child servers to compute under a uniform straggling pattern where each server is equally likely to straggle. We distribute tasks to the servers so that the master is able to receive most of the tasks even if a significant number of child servers fail to communicate. We first show that all balanced assignment schemes have the same expectation on the number of distinct tasks received and then study the variance. The variance or the second moment is a useful metric to study as there could be a high variation in the number of distinct tasks received. We show constructions using a generalization of “Balanced Incomplete Block Design”(BOSE, 1939; Sprott, 1955) minimizes the variance, and constructions based on repetition coding schemes attain the largest variance. Both minimum variance and maximum variance attaining designs have their own use cases depending on whether the master aims for a heavy-tailed or light-tailed distribution on the number of distinct jobs. We further show the equivalence between job and server-based assignment schemes when the number of jobs and child servers are equal. Block Designs, Repetition Codes, Coding Theory, Straggler Mitigation and Parallel Computing ## 1\. Introduction A distributed computing framework where the computation is done on multiple machines has been widely used for large-scale computations in (Azu, [n. d.]; AWS, [n. d.]; Goo, [n. d.]). This framework allows us to utilize the computation resources and memory of multiple machines often referred to as workers. Under a common and simple implementation of distributed computing, the master server divides the computation tasks to multiple child servers (workers). After each worker finishes its computation, it shares its results with the master server. The master aggregates the results received from different workers to finish its task. However, in real noisy communication frameworks, a subset of servers (workers) can be arbitrarily slow (often referred to as the stragglers) compared to the rest of the workers. Lately, there has been work to mitigate the issue of stragglers by introducing redundancies (Karakus et al., 2019; Aktas and Soljanin, 2019; Joshi et al., 2017). Coding theoretic techniques have often been used to introduce redundancies for straggler mitigation described extensively in (Li and Avestimehr, 2020) and used to split the data and assign different data parts to different servers. Distributed coding framework has been used in gradient computation (Tandon et al., 2017), matrix-matrix multiplication (Lee et al., 2017; Yu et al., 2017), polynomial computation (Yu et al., 2019) and convolution coding(Dutta et al., 2017) using techniques from coding theory namely MDS codes(Lee et al., 2017; Ferdinand and Draper, 2018), LDPC codes(Maity et al., 2019) and rateless codes(Mallick et al., 2020). The master server aggregates the computations transmitted by the non-straggling workers to compute the desired result. However, as happens typically in most modern cloud computing systems like Amazon EC2, some servers can operate with a significantly high throughput (Ananthanarayanan et al., 2013a; Zhang et al., 2013; Zhao et al., 2014). Statistical knowledge of communication and computation latency of each server can be used to design better assignment schemes and allocate the tasks more efficiently as studied in(Wang et al., 2019a; Sun et al., 2019; Yang et al., 2019). On a side note, there may be several scenarios where it may not be possible to split a computing task to multiple sub-parts, and job cloning is often used for straggler mitigation in such scenarios as studied in (Chen et al., 2014; Ananthanarayanan et al., 2013b; Joshi et al., 2015, 2017; Szajda et al., 2005). This is also popular in cloud computing (Joshi et al., 2017; Wang et al., 2015, 2014) where one task is assigned to multiple servers to combat the stragglers to obtain a low compute time. In this paper, we broadly study this problem of job assignment of multiple cloned jobs to multiple servers treating each server with identical communication latency where each server has a significant probability of straggling. These assignment schemes may be particularly useful in framework where there is high noise in the communication channel between the child servers and master server. ## 2\. Related work Block designs have been widely used in experiment design (Bose and Nair, 1939; Addelman, 1969; Shieh and Jan, 2004) where experiment designs are grouped into blocks and random treatments are applied to each block. Recently block designs and its variants have been used to construct LDPC codes (Ammar et al., 2004), gradient coding (Kadhe et al., 2019; Sakorikar and Wang, 2021) and error correcting codes (Smith, 1968). The most common amongst these are 2-designs or as often called balanced incomplete block designs (BIBDs), which are designs where every pair of points occur together in the same number of blocks (Bose and Nair, 1939; Colbourn and Dinitz, 2006). It has been shown that 2-designs (if exist) uniquely attain A-, D- and E optimality for experiment design (Kiefer, 1975; Kshirsagar, 1958), which was further generalized in (Yeh, 1986). However, our framework is different where we study variance on the number of distinct tasks received at the the master and show that repetition coding and a generalization of BIBDs can be utilized to attain the largest and the least variance respectively. In another direction, assignment policies of tasks to different servers under a distributed server system have been studied, check (Semchedine et al., 2011) for a survey on the same. Typically tasks arrive to a distributed server system stochastically in real time and this system typically distributes the tasks to various servers to minimise the response time as studied in (Colajanni et al., 1998; Harchol-Balter, 2000; Harchol-Balter et al., 1999). Recent observations show that the distribution of tasks follows a heavy-tailed distribution (Crovella and Bestavros, 1997; Arlitt and Williamson, 1996; Crovella et al., 1998; Williams et al., 2005), and designing assignment policies suited for load balancing across servers becomes difficult. In many cases, it may not be possible to recover the entire result but only partially recover the result as studied in coding theory (Babu and Kumar, 2015; Korhonen and Frossard, 2009). This setup has also been studied in distributed computing setups where the aim is not to recover the exact result but an approximation of the result suffices. This setup has been studied for matrix computation (Ozfatura et al., 2021) and for gradient coding (Sarmasarkar et al., 2022; Wang et al., 2019b). Our problem follows a similar setup where the master only aims to recover a high fraction of the jobs as it may be difficult for the server to recover all tasks in a noisy environment where a substantial fraction of servers may straggle. ## 3\. Our contributions In this paper, we study a setup where a master aims to compute $n$ jobs and has $c$ identical servers to do the computations. In our model, the arrival and assignment of tasks are not in real-time but instead, the set of tasks (jobs) to be computed and the set of servers is given apriori. Jobs are appropriately replicated and can be viewed as a way to mitigate the issue of slow (straggling) workers. We study a homogeneous setup treating all the servers as having identical computing and stochastic properties, that is, balanced schemes with each server being assigned $k$ jobs and each job being assigned to $r$ servers. We further assume that each server is equally likely to straggle. We consider a setup where a non-straggling worker can successfully transmit all tasks to the master as studied in (Tandon et al., 2017; Joshi et al., 2017; Ozfatura et al., 2019). We aim to design coding schemes where the master aims to maximise the number of computed distinct jobs that are received. Towards this aim, we study the mean and variance of the number of distinct jobs received in Section 5. Our contributions can be described as follows. a) We first show that when any set of $x$ servers is equally likely to straggle, then for every balanced assignment, the expected number of completed jobs that the master receives is the same (Theorem 1) and study the variance of the number of jobs received within the class of balanced assignment schemes. b) Typically, assignment schemes with the largest variance may be useful for systems where we aim to increase the likelihood of the master receiving a large number (or all) of jobs (heavier mass at tails), and schemes with the least variance may be useful for systems where one aims to reduce the likelihood of the master receiving a small number of jobs (lighter mass at the tails). This follows since every balanced assignment scheme has the same expectation on the jobs received. c) We show that certain special balanced assignments (called proximally compact and stretched compact designs) when they exist, are guaranteed to attain the least variance and largest variance respectively (Theorem 5 and Theorem 8 in Section 6). d) We show how our results generalise to the case where $x$ is sampled from a distribution. This would imply that results on least variance and largest variance would continue to hold when each worker is independently and equally likely to straggle with probability $p$ (Theorems 1, 2 and 3 in Section 7). e) Finally, we show that when the number of jobs equals the number of servers, proximally compact job assignments and server assignments (replacing the roles of jobs and servers) are identical and both of them attain the least variance in Theorem 2 in Appendix A. ## 4\. Preliminaries and Notation We consider a setup where a master has $n$ jobs to compute and has $c$ servers to do the computations. We further assume a noisy scenario where only a fraction of these servers are able to communicate back to the master. Therefore, for increasing reliability, redundancies are introduced in the setup by assigning each job to multiple servers. In this paper, we examine the expectation and variance on the number of distinct completed jobs that the master receives from the servers that were able to communicate, for various assignment schemes. In particular, we study assignment schemes that achieve certain desired variances on the number of received distinct jobs. Given a set of $n$ jobs and $c$ servers, we study various assignments of jobs to different servers by the master server. More formally, let us denote the $n$ jobs by $\mathcal{A}=\\{a_{1},\ldots,a_{n}\\}$ and $c$ servers by $\mathcal{S}=\\{s_{1},s_{2},\ldots,s_{c}\\}$. Any assignment ($D$) of jobs in $\mathcal{A}$ to servers in $\mathcal{S}$, can be equivalently represented by a bipartite graph $\mathcal{G}_{D}$ where the nodes denote the jobs and the servers. The edges of the graph exist between nodes representing job $a_{i}$ and server $s_{j}$ if job $a_{i}$ is assigned to server $s_{j}$. Alternately, for a job assignment ${D}$, we can define an assignment matrix $A_{{D}}\in\\{0,1\\}^{n\times c}$ as given below. ###### Definition 1. (Construction of $A_{D}$): Given an assignment of jobs in $\mathcal{A}$ to servers in $\mathcal{S}$, we define matrix $A_{D}\in\\{0,1\\}^{n\times c}$ as follows. : $\displaystyle{}A_{{D}}[i,j]$ $\displaystyle=1\text{ if job $a_{i}$ is assigned to server $s_{j}$}$ (1) $\displaystyle=0\text{ otherwise}$ Observe that the matrix $A_{D}$ represents the adjacency matrix for the bipartite graph $\mathcal{G}_{D}$. We specifically study balanced assignment schemes where each server is assigned the same number of jobs and each job is assigned to the same number of servers. More formally, we define this as follows. ###### Definition 2. (Balanced $(n,k,r,c)$ assignment): Given a set of $n$ jobs and $c$ servers, we call an assignment scheme of jobs a balanced $(n,k,r,c)$ assignment if the following conditions are satisfied. * • Each server is assigned precisely $k$ distinct jobs to compute. * • Each job is assigned to precisely $r$ distinct servers. Note that this assignment scheme ensures that $n\times r=k\times c$. We can equivalently define it in terms of matrix $A_{D}$ as follows. ###### Definition. (Balanced $(n,k,r,c)$ assignment in terms of $A_{D}$): Given a set of $n$ jobs and $c$ servers, we call the assignment scheme $D$ of jobs to servers a balanced $(n,k,r,c)$ assignment if each row of $A_{D}$ sums up to $r$ and each column sums up to $k$. Let us look at an example of a balanced $(9,3,2,6)$ assignment scheme. ###### Example 0. We describe a balanced assignment scheme with 9 jobs $\\{a_{1},a_{2},\ldots,a_{9}\\}$ and 6 servers $\\{s_{1},s_{2},\ldots,s_{6}\\}$ in Table 1. Note that each job is assigned to precisely $2$ servers and each server has exactly 3 jobs to compute. The assignment scheme is motivated from a cyclic assignment scheme. Jobs Servers | $s_{1}$ | $s_{2}$ | $s_{3}$ | $s_{4}$ | $s_{5}$ | $s_{6}$ ---|---|---|---|---|---|--- $a_{1}$ | 1 | 1 | | | | $a_{2}$ | | 1 | 1 | | | $a_{3}$ | | | 1 | 1 | | $a_{4}$ | | | | 1 | 1 | $a_{5}$ | | | | | 1 | 1 $a_{6}$ | 1 | | | | | 1 $a_{7}$ | 1 | 1 | | | | $a_{8}$ | | | 1 | 1 | | $a_{9}$ | | | | | 1 | 1 Table 1. Assignment of jobs to various servers in a balanced $(9,3,2,6)$ assignment scheme ## 5\. The Mean and the Variance We consider the number of distinct jobs $d$ received at the master when only a subset of $x$ servers manage to communicate with the master. We consider any subset of $\mathcal{S}$ with cardinality $x$ to be equally likely be the set of servers that communicates with the master. Note that with this definition, if $\hat{S}\subseteq\mathcal{S}$ (with $|\hat{S}|=x$) is the subset of servers that communicate with the master, then we can denote the number of distinct jobs received $d=|\cup_{j\in\hat{S}}\text{supp}(A_{D}[:,j])|$ where $\text{supp}(v)$ denotes the indices of the non-zero entries of the vector $v$. Now, consider the uniform distribution over all subsets of servers of cardinality $x$ which we denote by $\mathfrak{D}_{\mathcal{S},x}$ i.e. a sample from this distribution returns any subset of $\mathcal{S}$ of cardinality $x$ with probability $\frac{1}{{{|\mathcal{S}|\choose x}}}$. For a given assignment ${D}$ of jobs to servers, we denote the expectation of the number of distinct completed jobs received by the master when any subset of $x$ servers is able to communicate with master uniformly at random by $\mathbbm{E}_{{D},x}[d]$ and the corresponding variance by $\sigma_{{D},x}[d]$. The expectation and the variance on the number of distinct received jobs $d$ may be written as (2) ${}\mathbbm{E}_{D,x}[d]=\mathbbm{E}_{\hat{S}\sim\mathfrak{D}_{\mathcal{S},x}}\left[\left|\bigcup_{j\in\hat{S}}\text{supp}(A_{D}[:,j])\right|\right]\text{ and }\sigma_{D,x}[d]=\sigma_{\hat{S}\sim\mathfrak{D}_{\mathcal{S},x}}\left[\left|\bigcup_{j\in\hat{S}}\text{supp}(A_{D}[:,j])\right|\right]$ Note that the randomness in this setup is only in the set of servers that can communicate with the master. The assignment scheme has no randomness associated with it. Theorem 1 states that the expectation on the number of distinct jobs $\mathbb{E}_{D,x}[d]$ is the same for every balanced $(n,k,r,c)$ assignment. This expectation is a function of $n,k,r,c$ and $x$ and is independent of the specific balanced assignment $D$ we choose. Throughout the remainder of this paper, $\mathfrak{n}^{D}_{i,\hat{S}}$ denotes the number of servers in $\hat{S}$ to which job $a_{i}$ is assigned under the assignment scheme $D$. Observe that $\mathfrak{n}^{D}_{i,\hat{S}}$ can take any value from $0$ to $r$. ###### Theorem 1. Consider any balanced $(n,k,r,c)$ assignment $D$. The expectation of the number of distinct completed jobs $d$ received by the master when any subset of cardinality $x$ of the set of servers $\mathcal{S}$ is able to communicate with the master with equal probability is the same for every balanced $(n,k,r,c)$ assignment $D$ and is given by (3) $\mathbbm{E}_{{D},x}[d]=n\cdot\left(1-\frac{{c-r\choose x}}{{c\choose x}}\right)$ We present a proof sketch below. A detailed proof is presented in Appendix B. ###### Proof Sketch. The number of distinct jobs $d$ received by the master when servers in a subset $\hat{S}$ (with $|\hat{S}|=x$) is able to communicate with the master is given by (4) ${}d=\left|\bigcup_{j\in\hat{S}}\text{supp}(A_{D}[:,j])\right|=\left(k\times x-\sum\limits_{i=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}\right)$ Note that the term $\sum\limits_{i=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}$ excludes those jobs which have been received multiple times from various servers present in $\hat{S}$. (5) $\displaystyle{}\mathbb{E}_{D,x}[d]=\frac{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(k\times x-\sum\limits_{i=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})}{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}1}{=}$ $\displaystyle k\times x-\frac{n\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}}{{c\choose x}}$ Observe that for every job $a_{i}$ in a balanced $(n,k,r,c)$ assignment, the quantity $\sum\limits_{\hat{S}\subset\mathcal{S},|\hat{S}|=x}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}$ is the same, i.e., this summation is independent of $i$. We now show that the quantity $\sum\limits_{\hat{S}\subset\mathcal{S},|\hat{S}|=x}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}$ for any specified $x$, is the same for every balanced $(n,k,r,c)$ distribution ${D}$. We compute this sum by counting the number of subsets $\hat{S}\subset\mathcal{S}$ of cardinality $x$ which additionally satisfies the constraint on $\mathfrak{n}^{D}_{i,\hat{S}}=t$ (i.e. job $a_{i}$ is present in exactly $t$ servers from $\hat{S}$) for every $t$ from $2$ to $r$ (as these cases deal with the job $a_{i}$ appearing more than once in the subset $\hat{S}$). Equation (46) follows from multiplying two binomial expressions 111Detailed proof is presented in Appendix B and considering their coefficients. $\displaystyle{}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}{=}\sum\limits_{t=1}^{r-1}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x,\mathfrak{n}^{D}_{i,\hat{S}}=t+1\end{subarray}}t\overset{(a)}{=}$ $\displaystyle\sum\limits_{t=1}^{r-1}t{r\choose(t+1)}{(c-r)\choose(x-t-1)}$ (7) ${}\sum\limits_{t=1}^{r-1}t{r\choose(t+1)}{(c-r)\choose(x-t-1)}=r\times{{c-1\choose x-1}}+{{c-r\choose x}}-{{c\choose x}}$ Combining equations (B), (B) and (46), we get the desired result. ∎ A few comments are in order here. Note that for $x=1$, the expectation (as expected) is precisely $k$. Observe that for $x>c-r$, the expectation goes to $n$. In other words, if the number of servers that successfully communicates with the master is greater than $(c-r)$, then the master obtains at least one copy of every job $a_{i}\in\mathcal{A}$. This follows since every job is assigned to exactly $r$ servers and therefore for any job to be missed out, the $r$ servers to which that specific job was assigned, should fail to communicate with the master. Thus, if any job is missed out, then the number of servers that manage to communicate with the master $x$ can at most be $c-r$. We now calculate the variance for the number of distinct jobs $d$ received at the master for any balanced $(n,k,r,c)$ job assignment $D$. From the comments in the previous paragraph, it is clear that $\sigma_{D,1}[d]=0$ for the case $x=1$, since the master always receives precisely $k$ distinct jobs, if only one server manages to communicate with the master. Similarly, $\sigma_{D,x}[d]=0$ for $x>c-r$, as the master would receive all the $n$ jobs if more than $c-r$ servers communicate (for the sake of completeness, we formally calculate this in Corollary 1). For calculating the variance on the number of distinct jobs $d$ received by the master, observe $\sigma_{D,x}(d)=\sigma_{D,x}\left(k\times x-\sum\limits_{i}(\mathfrak{n}^{D}_{i,\hat{S}}-1\right)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})=\sigma_{D,x}\left(\sum\limits_{i}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}\right)$ The above follows since $\sigma(t-X)=\sigma(X)$ where $t$ is a constant and $X$ is a random variable. We now make use of the definition $\text{var}(X)=\mathbbm{E}[X^{2}]-(\mathbbm{E}[X])^{2}$. Therefore, $\displaystyle{}\sigma_{D,x}\left(\sum\limits_{i}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}\right)=$ $\displaystyle\frac{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}{\left(\sum\limits_{i=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}\right)}^{2}}{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}1}-\left(\frac{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};|\hat{S}|=x\end{subarray}}{(\sum\limits_{i=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})}}{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}1}\right)^{2}$ $\displaystyle\overset{(a)}{=}\frac{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1})}{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}1}-\left(\frac{\sum\limits_{i=1}^{n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}{((\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})}}{{c\choose x}}\right)^{2}$ $\displaystyle\overset{(b)}{=}\frac{\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}}{{c\choose x}}-\left(\frac{n\sum\limits_{t=1}^{r-1}t{r\choose(t+1)}{(c-r)\choose(x-t-1)}}{{c\choose x}}\right)^{2}$ (8) $\displaystyle\overset{(c)}{=}\frac{\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}}{{c\choose x}}-\left(\frac{n{{c-r\choose x}}}{{{c\choose x}}}+n\times\left(\frac{rx}{c}-1\right)\right)^{2}$ In the above set of equations, $(a)$ follows from the identity $(\sum\limits_{i}b_{i})^{2}=\sum\limits_{i}\sum\limits_{j}b_{i}b_{j}$. The first term in $(b)$ is obtained by interchanging the order of summations, whereas the second term comes from equation (B). Further, the second term in $(c)$ follows using the equation (46) given in the proof of Theorem 1. Observe that the second term in the final expression in equation (5) depends only on $n,r,c$ and $x$ and is independent of the specific balanced $(n,k,r,c)$ job assignment $D$. On the other hand, the first term in equation (5) depends on the particular assignment $D$. We now consider the numerator of the first term of equation (5) in more detail. We can break this expression into two parts, where one part is dependent on just one index $i$ and the other part is dependent on two distinct indices $i,j$. Thus, $\displaystyle\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}$ (9) $\displaystyle=$ $\displaystyle 2\sum\limits_{1\leq i<j\leq n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}$ (10) $\displaystyle\hskip 80.00012pt+\sum\limits_{1\leq i=j\leq n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}$ In equation (5), the second term can be rewritten as $\sum\limits_{1\leq i\leq n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}((\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})^{2}$. For every job $a_{i}$, this expression calculates $\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}((\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})^{2}$, which is independent of the choice of the job $a_{i}$ in any balanced $(n,k,r,c)$ assignment $D$. In fact, this second term of equation (5) is independent of the choice of $D$ and it depends only on the values of $n,c,r$ and $x$. We can compute this sum by counting the number of subsets $\hat{S}\subset\mathcal{S}$ of cardinality $x$ that additionally satisfy the constraint $\mathfrak{n}^{D}_{i,\hat{S}}=t$ (i.e. job $a_{i}$ is present in exactly $t$ servers from $\hat{S}$) for every $t$ from $2$ to $r$. Thus (11) $\sum\limits_{1\leq i\leq n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}((\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})^{2}=n\sum\limits_{t=1}^{r-1}t^{2}{r\choose(t+1)}{(c-r)\choose(x-t-1)}$ Note that the number of subsets $\hat{S}\subset\mathcal{S}$ of cardinality $x$ such that a particular job $a_{i}$ appears $t+1$ times in $\hat{S}$ is given by ${r\choose(t+1)}{(c-r)\choose(x-t-1)}$. As $\mathfrak{n}^{D}_{i,\hat{S}}=t+1$ for this particular $\hat{S}$, therefore $((\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})^{2}=t^{2}$. This explains the final expression in equation (11). A closed form expression for the sum $\sum\limits_{t=1}^{r-1}t^{2}{r\choose(t+1)}{(c-r)\choose(x-t-1)}$ can be obtained by considering the following binomial expressions (12) ${}r(r-1)y^{2}(1+y)^{r-2}-ry(1+y)^{r-1}-1+{(1+y)^{r}}=\sum_{t=1}^{r-1}t^{2}{{r\choose t+1}}y^{t+1}$ (13) ${}(1+y)^{c-r}=\sum_{v=0}^{c-r}{{c-r\choose v}}y^{v}$ Multiplying equations (12) and (13), one obtains $\sum\limits_{t=1}^{r-1}t^{2}{{r\choose t+1}}{{c-r\choose x-t-1}}$ to be the coefficient of $y^{x}$ in $r(r-1)y^{2}(1+y)^{c-2}-ry(1+y)^{c-1}-(1+y)^{c-r}+(1+y)^{c}$, thus, (14) ${}\sum\limits_{t=1}^{r-1}t^{2}{{r\choose t+1}}{{c-r\choose x-t-1}}=r(r-1){{c-2\choose x-2}}-r{{c-1\choose x-1}}-{{c-r\choose x}}+{{c\choose x}}$ Finally, we analyse the first term in equation (5), viz., $\sum\limits_{1\leq i<j\leq n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}$. If a specific pair of jobs $a_{i},a_{j}$ appear $(\alpha+1)$ and $(\beta+1)$ times respectively in some subset $\hat{S}\subset\mathcal{S}$ of cardinality $x$, then such a pair of jobs contribute $\alpha\beta$ towards this expression that we are analysing. One needs to add up such contributions from every distinct pair of jobs $(a_{i},a_{j})$ and every subset $\hat{S}\subset\mathcal{S}$ of cardinality $x$ to get the final value of this expression. The strategy that we adopt to compute this sum is as follows : we find $\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}$ for any given pair of distinct jobs $(a_{i},a_{j})$. Observe that this expression depends on how the pair of jobs $(a_{i},a_{j})$ are distributed amongst the $c$ servers, which in turn depends on the particular balanced $(n,k,r,c)$ job assignment $D$ that is under consideration. Now, given a particular pair of jobs $(a_{i},a_{j})$, how they are farmed to the servers can essentially differ only in the number of servers that are assigned both the jobs $a_{i},a_{j}$ simultaneously. The number of servers that are simultaneously assigned both the jobs $(a_{i},a_{j})$ can range from $0$ to $r$. If a pair of jobs $(a_{i},a_{j})$ are assigned together to precisely $m$ servers (with $0\leq m\leq r$), then the sum $\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}$ calculated for this particular pair of jobs is precisely equal to the corresponding sum for every other pair of jobs that are assigned together to precisely $m$ servers. We use the notation (15) ${}g(m,x)=\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}$ to indicate this particular sum that arises from a pair of jobs $a_{i},a_{j}$ that are assigned together to precisely $m$ servers. We now show that the values of $g(m,x)$ depends only on $c,r,m$ and $x$. We give the expression for $g(0,x)$ in Lemma 2 and give a recursion to compute $g(m,x)$ in Lemma 3. ###### Lemma 0. For a balanced $(n,k,r,c)$ assignment, the value of $g(0,x)$ is given by (16) $\displaystyle{}g(0,x)=$ $\displaystyle r^{2}{{c-2\choose x-2}}-2r{{c-1\choose x-1}}+{{c\choose x}}-2{{c-r\choose x}}+2r{{c-r-1\choose x-1}}+{c-2r\choose x}$ ###### Lemma 0. For a balanced $(n,k,r,c)$ assignment, the values for $g(m,x)$ are related in the following fashion (17) ${}g(m+1,x)-g(m,x)={c-2\choose x-1}-2{c-r-1\choose x-1}+{c-2r+m\choose x-1}$ We prove these Lemmas in Appendix E and F respectively using a careful application of techniques from combinatorics. Note that all the expressions for $g(m,x)$ depends on the values of $c,r,m$ and $x$ and is therefore independent of which balanced $(n,k,r,c)$ assignment $D$ we choose. Observe further that the expression for $g(m+1,x)-g(m,x)$ is an increasing function of $m$ when $x$ is fixed. This is clear from the fact that only the last term in (17) depends on $m$. Of course, if $x\leq c-2r$, then we can further conclude that $g(m+1,x)-g(m,x)$ is a strictly increasing function of $m$. We now define $\mathfrak{m}^{D}(m)$ as the number of distinct pairs of jobs $(a_{i},a_{j})$ with $1\leq i<j\leq n$ that are assigned together to precisely $m$ servers in the balanced $(n,k,r,c)$ assignment $D$. One can formally define this number for a specific balanced $(n,k,r,c)$ assignment $D$ using the assignment matrix $A_{D}$ as (18) ${}\mathfrak{m}^{D}(m)=\sum_{\begin{subarray}{c}(i_{1},i_{2})\\\ 1\leq i_{1}<i_{2}\leq n\end{subarray}}\mathbbm{1}_{\sum_{j=1}^{c}A_{D}[i_{1},j]A_{D}[i_{2},j]=m}$ Given a balanced $(n,k,r,c)$ assignment $D$, the numbers $\mathfrak{m}^{D}(m)$ have some additional properties (19) $\sum_{m=0}^{r}\mathfrak{m}^{D}(m)={n\choose 2}$ (20) $\sum_{m=0}^{r}m\mathfrak{m}^{D}(m)=c{k\choose 2}$ Equation (19) follows from the fact that there are a total of $n$ jobs and thus the total number of job pairs is given by ${{n\choose 2}}$. Equation (20) follows from the fact that the number of pairs of jobs that are assigned together to a fixed server $s_{i}$ is given by ${{k\choose 2}}$. Summing over all the servers in $\mathcal{S}$ gives us the RHS in (20). Note that we count each pair of jobs as many times as they appear together in a server and thus we say $\sum_{m=0}^{r}m\mathfrak{m}^{D}(m)=c{k\choose 2}$. Observe that the first term of equation (5) that we are evaluating can now be rewritten as (21) $\sum\limits_{1\leq i<j\leq n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}=\sum_{m=0}^{r}\mathfrak{m}^{D}(m)g(m,x)$ Putting all of the discussion together, we therefore conclude the following regarding evaluation of variance $\sigma_{D,x}(d)$ : ###### Theorem 4. Consider any assignment $D$ amongst balanced $(n,k,r,c)$ assignments. The variance on the number of distinct jobs $d$ $(\sigma_{D,x}(d))$ received by the master when any subset of servers $\mathcal{S}$ of cardinality $x>1$ is able to communicate to the master with equal probability, is equal to (22) ${}\sigma_{{D},x}(d)=\frac{2\sum\limits_{m=0}^{r}\mathfrak{m}^{D}(m)g(m,x)+T_{2}(n,r,c,x)}{{c\choose x}}-\left(T_{1}(n,r,c,x)\right)^{2}\\\ $ where $T_{1}(n,r,c,x)=\frac{n{{c-r\choose x}}}{{{c\choose x}}}+n\times(\frac{rx}{c}-1)$ and (23) $\displaystyle T_{2}(n,r,c,x)=$ $\displaystyle n\sum\limits_{t=1}^{r-1}t^{2}{r\choose(t+1)}{(c-r)\choose(x-t-1)}$ (24) $\displaystyle=$ $\displaystyle n\left(r(r-1){{c-2\choose x-2}}-r{{c-1\choose x-1}}-{{c-r\choose x}}+{{c\choose x}}\right)$ Thus, we have shown that while the mean of the number of received jobs $d$ is the same for all balanced $(n,k,r,c)$ assignments, the variance of $d$ is dependent on the frequency distribution of job pairs assigned to the same server. ## 6\. Results on extreme variance based on $[\mathfrak{m}^{D}(m)]_{m=0}^{r}$ It is clear from the previous sections, that while all $(n,k,r,c)$ balanced job assignments give the same mean, they display different variances based on the values of $\mathfrak{m}^{D}(m)$ that arise in the respective job assignments. Here $\mathfrak{m}^{D}(m)$ refers to number of pairs of distinct jobs assigned together to precisely $m$ servers as defined in (18). We now explore the range of variances that balanced $(n,k,r,c)$ job assignments can attain. ###### Definition 2. Given a balanced $(n,k,r,c)$ job assignment $D$, we define a shape vector $h_{D}\in\mathbb{N}^{r+1}$ associated to the assignment $D$ as (25) ${}h_{D}=[\mathfrak{m}^{D}(0),\mathfrak{m}^{D}(1),\ldots\mathfrak{m}^{D}(r)]^{T}$ Thus the shape vector lists the number of pairs of jobs that are assigned together to $m$ servers, for $m=0,1,\cdots,r$ and is a vector of length $r+1$. Clearly, the entries of $h_{D}$ are all non-negative integers. Two distinct balanced $(n,k,r,c)$ job assignments $D_{1},D_{2}$ would have the same mean and variance if and only if the corresponding shape vectors $h_{D_{1}}$ and $h_{D_{2}}$ are the same. We first characterize all possible candidate shape vectors in $\mathbb{N}^{r+1}$.222In our notation, $\mathbb{N}$ is the set of all non- negative integers including 0. Since the entries of the shape vectors are $\mathfrak{m}^{D}(m)$ arising out of balanced $(n,k,r,c)$ job assignment, they must satisfy equations (19) and (20). These can be rewritten as (26) ${}\begin{bmatrix}1&1&1&\cdots&1\\\ 0&1&2&\cdots&r\end{bmatrix}\begin{pmatrix}\mathfrak{m}^{D}(0)\\\ \mathfrak{m}^{D}(1)\\\ \vdots\\\ \mathfrak{m}^{D}(r)\end{pmatrix}=Hh_{D}=\begin{pmatrix}{n\choose 2}\\\ c{k\choose 2}\end{pmatrix}$ Here the matrix $H=\begin{bmatrix}1&1&1&\cdots&1\\\ 0&1&2&\cdots&r\end{bmatrix}$. Thus, two possible shape vectors differ by a vector in the kernel of the matrix $H$. Therefore, if one has a particular balanced $(n,k,r,c)$ job assignment $D$ with the corresponding shape vector $h_{D}$, then all other possible shape vectors can be characterized as vectors $(h_{D}+v)\in\mathbb{N}^{r+1}$ where $v\in\ker H$. A basis for the kernel of the matrix $H$ is given by the $r-1$ vectors (27) ${}\left\\{h_{1},h_{2},h_{3},\cdots,h_{r-2},h_{r-1}\right\\}=\left\\{\begin{pmatrix}1\\\ -2\\\ 1\\\ 0\\\ \vdots\\\ 0\\\ 0\\\ 0\end{pmatrix},\begin{pmatrix}0\\\ 1\\\ -2\\\ 1\\\ \vdots\\\ 0\\\ 0\\\ 0\end{pmatrix},\begin{pmatrix}0\\\ 0\\\ 1\\\ -2\\\ \vdots\\\ 0\\\ 0\\\ 0\end{pmatrix},\cdots,\begin{pmatrix}0\\\ 0\\\ 0\\\ 0\\\ \vdots\\\ -2\\\ 1\\\ 0\end{pmatrix},\begin{pmatrix}0\\\ 0\\\ 0\\\ 0\\\ \vdots\\\ 1\\\ -2\\\ 1\end{pmatrix}\right\\}$ Note that each of these basis vectors $h_{i}$ have only three nonzero entries. We make use of these basis vectors in determining extremal values of variances that a balanced $(n,k,r,c)$ job assignment can attain. Based on the shape vectors, we now define certain special kinds of balanced $(n,k,r,c)$ job assignments. ###### Definition 2. A balanced $(n,k,r,c)$ assignment $D$ is compact if the corresponding shape vector $h_{D}$ has at most two non-zero elements. Under certain special conditions, there is a possibility that the shape vector $h_{D}$ has only one nonzero entry. Of course, in this special case, every possible pair of jobs is assigned together to $m$ servers, where $m=\frac{ck(k-1)}{n(n-1)}=r\frac{k-1}{n-1}$. Clearly, the dependence of $m$ on the values of $n,k,r$ forces such a possibility to be rare. So in general, compact balanced $(n,k,r,c)$ job assignments have two nonzero entries. ###### Definition 2. A balanced assignment $D$ is proximally compact if the shape vector $h_{D}$ has either exactly one nonzero entry or has exactly two consecutive nonzero entries. ###### Lemma 1. For proximally compact $(n,k,r,c)$ assignment $D$, we have $\ell=\Bigl{\lfloor}\frac{r(k-1)}{n-1}\Bigr{\rfloor}$ where $\ell$ denotes the index of the smallest non-zero entry in the shape vector $h_{D}$. ###### Proof. For a proximally compact $(n,k,r,c)$ assignment $D$, if the shape vector has only one nonzero entry, then $\mathfrak{m}^{D}(\ell)={n\choose 2}$. As the total number of job pairs at the $c$ servers is $c{k\choose 2}$, therefore using (20), we have $\ell=\frac{c{k\choose 2}}{{n\choose 2}}=\frac{r(k-1)}{n-1}$. On the other hand, if a proximally compact $(n,k,r,c)$ assignment $D$, has a shape vector with exactly two consecutive nonzero entries, then $\mathfrak{m}^{D}(m)$ is zero for all $m\neq\ell,\ell+1$ for some $\ell$. Using (19), we have $\mathfrak{m}^{D}(\ell+1)={{n\choose 2}}-\mathfrak{m}^{D}(\ell)$ and by (20), we get $\ell(\mathfrak{m}^{D}(\ell))+(\ell+1)({{n\choose 2}}-\mathfrak{m}^{D}(\ell))=c{{k\choose 2}}$. Thus we can conclude that $\ell{{n\choose 2}}\leq c{{k\choose 2}}$ and $(\ell+1){{n\choose 2}}>c{{k\choose 2}}$ (as $\mathfrak{m}^{D}(\ell)>0$) and therefore $\ell=\Bigl{\lfloor}\frac{c.{{k\choose 2}}}{{{n\choose 2}}}\Bigr{\rfloor}=\Bigl{\lfloor}\frac{r(k-1)}{n-1}\Bigr{\rfloor}$. ∎ Given $n,k,r,c$ it is not clear whether a balanced $(n,k,r,c)$ job assignment which is proximally compact exists. By Lemma 1, it is clear that there can be only one shape vector in $\mathbb{N}^{r+1}$ that satisfies the condition of proximal compactness. One can therefore conclude that there is at most only one proximally compact balanced $(n,k,r,c)$ job assignment. The special case of the shape vector having only one nonzero entry corresponds to the case of balanced incomplete block designs (BIBD) (BOSE, 1939; Colbourn and Dinitz, 2006). Balanced incomplete block designs is a very well studied subject. For the sake of completeness, we give the definition of BIBD below. ###### Defintion 3. (BIBD $(v,b,r,k,\lambda)$ scheme as in (Colbourn and Dinitz, 2006)) - A balanced incomplete block design (BIBD) is a pair $(V,B)$ where V is a $v$-set and B is a collection of $b$ $k$-sized subsets of $V$ (blocks) such that each element of $V$ is contained in exactly $r$ blocks and any 2-subset of V is contained in exactly $\lambda$ blocks. Note that we can associate the set $V$ to the set of jobs $\mathcal{A}$. Thus $v$ is the same as $n$ that we have employed so far. Each $k$ sized subset of $V$ (or $\mathcal{A}$) can be identified to the set of jobs assigned to a server. The number $r$ has the same interpretation as in our case. Since $B$ is a collection of $b$ $k$-sized sets, we can think of the number of servers $c$ being equal to $b$. Finally $\lambda=\frac{r(k-1)}{n-1}$. Thus $(n,k,r,c)$ in our case is the same as $(v,b,r,k,\lambda)$ quoted in the definition of BIBD above. Even in cases where $n,k,r$ may lead to a $\lambda$ which is a positive integer, it is not known whether a BIBD always exists. However, multiple constructions of BIBD for various parameters have been described in (BOSE, 1939) using various techniques, like vector sub-spaces over finite fields etc. The famous Bruck-Ryser-Chowla theorem in (Sprott, 1955) gives some necessary conditions on $n,k,r$ that guarantees the existence of a BIBD. Proximally compact assignments may be thought of as a generalization of BIBDs that do not insist on a unique number $\lambda$, that represents the number of servers to be shared by every pair of jobs. Instead proximally compact assignments allow every pair of jobs to be assigned to either $\ell$ or $\ell+1$ servers. As remarked earlier, Lemma 1 ensures that given $n,k,r,c$ there is a unique shape vector that can be constructed and therefore at most only one proximally compact assignment can exist. We now provide an example of a proximally compact assignment scheme that is not a BIBD. ###### Example 0. Consider balanced $(9,3,3,9)$ assignment schemes. In this case, $\frac{r(k-1)}{n-1}=\frac{3}{4}$ and so there is no BIBD possible. Further, $\ell=0$ and the corresponding shape vector for a possible proximally compact assignment should be $h_{D}=[9,27,0,0]^{T}$. We display an assignment scheme in Table 2 whose shape vector is indeed $h_{D}$. Note that in this scheme, 27 pairs of jobs are assigned together to a server once and there are 9 pairs of jobs that were never assigned together. Jobs Servers | $s_{1}$ | $s_{2}$ | $s_{3}$ | $s_{4}$ | $s_{5}$ | $s_{6}$ | $s_{7}$ | $s_{8}$ | $s_{9}$ ---|---|---|---|---|---|---|---|---|--- $a_{1}$ | 1 | 1 | 1 | | | | | | $a_{2}$ | | | | 1 | 1 | 1 | | | $a_{3}$ | | | | | | | 1 | 1 | 1 $a_{4}$ | 1 | | | 1 | | | 1 | | $a_{5}$ | | 1 | | | 1 | | | 1 | $a_{6}$ | | | 1 | | | 1 | | | 1 $a_{7}$ | 1 | | | | 1 | | | | 1 $a_{8}$ | | 1 | | | | 1 | 1 | | $a_{9}$ | | | 1 | 1 | | | | 1 | Table 2. Assignment of jobs to servers in a proximally minimally compact $(9,3,3,9)$ assignment scheme ###### Theorem 5. If a proximally compact balanced $(n,k,r,c)$ job assignment exists, then it has the least variance amongst all balanced $(n,k,r,c)$ job assignments. A proof sketch is presented below. A detailed proof is given in Appendix C. Recall that in our notation $\mathbb{N}$ denotes the set of all non-negative integers including 0. ###### Proof Sketch. Let $h_{D}$ be the shape vector corresponding to the proximally compact balanced $(n,k,r,c)$ job assignment. Thus $h_{D}(i)=0$ for all $i\leq\ell$ and $i>\ell+2$ for $\ell$ as calculated in Lemma 1. Any balanced $(n,k,r,c)$ job assignment $D_{1}$ would have a shape vector $h_{D}+v$ where $v\in\ker H$ with matrix $H$ as defined in (26). Observe from the expression of variance in (22) that it is only the term $\sum\limits_{m=0}^{r}\mathfrak{m}^{D}(m)g(m,x)$ that varies amongst the different balanced $(n,k,r,c)$ assignments. Thus it is enough to show that for every permissible $v\in\ker H$ mentioned above, $\sum\limits_{m=0}^{r}v(m+1)g(m,x)\geq 0$, in order to conclude that the proximally compact balanced $(n,k,r,c)$ assignment has the least variance. We therefore first characterize $v\in\ker H$ that may appear from some balanced $(n,k,r,c)$ assignment. As both the shape vectors $h_{D},h_{D}+v\in\mathbb{N}^{r+1}$, therefore $v(i)\geq 0$ for all $i\neq\ell+1,\ell+2$. Further, as $v\in\ker H$, therefore $\sum\limits_{i=1}^{r+1}v(i)=0$ and so if $v$ is a nonzero vector, then at least one of $v(\ell+1),v(\ell+2)$ must be a negative integer. As $v\in\ker H$, therefore $v$ can be expressed in terms of the basis vectors $\left\\{h_{i}\right\\}$ listed in (27). Let $v=\sum\limits_{i=1}^{r-1}\alpha_{i}h_{i}$ and we show 333A detailed proof proving non-negativity of $\alpha_{i}$ is in Appendix C that all $\alpha_{i}\in\mathbb{N}$. As a result, we get $\displaystyle\sum\limits_{m=0}^{r}v(m+1)g(m,x)=\sum\limits_{i=1}^{r-1}\alpha_{i}\left(g(i-1,x)-2g(i,x)+g(i+1,x)\right)$ By Lemma 3, we know that $g(m+1,x)-g(m,x)={{c-2}\choose{x-1}}-2{{c-r-1}\choose{x-1}}+{{c-2r+m}\choose{x-1}}$ and therefore (28) $\displaystyle g(i-1,x)-2g(i,x)+g(i+1,x)$ $\displaystyle=$ $\displaystyle\left(g(i+1,x)-g(i,x)\right)-\left(g(i,x)-g(i-1,x)\right)$ $\displaystyle=$ $\displaystyle{{c-2r+i}\choose{x-1}}-{{c-2r+i-1}\choose{x-1}}$ $\displaystyle=$ $\displaystyle\frac{x-1}{c-2r+i-x+1}{{c-2r+i-1}\choose{x-1}}\geq 0$ Thus we have (29) $\displaystyle{}\sum\limits_{m=0}^{r}v(m+1)g(m,x)$ $\displaystyle=$ $\displaystyle\sum\limits_{i=1}^{r-1}\alpha_{i}\left(g(i-1,x)-2g(i,x)+g(i+1,x)\right)$ $\displaystyle=$ $\displaystyle\sum\limits_{i=1}^{r-1}\frac{\alpha_{i}(x-1)}{c-2r+i-x+1}{{c-2r+i-1}\choose{x-1}}\geq 0$ As the above is true for every permissible $v\in\ker H$ such that $h_{D}+v$ is a shape vector, therefore we conclude that proximally compact balanced $(n,k,r,c)$ assignment has the least variance amongst all balanced $(n,k,r,c)$ assignments. ∎ The above theorem guarantees that if a proximally compact balanced $(n,k,r,c)$ assignment exists, then it has the least variance. Also note from (48) that for the case $x=1$, the contribution due to every permissible $v\in\ker H$ is zero and so all shape vectors give the same variance, which is zero. That is consistent with what we have shown earlier. Similarly, note that for $x>c-r$, every ${{c-2r+i-1}\choose{x-1}}$ is zero and therefore from (48) we can conclude that every balanced $(n,k,r,c)$ assignment has the same variance when one considers the cases of $x>c-r$. We now provide an example of $(n,k,r,c)$ that does not have a proximally compact assignment. ###### Example 0. Consider balanced $(10,5,4,8)$ assignments. Here $\frac{r(k-1)}{n-1}=\frac{16}{9}$ and so the shape vector must have at least two nonzero entries. Moreover, $\ell=1$ and the corresponding shape vector for a possible proximally compact assignment should be $h_{D}=[0,10,35,0,0]^{T}$. We now show that an assignment with the shape vector $h_{D}$ does not exist. As every job is assigned to $r=4$ servers, every job is involved in $r(k-1)=16$ job pairs. Consider the job $a_{1}$ and let $x$ be the number of other jobs with whom $a_{1}$ shares only one server. So $9-x$ jobs share two servers each with $a_{1}$. Clearly $2(9-x)+x=16$ which implies that $x=2$. As $a_{1}$ was an arbitrary choice, we can conclude that every job shares one server with two other jobs and shares two servers with the other $7$ jobs. Let $\mathfrak{G}=\left\\{a_{1},a_{2},a_{3},\cdots a_{i},a_{1}\right\\}$ be a cycle of jobs such that each job shares only one server with the jobs that are its predecessor and successor in $\mathfrak{G}$. One can now argue that the only permissible lengths of these cycles can be either 5 or 10.444This follows since there cannot exist a cycle of size 3 i.e. there cannot exist jobs $a,b$ and $c$ such that pairs $(a,b)$, $(b,c)$ and $(c,a)$ share a server each. Then some more work proves that it is not possible to have an assignment with $h_{D}=[0,10,35,0,0]^{T}$ The actual balanced $(10,5,4,8)$ assignments that have shape vectors closest to $h_{D}$ have shape vectors $[1,8,36,0,0]^{T}$ and $[0,12,31,2,0]^{T}$ respectively. We now define another class of compact assignments. ###### Definition 3. A balanced assignment $D$ is stretched compact if the shape vector $h_{D}$ has non-zero elements only in the first and the last entries. If only the first and last entries of the shape vector $h_{D}$ are nonzero, then by (26) it is clear that the last entry of the shape vector is $\frac{c}{r}{k\choose 2}=\frac{n(k-1)}{2}$ and therefore the first entry of the shape vector is $\frac{n(n-k)}{2}$. Of course, if $n$ is a odd number and $k$ is even, then these calculated quantities are not integers and therefore for such cases, there is no possibility of existence of a stretched compact $(n,k,r,c)$ assignment. Even for the other cases, there is no guarantee that a stretched compact $(n,k,r,c)$ assignment exists even though the shape vector has integer entries. ###### Theorem 8. If a stretched compact balanced $(n,k,r,c)$ job assignment exists, then it has the largest variance amongst all balanced $(n,k,r,c)$ job assignments. This proof goes along very similar lines to that of Theorem 5 and we prove in Appendix D. ###### Example 0. Let us revisit the earlier example of $(10,5,4,8)$ assignments. It is clear that a shape vector corresponding to stretched compact assignment is permissible, namely $h_{D}=[25,0,0,0,20]^{T}$. Such an assignment is indeed possible. Divide the $10$ jobs into two sets of $5$ jobs. Assign each of these sets of jobs to $4$ servers. That results in a total of $20$ pairs sharing $4$ servers each and the rest $25$ pairs consisting of a job each from the two sets sharing no servers. This repetition assignment is a stretched compact assignment and therefore has the largest variance amongst all possible balanced $(10,5,4,8)$ assignments. It is clear that this sort of repetition assignment where multiple servers have the same set of jobs assigned is only possible if $k$ divides $n$. In such a situation, one can subdivide the jobs into $\frac{n}{k}$ groups of $k$ jobs each. Each of these groups are repeated at $r$ servers, thus accounting for $\frac{n}{k}r=c$ servers. Thus the number of job pairs that appear together $r$ times is equal to $\frac{n}{k}{k\choose 2}=\frac{n(k-1)}{2}$. And in all these cases, these repetition assignments would correspond to a stretched compact assignment which has the largest variance amongst all balanced $(n,k,r,c)$ assignments. ## 7\. Generalization where the number of servers that return ($x$) is random We now look at a scenario where each of the $c$ servers is independently and equally likely to communicate with the master with probability $p$. Note that under this setup, the distribution of the number of servers $x$ that could communicate is given by the binomial distribution $B(c,p)$. Observe that conditioned on $x$, every subset of $x$ servers is equally likely to communicate with the master. Under this setup, we can now state our results on mean and variance on the number of distinct jobs received by the master. ###### Theorem 1. Consider any balanced $(n,k,r,c)$ assignment $D$, where each server is independently and equally likely to communicate with the master with probability $p$. The expectation of the number of distinct completed jobs $d$ received is the same for every assignment $D$ amongst all balanced $(n,k,r,c)$ assignments and is given by (30) $\mathbbm{E}_{D}[d]=n-n(1-p)^{r}$ and the variance is given by (31) $\sigma_{D}(d)=\sum_{x=0}^{c}\sigma_{D,x}(d){{c\choose x}}p^{x}(1-p)^{c-x}$ where $\sigma_{D,x}(d)$ is given by the expression in Equation (22). ###### Proof. Observe that under this setup, the number of servers that communicates with the master $x$ is be given by the binomial distribution $B(c,p)$. Also observe that under this setup conditioned on $x$, any set of $x$ servers is equally likely to communicate with the master. We can thus say that (32) $\displaystyle\mathbbm{E}_{D}[d]=$ $\displaystyle\mathbbm{E}_{x\sim B(c,p)}\mathbbm{E}_{D,x}[d]$ (33) $\displaystyle\overset{(a)}{=}\sum\limits_{x=0}^{c}n\left(1-\frac{{c-r\choose x}}{{c\choose x}}\right){{c\choose x}}p^{x}(1-p)^{c-x}$ (34) $\displaystyle=n-n(1-p)^{r}\sum\limits_{x=0}^{c}{{c-r\choose x}}p^{x}(1-p)^{c-r-x}=n-n(1-p)^{r}$ Note $(a)$ follows from the expression of mean in Theorem 1. Using a very similar technique, we can prove a result of $\sigma_{D}(d)$ as well. ∎ We can actually generalize some of our results in Theorem 5 and 8 for a more generalized setup where the number of servers that return is not unique but is sampled from some distribution $\mathcal{P}$. However, we ensure that the subset $S_{1}$ is the set of servers that could communicate is equally likely as the subset of servers $S_{2}$ that could communicate if $|S_{1}|=|S_{2}|$. Formally, we study the setup where $x$ is sampled from a distribution $\mathcal{P}$ and conditioned on $x$, any subset of $x$ servers is equally likely to be the set of servers that could communicate with the master. This precisely captures the case where every server is independently able to communicate to the master with probability $p$, in which case $\mathcal{P}$ would be given by $B(c,p)$ ###### Theorem 2. Let us consider $x\sim\mathcal{P}$. Conditioned on $x$, we study the setup where any set of $x$ servers is equally likely to communicate with the master. Then the proximally compact assignment (if it exists) attains the least variance on the number of distinct jobs received at master amongst all balanced $(n,k,r,c)$ assignment schemes. ###### Proof. Let us denote the number of distinct jobs when any set of $x$ servers return uniformly at random by $d$. However, in our problem $x$ itself might be sampled from a distribution $\mathcal{P}$. Let us denote the variance in this set-up under this assignment of jobs to servers (say ${D}$) by $\sigma_{{D},x\sim\mathcal{P}}(d)$. Now using law of variances(Eve’s law), we can say that $\sigma_{{D},x\sim\mathcal{P}}(d)=\mathbb{E}_{x\sim\mathcal{P}}[\sigma_{{D},x}(d)]+\sigma_{x\sim\mathcal{P}}[\mathbb{E}_{{D},x}(d)]$ Now consider assignments $D$ and ${D}_{1}$ such that assignment $D$ is a proximally compact $(n,k,r,c)$ assignment scheme and assignment $D_{1}$ could be any balanced $(n,k,r,c)$ assignment scheme. However, we know from Theorem 5 that $\sigma_{{D},x}(d)\leq\sigma_{{D}_{1},x}(d)$ for every $x$ if $D$ is a proximally compact $(n,k,r,c)$ assignment scheme and assignment $D_{1}$ is any other balanced $(n,k,r,c)$ assignment scheme. We also know that $\mathbbm{E}_{{D},x}(d)=\mathbbm{E}_{{D}_{1},x}(d)$ from Theorem 1. Combining the two properties, we get that $\sigma_{{D},x\sim\mathcal{P}}(d)\leq\sigma_{{D}_{1},x\sim\mathcal{P}}(d)$ thus proving the theorem. ∎ Similarly, we can prove a result corresponding to that of Theorem 8 for this setup. ###### Theorem 3. Let us consider $x\sim\mathcal{P}$. Conditioned on $x$, if any set of $x$ servers is equally likely to communicate with the master, then the stretched compact $(n,k,r,c)$ assignment scheme (if it exists) attains the largest variance on the number of distinct jobs received at master amongst all balanced $(n,k,r,c)$ assignment schemes. ## 8\. Conclusion In this work, we study the mean and the variance of the number of distinct jobs received at server under various assignment schemes and show that assignment schemes based on the repetition coding and block designs attain the largest and least variance respectively. However, it is not always known whether such designs always exist and famous Bruck-Ryser-Chowla theorem in (Sprott, 1955) just gives some necessary conditions. In future works, it would be interesting to see if small modifications to designs are still close to extremal variance in case they do not exist. Another direction of work could be to extend results beyond the second moment to the $t^{th}$ moment and investigate if constructions based on $t$-designs can attain extremal results. ## 9\. Acknowledgements Most of this work was performed when SS was an undergraduate at Indian Institute of Technology, Bombay. SS would like to thank Siddharth Chandak for insightful discussions on Section 7. ## References * (1) * AWS ([n. d.]) [n. d.]. Distributed Computing AWS. https://aws.amazon.com/what-is/distributed-computing/. * Goo ([n. d.]) [n. d.]. Distributed Computing Google Cloud. https://cloud2data.com/distributed-computing-on-the-google-cloud-platform/. * Azu ([n. d.]) [n. d.]. Distributed Computing Microsoft Azure. https://cloud2data.com/distributed-computing-in-microsoft-azure/. * Addelman (1969) Sidney Addelman. 1969\. The generalized randomized block design. _The American Statistician_ 23, 4 (1969), 35–36. * Aktas and Soljanin (2019) Mehmet Fatih Aktas and Emina Soljanin. 2019. Straggler Mitigation at Scale. _CoRR_ abs/1906.10664 (2019). arXiv:1906.10664 http://arxiv.org/abs/1906.10664 * Ammar et al. (2004) Bassem Ammar, Bahram Honary, Yu Kou, Jun Xu, and Shu Lin. 2004. Construction of low-density parity-check codes based on balanced incomplete block designs. _IEEE Transactions on information Theory_ 50, 6 (2004), 1257–1269. * Ananthanarayanan et al. (2013a) Ganesh Ananthanarayanan, Ali Ghodsi, Scott Shenker, and Ion Stoica. 2013a. Effective Straggler Mitigation: Attack of the Clones. In _Proceedings of the 10th USENIX Conference on Networked Systems Design and Implementation_ (Lombard, IL) _(nsdi’13)_. USENIX Association, USA, 185–198. * Ananthanarayanan et al. (2013b) Ganesh Ananthanarayanan, Ali Ghodsi, Scott Shenker, and Ion Stoica. 2013b. Effective straggler mitigation: Attack of the clones. In _10th USENIX Symposium on Networked Systems Design and Implementation (NSDI 13)_. 185–198. * Arlitt and Williamson (1996) Martin F. Arlitt and Carey L. Williamson. 1996. Web Server Workload Characterization: The Search for Invariants. In _Proceedings of the 1996 ACM SIGMETRICS International Conference on Measurement and Modeling of Computer Systems_ (Philadelphia, Pennsylvania, USA) _(SIGMETRICS ’96)_. Association for Computing Machinery, New York, NY, USA, 126–137. https://doi.org/10.1145/233013.233034 * Babu and Kumar (2015) Balaji Srinivasan Babu and P. Vijay Kumar. 2015. On Partial Maximally-Recoverable and Maximally-Recoverable Codes. _CoRR_ abs/1501.07130 (2015). arXiv:1501.07130 http://arxiv.org/abs/1501.07130 * BOSE (1939) R. C. BOSE. 1939\. ON THE CONSTRUCTION OF BALANCED INCOMPLETE BLOCK DESIGNS. _Annals of Eugenics_ 9, 4 (1939), 353–399. https://doi.org/10.1111/j.1469-1809.1939.tb02219.x arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1469-1809.1939.tb02219.x * Bose and Nair (1939) Raj Chandra Bose and K Raghavan Nair. 1939. Partially balanced incomplete block designs. _Sankhyā: The Indian Journal of Statistics_ (1939), 337–372. * Chen et al. (2014) Shengbo Chen, Yin Sun, Longbo Huang, Prasun Sinha, Guanfeng Liang, Xin Liu, Ness B Shroff, et al. 2014\. When queueing meets coding: Optimal-latency data retrieving scheme in storage clouds. In _IEEE INFOCOM 2014-IEEE Conference on Computer Communications_. IEEE, 1042–1050. * Colajanni et al. (1998) Michele Colajanni, Philip S. Yu, and Daniel M Dias. 1998\. Analysis of task assignment policies in scalable distributed Web-server systems. _IEEE transactions on Parallel and Distributed Systems_ 9, 6 (1998), 585–600. * Colbourn and Dinitz (2006) Charles J. Colbourn and Jeffrey H. Dinitz. 2006. _Handbook of Combinatorial Designs, Second Edition (Discrete Mathematics and Its Applications)_. Chapman & Hall/CRC. * Crovella and Bestavros (1997) M.E. Crovella and A. Bestavros. 1997. Self-similarity in World Wide Web traffic: evidence and possible causes. _IEEE/ACM Transactions on Networking_ 5, 6 (1997), 835–846. https://doi.org/10.1109/90.650143 * Crovella et al. (1998) Mark E Crovella, Murad S Taqqu, and Azer Bestavros. 1998\. Heavy-tailed probability distributions in the World Wide Web. _A practical guide to heavy tails_ 1 (1998), 3–26. * Dutta et al. (2017) S. Dutta, V. Cadambe, and P. Grover. 2017. Coded convolution for parallel and distributed computing within a deadline. In _2017 IEEE International Symposium on Information Theory (ISIT)_. 2403–2407. https://doi.org/10.1109/ISIT.2017.8006960 * Ferdinand and Draper (2018) Nuwan Ferdinand and Stark C Draper. 2018. Hierarchical coded computation. In _2018 IEEE International Symposium on Information Theory (ISIT)_. IEEE, 1620–1624. * Harchol-Balter (2000) Mor Harchol-Balter. 2000\. Task assignment with unknown duration. In _Proceedings 20th IEEE international conference on distributed computing systems_. IEEE, 214–224. * Harchol-Balter et al. (1999) Mor Harchol-Balter, Mark E Crovella, and Cristina D Murta. 1999\. On choosing a task assignment policy for a distributed server system. _J. Parallel and Distrib. Comput._ 59, 2 (1999), 204–228. * Joshi et al. (2015) Gauri Joshi, Emina Soljanin, and Gregory Wornell. 2015\. Queues with redundancy: Latency-cost analysis. _ACM SIGMETRICS Performance Evaluation Review_ 43, 2 (2015), 54–56. * Joshi et al. (2017) Gauri Joshi, Emina Soljanin, and Gregory Wornell. 2017\. Efficient Redundancy Techniques for Latency Reduction in Cloud Systems. _ACM Trans. Model. Perform. Eval. Comput. Syst._ 2, 2, Article 12 (apr 2017), 30 pages. https://doi.org/10.1145/3055281 * Kadhe et al. (2019) Swanand Kadhe, O Ozan Koyluoglu, and Kannan Ramchandran. 2019\. Gradient coding based on block designs for mitigating adversarial stragglers. In _2019 IEEE International Symposium on Information Theory (ISIT)_. IEEE, 2813–2817. * Karakus et al. (2019) Can Karakus, Yifan Sun, Suhas Diggavi, and Wotao Yin. 2019\. Redundancy Techniques for Straggler Mitigation in Distributed Optimization and Learning. _Journal of Machine Learning Research_ 20, 72 (2019), 1–47. http://jmlr.org/papers/v20/18-148.html * Kiefer (1975) J. Kiefer. 1975\. Balanced Block Designs and Generalized Youden Designs, I. Construction (Patchwork). _The Annals of Statistics_ 3, 1 (1975), 109 – 118. https://doi.org/10.1214/aos/1176343002 * Korhonen and Frossard (2009) Jari Korhonen and Pascal Frossard. 2009. Flexible forward error correction codes with application to partial media data recovery. _Signal Processing: Image Communication_ 24, 3 (2009), 229–242. https://doi.org/10.1016/j.image.2008.12.005 * Kshirsagar (1958) A. M. Kshirsagar. 1958\. A Note on Incomplete Block Designs. _The Annals of Mathematical Statistics_ 29, 3 (1958), 907–910. http://www.jstor.org/stable/2237276 * Lee et al. (2017) Kangwook Lee, Maximilian Lam, Ramtin Pedarsani, Dimitris Papailiopoulos, and Kannan Ramchandran. 2017\. Speeding up distributed machine learning using codes. _IEEE Transactions on Information Theory_ 64, 3 (2017), 1514–1529. * Li and Avestimehr (2020) Songze Li and Salman Avestimehr. 2020. Coded computing. _Foundations and Trends® in Communications and Information Theory_ 17, 1 (2020). * Maity et al. (2019) Raj Kumar Maity, Ankit Singh Rawa, and Arya Mazumdar. 2019\. Robust gradient descent via moment encoding and LDPC codes. In _2019 IEEE International Symposium on Information Theory (ISIT)_. IEEE, 2734–2738. * Mallick et al. (2020) Ankur Mallick, Malhar Chaudhari, Utsav Sheth, Ganesh Palanikumar, and Gauri Joshi. 2020\. Rateless codes for near-perfect load balancing in distributed matrix-vector multiplication. In _Abstracts of the 2020 SIGMETRICS/Performance Joint International Conference on Measurement and Modeling of Computer Systems_. 95–96. * Ozfatura et al. (2019) Emre Ozfatura, Deniz Gündüz, and Sennur Ulukus. 2019\. Speeding Up Distributed Gradient Descent by Utilizing Non-persistent Stragglers. In _2019 IEEE International Symposium on Information Theory (ISIT)_. 2729–2733. https://doi.org/10.1109/ISIT.2019.8849684 * Ozfatura et al. (2021) Emre Ozfatura, Sennur Ulukus, and Deniz Gündüz. 2021\. Coded distributed computing with partial recovery. _IEEE Transactions on Information Theory_ 68, 3 (2021), 1945–1959. * Sakorikar and Wang (2021) Animesh Sakorikar and Lele Wang. 2021. Variants on block design based gradient codes for adversarial stragglers. In _2021 11th International Symposium on Topics in Coding (ISTC)_. IEEE, 1–5. * Sarmasarkar et al. (2022) Sahasrajit Sarmasarkar, V Lalitha, and Nikhil Karamchandani. 2022\. On gradient coding with partial recovery. _IEEE Transactions on Communications_ 71, 2 (2022), 644–657. * Semchedine et al. (2011) Fouzi Semchedine, Louiza Bouallouche-Medjkoune, and Djamil Aissani. 2011\. Task assignment policies in distributed server systems: A survey. _Journal of network and Computer Applications_ 34, 4 (2011), 1123–1130. * Shieh and Jan (2004) Gwowen Shieh and Show-Li Jan. 2004. The effectiveness of randomized complete block design. _Statistica Neerlandica_ 58, 1 (2004), 111–124. * Smith (1968) KJC Smith. 1968\. _An application of incomplete block designs to the construction of error-correcting codes_. Technical Report. North Carolina State University. Dept. of Statistics. * Sprott (1955) D. A. Sprott. 1955\. Balanced Incomplete Block Designs and Tactical Configurations. _Ann. Math. Statist._ 26, 4 (12 1955), 752–758. https://doi.org/10.1214/aoms/1177728433 * Sun et al. (2019) Yuxuan Sun, Junlin Zhao, Sheng Zhou, and Deniz Gunduz. 2019\. Heterogeneous Coded Computation across Heterogeneous Workers. In _2019 IEEE Global Communications Conference (GLOBECOM)_. 1–6. https://doi.org/10.1109/GLOBECOM38437.2019.9014006 * Szajda et al. (2005) Doug Szajda, Barry Lawson, and Jason Owen. 2005\. Toward an Optimal Redundancy Strategy for Distributed Computations. In _2005 IEEE International Conference on Cluster Computing_. 1–11. https://doi.org/10.1109/CLUSTR.2005.347045 * Tandon et al. (2017) Rashish Tandon, Qi Lei, Alexandros G. Dimakis, and Nikos Karampatziakis. 2017. Gradient Coding: Avoiding Stragglers in Distributed Learning. In _Proceedings of the 34th International Conference on Machine Learning_ _(Proceedings of Machine Learning Research, Vol. 70)_ , Doina Precup and Yee Whye Teh (Eds.). PMLR, 3368–3376. https://proceedings.mlr.press/v70/tandon17a.html * Wang et al. (2014) Da Wang, Gauri Joshi, and Gregory Wornell. 2014. Efficient Task Replication for Fast Response Times in Parallel Computation. _SIGMETRICS Perform. Eval. Rev._ 42, 1 (jun 2014), 599–600. https://doi.org/10.1145/2637364.2592042 * Wang et al. (2015) Da Wang, Gauri Joshi, and Gregory Wornell. 2015. Using Straggler Replication to Reduce Latency in Large-Scale Parallel Computing. _SIGMETRICS Perform. Eval. Rev._ 43, 3 (nov 2015), 7–11. https://doi.org/10.1145/2847220.2847223 * Wang et al. (2019a) Haozhao Wang, Song Guo, Bin Tang, Ruixuan Li, and Chengjie Li. 2019a. Heterogeneity-aware Gradient Coding for Straggler Tolerance. In _2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS)_. 555–564. https://doi.org/10.1109/ICDCS.2019.00062 * Wang et al. (2019b) Sinong Wang, Jiashang Liu, and Ness Shroff. 2019b. Fundamental Limits of Approximate Gradient Coding. _Proc. ACM Meas. Anal. Comput. Syst._ 3, 3, Article 52 (dec 2019), 22 pages. https://doi.org/10.1145/3366700 * Williams et al. (2005) Adepele Williams, Martin Arlitt, Carey Williamson, and Ken Barker. 2005. Web workload characterization: Ten years later. _Web content delivery_ (2005), 3–21. * Yang et al. (2019) Chien-Sheng Yang, Ramtin Pedarsani, and A. Salman Avestimehr. 2019\. Timely-Throughput Optimal Coded Computing over Cloud Networks. In _Proceedings of the Twentieth ACM International Symposium on Mobile Ad Hoc Networking and Computing_ (Catania, Italy) _(Mobihoc ’19)_. Association for Computing Machinery, New York, NY, USA, 301–310. https://doi.org/10.1145/3323679.3326528 * Yeh (1986) Ching-Ming Yeh. 1986\. Conditions for Universal Optimality of Block Designs. _Biometrika_ 73, 3 (1986), 701–706. http://www.jstor.org/stable/2336535 * Yu et al. (2019) Qian Yu, Songze Li, Netanel Raviv, Seyed Mohammadreza Mousavi Kalan, Mahdi Soltanolkotabi, and Salman A Avestimehr. 2019. Lagrange coded computing: Optimal design for resiliency, security, and privacy. In _The 22nd International Conference on Artificial Intelligence and Statistics_. PMLR, 1215–1225. * Yu et al. (2017) Qian Yu, Mohammad Ali Maddah-Ali, and A Salman Avestimehr. 2017\. Polynomial codes: an optimal design for high-dimensional coded matrix multiplication. In _Proceedings of the 31st International Conference on Neural Information Processing Systems_. 4406–4416. * Zhang et al. (2013) Qi Zhang, Mohamed Faten Zhani, Raouf Boutaba, and Joseph L. Hellerstein. 2013. Harmony: Dynamic Heterogeneity-Aware Resource Provisioning in the Cloud. In _2013 IEEE 33rd International Conference on Distributed Computing Systems_. 510–519. https://doi.org/10.1109/ICDCS.2013.28 * Zhao et al. (2014) Xu Zhao, Ling Liu, Qi Zhang, and Xiaoshe Dong. 2014\. Improving MapReduce Performance in a Heterogeneous Cloud: A Measurement Study. In _2014 IEEE 7th International Conference on Cloud Computing_. 400–407. https://doi.org/10.1109/CLOUD.2014.61 ## Appendix A Assignment equivalence when number of servers equals number of jobs So far, we have been viewing the $(n,k,r,c)$ balanced assignment as a task of assigning jobs to servers. Equivalently, one can also view the same balanced assignment as assigning servers to jobs. Since every job is assigned to $r$ distinct servers, there is a total of ${r\choose 2}$ server pairs that can be associated to each job. Like the shape vector $h_{D}$ that was built from information about frequency of job-pairs at the servers in a given assignment, one may build an equivalent shape vector based on frequency of server-pairs that are associated to the jobs. Thus every server-pair can possibly be associated to $m$ jobs, where $0\leq m\leq k$. Thus, the shape vector corresponding to server assignment viewpoint would be a vector in $\mathbb{N}^{k+1}$. One may now define compact balanced $(n,k,r,c)$ server assignments as those balanced assignments that have server shape vector having at most two nonzero entries. We may view proximally compact balanced $(n,k,r,c)$ server assignment, through an analogous definition. ###### Definition 4. Given a balanced $(n,k,r,c)$ assignment, we call it proximally compact $(n,k,r,c)$ server assignment if for every pair of distinct servers $s_{i}$ and $s_{j}$ in $\mathcal{S}$, the number of jobs assigned to both $s_{i}$ and $s_{j}$ simultaneously is exactly $\ell$ or $\ell+1$ for some integer $\ell$. ###### Lemma 2. For any proximally compact $(n,k,r,c)$ server assignment, we must have $\ell=\Bigl{\lfloor}\frac{n.{{r\choose 2}}}{{{c\choose 2}}}\Bigr{\rfloor}$ This lemma can be proved in a very similar way as that of Lemma 1. We may observe that in general, proximally compact server assignment schemes may not be proximally compact job assignment schemes. We demonstrate that with an Example 1 below. ###### Example 0. We now describe a proximally compact $(14,6,3,7)$ server assignment which is not a proximally compact job assignment in Table 3. Note that in this scheme, every pair of servers have exactly $2$ jobs in common and therefore proximally compact server assignment. However, some pairs of jobs appear together in one server like $(a_{1},a_{2})$, some pairs of jobs appear together in $2$ servers like pairs of jobs $(a_{1},a_{8})$ and some pairs of jobs like $(a_{3},a_{10})$ appear together in 3 servers and thus it is not proximally compact job assignment. Jobs Servers | $s_{1}$ | $s_{2}$ | $s_{3}$ | $s_{4}$ | $s_{5}$ | $s_{6}$ | $s_{7}$ ---|---|---|---|---|---|---|--- $a_{1}$ | | 1 | | 1 | | 1 | $a_{2}$ | 1 | | | 1 | 1 | | $a_{3}$ | | | 1 | 1 | | | 1 $a_{4}$ | 1 | 1 | 1 | | | | $a_{5}$ | | 1 | | | 1 | | 1 $a_{6}$ | 1 | | | | | 1 | 1 $a_{7}$ | | | 1 | | 1 | 1 | $a_{8}$ | 1 | | | 1 | | 1 | $a_{9}$ | | 1 | | 1 | 1 | | $a_{10}$ | | | 1 | 1 | | | 1 $a_{11}$ | 1 | 1 | 1 | | | | $a_{12}$ | 1 | | | | 1 | | 1 $a_{13}$ | | 1 | | | | 1 | 1 $a_{14}$ | | | 1 | | 1 | 1 | Table 3. Assignment of jobs to various servers in a proximally compact $(14,6,3,7)$ server assignment scheme However, when $n=c$ our subsequent result shows that proximally compact server assignment schemes and proximally compact job assignment schemes are equivalent. ###### Theorem 2. Amongst balanced $(n,k,k,n)$ assignments, every proximally compact $(n,k,k,n)$ job assignment is also a proximally compact server assignment and vice-versa. To prove this theorem, we first prove Lemma 3. We define a random variable $Y^{D}$ as follows for balanced $(n,k,r,c)$ assignment $D$ as the number of servers in which a pair of jobs chosen uniformly at random occur together. Formally, we can say that (35) ${}\mathbb{P}[Y^{D}=p]=\frac{\mathfrak{m}^{D}(p)}{{{n\choose 2}}}\text{for any integer $p\in[0,r]$}$ Observe that this is a valid distribution as $\sum\limits_{p=0}^{r}\mathfrak{m}^{D}(p)={{n\choose 2}}$ in Equation (19). ###### Lemma 2. For any balanced $(n,k,r,c)$ assignment $D$, the variance of $Y^{D}$ is linearly proportional to the variance of distinct jobs $d$ received at master when any 2 servers chosen uniformly at random return i.e. are able to communicate their results to the master. We can also state it as follows. (36) $\sigma_{D,2}(d)=\frac{{{n\choose 2}}\sigma(Y^{D})+\frac{\left(c{k\choose 2}\right)^{2}}{{n\choose 2}}+n\left({r\choose 2}\right)-c\left({k\choose 2}\right)}{{c\choose 2}}-\left(\frac{n{r\choose 2}}{{c\choose 2}}\right)^{2}$ Further for $n=c$, we can say that $\sigma(Y^{D})=\sigma_{D,2}(d)$ ###### Proof. Observe that $g(0,x)=\sum\limits_{i=2}^{r+1}\sum\limits_{j=2}^{r+1}(i-1)(j-1){(c-2r)\choose(x-i-j)}{r\choose i}{r\choose j}$ from equation (16) and therefore $g(0,2)=0$. Further, using equation (17), we have $g(m+1,2)-g(m,2)=(c-2)-2(c-r-1)+(c-2r+m)=m$. Therefore, $g(m,2)=1+2+\cdots+(m-1)=\frac{m(m-1)}{2}$. Consider the numerator of first term in $\sigma_{D,x}(d)$ in equation (22) which was shown to be $2\sum\limits_{m=0}^{r}\mathfrak{m}^{D}(m)g(m,x)+n\left(\sum\limits_{t=1}^{r-1}t^{2}{r\choose(t+1)}{(c-r)\choose(x-t-1)}\right)$. Note that here we consider $x=2$ as we are considering the case when only two servers communicate back to the master. So for $x=2$, $\displaystyle 2\sum\limits_{m=0}^{r}\mathfrak{m}^{D}(m)g(m,2)+n\left(\sum\limits_{t=1}^{k-1}t^{2}{r\choose(t+1)}{(c-r)\choose(2-t-1)}\right)=$ $\displaystyle\sum\limits_{m=0}^{r}\mathfrak{m}^{D}(m)m(m-1)+n{r\choose 2}$ $\displaystyle\overset{(a)}{=}$ $\displaystyle\sum\limits_{m=0}^{r}m^{2}\mathfrak{m}^{D}(m)+n{r\choose 2}-c{k\choose 2}$ $(a)$ follows since $\sum\limits_{m=0}^{r}m\mathfrak{m}^{D}(m)=c{k\choose 2}$ in Equation (20). Therefore, (37) $\displaystyle{}\sigma_{D,2}(d)=$ $\displaystyle\frac{\sum\limits_{m=0}^{r}m^{2}\mathfrak{m}^{D}(m)+n{r\choose 2}-c{k\choose 2}}{{c\choose 2}}-\left(\frac{n{r\choose 2}}{{c\choose 2}}\right)^{2}$ Observe $\mathbbm{E}[Y^{D}]=\frac{\sum_{m=0}^{r}m\mathfrak{m}^{D}(m)}{{{n\choose 2}}}=\frac{c{{k\choose 2}}}{{{n\choose 2}}}$ as $\sum\limits_{m=0}^{r}m\mathfrak{m}^{D}(m)=c{k\choose 2}$ by equation (20) and using the definition of $Y^{D}$ in Equation (35). $\displaystyle{}\sigma(Y^{D})=$ $\displaystyle\mathbbm{E}[(Y^{D})^{2}]-(\mathbbm{E}[Y^{D}])^{2}$ (38) $\displaystyle=$ $\displaystyle\frac{1}{{n\choose 2}}\sum_{m=0}^{r}m^{2}\mathfrak{m}^{D}(m)-\left(\frac{c{{k\choose 2}}}{{{n\choose 2}}}\right)^{2}$ Thus, using equations (A) and (37), we obtain the expression (36) in the statement of the lemma. Further, when $n=c$, then $r=k$ and the two equations (A) and (37) become equal. ∎ We now prove Theorem 2. ###### Proof. Consider the set of balanced $(n,k,k,n)$ assignment schemes. Now recall $Y^{D}$ denoted the number of servers where a pair of jobs chosen uniformly at random occurs together and for $n=c$ and $k=r$, we have (39) ${}\sigma_{D,2}(d)=\sigma(Y^{D}).$ Further $\mathbb{E}[Y^{D}]=\frac{n{{k\choose 2}}}{{{n\choose 2}}}$. Also, observe from Theorem 1 (under $x=2,n=c$ and $k=r$ ) that (40) ${}\mathbb{E}_{D,2}[d]=n\left(1-\frac{(n-k)(n-k-1)}{n(n-1)}\right)=\left(2k-n\frac{k(k-1)}{n(n-1)}\right)=\mathbb{E}_{D}[2k-Y^{D}]$ We first show that every proximally compact $(n,k,k,n)$ job assignment is also a proximally compact server assignment scheme. Suppose not and consider a balanced assignment scheme $D$ which is not a proximally compact server assignment but is a proximally compact job assignment. Now let us consider the scenario where $x=2$ (exactly 2 randomly chosen servers) are able to communicate with the master. Now if $D$ is not a proximally compact server assignment, then $d$ (the number of distinct jobs received) can take at least two distinct integral values which are non- consecutive, hence $2k-d$ also has a support of at least $2$ distinct non- consecutive integral values. However the random variable $Y^{D}$ has a support of at most 2 over two consecutive indices (since it is a proximally compact job assignment). Now observe that random $2k-d$ and $Y^{D}$ have the same expectation (shown above in Equation (40)). Therefore the variance of $2k-d$ is clearly more than that of $Y^{D}$ which contradicts equation (39). Now we consider the reverse situation where a balanced assignment scheme is a proximally compact server assignment but not proximally compact job assignment. Again consider the scenario where $x=2$ (exactly 2 randomly chosen servers) are able to communicate with the master. The number of distinct jobs received at the master $d$ can take at most 2 consecutive values (as it is a proximally compact server assignment) and therefore the random variable $2k-d$ has a support of at most 2 over two consecutive indices. Suppose the assignment scheme is not proximally compact job assignment, $Y^{D}$ has a support of at least 2 elements which are non consecutive. As the random variables $Y^{D}$ and $2k-d$ have the same expectation according to (40), hence the variance of $Y^{D}$ has to be clearly larger than that $2k-d$ which is turn contradicts equation (39). Thus, the set of proximally compact job assignments is identical to the set of proximally compact server assignments. ∎ Thus, Theorem 2 ensures that when the number of servers equals the number of jobs, proximally compact server assignment and proximally compact job assignments minimise the variance of the number of distinct jobs received at the server as shown in Theorem 2. ## Appendix B Proof of Theorem 1 ###### Proof. The number of distinct jobs $d$ received by the master when servers in a subset $\hat{S}$ (with $|\hat{S}|=x$) is able to communicate with the master is given by (41) ${}d=\left|\bigcup_{j\in\hat{S}}\text{supp}(A_{D}[:,j])\right|=\left(k\times x-\sum\limits_{i=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}\right)$ Note that the term $\sum\limits_{i=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}$ excludes those jobs which have been received multiple times from various servers present in $\hat{S}$. $\displaystyle{}\mathbb{E}_{D,x}[d]=\frac{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(k\times x-\sum\limits_{i=1}^{n}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1})}{\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}1}{=}$ $\displaystyle k\times x-\frac{\sum\limits_{i=1}^{n}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}}{{c\choose x}}$ (42) $\displaystyle{=}$ $\displaystyle k\times x-\frac{n\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}}{{c\choose x}}$ Observe that for every job $a_{i}$ in a balanced $(n,k,r,c)$ assignment, the quantity $\sum\limits_{\hat{S}\subset\mathcal{S},|\hat{S}|=x}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}$ is the same, i.e., this summation is independent of $i$. We now show that the quantity $\sum\limits_{\hat{S}\subset\mathcal{S},|\hat{S}|=x}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}$ for any specified $x$, is the same for every balanced $(n,k,r,c)$ distribution ${D}$. We compute this sum by counting the number of subsets $\hat{S}\subset\mathcal{S}$ of cardinality $x$ which additionally satisfies the constraint on $\mathfrak{n}^{D}_{i,\hat{S}}=t$ (i.e. job $a_{i}$ is present in exactly $t$ servers from $\hat{S}$) for every $t$ from $2$ to $r$ (as these cases deal with the job $a_{i}$ appearing more than once in the subset $\hat{S}$). $\displaystyle{}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}{=}\sum\limits_{t=1}^{r-1}\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x,\mathfrak{n}^{D}_{i,\hat{S}}=t+1\end{subarray}}t{=}\text{ }$ $\displaystyle\sum\limits_{t=1}^{r-1}t\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x,\mathfrak{n}^{D}_{i,\hat{S}}=t+1\end{subarray}}1$ (43) $\displaystyle\overset{(a)}{=}$ $\displaystyle\sum\limits_{t=1}^{r-1}t{r\choose(t+1)}{(c-r)\choose(x-t-1)}$ The last equality $(a)$ comes from counting the number of subsets $\hat{S}\subset\mathcal{S}$ of cardinality $x$ that contain precisely $t+1$ servers that were assigned the job $a_{i}$. Consider the following binomial expressions (44) ${}ry(1+y)^{r-1}+1-{(1+y)^{r}}=\sum\limits_{t=0}^{r-1}t{{r\choose t+1}}y^{t+1}$ (45) ${}(1+y)^{c-r}=\sum\limits_{u=0}^{c-r}{{c-r\choose u}}y^{u}$ Multiplying equations (44) and (45), one observes that $\sum\limits_{t=1}^{r-1}t{r\choose(t+1)}{(c-r)\choose(x-t-1)}$ is precisely the coefficient of $y^{x}$ in $ry(1+y)^{c-1}+{(1+y)^{c-r}}-{(1+y)^{c}}$. Thus, (46) ${}\sum\limits_{t=1}^{r-1}t{r\choose(t+1)}{(c-r)\choose(x-t-1)}=r\times{{c-1\choose x-1}}+{{c-r\choose x}}-{{c\choose x}}$ Combining equations (B), (B) and (46), we get $\mathbbm{E}_{D,x}[d]=k\times x-\frac{n\left(r\times{{c-1\choose x-1}}+{{c-r\choose x}}-{{c\choose x}}\right)}{{{c\choose x}}}=n\left(1-\frac{{{c-r\choose x}}}{{{c\choose x}}}\right)$ ∎ ## Appendix C Proof of Theorem 5 ###### Proof. Let $h_{D}$ be the shape vector corresponding to the proximally compact balanced $(n,k,r,c)$ job assignment. Thus $h_{D}(i)=0$ for all $i\leq\ell$ and $i>\ell+2$ for $\ell$ as calculated in Lemma 1. Any balanced $(n,k,r,c)$ job assignment $D_{1}$ would have a shape vector $h_{D}+v$ where $v\in\ker H$ with matrix $H$ as defined in (26). Observe from the expression of variance in (22) that it is only the term $\sum\limits_{m=0}^{r}\mathfrak{m}^{D}(m)g(m,x)$ that varies amongst the different balanced $(n,k,r,c)$ assignments. Thus it is enough to show that for every permissible $v\in\ker H$ mentioned above, $\sum\limits_{m=0}^{r}v(m+1)g(m,x)\geq 0$, in order to conclude that the proximally compact balanced $(n,k,r,c)$ assignment has the least variance. We therefore first characterize $v\in\ker H$ that may appear from some balanced $(n,k,r,c)$ assignment. As both the shape vectors $h_{D},h_{D}+v\in\mathbb{N}^{r+1}$, therefore $v(i)\geq 0$ for all $i\neq\ell+1,\ell+2$. Further, as $v\in\ker H$, therefore $\sum\limits_{i=1}^{r+1}v(i)=0$ and so if $v$ is a nonzero vector, then at least one of $v(\ell+1),v(\ell+2)$ must be a negative integer. As $v\in\ker H$, therefore $v$ can be expressed in terms of the basis vectors $\left\\{h_{i}\right\\}$ listed in (27). Let $v=\sum\limits_{i=1}^{r-1}\alpha_{i}h_{i}$. We now show that all $\alpha_{i}\in\mathbb{N}$. Consider the components of $v$ for $i\leq\ell$. Let $j$ be the smallest index where $v(j)>0$ and $j\leq\ell$. Then one can inductively argue that $\alpha_{i}=0$ for all $i<j$ by starting with $i=1$ as $v(i)=0$ for $i<j$. Further, $\alpha_{j}=v(j)>0$. Now, $v(j+1)=\alpha_{j+1}-2\alpha_{j}\geq 0$ implies that $\alpha_{j+1}\geq 2\alpha_{j}$. Similarly, $v(j+2)=\alpha_{j+2}-2\alpha_{j+1}+\alpha_{j}$, which implies $\alpha_{j+2}\geq 2\alpha_{j+1}-\alpha_{j}\geq 3\alpha_{j}$ as $v(j+2)\geq 0$. On the same lines, $\alpha_{j+3}\geq 2\alpha_{j+2}-\alpha_{j+1}\geq 2(2\alpha_{j+1}-\alpha_{j})-\alpha_{j+1}=3\alpha_{j+1}-2\alpha_{j}\geq 4\alpha_{j}$. Inductively, one can show that $\alpha_{j+k}\geq(k+1)\alpha_{j}$ for all $k\leq(\ell-j)$. Thus $\alpha_{i}\geq 0$ for all $1\leq i\leq\ell$. This accounts for all $v(i)\geq 0$ for $i\leq\ell$. Similarly, one can utilize $v(i)\geq 0$ for $i>\ell+2$ to conclude that $\alpha_{i}\geq 0$ for $\ell+1\leq i\leq r-1$, by proceeding from the other end. Let $j$ now be the largest index where $v(j)>0$ and $j>\ell+2$. If $j<r+1$, then $v(r+1)=0$ forces $\alpha_{r-1}=0$. Once again, one can inductively argue that $\alpha_{i}=0$ for $j-1\leq i\leq r-1$ as the corresponding $v(i+2)=0$. Further, $\alpha_{j-2}=v(j)>0$. Now, $v(j-1)=\alpha_{j-3}-2\alpha_{j-2}\geq 0$ implies that $\alpha_{j-3}\geq 2\alpha_{j-2}$. Using $v(j-2)\geq 0$, one obtains $\alpha_{j-4}\geq 2\alpha_{j-3}-\alpha_{j-2}\geq 3\alpha_{j-2}$. Reflecting the argument used before, one can conclude that $\alpha_{j-2-k}\geq(k+1)\alpha_{j-2}$ for $0\leq k\leq(j-\ell-3)$ and so $\alpha_{i}\geq 0$ for $\ell+1\leq i\leq r-1$. Thus all $\alpha_{i}\in\mathbb{N}$. As a result, we get $\displaystyle\sum\limits_{m=0}^{r}v(m+1)g(m,x)=\sum\limits_{i=1}^{r-1}\alpha_{i}\left(g(i-1,x)-2g(i,x)+g(i+1,x)\right)$ By Lemma 3, we know that $g(m+1,x)-g(m,x)={{c-2}\choose{x-1}}-2{{c-r-1}\choose{x-1}}+{{c-2r+m}\choose{x-1}}$ and therefore (47) $\displaystyle g(i-1,x)-2g(i,x)+g(i+1,x)$ $\displaystyle=$ $\displaystyle g(i+1,x)-g(i,x)-\left\\{g(i,x)-g(i-1,x)\right\\}$ $\displaystyle=$ $\displaystyle{{c-2r+i}\choose{x-1}}-{{c-2r+i-1}\choose{x-1}}$ $\displaystyle=$ $\displaystyle\frac{x-1}{c-2r+i-x+1}{{c-2r+i-1}\choose{x-1}}\geq 0$ Thus we have (48) $\displaystyle{}\sum\limits_{m=0}^{r}v(m+1)g(m,x)$ $\displaystyle=$ $\displaystyle\sum\limits_{i=1}^{r-1}\alpha_{i}\left(g(i-1,x)-2g(i,x)+g(i+1,x)\right)$ $\displaystyle=$ $\displaystyle\sum\limits_{i=1}^{r-1}\frac{\alpha_{i}(x-1)}{c-2r+i-x+1}{{c-2r+i-1}\choose{x-1}}\geq 0$ As the above is true for every permissible $v\in\ker H$ such that $h_{D}+v$ is a shape vector, therefore we conclude that proximally compact balanced $(n,k,r,c)$ assignment has the least variance amongst all balanced $(n,k,r,c)$ assignments. ∎ ## Appendix D Proof of Theorem 8 ###### Proof. Let $h_{D}$ be the shape vector corresponding to a stretched compact balanced $(n,k,r,c)$ job assignment. Then $h_{D}(i)=0$ for all $i\neq 1,r+1$. Any balanced $(n,k,r,c)$ job assignment $D_{1}$ has a shape vector $h_{D}+v$ where $v\in\ker H$ with $H$ as defined in (26). Similar to the proof of Theorem 5, we now characterize $v\in\ker H$ that can arise from some balanced $(n,k,r,c)$ assignment. Note that $v(i)\geq 0$ for all $i\neq 1,r+1$ and $\sum\limits_{i=1}^{r+1}v(i)=0$. Hence $v(1)\leq 0$ and $v(r+1)\leq 0$. As $v\in\ker H$, therefore $v=\sum\limits_{i=1}^{r-1}\alpha_{i}h_{i}$ where $\left\\{h_{i}\right\\}$ is the basis for $\ker H$ as listed in (27). Following the proof of Theorem 5, it is now enough to show that all $\alpha_{i}\leq 0$ for any permissible nonzero vector $v\in\ker H$ since the expression in (48) would then be rendered negative, thereby signalling a decrease in the variance for all balanced $(n,k,r,c)$ assignments with shape vector $h_{D}+v$. Observe that as $v(1)\leq 0$, therefore $\alpha_{1}=v(1)\leq 0$. If $r=2$, then $\ker H$ is one dimensional and therefore $v(1)=v(3)\leq 0$, while $v(2)=-2v(1)\geq 0\geq v(1)$. If $r>2$, then as $v(2)\geq 0$ and $v(2)=\alpha_{2}-2\alpha_{1}$, therefore $\alpha_{2}\geq 2\alpha_{1}$. Now if $r=3$, then $\ker H$ is two dimensional and $\alpha_{2}=v(4)\leq 0$. Further, $v(3)\geq 0$ implies $\alpha_{1}-2\alpha_{2}\geq 0$ and therefore $\alpha_{1}\geq 2\alpha_{2}$. Thus again, for $r=3$, both $\alpha_{1},\alpha_{2}$ are negative and their values are mutually bound by the constraints $\frac{\alpha_{1}}{2}\geq\alpha_{2}\geq 2\alpha_{1}$ and $\frac{\alpha_{2}}{2}\geq\alpha_{1}\geq 2\alpha_{2}$. Now, we consider the cases of $r>3$. In this case, $v(3)\geq 0$ translates to $2\alpha_{2}\leq\alpha_{1}+\alpha_{3}$. Therefore $4\alpha_{2}\leq 2\alpha_{1}+2\alpha_{3}\leq\alpha_{2}+2\alpha_{3}$ (by using the condition obtained from $v(2)\geq 0$) which in turn translates to $3\alpha_{2}\leq 2\alpha_{3}$. From the condition $v(4)\geq 0$, we get $2\alpha_{3}\leq\alpha_{2}+\alpha_{4}$ which can now be manipulated to $6\alpha_{3}\leq 3\alpha_{2}+3\alpha_{4}\leq 2\alpha_{3}+3\alpha_{4}$ which gives us $4\alpha_{3}\leq 3\alpha_{4}$. Following the steps mentioned above, one can use the subsequent $v(k)\geq 0$ to show that $k\alpha_{k-1}\leq(k-1)\alpha_{k}$ for $2\leq k\leq(r-1)$. Combining all these inequalities, one gets $\alpha_{1}\leq\frac{\alpha_{2}}{2}\leq\frac{\alpha_{3}}{3}\leq\cdots\leq\frac{\alpha_{k}}{k}\leq\cdots\leq\frac{\alpha_{r-1}}{r-1}\leq 0$. This proves that all the $\alpha_{i}\leq 0$. ∎ Interestingly, in the proof above, one could have started from the other end and as already shown for the case $r=3$, one can get another set of constraints $\alpha_{r-1}\leq\frac{\alpha_{r-2}}{2}\leq\cdots\leq\frac{\alpha_{r-k}}{k}\leq\cdots\leq\frac{\alpha_{1}}{r-1}\leq 0$. These interwoven constraints restrict the possible values for the $\alpha_{i}$ where $v=\sum\limits_{i=1}^{r-1}\alpha_{i}h_{i}$. ## Appendix E Proof of Claim 2 ###### Proof. Consider a pair of jobs $(a_{i},a_{j})$ such that no server has been assigned both $a_{i}$ and $a_{j}$ together. Therefore there are precisely $r$ servers that have been assigned $a_{i}$ and not $a_{j}$. Another $r$ servers that are assigned $a_{j}$ but not $a_{i}$ while the remaining $c-2r$ servers are assigned neither $a_{i}$ nor $a_{j}$. Then (49) $g(0,x)=\sum\limits_{\begin{subarray}{c}\hat{S}\subset\mathcal{S};\\\ |\hat{S}|=x\end{subarray}}(\mathfrak{n}^{D}_{i,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{i,\hat{S}}>1}(\mathfrak{n}^{D}_{j,\hat{S}}-1)\mathbbm{1}_{\mathfrak{n}^{D}_{j,\hat{S}}>1}=\sum_{t=2}^{r}\sum_{u=2}^{r}(t-1)(u-1){r\choose t}{r\choose u}{c-2r\choose x-t-u}$ Clearly any subset of servers $\hat{S}\subset\mathcal{S}$ of cardinality $x$ that has at most only one instance of the job $a_{i}$ assigned amongst its members does not contribute to the sum. Ditto for $a_{j}$. Therefore, one needs to consider only those subsets $\hat{S}$ of servers that contain at least two servers that are assigned $a_{i}$ and at least two servers that are assigned $a_{j}$. In the final expression of equation (49), ${r\choose t}{r\choose u}{c-2r\choose x-t-u}$ counts the number of subsets of servers $\hat{S}$ of cardinality $x$ that contain $t$ servers assigned $a_{i}$, $u$ servers assigned $a_{j}$ and $x-t-u$ servers that have been assigned neither. The summation limits ensure that there are at least $2$ servers assigned $a_{i}$ and at least $2$ servers assigned $a_{j}$. The expression $(t-1)(u-1)$ is the contribution of each subset $\hat{S}$ that contains $t$ copies of $a_{i}$ and $u$ copies of $a_{j}$ assigned to its members. A closed form solution of the expression for $g(0,x)$ can be obtained by considering (50) ${}ry(1+y)^{r-1}+1-(1+y)^{r}=\sum_{t=1}^{r}(t-1){{r\choose t}}y^{t}$ (51) ${}ry(1+y)^{r-1}+1-(1+y)^{r}=\sum_{u=1}^{r}(u-1){{r\choose u}}y^{u}$ (52) ${}(1+y)^{c-2r}=\sum_{v=0}^{c-2r}{{c-2r\choose v}}y^{v}$ Multiplying these three expressions (50), (51) and (52), we get $\sum_{t=2}^{r}\sum_{u=2}^{r}(t-1)(u-1){{c-2r\choose x-t-u}}{{r\choose t}}{{r\choose u}}$ to be the coefficient of $y^{x}$ in $\left(ry(1+y)^{r-1}+1-(1+y)^{r}\right)^{2}(1+y)^{c-2r}$. $\displaystyle\sum_{t=2}^{r}\sum_{u=2}^{r}(t-1)(u-1){{c-2r\choose x-t-u}}{{r\choose t}}{{r\choose u}}$ (53) $\displaystyle=$ $\displaystyle r^{2}{{c-2\choose x-2}}-2r{{c-1\choose x-1}}+{{c\choose x}}-2{{c-r\choose x}}+2r{{c-r-1\choose x-1}}+{c-2r\choose x}$ ∎ ## Appendix F Proof of Claim 3 ###### Proof. Let us consider a pair of jobs $(a_{i},a_{j})$ that have been assigned together to precisely $m$ servers. Without loss of generality, let $s_{1},s_{2},\cdots,s_{m}$ be the servers that are assigned both the jobs $a_{i},a_{j}$. Let $s_{m+1},s_{m+2},\cdots,s_{r}$ be the servers that have been assigned $a_{i}$ but not $a_{j}$. Assume servers $s_{r+1},s_{r+2},\cdots,s_{2r-m}$ are the servers assigned $a_{j}$ but not $a_{i}$. The last $c-2r+m$ servers $s_{2r-m+1},s_{2r-m+2},\cdots,s_{c}$ are the ones that have not been assigned $a_{i}$ or $a_{j}$. Let another pair of jobs $(a_{i_{1}},a_{j_{1}})$ be such that they have been assigned together to precisely $m+1$ servers. We now consider a bijective map $f:\mathcal{S}\rightarrow\mathcal{S}$ described in the following fashion. Let $f(s_{\ell})$ for $1\leq\ell\leq m+1$ be servers that have been assigned both the jobs $a_{i_{1}}$ and $a_{j_{1}}$. Let $f(s_{\ell})$ for $m+2\leq\ell\leq r$ be servers that have been assigned $a_{i_{1}}$ but not $a_{j_{1}}$. Further let $f(s_{\ell})$ for $r+2\leq\ell\leq 2r-m$ be servers that have been assigned $a_{j_{1}}$ but not $a_{i_{1}}$. The rest of $f(s_{\ell})$ have not been assigned $a_{i_{1}}$ or $a_{j_{1}}$. Thus there are two special servers, namely $s_{m+1}$ (which does job $a_{i}$ but not $a_{j}$) and $s_{r+1}$ (which does job $a_{j}$ but not $a_{i}$), and whose images $f(s_{m+1})$ (which does both the jobs $a_{i_{1}}$ and $a_{j_{1}}$) and $f(s_{r+1})$ (which does neither $a_{i_{1}}$ nor $a_{j_{1}}$) that we shall pay special attention to. For any $\hat{S}\subset\mathcal{S}$ of cardinality $x$, let us compare its contribution to the sum $g(m,x)$ with the contribution of $f(\hat{S})$ towards $g(m+1,x)$. Clearly, if $\hat{S}\subset\mathcal{S}\setminus\left\\{s_{m+1},s_{r+1}\right\\}$, then the contribution of $\hat{S}$ towards $g(m,x)$ is exactly the same as the contribution of $f(\hat{S})$ to $g(m+1,x)$. Similarly, if $s_{m+1},s_{r+1}\in\hat{S}$, then contribution of $\hat{S}$ towards $g(m,x)$ and that of $f(\hat{S})$ towards $g(m+1,x)$ is exactly the same. Therefore it suffices to only consider those subsets $\hat{S}$ of cardinality $x$ that contain exactly one of the two special servers $\left\\{s_{m+1},s_{r+1}\right\\}$ to evaluate the difference $g(m+1,x)-g(m,x)$. Hence we look at subsets $\hat{S}$ that are formed by taking either $s_{m+1}$ or $s_{r+1}$ along with $\bar{S}\subset\mathcal{S}\setminus\left\\{s_{m+1},s_{r+1}\right\\}$ of cardinality $x-1$. Let $\bar{S}\subset\mathcal{S}\setminus\left\\{s_{m+1},s_{r+1}\right\\}$ of cardinality $x-1$ contain $\alpha>0$ instances of job $a_{i}$ and $\beta>0$ instances of job $a_{j}$ assigned to its servers. Then $\bar{S}\cup\\{s_{m+1}\\}$ contributes $\alpha(\beta-1)$ towards $g(m,x)$, whereas $\bar{S}\cup\\{s_{r+1}\\}$ contributes $(\alpha-1)\beta$ towards $g(m,x)$. At the same time, $f(\bar{S})\cup\\{f(s_{m+1})\\}$ contributes $\alpha\beta$ towards $g(m+1,x)$, whereas $f(\bar{S})\cup\\{f(s_{r+1})\\}$ contributes $(\alpha-1)(\beta-1)$ towards $g(m+1,x)$. Thus, one can evaluate the contribution of $\bar{S}$ towards the difference $g(m+1,x)-g(m,x)$ to be $\alpha\beta+(\alpha-1)(\beta-1)-\alpha(\beta-1)-(\alpha-1)\beta=1$. So every subset $\bar{S}\subset\mathcal{S}\setminus\\{s_{m+1},s_{r+1}\\}$ of cardinality $x-1$, whose servers have at least one instance each of jobs $a_{i}$ and $a_{j}$ assigned to them, contributes a net change of $1$ towards the difference $g(m+1,x)-g(m,x)$. One needs to just count the number of subsets $\bar{S}$ of cardinality $x-1$ that satisfy these conditions to find $g(m+1,x)-g(m,x)$. Total number of subsets of cardinality $x-1$ of the set $\mathcal{S}\setminus\\{s_{m+1},s_{r+1}\\}$ is given by ${c-2\choose x-1}$. If the subset $\bar{S}$ is one of the ${c-r-1\choose x-1}$ subsets chosen from the servers $\\{s_{r+2},s_{r+3},\cdots s_{c}\\}$, then the job $a_{i}$ is not assigned to any of its servers. Similarly, if $\bar{S}$ is one of the ${c-r-1\choose x-1}$ chosen from the servers $\\{s_{m+2},s_{m+3},\cdots,s_{r}\\}\cup\\{s_{2r-m+1},s_{2r-m+2},\cdots,s_{c}\\}$, then it does not have any instance of the job $a_{j}$ assigned to its servers. As these subsets $\bar{S}$ do not contribute to the difference $g(m+1,x)-g(m,x)$, their numbers have to be subtracted from ${c-2\choose x-1}$. In the process, $\bar{S}\subset\\{s_{2r-m+1},s_{2r-m+2},\cdots,s_{c}\\}$ have been subtracted twice and therefore ${c-2r+m\choose x-1}$ needs to added back (inclusion- exclusion principle), thereby giving $g(m+1,x)-g(m,x)={c-2\choose x-1}-2{c-r-1\choose x-1}+{c-2r+m\choose x-1}$ ∎ ## Appendix G Proof of Corollary 1 ###### Corollary 1. For $x>c-r$ , the expression of $\sigma_{D,x}(d)$ in Equation (22) in Theorem 4 goes to zero. ###### Proof. Recall the expression of $\sigma_{D,x}(d)$ from Equation (22). Observe that expression $g(m+1,x)-g(m,x)$ from Equation (17) would be ${{c-2\choose x-1}}$ for $x>c-r$ as the second and third term in equation (17) goes to zero since $p\leq r$ and $x>c-r$. (54) ${}g(m+1,x)-g(m,x)={{c-2\choose x-1}}$ Let us now compute $g(0,x)$ using the expression in (49) and (16) for $x>c-r$. (55) ${}g(0,x)=\sum_{i=2}^{r+1}\sum_{j=2}^{r+1}(i-1)(j-1){{c-2r\choose x-i-j}}=r^{2}{{c-2\choose x-2}}-2r{{c-1\choose x-1}}+{{c\choose x}}$ Thus, from equations (54) and (55), we get (56) ${}g(m,x)=r^{2}{{c-2\choose x-2}}-2r{{c-1\choose x-1}}+{{c\choose x}}+m\times{{c-2\choose x-1}}$ Since, $x>c-r$, we may claim that the term $T_{2}(n,k,r,c)$ in Equation (22) in Thoerem 4 goes as follows. (57) ${}T_{2}(n,k,r,c)=\sum\limits_{t=1}^{r-1}t^{2}{r\choose(t+1)}{(c-r)\choose(x-t-1)}=\left(r(r-1){{c-2\choose x-2}}-r{{c-1\choose x-1}}+{{c\choose x}}\right)$ Also observe that since $x>c-r$ the term ${{c-r\choose x}}$ goes to zero, hence not written in equation (14). Thus the numerator of the first term in equation (22) in Thoerem 4 is given by (from equations (55) and (56) and (57)) $\displaystyle 2.\sum\limits_{m=0}^{r}\mathfrak{m}^{D}(m)g(m,x)+n\sum\limits_{t=1}^{r-1}t^{2}{r\choose(t+1)}{(c-r)\choose(x-t-1)}$ $\displaystyle=$ $\displaystyle\sum\limits_{m=0}^{r}\Biggl{(}2\mathfrak{m}^{D}(m)\left(r^{2}{{c-2\choose x-2}}-2r{{c-1\choose x-1}}+{{c\choose x}}\right)+2m\mathfrak{m}^{D}(m){{c-2\choose x-1}}\Biggr{)}$ $\displaystyle\hskip 180.00027pt+n\left(r(r-1){{c-2\choose x-2}}-r{{c-1\choose x-1}}+{{c\choose x}}\right)$ $\displaystyle\overset{(a)}{=}\left(r^{2}{{c-2\choose x-2}}-2r{{c-1\choose x-1}}+{{c\choose x}}\right)n(n-1)+{{c-2\choose x-1}}ck(k-1)$ $\displaystyle\hskip 180.00027pt+n\left(r(r-1){{c-2\choose x-2}}-r{{c-1\choose x-1}}+{{c\choose x}}\right)$ $\displaystyle\overset{(b)}{=}{{c-2\choose x-2}}(nr(nr-1))+{{c-2\choose x-1}}ck(k-1)-{{c-1\choose x-1}}(nr(2n-1))+n^{2}{{c\choose x}}$ $\displaystyle\overset{(c)}{=}{{c-2\choose x-2}}n^{2}r^{2}+{{c-2\choose x-1}}ck^{2}-{{c-1\choose x-1}}(nr(2n))+n^{2}{{c\choose x}}$ $\displaystyle\overset{(d)}{=}{{c-1\choose x-1}}{nr\times kx}-{{c-1\choose x-1}}(nr(2n))+n^{2}{{c\choose x}}$ $\displaystyle\overset{(e)}{=}{{c\choose x}}\left(\left(\frac{nrx}{c}\right)^{2}-2n\left(\frac{nrx}{c}\right)+n^{2}\right)$ (58) $\displaystyle\overset{(f)}{=}{{c\choose x}}\left(n\times\left(\frac{rx}{c}-1\right)\right)^{2}$ We now argue for each of the steps below. * • $(a)$ follows since $\sum\limits_{m=0}^{r}m\times\mathfrak{m}^{D}(m)=c{{k\choose 2}}$ and $\sum\limits_{m=0}^{r}\mathfrak{m}^{D}(m)={{n\choose 2}}$ in Equations (19) and (20) * • $(b)$ follows by combining the coefficints of ${{c-2\choose x-2}}$, ${c-1\choose x-1}$ and ${c\choose x}$. * • $(c)$ follows as $nr{{c-2\choose x-2}}+kc{{c-2\choose x-1}}=nr{{c-1\choose x-1}}$. This can be explained by the fact that $n\times r=k\times c$. * • $(d)$ follows from the following set of equalities $\displaystyle{{c-2\choose x-2}}n^{2}r^{2}+{{c-2\choose x-1}}ck^{2}=$ $\displaystyle\frac{(c-1)!}{(x-2)!}{(c-x-1)!}\left(\frac{ck}{c-x}+\frac{k}{x-1}\right)$ $\displaystyle=$ $\displaystyle\frac{(c-1)!\times nr\times kx(c-1)}{(x-2)!(c-x)(x-1)}=nr\times kx{{c-1\choose x-1}}$ * • $(e)$ and $(f)$ follow from the fact that $n\times r=k\times c$ Now, observe the second term of $\sigma_{D,x}(d)$ in equation (22) and we see that $T_{1}(n,k,r,c)=n\times\left(\frac{rx}{c}-1\right)$ as $x>c-r$. Thus, using equation (G), we can say that $\sigma_{D,x}(d)=0$ for $x>c-r$. ∎
# Is asymptotically safe inflation eternal? J. Chojnacki,**footnotetext: Corresponding author. J. Krajecka J. H. Kwapisz O. Slowik A. Strag (December 2020) ###### Abstract Recently, based on swampland considerations in string theory, the (no) eternal inflation principle has been put forward. The natural question arises whether similar conditions hold in other approaches to quantum gravity. In this article, the asymptotic safety hypothesis is considered in the context of eternal inflation. As exemplary inflationary models the SU(N) Yang-Mills in the Veneziano limit and various RG-improvements of the gravitational action are studied. The existence of UV fixed point generically flattens the potential and our findings suggest no tension between eternal inflation and asymptotic safety, both in the matter and gravitational sector in contradistinction to string theory. Moreover, the eternal inflation cannot take place in the range of applicability of effective field quantum gravity theory. We employ the analytical relations for eternal inflation to some of the models with single minima, such as Starobinsky inflation, alpha-attractors, or the RG-improved models and verify them with the massive numerical simulations. The validity of these constraints is also discussed for a multi-minima model. ## 1 Introduction Our Universe consists of 4 fundamental forces. Three of these forces have been consistently described on the quantum level and combined into the Standard Model of particle physics. Only quantum gravity seems to be elusive and has not been fully described in terms of quantum theory. This is not only because gravity is power counting non-renormalizable but also due to the fact that direct quantum gravity regime cannot be accessed experimentally (for example an accelerator measuring the quantum gravity effects would have to be big as our Solar System). In recent years an alternative strategy has been put forward, namely one formulates a fundamental quantum gravity theory and then tests, which of the low energy effective theories can be UV completed by this quantum gravity model. In string theory this goes under the name of swampland conjectures [2, 3]. Recently widely discussed is so-called de-Sitter conjecture [4, 5] which states that string theory cannot have de-Sitter vacua and is in tension with single field inflation [6, 7]. There seems to be also a tension between standard S-matrix formulation of quantum gravity and existence of stable de- Sitter space [8, 9, 10, 11]. However it is not established whether asymptotic safety admits a standard S-matrix formulation [12] due to fractal spacetime structure in the deep quantum regime [13, 14] In line of these swampland criteria the no eternal inflation principle has been put forward [1], see also the further discussions on the subject of eternal inflation [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26]. On the other hand, the theory of inflation is a well-established model providing an answer to problems in classical cosmology, such as the flatness problem, large-scale structures formation, homogeneity, and isotropy of the universe. A handful of models is in an agreement with the CMB observations. In the inflationary models, quantum fluctuations play a crucial role in primordial cosmology, providing a seed for the large-scale structure formation after inflation and giving a possibility for the eternally inflating multiverse. Initial fluctuations in the early universe may cause an exponential expansion in points scattered throughout the space. Such regions, rapidly grow and dominate the volume of the universe, creating ever-inflating, disconnected pockets. Since so far there is no way to verify the existence of the other pockets, we treat them as potential autonomous universes, being part of the multiverse. In the light of this tension between string theory and inflationary paradigm [1], one can be interested in how robust are the swampland criteria for the various quantum gravity models. In accordance with the inflation theory, we anticipate that the dynamics of the universe are being determined by the quantum corrections to the general relativity stemming from the concrete UV model. The effective treatment led Starobinsky to create a simple inflationary model taking into account the anomaly contributions to the energy-momentum tensor. As pointed out by Donoghue [27] below the Planck scale, for quantum gravity one can safely take the effective field theory perspective. Yet these quantum gravity effects can be important below the Planck scale by the inclusion of higher dimensional operators. The gravitational constant $G_{N}$ has a vanishing anomalous dimension below the Planck scale and various logarithmic corrections to the $R^{2}$ have been considered, capturing the main quantum effects [28, 29, 30, 31]. Yet in order to get the correct 60 e-fold duration of inflationary period one has to push the scalar field value in the Einstein frame beyond the Planck mass [1]. Furthermore most of these models do not possess a flat potential limit (either diverge or have a runaway solutions), suggesting that eternal inflation can be investigated only if one takes into account the full quantum corrections to the Starobinsky inflation. In the effective field theory scheme, the predictive power of the theory is limited, as the description of gravity at transplanckian scales requires fixing infinitely many coupling constants from experiments. The idea of asymptotic safety [32] was introduced by Stephen Weinberg in 1978 as a UV completion of the quantum theory of gravity. The behavior of an asymptotically safe theory is characterized by scale invariance in the high-momentum regime. Scale invariance requires the existence of a non-trivial Renormalization Group fixed point for dimensionless couplings. There are many possible realizations of such non-trivial fixed point scenario, such as the canonical vs anomalous scaling (gravitational fixed point [33, 34, 35, 36]) and one-loop vs two-loop contributions or gauge vs Yukawa contributions, see [37] for further details and [38] for current status of asymptotically safe gravity. The existence of an interacting fixed point and hence the flatness of the potential in the Einstein frame led Weinberg to discuss [39] cosmological inflation as a consequence of Asymptotically Safe Gravity, see also [40, 41] for discussion of AS cosmology. Following this suggestion, we study two types of models. The first type relies on the RG-improvement of the gravitational actions and is based on the asymptotic safety hypothesis that gravity admits a non-trivial UV fixed point. Since asymptotically safe gravity flattens the scalar field potentials [42], one can expect that it will result in the eternal inflation for large enough initial field values. On the other hand, RG-improved actions can serve as a UV completion of the Starobinsky model. One should also note that asymptotically safe swampland has been vastly studied [43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69]. The other model relies on the non-trivial fixed point in the pure matter sector governed by the Yang-Mills dynamics in the Veneziano limit [70, 71], see also [72, 73, 74, 75, 75]. In this model, we have uncovered a new type of eternal inflation scenario relying on tunneling to a false vacuum - in the opposite direction as it was considered in the old-inflation proposal [76]. In contradistinction to string theory, the couplings in the asymptotic safety paradigm are predicted from the RG-flow of the theory and their fixed point values rather than as vacuum expectation values (vev’s) of certain scalar fields. Hence, the asymptotically safe eternally inflating multiverses landscape is much less vast than the one stemming from the string theory, making these models much less schismatic [77]. Finally, let us note that asymptotic safety can argue for the homogeneous and isotropic initial conditions on its own using the finite action principle [78]. Our work is organized as follows. In Chapter 2 we introduce the idea of eternal inflation and multiverse. We discuss necessary conditions for eternal inflation to occur based on the Fokker-Planck equation. In Chapter 3 we show, how the developed tools work in practice with the two popular inflationary models. Chapter 4 is devoted to the presence of eternal inflation in Asymptotically Safe models. In Chapter 5 the results are discussed and concluded. ## 2 How inflation becomes eternal? In this section we discuss, under what circumstances the inflation becomes eternal. Our discussion follows closely [1]. ### 2.1 Fokker-Planck equation Consider a scalar field in the FLRW metric $\displaystyle S=\int d^{4}\sqrt{-g}\left(\frac{1}{2}M_{Pl}^{2}R+\frac{1}{2}g^{\mu\nu}\partial_{\mu}\phi\partial_{\nu}\phi-V(\phi)\right),$ (2.1) with $\phi(t,\vec{x})=\phi(t)$ one obtains the following equations of motion: $\displaystyle\ddot{\phi}+3H\phi+\frac{\partial V}{\partial\phi}=0,\quad H^{2}M_{P}^{2}=\frac{1}{3}\left(\frac{1}{2}\dot{\phi}^{2}+V(\phi)\right),$ (2.2) which in the slow-roll approximation become [79]: $\displaystyle 3H\dot{\phi}+\frac{\partial V}{\partial\phi}\approx 0,\quad$ $\displaystyle H^{2}M_{Pl}^{2}\approx\frac{1}{3}V\left(\phi\right).$ (2.3) Inflation ends once one of the so-called slow-roll parameters becomes of order one $\displaystyle\epsilon\simeq\frac{M_{P}^{2}}{2}\left(\frac{V_{,\phi}}{V}\right)^{2},\quad\eta\simeq M_{P}^{2}\frac{V_{,\phi\phi}}{V},$ (2.4) and enters the oscillatory, reheating phase. The standard treatment of eternal inflation relies on the stochastic inflation approach [80]. One splits the field into classical background and short-wavelength quantum field $\displaystyle\phi\left(t,\vec{x}\right)=\phi_{cl}\left(t,\vec{x}\right)+\delta\phi\left(t,\vec{x}\right).$ (2.5) Due to the fact that action is quadratic in the fluctuations, their spatial average over the Hubble volume is normally distributed. Hence from now on we shall assume that both background and fluctuations are homogeneous, which is the standard treatment of eternal inflation (if not otherwise specified). In the large e-fold limit, equation of motion for the full field takes form of slow-roll equation with additional classical noise term [1, 81, 82], known as the Langevin Equation: $\displaystyle 3H\dot{\phi}+\frac{\partial V}{\partial\phi}=N\left(t\right),$ (2.6) where $N\left(t\right)$ is a Gaussian distribution with mean equal 0 and variance $\sigma=\frac{H^{3}t}{4\pi^{2}}$ [83]. Then the probability density of the inflaton field is then given by the Fokker-Planck equation [1]: $\displaystyle\dot{P}[\phi,t]=\frac{1}{2}\left(\frac{H^{3}}{4\pi^{2}}\right)\frac{\partial^{2}P[\phi,t]}{\partial\phi\partial\phi}+\frac{1}{3H}\partial_{i}\left(\partial^{i}V\left(\phi\right)P[\phi,t]\right),$ (2.7) where $\dot{P}[\phi,t]:=\frac{\partial}{\partial t}P[\phi,t]$. ### 2.2 Analytic solutions To understand better the Fokker-Planck equation, let us now briefly discuss the analytical solutions. #### Case 1. Constant potential $\displaystyle V\left(\phi\right)=V_{0},$ (2.8) the Fokker-Planck equation reduces to $\displaystyle\dot{P}[\phi,t]=\frac{1}{2}\left(\frac{H^{3}}{4\pi^{2}}\right)\frac{\partial^{2}P[\phi,t]}{\partial\phi\partial\phi},$ (2.9) furthermore $H^{2}=\textrm{const}$ by the Friedman equations. Then the Fokker Planck equation reduces to the standard heat equation, which has a solution given by a Gaussian distribution: $\displaystyle P[\phi,t]=\frac{1}{\sigma\left(t\right)\sqrt{2\pi}}\exp\left[-\frac{\left(\phi-\mu\left(t\right)\right)^{2}}{2\sigma\left(t\right)^{2}}\right],$ (2.10) with $\displaystyle\mu\left(t\right)=0,\quad$ $\displaystyle\sigma^{2}\left(t\right)=\frac{H^{3}}{4\pi^{2}}t.$ (2.11) A delta-function distribution initially centered at $\phi$ = 0 will remain centered at $\phi=0$ for all time. It will however, spread out by the amount $\sigma\left(t=H^{-1}\right)=H/2\pi$ after a Hubble time. This represents the standard “Hubble-sized” quantum fluctuations that are well-known in the context of inflation, famously imprinted in the CMB and ultimately seeding the observed large-scale structure. #### Case 2. Linear potential For the linear hilltop model the potential is given by $\displaystyle V\left(\phi\right)=V_{0}-\alpha\phi.$ (2.12) Fokker-Planck equation is analogously solved by the Gaussian distribution (2.10) with: $\displaystyle\mu\left(t\right)=\frac{\alpha}{3H}t,\quad$ $\displaystyle\sigma^{2}\left(t\right)=\frac{H^{3}}{4\pi^{2}}t.$ (2.13) The time-dependence of $\mu\left(t\right)$ is due to the classical rolling of the field in the linear potential. The time-dependence of $\sigma^{2}\left(t\right)$ is purely due to Hubble-sized quantum fluctuations, and it precisely matches the result in the constant case. In general, for a linear and quadratic potential the equation simplifies to the heat equation, hence the solutions are Gaussian. Furthermore, if the potential is asymptotically flat, a finite limit at infinity exists. One may employ a series expansion around this point at infinity, approximating the potential up to the linear term. It is then expected that the probability density is approximately Gaussian (2.13). In the next section, we describe in detail how the Gaussian distribution causes the inflaton to decay exponentially. ### 2.3 Eternal inflation conditions Given an arbitrary field value $\phi_{c}$, one can ask what is the probability that quantum field $\phi=\phi(t)$ is above this value: $\displaystyle\mathrm{Pr}[\phi>\phi_{c},t]=\int^{\infty}_{\phi_{c}}d{\phi}P[\phi,t].$ (2.14) Since the distribution is Gaussian, then for $\phi_{c}$ large enough the $Pr[\phi>\phi_{c},t]$ can be approximated by an exponential decay: $\displaystyle\mathrm{Pr}[\phi>\phi_{c},t]\approx C(t)\exp(-\Gamma t),$ (2.15) where $C(t)$ is polynomial in $t$ and all of the dependence on $\phi_{c}$ is contained in $C(t)$. Then it seems that inflation cannot last forever since $\displaystyle\lim_{t\to\infty}\mathrm{Pr}[\phi>\phi_{c},t]=0.$ (2.16) However, there is an additional effect to be included: expansion of the universe during inflation. The size of the universe depends on time according to: $\displaystyle U\left(t\right)=U_{0}e^{3Ht},$ (2.17) where $U_{0}$ is the initial volume of the pre-inflationary universe. One can interpret the probability $Pr[\phi>\phi_{c},t]$ as fraction of the volume $U_{inf}\left(t\right)$ still inflating, that is: $\displaystyle U_{inf}\left(t\right)=U_{0}e^{3Ht}Pr[\phi>\phi_{c},t],$ (2.18) then in order for the Universe to inflate eternally, the positive exponential factor $3H$ in Eq. (2.18) and the negative exponential factor $-\Gamma$ in (2.15) must satisfy: $\displaystyle 3H>\Gamma.$ (2.19) We shall illustrate this general property on an example of linear potential. Evaluating the integral for probability density, in the linear case gives: $\displaystyle Pr[\phi>\phi_{c},t]=\frac{1}{2}\textrm{erfc}\left({\frac{\frac{\alpha}{3H}t-\phi_{c}}{\frac{H}{2\pi}\sqrt{2Ht}}}\right).$ (2.20) The error function may be approximated by an exponential: $\displaystyle Pr[\phi>\phi_{c},t]=C\left(t\right)\textrm{exp}\left(-\frac{4\pi^{2}\alpha^{2}}{18H^{5}}t\right),$ (2.21) where $C\left(t\right)$ is power-law in $t$ and $\phi_{c}$ vanished from the final approximation of the probability, which is a generic feature. By comparing the exponents we can check, whether $U_{inf}$ will grow or tend to zero. The condition for eternal inflation to occur becomes: $\displaystyle 3H>\frac{4\pi^{2}\alpha^{2}}{18H^{5}}.$ (2.22) For linear potential $\alpha=V^{\prime}\left(\phi\right)$ using the slow-roll equations equation (2.3), above condition can be rewritten: $\displaystyle\frac{|V^{\prime}|}{V^{\frac{3}{2}}}<\frac{\sqrt{2}}{2\pi}\frac{1}{M^{2}_{Pl}}.$ (2.23) This can be interpreted as quantum fluctuations dominating over classical field rolling. For linear potential, this is satisfied for a large $\phi$. Similarly, the second condition for the eternal inflation may be derived from the quadratic hilltop potential: $-\frac{V^{\prime\prime}}{V}<\frac{3}{M^{2}_{Pl}}.$ (2.24) Further necessary conditions on p-th derivative with $p>2$ have been derived in [1] and give: $\displaystyle[-\textrm{sgn}\left(\partial^{p}V\right)]^{p+1}\frac{|\partial^{p}V|}{V^{(4-p}/2}<\mathcal{N}_{p}M_{Pl}^{p-4},$ (2.25) where $\mathcal{N}_{p}\gg 1$ is numerically determined coefficient. Eternal inflation can be understood as a random walk of a field and a diffusion process on top of the classical motion [Vilenkin, Guth:2000ka, 15]. In order to cross check the formulas (2.23, 2.24) the numerical simulation has been developed. To reconstruct the probability distribution one simulates the discretized version of equation (2.6): $\displaystyle\phi_{n}=\phi_{n-1}-\frac{1}{3H}V^{\prime}\left(\phi_{n-1}\right)\delta t+\delta\phi_{q}\left(\delta t\right),$ (2.26) with $\delta\phi_{q}\left(\delta t\right)$ being random number taken from the gaussian distribution with mean equal zero, and variance $\frac{H^{3}}{4\pi^{2}}\delta t$. We further assume the Hubble parameter to be constant and respecting the slow-roll regime $H=\frac{1}{M_{Pl}}\sqrt{\frac{V(\phi_{0})}{3}}$, where $V(\phi_{0})$ is a value of the potential at the start of the simulation. We verified, that the change of $H$ caused by the field fluctuation does not affect the conclusions for eternal inflation. The simulation starts at the user-given value $\phi_{0}$ and follows the Langevin discretized equation (2.26). If the inflation occurs, the corresponding timestep $t_{n}$ is added to a list. This happens while the slow-roll conditions are satisfied, meaning $\epsilon(t_{n})$ and $\eta(t_{n})$ are smaller than one. Violation of one of these conditions resets the simulation. However, the list containing information about time $t_{n}$ of the ongoing inflation is stored in the memory. This large time list is appended in the similar way each evolution. Its size may be estimated by $N\frac{T_{c}}{\delta t}$, where $N$ is the total number of simulations and $T_{c}$ is the time of the classical slow-roll inflation starting at $\phi_{0}$. It is important to stress, that the duration of a particular evolution may be too long to compute in any practical time. We employ a large timeout ending the evolution. This is a good approximation for our purposes. Finally, the list containing information about every timestep at which inflation was ongoing in $N$ independent simulations is sorted in an ascending order. A normalized histogram with 1000 equal-width bins is created from the list. The number of counts is related to the probability of ongoing inflation, while the bins correspond to inflationary time. The field’s evolution supports the Fokker-Planck result (2.15). In the slow-roll regime, the inflaton decays exponentially with decay parameter $\Gamma$. This is true for every numerically investigated potential in this work. In order to recognize the eternally inflating models we search for such initial value of the field $\phi_{0}$ that $\Gamma<3H$. In the numerical analysis we eliminate $\phi_{c}$ in equation (2.14) and instead check for the slow-roll conditions violation at each step. ### 2.4 Tunneling and eternal inflation Most of the inflationary potentials are of the single-minimum type such as Starobinsky inflation and alpha-attractors. There are however, potentials which are of type depicted on figure 1 and possess a various minima. In such models, the inflation can become eternal due to tunnelling to the false vacua. When the vacua are degenerate enough, the tunneling dominates over quantum uphill rolling. The tunnelling goes in the opposing direction to the old inflation scenario [76] as showed on figure 1. As it will turn out this is the dominant effect for the model discussed in section 4.3. The eternal inflation mechanism discussed in the previous sections relies on the local shape of the potential and cannot provide an accurate description in that case. In order to quantitatively derive predictions for this new effect, we shall rely on the first passage formalism [84, 85] instead, and apply it to the eternal inflation considerations. Figure 1: Left: field initially placed at the maximum of the potential may decay towards one of the two vacua: at $\phi_{-}$ with probability $p_{-}$ and at $\phi_{+}$ with probability $p_{+}$. Right: field initially placed at $\phi_{0}<\phi_{max}$ may tunnel trough the barrier towards $\phi_{+}$ with probability $p_{+}(\phi_{0})$. Analogous tunneling from "plus" to "minus" side is also possible. Given the initial value of the field $\phi_{0}$ being between $\phi_{-}$ and $\phi_{+}$, the probability that it reaches $\phi_{+}$ before $\phi_{-}$ and $\phi_{-}$ before $\phi_{+}$, denoted respectively $p_{+}(\phi_{0})$ and $p_{-}(\phi_{0})$, obeys the following equation: $\displaystyle vp^{\prime\prime}_{\pm}(\phi)-\frac{v^{\prime}}{v}p^{\prime}_{\pm}(\phi)=0,$ (2.27) with initial conditions: $p_{\pm}(\phi_{\pm})=1$, $p_{\pm}(\phi_{\mp})=0$, where $v=v(\phi)$ is the dimensionless potential: $\displaystyle v(\phi):=\frac{V(\phi)}{24\pi^{2}M_{Pl}^{4}}.$ (2.28) The analytical solution is: $\displaystyle p_{\pm}(\phi_{0})=\pm\frac{\int^{\phi_{0}}_{\phi_{\pm}}e^{-\frac{1}{v(\phi)}}d\phi}{\int^{\phi_{+}}_{\phi_{-}}e^{-\frac{1}{v(\phi)}}d\phi}.$ (2.29) One may also define the probability ratio $R$: $\displaystyle R(\phi_{0}):=\frac{p_{+}(\phi_{0})}{p_{-}(\phi_{0})}=\frac{\int^{0}_{\phi_{-}}e^{-\frac{1}{v(\phi)}}d\phi}{\int^{\phi_{+}}_{\phi_{0}}e^{-\frac{1}{v(\phi)}}d\phi}.$ (2.30) The above integrals may be evaluated numerically. However, if the amplitude of $v(\phi)$ is much smaller than 1, the term $e^{-1/v(\phi)}$ will be extremely small, possibly exceeding machine precision in the computation. Yet, one can use the steepest descent approximation. Consider a potential with the field located initially on the local maximum $\phi_{max}$ with a minimum on each side, as depicted on figure 1. Then the probability ratio $R$ may be evaluated approximately, where the leading contributions to (2.30) come from the values of the field in the neighborhood of $\phi_{max}$. $p_{+}$ and $p_{-}$ gives a probability of evolution realised respectively by the red and the green ball. We get:111For the details of the calculation consult [85]. $\displaystyle R(\phi_{max})\approx 1-\frac{2}{3}\frac{\sqrt{2}}{\pi}\frac{v(\phi_{max})v^{\prime\prime\prime}(\phi_{max})}{|v^{\prime\prime}(\phi_{max})|^{3/2}}.$ (2.31) In this regime, probability of descending into each of the minima $\phi_{-}$ and $\phi_{+}$ is similar, giving $|1-R|\ll 1$. It is possible to start the inflation in the subset of $[\phi_{-},\phi_{+}]$ that would lead to the violation of the slow-roll conditions and tunnel through the potential barrier to the sector dominated by eternal inflation, schematically shown on figure 1. We further analyze this possibility in Sec. 4.3 for a particular effective potential with two vacua, stemming from an asymptotically safe theory. We use equation (2.31) to find the dependence of $R$ on the parameters of the theory and verify the result with direct numerical simulation of the Langevin Equation (2.6) given set of parameters. ## 3 Exemplary models In this section we show the basic application of the conditions (2.23, 2.24) to simple effective potentials stemming from the $\alpha$-attractor models and the Starobinsky inflation. ### 3.1 Alpha-attractor models We start our investigation with the $\alpha$-attractor models [86], a general class of the inflationary models, originally introduced in the context of supergravity. They are consistent with the CMB data, and their preheating phase has been studied on a lattice in [87]. The phenomenological features of these models are described by the Lagrangian: $\displaystyle\frac{1}{\sqrt{-g}}\mathcal{L}_{T}=\frac{1}{2}R-\frac{1}{2}\frac{\partial\phi^{2}}{\left(1-\frac{\phi^{2}}{6\alpha}\right)^{2}}-V(\phi).$ (3.1) Here, $\phi$ is an inflaton and $\alpha$ can take any real, positive value. At the limit $\alpha\xrightarrow{}\infty$ the scalar field becomes canonically normalized, and the theory coincides with the chaotic inflation. Canonical and non-canonical fields are related by the transformation: $\displaystyle\phi=\sqrt{6\alpha}\tanh{\frac{\varphi}{\sqrt{6\alpha}}}.$ (3.2) We further consider T-models, in which the potential of canonically normalised field is given by: $\displaystyle V(\phi)=\alpha\mu^{2}\tanh^{2n}{\frac{\varphi}{\sqrt{6\alpha}}},$ (3.3) where parameter $\mu$ is of order $10^{-5}$. The shape of the potential for $n=1$ was plotted on figure 2. At the large $\phi$ the potential (3.3) is asymptotically flat, this creates the possibility for eternal inflation to occur. Using the first condition (2.23) we have verified, that generally the space is eternally inflating for all initial values of the $\phi$, above certain $\phi_{EI}$. The second eternal condition (2.24) as well as the higher order conditions are satisfied for almost all values of $\phi_{0}$ above $0$, providing no new information. This is a generic feature for all of the models we investigate. For every $\alpha$, $\phi_{0}$ necessary to produce 60 e-folds is safely below $\phi_{EI}$. We have found $\phi_{0}$ by solving slow-roll equation numerically. It is shown on figure 2. The time of inflation larger than 60 Planck-times is unlikely according to the Planck Collaboration data [88]. The values of $\phi_{EI}$ change only slightly with $n$. We may therefore conclude, that $\alpha$-attractor models are consistent with the beginning of "our" pocket universe. However, it is not inconceivable that the field fluctuations in other parts of the early universe had values $\phi_{0}>\phi_{EI}$, driving the eternal inflation. Figure 2: Left: the T-model potential with $n=1$ and various $\alpha$ were depicted. Right: plot of the initial value $\phi_{0}$ necessary for 60 e-folds, as a function of $\alpha$ (blue), as well as the lowest initial value $\phi_{EI}$ of the field, at which the eternal inflation "kicks in" (yellow). ### 3.2 Starobinsky model Solutions stemming from Einstein-Hilbert action predict initial singularity. In 1980 Starobinsky proposed a model [89], where pure modified gravitational action can cause non-singular evolution of the universe, namely: $\displaystyle S=\frac{1}{2}\int\sqrt{|g|}d^{4}x\left(M^{2}_{p}R+\frac{1}{6M^{2}}R^{2}\right),$ (3.4) this can be rewritten to the effective potential form, with: $\displaystyle V\left(\phi\right)=V_{0}\left(1-\exp\left(-\sqrt{\frac{2}{3}}\frac{\phi}{M_{Pl}}\right)\right)^{2}.$ (3.5) The inflation begins on a plateau at large $\phi$. The field rolls towards a minimum at $\phi=0$, where the oscillatory reheating phase occurs. It has been estimated from the CMB data, that during the inflation the volume of the universe has grown by approximately 60 e-folds. This corresponds to the initial condition $\phi_{0}=5.5$ $M_{Pl}$, without taking into account quantum gravity effects. It is possible to perturbatively recover information about the shape of the potential from the CMB, for details see [90]. Amplitude of the scalar power spectrum $A_{s}=2\times 10^{-9}$ fixes the value $V_{0}=8.12221\times 10^{-11}$ $M^{4}_{Pl}$, via the relation: $\displaystyle V_{0}=24\pi^{2}\epsilon(\phi_{0})A_{s}.$ (3.6) Applying the analytical eternal inflation conditions (2.23, 2.24) to the Starobinsky potential, the initial value of the field, above which the eternal inflation occurs, has been estimated to be $\phi_{0}=16.7$ $M_{Pl}$. It has been found, that decay rate $\Gamma$ decreases approximately exponentially with $\phi_{0}$. Our numerical simulation confirms the analytical prediction discussed in [1] within a sample of 10000 simulations. They are performed as described in the previous section. An exemplary numerical evolution of the Langevin equation (2.6) is shown in the figure 3. The linear fit in the early times shows the exponential decay of the inflation in the slow-roll regime. Nevertheless, for eternal inflation scenario and realistic phenomenology, this model requires the transplanckian values of the fields in order to reproduce the correct tensor to scalar ratio $r$, amplitudes and spectral tilt $n_{s}$. Hence the Starobinsky inflation will be affected, possibly invalidated by the quantum gravity fluctuations. The leading log corrections have been studied in [31], which we as well study in the context of eternal inflation. Yet, due to the large field values required for eternal inflation to occur, one should seek a theory predictive up to an arbitrary large energy scale, which we discuss in the next section. Figure 3: Left: exemplary field evolution for the Starobinsky model has been plotted. Green plot shows the solution to the classical slow-roll equation, and the black plot is the Langevin solution. The inflation ends, when the slow-roll parameter reaches 1. Values of the field are given in $M_{Pl}$. Right: the time dependence of the probability, that inflation still occurs. In the slow-roll regime, the probability decays exponentially. Linear fit slope, the decay rate is around $\Gamma=0.15$. Red, dashed line denotes eternal inflation threshold with slope $3H$ of order $10^{-5}$. Since $3H<\Gamma$, initial condition $\phi_{0}=3$ $M_{Pl}$ is an example of non-eternally inflating universe. ## 4 Eternal inflation in asymptotically safe models In this chapter, as a warm up we study effective corrections to the Starobinsky inflation, providing a different behavior at the large field values. Later we show, that RG-improvement of Starobinsky model proposed in [91], see also [92, 93] for review, closely related to the $R+R^{2}$ renormalizable Fradkin-Tseytlin gravity [94] produces a branch of inflationary potential entirely dominated by eternal inflation. We find the initial values of the inflaton such that inflation becomes eternal, for the remaining branch, as a function of theory parameters. Finally we show, that the possibility of tunneling through the potential barrier present in [95] becomes a new mechanism for eternal inflation. In all of the asymptotically safe inflationary theories, eternal inflation is present as a consequence of asymptotic flatness of the effective potential. ### 4.1 Quantum corrections to the Starobinsky model Below Planck scale the gravitational constant $G_{N}$ has a vanishing anomalous dimension and the $R^{2}$ has a coefficient that runs logarithmically [96] (this comes from the fact that $R^{2}$ is dimensionless in $4$ dimensions). Hence, one can motivate various quantum corrected inflationary models, such as [28, 29, 30, 31]. In particular, the leading-log corrections to the Starobinsky model are given by [31]: $\displaystyle\mathcal{L}_{eff}=\frac{M_{Pl}^{2}R}{2}+\frac{\frac{a}{2}R^{2}}{1+b\ln(\frac{R}{\mu^{2}})}+\mathcal{O}(R^{3}).$ (4.1) In order to find the Einstein frame potential for this model a few steps need to be taken. First, we use the conformal transformation [97]. Then, following the next transformation for the Ricci scalar and the metric determinant we get the Einstein frame action: $\displaystyle S=\int d^{4}x\sqrt{-g_{E}}\left[\frac{M_{Pl}^{2}}{2}R_{E}-\frac{1}{2}g_{E}^{\mu\nu}\left(\partial_{\mu}\phi_{E}\right)\left(\partial_{\nu}\phi_{E}\right)-V_{E}(\phi_{E})\right],$ (4.2) which depends on the sought potential $V_{E}$, that can be further obtained:222Here, the effective action differs from [31] due to the introduction of auxiliary numerical parameter $e\approx 2.81$. Nevertheless, the dynamics stemming from each of the potentials are equivalent. $\displaystyle V_{E}(\Phi)=\frac{M_{Pl}^{4}}{2}\frac{a\Phi^{2}\left(1+b\ln\left(\frac{\Phi}{\mu^{2}}\right)\right)^{2}\left(1+b\ln\left(\frac{\Phi}{e\mu}^{2}\right)\right)}{\left\\{M_{p}^{2}\left(1+b\ln\left(\frac{\Phi}{\mu^{2}}\right)\right)^{2}+2a\Phi\left(1+b\ln\left(\frac{\Phi}{\sqrt{e}\mu^{2}}\right)\right)\right\\}^{2}},$ (4.3) with $\phi_{E}$ given by $F(\phi_{E})=M_{Pl}^{2}\exp\left(\sqrt{\frac{2}{3}}\frac{\phi_{E}}{M_{Pl}}\right)=M_{P}^{2}+\frac{a\Phi[2-b+2b\ln(\Phi/\mu^{2})]}{[1+2b\ln(\Phi/\mu^{2})]^{2}},$ (4.4) yet the transformation between $\Phi$ and $\phi_{E}$ is non-invertible. By taking into account the COBE normalization, we can treat $b$ as a free parameter and fix $a(b)$. For $b=0$ one obtains the usual Starobinsky model, and for $b\ll 1$ one gets the model discussed in [29] $R^{2}(1+\beta\ln R)$ with the potential given by Lambert W function (so the same as for model discussed in section 4.3) and approximated in the limit $\beta\ll 1$ as $V\approx\frac{V_{s}}{1+b/(2\alpha)+\beta/\alpha\ln[(e^{\tilde{\chi}}-1)/2\alpha]},$ (4.5) where $V_{s}$ is the Starobinsky potential, $\tilde{\chi}$ is the Einstein frame field and $\alpha(\beta)$, where we have kept the original notation. From the plots 4, 4, one can see that both of the models should give similar inflationary observables as in the Starobinsky inflation. On the other hand the eternal inflation in this models will be quite different. These models for $\beta<0$ and $b>0$ have the potentials that are non-flat for large field values, while for $\beta>0$ potential depicted on Figure 4 have the runaway behaviour, so different asymptotic behaviour, which is discussed in Section 4.3. This makes those potentials quantitatively different from the Starobinsky model in the context of eternal inflation and suggest, that eternal inflation cannot take place in those models. To be concrete, we have checked that eternal inflation for model described by 4.3 takes place at $\Phi\approx 1000\,M_{Pl}$, which is far beyond the applicability of the model. So now we turn to the inflationary models stemming from the asymptotic safety. Figure 4: Left: Plots of the potential (4.3) for various $b$ parameters. Right: Potential (4.5) in dependence of $\beta$ parameter values. Approach towards infinity in case of $\beta>0$ is visible. ### 4.2 RG-improved Starobinsky inflation Renormalization Group improvement is a procedure of identifying and replacing the RG scale $k^{2}$ with a physical scale. It incorporates leading-order quantum effects in the dynamics of classical system. In the case of gravity, running of coupling constants in Einstein-Hilbert action results in additional contribution to the field equations from the gravitational energy-momentum tensor [93]. In the de Sitter-type setting $k^{2}\sim R$ is the unique identification of the physical scale dictated by Bianchi identities [93]. Such replacement in the scale-dependent Einstein-Hilbert action generates an effective $f(R)$ action, whose analytical expression is determined by running of the gravitational couplings. RG-improvement could solve classical black hole singularity problem [98, 99], gives finite entanglement entropy [100] and generates inflationary regime in quantum gravity [37]. In this section, we study the asymptotically safe inflation based on RG- improved quadratic gravity Lagrangian, considered in [91, 93]: $\mathcal{L}_{k}=\frac{1}{16\pi g_{k}}\left(R-2\lambda_{k}k^{2}\right)-\beta_{k}R^{2},$ (4.6) with the running dimensionless couplings $g_{k}$, $\lambda_{k}$, $\beta_{k}$ being the three relevant directions of the theory, with running given by [101]: $g_{k}=\frac{6\pi c_{1}k^{2}}{6\pi\mu^{2}+23c_{1}(k^{2}-\mu^{2})},\quad\quad\beta_{k}=\beta_{\ast}+b_{0}\left(\frac{k^{2}}{\mu^{2}}\right)^{-\theta_{3}/2},$ (4.7) where $\mu$ is the infrared renormalization point such that $c_{1}=g_{k}(k=\mu)$ and $c_{1}$ and $b_{0}$ are integration constants. We introduce a parameter $\alpha$ as $\alpha=-2\mu^{\theta_{3}}b_{0}/M_{P}^{2},$ (4.8) that measures the departure from the non-gaussian fixed point (NGFP). One may find the behavior of the couplings near the NGFP and substitute the appropriate expressions to the Lagrangian using the RG-improvement and the following identification of scale $k^{2}=\xi R$, where $\xi$ is an arbitrary parameter of order one. Figure 5: Left: $V_{+}(\phi)$ plot for various $\alpha$ and fixed $\Lambda=1$ is shown. Right, logarithmic dependence of initial field value above which eternal inflation occurs on parameter $\alpha$ has been found. Blue points were evaluated via (2.23). Following [91] we shall assume $\theta_{3}=1$, then the transformation from the Jordan to the Einstein frame yields an effective potential [91, 93]: $\displaystyle\begin{split}V_{\pm}=&\frac{m^{2}e^{-2\sqrt{\frac{2}{3}}\kappa\phi}}{256\kappa}\Bigg{\\{}\vphantom{6\alpha^{3}\sqrt{\alpha^{2}+16e^{\sqrt{\frac{2}{3}}\kappa\phi}-16}}192(e^{\sqrt{\frac{2}{3}}\kappa\phi}-1)^{2}-3\alpha^{4}+128\Lambda\\\ &-\sqrt{32}\alpha\left[(\alpha^{2}+8e^{\sqrt{\frac{2}{3}}\kappa\phi}-8)\pm\alpha\sqrt{\alpha^{2}+16e^{\sqrt{\frac{2}{3}}\kappa\phi}-16}\right]^{\frac{3}{2}}\\\ &-3\alpha^{2}(\alpha^{2}+16e^{\sqrt{\frac{2}{3}}\kappa\phi}-16)\mp 6\alpha^{3}\sqrt{\alpha^{2}+16e^{\sqrt{\frac{2}{3}}\kappa\phi}-16}\Bigg{\\}}\end{split},$ (4.9) where the only free parameters are cosmological constant $\Lambda$ and $\alpha$ after the CMB normalization we perform below. The $V_{+}$ branch predicts the reheating phase, figure 5 shows its plot for various $\alpha$. Similarly as in the case of Starobinsky inflation, we have denoted $V_{0}$ the constant part of the potential at infinity $V(\phi\xrightarrow{}\infty)=V_{0}=\frac{3m^{2}}{4\kappa^{2}}$ and fixed it with CMB data by the relation (3.6). For example, given $\alpha=2.8$, $\Lambda=1$ the plateau value is equal to $V_{0}=1.99\times 10^{-10}$ $M^{4}_{Pl}$, hence one may fix the mass parameter $m=2\times 10^{14}$ GeV. Now we shall investigate the eternal inflation conditions given by (2.23, 2.24). These conditions restricts the initial value of the field. We search for $\phi_{0}$ above which the eternal inflation occurs, as a function of the theory parameters. We have also found, that initial value above which eternal inflation occurs does not depend on the cosmological constant. It is due to the fact that $\Lambda$ only shifts the minimum of the potential and does not affect the large-field behavior of the system. Analytical conditions for EI have been checked for a set of $\alpha$ and depicted on figure 5. The initial value of the field, depends logarithmically on $\alpha$. The reason for that behaviour is the following. In the large field expansion: $\displaystyle V_{\pm}(\phi)=V_{plateau}-128V_{0}\alpha e^{-\frac{1}{2}\sqrt{\frac{3}{2}}\phi},$ (4.10) and by the substitution $\tilde{\phi}=e^{-\frac{1}{2}\sqrt{\frac{3}{2}}\phi}$ the potential reduces to the linear hilltop model, which justifies the usage of formulae (2.23, 2.24) and the functional form of $\phi_{0}(\alpha)$. The results were also confirmed by the numerical simulations. For example, given $\Lambda=1,\,\alpha=1.6$, the analytical considerations predict $\phi_{EI}=22.6\,M_{Pl}$. The direct numerical simulation for this set of parameters yields $\Gamma=0.0001$ $M_{Pl}$, and $3H=0.0003$ $M_{Pl}$, meaning that the eternal inflation begins slightly below the expected value $\phi_{EI}$. The plateau of (4.9) at large field values is a characteristic feature of effective inflationary potentials stemming from the asymptotically safe theories. It is dominated by eternal inflation and may suggest a deeper relation between the asymptotic safety of quantum gravity and multiverse. ### 4.3 Large N-dynamics and (eternal) inflation In this section we investigate model in which inflation is driven by an ultraviolet safe and interacting scalar sector stemming from a new class of non-supersymmetric gauge field theories. We consider a $\mathrm{SU}(N_{C})$ gauge theory, with $N_{F}$ Dirac fermions and interacting with an $N_{F}$ $\times$ $N_{F}$ complex scalar matrix $H_{ij}$ that self interacts, described in [95]. The Veneziano limit ($N_{F}\to+\infty$, $N_{C}\to+\infty$, $N_{F}/N_{C}=\mathrm{const}$) is taken such that the ratio $N_{F}/N_{C}$ becomes a continuous parameter [70]. The action in Jordan frame has the following form: $S_{J}=\int d^{4}x\sqrt{-g}\left\\{-\frac{M^{2}+\xi\phi^{2}}{2}R+\frac{g^{\mu\nu}}{2}\partial_{\mu}\phi\partial_{\nu}\phi- V_{\mathrm{iUVFP}}\right\\},$ (4.11) where the leading logarithmically resummed potential $V_{iUVFP}$ is given by: $V_{\mathrm{iUVFP}}(\phi)=\frac{\lambda_{*}\phi^{4}}{4N_{f}^{2}\left(1+W(\phi)\right)}\left(\frac{W(\phi)}{W(\mu_{0})}\right)^{\frac{18}{13\delta}},$ (4.12) where $\lambda_{*}=\delta\frac{16\pi^{2}}{19}(\sqrt{20+6\sqrt{23}}-\sqrt{23}-1)$ is positive quartic coupling at the fixed point, $\phi$ is the real scalar field along the diagonal of $H_{ij}=\phi\delta_{ij}/\sqrt{2N_{f}}$ and $\delta=N_{F}/N_{C}-11/2$ is the positive control parameter, $W(\phi)$ is the Lambert function solving the transcendent equation $z=W\exp W,$ (4.13) with $z(\mu)=\left(\frac{\mu_{0}}{\mu}\right)^{\frac{4}{3}\delta\alpha*}\left(\frac{\alpha*}{\alpha_{0}}-1\right)\exp\left[\frac{\alpha*}{\alpha_{0}}-1\right].$ (4.14) The parameter $\alpha*=\frac{26}{57}\delta+O(\delta^{2})$ is the gauge coupling at its UV fixed point value and $\alpha_{0}=\alpha(\mu_{0})$ is the same coupling at a reference scale $\mu_{0}$. A conformal transformations allows to rewrite the action from Jordan to Einstein frame. Assuming a single field slow-roll inflation, we examine inflationary predictions of the potential and compute the slow-roll parameters: $\epsilon=\frac{M_{Pl}^{2}}{2}\left(\frac{dU/d\chi}{U}\right)^{2},\quad\quad\eta=M_{Pl}^{2}\frac{d^{2}U/d\chi^{2}}{U},$ (4.15) where $U=V_{\mathrm{iUVFP}}/\Omega^{4}$, with $\Omega^{2}=(M^{2}+\xi\phi^{2})/M_{Pl}^{2}$ being the conformal transformation of the metric, and $\chi$ is the canonically normalized field in the Einstein frame. We assume, that $M=M_{Pl}$. Inflation ends when the slow-roll conditions are violated, that is when $\epsilon(\phi_{end})$ or |$\eta(\phi_{end})$| = 1. We analyze the non-minimal case, where the coupling $\xi$ is non-vanishing. The potential $U$ is given by: $U=\frac{V_{\mathrm{iUVFP}}}{\Omega^{4}}\approx\frac{\lambda_{*}\phi^{4}}{4N_{F}^{2}\left(1+\frac{\xi\phi^{2}}{M_{Pl}^{2}}\right)^{2}}\left(\frac{\phi}{\mu_{0}}\right)^{-\frac{16}{19}\delta}\mathrm{.}$ (4.16) Figure 6: Left: the non-minimally coupled potential as a function of $\phi$ for $\delta=0.1$, $\xi=1/6$ and $\mu_{0}=10^{-3}M_{Pl}$. There is a maximum at $\phi_{max}=16.7\,M_{Pl}$. Right: For the same set of parameters, we plot the first eternal inflation condition as a function of $\phi$ (blue curve) and the eternal inflation bound (yellow curve). Inflation becomes eternal if the blue curve is below the yellow one. At $\phi_{max}$ the first derivative of the potential vanishes and (2.23) predicts a narrow window for the eternal inflation. In the large field limit $\phi$ $\gg$ $M_{Pl}/\sqrt{\xi}$ the $\phi^{4}$ term in the numerator cancels against the term in the denominator. In this limit, the quantum corrections dictate the behaviour of the potential, which is found to decrease as: $\frac{\lambda_{*}M_{Pl}^{4}}{4N_{F}^{2}\xi^{2}}\left(\frac{\phi}{\mu_{0}}\right)^{-\frac{16}{19}\delta}\mathrm{.}$ (4.17) The non-minimal coupled potential has one local maximum and two minima. The region to the left of the maximum is the region, where the inflation can be brought to an end and the reheating takes place [102]. To the right of the maximum, the inflation becomes classically eternal. For large values of $\phi$ the potential flattens out and the slow-roll conditions are not violated. Numerical solutions to the FP equation shows that the is no possibility of eternal inflation in that region, since it is an unstable maximum and hence any quantum fluctuation will drop it from that position. Furthermore due to steepness of the potential around this maxima there is no possibility for the field to remain in that region. Let us now investigate the analytical eternal inflation conditions. Similarly as in the Starobinsky model, the second condition (2.24) is always satisfied. The first condition (2.23) is illustrated on the figure 6. There is a peak for $\phi$ = $\phi_{max}$ = 16.7 $M_{Pl}$, due to the vanishing derivative and if we "zoom in", the analytical condition allows for eternal inflation in the close neighbourhood of $\phi_{max}$. We have verified numerically this is not a sustainable attractor of eternal inflation. A field that starts evolution at $\phi_{max}$ will leave this region, as it cannot climb further up-hill. Nevertheless, eternal inflation may still occur due to the quantum tunnelling through the potential barrier. #### Tunneling through the potential barrier As described in section 2.4, if the potential has multiple vacua, quantum tunneling through the potential barrier is expected. The non-minimal coupling potential (4.16) belongs in this class. The question is, whether tunnelling from the non-eternal inflation region of $\phi<\phi_{max}$ to the region of classical eternal inflation $\phi>\phi_{max}$ is possible. We start with investigating the fate of the field initially placed at the peak $\phi_{0}=\phi_{max}$ of the potential depicted on figure 6. By the virtue of the steepest descent approximation at the maximum equation (2.31) may be employed. The resulting ratio of probabilities of the right-side descent to the left-side descent $R(\alpha,\xi)=\frac{p_{+}}{p_{-}}$, as a function of parameter $\alpha$ and the non-minimal coupling constant $\xi$ was calculated directly from the formula (2.31) without the need of numerical simulation of the Langevin equation. It is presented on figure 7. Due to the complexity of the potential (4.16) its maximum was found numerically and then employed in (2.31). Figure 7 shows how the maximum changes with parameters. As expected, the ratio $R(\alpha,\xi)$ is close to 1 and favors the right-side (left-side) of the potential for large (small) values of the parameters. The biggest ratio emerges at large values of the parameters $\xi$ and $\delta$, since the potential is "step-like" and highly asymmetric. It is monotonically decreasing with the values of $\phi_{max}$, for which the potential has a maximum, see figure 8. Figure 7: Left: probability ratio $R=\frac{p_{+}}{p_{-}}$ of descending from the maximum towards the right minimum ($p_{+}$) and the left minimum ($p_{-}$), as a function of theory parameters $\xi$ and $\delta$. For the small values of the parameters it is more probable to fall from the maximum towards $\phi_{-}=0$ with non-eternal inflation, while for the large values of parameters, minimum at $\phi_{+}=\infty$ is favored, resulting in eternally inflating universe. Right: the value of the field at which, the potential is maximal. The above figures are qualitatively similar because for the small values of $\phi_{max}$, the effective potential is highly asymmetric ("step-like"). This brakes the symmetry between the right and left descend probability. In order to verify the accuracy of the relation (2.31), we have performed numerical simulation of the discretized Langevin equation (2.26) with initial condition $\phi_{0}=\phi_{max}$. For example it was found, that for the set of the parameters $N_{F}=10$, $\mu=10^{-3}M_{Pl}$, $\xi=\frac{1}{6}$, $\delta=0.1$ the steepest descent approximation yields $R=0.92$, and the numerical analysis results in $R=0.97$, which proves good accuracy of the analytical formula. One may wonder, how does the probability $p_{\pm}(\phi_{0})$ depends on the departure from the maximum $\phi_{0}\neq\phi_{max}$. The analytical answer is given by (2.29). As we have checked numerically inflation becomes eternal, when tunneling probability is non-zero, as depicted on Figure 8. To bypass the numerical calculation of the integral (2.29) we employ the direct numerical simulation of the Langevin equation. This time however, we do not seek for the time evolution of the inflaton. Rather than creating histograms of the count of inflationary events at a given timestep, we simply track the probabilities $p_{+}$ and $p_{-}$. We say that the particle tunnelled through the potential barrier contributing to $p_{+}$ if the evolution starts at $\phi_{0}<\phi_{max}$ and proceeds to arbitrarily large field values after a long time. For each point at figure 8 the probability has been calculated on the sample of 10000 simulations. As expected, choosing values of $\phi_{0}$ smaller than $\phi_{max}$ lowers the probability of the tunneling to the right side of the barrier. Moreover, the probability of tunneling decreases linearly with the distance to the maximum. The result of the simulation for the set of parameters $N_{F}=10,\quad\mu=10^{-3}M_{Pl},\quad\xi=\frac{1}{6},\quad\delta=0.1$ (4.18) is shown on figure 8. The green line corresponds to the green ball on figure 1 and shows the probability of tunneling through the barrier (as in figure 1) as a function of proximity to the maximum $\phi_{0}\neq\phi_{max}$. The red line corresponds to the red ball on figure 1 and shows the probability of rolling towards the minimum at infinity. The rolling is also a stochastic process, as the tunneling in the opposite direction is possible. The probability distribution of tunneling in either direction is not a symmetric process. Notice, that the initial condition, for which $p_{+}=\frac{1}{2}$ is shifted to the right of $\phi_{max}$. This means, that starting from the maximum, it is slightly more probable to land in $\phi_{-}$. There is a point, below which the green ball cannot tunnel $p_{+}=0$ (for the set of parameters given by (4.18), at $9.2M_{Pl}$), and the other limiting case (at $24.8$ $M_{P}$), when the red ball cannot tunnel $p_{-}=0$. Hence, for every initial value of the field above $9.2$ Planck Masses, there is a non-zero probability of eternal inflation. On the other hand, for $\phi_{0}=9.2\,M_{Pl}$ and parameters given by (4.18), the inflation classically produces roughly 54 e-folds, depending on the reheating time [102], and is in the agreement with CMB data. This shows that the model is on a verge of being eternally inflating, which may point out to the interesting phenomenology. To sum up the critical point of our analysis is that the analytical conditions (2.23, 2.24) did not allow for eternal inflation, even though the tunneling process may evolve any initial point above 9.2 Planck Masses to $\phi_{+}=\infty$, and not violate the slow-roll conditions. This shows that conditions (2.23,2.24) are not well suited for the multiple minima models and cannot contain the full information of the global influence of quantum fluctuations in the early universe. Figure 8: Left: Linear probability distribution of tunneling (green side) and rolling (red side) towards the minimum at infinity as. The data points have been directly simulated. Right: Probability ratio $R$, evaluated with (2.31), is a monotonically decreasing with the value of the maximum of the potential in the steepest descent approximation ## 5 Conclusions Eternal inflation remains a conceptual issue of the inflationary paradigm. The creation of scattered, causally disconnected regions of spacetime - the multiverse is not confirmed observationally and raises question about the inflationary predictions [77]. Hence, one may impose the no eternal inflation principle [1] to restrict free parameters and the initial conditions. We have investigated popular inflationary models, and have found that in principle, eternal inflation is present at every asymptotically flat effective potential for large field values, assuming the ergodicity of the system. The finite inflationary time of our pocket universe serves as a consistency condition of the multiverse predictions. In section 3.1, we verified that $\alpha$-attractor T-models are consistent from this point of view. If the initial value of the scalar field driving inflation is above the Planck scale, a UV-completeness of a given model is necessary. Starobinsky inflation stemming from $R^{2}$ gravity gives around 60 e-folds for $\phi_{0}$=5.5 $M_{pl}$. We have considered the effective quantum corrections to the Starobinsky inflation based on the qualitative behavior of the running coupling constants. Next, the RG-improvement of $R+R^{2}$ Lagrangian was studied in this case and we have found that field values required for eternal inflation are typically higher than the ones for the Starobinsky case. The flatness of the potential and possibility of eternal inflation seems to be a signature mark of asymptotically safe UV completions, in contradistinction to the effective theory corrections. We have checked that [31] $\Phi\sim 1000$ $M_{Pl}$ in order to get eternal inflation, which is far beyond the applicability of the model. Furthermore we have found, that for potentials with multiple vacua, tunneling through potential barriers provides a new mechanism for eternal inflation. So in order to understand the inflationary dynamics one cannot simply cut the potential at the maximum. The $\mathrm{SU}(N)$ Gauge theory with Dirac fermions provides an example for such behavior. The probability of tunneling to the side dominated by eternal inflation becomes negligible few Planck Masses away from the peak of the potential. Yet the fixed point values of the couplings and possibly the shape of the potential can be obscured by the quantum gravity effects and this shall be investigated elsewhere. Our analysis reveals that there is no obstruction for the multiverse scenario in the asymptotically safe models. Yet its occurrence depends on the initial conditions for the inflationary phase and the matching to the observational data, tying these three profound issues together. On the other hand, in AS models these questions can have intriguing answers by the finite action principle [78]. ## Acknowledgments We thank J. Reszke and J. Łukasik for participating in the early stages of this project. We thank G. Dvali, A. Eichhorn, M. Pauli, A. Platania, T. Rudelius, S. Vagnozzi and Z.W. Wang for fruitful discussions and extensive comments on the manuscript. Work of J.H.K. was supported by the Polish National Science Center (NCN) grant 2018/29/N/ST2/01743. J.H.K. would like to acknowledge the CP3-Origins hospitality during this work. The computational part of this research has been partially supported by the PL-Grid Infrastructure. ## References * [1] T. Rudelius, “Conditions for (No) Eternal Inflation,” JCAP, vol. 08, p. 009, 2019. * [2] C. Vafa, “The String landscape and the swampland,” 9 2005. * [3] H. Ooguri and C. Vafa, “On the Geometry of the String Landscape and the Swampland,” Nucl. Phys. B, vol. 766, pp. 21–33, 2007. * [4] G. Obied, H. Ooguri, L. Spodyneiko, and C. Vafa, “De Sitter Space and the Swampland,” 6 2018. * [5] P. Agrawal, G. Obied, P. J. Steinhardt, and C. Vafa, “On the Cosmological Implications of the String Swampland,” Phys. Lett. B, vol. 784, pp. 271–276, 2018. * [6] A. Achúcarro and G. A. Palma, “The string swampland constraints require multi-field inflation,” JCAP, vol. 02, p. 041, 2019. * [7] W. H. Kinney, S. Vagnozzi, and L. Visinelli, “The zoo plot meets the swampland: mutual (in)consistency of single-field inflation, string conjectures, and cosmological data,” Class. Quant. Grav., vol. 36, no. 11, p. 117001, 2019. * [8] G. Dvali and C. Gomez, “Quantum Compositeness of Gravity: Black Holes, AdS and Inflation,” JCAP, vol. 01, p. 023, 2014. * [9] G. Dvali and C. Gomez, “Quantum Exclusion of Positive Cosmological Constant?,” Annalen Phys., vol. 528, pp. 68–73, 2016. * [10] G. Dvali, C. Gomez, and S. Zell, “Quantum Break-Time of de Sitter,” JCAP, vol. 06, p. 028, 2017. * [11] G. Dvali, “$S$-Matrix and Anomaly of de Sitter,” Symmetry, vol. 13, no. 1, p. 3, 2020. * [12] J. H. Kwapisz and K. A. Meissner, “Asymptotic safety and quantum gravity amplitudes,” 5 2020. * [13] O. Lauscher and M. Reuter, “Fractal spacetime structure in asymptotically safe gravity,” JHEP, vol. 10, p. 050, 2005. * [14] O. Lauscher and M. Reuter, “Asymptotic safety in quantum Einstein gravity: Nonperturbative renormalizability and fractal spacetime structure,” in 14th Oporto Meeting on Geometry, Topology and Physics: Mathematical Aspects of Quantum Field Theory, pp. 293–313, 11 2005. * [15] A. H. Guth, “Eternal inflation and its implications,” J. Phys. A, vol. 40, pp. 6811–6826, 2007. * [16] M. C. Johnson and J.-L. Lehners, “Cycles in the Multiverse,” Phys. Rev. D, vol. 85, p. 103509, 2012. * [17] J.-L. Lehners, “Eternal Inflation With Non-Inflationary Pocket Universes,” Phys. Rev. D, vol. 86, p. 043518, 2012. * [18] G. León, “Eternal inflation and the quantum birth of cosmic structure,” Eur. Phys. J. C, vol. 77, no. 10, p. 705, 2017. * [19] H. Matsui and F. Takahashi, “Eternal Inflation and Swampland Conjectures,” Phys. Rev. D, vol. 99, no. 2, p. 023533, 2019. * [20] Z. Wang, R. Brandenberger, and L. Heisenberg, “Eternal Inflation, Entropy Bounds and the Swampland,” Eur. Phys. J. C, vol. 80, no. 9, p. 864, 2020\. * [21] J. J. Blanco-Pillado, H. Deng, and A. Vilenkin, “Eternal Inflation in Swampy Landscapes,” JCAP, vol. 05, p. 014, 2020. * [22] C.-M. Lin, “Topological Eternal Hilltop Inflation and the Swampland Criteria,” JCAP, vol. 06, p. 015, 2020. * [23] O. Hohm and B. Zwiebach, “Non-perturbative de Sitter vacua via $\alpha^{\prime}$ corrections,” Int. J. Mod. Phys. D, vol. 28, no. 14, p. 1943002, 2019\. * [24] O. Hohm and B. Zwiebach, “Duality invariant cosmology to all orders in $\alpha$’,” Phys. Rev. D, vol. 100, no. 12, p. 126011, 2019. * [25] T. Banks, “On the Limits of Effective Quantum Field Theory: Eternal Inflation, Landscapes, and Other Mythical Beasts,” 10 2019. * [26] M.-S. Seo, “Eternal inflation in light of Wheeler-DeWitt equation,” JCAP, vol. 11, p. 007, 2020. * [27] J. F. Donoghue, “General relativity as an effective field theory: The leading quantum corrections,” Phys. Rev. D, vol. 50, pp. 3874–3888, 1994. * [28] A. Codello, J. Joergensen, F. Sannino, and O. Svendsen, “Marginally Deformed Starobinsky Gravity,” JHEP, vol. 02, p. 050, 2015. * [29] I. Ben-Dayan, S. Jing, M. Torabian, A. Westphal, and L. Zarate, “$R^{2}\log R$ quantum corrections and the inflationary observables,” JCAP, vol. 09, p. 005, 2014. * [30] K. Bamba, G. Cognola, S. D. Odintsov, and S. Zerbini, “One-loop modified gravity in a de Sitter universe, quantum-corrected inflation, and its confrontation with the Planck result,” Phys. Rev. D, vol. 90, no. 2, p. 023525, 2014. * [31] L.-H. Liu, T. Prokopec, and A. A. Starobinsky, “Inflation in an effective gravitational model and asymptotic safety,” Phys. Rev. D, vol. 98, no. 4, p. 043505, 2018. * [32] S. Weinberg, Ultraviolet divergences in quantum theories of gravitation, pp. 790–831. 1 1980. * [33] M. Reuter, “Nonperturbative evolution equation for quantum gravity,” Phys. Rev. D, vol. 57, pp. 971–985, 1998. * [34] W. Souma, “Nontrivial ultraviolet fixed point in quantum gravity,” Prog. Theor. Phys., vol. 102, pp. 181–195, 1999. * [35] O. Lauscher and M. Reuter, “Ultraviolet fixed point and generalized flow equation of quantum gravity,” Phys. Rev. D, vol. 65, p. 025013, 2002. * [36] M. Reuter and F. Saueressig, “Renormalization group flow of quantum gravity in the Einstein-Hilbert truncation,” Phys. Rev. D, vol. 65, p. 065016, 2002. * [37] A. Eichhorn, “An asymptotically safe guide to quantum gravity and matter,” 2019\. * [38] N. Dupuis, L. Canet, A. Eichhorn, W. Metzner, J. Pawlowski, M. Tissier, and N. Wschebor, “The nonperturbative functional renormalization group and its applications,” 6 2020. * [39] S. Weinberg, “Asymptotically safe inflation,” Physical Review D, vol. 81, Apr 2010. * [40] M. Reuter and F. Saueressig, “From big bang to asymptotic de Sitter: Complete cosmologies in a quantum gravity framework,” JCAP, vol. 09, p. 012, 2005\. * [41] M. Reuter and F. Saueressig, “Quantum Einstein Gravity,” New J. Phys., vol. 14, p. 055022, 2012. * [42] A. Eichhorn, Y. Hamada, J. Lumma, and M. Yamada, “Quantum gravity fluctuations flatten the Planck-scale Higgs potential,” Phys. Rev. D, vol. 97, no. 8, p. 086004, 2018. * [43] M. Shaposhnikov and C. Wetterich, “Asymptotic safety of gravity and the Higgs boson mass,” Phys. Lett. B, vol. 683, pp. 196–200, 2010. * [44] O. Zanusso, L. Zambelli, G. Vacca, and R. Percacci, “Gravitational corrections to Yukawa systems,” Phys. Lett. B, vol. 689, pp. 90–94, 2010\. * [45] J.-E. Daum, U. Harst, and M. Reuter, “Running Gauge Coupling in Asymptotically Safe Quantum Gravity,” JHEP, vol. 01, p. 084, 2010. * [46] S. Folkerts, D. F. Litim, and J. M. Pawlowski, “Asymptotic freedom of Yang-Mills theory with gravity,” Phys. Lett. B, vol. 709, pp. 234–241, 2012. * [47] N. Christiansen, D. F. Litim, J. M. Pawlowski, and A. Rodigast, “Fixed points and infrared completion of quantum gravity,” Phys. Lett. B, vol. 728, pp. 114–117, 2014. * [48] Z.-W. Wang, F. S. Sage, T. G. Steele, and R. B. Mann, “Asymptotic Safety in the Conformal Hidden Sector?,” J. Phys. G, vol. 45, no. 9, p. 095002, 2018\. * [49] A. Eichhorn, A. Held, and J. M. Pawlowski, “Quantum-gravity effects on a Higgs-Yukawa model,” Phys. Rev. D, vol. 94, no. 10, p. 104027, 2016. * [50] F. Grabowski, J. H. Kwapisz, and K. A. Meissner, “Asymptotic safety and Conformal Standard Model,” Phys. Rev. D, vol. 99, no. 11, p. 115029, 2019\. * [51] J. H. Kwapisz, “Asymptotic safety, the Higgs boson mass, and beyond the standard model physics,” Phys. Rev. D, vol. 100, no. 11, p. 115001, 2019\. * [52] A. Eichhorn and A. Held, “Top mass from asymptotic safety,” Phys. Lett. B, vol. 777, pp. 217–221, 2018. * [53] A. Eichhorn and A. Held, “Mass difference for charged quarks from asymptotically safe quantum gravity,” Phys. Rev. Lett., vol. 121, no. 15, p. 151302, 2018. * [54] A. Eichhorn, A. Held, and C. Wetterich, “Quantum-gravity predictions for the fine-structure constant,” Phys. Lett. B, vol. 782, pp. 198–201, 2018\. * [55] A. Eichhorn and S. Lippoldt, “Quantum gravity and Standard-Model-like fermions,” Phys. Lett. B, vol. 767, pp. 142–146, 2017. * [56] N. Christiansen, D. F. Litim, J. M. Pawlowski, and M. Reichert, “Asymptotic safety of gravity with matter,” Phys. Rev. D, vol. 97, no. 10, p. 106012, 2018. * [57] A. Eichhorn, S. Lippoldt, and M. Schiffer, “Zooming in on fermions and quantum gravity,” Phys. Rev. D, vol. 99, no. 8, p. 086002, 2019. * [58] N. Christiansen and A. Eichhorn, “An asymptotically safe solution to the U(1) triviality problem,” Phys. Lett. B, vol. 770, pp. 154–160, 2017. * [59] A. Eichhorn, A. Held, and C. Wetterich, “Predictive power of grand unification from quantum gravity,” JHEP, vol. 08, p. 111, 2020. * [60] A. Eichhorn and M. Schiffer, “$d=4$ as the critical dimensionality of asymptotically safe interactions,” Phys. Lett. B, vol. 793, pp. 383–389, 2019. * [61] R. Alkofer, A. Eichhorn, A. Held, C. M. Nieto, R. Percacci, and M. Schröfl, “Quark masses and mixings in minimally parameterized UV completions of the Standard Model,” Annals Phys., vol. 421, p. 168282, 2020. * [62] J. Daas, W. Oosters, F. Saueressig, and J. Wang, “Asymptotically safe gravity with fermions,” Phys. Lett. B, vol. 809, p. 135775, 2020. * [63] A. Held, “Effective asymptotic safety and its predictive power: Gauge-Yukawa theories,” Front. in Phys., vol. 8, p. 341, 2020. * [64] Y. Hamada, K. Tsumura, and M. Yamada, “Scalegenesis and fermionic dark matters in the flatland scenario,” Eur. Phys. J. C, vol. 80, no. 5, p. 368, 2020. * [65] M. Reichert and J. Smirnov, “Dark Matter meets Quantum Gravity,” Phys. Rev. D, vol. 101, no. 6, p. 063015, 2020. * [66] A. Eichhorn and M. Pauly, “Safety in darkness: Higgs portal to simple Yukawa systems,” 5 2020. * [67] A. Eichhorn and M. Pauly, “Constraining power of asymptotic safety for scalar fields,” 9 2020. * [68] Y. Hamada, J. M. Pawlowski, and M. Yamada, “Gravitational instantons and anomalous chiral symmetry breaking,” 9 2020. * [69] G. P. de Brito, A. Eichhorn, and M. Schiffer, “Light charged fermions in quantum gravity,” 10 2020. * [70] D. F. Litim and F. Sannino, “Asymptotic safety guaranteed,” Journal of High Energy Physics, vol. 2014, Dec 2014. * [71] D. F. Litim, M. Mojaza, and F. Sannino, “Vacuum stability of asymptotically safe gauge-Yukawa theories,” JHEP, vol. 01, p. 081, 2016. * [72] R. Mann, J. Meffe, F. Sannino, T. Steele, Z.-W. Wang, and C. Zhang, “Asymptotically Safe Standard Model via Vectorlike Fermions,” Phys. Rev. Lett., vol. 119, no. 26, p. 261802, 2017. * [73] O. Antipin, N. A. Dondi, F. Sannino, A. E. Thomsen, and Z.-W. Wang, “Gauge-Yukawa theories: Beta functions at large $N_{f}$,” Phys. Rev. D, vol. 98, no. 1, p. 016003, 2018. * [74] E. Molinaro, F. Sannino, and Z. W. Wang, “Asymptotically safe Pati-Salam theory,” Phys. Rev. D, vol. 98, no. 11, p. 115007, 2018. * [75] Z.-W. Wang, A. Al Balushi, R. Mann, and H.-M. Jiang, “Safe Trinification,” Phys. Rev. D, vol. 99, no. 11, p. 115017, 2019. * [76] A. H. Guth, “The Inflationary Universe: A Possible Solution to the Horizon and Flatness Problems,” Adv. Ser. Astrophys. Cosmol., vol. 3, pp. 139–148, 1987. * [77] A. Ijjas, P. J. Steinhardt, and A. Loeb, “Inflationary schism,” Phys. Lett. B, vol. 736, pp. 142–146, 2014. * [78] J.-L. Lehners and K. Stelle, “A Safe Beginning for the Universe?,” Phys. Rev. D, vol. 100, no. 8, p. 083540, 2019. * [79] J. H. Kwapisz, “Conformal standard model and inflation,” 11 2019. * [80] A. D. Linde, “Hard art of the universe creation (stochastic approach to tunneling and baby universe formation),” Nucl. Phys. B, vol. 372, pp. 421–442, 1992. * [81] C. Kiefer, D. Polarski, and A. A. Starobinsky, “Quantum to classical transition for fluctuations in the early universe,” Int. J. Mod. Phys. D, vol. 7, pp. 455–462, 1998. * [82] C. Kiefer and D. Polarski, “Why do cosmological perturbations look classical to us?,” Adv. Sci. Lett., vol. 2, pp. 164–173, 2009. * [83] A. Linde, “Stochastic approach to tunneling and baby universe formation,” Nuclear Physics B, vol. 372, p. 421–442, Mar 1992. * [84] V. Vennin and A. A. Starobinsky, “Correlation functions in stochastic inflation,” The European Physical Journal C, vol. 75, Sep 2015. * [85] M. Noorbala, V. Vennin, H. Assadullahi, H. Firouzjahi, and D. Wands, “Tunneling in stochastic inflation,” Journal of Cosmology and Astroparticle Physics, vol. 2018, p. 032–032, Sep 2018. * [86] J. J. M. Carrasco, R. Kallosh, and A. Linde, “Cosmological attractors and initial conditions for inflation,” Physical Review D, vol. 92, Sep 2015\. * [87] T. Krajewski, K. Turzyński, and M. Wieczorek, “On preheating in $\alpha$-attractor models of inflation,” The European Physical Journal C, vol. 79, Aug 2019. * [88] Planck Collaboration, “Planck 2018 results - x. constraints on inflation,” A&A, vol. 641, p. A10, 2020. * [89] A. A. Starobinsky, “A new type of isotropic cosmological models without singularity,” Physics Letters B, vol. 91, pp. 99–102, Mar. 1980. * [90] J. E. Lidsey, A. R. Liddle, E. W. Kolb, E. J. Copeland, T. Barreiro, and M. Abney, “Reconstructing the inflaton potential—an overview,” Reviews of Modern Physics, vol. 69, p. 373–410, Apr 1997. * [91] A. Bonanno and A. Platania, “Asymptotically safe inflation from quadratic gravity,” Phys. Lett. B, vol. 750, pp. 638–642, 2015. * [92] A. Bonanno and F. Saueressig, “Asymptotically safe cosmology – A status report,” Comptes Rendus Physique, vol. 18, pp. 254–264, 2017. * [93] A. Platania, “From renormalization group flows to cosmology,” Front. in Phys., vol. 8, p. 188, 2020. * [94] E. Fradkin and A. A. Tseytlin, “Renormalizable Asymptotically Free Quantum Theory of Gravity,” Phys. Lett. B, vol. 104, pp. 377–381, 1981. * [95] N. G. Nielsen, F. Sannino, and O. Svendsen, “Inflation from Asymptotically Safe Theories,” Phys. Rev. D, vol. 91, p. 103521, 2015. * [96] M. Demmel, F. Saueressig, and O. Zanusso, “A proper fixed functional for four-dimensional Quantum Einstein Gravity,” JHEP, vol. 08, p. 113, 2015\. * [97] D. I. Kaiser, “Conformal Transformations with Multiple Scalar Fields,” Phys. Rev. D, vol. 81, p. 084044, 2010. * [98] A. Bonanno and M. Reuter, “Quantum gravity effects near the null black hole singularity,” Phys. Rev. D, vol. 60, p. 084011, 1999. * [99] A. Bonanno and M. Reuter, “Spacetime structure of an evaporating black hole in quantum gravity,” Phys. Rev. D, vol. 73, p. 083005, 2006. * [100] C. Pagani and M. Reuter, “Finite Entanglement Entropy in Asymptotically Safe Quantum Gravity,” JHEP, vol. 07, p. 039, 2018. * [101] A. Bonanno and M. Reuter, “Proper time flow equation for gravity,” JHEP, vol. 02, p. 035, 2005. * [102] O. Svendsen, H. Bazrafshan Moghaddam, and R. Brandenberger, “Preheating in an Asymptotically Safe Quantum Field Theory,” Phys. Rev. D, vol. 94, no. 8, p. 083527, 2016.
# Regressive Side Effects of Training Language Models to Mimic Student Misconceptions Shashank Sonkar Rice University Houston, TX <EMAIL_ADDRESS> &Naiming Liu Rice University Houston, TX <EMAIL_ADDRESS> &Richard G. Baraniuk Rice University Houston, TX <EMAIL_ADDRESS> ###### Abstract This paper presents a novel exploration into the regressive side effects of training Large Language Models (LLMs) to mimic student misconceptions for personalized education. We highlight the problem that as LLMs are trained to more accurately mimic student misconceptions, there is a compromise in the factual integrity and reasoning ability of the models. Our work involved training an LLM on a student-tutor dialogue dataset to predict student responses. The results demonstrated a decrease in the model’s performance across multiple benchmark datasets, including the ARC reasoning challenge and TruthfulQA, which evaluates the truthfulness of model’s generated responses. Furthermore, the HaluEval Dial dataset, used for hallucination detection, and MemoTrap, a memory-based task dataset, also reported a decline in the model accuracy. To combat these side effects, we introduced a “hallucination token” technique. This token, appended at the beginning of each student response during training, instructs the model to switch between mimicking student misconceptions and providing factually accurate responses. Despite the significant improvement across all datasets, the technique does not completely restore the LLM’s baseline performance, indicating the need for further research in this area. This paper contributes to the ongoing discussion on the use of LLMs for student modeling, emphasizing the need for a balance between personalized education and factual accuracy. ## 1 Introduction Personalized education, an approach that caters to the unique learning needs of individuals, is increasingly becoming a key aspiration in educational technology [1, 2, 3]. With the advent of advanced Large Language Models (LLMs), this aspiration is inching closer to reality. LLMs, such as Llama [4] and GPT [5, 6] models, are playing a pivotal role in this domain, demonstrating significant potentials in various applications, including the simulation of student behavior [7] and learning patterns [8]. However, the road to leveraging LLMs for personalized education is challenging [9, 10, 11]. In this paper, we have identified regressive side effects arising from training LLMs to mimic student behavior. We find that training LLMs to replicate student misconceptions accurately is a double-edged sword. On one hand, it creates a model that can mimic student behavior, making it a potentially effective tool for personalized learning. On the other hand, it significantly compromises the model’s factual integrity and reasoning ability. These regressive side effects are a critical issue, as the primary role of any educational model is to provide accurate and reliable information. Figure 1: This figure illustrates a typical student-tutor interaction from the CLASS [12] dataset. Unlike the CLASS methodology, which focuses on training a tutor model, our study trains a ‘student model’, with the LLM predicting student responses. This approach is motivated by the potential of personalized education, where understanding and mimicking student behavior can lead to more effective learning interventions. However, this approach, while effectively mimicking student misconceptions, leads to regressive side effects such as compromising the model’s factual integrity and diminishing its reasoning abilities. The conversation shown here exemplifies this issue, where the student’s response, while partially correct, contains misconceptions. To mitigate these side effects, we introduce hallucination tokens ([hal] and [/hal]) appended to student responses during training. These tokens instruct the model to switch between mimicking student misconceptions and providing factually accurate responses. Despite significant improvements, the technique does not fully restore the model’s baseline performance, highlighting the complexity of the issue and the need for further research. To investigate this issue further, we conducted a comprehensive exploration involving training LLMs on a student-tutor dialogue dataset. This dataset, derived from the CLASS framework [12, 13, 2], comprises dialogues on biology questions sourced from college-level textbooks. An example of the student- tutor interaction from the dataset is illustrated in figure 1. It provides a realistic representation of student learning patterns, featuring student misconceptions and the tutor’s rectifications. We used the dataset to train the latest Vicuna models (7B and 13B) [14], state-of-the-art Llama [4] variants, to mimic student responses. However, the training process significantly decreased the model’s performance across various benchmark datasets, including the ARC reasoning challenge [15], TruthfulQA [16], Hallucination Evaluation Dialogue [17], and MemoTrap [18]. We present a detailed analysis across nine key benchmarks using the Eleuther AI Language Model Evaluation Harness [19], a widely used framework [20] which provides a thorough and fair assessment of generative models across a spectrum of reasoning and general knowledge tasks. We present a detailed analysis across nine key benchmarks using the Eleuther AI Language Model Evaluation Harness [19], a widely used framework [20] to test generative language models on a large number of different evaluation tasks. To further understand the regressive side effects, we conducted a control experiment to compare the model trained to predict tutor responses versus one trained to predict student responses. The results showed that training the LLM on tutor responses did not lead to any performance decline observed when mimicking student responses. This trends highlight that the regressive side effects are a unique challenge specific to training LLMs to replicate student misconceptions. To counteract the side effects, we propose to incorporate novel start and end hallucination tokens ([hal] and [/hal]) into the LLM training process. These tokens, placed at the beginning and end of each student response, serve as cues to the model, instructing it when to differentiate between providing accurate responses and replicating student misconceptions. Our results indicate a substantial improvement in the model’s performance across all datasets after introducing this token. However, these tokens do not fully restore the model’s baseline performance, underscoring the complexity of the issue. Through our research, we have brought to the following critical contributions in the realm of personalized education leveraging LLMs: 1\. We have uncovered and thoroughly investigated regressive side effects in the LLMs trained for student modeling. This research highlights the paradoxical challenge when LLMs are trained to mimic student misconceptions, potentially compromising their factual integrity and reasoning ability. 2\. We have proposed hallucination tokens to mitigate these regressive effects. These tokens, added to the training process, instruct the LLMs to differentiate between mimicking student misconceptions and providing factually accurate responses, substantially improving the model’s performance. 3\. Despite the improvements achieved with the hallucination tokens, our research indicates that it does not fully counteract the regressive side effects. This points to the complexity of this issue and underscores the need for further research in this area. Our research marks a significant step towards understanding the complexities of using LLMs for student modeling. The findings and contributions of this study will fuel further exploration and innovation in this domain, ultimately refining the use of LLMs in personalized learning environments. ## 2 Related Work The intersection of artificial intelligence and education has been an area of active research, with a focus on developing systems that can adapt to and support individual learners. Our work touches upon several research domains, including student modeling, the design of intelligent tutoring systems, and the deployment of Large Language Models (LLMs) in educational contexts. ### 2.1 Student Modeling Student modeling has long been the cornerstone of personalized learning, with early attempts using rule-based and Bayesian systems to predict student knowledge and behaviors [21]. Recent advancements have shifted towards utilizing machine learning to create more sophisticated models that can adapt to student learning patterns over time [22, 7]. Our work builds upon these foundations by exploring how LLMs can simulate not only the knowledge but also the typical errors and misconceptions students have during the learning process. ### 2.2 Intelligent Tutoring Systems (ITS) Intelligent tutoring systems have been designed to provide immediate and personalized instruction or feedback to learners without human intervention [23]. The application of LLMs in ITS presents a novel opportunity to create systems that can engage in more natural and meaningful dialogues with students [3, 12]. Our approach diverges from traditional ITS by focusing on the intentional generation of errors to mimic a student’s learning trajectory, rather than solely providing expert-level instructions [24]. ### 2.3 Large Language Models in Education The use of LLMs like GPT [5] in education is a relatively new but rapidly growing field of study [25]. These models have been employed for various educational purposes, from generating educational content to serving as conversational agents [26, 12]. However, the challenge of ensuring the truthfulness and reliability of the information provided by LLMs is a recurring concern [27]. Our research contributes to this dialogue by investigating the impact of training LLMs to produce student-like errors and proposing a novel ‘hallucination token’ to manage this trade-off. ### 2.4 Truthfulness and Reliability in AI The TruthfulQA benchmark has been instrumental in highlighting the issues of truthfulness in AI-generated content [28]. The ARC challenge further emphasizes the complexity of reasoning required from AI systems beyond simple fact retrieval [29]. Our work is aligned with these challenges, as we seek to understand and improve the truthfulness and reasoning capacity of LLMs when they are trained to replicate student behaviors. In conclusion, our study intersects with and contributes to the existing body of work in these areas by addressing the unique challenge of training LLMs to authentically mimic student learning processes, including the generation of errors. Our introduction of the “hallucination token” represents a step forward in this domain, suggesting a new direction for future research and development. ## 3 Methodology Our methodology is divided into three main parts: data preparation, model training, and the incorporation of hallucination tokens. ### 3.1 Data Preparation The first step in our methodology involves preparing the dataset for training the LLMs. We denote the conversation dataset as $\mathcal{D}$, which consists of ordered pairs of tutor-student conversational turns: $\mathcal{D}=\\{(\mathbf{x}_{1},\mathbf{y}_{1}),(\mathbf{x}_{2},\mathbf{y}_{2}),\ldots,(\mathbf{x}_{N},\mathbf{y}_{N})\\}$, where $N$ is the total number of conversational turns. Each $\mathbf{x}$ represents a sequence of tutor utterances, and each corresponding $\mathbf{y}$ represents the student response. The dataset is derived from the CLASS framework [12], which provides a realistic representation of student learning patterns, featuring student misconceptions and the tutor’s rectifications. This dataset provides a rich source of student-tutor dialogues on biology questions sourced from college textbooks. ### 3.2 Model Training The second step in our methodology involves training LLMs. The LLMs are designed to predict the next utterance given the previous conversational context. Unlike traditional approaches that focus on the correct responses typically output by a tutoring system, our model centers on student outputs, which may possess a mix of correctness and misconceptions. For an input sequence $\mathbf{x}_{i}$, the LLM aims to generate an output sequence $\hat{\mathbf{y}}_{i}$ that resembles a student’s response. The language modeling loss for a single data pair is defined by the negative log likelihood: $\mathcal{L}(\mathbf{y}_{i},\hat{\mathbf{y}}_{i})=-\sum_{t=1}^{|\mathbf{y}_{i}|}\log p\left(y_{i,t}\middle|\mathbf{x}_{i},\mathbf{y}_{i,<t};\theta\right)$ where $\mathbf{y}_{i,<t}$ indicates the tokens in the true response preceding the current token $y_{i,t}$, and $\theta$ encapsulates the parameters of the LLM. The overall training loss is the sum over the entire dataset: $\mathcal{L}_{\text{total}}=\sum_{i=1}^{N}\mathcal{L}\left(\mathbf{y}_{i},\hat{\mathbf{y}}_{i}\right)$ ### 3.3 Incorporation of Hallucination Tokens The third step in our methodology involves the incorporation of hallucination tokens. To enhance the LLM’s ability to generate responses that simulate student behaviors, including providing incorrect or uncertain information, we introduce hallucination token markers. Each student response in the dataset is enriched with these markers to indicate the beginning and the end of the potentially inaccurate content. Let $\mathbf{y}_{i}$ be an original student response sequence from the dataset. The augmented student response $\tilde{\mathbf{y}}_{i}$ used for training is constructed by prepending and appending hallucination tokens [hal] and [/hal], respectively: $\tilde{\mathbf{y}}_{i}=\left[\texttt{[hal]},\mathbf{y}_{i,1},\mathbf{y}_{i,2},\ldots,\mathbf{y}_{i,|\mathbf{y}_{i}|},\texttt{[/hal]}\right]$ In the modified training regime, the LLM predicts the sequence $\hat{\mathbf{y}}_{i}$ such that it learns to include these tokens, effectively grasping the context of student uncertainty or errors. These tokens serve as cues to the model, instructing it when to differentiate between providing accurate responses and replicating student misconceptions. ## 4 Experiments and Discussion Table 1: Performance of Vicuna models on TruthfulQA tasks. The table compares the performance of the original vicuna model, the control model trained to mimic tutor responses in biology (tutor) the model trained to mimic tutor responses in biology (tutor), and the model trained with hallucination tokens (student-hal). The results are presented for three different settings: MC1, MC2, and Generation. MC1 refers to a setting where there is only one correct answer to a question, while MC2 refers to a setting where there are multiple correct answers. For these settings, the performance is measured in terms of accuracy. The generation setting involves the model generating 1-2 sentence answers, with performance evaluated using BLEU and ROUGE scores. The results highlight the significant drop in performance when the model is trained to mimic student responses, demonstrating a regressive side effect in terms of truthfulness. However, the substantial recovery in performance with the introduction of hallucination tokens suggests a promising strategy to mitigate these regressive effects. Dataset | | TQA MC1 --- (Single-true) | TQA MC2 --- (Multi-true) TruthfulQA (TQA) Generation Metric | Accuracy | Accuracy | BLEU | | ROUGE --- (unigram) | ROUGE --- (bigram) | ROUGE --- (LCS) vicuna-7b-v1.5 | 32.93 | 50.37 | 49.69 | 51.41 | 45.90 | 50.55 tutor-7b | 34.64 | 52.43 | 42.72 | 47.12 | 37.94 | 45.29 student-7b | 23.75 | 36.14 | 24.60 | 29.74 | 14.32 | 28.89 student-hal-7b | 29.25 | 44.68 | 43.94 | 47.61 | 36.47 | 45.53 vicuna-13b-v1.5 | 35.01 | 50.87 | 47.12 | 50.18 | 44.92 | 49.08 tutor-13b | 34.76 | 52.20 | 42.84 | 48.71 | 38.80 | 46.76 student-13b | 22.15 | 33.93 | 15.18 | 18.12 | 6.12 | 17.75 student-hal-13b | 27.91 | 41.46 | 39.29 | 42.35 | 33.66 | 42.96 In this section, we present our experimental methodology and discuss the findings in detail. The experiments were designed to explore the regressive side effects of training LLMs to mimic student behavior and to assess the effectiveness of our proposed hallucination tokens in mitigating these effects. ### 4.1 Experimental Setup We trained the Vicuna 7B and 13B models [14], one of the best open-source LLMs, on a student-tutor dialogue dataset derived from the CLASS [12] framework. This dataset, which provides a realistic representation of student learning patterns, misconceptions, and the tutor’s rectifications, was used to fine-tune the models to generate outputs that mimic student dialogue. The dataset contains $648$ conversations, which sums up to a total of 20K student- tutor interactions. Average length of conversations is around $~{}400$ words, only including the student and tutor fields in the conversation template. The models were evaluated across seven key benchmarks using the Eleuther AI Language Model Evaluation Harness [19]. These benchmarks include the TruthfulQA [16], ARC [15], HellaSwag [30], Winogrande [31], MMLU [32], HaluEval Dialogue [17], and MemoTrap [18]. Each of these benchmarks tests different aspects of the model’s performance, including its truthfulness, reasoning abilities, ability to recognize hallucinations, and memory-based task performance. ### 4.2 In-depth Analysis: TruthfulQA In the realm of educational technology, the veracity of information provided by a model is of paramount importance. Misinformation or misconceptions can lead to significant learning detriments, making the truthfulness of a model’s responses a critical factor in its effectiveness as an educational tool. Therefore, we chose to conduct an in-depth analysis of our models’ performance on the TruthfulQA benchmark. TruthfulQA is a benchmark specifically designed to measure the truthfulness of a language model’s responses across a wide range of categories. It tests the model’s ability to avoid generating false answers learned from imitating human texts, a challenge that is particularly relevant to our study. Given the importance of truthfulness in educational contexts and the unique challenges posed by training models to mimic student misconceptions, we believe that a rigorous analysis of our models’ performance on TruthfulQA is warranted. In this section, we present our findings from the TruthfulQA benchmark, exploring the impact of training models to mimic student behavior and the effectiveness of our proposed hallucination tokens in mitigating any negative effects. We delve into the results from the multiple-choice and generation tasks within TruthfulQA, providing a comprehensive view of our models’ truthfulness in different contexts. TruthfulQA Multiple-Choice Setting 1 (MC1) Findings. In the first multiple- choice setting, where there is a single correct label, the student-7b model’s accuracy decreased by 15 points compared to the vicuna-7b model. However, the introduction of hallucination tokens led to a significant recovery in performance. This finding is particularly relevant in the context of education, where maintaining the truthfulness of responses is crucial. The improvement with hallucination tokens suggests that it is possible to train models that can both simulate student behavior and adhere to factual accuracy, a key consideration for deploying LLMs in educational settings. TruthfulQA Multiple-Choice Setting 2 (MC2) Findings. In the second multiple- choice setting, where multiple correct labels are possible, we observed a similar trend to the MC1 setting. The student-7b model experienced a significant drop in accuracy, from 50.37% in the vicuna-7b model to 36.14% when trained to mimic student responses. However, the introduction of hallucination tokens led to a notable improvement in performance, with the student-7b model’s accuracy recovering to 44.68%. This recovery is particularly relevant in the context of education, where multiple perspectives or answers might be correct. The ability of the model to navigate such complexities while maintaining truthfulness is crucial. The improvement with hallucination tokens suggests that it is possible to train models that can both simulate student behavior and adhere to factual accuracy, a key consideration for deploying LLMs in educational settings. TruthfulQA Generation Findings. For the TruthfulQA generation task, where the model is tasked with generating 1-2 sentence answers, we employed ROUGE scores to evaluate performance due to the generative nature of the task. The student-7b model saw a significant decrease in ROUGE scores, from 51.41 in the vicuna-7b model to 29.74, indicating a substantial loss in the ability to generate truthful, relevant responses. However, the introduction of hallucination tokens led to a significant recovery in performance, with ROUGE scores improving to 47.61. This finding is crucial for educational technology as LLMs are increasingly used as generative agents to create educational content, provide explanations, and engage in dialogue with students. The ability to generate truthful, accurate responses is fundamental to their utility in these contexts. The recovery observed with hallucination tokens highlights their potential to enable LLMs to simulate student misconceptions for personalized learning without sacrificing the quality and truthfulness of their output. Table 2: Comparative performance of Large Language Models (LLMs) on various benchmarks before and after the introduction of hallucination tokens, with a control experiment involving tutor models. The table presents the performance of Vicuna 7B models across five key benchmarks: ARC Reasoning, Hallucination Evaluation Dialogue (HaluDial), Hallucination Memorization Trap (MemoTrap), TruthfulQA (TQA), HellaSwag (HSwag), MMLU, and Winogrande (WinoG). The numbers in parentheses (e.g., 25-S in ARC) represent the number of few-shot examples provided to the model during evaluation. The performance is measured in terms of accuracy percentage. The table compares the performance of the original vicuna models, tutor models, student models, and student models trained with hallucination tokens (student-hal). The results highlight the significant drop in performance when the model is trained to mimic student responses, demonstrating regressive side effects across multiple tasks. However, the introduction of hallucination tokens leads to a substantial recovery in performance across all benchmarks, underscoring their potential in mitigating these regressive effects. Model | Avg | | ARC --- (25-S) | HaluDial --- (0-S) | MemoTrap --- (0-S) | TQA --- (6-S) | HSwag --- (10-S) | MMLU --- (5-S) | WinoG --- (5-S) vicuna-7b-v1.5 | 60.8 | 53.24 | 69.08 | 68.48 | 50.34 | 77.39 | 51.04 | 72.14 tutor-7b | 61.0 | 52.13 | 68.81 | 69.23 | 52.3 | 78.07 | 51.32 | 71.19 student-7b | 55.4 | 40.61 | 65.39 | 65.28 | 36.87 | 76.72 | 50.77 | 71.9 student-hal-7b | 58.0 | 45.48 | 70.73 | 66.88 | 44.83 | 77.21 | 51.54 | 72.03 vicuna-13b-v1.5 | 64.2 | 57.08 | 73.78 | 67.2 | 51.51 | 81.24 | 56.67 | 74.66 tutor-13b | 64.7 | 57.34 | 73.92 | 66.13 | 52.99 | 81.51 | 57.02 | 74.35 student-13b | 58.2 | 46.5 | 66.97 | 65.81 | 35.0 | 80.36 | 57.06 | 72.22 student-hal-13b | 60.3 | 48.63 | 72.98 | 66.13 | 42.75 | 80.28 | 56.4 | 73.16 ### 4.3 Benchmark Evaluation Following the exploration of TruthfulQA settings, we delve into the performance of our models across a broader range of benchmarks as detailed in Table 2. These benchmarks—ARC, HaluEval Dial, MemoTrap, MMLU, HellaSwag, and Winogrande—offer a comprehensive view of the models’ capabilities in reasoning, detecting hallucinations, avoiding memorization traps, and understanding commonsense, respectively. AI2 Reasoning Challenge (ARC) Findings. ARC serves as a rigorous benchmark to evaluate a model’s reasoning capabilities through a set of grade-school science questions. These questions are designed to test not just the factual knowledge of the models but also their ability to apply this knowledge in reasoning through complex, multi-step problems. The ARC dataset is particularly relevant in educational contexts as it mirrors the type of critical thinking and problem-solving skills students are expected to develop. In our experiments, the performance of models trained to mimic student responses on the ARC benchmark experienced a notable decline. Specifically, the vicuna-7b model saw its accuracy decrease from 53.24% to 40.61% when trained on student dialogues. This significant drop in performance highlights a critical concern: training LLMs to replicate student behavior, including misconceptions, can severely impair their reasoning abilities. However, our introduction of hallucination tokens into the training process presents a silver lining. Our approach led to a partial recovery in the ARC performance, with accuracy improving to 45.48%. While this does not fully restore the model’s baseline performance, it represents a significant step towards mitigating the regressive side effects of training LLMs on student data. Hallucination Evaluation (HaluEval) Dialogue Findings. The HaluEval Dial benchmark is designed to assess a model’s ability to recognize and avoid hallucinations in generated responses, particularly in the context of knowledge grounded dialogue tasks. Hallucinations in this context refer to the model generating information that is not supported by the input data or general knowledge, a critical issue when models are used in educational settings where accuracy is paramount. Our findings indicate that training models to mimic student responses led to a decrease in performance on the HaluEval Dial benchmark. Specifically, the vicuna-7b model saw its accuracy drop from 69.0% to 65.39%. However, the introduction of hallucination tokens demonstrated a remarkable ability to counteract this effect, with the student-7b model’s accuracy improving to 70.73%. Memorization Traps (MemoTrap) Findings. MemoTrap is a benchmark designed to test whether language models can avoid memorization traps by prompting them to complete well-known proverbs with endings that deviate from the commonly used ones. This benchmark is particularly relevant for evaluating a model’s ability to generate creative and contextually appropriate responses rather than relying on rote memorization. In our experiments, training models to mimic student responses resulted in a decrease in performance on the MemoTrap benchmark. The vicuna-7b model’s accuracy decreased from 68.48% to 65.28%, indicating that training on student dialogues might encourage the model to rely more on memorization rather than understanding and applying knowledge flexibly. The introduction of hallucination tokens led to a slight improvement, with accuracy increasing to 66.88%. MMLU, HellaSwag, and Winogrande Findings. The performance of models on the MMLU, HellaSwag, and Winogrande benchmarks remained relatively stable, regardless of whether they were trained to mimic tutor or student responses. The nuanced impact observed in other benchmarks underscores the importance of carefully considering the training data and methodologies used when developing LLMs for educational purposes. The introduction of hallucination tokens emerges as a promising strategy for mitigating some of the regressive side effects associated with training models to mimic student behavior, ensuring that they can still serve as effective tools for personalized learning without compromising on factual accuracy or reasoning capabilities. ### 4.4 Control Models: Tutor Models To further understand the regressive side effects of training LLMs to mimic student behavior, we conducted a control experiment by training models to predict tutor responses. This experiment aimed to compare the performance of models trained to predict tutor responses versus those trained to predict student responses. The tutor models were trained using the same student-tutor dialogue dataset derived from the CLASS framework [12]. However, instead of training the models to mimic student responses, we trained them to predict the responses of the tutor. Our findings, as shown in Table 2, revealed that training the LLMs on tutor responses did not lead to the same performance decline observed when mimicking student responses. This result underscores that the regressive side effects are a unique challenge specific to training LLMs to replicate student misconceptions. ## 5 Conclusion In this study, we have delved into the challenges of training LLMs to mimic student behavior, with a particular focus on the regressive side effects that emerge in this process. Our findings reveal a complex paradox: as LLMs become more adept at replicating student misconceptions, they tend to compromise on their factual integrity and reasoning ability. Our experiments demonstrated a notable decrease in the model’s performance across various key benchmark datasets like ARC Reasoning Challenge and TruthfulQA. To mitigate these regressive side effects, we introduced a novel technique involving the use of hallucination tokens during the training process. Our results indicate that the introduction of these tokens leads to a substantial improvement in the model’s performance across all datasets. However, it’s important to note that despite the significant improvements achieved with the hallucination tokens, they do not fully restore the model’s baseline performance. This outcome underscores the complexity of the problem and highlights the need for a more nuanced approach when training LLMs to mimic student behavior. While we have made some strides in addressing the regressive side effects, our work is just the beginning. We believe that our findings will pave the way for further research in this domain, ultimately contributing to the refinement of LLMs in personalized learning environments. ## 6 Ethics The ethical implications of training models on incorrect data are profound and demand conscientious exploration in future work. As our models find application in real-world educational settings, the delineation between the effective simulation of student behaviors and the propagation of misinformation will need to be continually assessed and refined. Our research has thus laid the groundwork for a new pedagogical paradigm, where AI becomes a symbiotic partner in the complex choreography of learning and teaching. ## Acknowledgments This work was supported by NSF grants 1842378, ONR grant N0014-20-1-2534, AFOSR grant FA9550-22-1-0060, and a Vannevar Bush Faculty Fellowship, ONR grant N00014-18-1-2047. ## References * [1] Sankalan Pal Chowdhury, Vilém Zouhar, and Mrinmaya Sachan. Scaling the authoring of autotutors with large language models. arXiv preprint arXiv:2402.09216, 2024. * [2] Shashank Sonkar, Kangqi Ni, Sapana Chaudhary, and Richard G Baraniuk. Pedagogical alignment of large language models. arXiv preprint arXiv:2402.05000, 2024. * [3] Robin Schmucker, Meng Xia, Amos Azaria, and Tom Mitchell. Ruffle&riley: Towards the automated induction of conversational tutoring systems. arXiv preprint arXiv:2310.01420, 2023. * [4] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. * [5] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of Artificial General Intelligence: Early experiments with GPT-4. arXiv preprint arXiv:2303.12712, 2023. * [6] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022. * [7] Naiming Liu, Zichao Wang, Richard Baraniuk, and Andrew Lan. Open-ended knowledge tracing for computer science education. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3849–3862, 2022. * [8] Enkelejda Kasneci, Kathrin Seßler, Stefan Küchemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke Hüllermeier, et al. Chatgpt for good? on opportunities and challenges of large language models for education. Learning and Individual Differences, 103:102274, 2023. * [9] Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D Hwang, et al. Faith and fate: Limits of transformers on compositionality. arXiv preprint arXiv:2305.18654, 2023. * [10] Shashank Sonkar and Richard G. Baraniuk. Deduction under perturbed evidence: Probing student simulation (knowledge tracing) capabilities of large language models. CEUR Workshop Proceedings, 3487:26–33, 2023. * [11] Allison Wang, Ethan Prihar, and Neil Heffernan. Assessing the ality of large language models in generating mathematics explanations. Learning@ Scale (L@ S’23), July 20–22, 2023, Copenhagen, Denmark., 2023. * [12] Shashank Sonkar, Naiming Liu, Debshila Mallick, and Richard Baraniuk. CLASS: A Design Framework for Building Intelligent Tutoring Systems Based on Learning Science principles. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1941–1961, Singapore, December 2023. * [13] Shashank Sonkar, Xinghe Chen, Myco Le, Naiming Liu, Debshila Basu Mallick, and Richard Baraniuk. Code soliloquies for accurate calculations in large language models. In Proceedings of the 14th Learning Analytics and Knowledge Conference, pages 828–835, 2024. * [14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023. * [15] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge, 2018. * [16] Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods, 2022. * [17] Junyi Li, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. HaluEval: A large-scale hallucination evaluation benchmark for large language models. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6449–6464, Singapore, December 2023. Association for Computational Linguistics. * [18] Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. The inverse scaling prize, 2022. * [19] Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation, 12 2023. * [20] Edward Beeching, Clémentine Fourrier, Nathan Habib, Sheon Han, Nathan Lambert, Nazneen Rajani, Omar Sanseviero, Lewis Tunstall, and Thomas Wolf. Open llm leaderboard, 2023. * [21] Martha C Polson and J Jeffrey Richardson. Foundations of intelligent tutoring systems. Psychology Press, 2013. * [22] Ryan SJD Baker, Kalina Yacef, et al. The state of educational data mining in 2009: A review and future visions. Journal of educational data mining, 1(1):3–17, 2009. * [23] Beverly Park Woolf. Building intelligent interactive tutors: Student-centered strategies for revolutionizing e-learning. Morgan Kaufmann, 2010. * [24] Kurt VanLehn. The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational psychologist, 46(4):197–221, 2011. * [25] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020. * [26] Neil T Heffernan and Cristina Lindquist Heffernan. The assistments ecosystem: Building a platform that brings scientists and teachers together for minimally invasive research on human learning and teaching. International Journal of Artificial Intelligence in Education, 24:470–497, 2014. * [27] Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958, 2021. * [28] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. * [29] Oren Etzioni, Anthony Fader, Janara Christensen, Stephen Soderland, and Mausam Mausam. Open information extraction: The second generation. In IJCAI, volume 11, pages 3–10, 2011. * [30] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence?, 2019. * [31] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WINOGRANDE: an adversarial winograd schema challenge at scale, 2019. * [32] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. CoRR, abs/2009.03300, 2020.
OPES, MetaD, CVs, FES # Exploration vs Convergence Speed in Adaptive-bias Enhanced Sampling Michele Invernizzi Freie Universität Berlin, 14195 Berlin, Germany <EMAIL_ADDRESS>Michele Parrinello Italian Institute of Technology, 16163 Genova, Italy ###### Abstract In adaptive-bias enhanced sampling methods, a bias potential is added to the system to drive transitions between metastable states. The bias potential is a function of a few collective variables and is gradually modified according to the underlying free energy surface. We show that when the collective variables are suboptimal, there is an exploration-convergence tradeoff, and one must choose between a quickly converging bias that will lead to fewer transitions, or a slower to converge bias that can explore the phase space more efficiently but might require a much longer time to produce an accurate free energy estimate. The recently proposed On-the-fly Probability Enhanced Sampling (OPES) method focuses on fast convergence, but there are cases where fast exploration is preferred instead. For this reason, we introduce a new variant of the OPES method that focuses on quickly escaping metastable states, at the expense of convergence speed. We illustrate the benefits of this approach on prototypical systems and show that it outperforms the popular metadynamics method. ###### keywords: opes, metadynamics, umbrella sampling, collective variables, free energy ## 1 Introduction Molecular dynamics has become a valuable tool in the study of a variety of phenomena in physics, chemistry, biology, and materials science. One of the long-standing challenges in this important field is the sampling of rare events, such as chemical reactions or conformational changes in biomolecules. To simulate effectively such systems, many enhanced sampling methods have been developed. An important class of such methods is based on an adaptive-bias approach and includes adaptive umbrella sampling1, metadynamics (MetaD) 2, 3, and the recently developed on-the-fly probability enhanced sampling (OPES)4, 5, 6. Adaptive-bias methods operate by adding to the system’s energy $U(\mathbf{R})$ an external bias potential $V=V(\mathbf{s})$, that is a function of a set of collective variables (CVs), $\mathbf{s}$. The CVs, $s=s(\mathbf{R})$, depend on the atomic coordinates $\mathbf{R}$ and are meant to describe the slow modes associated with the rare event under study. They also define a free energy surface (FES), $F(\mathbf{s})=-\frac{1}{\beta}\log P(\mathbf{s})$, where $\beta=(k_{B}T)^{-1}$ is the inverse Boltzmann factor and $P(\mathbf{s})$ the marginal $\mathbf{s}$ distribution, $P(\mathbf{s})\propto\int e^{-\beta U(\mathbf{R})}\delta[\mathbf{s}-\mathbf{s}(\mathbf{R})]\,d\mathbf{R}$. The bias is periodically updated until it converges to a chosen form. A popular choice is to have it exactly offset the underlying FES, $V(\mathbf{s})=-F(\mathbf{s})$, so that the resulting $\mathbf{s}$ distribution is uniform. The main limitation of adaptive-bias methods is that finding good collective variables is sometimes difficult and a bad choice of CVs might not promote the desired transitions in an affordable computer time. In practical applications one generally has to live with suboptimal CVs7 that still can drive transitions, but do not include some of the slow modes. In this case, applying a static bias cannot speed up the slow modes that are not accounted for, and thus transitions remain quite infrequent. It is sometimes possible to achieve a faster transition rate by using a rapidly changing bias, which can push the system out of a metastable state through a high free energy pathway, different from the energetically favoured one. However, unless one wishes to deal explicitly with out-of-equilibrium statistics8, 9, 10, it is not possible to obtain reliable information about the system while the bias changes in a non- adiabatic fashion. To estimate the FES and other observables one must let the adaptive-bias method approach convergence, and as the bias becomes quasi- static, transitions inevitably become less frequent. We refer to this situation as an exploration-convergence tradeoff that every adaptive-bias enhanced sampling method has to deal with, when suboptimal CVs are used. Some methods, like OPES, focus more on quickly converging to a quasi-static bias potential and thus obtaining an efficiently reweighted FES, while others, like metadynamics, focus more on escaping metastable states and exploring the phase space. We will demonstrate this qualitative difference on some prototypical systems. For simplicity, in the paper we only consider the well-tempered variant of metadynamics3, but in the Supporting Information (SI) we provide examples that use the original non-tempered MetaD2 and other popular variants, such as parallel-bias MetaD11. We propose here a variant of OPES, named OPES-explore, that focuses on rapid exploration, rather than on fast convergence. It shares many features with the original OPES, and is designed to be an easy-to-use tool requiring few input parameters. To this end, we also introduce an adaptive bandwidth algorithm that can be used in both OPES variants, and further reduces the number of input parameters that need to be specified. The detailed description of the adaptive bandwidth algorithm is left to the SI. All OPES simulations presented make use of this algorithm. ## 2 The OPES method The enhanced sampling method OPES works by adding an adaptive-bias potential to the energy of the system, so as to modify the Boltzmann probability distribution into a desired target one. Most adaptive-bias methods aim at sampling uniformly the CV space, but it has been shown that choosing a different target distribution could be advantageous12, 13. There are two different classes of target distributions that can be sampled with OPES; metadynamics-like and replica-exchange-like. We will consider here only the former type, introduced in Ref. 4, but the interested reader can find in Ref. 5 information about OPES for replica-exchange-like sampling. To define a metadynamics-like target distribution, one has to choose a set of collective variables, $s=s(\mathbf{R})$. As stated in the introduction, the unbiased marginal probability along such CVs is $P(\mathbf{s})\propto\int e^{-\beta U(\mathbf{R})}\delta[\mathbf{s}-\mathbf{s}(\mathbf{R})]\,d\mathbf{R}$, where $U(\mathbf{R})$ is the potential energy. The target distribution is then defined by requiring a specific marginal probability distribution over the CVs, $p^{\text{tg}}(\mathbf{s})$. Consequently, the desired bias potential is written as: $V(\mathbf{s})=-\frac{1}{\beta}\log\frac{p^{\text{tg}}(\mathbf{s})}{P(\mathbf{s})}\,,$ (1) so that $\int e^{-\beta[U(\mathbf{R})+V(\mathbf{s})]}\delta[\mathbf{s}-\mathbf{s}(\mathbf{R})]\,d\mathbf{R}\propto p^{\text{tg}}(\mathbf{s})$. A typical choice for $p^{\text{tg}}(\mathbf{s})$ is the well-tempered distribution3: $p^{\text{WT}}(\mathbf{s})\propto[P(\mathbf{s})]^{1/\gamma}\,,$ (2) where the bias factor $\gamma>1$ controls how much the original distribution is smoothed out. In the limit of $\gamma=\infty$ one targets a uniform distribution. The core idea of OPES is to update self-consistently the estimate of the probability distributions and of the bias potential, in an on-the-fly fashion similar to self-healing umbrella sampling 14. The estimate of the unbiased probability is obtained via a weighted kernel density estimation, so that at step $n$ one has: $P_{n}(\mathbf{s})=\frac{\sum_{k}^{n}w_{k}G(\mathbf{s},\mathbf{s}_{k})}{\sum_{k}^{n}w_{k}}\,,$ (3) where the weights $w_{k}$ are given by $w_{k}=e^{\beta V_{k-1}(\mathbf{s}_{k})}$, and the Gaussian kernels $G(\mathbf{s},\mathbf{s}^{\prime})=h\exp\left[-\frac{1}{2}(\mathbf{s}-\mathbf{s}^{\prime})^{T}\boldsymbol{\Sigma}^{-1}(\mathbf{s}-\mathbf{s}^{\prime})\right]$ have a diagonal covariance matrix $\Sigma_{ij}=\sigma^{2}_{i}\delta_{ij}$ and fixed height $h=\prod_{i}\left(\sigma_{i}\sqrt{2\pi}\right)^{-1}$. The number of kernels to represent $P_{n}(\mathbf{s})$ would grow linearly with simulation time, but this is avoided thanks to an on-the-fly kernel compression algorithm15, as described in detail in the supporting information of Ref. 4. The compression algorithm also allows for the bandwidth of the kernels to shrink over time, as the effective sample size $N_{\text{eff}}^{(n)}=\left(\sum_{k}^{n}w_{k}\right)^{2}/\sum_{k}^{n}w_{k}^{2}$ grows. The idea is to start with a coarse estimate of $P(\mathbf{s})$ and then refine it as more data are available. The kernel bandwidth of the $i$-th CV at step $n$ is: $\sigma_{i}^{(n)}=\sigma_{i}^{(0)}[N_{\text{eff}}^{(n)}(d+2)/4]^{-1/(d+4)}\,,$ (4) where $d$ is the total number of CVs. The instantaneous bias is based on the probability estimate $P_{n}(\mathbf{s})$, following Eq. (1) and using the approximation $p^{\text{WT}}(\mathbf{s})\propto[P_{n}(\mathbf{s})]^{1/\gamma}$, one has: $V_{n}(\mathbf{s})=(1-1/\gamma)\frac{1}{\beta}\log\left(\frac{P_{n}(\mathbf{s})}{Z_{n}}+\epsilon\right)\,,$ (5) where $\epsilon$ is a regularization term that limits the maximum possible absolute value of the bias potential, and $Z_{n}$ can be understood as a normalization of $P_{n}(\mathbf{s})$ over the CV space thus far explored, $\Omega_{n}$: $Z_{n}=\frac{1}{|\Omega_{n}|}\int_{\Omega_{n}}P_{n}(\mathbf{s})\,d\mathbf{s}\,.$ (6) This integral is calculated approximately as a sum of $P_{n}$ over the compressed kernels, as explained in the supplementary information of Ref. 4. The intuitive idea is that new kernels are added to the compressed representation only when a new region of CV space is sampled (otherwise they are merged with existing ones), thus the explored CV-space volume $|\Omega_{n}|$, is approximately proportional to the total number of compressed kernels. The introduction of the $Z_{n}$ term is one of the key innovations of OPES. In similar methods, once a new metastable state is found one often sees a dramatic increase of the exit time, compared to the first one16 (see SI, Fig. S5). This exit time problem is present also when the CVs are optimal, and should not be confused with the exploration-convergence tradeoff that is the primary concern of this paper. Other convergence-focused methods introduce extra parameters to tackle this problem, for example in transition-tempered metadynamics17 prior knowledge of the position of all metastable states is required. Instead, OPES avoids the exit time problem by taking into account the expansion of the CV space via the $Z_{n}$ term, which allows the bias to adjust more quickly when a new CV-space region is sampled4. At the start of an OPES simulation only a handful of parameters needs to be chosen, namely the initial kernel bandwidth, the pace at which the bias is updated, and the approximate FES barrier height that needs to be overcome. From this last information a prescription is given to automatically set the values $\gamma$ and $\epsilon$. The number of parameters can be reduced even further if one uses, as we shall do here, the adaptive bandwidth algorithm discussed in the SI. ## 3 An exploratory OPES variant We present now a new OPES variant called OPES-explore which, compared to the original OPES formulation, leads to a faster exploration of the phase space at the cost of a slower convergence. We have recalled that the $Z_{n}$ term allows OPES to quickly adapt the bias when a metastable state is found in a previously unexplored region of CV space. However, if the CVs used are suboptimal, it may happen that a new metastable state is found in an already explored $\mathbf{s}$ region18, 19. In such a case, the $Z_{n}$ term remains constant and is therefore ineffective in accelerating the exit time. Instead, to encourage a rapid exit, one would need a method that allows the bias to significantly change shape again. Fortunately, it is possible to achieve this exploratory behaviour simply by making a minimal change to the OPES protocol, which gives rise to the OPES-explore variant. In formulating OPES-explore, we restrict ourselves to the case of using as target the well-tempered distribution, $p^{\text{tg}}(\mathbf{s})=p^{\text{WT}}(\mathbf{s})$, Eq. (2). In OPES, the bias is expressed as a function of $P_{n}(\mathbf{s})$, the on-the-fly estimate of the unknown equilibrium distribution $P(\mathbf{s})$. At the beginning of the simulation this estimate is not reliable, but it improves over time and converges in a self-consistent way. In OPES-explore instead, one builds the bias starting from the on-the-fly estimate of the distribution that is being sampled in the biased simulation: $p^{\text{WT}}_{n}(\mathbf{s})=\frac{1}{n}\sum_{k}^{n}G(\mathbf{s},\mathbf{s}_{k})\,,$ (7) where $\mathbf{s}_{k}$ is the CVs value sampled at step $k$. As the simulation converges, $p^{\text{WT}}_{n}(\mathbf{s})$ approaches the target well-tempered distribution $p^{\text{WT}}(\mathbf{s})$. Thus, analogously to Sec. 2, we use the approximation $P(\mathbf{s})\propto[p^{\text{WT}}_{n}(\mathbf{s})]^{\gamma}$ and write the bias according to Eq. (1): $V_{n}(\mathbf{s})=(\gamma-1)\frac{1}{\beta}\log\left(\frac{p^{\text{WT}}_{n}(\mathbf{s})}{Z_{n}}+\epsilon\right)\,,$ (8) where $\epsilon$ and $Z_{n}$ have been added for the same reasons as in Eq. (5). We notice that the expressions in Eqs. (3) and (7), which define the probability estimates used in the two OPES schemes, converge respectively to $P(\mathbf{s})$ and $p^{\text{WT}}(\mathbf{s})$ only within the self- consistent scheme where the simulation runs with a bias that is updated on- the-fly according to Eqs. (5) and (8) respectively. Both OPES variants are applications of the general Eq. (1), but OPES estimates on-the-fly $P(\mathbf{s})$ and uses it to calculate the bias, while OPES-explore does the same but with $p^{\text{WT}}(\mathbf{s})\propto[P(\mathbf{s})]^{1/\gamma}$. The free energy surface as a function of the CVs can be estimated in two distinct ways, either directly from the probability estimate, $F_{n}(\mathbf{s})=-\gamma\frac{1}{\beta}\log p_{n}^{\text{WT}}(\mathbf{s})$, or via importance sampling reweighting, e.g. using a weighted kernel density estimation, $F_{n}(\mathbf{s})=-\frac{1}{\beta}\log\sum_{k}^{n}e^{\beta V_{k-1}(\mathbf{s}_{k})}G(\mathbf{s},\mathbf{s}_{k})\,.$ (9) In standard OPES these two estimates are equivalent, while in OPES-explore (similarly to MetaD) they can differ significantly in the first part of the simulation until they eventually converge to the same estimate. | ---|--- (a) OPES | (b) OPES EXPLORE Figure 1: Time evolution of a typical simulation of alanine dipeptide in vacuum using the two OPES variants with the dihedral angles $\phi$ and $\psi$ as CVs. For each method, the compressed kernels are shown on the left with the point size indicating the adaptive bandwidth, and the corresponding free energy estimate $F_{n}(\phi,\psi)$ on the right. (a) In the original OPES, kernels make up the unbiased distribution estimate $P_{n}(\phi,\psi)$ and $F_{n}(\phi,\psi)=-\frac{1}{\beta}\log P_{n}(\phi,\psi)$, while (b) in OPES- explore kernels make up the sampled distribution estimate $p^{\text{WT}}_{n}(\phi,\psi)$ and $F_{n}(\phi,\psi)=-\gamma\frac{1}{\beta}\log p^{\text{WT}}_{n}(\phi,\psi)$. All $F_{n}(\phi,\psi)$ are shifted to have zero minimum. Notice how OPES- explore requires fewer kernels and visits higher FES regions. In figure 1 we contrast an OPES and OPES-explore simulation of alanine dipeptide in vacuum, which has become a standard test for enhanced sampling methods. Both simulations have the same input parameters and use the adaptive bandwidth scheme described in the SI. The bias is initially quite coarse, but the width of the kernels reduces as the simulation proceeds and the details of the FES are increasingly better described. It can clearly be seen that the OPES-explore variant employs fewer kernels compared to the original OPES. This is due to the fact that in OPES-explore the kernel density estimation is used for $p^{\text{WT}}(\mathbf{s})\propto[P(\mathbf{s})]^{1/\gamma}$ that is a smoothed version of $P(\mathbf{s})$, and thus requires less details. This more compact representation can be useful especially in higher dimensions, where the number of kernels can greatly increase despite the compression algorithm. However, as a drawback it can result in a less accurate bias estimate, especially for large values of $\gamma$. ## 4 Fewer transitions can lead to better convergence Figure 2: (a) The Müller potential energy surface, $U(x,y)$. (b) The free energy surface along the $x$ coordinate, $F(x)$, with and without the addition of the bias potential $V(x)=-(1-1/\gamma)F(x)$, where $\gamma=20$. (c) The potential energy modified by the bias potential, $U(x,y)+V(x)$. It can be seen that, despite the almost flat profile along $x$, the transition region between the states remains at high energy. The difference in performance between OPES and OPES-explore cannot be judged from the alanine dipeptide example, because in this case the CVs chosen are extremely efficient. In order to highlight the difference between the two methods, we study a simple two-dimensional model potential that is known as the Müller potential20, see Fig. 2a, using the $x$ coordinate as collective variable. This is a clear example of suboptimal CV, since it can discriminate the metastable states, but not the transition state. For two-dimensional systems the free energy along the CV, $F(x)$, can be computed precisely with numerical integration, Fig. 2b. From $F(x)$, the free energy difference between the two metastable states can be calculated as $\Delta F=-\frac{1}{\beta}\log\frac{\int_{0}^{1}e^{-\beta F(x)}dx}{\int_{-1.3}^{0}e^{-\beta F(x)}dx}\,.$ (10) While it is possible to distinguish better the two states by using also the $y$ coordinate, this does not result in a significant difference in the $\Delta F$ value (see SI, Sec. S3). On the other hand, $x$ does a poor job of identifying the transition state, which is around $x\approx-0.7$ and $y\approx 0.6$, and not at $x\approx 0$ as it would seem from $F(x)$. As a consequence, it is not possible to significantly increase the transition rate between the states using a static bias that is a function of $x$ only. To show this, we consider the effect of adding to the system the converged well-tempered bias $V(x)=-(1-1/\gamma)F(x)$, with $\gamma=20$. In Fig. 2b, we can see the effect of the bias on the FES along $x$, which becomes almost completely flat. However, when we consider the full 2D landscape, Fig. 2c, we can see that such bias does not really remove the barrier between the two states. From the height of the barrier at the transition state, one can roughly estimate that adding $V(x)$ improves the transition rate of about one order of magnitude. Nevertheless, transitions remain quite rare, around one every $10^{6}$ uncorrelated samples (see SI, Sec. S3). We want to compare the two OPES variants and well-tempered metadynamics in this challenging setting, where CVs are suboptimal and the total simulation time is not enough to reach full convergence. This type of situation is not uncommon in practical applications, and it is thus of great interest. Given enough time, all the method considered converge to the same bias potential and sample the same target distribution, but we shall see that before reaching this limit they behave very differently. | ---|--- (a) $x$ trajectory | (b) $\Delta F$ estimate Figure 3: Typical simulations of the Müeller potential using different methods for biasing the $x$ coordinate. Given more time, the three methods will converge to the same bias potential and will sample the same target distribution. In (a) is the trajectory along the CV and in (b) is the corresponding $\Delta F_{n}$, Eq. 10, calculated using the FES estimate obtained directly from the applied bias, $F_{n}(x)=-(1-1/\gamma)^{-1}V_{n}(x)$. The correct $\Delta F$ value is highlighted by a blue stripe 1 k${}_{\text{B}}$T thick. Figure 3a shows a typical run of the Müller potential obtained by biasing the $x$ coordinate with OPES, OPES-explore or MetaD. As a simple way to visualize the evolution of the bias, we also report in Fig. 3b the $\Delta F_{n}$ estimate obtained directly from the applied bias, by using $F_{n}(x)=-(1-1/\gamma)^{-1}V_{n}(x)$ in Eq. (10). We can see a qualitative difference between OPES and the other two methods. OPES reaches a quasi-static bias that is very close to the converged one, but samples a distribution that is far from the well-tempered one, where the two basins would be about equally populated. On the other hand, the $x$ distribution sampled by OPES-explore is closer to the target well-tempered one, but its bias is far from converged, and makes ample oscillations around the correct value. Metadynamics behaves similarly to OPES-explore. This is the exploration-convergence tradeoff described in the introduction. Since the CV is suboptimal, even when using the converged bias $V(x)$, to see a transition occur one has to wait for an average number of steps $\tau\approx 10^{6}$, which is more than the total length of the simulation. However, it is possible to greatly accelerate transitions by using a time-dependent bias that forces the system into higher energy pathways, that are not accessible at equilibrium. In OPES-explore the bias is based on the estimate of the sampled probability $p^{\text{WT}}_{n}(\mathbf{s})$, and pushes to make it similar to the almost flat well-tempered target. This means that in order to have a quasi-static bias, $p^{\text{WT}}_{n}(\mathbf{s})$ should both be almost flat and not change significantly as the simulation proceeds. Clearly, this cannot happen unless the simulation is longer than $\tau$, otherwise most of the time would be spent in the same basin and $p^{\text{WT}}_{n}(\mathbf{s})$ would be far from flat. On the contrary, in OPES the bias is based on the reweighted estimate $P_{n}(\mathbf{s})$, and thus it can reach a quasi-static regime even before sampling the target distribution. | ---|--- (a) Direct $\Delta F$ | (b) Reweighted $\Delta F$ Figure 4: Estimate of the free energy difference $\Delta F$ for the Müeller potential obtained by averaging 25 independent runs for each biasing method. The standard deviation is also shown for each estimate. Given more time, all these estimates will converge to the correct $\Delta F$. All simulations start from the main basin, $x<0$, but with different initial conditions. In (a) is the estimate obtained directly from the applied bias, as in Fig. 3b, while in (b) is the corresponding estimate obtained via reweighting. For metadynamics two different reweighting schemes are considered, bias-offset 21, 22 and last- bias reweighting23, 19. The correct $\Delta F$ value is highlighted by a blue stripe 1 k${}_{\text{B}}$T thick. In figure 4a we show the $\Delta F_{n}$ estimate averaged over 25 independent runs, all starting from the main basin $x<0$. We can see that on average OPES provides the best $\Delta F_{n}$ estimate at any $n$ in spite of the fact that it induces far less transitions. In fact, most of the time only one full back- and-forth transition is observed (see SI). One should notice that after a single transition the $\Delta F_{n}$ estimate is far from being accurate (see Fig. 3b) but, since the bias quickly becomes quasi-static, it is possible to collect equilibrium samples and reliably reweight them, and the average estimate becomes more accurate the more simulations are run. Instead in OPES- explore and MetaD, despite starting from independent initial conditions, the runs are highly correlated, due to the transitions being mostly driven by the strong changes in the bias rather than the natural fluctuations of the system. As a further consequence of this, a systematic error is present in the average estimate, even if $\Delta F_{n}$ is further averaged over time, to remove the oscillatory behaviour of OPES-explore and MetaD. Such systematic error depends on the characteristic of the system and the chosen CVs, and is hard to predict weather it will be relevant or small. Nevertheless, one can be sure that it reduces over time as the bias converges24. Estimates of $\Delta F_{n}$ using different reweighting schemes are shown in Fig. 4b. For OPES and OPES-explore the simple Eq. (9) has been used, while for MetaD we consider two of the most popular reweighiting schemes, namely last- bias reweighting23, 19 and bias-offset reweighting21, 22. As expected, the reweighting estimate of OPES is virtually identical to the direct estimate obtained from the bias, while for the other two methods the two estimates differ. The reweighing of OPES-explore has very small statistical uncertainty, which further highlights the presence of a systematic error in the free energy difference estimate. Like others before us22, 25, 26, we observe empirically that the last-bias reweighting for MetaD tends to always be in agreement with the direct estimate, even when the simulation is far from converged, while the bias-offset reweighting provides a very unreliable estimate if the MetaD bias has not reached a quasi-static regime and the initial part of the simulation is not discarded. Once again, it must be noted that the simulations considered here are not fully converged, otherwise all the different estimates of the various methods would have yielded the correct result, without systematic errors. However, for most practical purposes they behave very differently, thus it is important to choose between an exploration-focused or a convergence-focused enhanced sampling method, depending on the specific aim of the simulation. ## 5 Sometimes exploration is what matters In the examples of the previous paragraph, it was shown in that OPES converges to a quasi-static bias faster than OPES-explore and provides more accurate FES estimates. However, FES estimation is not the only goal of an enhanced sampling simulation. In complex systems where good CVs are not available, convergence can remain out of reach, still one might be interested in exploring the phase space and find all the relevant metastable basins. In such situation, OPES-explore can be a useful tool. Figure 5: The eight metastable basins of alanine tetrapeptide in vacuum sampled via OPES-explore by biasing the three $\psi$ angles, a suboptimal set of CVs. Each basin is identified by the sign of the three $\phi$ angles, for a total of $2^{3}$ possible combinations. The most stable basin has $\phi_{1},\phi_{2},\phi_{3}<0$, while for the least stable $\phi_{1},\phi_{2},\phi_{3}>0$. Figure 6: Exploration time of the eight metastable basins of alanine tetrapeptide over 100 ns. The lines are an average over 10 independent runs for each method, showing the total number of visited basins. In (a) the bias is a function of the three $\phi$ angles, $V=V(\phi_{1},\phi_{2},\phi_{3})$, while in (b) the three $\psi$ angles are used, $V=V(\psi_{1},\psi_{2},\psi_{3})$. See SI for results with different input parameters and other MetaD variants, such as parallel-bias MetaD11. We consider here as test system alanine tetrapeptide in vacuum, as in Ref. 4. It has three $\phi$ dihedral angles, each of them can change from positive to negative values and vice versa with a relatively low probability. This leads to $2^{3}=8$ distinct metastable basins, each corresponding to a different combination of $\phi$ angles signs, as shown in Fig. 5. Here we are not interested in estimating the FES, but rather we want to compare the ability of different methods to explore this space and discover all metastable states. Figure 6 shows the number of explored basins averaged over 10 independent simulations for each enhanced sampling method. The simulations in the top panel (Fig. 6a) use as CVs the $\phi$ angles, $V=V(\phi_{1},\phi_{2},\phi_{3})$, which are good CVs, while in the bottom (Fig. 6b) the suboptimal $\psi$ angles are used, $V=V(\psi_{1},\psi_{2},\psi_{3})$. In all methods, the exploration time increases approximately by two orders of magnitude when suboptimal CVs are used (please note the horizontal logarithmic scale). As expected, OPES and OPES-explore have similar exploration speed when using good CVs, while with suboptimal CVs OPES struggles to find all the metastable basins. This is because the same region of CV space might correspond to two different metastable basins, or to a basin and a transition state, as for the Müller potential18, 19. In this situation, the previously estimated bias must change considerably for the simulation to escape quickly the current metastable state. The exploration speed of MetaD depends critically on the input parameters and requires a trial-and-error tuning. We report here only the outcome of MetaD simulations in which a standard choice of the input parameters has been made. As can be seen in Fig. 6, in these simulations the exploration speed is roughly one order of magnitude slower than that of OPES-explore. However, the performance of MetaD simulations can be improved by using different settings, as shown in the SI. In the SI we also report and briefly discuss results obtained with non-tempered metadynamics2, adaptive-Gaussians metadynamics23 and parallel-bias metadynamics11. None of these MetaD variants significantly improve the exploration speed, and some make it even worse. Finally, in the SI we show how a preliminary OPES-explore run can be combined with a multithermal OPES simulation5, 6 to sample efficiently alanine tetrapeptide and reach a converged FES, even without explicitly biasing the $\phi$ angles. ## 6 Conclusion We have shown with the help of model systems that there is an exploration- convergence tradeoff in adaptive-bias methods when suboptimal CVs are used. This tradeoff should not be confused with the exit time problem, that is present also with optimal CVs, and is discussed in Sec. 2 and Refs. 17, 16. Contrary to the exit time problem, the exploration-convergence tradeoff cannot be solved and is an intrinsic limitation of CV-based adaptive-bias methods, that is a consequence of suboptimal CVs. We believe the best way to handle this tradeoff is to have separate methods that clearly focus on one or the other aspect, so that they can be used depending on the application. In a convergence-focused method the bias soon becomes quasi-static to allow for accurate reweighting and free energy estimation. However, with suboptimal CVs this leads to a slow transition rate and a long time is required to sample the target distribution. As discussed, even if one knows the true $F(\mathbf{s})$ and directly applies the converged bias, one would not obtain a faster exploration. In an exploration-focused method, it is possible to improve the exploration speed by letting the bias change substantially even in a CV region that has already been visited. While this may increase the number of transitions, it comes at the cost of a less accurate estimate of the free energy. The original OPES method focuses on fast convergence to provide an accurate estimate of the free energy surface and reweighted observables. As a consequence, it is very sensitive to the quality of the CVs (see e.g. Fig. 3a) and any improvement in the CVs results in a clear acceleration of the transition rate. This is a particularly useful property when developing machine learning-based CVs, and in fact OPES has already been used several times in this context27, 28, 29, 30, 31, 32. In other situations, improving the CVs may require first a better exploration of the phase space33, 34, 35, 28. Furthermore, one may be interested simply in exploring the metastable states of a system rather than estimating an accurate FES36, 37, 38, 39. For this reason, we have introduced a variant of the OPES method, OPES-explore, that focuses on quickly sampling the target distribution and exploring the phase space. We have shown that also well-tempered metadynamics is an exploration-focused method. One of the main advantages of OPES-explore over MetaD is that it is easier to use, since it requires fewer input parameters and it has a more straightforward reweighting scheme (but more advanced ones can also be used25, 40). Another important difference between the two methods is that OPES- explore, similarly to OPES, by default provides a maximum threshold to the applied bias potential, thus it avoids unreasonably high free energy regions. To obtain the same effect with MetaD, one typically has to define some ad hoc static bias walls by trial and error. This last feature of OPES-explore has been recently leveraged by Raucci et al. to systematically discover reaction pathways in chemical processes41. Finally, we should clarify that OPES-explore, just as metadynamics, might not be able to exit any metastable state if the CVs are too poor42, 19, and its improved exploration capability can only be harnessed if the CVs are close enough to the correct ones to make such transitions possible. The speed and small number of input parameters of OPES-explore are extremely helpful for quickly testing several candidate CVs, to find out which can drive transitions and discard the bad ones. We believe that OPES-explore is an important addition to the OPES family of methods and will become a useful tool for researchers as it pushes forward the trend for more robust and reliable enhanced sampling methods. We thank Valerio Rizzi and Umberto Raucci for useful discussions. M.I. acknowledges support from the Swiss National Science Foundation through an Early Postdoc.Mobility fellowship. Calculations were carried out on Euler cluster at ETH Zurich and on workstations provided by USI Lugano. ## Data availability An open-source implementation of the OPES and OPES-explore methods is available in the enhanced sampling library PLUMED from version 2.843. All the data and input files needed to reproduce the simulations presented in this paper are available on PLUMED-NEST (www.plumed-nest.org), the public repository of the PLUMED consortium 44, as plumID:22.003 . Description of the adaptive bandwidth algorithm, computational details regarding the Müller potential and further biased trajectories, exploration speed for alanine tetrapeptide using other methods, and description of a multithermal-multiumbrella simulation to improve upon the OPES-explore run. ## References * Mezei 1987 Mezei, M. Adaptive umbrella sampling: Self-consistent determination of the non-Boltzmann bias. _Journal of Computational Physics_ 1987, _68_ , 237–248 * Laio and Parrinello 2002 Laio, A.; Parrinello, M. Escaping free-energy minima. _Proceedings of the National Academy of Sciences_ 2002, _99_ , 12562–12566 * Barducci et al. 2008 Barducci, A.; Bussi, G.; Parrinello, M. Well-Tempered Metadynamics: A Smoothly Converging and Tunable Free-Energy Method. _Physical Review Letters_ 2008, _100_ , 020603 * Invernizzi and Parrinello 2020 Invernizzi, M.; Parrinello, M. Rethinking Metadynamics: From Bias Potentials to Probability Distributions. _The Journal of Physical Chemistry Letters_ 2020, _11_ , 2731–2736 * Invernizzi et al. 2020 Invernizzi, M.; Piaggi, P. M.; Parrinello, M. Unified Approach to Enhanced Sampling. _Physical Review X_ 2020, _10_ , 41034 * Invernizzi 2021 Invernizzi, M. OPES: On-the-fly Probability Enhanced Sampling Method. _Il Nuovo Cimento C_ 2021, _44_ , 112 * Invernizzi and Parrinello 2019 Invernizzi, M.; Parrinello, M. Making the Best of a Bad Situation: A Multiscale Approach to Free Energy Calculation. _Journal of Chemical Theory and Computation_ 2019, _15_ , 2187–2194 * Jarzynski 1997 Jarzynski, C. Nonequilibrium Equality for Free Energy Differences. _Physical Review Letters_ 1997, _78_ , 2690–2693 * Donati and Keller 2018 Donati, L.; Keller, B. G. Girsanov reweighting for metadynamics simulations. _The Journal of Chemical Physics_ 2018, _149_ , 072335 * Bal 2021 Bal, K. M. Reweighted Jarzynski Sampling: Acceleration of Rare Events and Free Energy Calculation with a Bias Potential Learned from Nonequilibrium Work. _Journal of Chemical Theory and Computation_ 2021, _17_ , 6766–6774 * Pfaendtner and Bonomi 2015 Pfaendtner, J.; Bonomi, M. Efficient Sampling of High-Dimensional Free-Energy Landscapes with Parallel Bias Metadynamics. _Journal of Chemical Theory and Computation_ 2015, _11_ , 5062–5067 * Valsson and Parrinello 2014 Valsson, O.; Parrinello, M. Variational Approach to Enhanced Sampling and Free Energy Calculations. _Physical Review Letters_ 2014, _113_ , 090601 * White et al. 2015 White, A. D.; Dama, J. F.; Voth, G. A. Designing Free Energy Surfaces That Match Experimental Data with Metadynamics. _Journal of Chemical Theory and Computation_ 2015, _11_ , 2451–2460 * Marsili et al. 2006 Marsili, S.; Barducci, A.; Chelli, R.; Procacci, P.; Schettino, V. Self-healing Umbrella Sampling: A Non-equilibrium Approach for Quantitative Free Energy Calculations. _The Journal of Physical Chemistry B_ 2006, _110_ , 14011–14013 * Sodkomkham et al. 2016 Sodkomkham, D.; Ciliberti, D.; Wilson, M. A.; Fukui, K.-I.; Moriyama, K.; Numao, M.; Kloosterman, F. Kernel density compression for real-time Bayesian encoding/decoding of unsorted hippocampal spikes. _Knowledge-Based Systems_ 2016, _94_ , 1–12 * Fort et al. 2017 Fort, G.; Jourdain, B.; Lelièvre, T.; Stoltz, G. Self-healing umbrella sampling: convergence and efficiency. _Statistics and Computing_ 2017, _27_ , 147–168 * Dama et al. 2014 Dama, J. F.; Rotskoff, G.; Parrinello, M.; Voth, G. A. Transition-tempered metadynamics: Robust, convergent metadynamics via on-the-fly transition barrier estimation. _Journal of Chemical Theory and Computation_ 2014, _10_ , 3626–3633 * Pietrucci 2017 Pietrucci, F. Strategies for the exploration of free energy landscapes: Unity in diversity and challenges ahead. _Reviews in Physics_ 2017, _2_ , 32–45 * Bussi and Laio 2020 Bussi, G.; Laio, A. Using metadynamics to explore complex free-energy landscapes. _Nature Reviews Physics_ 2020, * Müller and Brown 1979 Müller, K.; Brown, L. D. Location of saddle points and minimum energy paths by a constrained simplex optimization procedure. _Theoretica Chimica Acta_ 1979, _53_ , 75–93 * Tiwary and Parrinello 2015 Tiwary, P.; Parrinello, M. A time-independent free energy estimator for metadynamics. _Journal of Physical Chemistry B_ 2015, _119_ , 736–742 * Valsson et al. 2016 Valsson, O.; Tiwary, P.; Parrinello, M. Enhancing Important Fluctuations: Rare Events and Metadynamics from a Conceptual Viewpoint. _Annual Review of Physical Chemistry_ 2016, _67_ , 159–184 * Branduardi et al. 2012 Branduardi, D.; Bussi, G.; Parrinello, M. Metadynamics with Adaptive Gaussians. _Journal of Chemical Theory and Computation_ 2012, _8_ , 2247–2254 * Dama et al. 2014 Dama, J. F.; Parrinello, M.; Voth, G. A. Well-Tempered Metadynamics Converges Asymptotically. _Physical Review Letters_ 2014, _112_ , 240602 * Marinova and Salvalaglio 2019 Marinova, V.; Salvalaglio, M. Time-independent free energies from metadynamics via mean force integration. _The Journal of Chemical Physics_ 2019, _151_ , 164115 * Giberti et al. 2020 Giberti, F.; Cheng, B.; Tribello, G. A.; Ceriotti, M. Iterative Unbiasing of Quasi-Equilibrium Sampling. _Journal of Chemical Theory and Computation_ 2020, _16_ , 100–107 * Bonati et al. 2020 Bonati, L.; Rizzi, V.; Parrinello, M. Data-Driven Collective Variables for Enhanced Sampling. _The Journal of Physical Chemistry Letters_ 2020, _11_ , 2998–3004 * Bonati et al. 2021 Bonati, L.; Piccini, G.; Parrinello, M. Deep learning the slow modes for rare events sampling. _Proceedings of the National Academy of Sciences_ 2021, _118_ , e2113533118 * Karmakar et al. 2021 Karmakar, T.; Invernizzi, M.; Rizzi, V.; Parrinello, M. Collective variables for the study of crystallisation. _Molecular Physics_ 2021, _119_ , e1893848 * Trizio and Parrinello 2021 Trizio, E.; Parrinello, M. From Enhanced Sampling to Reaction Profiles. _The Journal of Physical Chemistry Letters_ 2021, _12_ , 8621–8626 * Rizzi et al. 2021 Rizzi, V.; Bonati, L.; Ansari, N.; Parrinello, M. The role of water in host-guest interaction. _Nature Communications_ 2021, _12_ , 93 * Ansari et al. 2021 Ansari, N.; Rizzi, V.; Carloni, P.; Parrinello, M. Water-Triggered, Irreversible Conformational Change of SARS-CoV-2 Main Protease on Passing from the Solid State to Aqueous Solution. _Journal of the American Chemical Society_ 2021, _143_ , 12930–12934 * Branduardi et al. 2007 Branduardi, D.; Gervasio, F. L.; Parrinello, M. From A to B in free energy space. _The Journal of Chemical Physics_ 2007, _126_ , 054103 * McCarty and Parrinello 2017 McCarty, J.; Parrinello, M. A variational conformational dynamics approach to the selection of collective variables in metadynamics. _The Journal of Chemical Physics_ 2017, _147_ , 204109 * Mendels et al. 2018 Mendels, D.; Piccini, G.; Parrinello, M. Collective Variables from Local Fluctuations. _Journal of Physical Chemistry Letters_ 2018, _9_ , 2776–2781 * Piaggi and Parrinello 2018 Piaggi, P. M.; Parrinello, M. Predicting polymorphism in molecular crystals using orientational entropy. _Proceedings of the National Academy of Sciences_ 2018, _115_ , 10251–10256 * Capelli et al. 2019 Capelli, R.; Carloni, P.; Parrinello, M. Exhaustive Search of Ligand Binding Pathways via Volume-Based Metadynamics. _The Journal of Physical Chemistry Letters_ 2019, _10_ , 3495–3499 * Ahlawat et al. 2021 Ahlawat, P.; Hinderhofer, A.; Alharbi, E. A.; Lu, H.; Ummadisingu, A.; Niu, H.; Invernizzi, M.; Zakeeruddin, S. M.; Dar, M. I.; Schreiber, F.; Hagfeldt, A.; Grätzel, M.; Rothlisberger, U.; Parrinello, M. A combined molecular dynamics and experimental study of two-step process enabling low-temperature formation of phase-pure $\alpha$-FAPbI 3. _Science Advances_ 2021, _7_ , eabe3326 * Francia et al. 2021 Francia, N. F.; Price, L. S.; Salvalaglio, M. Reducing crystal structure overprediction of ibuprofen with large scale molecular dynamics simulations. _CrystEngComm_ 2021, _23_ , 5575–5584 * Carli and Laio 2021 Carli, M.; Laio, A. Statistically unbiased free energy estimates from biased simulations. _Molecular Physics_ 2021, _119_ * Raucci et al. 2022 Raucci, U.; Rizzi, V.; Parrinello, M. Discover, Sample, and Refine: Exploring Chemistry with Enhanced Sampling Techniques. _The Journal of Physical Chemistry Letters_ 2022, 1424–1430 * Bussi and Branduardi 2015 Bussi, G.; Branduardi, D. _Free-Energy Calculations with Metadynamics: Theory and Practice_ ; John Wiley & Sons, Inc, 2015; pp 1–49 * Tribello et al. 2014 Tribello, G. A.; Bonomi, M.; Branduardi, D.; Camilloni, C.; Bussi, G. PLUMED 2: New feathers for an old bird. _Computer Physics Communications_ 2014, _185_ , 604–613 * The PLUMED consortium 2019 The PLUMED consortium, Promoting transparency and reproducibility in enhanced molecular simulations. _Nature Methods_ 2019, _16_ , 670–673 See pages - of SupportingInformation.pdf
# The Based Rings of Two-sided cells in an Affine Weyl group of type $\tilde{B}_{3}$, II Yannan Qiu∗ and Nanhua XI† ∗ School of Mathematical Sciences Zhejiang University Zhejiang 310058, China China<EMAIL_ADDRESS>† Academy of Mathematics and Systems Science Chinese Academy of Sciences Beijing 100190, China and School of Mathematical Sciences University of Chinese Academy of Sciences Chinese Academy of Sciences Beijing 100049, China<EMAIL_ADDRESS>Dedicated to George Lusztig with greatest respect. ###### Abstract. We compute the based rings of two-sided cells corresponding to the unipotent classes in $Sp_{6}(\mathbb{C})$ with Jordan blocks (33), (411), (222), respectively. The results for the first two two-sided cells also verify Lusztig’s conjecture on the structure of the based rings of two-sided cells of an affine Weyl group. The result for the last two-sided cell partially suggests a modification of Lusztig’s conjecture on the structure of the based rings of two-sided cells of an affine Weyl group. Y. Qiu was partially supported by National Natural Science Foundation of China, No. 12171030. N. Xi was partially supported by National Key R&D Program of China, No. 2020YFA0712600, and by National Natural Science Foundation of China, No. 11688101. We are concerned with the based rings of two-sided cells in an affine Weyl group of type $\tilde{B}_{3}$. In a previous paper we discussed the based ring of the two-sided cell corresponding to the nilpotent element in $Sp_{6}(\mathbb{C})$ with 3 equal Jordan blocks and showed that Lusztig’s conjecture on the structure of the based rings of the two-sided cells of an affine Weyl group needs modification (see section 4 in [QX]). In this paper we compute the based rings of two-sided cells corresponding to the unipotent classes in $Sp_{6}(\mathbb{C})$ with Jordan blocks (411), (33), (222), respectively. The results for the first two two-sided cells also verify Lusztig’s conjecture on the structure of the based rings of two-sided cells of an affine Weyl group. The result for the last two-sided cell partially suggests a modification of Lusztig’s conjecture on the structure of the based rings of two-sided cells of an affine Weyl group. For the first two two-sided cells, the validity of Lusztig’s conjecture on the based rings is already included in the main theorem in [BO]. Here we construct the bijection in Lusztig’s conjecture explicitly so that the results in this paper can be used for computing certain irreducible representations of affine Hecke algebras of type $\tilde{B}_{3}$. In section 5 we give a description for the based ring of the two-sided cell corresponding to the nilpotent element in $Sp_{6}(\mathbb{C})$ with Jordan blocks (222), which can also be used to compute certain irreducible representations of affine Hecke algebras of type $\tilde{B}_{3}$. The contents of the paper are as follows. Section 1 is devoted to preliminaries, which include some basic facts on (extended) affine Weyl groups and their Hecke algebras and formulation of Lusztig’s conjecture on the structure of the based ring of a two-sided cell in an affine Weyl group. In section 2 we recall some results on cells of the (extended) affine Weyl group of type $\tilde{B}_{3}$, which are due to J. Du. Sections 3, 4, 5 are devoted to discussing based rings of two-sided cells corresponding to the unipotent classes in $Sp_{6}(\mathbb{C})$ with Jordan blocks (411), (33), (222), respectively. ## 1\. Affine Weyl groups and their Hecke algebras In this section we fix some notations and refer to [KL, L1, L2, L3, QX] for more details. 1.1. Extended affine Weyl groups and their Hecke algebras Let $G$ be a connected reductive algebraic group over the field $\mathbb{C}$ of complex numbers. Let $W_{0}$ be the Weyl group of $G$ and $W$ the affine Weyl group attached to $G$. The set of simple reflections of $W$ is denoted by $S$. We shall denote the length function of $W$ by $l$ and use $\leq$ for the Bruhat order on $W$. We also often write $y<w$ or $w>y$ if $y\leq w$ and $y\neq w$. Let $H$ be the Hecke algebra of $(W,S)$ over $\mathcal{A}=\mathbb{Z}[q^{\frac{1}{2}},q^{-\frac{1}{2}}]$ $(q$ an indeterminate) with parameter $q$. Let $\\{T_{w}\\}_{w\in W}$ be its standard basis. Then we have $(T_{s}-q)(T_{s}+1)=0$ and $T_{w}T_{u}=T_{wu}$ if $l(wu)=l(w)+l(u)$. Let $C_{w}=q^{-\frac{l(w)}{2}}\sum_{y\leq w}P_{y,w}T_{y},\ w\in W$ be the Kazhdan-Lusztig basis of $H$, where $P_{y,w}$ are the Kazhdan- Lusztig polynomials. The degree of $P_{y,w}$ is less than or equal to $\frac{1}{2}(l(w)-l(y)-1)$ if $y<w$ and $P_{w,w}=1$. Convention: set $P_{y,w}=0$ if $y\not\leq w$. If $y<w$, we write $P_{y,w}=\mu(y,w)q^{\frac{1}{2}(l(w)-l(y)-1)}+\text{lower degree terms.}$ We shall write $y\prec w$ if $\mu(y,w)\neq 0$. We have (a) Let $y\leq w$. Assume that $sw\leq w$ for some $s\in S$. Then $\displaystyle P_{y,w}$ $\displaystyle=P_{sy,w},\ \text{if}\ sy>y;$ $\displaystyle P_{y,w}$ $\displaystyle=q^{1-c}P_{sy,sw}+q^{c}P_{y,sw}-\sum_{\stackrel{{\scriptstyle\stackrel{{\scriptstyle z\in W}}{{y\leq z\prec sw}}}}{{sz<z}}}\mu(z,sw)q^{\frac{l(w)-l(z)}{2}}P_{y,z},$ where $c=1$ if $sy<y$ and $c=0$ if $sy>y$. (b) Let $y\leq w$. Assume that $ws\leq w$ for some $s\in S$. Then $\displaystyle P_{y,w}$ $\displaystyle=P_{ys,w},\ \text{if}\ ys>y;$ $\displaystyle\ P_{y,w}$ $\displaystyle=q^{1-c}P_{ys,ws}+q^{c}P_{y,ws}-\sum_{\stackrel{{\scriptstyle\stackrel{{\scriptstyle z\in W}}{{y\leq z\prec ws}}}}{{zs<z}}}\mu(z,ws)q^{\frac{l(w)-l(z)}{2}}P_{y,z},$ where $c=1$ if $ys<y$ and $c=0$ if $ys>y$. From the two formulas above one gets (see [KL]) (c) Let $y,w\in W$ and $s\in S$ be such that $y<w,\ sw<w,$ and $sy>y$. Then $y\prec w$ if and only if $w=sy$. Moreover this implies that $\mu(y,w)=1$. (d) Let $y,w\in W$ and $s\in S$ be such that $y<w,\ ws<w$, and $ys>y$. Then $y\prec w$ if and only if $w=ys$. Moreover this implies that $\mu(y,w)=1$. The following formulas for computing $C_{w}$ (see [KL]) will be used in sections 3, 4, 5. (e) Let $w\in W$ and $s\in S$. Then (1) $\displaystyle C_{s}C_{w}=\begin{cases}\displaystyle(q^{\frac{1}{2}}+q^{-\frac{1}{2}})C_{w},\quad&\text{if\ }sw<w,\\\ \displaystyle C_{sw}+\sum_{\stackrel{{\scriptstyle z\prec w}}{{sz<z}}}\mu(z,w)C_{z},\quad&\text{if\ }sw\geq w.\end{cases}$ (2) $\displaystyle C_{w}C_{s}=\begin{cases}\displaystyle(q^{\frac{1}{2}}+q^{-\frac{1}{2}})C_{w},\quad&\text{if\ }ws<w,\\\ \displaystyle C_{ws}+\sum_{\stackrel{{\scriptstyle z\prec w}}{{zs<z}}}\mu(z,w)C_{z},\quad&\text{if\ }ws\geq w.\end{cases}$ 1.2. Cells of affine Weyl groups We refer to [KL] for definition of left cells, right cells and two-sided cells of $W$. For $h,\,h^{\prime}\in H$ and $x\in W$, write $hC_{x}=\sum_{y\in W}a_{y}C_{y},\quad C_{x}h=\sum_{y\in W}b_{y}C_{y},\quad hC_{x}h^{\prime}=\sum_{y\in W}c_{y}C_{y},\quad a_{y},b_{y},c_{y}\in\mathcal{A}.$ Define $y\underset{L}{\leq}x$ if $a_{y}\neq 0$ for some $h\in H$, $y\underset{R}{\leq}x$ if $b_{y}\neq 0$ for some $h\in H$, and $y\underset{LR}{\leq}x$ if $c_{y}\neq 0$ for some $h,h^{\prime}\in H$. We write $x\underset{L}{\sim}y$ if $x\underset{L}{\leq}y\underset{L}{\leq}x$, $x\underset{R}{\sim}y$ if $x\underset{R}{\leq}y\underset{R}{\leq}x$, and $x\underset{LR}{\sim}y$ if $x\underset{LR}{\leq}y\underset{LR}{\leq}x$. Then $\underset{L}{\sim},\ \underset{R}{\sim},\ \underset{LR}{\sim}$ are equivalence relations on $W$. The equivalence classes are called left cells, right cells, and two-sided cells of $W$ respectively. Note that if $\Gamma$ is a left cell of $W$, then $\Gamma^{-1}=\\{w^{-1}\,|\,w\in\Gamma\\}$ is a right cell. For $w\in W$, set $R(w)=\\{s\in S\,|\,ws\leq w\\}$ and $L(w)=\\{s\in S\,|\,sw\leq w\\}.$ Then we have (see [KL]) (a) $R(w)\subset R(u)$ if $u\underset{L}{\leq}w$ and $L(w)\subset L(u)$ if $u\underset{R}{\leq}w.$ In particular, $R(w)=R(u)$ if $u\underset{L}{\sim}w$ and $L(w)=L(u)$ if $u\underset{R}{\sim}w.$ 1.3. $*$-operations The $*$-operation introduced in [KL] and generalized in [L1] is a useful tool in the theory of cells of Coxeter groups. Let $s,t$ be simple reflections in $S$ and assume that $st$ has order $m\geq 3$. Let $w\in W$ be such that $sw\geq w,\ tw\geq w$. The $m-1$ elements $sw,\ tsw,\ stsw,\ ...,$ is called a left string (with respect to $\\{s,t\\})$, and the $m-1$ elements $tw,\ stw,\ tstw,\ ...,$ is also called a left string (with respect to $\\{s,t\\})$. Similarly we define right strings (with respect to $\\{s,t\\}$). Then (see [L1]) (a) A left string in $W$ is contained in a left cell of $W$ and a right string in $W$ is contained in a right cell of $W$. Assume that $x$ is in a left (resp. right ) string (with respect to $\\{s,t\\})$ of length $m-1$ and is the $i$th element of the left (resp. right) string, define ${}^{*}x$ (resp. $x^{*}$) to be the $(m-i)$th element of the string, where $*=\\{s,t\\}$. The following result is proved in [X2]. (b) Let $x$ be in $W$ such that $x$ is in a left string with respect to $*=\\{s,t\\}$ and is also in a right string with respect to $\star=\\{s^{\prime},t^{\prime}\\}$. Then ${}^{*}x$ is in a right string with respect to $\\{s^{\prime},t^{\prime}\\}$ and $x^{\star}$ is in a left string with respect to $\\{s,t\\}$. Moreover ${}^{*}(x^{\star})=({}^{*}x)^{\star}$. We shall write ${}^{*}x^{\star}$ for ${}^{*}(x^{\star})=({}^{*}x)^{\star}$. The following result is due to Lusztig [L1]. (c) Let $\Gamma$ be a left cell of $W$ and an element $x\in\Gamma$ is in a right string $\sigma_{x}$ with respect to $*=\\{s,t\\}$. Then any element $w\in\Gamma$ is in a right string $\sigma_{w}$ with respect to $*=\\{s,t\\}$. Moreover $\Gamma^{*}=\\{w^{*}\,|\,w\in\Gamma\\}$ is a left cell of $W$ and $\Omega=\displaystyle\left(\cup_{w\in\Gamma}\sigma_{w}\right)-\Gamma$ is a union of at most $m-2$ left cells. Following Lusztig [L1] we set $\tilde{\mu}(y,w)=\mu(y,w)$ if $y<w$ and $\tilde{\mu}(y,w)=\mu(w,y)$ if $w<y$. For convenience we also set $\tilde{\mu}(y,w)=0$ if $y\nless w$ and $w\nless y$. Assume that $x_{1},x_{2},...,x_{m-1}$ and $y_{1},y_{2},...,y_{m-1}$ are two left strings with respect to $*=\\{s,t\\}$. Define $a_{ij}=\begin{cases}\tilde{\mu}(x_{i},y_{j}),\quad&\text{if\ }\\{s,t\\}\cap L(x_{i})=\\{s,t\\}\cap L(y_{j}),\\\ 0,\quad&\text{otherwise}.\end{cases}$ Lusztig proved the following identities (see Subsection 10.4 in [L1]). (d) If $m=3$, then $a_{11}=a_{22}$ and $a_{12}=a_{21}$. (e) If $m=4$, then (3) $\displaystyle a_{11}=a_{33},\ a_{13}=a_{31},\ a_{22}=a_{11}+a_{13},\ a_{12}=a_{21}=a_{23}=a_{32}.$ 1.4. Lusztig’s $a$-function For $x,y\in W$, write $C_{x}C_{y}=\sum_{z\in W}h_{x,y,z}C_{z},\qquad h_{x,y,z}\in\mathcal{A}=\mathbb{Z}[q^{\frac{1}{2}},q^{-\frac{1}{2}}].$ Following Lusztig ([L1]), we define $a(z)={\rm min}\\{i\in\mathbf{N}\ |\ q^{-\frac{i}{2}}h_{x,y,z}\in\mathbb{Z}[q^{-\frac{1}{2}}]{\rm\ for\ all\ }x,y\in W\\}.$ If for any $i$, $q^{-\frac{i}{2}}h_{x,y,z}\not\in\mathbb{Z}[q^{-\frac{1}{2}}]{\rm\ for\ some\ }x,y\in W$, we set $a(z)=\infty.$ The following properties are proved in [L1]. (a) We have $a(w)\leq l(w_{0})$ for any $w\in W$, where $w_{0}$ is the longest element in the Weyl group $W_{0}$. (b) $a(x)\geq a(y)$ if $x\underset{LR}{\leq}y$. In particular, $a(x)=a(y)$ if $x\underset{LR}{\sim}y$. (c) $x\underset{L}{\sim}y$ (resp. $x\underset{R}{\sim}y,\ x\underset{LR}{\sim}y$) if $a(x)=a(y)$ and $x\underset{L}{\leq}y$ (resp. $x\underset{R}{\leq}y,\ x\underset{LR}{\leq}y$). (d) If $h_{x,y,z}\neq 0$, then $z\underset{R}{\leq}x$ and $z\underset{L}{\leq}y$. In particular, $a(z)>a(x)$ if $z\not\underset{R}{\sim}x$, and $a(z)>a(y)$ if $z\not\underset{L}{\sim}y$. Following Lusztig, we define $\gamma_{x,y,z}$ by the following formula, $h_{x,y,z}=\gamma_{x,y,z}q^{\frac{a(z)}{2}}+{\rm\ lower\ degree\ terms}.$ Springer showed that $l(z)\geq a(z)$ (see [L2]). Let $\delta(z)$ be the degree of $P_{e,z}$, where $e$ is the neutral element of $W$. Then actually one has $l(z)-a(z)-2\delta(z)\geq 0$ (see [L2]). Set $\mathcal{D}=\\{z\in W\ |\ l(z)-a(z)-2\delta(z)=0\\}.$ The elements of $\mathcal{D}$ are involutions, called distinguished involutions of $(W,S)$ (see [L2]). The following properties are due to Lusztig [L2] except the (j) (which is trivial) and (k) (proved in [X2]). (e) $\gamma_{x,y,z}\neq 0\Longrightarrow x\underset{L}{\sim}y^{-1},\ y\underset{L}{\sim}z,\ x\underset{R}{\sim}z.$ (f) $x\underset{L}{\sim}y^{-1}$ if and only if $\gamma_{x,y,z}\neq 0$ for some $z\in W$. (g) $\gamma_{x,y,z}=\gamma_{y,z^{-1},x^{-1}}=\gamma_{z^{-1},x,y^{-1}}$. (h) $\gamma_{x,d,x}=\gamma_{d,x^{-1},x^{-1}}=\gamma_{x^{-1},x,d}=1$ if $x\underset{L}{\sim}d$ and $d$ is a distinguished involution. (i) $\gamma_{x,y,z}=\gamma_{y^{-1},x^{-1},z^{-1}}.$ (j) If $\omega,\tau\in W$ has length 0, then $\gamma_{\omega x,y\tau,\omega z\tau}=\gamma_{x,y,z},\ \ \gamma_{x\omega,\tau y,z}=\gamma_{x,\omega\tau y,z}.$ (k) Let $x,y,z\in W$ be such that (1) $x$ is in a left string with respect to $*=\\{s,t\\}$ and also in a right string with respect to $\\#=\\{s^{\prime},t^{\prime}\\}$, (2) $y$ is in a left string with respect to $\\#=\\{s^{\prime},t^{\prime}\\}$ and also in a right string with respect to $\star=\\{s^{\prime\prime},t^{\prime\prime}\\}$, (3) $z$ is in a left string with respect to $*=\\{s,t\\}$ and also in a right string with respect to $\star=\\{s^{\prime\prime},t^{\prime\prime}\\}$. Then $\gamma_{x,y,z}=\gamma_{{}^{*}x^{\\#},{}^{\\#}y^{\star},{}^{*}z^{\star}}.$ For $w\in W$, set $\tilde{T}_{w}=q^{-l(w)/2}T_{w}$. For $x,y\in W$, write $\tilde{T}_{x}\tilde{T}_{y}=\sum_{z\in W}f_{x,y,z}\tilde{T}_{z},\qquad f_{x,y,z}\in\mathcal{A}=\mathbb{Z}[q^{\frac{1}{2}},q^{-\frac{1}{2}}].$ (l) If $x,y,w$ are in a two-sided cell of $W$, $f_{x,y,w}=\lambda q^{\frac{a(w)}{2}}+$ lower degree terms and as Laurent polynomials in $q^{\frac{1}{2}}$, deg$f_{x,y,z}\leq a(w)$ for all $z\in W$, then $\gamma_{x,y,w}=\lambda.$ (m) Each left cell (resp. each right cell) of $W$ contains a unique distinguished involution. (n) Each two-sided cell of $W$ contains only finitely many left cells. (o) Let $I$ be a subset of $S$ such that the subgroup $W_{I}$ of $W$ generated by $I$ is finite. Then the longest element $w_{I}$ is a distinguished involution. Let $d$ be a distinguished involution in $W$. (p) For any $\omega\in\Omega$, the element $\omega d\omega^{-1}$ is a distinguished involution. (q) Suppose $s,t\in S$ and $st$ has order 3. Then $d\in D_{L}(s,t)$ if and only if $d\in D_{R}(s,t)$. If $d\in D_{L}(s,t)$, then ${}^{*}d^{*}$ is a distinguished involution. 1.5. Assume $s,t\in S$ and $st$ has order 4. Let $w,u,v$ be in $W$ such that $l(ststw)=l(w)+4$ and $l(ststv)=l(v)+4$. We have (see [X2, 1.6.3]) * (a) $\gamma_{tsw,u,tv}=\gamma_{sw,u,stv},$ * (b) $\gamma_{tsw,u,tsv}=\gamma_{sw,u,sv}+\gamma_{sw,u,stsv},$ * (c) $\gamma_{tsw,u,tstv}=\gamma_{sw,u,stv},$ * (d) $\gamma_{tstw,u,tv}+\gamma_{tw,u,tv}=\gamma_{stw,u,stv},$ * (e) $\gamma_{tstw,u,tsv}=\gamma_{stw,u,stsv},$ * (f) $\gamma_{tstw,u,tstv}+\gamma_{tw,u,tstv}=\gamma_{stw,u,stv}.$ Assume $s,t\in S$ and $st$ has order 4. Let $w,u,v$ be in $W$ such that $l(ustst)=l(u)+4$ and $l(vstst)=l(v)+4$. We have (loc.cit) * (a’) $\gamma_{w,ut,vst}=\gamma_{w,uts,vs},$ * (b’) $\gamma_{w,ust,vst}=\gamma_{w,us,vs}+\gamma_{w,usts,vs},$ * (c’) $\gamma_{w,utst,vst}=\gamma_{w,uts,vs},$ * (d’) $\gamma_{w,ut,vtst}+\gamma_{w,ut,vt}=\gamma_{w,uts,vts},$ * (e’) $\gamma_{w,ust,vtst}=\gamma_{w,usts,vts},$ * (f’) $\gamma_{w,utst,vtst}+\gamma_{w,utst,vt}=\gamma_{w,uts,vts}.$ 1.6. The based ring of a two-sided cell For each two-sided cell $c$ of $W$, let $J_{c}$ be the free $\mathbb{Z}$-module with a basis $t_{w},\ w\in c$. Define $t_{x}t_{y}=\sum_{z\in c}\gamma_{x,y,z}t_{z}.$ Then $J_{c}$ is an associative ring with unit $\sum_{d\in\mathcal{D}\cap c}t_{d}.$ The ring $J=\bigoplus_{c}J_{c}$ is a ring with unit $\sum_{d\in\mathcal{D}}t_{d}$. Sometimes $J$ is called asymptotic Hecke algebra since Lusztig established an injective $\mathcal{A}$-algebra homomorphism $\phi:H\to J\otimes\mathcal{A},\quad C_{x}\mapsto\sum_{\stackrel{{\scriptstyle\stackrel{{\scriptstyle d\in\mathcal{D}}}{{w\in W}}}}{{w\underset{L}{\sim}d}}}h_{x,d,w}t_{w}.$ 1.7. Lusztig’s conjecture on the structure of $J_{c}$ In [L3] Lusztig states a conjecture on $J_{c}$ using equivariant $K$-groups on finite sets. Let $G$ be a connected reductive group over $\mathbb{C}$. Lusztig establishes a bijection between the two-sided cells of the extended affine Weyl group $W$ and the unipotent classes of $G$. For each two-sided cell $c$ of $W$, let $u$ be a unipotent element in the unipotent class corresponding to $c$ and let $F_{c}$ be a maximal reductive subgroup of the centralizer of $u$ in $G$. Conjecture (Lusztig [L3]): Assume that $G$ is a simply connected simple algebraic group over $\mathbb{C}$. Then there exists a finite set $Y$ with an algebraic action of $F_{c}$ and a bijection $\pi:c\to\text{{the set of} isomorphism classes of irreducible $F_{c}$-vector bundles on}\ Y\times Y.$ such that (i) The bijection $\pi$ induces a ring homomorphism $\pi:J_{c}\to K_{F_{c}}(Y\times Y),\ \ t_{x}\mapsto\pi(x).$ (ii) $\pi(x^{-1})_{(a,b)}=\pi(x)_{(b,a)}^{*}$ is the dual representation of $\pi(x)_{(b,a)}.$ ## 2\. Cells in an extended affine Weyl group of type $\tilde{B}_{3}$ In this section $G=Sp_{6}(\mathbb{C})$, so that the extended affine Weyl group $W$ attached to $G$ is of type $\tilde{B}_{3}$. The left cells and two-sided cells are described by J. Du (see [D]). We recall his results. 2.1. The Coxeter graph of $W$. As usual, we number the 4 simple reflections $s_{0},\ s_{1},\ s_{2},\ s_{3}$ in $W$ so that $\displaystyle s_{0}s_{1}=s_{1}s_{0},\quad s_{0}s_{3}=s_{3}s_{0},\quad s_{1}s_{3}=s_{3}s_{1},$ $\displaystyle(s_{0}s_{2})^{3}=(s_{1}s_{2})^{3}=e,\quad(s_{2}s_{3})^{4}=e,$ where $e$ is the neutral element in $W$. The relations among the simple reflections can be read through the following Coxeter graph: $\tilde{B}_{3}:$$2$$0$$1$$3$ There is a unique nontrivial element $\tau$ in $W$ with length 0. We have $\tau^{2}=e,\ \tau s_{0}\tau=s_{1},\ \tau s_{i}\tau=s_{i}$ for $i=2,3.$ Note that $s_{1},s_{2},s_{3}$ generate the Weyl group $W_{0}$ of type $B_{3}$ and $s_{0},s_{1},s_{2},s_{3}$ generate an affine Weyl group $W^{\prime}$ of type $\tilde{B}_{3}$. And $W$ is generated by $\tau,\ s_{0},s_{1},s_{2},s_{3}$. 2.2. Cells in $W$ According to [D], the extended affine Weyl group $W$ attached to $Sp_{6}(\mathbb{C})$ has 8 two-sided cells: $A,\quad B,\quad C,\quad D,\quad E,\quad F,\quad G,\quad H.$ The following table displays some useful information on these two-sided cells. | | Number | Size of Jordan blocks | Maximal reductive subgroup ---|---|---|---|--- | | of left | of the corresponding | of the centralizer of a unipotent $X$ | ${a(X)}$ | cells in $X$ | unipotent class in $Sp_{6}(\mathbb{C})$ | element in the corresp. unipotent class $A$ | 9 | 48 | (111111) | $Sp_{6}(\mathbb{C})$ $B$ | 6 | 24 | (21111) | $Sp_{4}(\mathbb{C})\times\mathbb{Z}/2\mathbb{Z}$ $C$ | 4 | 18 | (2211) | $SL_{2}(\mathbb{C})\times O_{2}(\mathbb{C})$ $D$ | 3 | 12 | (222) | $O_{3}(\mathbb{C})$ $E$ | 2 | 8 | (411) | $SL_{2}(\mathbb{C})\times\mathbb{Z}/2\mathbb{Z}$ $F$ | 2 | 6 | (33) | $SL_{2}(\mathbb{C})$ $G$ | 1 | 4 | (42) | $\mathbb{Z}/2\mathbb{Z}\times\mathbb{Z}/2\mathbb{Z}$ $H$ | 0 | 1 | (6) | $\mathbb{Z}/2\mathbb{Z}$ The notations for two-sided cells in the table are the same as those in [D], which will be replaced by other notations in subsequent sections, otherwise confusion would happen since notations $C,F,G$ are already used for other objects. In subsequent sections, for a reduced expression $s_{i_{1}}s_{i_{2}}\cdots s_{ik}$ of an element in $W$, we often write $i_{1}i_{2}\cdots i_{k}$ instead of the reduced expression. In the rest of the paper, $W$ always stands for the affine Weyl group attached to $Sp_{6}(\mathbb{C})$, $\tau,\ s_{i}$ are as in Subsection 2.1, and all representations in this paper are rational representations of algebraic groups. ## 3\. The based ring of the two-sided cell containing $s_{0}s_{1}$ 3.1. In this section $c$ stands for the two-sided cell of $W$ containing $s_{0}s_{1}$. According to [D, Figure I, Theorem 6.4], $c$ has six left cells. We list the six left cells and representative elements in the left cells given in [D, figure I]: $\Gamma_{1},\ 01;\quad\Gamma_{2},\ 012;\quad\Gamma_{3},\ 0123;\quad\Gamma_{4},\ 01232;\quad\Gamma_{5},\ 012321;\quad\Gamma_{6},\ 012320.$ The values of $a$-function on $c$ is 2. The corresponding unipotent class in $Sp_{6}(\mathbb{C})$ has Jordan block sizes (411). Maximal reductive subgroup of the centralizer of an element in the unipotent class is $F_{c}=\mathbb{Z}/2\mathbb{Z}\times SL_{2}(\mathbb{C})$. Let $\epsilon$ be the nontrivial one dimensional representation of $\mathbb{Z}/2\mathbb{Z}$ and $V(k)$ be an irreducible representation of $SL_{2}(\mathbb{C})$ with highest weight $k$. They can be regarded as irreducible representations of $F_{c}$ naturally. Up to isomorphism, the irreducible representations of $F_{c}$ are $V(k),\ \epsilon\otimes V(k),\ k=0,1,2,3,...$. We will denote $\epsilon\otimes V(k)$ by $\epsilon V(k)$. Let $x_{k}=(s_{0}s_{1}s_{2}s_{3}s_{2})^{k}s_{1}s_{0},\ u_{1}=e,\ u_{2}=s_{2},\ u_{3}=s_{3}s_{2},\ u_{4}=s_{2}s_{3}s_{2},\ u_{5}=s_{1}s_{2}s_{3}s_{2},\ u_{6}=s_{0}s_{2}s_{3}s_{3}.$ According to [D, Theorem 6.4], we have (a) $c=\\{u_{i}x_{k}u_{j}^{-1},\ u_{i}\tau x_{k}u_{j}^{-1}\,|\,1\leq i,j\leq 6,\ k=0,1,2,3,...\\}.$ (b) $\Gamma_{j}=\\{u_{i}x_{k}u_{j}^{-1},\ u_{i}\tau x_{k}u_{j}^{-1}\,|\,1\leq i\leq 6,\ k=0,1,2,3,...\\},\ j=1,2,3,4,5,6.$ Let $Y=\\{1,\ 2,\ ...,\ 6\\}$ and let $F_{c}$ act on $Y$ trivially. Then $K_{F_{c}}(Y\times Y)$ is isomorphic to the $6\times 6$ matix ring $M_{6}(\text{Rep\,}F_{c})$, where $\text{Rep\,}F_{c}$ is the representation ring of $F_{c}$. Recall that $F_{c}=\mathbb{Z}/2\mathbb{Z}\times SL_{2}(\mathbb{C})$ in this section. The main result in this section is the following theorem. Theorem 3.2. Let $c$ be the two-sided cell of $W$ (the affine Weyl group $W$ attached to $Sp_{6}(\mathbb{C})$) containing $s_{0}s_{1}$. Then the map $\pi:c\to M_{6}(\text{Rep\,}F_{c}),\quad u_{i}x_{k}u_{j}^{-1}\mapsto V(k)_{ij},\ u_{i}\tau x_{k}u_{j}^{-1}\mapsto\epsilon V(k)_{ij}\ $ induces a ring isomorphism $\pi:J_{c}\to M_{6}(\text{Rep\,}F_{c}),\quad t_{u_{i}x_{k}u_{j}^{-1}}\mapsto V(k)_{ij},\quad t_{u_{i}\tau x_{k}u_{j}^{-1}}\mapsto\epsilon V(k)_{ij},$ where $V(k)_{ij}$ (resp. $\epsilon V(k)_{ij})$ is the matrix in $M_{6}(\text{Rep\,}F_{c})$ whose entry at $(p,q)$ is $V(k)$ (resp. $\epsilon V(k)$) if $(p,q)=(i,j)$ and is 0 otherwise. Remark: The Theorem 4 in [BO] implies that Lusztig’s conjecture on the structure of $J_{c}$ is true. Since under the isomorphism $K_{F_{c}}(Y\times Y)\simeq M_{6}(\text{Rep\,}F)$, irreducible ${F_{c}}$-vector bundles on $Y\times Y$ correspond to those $V(k)_{ij},\ \epsilon V(k)_{ij}$, hence Theorem 3.2 provides a computable verification for Lusztig conjecture on the structure of $J_{c}$. We prove Theorem 3.2 by establishing three lemmas. Lemma 3.3. Let $1\leq i,j,m,n\leq 6$ and $k,l$ be nonnegative integers. For $z_{k}=x_{k}$ or $\tau x_{k}$, $z_{l}=x_{l}$ or $\tau x_{l}$, and $z_{p}=x_{p}$ or $\tau x_{p}$, we have (a) $\gamma_{u_{i}z_{k}u_{j}^{-1},u_{m}z_{l}u_{n}^{-1},z}=0\quad\text{if}\ j\neq m\ \text{or}\ z\neq u_{i}\tau^{a}z_{p}u^{-1}_{n},\ a=0,1,\ \text{for some }p;$ (b) $\gamma_{u_{i}z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},u_{i}z_{p}u_{n}^{-1}}=\gamma_{z_{k},z_{l},z_{p}},\quad\text{for any nonnegative integer }p.$ Proof. Note that $z_{l}^{-1}=z_{l}$. If $\gamma_{u_{i}z_{k}u_{j}^{-1},u_{m}z_{l}u_{n}^{-1},z}\neq 0$, then by 1.4(e) we get $u_{i}z_{k}u_{j}^{-1}\underset{L}{\sim}(u_{m}z_{l}u_{n}^{-1})^{-1}=u_{n}z_{l}u_{m}^{-1}$, $u_{i}z_{k}u_{j}^{-1}\underset{R}{\sim}z,\ u_{m}z_{l}u_{n}^{-1}\underset{L}{\sim}z$. By (b) in Subsection 3.1 we see that the first assertion is true. Now we prove the second assertion. Let $*=\\{s_{1},s_{2}\\},\ \\#=\\{s_{2},s_{3}\\}$ and $\star=\\{s_{0},s_{2}\\}$. Then (c) $\Gamma_{2}=\Gamma_{1}^{*},\quad\Gamma_{4}=\Gamma_{2}^{\\#},\quad\Gamma_{5}=\Gamma_{4}^{*},\quad\Gamma_{6}=\Gamma_{4}^{\star}.$ Applying 1.4 (k) we see that (b) is true if none of $i,j,n$ is 3. Now assume that $i=3$. By 1.5 (b) we get $\gamma_{u_{3}z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},u_{3}z_{p}u_{n}^{-1}}=\gamma_{u_{2}z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},u_{2}z_{p}u_{n}^{-1}}+\gamma_{u_{2}z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},u_{4}z_{p}u_{n}^{-1}}.$ By Part (a) of the lemma, we have $\gamma_{u_{2}z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},u_{4}z_{p}u_{n}^{-1}}=0.$ Then using (c) above and 1.4 (k) we get $\gamma_{u_{3}z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},u_{3}z_{p}u_{n}^{-1}}=\gamma_{u_{2}z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},u_{2}z_{p}u_{n}^{-1}}=\gamma_{z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},z_{p}u_{n}^{-1}}.$ Similarly, if $n=3$, we have $\gamma_{u_{i}z_{k}u_{j}^{-1},u_{j}z_{l}u_{3}^{-1},u_{i}z_{p}u_{3}^{-1}}=\gamma_{u_{i}z_{k}u_{j}^{-1},u_{j}z_{l}u_{2}^{-1},u_{i}z_{p}u_{2}^{-1}}=\gamma_{u_{i}z_{k}u_{j}^{-1},u_{j}z_{l},u_{i}z_{p}}.$ We have showed for any $1\leq i,n\leq 6$ the following identity holds: $\gamma_{u_{i}z_{k}u_{j}^{-1},u_{j}z_{l}u_{n}^{-1},u_{i}z_{p}u_{n}^{-1}}=\gamma_{z_{k}u_{j}^{-1},u_{j}z_{l},z_{p}}.$ Note that $z_{k}^{-1}=z_{k}$. By above identity and 1.4 (g), we get $\gamma_{z_{k}u_{j}^{-1},u_{j}z_{l},z_{p}}=\gamma_{u_{j}z_{l},z_{p}^{-1},u_{j}z_{k}^{-1}}=\gamma_{z_{l},z_{p},z_{k}}=\gamma_{z_{k},z_{l},z_{p}}.$ Assertion (b) is proved and the lemma is proved.∎ Lemma 3.4. For nonnegative integers, and $a,b=0,1$, we have $\gamma_{\tau^{a}x_{k},\tau^{b}x_{l},\tau^{a+b}x_{p}}=\gamma_{x_{k},x_{l},x_{p}},\quad\gamma_{\tau^{a}x_{k},\tau^{b}x_{l},\tau^{c}x_{p}}=0\ \text{if }\tau^{c}\neq\tau^{a+b}.$ Proof. The assertion follows from 1.4 (j).∎ Lemma 3.5. For nonnegative integers $k,l$ we have $t_{x_{k}}t_{x_{l}}=\sum_{0\leq p\leq\min\\{k,l\\}}t_{{x_{k+l-2i}}}.$ Proof. If $k=0$ or $l=0$, the identity above is trivial since $x_{0}$ is a distinguished involution. Now assume that $k=1$ and $l\geq 1$. Let $\zeta=q^{\frac{1}{2}}-q^{-\frac{1}{2}}$. By a simple computation we see $\tilde{T}_{x_{1}}\tilde{T}_{x_{l}}=\zeta^{2}(\tilde{T}_{x_{l+1}}+\tilde{T}_{x_{l-1}}+\tilde{T}_{s_{0}s_{1}s_{3}(s_{2}s_{3}s_{2}s_{0}s_{1})^{l}}+\tilde{T}_{s_{0}s_{2}s_{3}s_{2}s_{1}(s_{2}s_{3}s_{2}s_{0}s_{1})^{l}})+\text{lower degree terms,}$ Since $a(s_{0}s_{1}s_{3}(s_{2}s_{3}s_{2}s_{0}s_{1})^{l})\geq a(s_{0}s_{1}s_{3})=3,\ a(s_{0}s_{2}s_{3}s_{2}s_{1}(s_{2}s_{3}s_{2}s_{0}s_{1})^{l})\geq a(s_{2}s_{1}s_{2})=3$, we see that $s_{0}s_{1}s_{3}(s_{2}s_{3}s_{2}s_{0}s_{1})^{l}$ and $s_{0}s_{2}s_{3}s_{2}s_{1}(s_{2}s_{3}s_{2}s_{0}s_{1})^{l}$ are not in the two- sided cell $c$. By 1.4 (l), we have $t_{x_{1}}t_{x_{l}}=t_{x_{l+1}}+t_{x_{l-1}}.$ For $k\geq 2$, since $t_{x_{k}}=t_{x_{1}}t_{x_{k-1}}-t_{x_{k-2}}$, we can use induction on $k$ to prove the lemma. The argument is completed.∎ Proof of Theorem 3.2. Combining Lemmas 3.3, 3.4 and 3.5 we see that Theorem 3.2 is true. ## 4\. The based ring of the two-sided cell containing $s_{1}s_{3}$ 4.1. In this section $c$ stands for the two-sided cell of $W$ containing $s_{1}s_{3}$. According to [D, Figure I, Theorem 6.4], $c$ has eight left cells. We list the eight left cells and representative elements in the left cells given in [D, Figure I]: $\begin{array}[]{lllllllll}&\Gamma_{1},&13;&\Gamma_{2},&132;&\Gamma_{3},&1323;&\Gamma_{4},&1320;\\\ &\Gamma_{5},&03;&\Gamma_{6},&032;&\Gamma_{7},&0323;&\Gamma_{8},&0321.\end{array}$ The values of $a$-function on $c$ is 2. The corresponding unipotent class in $Sp_{6}(\mathbb{C})$ has Jordan block sizes (33). Maximal reductive subgroup of the centralizer of an element in the unipotent class is ${F_{c}}=SL_{2}(\mathbb{C})$. Let $V(k)$ be an irreducible representation of ${F_{c}}=SL_{2}(\mathbb{C})$ with highest weight $k$. Up to isomorphism, the irreducible representations of ${F_{c}}$ are $V(k),\ k=0,1,2,3,...$. Let $x_{k}=(\tau s_{0}s_{3}s_{2})^{k}s_{1}s_{3},\ u_{1}=e,\ u_{2}=s_{2},\ u_{3}=s_{3}s_{2},\ u_{4}=s_{0}s_{2},\ u_{5}=\tau,\ u_{6}=\tau s_{2},\ u_{7}=\tau s_{3}s_{2},\ u_{8}=\tau s_{0}s_{2}.$ According to [D, Theorem 6.4], we have (a) $c=\\{u_{i}x_{k}u_{j}^{-1}\,|\,1\leq i,j\leq 8,\ k=0,1,2,3,...\\}.$ (b) $\Gamma_{j}=\\{u_{i}x_{k}u_{j}^{-1}\,|\,1\leq i\leq 8,\ k=0,1,2,3,...\\},\ j=1,2,3,4,5,6,7,8.$ Let $Y=\\{1,\ 2,\ ...,\ 7,\ 8\\}$ and let ${F_{c}}$ act on $Y$ trivially. Then $K_{F_{c}}(Y\times Y)$ is isomorphic to the $8\times 8$ matrix ring $M_{8}(\text{Rep\,}{F_{c}})$, where $\text{Rep\,}{F_{c}}$ is the representation ring of ${F_{c}}=SL_{2}(\mathbb{C})$. The main result in this section is the following. Theorem 4.2. Let $c$ be the two-sided cell of $W$ (the extended affine Weyl group attached to $Sp_{6}(\mathbb{C})$) containing $s_{1}s_{3}$. Then the map $\pi:c\to M_{8}(\text{Rep\,}F_{c}),\quad u_{i}x_{k}u_{j}^{-1}\mapsto V(k)_{ij}$ induces a ring isomorphism $\pi:J_{c}\to M_{8}(\text{Rep\,}F_{c}),\quad t_{u_{i}x_{k}u_{j}^{-1}}\mapsto V(k)_{ij},$ where $V(k)_{ij}$ is the matrix in $M_{8}(\text{Rep\,}F_{c})$ whose entry at $(p,q)$ is $V(k)$ if $(p,q)=(i,j)$ and is 0 otherwise. Remark: The Theorem 4 in [BO] implies that Lusztig’s conjecture on the structure of $J_{c}$ is true. Since under the isomorphism $K_{F_{c}}(Y\times Y)\simeq M_{8}(\text{Rep\,}F_{c})$, irreducible ${F_{c}}$-vector bundles on $Y\times Y$ correspond to $V(k)_{ij}$’s, Theorem 4.2 provides a computable verification for Lusztig’s conjecture on the structure of $J_{c}$. We prove Theorem 4.2 by establishing two lemmas. Lemma 4.3. Let $1\leq i,j,m,n\leq 8$ and $k,l$ be nonnegative integers. Then (a) $\gamma_{u_{i}x_{k}u_{j}^{-1},u_{m}x_{l}u_{n}^{-1},z}=0\quad\text{if}\ j\neq m\ \text{or}\ x\neq u_{i}x_{p}u^{-1}_{n}\ \text{for some }p;$ (b) $\gamma_{u_{i}x_{k}u_{j}^{-1},u_{j}x_{l}u_{n}^{-1},u_{i}x_{p}u_{n}^{-1}}=\gamma_{x_{k},x_{l},x_{p}},\quad\text{for any nonnegative integer }p.$ Proof. Note that $x_{l}^{-1}=x_{l}$. If $\gamma_{u_{i}x_{k}u_{j}^{-1},u_{m}x_{l}u_{n}^{-1},z}\neq 0$, then by 1.4(e) we get $u_{i}x_{k}u_{j}^{-1}\underset{L}{\sim}(u_{m}x_{l}u_{n}^{-1})^{-1}=u_{n}x_{l}u_{m}^{-1}$, $u_{i}x_{k}u_{j}^{-1}\underset{R}{\sim}z,$ and $u_{m}x_{l}u_{n}^{-1}\underset{L}{\sim}z$. By (b) in Subsection 4.1 we see that the first assertion is true. Now we prove the second assertion. Let $*=\\{s_{1},s_{2}\\},\ \\#=\\{s_{2},s_{3}\\}$ and $\star=\\{s_{0},s_{2}\\}$. Then (c) $\Gamma_{2}=\Gamma_{1}^{*},\quad\Gamma_{3}=\Gamma_{1}^{\\#},\quad\Gamma_{4}=\Gamma_{2}^{\star},\quad\Gamma_{5}=\tau\Gamma_{1}\tau,\ \Gamma_{6}=\tau\Gamma_{2}\tau,\ \Gamma_{7}=\tau\Gamma_{3}\tau,\ \Gamma_{8}=\tau\Gamma_{4}\tau.$ Applying 1.4 (j) and 1.4 (k) (repeatedly if necessary) we get the following identity. $\gamma_{u_{i}x_{k}u_{j}^{-1},u_{j}x_{l}u_{n}^{-1},u_{i}x_{p}u_{n}^{-1}}=\gamma_{x_{k}u_{j}^{-1},u_{j}x_{l},x_{p}}.$ Note that $\tau^{2}=e$. Again using 1.4 (j) and 1.4 (k) (repeatedly if necessary) we get the following identity. $\gamma_{x_{k}u_{j}^{-1},u_{j}x_{l},x_{p}}=\gamma_{x_{k},x_{l},x_{p}}.$ Part (b) is proved and the lemma is proved. ∎ Lemma 4.4. For nonnegative integers $k,l$, we have $t_{x_{k}}t_{x_{l}}=\sum_{0\leq p\leq\min\\{k,l\\}}t_{x_{k+l-2i}}.$ Proof. If $k=0$ or $l=0$, the identity above is trivial since $x_{0}$ is a distinguished involution. Now assume that $k=1$ and $l\geq 1$. Put $\xi=q^{\frac{1}{2}}+q^{-\frac{1}{2}}$. Let $H^{<13}$ be the $\mathcal{A}$-submodule of $H$ spanned by all $C_{w}$ with $a(w)\geq 3$. By Subsection 1.2 and 1.4 (b) we know that $H^{<13}$ is a two-sided ideal of $H$. Before continuing, we make a convention: we shall use the symbol $\Box$ for any element in the two-sided ideal $H^{<13}$ of $H$. Then $\Box+\Box=\Box$ and $h\Box=\Box$ for any $h\in H$. First we have $C_{x_{1}}=C_{\tau s_{3}s_{0}s_{2}s_{1}s_{3}}=C_{\tau s_{3}}C_{s_{0}}C_{s_{2}}C_{s_{1}}C_{s_{3}}-C_{\tau s_{0}s_{1}s_{3}},\ \ \text{and}\ C_{\tau s_{0}s_{1}s_{3}}\in H^{<13}.$ Note that ${L}(x_{k})=\\{s_{1},s_{3}\\}$. Hence (4) $\displaystyle C_{x_{1}}C_{x_{l}}=C_{\tau s_{3}}C_{s_{0}}C_{s_{2}}C_{s_{1}}C_{s_{3}}C_{x_{l}}+\Box=\xi^{2}C_{\tau s_{3}}C_{s_{0}}C_{s_{2}}C_{x_{l}}+\Box\in\xi^{2}C_{\tau s_{3}}C_{s_{0}}C_{s_{2}}C_{x_{l}}+H^{<13}.$ We compute $C_{\tau s_{3}}C_{s_{0}}C_{s_{2}}C_{x_{l}}$ step by step. Step 1. Compute $C_{s_{2}}C_{x_{l}}.$ We have $C_{s_{2}}C_{x_{l}}=C_{s_{2}x_{l}}+\sum\limits_{\begin{subarray}{c}z\prec x_{l}\\\ s_{2}z<z\end{subarray}}\mu(z,x_{l})C_{z}.$ Note that $L(x_{l})=\\{s_{1},s_{3}\\}$. Assume that $z\prec x_{l}$ and $s_{2}z\leq z$. If $s_{1}z\leq z$, then $\\{s_{1},s_{2}\\}\subset L(z)$ and $a(z)\geq a(s_{1}s_{2}s_{1})=3$. In this case, we have $C_{z}\in H^{<13}$. If $s_{1}z\geq z$, by 1.1(c) we must have $s_{1}z=x_{l}$. Then $z=\tau s_{3}s_{2}x_{l-1}$. This contradicts $s_{2}z\leq z$. Therefore we have (5) $\displaystyle C_{s_{2}}C_{x_{l}}=C_{s_{2}x_{l}}+\Box\in C_{s_{2}x_{l}}+H^{<13}.$ Step 2. Compute $C_{s_{0}}C_{s_{2}x_{l}}.$ We have $C_{s_{0}}C_{s_{2}x_{l}}=C_{s_{0}s_{2}x_{l}}+\sum\limits_{\begin{subarray}{c}z\prec s_{2}x_{l}\\\ s_{0}z<z\end{subarray}}\mu(z,s_{2}x_{l})C_{z}.$ Note that $L(s_{2}x_{l})=\\{s_{2}\\}$. Assume that $z\prec s_{2}x_{l}$ and $s_{0}z\leq z$. If $s_{2}z\leq z$, then $\\{s_{0},s_{2}\\}\subset{L}(z)$ and $a(z)\geq a(s_{0}s_{2}s_{0})=3$. In this case, we have $C_{z}\in H^{<13}$. If $s_{2}z\geq z$, by 1.1(c) we must have $z=x_{l}$. This contradicts $s_{0}z\leq z$. Therefore we have (6) $\displaystyle C_{s_{0}}C_{s_{2}x_{l}}=C_{s_{0}s_{2}x_{l}}+\Box\in C_{s_{0}s_{2}x_{l}}+H^{<13}.$ Step 3. Compute $C_{\tau s_{3}}C_{s_{0}s_{2}x_{l}}.$ We have $C_{\tau s_{3}}C_{s_{0}s_{2}x_{l}}=C_{x_{l+1}}+\sum\limits_{\begin{subarray}{c}z\prec s_{0}s_{2}x_{l}\\\ s_{3}z<z\end{subarray}}\mu(z,s_{0}s_{2}x_{l})C_{\tau z}.$ Assume that $z\prec s_{0}s_{2}x_{l}$ and $s_{3}z\leq z$. Using 1.4 (b), 1.4 (c) and 1.4 (d), we see that $a(\tau z)\geq 2$ , and if $a(\tau z)=2$ then $x_{l}\underset{L}{\sim}\tau z\underset{R}{\sim}x_{1}$. We are only concerned with those $C_{\tau z}$ in above summation with $a(\tau z)=2$. Then $\tau z=x_{m}$ for some $m<l$ and $L(z)=\\{s_{0},s_{3}\\}$. Note that $L(s_{0}s_{2}x_{l})=\\{s_{0}\\}$. We then have $\mu(z,s_{0}s_{2}x_{l})={\tilde{\mu}({}^{\star}z,{}^{\star}(s_{0}s_{2}x_{l}))}=\tilde{\mu}(s_{2}z,s_{2}x_{l})=\tilde{\mu}(^{*}(s_{2}z),^{*}(s_{2}x_{l}))=\tilde{\mu}(s_{1}s_{2}z,x_{l})$, where $\star=\\{s_{0},s_{2}\\},\ *=\\{s_{1},s_{2}\\}$. Since $m<l$, we have $\tilde{\mu}(s_{1}s_{2}z,x_{l})=\mu(s_{1}s_{2}z,x_{l})$. Noting that $s_{3}s_{1}s_{2}z=s_{3}s_{1}s_{2}\tau x_{m}\geq s_{1}s_{2}\tau x_{m}$ and $s_{3}x_{l}\leq x_{l}$, by 1.1(c) we see $s_{3}s_{1}s_{2}z=x_{l}$, which implies that $\tau z=x_{l-1}$. In conclusion, if $z\prec s_{0}s_{2}x_{l}$ and $s_{3}z\leq z$, then either $C_{z}\in H^{<13}$ or $z=\tau x_{l-1}$. Hence we have (7) $\displaystyle C_{\tau s_{3}}C_{s_{0}s_{2}x_{l}}=C_{x_{l+1}}+C_{x_{l-1}}+\Box.$ Combining formulas (4)-(7) we get $C_{x_{1}}C_{x_{l}}=\xi^{2}(C_{x_{l+1}}+C_{x_{l-1}})+\Box\in\xi^{2}(C_{x_{l+1}}+C_{x_{l-1}})+H^{<13}.$ Therefore we have $t_{x_{1}}t_{x_{l}}=t_{x_{l+1}}+t_{x_{l-1}}.$ For $k\geq 2$, since $t_{x_{k}}=t_{x_{1}}t_{x_{k-1}}-t_{x_{k-2}}$, we can use induction on $k$ to prove the lemma. The argument is completed. ∎ Proof of Theorem 4.2. Combining Lemmas 4.3 and 4.4 we see that Theorem 4.2 is true. ## 5\. The based ring of the two-sided cell containing $s_{1}s_{2}s_{1}$ 5.1. In this section we consider the two-sided cell in $W$ containing $s_{1}s_{2}s_{1}$. In [QX] we showed that Lusztig’s conjecture on the structure of the based ring of the two-sided cell needs modification. In this section we give a description of the based ring. For consistence, we keep the notations in [QY] for the two-sided cell of $W$ containing $s_{1}s_{2}s_{1}$. In particular, we denote $D$ for the two-sided cell of $W$ containing $s_{1}s_{2}s_{1}$. According to [D,Figure I, Theorem 6.4], we have the following result. (a) There are 12 left cells in the two-sided cell $D$ and a representative of each left cell in $D$ are: $\begin{array}[]{lllllllll}&D_{013},&013;&D_{2},&0132;&D_{02},&01320;&D_{12},&01321;\\\ &&&&&&&&\\\ &D_{3},&01323;&D_{03},&013203;&D_{01},&013201;&D_{13},&013213;\\\ &&&&&&&&\\\ &D^{\prime}_{2},&0132032;&\widehat{D^{\prime}_{2}},&0132132;&D_{1},&01320321;&D_{0},&01321320.\end{array}$ The value of $a$-function on $D$ is 3. Let $\Gamma$ and $\Gamma^{\prime}$ be two left cells of $W$. If $\Gamma^{\prime}=\Gamma^{*}$ for some $*=\\{s,t\\}$ (see Subsection 1.3 for definition of $*$-operation), then we write $\Gamma\ \overset{\\{s,t\\}}{\text{------}}\ \Gamma^{\prime}$. The following result is easy to verify. Lemma 5.2. Keep the notations as above. Then we have $D_{3}\ \overset{\\{s_{2},s_{3}\\}}{\text{------}}\ D_{013}\ \overset{\\{s_{1},s_{2}\\}}{\text{------}}\ D_{2}.$ $\displaystyle D_{0}\ \overset{\\{s_{0},s_{2}\\}}{\text{------}}$ $\displaystyle\widehat{D^{\prime}_{2}}\ \overset{\\{s_{2},s_{3}\\}}{\text{------}}\ D_{12}\ \overset{\\{s_{0},s_{2}\\}}{\text{------}}\ D_{01}\ \overset{\\{s_{1},s_{2}\\}}{\text{------}}\ D_{02}\ \overset{\\{s_{2},s_{3}\\}}{\text{------}}\ D^{\prime}_{2}\ \overset{\\{s_{1},s_{2}\\}}{\text{------}}\ D_{1}.$ $\displaystyle\ |{\scriptstyle{\\{s_{1},s_{2}\\}}}\qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad\ |{\scriptstyle{\\{s_{0},s_{2}\\}}}$ $\displaystyle D_{13}\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\ D_{03}$ 5.3. Let $\begin{array}[]{lllllll}u_{k}=(s_{0}s_{1}s_{3}s_{2})^{k}s_{0}s_{1}s_{3},&&\\\ x_{k}=(s_{1}s_{2}s_{3}s_{0})^{k}s_{1}s_{2}s_{1},&\ x^{\prime}_{0}=\tau s_{2}s_{0}s_{1}s_{2}s_{1},&\ x^{\prime}_{k+1}=\tau s_{0}s_{2}s_{3}s_{0}x_{k},\\\ p_{1}=e,&p_{2}=s_{2},&p_{3}=s_{3}s_{2},\\\ p_{4}=s_{1}s_{2},&p_{5}=s_{0}s_{2},&p_{6}=s_{0}s_{1}s_{2},\\\ p_{7}=s_{3}s_{1}s_{2},&p_{8}=s_{3}s_{0}s_{2},&p_{9}=s_{2}s_{3}s_{1}s_{2},\\\ p_{10}=s_{2}s_{3}s_{0}s_{2},&p_{11}=s_{0}s_{2}s_{3}s_{1}s_{2},&p_{12}=s_{1}s_{2}s_{3}s_{0}s_{2};\\\ q_{4}=p_{1}=e,&q_{5}=\tau,&q_{6}=s_{0},\\\ q_{7}=s_{3},&q_{8}=s_{3}\tau,&q_{9}=s_{2}s_{3},\\\ q_{10}=s_{2}s_{3}\tau,&q_{11}=s_{0}s_{2}s_{3},&q_{12}=s_{1}s_{2}s_{3}\tau.\end{array}$ According to [D, Theorem 6.4], we have (a) The two-sided cell $D$ consists of the following elements: $p_{i}u_{k}p_{j}^{-1},\ p_{i}\tau u_{k}p_{j}^{-1},\ q_{l}x_{0}q_{m}^{-1},\ q_{l}x_{0}q_{6}^{-1},\ q_{l}x_{0}q_{6}^{-1}\tau,\ q_{l}x^{\prime}_{0}q_{m}^{-1},$ where $1\leq i,j\leq 12,\ 4\leq l\leq 12,\ 4\leq m\neq 6\leq 12,\ k\geq 0.$ For convenience, we number the left cells in $D$ as follows: $\begin{array}[]{lllllll}\Gamma_{1}=D_{013},&\ \Gamma_{2}=D_{2},&\ \Gamma_{3}=D_{3},&\ \Gamma_{4}=D_{12},&\ \Gamma_{5}=D_{02},&\ \Gamma_{6}=D_{01},\\\ \Gamma_{7}=D_{13},&\ \Gamma_{8}=D_{03},&\ \Gamma_{9}=\widehat{D^{\prime}_{2}},&\ \Gamma_{10}=D^{\prime}_{2},&\ \Gamma_{11}=D_{0},&\ \Gamma_{12}=D_{1}.\end{array}$ Then (loc. cit) we have (b1) For $j=1,\ 2,\ 3$, the left cell $\Gamma_{j}$ consists of the following elements: $p_{i}u_{k}p_{j}^{-1},\quad p_{i}\tau u_{k}p_{j}^{-1},\qquad 1\leq i\leq 12,\ k\geq 0.$ (b2) For $j=4,\ 5,\ 7,\ 8,\ ...,\ 12$, the left cell $\Gamma_{j}$ consists of the following elements: $p_{i}u_{k}p_{j}^{-1},\quad p_{i}\tau u_{k}p_{j}^{-1},\quad q_{l}x_{0}q_{j}^{-1},\quad q_{l}x^{\prime}_{0}q_{j}^{-1},\qquad{1\leq i\leq 12},\ 4\leq l\leq 12,\ k\geq 0.$ Note that $p_{4}u_{k}p_{4}^{-1}=x_{k+1},\ p_{4}\tau u_{k}p_{4}^{-1}=x^{\prime}_{k+1}.$ (b3) The left cell $\Gamma_{6}$ consists of the following elements: $p_{i}u_{k}p_{6}^{-1},\quad p_{i}\tau u_{k}p_{6}^{-1},\quad q_{l}x_{0}q_{6}^{-1},\quad q_{l}x_{0}q_{6}^{-1}\tau,\qquad{1\leq i\leq 12},\ 4\leq l\leq 12,\ k\geq 0.$ 5.4. For the two-sided cell $D$, the corresponding unipotent class in $Sp_{6}(\mathbb{C})$ has Jordan block sizes (222). Maximal reductive subgroup of the centralizer of an element in the unipotent class is ${F_{c}}=O_{3}(\mathbb{C})=\mathbb{Z}/2\mathbb{Z}\times SO_{3}(\mathbb{C})$. Let $\tilde{F}_{c}=\mathbb{Z}/2\mathbb{Z}\times SL_{2}(\mathbb{C})$ be the simply connected covering of $F_{c}$. Let $Y$ be a set of 12 elements and let $\tilde{F}_{c}$ act on $Y$ trivially. Then $K_{\tilde{F}_{c}}(Y\times Y)$ is isomorphic to the $12\times 12$ matrix ring $M_{12}(\text{Rep\,}\tilde{F}_{c})$, where $\text{Rep\,}\tilde{F}_{c}$ is the representation ring of $\tilde{F}_{c}$. For nonnegative integer $k$, let $V(k)$ be an irreducible representation of $SL_{2}(\mathbb{C})$ with highest weight $k$. Let $\epsilon$ be the sign representation of $\mathbb{Z}/2\mathbb{Z}$. Regarding $V(k)$ and $\epsilon$ as representations of $\tilde{F}_{c}$ naturally, then, up to isomorphism, the irreducible representations of $\tilde{F}_{c}$ are $V(k),\ \epsilon V(k),\ k=0,\ 1,\ 2,\ ....$ When $k$ is even, $V(k)$ and $\epsilon V(k)$ are also irreducible representation of $F_{c}$. Let $V(k)_{ij}\in M_{12}(\text{Rep\,}\tilde{F}_{c})$ be the matrix whose entry at $(i,j)$ is $V(k)$ and is 0 elsewhere. Similarly we define $\epsilon V(k)_{ij}$. The main result in this section is the following theorem. Theorem 5.5. There is a natural injection $\displaystyle\pi:c\hookrightarrow$ $\displaystyle M_{12}(\text{Rep\,}\ \tilde{F}_{c}),$ $\displaystyle p_{i}u_{k}p_{j}^{-1}\longmapsto$ $\displaystyle\begin{cases}{V(2k)_{ij},}&{1\leq i,j\leq 3,}\\\ {V(2k+2)_{ij},}&{4\leq i,j\leq 12,}\\\ {V(2k+1)_{ij},}&{\text{otherwise};}\end{cases}$ $\displaystyle p_{i}\tau u_{k}p_{j}^{-1}{\longmapsto}$ $\displaystyle\begin{cases}{\epsilon V(2k)_{ij},}&{1\leq i,j\leq 3,}\\\ {\epsilon V(2k+2)_{ij},}&{4\leq i,j\leq 12,}\\\ {\epsilon V(2k+1)_{ij},}&{\text{otherwise};}\end{cases}$ $\displaystyle y{\longmapsto}$ $\displaystyle V(0)_{lm},\quad\text{if $y$ can be obtained from $x_{0}$ by a}$ sequence of left and/or right star operations, $\displaystyle y{\longmapsto}$ $\displaystyle\epsilon V(0)_{lm},\quad\text{if $y$ can be obtained from $x^{\prime}_{0}$ by a}$ sequence of left and/or right star operations, where $y=q_{l}x_{0}q_{m}^{-1}$ or $q_{l}x^{\prime}_{0}q_{m}^{-1}\ (m\neq 6)$ or $q_{l}x_{0}q_{6}^{-1}$ or $q_{l}x_{0}q_{6}^{-1}\tau$, $4\leq l,m\leq 12$. The injection $\pi$ induces an injective ring homomorphism $\Pi:J_{c}\rightarrow M_{12}(\text{Rep\,}\tilde{F}_{c})\simeq K_{\tilde{F}_{c}}(Y\times Y),\quad t_{w}\mapsto\pi(w),$ where $Y$ is a set of 12 elements with trivial $\tilde{F}_{c}$ action. Proof: We need to prove that (8) $\Pi(t_{w}t_{u})=\pi(w)\cdot\pi(u),\quad\text{for all }w,u\in D.$ Since $D$ is the union of all $\Gamma_{i},\ 1\leq i\leq 12$ and $D=D^{-1}$, we know that $D$ is the union of all $\Gamma_{i}^{-1}\cap\Gamma_{j},\ 1\leq i,j\leq 12.$ Assume that $w\in\Gamma_{i}^{-1}\cap\Gamma_{j}$ and $u\in\Gamma_{k}^{-1}\cap\Gamma_{l}$. Using 1.4(j), 1.4(k) and Lemma 5.2, we know that it suffices to prove formula (8) for $w\in\Gamma_{i}^{-1}\cap\Gamma_{j}$, $u\in\Gamma_{k}^{-1}\cap\Gamma_{l}$, $i,j,k,l\in\\{1,4\\}$. When $j\neq k$, by 1.4 (e) we see that $t_{w}t_{u}=0$, hence formula (8) holds in this case. When $i=j=k=l$, according to Theorem 3.1 in [QX], we know that formula (8) holds in this case. To complete the proof of the theorem we need to prove formula (8) for the following cases: (i) $i=1,\ j=k=1,\ l=4$; (ii) $i=1,\ j=k=4,\ l=4;$ (iii) $i=1,\ j=k=4,\ l=1;$ (iv) $i=4,\ j=k=1,\ l=1;$ (v) $i=4,\ j=k=1,\ l=4;$ (vi) $i=4,\ j=k=4,\ l=1.$ Keep the notations in the above paragraph. Applying 1.4 (g) and 1.4 (i) we see that to prove formula (8) we only need to prove it for the following two cases: ($\clubsuit$) $w\in\Gamma_{4}^{-1}\cap\Gamma_{1}$ and $u\in\Gamma_{1}^{-1}\cap\Gamma_{1}$; ($\spadesuit$) $w\in\Gamma_{4}^{-1}\cap\Gamma_{1}$ and $u\in\Gamma_{1}^{-1}\cap\Gamma_{4}$. Lemma $\clubsuit$: We have (a) $\Gamma_{4}^{-1}\cap\Gamma_{1}=\\{s_{1}s_{2}u_{k},\ s_{1}s_{2}\tau u_{k}\,|\,k\geq 0\\}$ and $\Gamma_{1}^{-1}\cap\Gamma_{1}=\\{u_{k},\ \tau u_{k}\,|\,k\geq 0\\}$. (b) $\displaystyle t_{s_{1}s_{2}u_{k}}t_{u_{l}}=t_{s_{1}s_{2}\tau u_{k}}t_{\tau u_{l}}=\sum_{0\leq i\leq\min\\{2k+1,2l\\}}t_{s_{1}s_{2}u_{k+l-i}}.$ (c) $\displaystyle t_{s_{1}s_{2}\tau u_{k}}t_{u_{l}}=t_{s_{1}s_{2}u_{k}}t_{\tau u_{l}}=\sum_{0\leq i\leq\min\\{2k+1,2l\\}}t_{s_{1}s_{2}\tau u_{k+l-i}}.$ ###### Proof. Part (a) is obtained from 5.3 (b1) and 5.3 (b2). Now we prove (b). Since $u_{0}$ is a distinguished involution, (b) is true for $l=0$. Assume that $l>0$. First we will prove (9) $\displaystyle t_{s_{1}s_{2}u_{0}}t_{u_{l}}=t_{s_{1}s_{2}u_{l}}+t_{s_{1}s_{2}u_{l-1}},\quad{\text{for any}}\ l>0.$ Let $\xi=q^{\frac{1}{2}}+q^{-\frac{1}{2}}$. By a simple computation we get (10) $\displaystyle C_{s_{1}s_{2}u_{0}}=(C_{s_{1}}C_{s_{2}}-1)C_{s_{0}s_{1}s_{3}}$ (11) $\displaystyle C_{s_{0}s_{1}s_{3}}C_{u_{l}}=\xi^{3}C_{u_{l}}.$ Hence (12) $\displaystyle C_{s_{1}s_{2}u_{0}}C_{u_{l}}=\xi^{3}(C_{s_{1}}C_{s_{2}}-1)C_{u_{l}}.$ Before continuing, we make a convention: we shall use the symbol $\Box$ for any element in the two-sided ideal $H^{<013}$ of $H$ spanned by all $C_{w}$ with $a(w)>3$. Then $\Box+\Box=\Box$ and $h\Box=\Box$ for any $h\in H$. In [QX, subsection 3.3, Step 1], we have shown the following identity: (13) $\displaystyle C_{s_{2}}C_{u_{l}}=C_{s_{2}u_{l}}+\Box\in C_{s_{2}u_{l}}+H^{<013}.$ Now we compute $C_{s_{1}}C_{s_{2}u_{l}}$. By formula (1) in 1.1 (e), we have $C_{s_{1}}C_{s_{2}u_{l}}=C_{s_{1}s_{2}u_{l}}+\sum\limits_{\begin{subarray}{c}y\prec s_{2}u_{l}\\\ s_{1}y<y\end{subarray}}\mu(y,s_{2}u_{l})C_{y}.$ Note that ${L}(s_{2}u_{l})=\\{s_{2}\\}$. First we have $\mu(u_{l},s_{2}u_{l})=1$ and $s_{1}u_{l}<u_{l}$. By 1.4 (c) and 1.4 (d), if $C_{y}$ appears in the above summation with nonzero coefficient, $y\neq u_{l}$ and $C_{y}\not\in H^{<013}$, then $y\in\Gamma_{4}^{-1}\cap\Gamma_{1}$. Assume $y\prec s_{2}u_{l}$, $s_{1}y<y$ and $y\in\Gamma_{4}^{-1}\cap\Gamma_{1}$. By (a) we must have $y{=s_{1}s_{2}u_{k}}=s_{1}s_{2}(s_{0}s_{1}s_{3}s_{2})^{k}s_{0}s_{1}s_{3}$ for some nonnegative integer $k\leq l-1$. Since $s_{2}s_{0}y\geq s_{0}y\geq y$, by 1.3 (d) we get $\mu(y,s_{2}u_{l})=\mu(s_{0}y,u_{l})$. Now $s_{3}u_{l}\leq u_{l}$ and $s_{3}s_{0}y\geq s_{0}y$, by 1.1(c) we get $s_{3}s_{0}y=u_{l}$. Hence $y=s_{1}s_{2}u_{l-1}$. We have shown (14) $\displaystyle C_{s_{1}}C_{s_{2}u_{l}}=C_{s_{1}s_{2}u_{l}}+C_{u_{l}}+C_{s_{1}s_{2}u_{l-1}}+\Box.$ Combining formulas (12) (13) and (14), we get formula (9). Recall the following formula in [QX, 3.3]: (15) $\displaystyle t_{u_{k}}t_{u_{l}}=\sum_{0\leq i\leq\min\\{2k,2l\\}}t_{u_{k+l-i}}.$ Now we employ formulas (9) and (15) to prove the identity in (b). We use induction on $k$. When $k=0$, it is just formula (9). Assume that the formula in (b) is true for nonnegative integer less than $k$. We have $\displaystyle t_{s_{1}s_{2}u_{k}}t_{u_{l}}=$ $\displaystyle(t_{s_{1}s_{2}u_{0}}t_{u_{k}}-t_{s_{1}s_{2}u_{k-1}})t_{u_{l}}$ $\displaystyle=$ $\displaystyle t_{s_{1}s_{2}u_{0}}\cdot\sum_{0\leq i\leq\min\\{2k,2l\\}}t_{u_{k+l-i}}-t_{s_{1}s_{2}u_{k-1}}t_{u_{l}}$ $\displaystyle=$ $\displaystyle\sum_{0\leq i\leq\min\\{2k,2l\\}}t_{s_{1}s_{2}u_{k+l-i}}+\sum_{0\leq i\leq\min\\{2k,2l\\}}t_{s_{1}s_{2}u_{k+l-i-1}}-\sum_{0\leq j\leq\min\\{2k-1,2l\\}}t_{s_{1}s_{2}u_{k+l-1-j}}$ $\displaystyle=$ $\displaystyle\sum_{0\leq i\leq\min\\{2k,2l\\}}t_{s_{1}s_{2}u_{k+l-i}}+\sum_{1\leq i\leq\min\\{2k+1,2l+1\\}}t_{s_{1}s_{2}u_{k+l-i}}-\sum_{1\leq j\leq\min\\{2k,2l+1\\}}t_{s_{1}s_{2}u_{k+l-j}}$ $\displaystyle=$ $\displaystyle\sum_{0\leq i\leq\min\\{2k+1,2l\\}}t_{s_{1}s_{2}u_{k+l-i}}$ Since $\tau u_{k}=u_{k}\tau$, by 1.4 (j) we have $t_{s_{1}s_{2}\tau u_{k}}t_{\tau u_{l}}=t_{s_{1}s_{2}u_{k}}t_{u_{l}}$. Part (b) is proved. Since $\tau u_{k}=u_{k}\tau$ and $\tau u_{k+l-i}=u_{k+l-i}\tau$, using 1.4 (j) we see that Part (c) follows from Part (b). The proof is completed. ∎ Lemma $\spadesuit$. (a) For $k\geq 0$, we have $s_{1}s_{2}u_{k}s_{2}s_{1}=x_{k+1}$ and $s_{1}s_{2}\tau u_{k}s_{2}s_{1}=x^{\prime}_{k+1}$. Moreover, $\Gamma_{4}^{-1}\cap\Gamma_{4}=\\{x_{k},\ x^{\prime}_{k}\,|\,k\geq 0\\}$. For nonnegative integers $k,l$ we have (b) $\displaystyle t_{s_{1}s_{2}u_{k}}t_{u_{l}s_{2}s_{1}}=t_{s_{1}s_{2}\tau u_{k}}t_{u_{l}\tau s_{2}s_{1}}=\sum_{0\leq i\leq\min\\{2k+1,2l+1\\}}t_{x_{k+l+1-i}}.$ (c) $\displaystyle t_{s_{1}s_{2}\tau u_{k}}t_{u_{l}s_{2}s_{1}}=t_{s_{1}s_{2}u_{k}}t_{u_{l}\tau s_{2}s_{1}}=\sum_{0\leq i\leq\min\\{2k+1,2l+1\\}}t_{x^{\prime}_{k+l+1-i}}.$ ###### Proof. Part (a) follows from the discussion in Subsection 5.3. Now we prove Part (b). First we prove (16) $\displaystyle t_{s_{1}s_{2}u_{0}}t_{u_{l}s_{2}s_{1}}=t_{x_{l+1}}+t_{x_{l}}.$ In [QX, Subsection 4.2], it is shown $t_{s_{1}s_{2}u_{0}}t_{u_{0}s_{2}s_{1}}=t_{x_{1}}+t_{x_{0}}$. Now assume that $l\geq 1$. As before, $\xi=q^{\frac{1}{2}}+q^{-\frac{1}{2}}$. Since $C_{s_{0}s_{1}s_{3}}C_{u_{l}s_{2}s_{1}}=\xi^{3}C_{u_{l}s_{2}s_{1}}$, using formula (10) we get (17) $\displaystyle C_{s_{1}s_{2}u_{0}}C_{u_{l}s_{2}s_{1}}=\xi^{3}(C_{s_{1}}C_{s_{2}}-1)C_{u_{l}s_{2}s_{1}}.$ We compute the right hand side of equality (17) step by step. As in the proof of Lemma $\clubsuit$, we shall use the symbol $\Box$ for any element in the two-sided ideal $H^{<013}$ of $H$ spanned by all $C_{w}$ with $a(w)>3$. Step 1: Compute $C_{s_{2}}C_{u_{l}s_{2}s_{1}}$. By formula (1) in 1.1 (e), we have $C_{s_{2}}C_{u_{l}s_{2}s_{1}}=C_{s_{2}u_{l}s_{2}s_{1}}+\sum\limits_{\begin{subarray}{c}y\prec u_{l}s_{2}s_{1}\\\ s_{2}y<y\end{subarray}}\mu(y,u_{l}s_{2}s_{1})C_{y}.$ Note that ${L}(u_{l}s_{2}s_{1})=\\{s_{0},s_{1},s_{3}\\}$. Assume $y\prec u_{l}s_{2}s_{1}$ and $s_{2}y<y$. * • If $s_{0}y>y$, then by 1.1(c) we get $s_{0}y=u_{l}s_{2}s_{1}$. This contradicts the assumption. So $s_{0}y>y$ would not occur. * • If $s_{1}y>y$, then by 1.1(c) we get $s_{1}y=u_{l}s_{2}s_{1}$. This contradicts the assumption. So $s_{1}y>y$ would not occur. * • If $s_{0}y<y,s_{1}y<y$ and $s_{2}y<y$, then $a(y)\geq a(w_{012})=6$. So $C_{y}\in H^{<013}.$ Therefore, (18) $\displaystyle C_{s_{2}}C_{u_{l}s_{2}s_{1}}=C_{s_{2}u_{l}s_{2}s_{1}}+\Box.$ Step 2: Similar to the proof for formula (14), we have (19) $\displaystyle C_{s_{1}}C_{s_{2}u_{l}s_{2}s_{1}}=C_{s_{1}s_{2}u_{l}s_{2}s_{1}}+C_{u_{l}s_{2}s_{1}}+C_{s_{1}s_{2}u_{l-1}s_{2}s_{1}}+\Box.$ Note $s_{1}s_{2}u_{l}s_{2}s_{1}=x_{l+1}$. Combining formulas (17)-(19), we get (16). Now we can prove part (b) using induction on $k$. By 1.4 (j), we know $t_{s_{1}s_{2}u_{k}}t_{u_{l}s_{2}s_{1}}=t_{s_{1}s_{2}\tau u_{k}}t_{u_{l}\tau s_{2}s_{1}}$. Thus for $k=0$, Part (b) is equivalent to formula (16), which is true. Now assume that $k\geq 1$ and Part (b) is true for $k-1$. Using Lemma $\clubsuit$ and 1.4(i), induction hypothesis and formula (16), we get $\displaystyle t_{s_{1}s_{2}u_{k}}t_{u_{l}s_{2}s_{1}}=$ $\displaystyle(t_{s_{1}s_{2}u_{0}}t_{u_{k}}-t_{s_{1}s_{2}u_{k-1}})t_{u_{l}s_{2}s_{1}}$ $\displaystyle=$ $\displaystyle{t_{s_{1}s_{2}u_{0}}\cdot}\sum_{0\leq i\leq\min\\{2l+1,2k\\}}t_{u_{k+l-i}s_{2}s_{1}}-t_{s_{1}s_{2}u_{k-1}}t_{u_{l}s_{2}s_{1}}$ $\displaystyle=$ $\displaystyle\sum_{0\leq i\leq\min\\{2l+1,2k\\}}(t_{x_{k+l+1-i}}+t_{x_{k+l-i}})-\sum_{0\leq i\leq\min\\{2l+1,2k-1\\}}t_{x_{k+l-i}}$ $\displaystyle=$ $\displaystyle\sum_{0\leq i\leq\min\\{2l+1,2k+1\\}}t_{x_{k+l+1-i}}.$ This completes the proof for Part (b). Proof for Part (c) is similar. First, it is easy to check that $C_{s_{0}s_{2}u_{0}}C_{u_{0}s_{2}s_{1}}=\xi^{3}(C_{\tau x^{\prime}_{1}}+C_{\tau x^{\prime}_{0}}),$ which implies $t_{s_{1}s_{2}\tau u_{0}}t_{u_{0}s_{2}s_{1}}=t_{x^{\prime}_{1}}+t_{x^{\prime}_{0}}.$ Further, we prove that $t_{s_{1}s_{2}\tau u_{0}}t_{u_{l}s_{2}s_{1}}=t_{x^{\prime}_{1+1}}+t_{x^{\prime}_{l}}.$ Then using induction on $k$, as the proof for Part (b), we prove Part (c). The proof for Lemma $\spadesuit$ is completed. ∎ We have completed the proof for Theorem 5.5. 5.6. Motivated by Theorem 5.5 and the discussion of the cocenter of $J$ in [BDD, Section 5] and some other evidences, we suggest a modification of Lusztig’s conjecture on the structure of $J_{c}$ which is stated for any connected reductive groups over $\mathbb{C}$. Let $W$ be the extended affine Weyl group attached to a connected reductive group over $\mathbb{C}$ (see Subsection 1.1) and let $c$ be a two-sided cell of $W$. Let $F_{c}$ be a maximal reductive subgroup of the centralizer of an element in the corresponding unipotent class of $G$. Then there should exist a reductive group $\tilde{F}_{c}$ with the following properties: (i) The reductive group $\tilde{F}_{c}$ is a simply connected covering of $F_{c}$. That is, the identity component $\tilde{F}_{c}^{\circ}$ has simply connected derived group, and there is a natural surjective homomorphism $\tilde{F}_{c}\to F_{c}$ with finite kernel. In particular, if $F_{c}^{\circ}$ has simply connected derived group, then $\tilde{F}_{c}=F_{c}$. (ii) There exists a finite set $Y$ with an algebraic action of $\tilde{F}_{c}$ and an injection $\pi:c\hookrightarrow\text{isomorphism classes of irreducible $\tilde{F}_{c}$-vector bundles on}\ Y\times Y.$ such that (iii) The injection $\pi$ induces a ring injection $\Pi:J_{c}\to K_{\tilde{F}_{c}}(Y\times Y),\ \ t_{x}{\mapsto}\pi(x).$ (iv) $\pi(x^{-1})_{(a,b)}=\pi(x)_{(b,a)}^{*}$ is the dual representation of $\pi(x)_{(b,a)}.$ (v) $K_{\tilde{F}_{c}}(Y\times Y)$ is a finitely generated left (and right as well) $\Pi(J_{c})$-module. It seems natural that the $F_{c}$-set $\mathbf{B}_{e}$ defined in a recent paper (see [L4]) would have an $\tilde{F}_{c}$-action compatible with the $F_{c}$-action and then $\mathbf{B}_{e}$ should be a good candidate for the set $Y$ above. Acknowledgement: Part of the work was done during YQ’s visit to the Academy of Mathematics and Systems Science, Chinese Academy of Sciences. YQ is very grateful to the AMSS for hospitality and for financial supports. ## References * [B] R. Bezrukavnikov, On tensor categories attached to cells in affine Weyl groups, In ”Representation Theory of Algebraic Groups and Quantum Groups”, Adv. Stud. Pure Math., 40, Math. Soc. Japan, Tokyo, 2004, pp. 69-90. * [BDD] R. Bezrukavnikov, S. Dawydiak, G. Dobrovolska, On the structure of the affine asymptotic Hecke algebras, arXiv:2110.15903. * [BO] R. Bezrukavnikov and V. Ostrik, On tensor categories attached to cells in affine Weyl groups II, In ”Representation Theory of Algebraic Groups and Quantum Groups”, Adv. Stud. Pure Math., 40, Math. Soc. Japan, Tokyo, 2004, pp.101-119. * [DLP] C. De Concini, G. Lusztig, C. Procesi,Homology of the zero-set of a nilpotent vector field on a flag manifold, J. Amer. Math. Soc. 1 (1988), 15-34. * [D] J. Du, The decomposition into cells of the affine Weyl group of type $\tilde{B_{3}}$, Communications in Algebra, 16 (1988), no.7, 1383–1409. * [KL] D. Kazhdan and G. Lusztig, Representations of Coxeter groups and Hecke algebras, Invent. Math. 53 (1979), 165-184. * [L1] G. Lusztig, Cells in affine Weyl groups, in “Algebraic groups and related topics”, Advanced Studies in Pure Math., vol. 6, Kinokunia and North Holland, 1985, pp. 255-287. * [L2] G. Lusztig, Cells in affine Weyl groups, II, J. Alg. 109 (1987), 536-548. * [L3] G. Lusztig, Cells in affine Weyl groups, IV, Journal of The Faculty of Science, 36 (1989), no.2, 297-328. * [L4] G. Lusztig, Discretization of Speinger fibers, arXiv:1712.07530v3, 2021. * [QX] Yannan Qiu and Nanhua Xi, The based ring of two-sided cells in an affine Weyl group of type $\tilde{B}_{3}$, I. Sci. China Math., to appear. arxiv: 2107.08983. * [X1] N. Xi, Representations of Affine Hecke Algebras, volume 1587, Springer Lecture Notes in Math., 1994. * [X2] N. Xi, The based ring of two-sided cells of affine Weyl groups of type ${\tilde{A}_{n-1}}$, volume 749, American Mathematical Soc., 2002.
# Classical and Quantum Algorithms for Tensor Principal Component Analysis Matthew B. Hastings Station Q, Microsoft Research, Santa Barbara, CA 93106-6105, USA Microsoft Quantum and Microsoft Research, Redmond, WA 98052, USA ###### Abstract We present classical and quantum algorithms based on spectral methods for a problem in tensor principal component analysis. The quantum algorithm achieves a quartic speedup while using exponentially smaller space than the fastest classical spectral algorithm, and a super-polynomial speedup over classical algorithms that use only polynomial space. The classical algorithms that we present are related to, but slightly different from those presented recently in Ref. [1]. In particular, we have an improved threshold for recovery and the algorithms we present work for both even and odd order tensors. These results suggest that large-scale inference problems are a promising future application for quantum computers. ## 1 Introduction Principal component analysis is a fundamental technique that finds applications in reducing the dimensionality of data and denoising. While an optimal choice of principal components for a matrix can be computed efficiently using linear algebra, the corresponding problem for tensors is much less well-understood. Ref. [2] introduced a simple statistical model for tensor principal component analysis, termed the “spiked tensor” problem, and this paper has lead to a large amount of follow-up research. The model consists of (see below for more precise definitions) randomly choosing some unknown “signal vector” $v_{\rm sig}\in{\mathbb{R}}^{N}$; then, the $p$-th order tensor $T_{0}=\lambda v_{\rm sig}^{\otimes p}+G$ (1) is formed, where $G\in({\mathbb{R}^{N}})^{\otimes p}$ is noise chosen from some random distribution and where $\lambda$ is some scalar representing a signal-to-noise ratio. One task called recovery is to infer $v_{\rm sig}$ (to some accuracy) given $T_{0}$. A simpler task called detection is to distinguish the case $\lambda=0$ from $\lambda=\overline{\lambda}$ for some $\overline{\lambda}>0$, again just given $T_{0}$. Ref. [2] presented a variety of algorithms for this problem. Following the normalization of Ref. [1], the entries of $G$ are chosen independently from a Gaussian distribution of zero mean and unit variance, with $|v_{\rm sig}|=\sqrt{N}$. Then, information theoretically, it is possible to recover for $\lambda$ much larger than $N^{(1-p)/2}$ [2, 3]. However, no polynomial time algorithm is known that achieves this performance. Rather, the two best known algorithms are spectral and sum-of-squares. Spectral algorithms were first suggested in Ref. [2]. There, a matrix is formed from $T_{0}$ (if $p$ is even, the matrix is $N^{p/2}$-by-$N^{p/2}$, with its entries given by entries of $T_{0}$) and the leading eigenvector of the matrix is used to determine $v_{\rm sig}$. For even $p$, this method works for $\lambda$ much larger than $N^{-p/4}$, and a variant of it is conjectured to perform similarly for odd $p$. Methods based on the sum-of-squares also perform similarly to the spectral method. The sum-of-squares method [4, 5] for this problem gives rise to a sequence of algorithms [6, 7], in which one can recover at $\lambda$ smaller than $N^{-p/4}$ at the cost of runtime and space increasing exponentially in ${\rm polylog}(N)N^{-p/4}/\lambda$. In Ref. [1], a sequence of spectral algorithms with similar performance was shown. In this paper, we present another spectral algorithm for the problem. Our spectral algorithm for even $p$ is closely related to that of Ref. [1] which we became aware of while this paper was in preparation, and we use the normalization in that paper. However, we make several changes. Our main technical results are the following. First, we prove an improved threshold for recovery for even $p$ using our algorithm; the improvement is by a constant factor and relies on a randomized method of recovering. Second, we provide a different algorithm for odd $p$ with provable guarantees on while no guarantees were given in Ref. [1] for odd $p$. For both even and odd $p$, we have provable bounds on recovery for $\lambda$ of order $N^{-p/4}$ (without any polylogarithmic factors) and we have a sequence of algorithms similar to that above for $\lambda$ small compared to $N^{-p/4}$. Third, we give a quantum algorithm for our spectral method, achieving a quartic speedup and exponential reduction in space. This quantum algorithm involves two main ideas. The first uses phase estimation and amplitude amplification to obtain a quadratic speedup in computing the largest eigenvector. The second idea uses a chosen input state to obtain a further quadratic speedup, giving the overall quartic speedup. We emphasize that the quantum speedup is quartic compared to classical spectral algorithms presented here and in previous work. We are not able to make an accurate comparison of the runtime to sum-of-squares methods. In part, given that the runtime of all of these algorithms increases exponentially in $\lambda^{-1}$, a change in prefactors in some estimates for threshold can give rise to polynomial changes in runtime. We expect that many of these estimates of thresholds are not tight (indeed, we expect that they are off by a polylogarithmic factor), and so either improved analytic methods or numerical simulations are needed to give an accurate comparison. At a more heuristic level, we present a rather different motivation for our spectral algorithm compared to Ref. [1]. Rather than being motivated by the so-called Kikuchi free energy, we instead are motivated by mean-field approximations to quantum many-body systems. We consider a system of a some number ${n_{bos}}$ of qudits, each of dimension $N$, and use the tensor $T_{0}$ to construct a quantum Hamiltonian on these qudits. Increasing ${n_{bos}}$ gives rise to a similar sequence of algorithms as above, with increased runtime but improved performance: the required ${n_{bos}}$ increases polynomially in $\lambda^{-1}$ as ${\rm polylog}(N)(N^{-p/4}/\lambda)^{4/(p-2)}$, but the runtime increases exponentially. Restricting to the symmetric subspace, these ${n_{bos}}$ qudits can be thought of as a system of bosons. In the case $p=4$, for example, our Hamiltonian has pairwise interaction terms for all pairs of qudits. It is natural from the viewpoint of mean-field theory in physics then to expect that the leading eigenvector of the problem, for large ${n_{bos}}$, can be approximated by a product state. While the bounds for arbitrary pairwise Hamiltonians would require rather large ${n_{bos}}$ for given $N$ in order for such a mean-field approximation to be accurate [8, 9, 10], we will be able to prove that for the statistical model above the mean-field approximation becomes accurate with high probablity at much smaller ${n_{bos}}$, depending upon the value of $\lambda$. In this mean-field regime, the product state is an ${n_{bos}}$-fold tensor product of a single particle state, and this single particle state is given by $v_{\rm sig}$ in an obvious way, regarding the vector $v_{\rm sig}$ as a vector in the single-particle Hilbert space. While we will not prove that this state is a good approximation to the leading eigenvector, it will be a good approximation to some state in an eigenspace with large eigenvalue. Then, the single particle density matrix allows one to infer $v_{\rm sig}$ (a similar matrix was used in Ref. [1] where it was termed a voting matrix). Classically, implementing this spectral algorithm requires high-dimensional linear algebra, in particular finding the leading eigenvector of a matrix of dimension $\approx N^{n_{bos}}$. This makes it a natural candidate for a quantum algorithm. Since the Hamiltonian here is fairly simple, it can be simulated efficiently using standard techniques in the literature reviewed later. This allows us to give a simple algorithm based on preparing a random initial state and then phase estimating in an attempt to project onto the leading eigenvector. The probability of success in this projection is inverse in the dimension of the matrix, so this simple algorithm leads to no speedup over classical. However, we show that it is possible to apply amplitude amplification to give a quadratic speedup over classical. More surprisingly, we show that one can use the tensor $T_{0}$ to prepare an input state to the algorithm with improved overlap with the leading eigenvector, giving the quantum algorithm a quartic speedup over classical. Here, when comparing to classical we are considering classical algorithms based on the power method or similar algorithms such as Lanczos; these algorithms require exponential space while the quantum algorithm uses only polynomial space. We also consider classical algorithms based on ideas in Ref. [11] which use polynomial space but the quantum algorithm is super-polynomially faster than these algorithms. We also present some minor improvements to the quantum algorithm which may be useful in practice. ### 1.1 Definitions, Random Ensembles, and Notation Let us make some formal definitions. A tensor $T$ of order $p$ and dimension $N$ is a multi-dimensional array. The entries of the tensor are written $T_{\mu_{1},\mu_{2},\ldots,\mu_{p}}$ where $p\geq 1$ is an integer and each $\mu_{a}$ ranges from $1,\ldots,N$. Generalizing previous work on this problem, we consider two possible cases, one in which entries of a tensor are chosen to be real numbers, and one in which they may be complex numbers, so that either $T\in({\mathbb{R}^{N}})^{\otimes p}$ or $T\in({\mathbb{C}}^{N})^{\otimes p}$; we explain later the reason for this generalization; a tensor with all entries real will be called a real tensor. A symmetric tensor is one that is invariant under permutation of its indices. The symmetrization of a tensor is equal to $1/p!$ times the sum of tensors given by permuting indices. The spiked tensor model for given $N,p$ is defined as follows. Let $v_{\rm sig}$ be a vector in ${\mathbb{R}}^{N}$, normalized by $|v_{\rm sig}|=\sqrt{N}$, chosen from some probability distribution; this is the “signal vector”. Let $G$ be a real tensor of order $p$ with entries chosen from a Gaussian distribution with vanishing mean. We let $T_{0}=\lambda v_{\rm sig}^{\otimes p}+G$ as above, where $v_{\rm sig}^{\otimes p}$ is defined to be the tensor with entries $(v_{\rm sig}^{\otimes p})_{\mu_{1},\ldots,\mu_{p}}=\prod_{a=1}^{p}(v_{\rm sig})_{\mu_{a}}.$ Here we use the notation that a subscript on a vector denotes an entry of that vector; we use a similar notation for matrices later. Remark: some of the best sum-of-squares results are for a different distribution in which the entries of $T_{0}$ are chosen from a biased distribution on $\\{-1,+1\\}$, rather than for the Gaussian distribution. We expect that using that distribution would not affect the results here too much, but we avoid treating that case also for simplicity. Since the tensor $v_{\rm sig}^{\otimes p}$ is symmetric, of course it is natural to replace $T_{0}$ by its symmetrization. Indeed, no information can be lost by this replacement since given a tensor $T_{0}$ one can symmetrize the tensor, and then add back in Gaussian noise chosen to vanish under symmetrization to obtain a tensor drawn from the same distribution as $T_{0}$ was. That is, the cases in which $G$ is symmetrized or not can be reduced to each other. A generalization of this problem is the case in which $G$ is chosen to have complex entries, with each entry having real and imaginary parts chosen from a Gaussian distribution with vanishing mean and variance $1/2$. We refer to this as the complex ensemble, while we refer to the case where $G$ has real entries as the real ensemble; the choice of reducing the variance to $1/2$ is a convenient normalization for later. It is clear that since $v_{\rm sig}$ is real, the case of complex $G$ can be reduced to the real case (up to an overall rescaling for the different variance) simply by taking the real part of $T_{0}$, and similarly the real case can be reduced to the complex case (again up to an overall rescaling) by adding Gaussian distributed imaginary terms to the entries of $T_{0}$. We will see that for odd $p$, at least for reasons of analyzing the algorithms, it is convenient not to symmetrize $T_{0}$ and to take complex $G$, while for even $p$ this is not necessary. It may be possible to avoid doing this for odd $p$ (which may improve the detection and recovery threshold of the algorithm by constant factors) and we comment on this later. We treat $p$ as fixed in the asymptotic notation, but consider the dependence on $N,{n_{bos}}$. So, throughout this paper, when we refer to a polynomial in $N$, we mean a polynomial independent of the parameter ${n_{bos}}$. The polynomial may, however, depend on $p$, such as $N^{p}$. We make additionally the following assumptions: ###### Assumption 1. We assume that ${n_{bos}}=O(N^{\theta})$ for some $p$-dependent constant $\theta>0$ chosen sufficiently small. We will also assume that $\lambda$ is $\Omega(N^{-\theta^{\prime}})$ for some $p$-dependent constant $\theta^{\prime}>p/4$. Finally, we assume that $\lambda=O(N^{-p/4})$. Remark: there is of course no reason to consider $\lambda$ larger than this since simple spectral methods succeed if $\lambda$ is $\omega(N^{-p/4})$, but we state this assumption explicitly as it simplifies some of the big-O notation. We will explicitly state this Assumption 1 in all theorems where it is needed; the assumption will be implicit in the statement of the lemmas and will not be explicitly stated to avoid cluttering the statement of the results. The first of these assumptions, that ${n_{bos}}=O(N^{\theta})$, is useful to simplify some of the statements of the results to avoid having to specify the allowed range of ${n_{bos}}$ in each case. For example, we will say that a quantity such as ${n_{bos}}^{p}/N$ is $o(1)$, meaning that we must take $\theta<1/p$. We do not specify the exact value of $\theta$ but it can be deduced from the proofs if desired. The second of these assumptions, that $\lambda$ is $\Omega(N^{-\theta^{\prime}})$, also helps simplify some of the statements of the results. Since we have assumed that ${n_{bos}}=O(N^{\theta})$ and we will see that the required ${n_{bos}}$ increases polynomially with $\lambda^{-1}$, this assumed lower bound on $\lambda$ is not a further restriction on our results. We write $\mathbb{E}[\ldots]$ to denote expectation values and ${\rm Pr}[\ldots]$ to denote a probability. Usually these are expectation values or probabilities over choices of $G$, though in some cases we consider expectation values over other random variables. We use $\|\ldots\|$ to denote the operator norm of an operator, i.e., the largest singular value of that operator. We use $|\ldots|$ to denote either the $\ell_{2}$ norm of a tensor or the $\ell_{2}$ norm of a quantum state. All logarithms are natural logarithms. We use $\langle\ldots|\ldots\rangle$ to denote inner products and use bra-ket notation both for vectors and for quantum mechanical states. We will need to compute expectation values over random choices of $G$ and also compute expectation values of certain operators in quantum states, such as $\langle\psi|O|\psi\rangle$ for some state $\psi$ and operator $O$. We refer to the latter as a quantum expectation value of $O$ in state $\psi$ to distinguish it from an expectation value over random variables. ### 1.2 Outline In section 2, we review some results on recovery and boosting from Ref. [1] and present a randomized recovery procedure that will help in improving the recovery threshold. In section 3, we give spectral algorithms for the spiked tensor problem for the case of both even and odd $p$. In that section, we present algorithms in terms of eigenvalues and eigenvectors of a matrix (more precisely, vectors in some eigenspace and quantum expectation values of operators in those vectors) that we call a Hamiltonian. We leave the method to compute these eigenvalues and expectation values for later in section 5, where we give classical and quantum algorithms for this and give time bounds for those algorithms. In section 4, we give some results on the spectra of random tensors needed for the analysis of these algorithms. A key idea here is reducing the case of a $p$-th order tensor for odd $p$ to the case of a $q$-th order tensor for even $q=2(p-1)$. One interesting corollary of this technique, see corollary 2, is that for odd $p$ and for the minimal value of ${n_{bos}}$ we are able remove a logarithmic factor in some of the bounds (a similar logarithmic factor has been removed also using in [4] using what they termed an “unfolding” algorithm). An appendix A gives an introduction to some techniques used to evaluate expectation values of tensors networks whose tensors are chosen from a Gaussian distribution; these techniques are used earlier in the paper. In section 6, we further discuss tensor networks and use this to consider limitations of certain algorithms and also to explain further some of the motivation for this algorithm. In section 7, we discuss some extensions of the results. The proof of detection is in theorem 2 for the even $p$ case and 4 for the odd $p$ case. The proof of recovery is in theorem 3 for the even $p$ case and theorem 5 for the odd $p$ case. The runtime bound for the fastest quantum algorithm is in theorem 6. This theorem gives a quartic improvement in the runtime compared to the fastest classical spectral algorithm; more precisely the log of the runtime with the quantum algorithm divided by the log of the runtime of the classical algorithm approaches $1/4$ as $N\rightarrow\infty$ at fixed $N^{-p/4}/\lambda$. ## 2 Recovery In this section we discuss recovery and define randomized procedures for recovery that will be useful in boosting the threshold. Following the notation of Ref. [1], define the correlation between two vectors by a normalized overlap ${\rm corr}(x,y)=\frac{|\langle x|y\rangle|}{|x|\cdot|y|}.$ (2) The goal of an algorithm is to produce a vector $x$ with large ${\rm corr}(x,v_{\rm sig})$. Note that we take the absolute value in the correlation, ignoring the sign. For even $p$, the sign of $v_{\rm sig}$ is irrelevant, while for odd $p$ it is easy, given a guess of $v_{\rm sig}$ up to sign, to try both choices of sign and see which is most likely. Strong recovery means that ${\rm corr}(x,v_{\rm sig})=1-o(1)$. Proposition 2.6 of Ref. [1], which is noted in that reference as being implicit in Ref. [2], shows how to “boost” a weaker correlation to strong recovery. It is shown that given a vector $u$ one can apply a single iteration of the the tensor power algorithm to obtain a new vector $x$ such that, with high probability, ${\rm corr}(x,v_{\rm sig})\geq 1-c\lambda^{-1}{\rm corr}(u,v_{\rm sig})^{1-p}N^{(1-p)/2},$ (3) where $c$ is a constant depending on $p$. So, for any $\lambda\omega(N^{(1-p)/2}$, if ${\rm corr}(u,v_{\rm sig})=\Omega(1)$, we have ${\rm corr}(x,v_{\rm sig})=1-o(1)$. Thus, it suffices to construct which outputs some vector $u$ which has, with high probability, ${\rm corr}(u,v_{\rm sig})=\Omega(1)$. This is termed weak recovery; indeed, one can be satisfied with even weaker assumptions depending on the value of $\lambda$; for $\lambda$ close to $N^{-p/4}$, one may have polynomially small ${\rm corr}(u,v_{\rm sig})$. The spectral algorithms that we construct later will output a matrix that we will write $\rho_{1}$. This matrix will be a positive semi-definite Hermitian matrix with trace $1$. In physics terms, such a matrix is termed a density matrix. For sufficiently large $\lambda$, the leading eigenvector of the matrix will have a large overlap with $v_{\rm sig}$. However, for smaller $\lambda$, we will still have some lower bound that, with high probability, $\langle v_{\rm sig}|\rho_{1}|v_{\rm sig}\rangle=\Omega(1)|v_{\rm sig}|^{2}$. The randomized algorithm 1 below shows how to use this to obtain weak recovery. This randomized algorithm allows us to improve the threshold over the recovery method [1] of simply using the leading eigenvector $\rho_{1}$ since it works even in some cases where the leading eigenvector has small correlation with $v_{\rm sig}$. Putting these results together we find that ###### Corollary 1. Given an algorithm that, with high probability, outputs a density matrix $\rho$ with $\frac{\langle v_{\rm sig}|\rho_{1}|v_{\rm sig}\rangle}{N}=\Omega(1),$ then in polynomial time, with high probability, strong recovery is possible. We will simply say that such an algorithm achieves recovery. We present the algorithm for the case that the matrix $\rho_{1}$ may be complex; if the matrix is real, one can instead sample from a real Gaussian distribution and the proof of lemma 1 goes through will slightly different constants. Algorithm 1 Input: density matrix $\rho$. Output, some vector $w$ obeying bounds described above * 1. Randomly sample a vector $u$ with entries chosen from a correlated complex Gaussian distribution with zero mean and with covariance $\mathbb{E}[\overline{u}_{i}u_{j}]=(\rho)_{ij},$ with $\mathbb{E}[u_{i}u_{j}]=\mathbb{E}[\overline{u}_{i}\overline{u}_{j}]=0$. * 2. Let $w=u/|u|$. We have ###### Lemma 1. For algorithm 1, with probability at least $1/2$, $|\langle w|v_{\rm sig}\rangle|\geq c^{\prime}\sqrt{\langle v_{\rm sig}|\rho|v_{\rm sig}\rangle},$ (4) for some scalar $c^{\prime}>0$. ###### Proof. We have $\mathbb{E}[|u|^{2}]={\rm tr}(\rho)=1$. Hence, with probability at least $3/4$ we have that $|u|^{2}\leq 4/3$. We have $\mathbb{E}[|\langle u|v_{\rm sig}\rangle|^{2}]=\langle v_{\rm sig}|\rho v_{\rm sig}\rangle$ and since $\langle u|v_{\rm sig}\rangle$ is a Gaussian random variable with mean $0$, with probability at least $3/4$ its absolute value is at least some positive constant $c^{\prime\prime}$ (the exact constant can be deduced from the error function) times its standard deviation. Hence, the lemma follows for $c^{\prime}=(3/4)c^{\prime\prime}$. ∎ ## 3 Spectral Algorithm We now give the spectral algorithms for the spiked tensor problem for the case of both even and odd $p$. In subsection 3.1, we define a Hamiltonian $H(T)$ given an arbitrary tensor $T$. Then, in subsection 3.2, we present the spectral algorithm in terms of $H(T_{0})$. For even $p$, the Hamiltonian that we present is very similar to the matrix $Y$ given in Ref. [1] but it has some minor differences. In our language (explained below), this matrix $Y$ is obtained by projecting our Hamiltonian of Eq. (5) into the subspace of the “symmetric subspace” spanned by $|\mu_{1}\rangle\otimes|\mu_{2}\rangle\otimes\ldots\otimes|\mu_{{n_{bos}}}\rangle$ with $\mu_{1},\ldots,\mu_{{n_{bos}}}$ all distinct. The relative reduction in the size of the matrix is only $O(1/N)$ in the limit of large $N$. Also, in our method, we have an $O(N)$ rotational symmetry of the basis which is very useful in analysis, for example showing that the eigenvalues of $H(\lambda v_{\rm sig}^{\otimes p})$ are independent of choice of $v_{\rm sig}$. For the matrix $Y$ of [1], this is not obvious to us and we do not fully understand the claimed behavior of the largest eigenvalue in that case. We will use a different notation, using creation and annihilation operators, which will help make this rotational symmetry more explicit. For odd $p$, the Hamiltonian that we use is unrelated to that of Ref. [1]. ### 3.1 Hamiltonian Definition For even $p$, given a tensor $T$ we define an linear operator $H(T)$ that we call a Hamiltonian as follows. This Hamiltonian is a linear operator on a vector space $({\mathbb{R}}^{N})^{\otimes{n_{bos}}}$ or $({\mathbb{C}}^{N})^{\otimes{n_{bos}}}$, for integer ${n_{bos}}\geq 1$ chosen below. We write basis elements of this space as $|\mu_{1}\rangle\otimes|\mu_{2}\rangle\otimes\ldots\otimes|\mu_{n_{bos}}\rangle$, and we call this space the full Hilbert space. We define $H(T)=\frac{1}{2}\sum_{i_{1},\ldots,i_{p/2}}\Bigl{(}\sum_{\mu_{1},\ldots,\mu_{p}}T_{\mu_{1},\mu_{2},\ldots,\mu_{p}}|\mu_{1}\rangle_{i_{1}}\langle\mu_{1+p/2}|\otimes|\mu_{2}\rangle_{i_{2}}\langle\mu_{2+p/2}|\otimes\ldots\otimes|\mu_{p/2}\rangle_{i_{p/2}}\langle\mu_{p}|+{\rm h.c.}\Bigr{)},$ (5) where the sum is over distinct $i_{1},i_{2},\ldots,i_{p/2}$ so that there are $(p/2!){{n_{bos}}\choose p/2}$ terms in the sum and where ${\rm h.c.}$ means adding the Hermitian conjugate of the given terms, so that $H(T)$ is Hermitian and where $|\mu\rangle_{i}\langle\nu|$ denotes the operator $|\mu\rangle\langle\nu|$ on qudit $i$. We require that ${n_{bos}}\geq p/2$ or else $H(T)$ is trivially zero. Note of course that if $T$ is real and symmetric, then the term $\sum_{\mu_{1},\ldots,\mu_{p}}T_{\mu_{1},\mu_{2},\ldots,\mu_{p}}|\mu_{1}\rangle_{i_{1}}\langle\mu_{1+p/2}|\otimes|\mu_{2}\rangle_{i_{2}}\langle\mu_{2+p/2}|\otimes\ldots\otimes|\mu_{p/2}\rangle_{i_{p/2}}\langle\mu_{p}|$ is already Hermitian. $H(T)$ can be regarded as a Hamiltonian acting on a space of ${n_{bos}}$ qudits, each of dimension $N$, and with interaction between sets of $p/2$ particles at a time. Even if $T$ is not symmetrized, $H(T)$ is unchanged if one applies an arbitrary permutation to the first $p/2$ indices of $T$ and applies the same permutation to the last $p/2$ indices of $T$. We may restrict to the symmetric subspace of this Hilbert space. We write $D(N,{n_{bos}})$ to indicate the dimension of this subspace. For $N\gg{n_{bos}}$, we can approximate $D(N,{n_{bos}})\approx N^{n_{bos}}/{n_{bos}}!$. Within the symmetric subspace, we can write this Hamiltonian in a so-called “second-quantized” form: $H(T)=\frac{1}{2}\Bigl{(}\sum_{\mu_{1},\ldots,\mu_{p}}T_{\mu_{1},\mu_{2},\ldots,\mu_{p}}\Bigl{(}\prod_{i=1}^{p/2}a^{\dagger}_{\mu_{i}}\Bigr{)}\Bigl{(}\prod_{i=p/2+1}^{p}a_{\mu_{i}}\Bigr{)}+{\rm h.c.}\Bigr{)}.$ (6) This replacement by a second-quantized Hamiltonian is simply a convenient notation. The operators $a^{\dagger}_{\mu},a_{\mu}$ are bosonic creation and annihilation operators, obeying canonical commutation relations $[a_{\mu},a^{\dagger}_{\nu}]=\delta_{\mu,\nu}$. See appendix B for a brief review of this formalism. We restrict to the subspace with a total of ${n_{bos}}$ bosons, i.e., we define the number operator $n$ by $n\equiv\sum_{\mu}a^{\dagger}_{\mu}a_{\mu},$ (7) and restrict to $n={n_{bos}}.$ An orthonormal basis of states for this symmetric subspace is given by all states equal to some normalization constant multiplying $a^{\dagger}_{\mu_{1}}a^{\dagger}_{\mu_{2}}\ldots a^{\dagger}_{\mu_{n_{bos}}}|0\rangle$, where $|0\rangle$ is the vacuum state (i.e., the state annihilated by $a_{\mu}$ for all $\mu$), and where $\mu_{1}\leq\mu_{2}\leq\ldots\leq\mu_{n_{bos}}$ is some sequence. The second quantized Hamiltonian for the symmetric subspace is unchanged under arbitrary permutation of the first $p/2$ indices of $T$ and arbitrary (not necessarily the same) permutation of the last $p/2$ indices of $T$. For odd $p$, we define the Hamiltonian $H(T)$ as follows. Given a tensor $T$ of odd order $p$, define a new tensor $\tilde{T}$ of even order $q=2(p-1)$ with components $\tilde{T}_{\mu_{1},\ldots,\mu_{(p-1)/2},\nu_{1},\ldots,\nu_{(p-1)/2},\mu_{(p-1)/2+1},\ldots,\mu_{p-1},\nu_{(p-1)/2},\ldots,\nu_{p-1}}=\sum_{\sigma}T_{\mu_{1},\ldots,\mu_{p-1},\sigma}T_{\nu_{1},\ldots,\nu_{p-1},\sigma}.$ (8) Then define $H(T)=H(\tilde{T})$, using the definition (5) for $H(\tilde{T})$. Note the order of indices on the left-hand side of Eq. (8). Using the second- quantized notation, this gives for odd $p$: $\displaystyle H(T)$ $\displaystyle=$ $\displaystyle\frac{1}{2}\Bigl{(}\sum_{\mu_{1},\ldots,\mu_{p-1}}\sum_{\nu_{1},\ldots,\nu_{p-1}}\sum_{\sigma}T_{\mu_{1},\mu_{2},\ldots,\mu_{p-1},\sigma}T_{\nu_{1},\nu_{2},\ldots,\nu_{p-1},\sigma}\Bigl{(}\prod_{i=1}^{(p-1)/2}a^{\dagger}_{\mu_{i}}a^{\dagger}_{\nu_{i}}\Bigr{)}\Bigl{(}\prod_{i=(p-1)/2+1}^{p-1}a_{\mu_{i}}a_{\nu_{i}}\Bigr{)}+{\rm h.c.}\Bigr{)},$ Now we require that ${n_{bos}}\geq p-1$ as otherwise $H(T)$ is trivially zero. For this Hamiltonian, it is convenient to take $G$ from the complex ensemble because, as we explain more below, it makes $\mathbb{E}[H(G)]$ equal to zero, as well as canceling out certain terms in higher order moments, making the proof of the spectral properties of $H(G)$ simpler. We discus later to what extent we can avoid using the complex ensemble. ### 3.2 Spectral Algorithms The spectral algorithm for detection and recovery is algorithm 2. In this subsection we prove correctness of this algorithm, using statistical properties of $H(G)$ proven later. This algorithm uses quantities $E_{0}$ and $E_{max}$ defined later; roughly $E_{max}$ is an upper bound on the eigenvalues of $H(G)$ and $E_{0}$ is the largest eigenvalue of $H(\lambda v^{\otimes p})$. The algorithm can then achieve detection by verifying that the largest eigenvalue is significantly larger than $E_{max}$, which occurs when $E_{0}$ is large enough. Indeed, we will see that it suffices to have $E_{0}=(1+c)E_{max}$ for any fixed $c>0$ (some results can be extended to the case that $c$ decays slowly with $N$ but we omit this for brevity). This fixes the scaling of ${n_{bos}}$ with $\lambda$ so that we need (up to polylogarithmic factors) ${n_{bos}}\gtrsim(N^{-p/4}/\lambda)^{4/(p-2)}$. One interesting feature of the algorithm is that in step 3, we compute the density matrix of the leading eigenvector or of any vector in the eigenspace of eigenvalue $\geq E_{cut}$, for $E_{cut}=(E_{0}+E_{max})/2$ defined in the algorithm. This might seem surprising, given that the leading eigenvector is computed in step 1; one might wonder why some other vector should be taken. We describe the algorithm in this way since, in later classical and quantum algorithms that we give to compute the spectral properties of the matrix, we might not extract the leading eigenvector but instead extract only some vector in this eigenspace due to use of the power method in a classical algorithm or due to approximations in phase estimation in a quantum algorithm. Thus, much of our analysis is given to showing that some eigenvalue larger than $E_{cut}$ exists by lower bounding the leading eigenvalue $\lambda_{1}$, but given that some such eigenvalue exists, we do not worry too much about exactly what mixture of eigenvectors in the given eigenspace we compute. Algorithm 2 Spectral algorithm. This algorithm takes a tensor $T_{0}$ as input and also a scalar $\overline{\lambda}$ and an integer ${n_{bos}}$. The output is a decision about whether $\lambda=\overline{\lambda}$ or $\lambda=0$, and, if the algorithm reports that $\lambda=\overline{\lambda}$, it also returns an approximation of $v_{\rm sig}$ (up to an overall sign). The quantity ${n_{bos}}$ is chosen depending upon the value of $\overline{\lambda}$; smaller values of $\lambda$ require larger values of ${n_{bos}}$ in order for $E_{0}$ to be sufficiently larger than $E_{max}$ for the algorithm to be accurate. See theorems 2,4,3,5. For $E_{0}\geq(1+c)E_{max}$ for any $c>0$, the algorithm achieves recovery. * 1. Compute the eigenvector of $H(T_{0})$ and the leading eigenvalue, denoted $\lambda_{1}$. * 2. (Detection) If $\lambda_{1}>E_{cut}\equiv(E_{0}+E_{max})/2,$ where $E_{0}=\overline{\lambda}(p/2)!{{n_{bos}}\choose p/2}N^{p/2}$ for even $p$, and $E_{0}=\overline{\lambda}^{2}(p-1)!{{n_{bos}}\choose p-1}N^{p}$ for odd $p$, and where $E_{max}$ is defined in theorem 1, then report that $\lambda=\overline{\lambda}$. Otherwise report $\lambda=0$. * 3. (Recovery) Compute the single particle density matrix (defined below) of the leading eigenvector or of any vector in the eigenspace of eigenvalue $\geq E_{cut}$. Apply algorithm 1 to recover an approximation to $v_{\rm sig}$. In section 4, we consider the largest eigenvalue $\lambda_{1}$ of $H(G)$ and show the following theorem which summarizes the results of lemmas 5,7. Roughly, up to prefactors, the result for odd $p$ is given by considering the result for a $q$-th order tensor for even $q=2(p-1)$ and then multiplying the largest eigenvalue by a factor of $\sqrt{N}$. ###### Theorem 1. Let $\lambda_{1}$ be the largest eigenvalue of $G$. For even $p$ let $E_{max}=\sqrt{2J\log(N)}{n_{bos}}^{p/4+1/2}N^{p/4},$ and for odd $p$ let $E_{max}=2\sqrt{J\log(N)}{n_{bos}}^{p/2}N^{p/2},$ where $J$ is a scalar depends that implicitly on $p,{n_{bos}},N$ and tends to some function depending only on $p$ for large ${n_{bos}},N$. More precisely, for even $p$, $J$ is equal to $(p/2)!{{n_{bos}}\choose p/2!}/{n_{bos}}^{p/2}+o(1)$ for the real ensemble and is twice that for the complex ensemble, and for odd $p$, $J$ is equal to that for the even case for $2(p-1)$. Then, for any $x$, assuming Assumption 1, ${\rm Pr}[\lambda_{1}\geq x]\leq\exp\Bigl{(}-\frac{x-E_{max}}{\xi}\Bigr{)},$ (10) with for even $p$ $\xi=\frac{\sqrt{J}{n_{bos}}^{p/4-1/2}N^{p/4}}{\sqrt{2\log(N)}}$ (11) and for odd $p$ $\xi=\frac{\sqrt{J}{n_{bos}}^{p/2-1}N^{p/2}}{\sqrt{\log(N)}}.$ (12) So, for any $E^{\prime}$ which is $\omega(\xi)$, with high probability $\lambda_{1}\leq E_{max}+E^{\prime}$. Now consider the eigenvectors and eigenvalues of $H(T_{0})$. For any symmetric tensor $T$ of order ${n_{bos}}$, let $|T\rangle$ be the vector on ${n_{bos}}$ qudits (each of dimension $N$) with amplitudes given by the entries of the tensor in the obvious way: $|T\rangle=\sum_{\mu_{1},\ldots,\mu_{{n_{bos}}}}T_{\mu_{1},\ldots,\mu_{{n_{bos}}}}|\mu_{1}\rangle\otimes\ldots\otimes|\mu_{{n_{bos}}}\rangle.$ This vector is only normalized if $|T|=1$. So, $\Psi_{\rm sig}\equiv N^{-{n_{bos}}/2}|v_{\rm sig}^{\otimes{n_{bos}}}\rangle$ is a normalized vector. We have the following simple property: ###### Lemma 2. Let $\lambda_{1}$ be the largest eigenvalues of $H(T_{0})$. Then, $\lambda_{1}\geq E\equiv\langle\Psi_{\rm sig}|H(T)|\Psi_{\rm sig}\rangle$. ###### Proof. Immediate from the variational principle. ∎ #### 3.2.1 Even $p$ Case We now show correctness of the algorithm. All results in this subsubsection refer to even $p$ even if we do not state it explicitly. First we estimate $E$ of lemma 2 to show detection. Then, we show recovery. We have $\displaystyle E$ $\displaystyle=$ $\displaystyle\langle\Psi_{\rm sig}|H(\lambda v_{\rm sig}^{\otimes p})|\Psi_{\rm sig}\rangle+\langle\Psi_{\rm sig}|H(G)|\Psi_{\rm sig}\rangle$ (13) $\displaystyle=$ $\displaystyle E_{0}+\langle\Psi_{\rm sig}|H(G)|\Psi_{\rm sig}\rangle,$ (14) where $E_{0}=\lambda(p/2)!{{n_{bos}}\choose p/2}N^{p/2}.$ (15) To evaluate $\langle\Psi_{\rm sig}|H(G)|\Psi_{\rm sig}\rangle$ it is convenient to exploit a rotational invariance of this problem. We can apply a rotation using any matrix $O$ in the orthogonal group $O(N)$, rotating the creation and annihilation operators by making the replacement $a^{\dagger}_{\mu}\rightarrow\sum_{\nu}O_{\mu\nu}a^{\dagger}_{\nu}$ and $a_{\mu}\rightarrow\sum_{\nu}O_{\mu\nu}a_{\nu}.$ This rotation preserves the canonical commutation relations and is equivalent to rotating the basis states on each qudit by $O$. To preserve the Hamiltonian $H(T_{0})$, we rotate each leg of the tensor $T_{0}$: $(T_{0})_{\mu_{1},\ldots,\mu_{p}}\rightarrow\sum_{\nu_{1},\ldots,\nu_{p}}(T_{0})_{\nu_{1},\ldots,\nu_{p}}\prod_{i}O_{\mu_{i},\nu_{i}}.$ This rotation preserves the Gaussian measure on $G$ but changes $v_{\rm sig}$. So, we can rotate so that $v_{\rm sig}$ is some fixed vector, say $(\sqrt{N},0,0,\ldots)$ so $(v_{\rm sig})_{1}=\sqrt{N}$. Then $\langle\Psi_{\rm sig}|H(G)|\Psi_{\rm sig}\rangle$ is equal to $(p/2)!{{n_{bos}}\choose p/2}$ multiplied by a single entry of $G$, i.e., the entries with all indices equal to $1$, which is some quantity chosen from a Gaussian distribution with zero mean and unit variance. So, with high probability, $E=\lambda(p/2)!{{n_{bos}}\choose p/2}N^{p/2}+{n_{bos}}^{p/2}O(N^{\eta})$ for any $\eta>0$. Hence, ###### Theorem 2. If $\lambda(p/2)!{{n_{bos}}\choose p/2}N^{p/2}-E_{max}=\omega\Bigl{(}\frac{\sqrt{J}{n_{bos}}^{p/4-1/2}N^{p/4}}{\sqrt{2\log(N)}}\Bigr{)}$, and if Assumption 1 holds, then with high probability algorithm 2 correctly determines whether $\lambda=0$ or $\lambda=\overline{\lambda}$. ###### Proof. This follows from the estimate of $E$ which lowers bounds the largest eigenvalue when $\lambda=\overline{\lambda}$ and from theorem 1 which upper bounds $\|H(G)\|$. ∎ Given any state, define the single particle density matrix $\rho_{1}$ in the basis for the full Hilbert space used in Eq. (5) to be the reduced density matrix on any of the qudits. Equivalently, this single particle density matrix can be expressed in terms of creation and annihilation operators by $(\rho_{1})_{\mu\nu}=\frac{1}{{n_{bos}}}a^{\dagger}_{\mu}a_{\nu}.$ (16) Note that $\rho_{1}$ is positive semi-definite, Hermitian, and trace $1$. We have ###### Theorem 3. Let Assumption 1 hold. Given any vector $\Psi$ such that $\langle\Psi|H(T_{0})|\Psi\rangle\geq(1+c^{\prime})E_{max}$ for any scalar $c^{\prime}>0$, then with high probability the corresponding single particle density matrix $\rho_{1}$ obeys $\frac{\langle v_{\rm sig}|\rho_{1}|v_{\rm sig}\rangle}{N}\geq(c^{\prime}-o(1))\frac{E_{max}}{E_{0}}.$ (17) In particular, for $\langle\Psi|H(T_{0})|\Psi\rangle\geq E_{cut}$ with $E_{0}\geq(1+c)E_{max}$ we have $c^{\prime}\geq c/2$ so with high probability $\frac{\langle v_{\rm sig}|\rho_{1}|v_{\rm sig}\rangle}{N}\geq\frac{1}{2}\frac{c}{1+c}.$ (18) Hence, algorithm 2 achieves recovery. ###### Proof. We have $\langle\Psi|H(\lambda v_{\rm sig}^{\otimes p}|\Psi\rangle+\langle\Psi|H(G)|\Psi\rangle\geq(1+c)E_{max}$. By theorem 1, with high probability $\langle\Psi|H(G)|\Psi\rangle\leq(1+o(1))E_{max}$. Hence, with high probability $\langle\Psi|H(\lambda v_{\rm sig}^{\otimes p})|\Psi\rangle\geq(c-o(1))E_{max}$. Rotate so that $v_{\rm sig}=(\sqrt{N},0,0,\ldots)$. Then, the Hamiltonian $H(\lambda v_{\rm sig}^{\otimes p})$ is diagonal in the computational basis for the full Hilbert space. Let $P(m)$ be the probability, for state $\Psi$, that exactly $m$ out of ${n_{bos}}$ qudits are in state $|1\rangle$. Then, $\langle\Psi|H(\lambda v_{\rm sig}^{\otimes p})|\Psi\rangle=\sum_{m}\lambda(p/2)!{m\choose p/2}N^{p/2}P(m)\leq\sum_{m}m(E_{0}/{n_{bos}})P(m)=E_{0}\langle v_{\rm sig}|\rho_{1}|v_{\rm sig}\rangle/N$. ∎ Remark: more generally, the same result holds for any mixed state: given any mixed state $\sigma$ such that ${\rm tr}(\sigma H(T_{0}))\geq(1+c)E_{max}$ for any scalar $c>0$, then with high probability the corresponding single particle density matrix $\rho_{1}$ obeys Eq. (17). #### 3.2.2 Odd $p$ Case We now show correctness of the algorithm for odd $p$. All results in this subsubsection refer to odd $p$ even if we do not state it explicitly. Let us first estimate $E$ to show detection. We have $E=\langle\Psi_{\rm sig}|H(T_{0})|\Psi_{\rm sig}\rangle$, but now $H(T_{0})$ is a quadratic function of $T_{0}$ so there are some cross-terms. We have $\langle\Psi_{\rm sig}|H(v_{\rm sig}^{\otimes p})|\Psi_{\rm sig}\rangle=\lambda^{2}(p-1)!{{n_{bos}}\choose p-1}N^{p}\equiv E_{0}.$ (19) The cross-term is $\lambda\langle\Psi_{\rm sig}|\sum_{\mu_{1},\ldots,\mu_{p-1}}\sum_{\nu_{1},\ldots,\nu_{p-1}}\sum_{\sigma}(v_{\rm sig}^{\otimes}p)_{\mu_{1},\mu_{2},\ldots,\mu_{p-1},\sigma}G_{\nu_{1},\nu_{2},\ldots,\nu_{p-1},\sigma}\Bigl{(}\prod_{i=1}^{(p-1)/2}a^{\dagger}_{\mu_{i}}a^{\dagger}_{\nu_{i}}\Bigr{)}\Bigl{(}\prod_{i=(p-1)/2+1}^{p-1}a_{\mu_{i}}a_{\nu_{i}}\Bigr{)}+{\rm h.c.}|\Psi_{\rm sig}\rangle.$ Exploiting the same rotational invariance as in the even $p$ case, this is equal to $2\lambda(p-1)!{{n_{bos}}\choose p-1}N^{p/2}$ multiplied by the real part of a single entry of $G$, i.e., the entry with all indices equal to $1$. So, with high probability, this cross-term is bounded by $\lambda{n_{bos}}^{p-1}O(N^{p/2+\eta})$ for any $\eta>0$. The term quadratic in $G$ is $\frac{1}{2}\langle\Psi_{\rm sig}|\sum_{\mu_{1},\ldots,\mu_{p-1}}\sum_{\nu_{1},\ldots,\nu_{p-1}}\sum_{\sigma}G_{\mu_{1},\mu_{2},\ldots,\mu_{p-1},\sigma}G_{\nu_{1},\nu_{2},\ldots,\nu_{p-1},\sigma}\Bigl{(}\prod_{i=1}^{(p-1)/2}a^{\dagger}_{\mu_{i}}a^{\dagger}_{\nu_{i}}\Bigr{)}\Bigl{(}\prod_{i=(p-1)/2+1}^{p-1}a_{\mu_{i}}a_{\nu_{i}}\Bigr{)}+{\rm h.c.}|\Psi_{\rm sig}\rangle.$ Exploiting the same rotational invariance as before, fixing $v_{\rm sig}=(\sqrt{N},0,0,\ldots)$, we must have all $\mu_{i},\nu_{i}=1$. So, this is a sum of squares of $N$ different entries of $G$, corresponding to $N$ possible choices of $\sigma$; since we use the complex ensemble, this has vanishing mean. So, with high probability this term is ${n_{bos}}^{p-1}O(N^{1/2+\eta})$ for any $\eta>0$. So, with high probability, $E\geq E_{0}-\lambda{n_{bos}}^{p-1}O(N^{p/2+\eta})-{n_{bos}}^{p-1}O(N^{1/2+\eta})$ (20) for any $\eta>0$. Hence, ###### Theorem 4. Let Assumption 1 hold. If $\lambda^{2}(p-1)!{{n_{bos}}\choose p-1}N^{p}-E_{max}=\omega\Bigl{(}\frac{\sqrt{J}{n_{bos}}^{p/2-1}N^{p/2}}{\sqrt{\log(N)}}\Bigr{)}$, then with high probability algorithm 2 correctly determines whether $\lambda=0$ or $\lambda=\overline{\lambda}$. ###### Proof. This follows from the estimate of $E$ in Eq. (20) which lowers bounds the largest eigenvalue when $\lambda=\overline{\lambda}$ and from theorem 1 which upper bounds $\|H(G)\|$. Note that for $\lambda=O(N^{-p/4})$, the terms $-\lambda{n_{bos}}^{p-1}O(N^{p/2+\eta})-{n_{bos}}^{p-1}O(N^{1/2+\eta})$ on the right-hand side of this equation are negligible as they are $o(\Bigl{(}\frac{\sqrt{J}{n_{bos}}^{p/2-1}N^{p/2}}{\sqrt{\log(N)}}\Bigr{)}$. ∎ We now consider recovery. We will first give a more general bound on cross- terms that will be useful. We have $H(T_{0})=H(\lambda v_{\rm sig}^{\otimes p})+H(G)+{\rm cross\,terms}.$ Let us first bound the cross-terms. ###### Lemma 3. With high probability, the operator norm of the cross-terms is bounded by $\lambda N^{p/2}O(N^{(p-1)/4}){n_{bos}}^{p-1}$. ###### Proof. The cross-terms are equal to $2H(T_{cross})$ where $T_{cross}$ is a tensor for even $q=2(p-1)$ which has components $\Bigl{(}T_{cross}\Bigr{)}_{\mu_{1},\ldots,\mu_{(p-1)/2},\nu_{1},\ldots,\nu_{(p-1)/2},\mu_{(p-1)/2+1},\ldots,\mu_{p},\nu_{(p-1)/2},\ldots,\nu_{p}}=\lambda\sum_{\sigma}T_{\mu_{1},\ldots,\mu_{p-1},\sigma}\cdot\Bigl{(}\prod_{i=1}^{p-1}(v_{\rm sig})_{\nu_{i}}\Bigr{)}\cdot(v_{\rm sig})_{\sigma}.$ (21) Rotating so that $v_{\rm sig}=(\sqrt{N},0,0,\ldots)$, we have $\Bigl{(}T_{cross}\Bigr{)}_{\mu_{1},\ldots,\mu_{(p-1)/2},\nu_{1},\ldots,\nu_{(p-1)/2},\mu_{(p-1)/2+1},\ldots,\mu_{p},\nu_{(p-1)/2},\ldots,\nu_{p}}=\lambda N^{p/2}\sum_{\sigma}T_{\mu_{1},\ldots,\mu_{p-1},1}\prod_{i=1}^{p-1}\delta_{\nu_{i},1}.$ (22) Clearly, $\|H(T_{cross})\|\leq{n_{bos}}^{p-1}\|M_{cross}\|,$ where $M_{cross}$ is an $N^{p-1}$-by-$N^{p-1}$ matrix, whose entries are determined by the entries of $T_{cross}$ in an obvious way so that the first $p-1$ indices of $T_{cross}$ determine a column index in $M$ and the last $p-1$ indices of $T_{cross}$ determine a row index of $M$. Regard $T_{\mu_{1},\ldots,\mu_{p-1},1}$ as being the entries of some tensor of order $p-1$ and let $M^{\prime}$ by the $N^{(p-1)/2}$-by-$N^{(p-1)/2}$ matrix whose entries are determined by the entries of this tensor, again in the obvious way so that the first $(p-1)/2$ indices of the tensor determine a column index and the last $(p-1)/2$ indices determine a row index. Then, $\|M_{cross}\|=\lambda N^{p/2}\|M^{\prime}\|.$ (23) However, since the entries of $M^{\prime}$ are independent Gaussian random variables, it follows from standard random matrix theory results [12] that with high probability $\|M^{\prime}\|=O(N^{(p-1)/4})$. Remark: the dependence on ${n_{bos}}$ is not tight here. ∎ So as in the even case, we have ###### Theorem 5. Let Assumption 1 hold. Given any vector $\Psi$ such that $\langle\Psi|H(T_{0})|\Psi\rangle\geq(1+c^{\prime})E_{max}$ for any scalar $c^{\prime}>0$, then with high probability the corresponding single particle density matrix $\rho_{1}$ obeys $\frac{\langle v_{\rm sig}|\rho_{1}|v_{\rm sig}\rangle}{N}\geq(c^{\prime}-o(1))\frac{E_{max}}{E_{0}}.$ (24) In particular, for $\langle\Psi|H(T_{0})|\Psi\rangle\geq E_{cut}$ with $E_{0}\geq(1+c)E_{max}$ we have $c^{\prime}\geq c/2$ so with high probability $\frac{\langle v_{\rm sig}|\rho_{1}|v_{\rm sig}\rangle}{N}\geq\frac{1}{2}\frac{c}{1+c}.$ (25) Hence, algorithm 2 achieves recovery. ###### Proof. We use lemma 3 to bound $\|H(T_{cross})\|$. This bound is asymptotically negligible compared to $E_{max}$. So, we have $\langle\Psi|H(\lambda v_{\rm sig}^{\otimes p}|\Psi\rangle+\langle\Psi|H(G)|\Psi\rangle\geq(1+c-o(1))E_{max}$ where the $o(1)$ denotes the $o(1)$ contribution from the cross-terms. Then, the rest of the proof is the same as in the even case, except for a replacement of $p/2$ by $p-1$. In detail: by theorem 1, with high probability $\langle\Psi|H(G)|\Psi\rangle\leq(1+o(1))E_{max}$. Hence, $\langle\Psi|H(\lambda v_{\rm sig}^{\otimes p}|\Psi\rangle\geq(c-o(1))E_{max}$. Rotate so that $v_{\rm sig}=(\sqrt{N},0,0,\ldots)$. Then, the Hamiltonian $H(\lambda v_{\rm sig}^{\otimes p})$ is diagonal in the computational basis for the full Hilbert space. Let $P(m)$ be the probability, for state $\Psi$, that exactly $m$ out of ${n_{bos}}$ qudits are in state $|1\rangle$. Then, $\langle\Psi|H(\lambda v_{\rm sig}^{\otimes p})|\Psi\rangle=\sum_{m}\lambda^{2}(p-1)!{m\choose p-1}N^{p}P(m)\leq\sum_{m}m(E_{0}/{n_{bos}})P(m)=E_{0}\langle v_{\rm sig}|\rho_{1}|v_{\rm sig}\rangle/N$. ∎ ## 4 Spectrum of Random Hamiltonian In this section, we will estimate the eigenvalues of $H(G)$. We consider first the case of even $p$. Here our proof is very similar to the of Ref. [1], though the method here also suggests some heuristics that may lead to a tighter bound in the future. Then, we consider the case of odd $p$ by reducing it to the case of even $p$. ### 4.1 Even $p$ We first consider the case of even $p$. Let $Z(\tau,p,N,{n_{bos}})$ denote $\mathbb{E}[{\rm tr}(\exp\\{\tau H(G)\\})]$ for a tensor $G$ of rank $p$, with entries of $G$ chosen from the Gaussian distribution, with given $N,{n_{bos}}$, for some real scalar $\tau$. In this subsection, $G$ may be symmetrized or not, and may be chosen from either the real or complex ensemble. The main result in this section is the following lemma: ###### Lemma 4. For each ensemble, real or complex, symmetrized or not, we have $Z(\tau,p,N,{n_{bos}})\leq D(N,{n_{bos}})\exp(\tau^{2}(J/2){n_{bos}}^{p/2}N^{p/2}),$ (26) where $J$ is a scalar depends that implicitly on $p,{n_{bos}},N$ and tends to some function depending only on $p$ for large ${n_{bos}},N$. More precisely, $J$ is equal to $(p/2)!{{n_{bos}}\choose p/2!}/{n_{bos}}^{p/2}+o(1)$ for the real ensemble and is twice that for the complex ensemble. ###### Proof. We first give a brief derivation of Eq. (4) below which is a standard result using quantum field theory techniques. Note that $\tau H(G)=H(\tau G)$ and $\tau G$ is a tensor with entries chosen from a Gaussian distribution with zero mean and variance $\tau^{2}$. Hence for any $\tau^{\prime}>\tau$ we have $Z(\tau^{\prime},p,N,{n_{bos}})=\mathbb{E}_{G,\eta}[{\rm tr}(\exp\\{H(\tau G+\eta)\\})],$ (27) where $\mathbb{E}_{G,\eta}[\ldots]$ denotes the expectation value over $G$ and $\eta$, with the tensor $\eta$ having Gaussian entries with zero mean and variance $(\tau^{\prime})^{2}-\tau^{2}$. Taking the expectation value over $\eta$, for $\tau^{\prime 2}-\tau^{2}$ small, we need to keep only the zeroth and second order terms on the right-hand side. So, we find that $\displaystyle\partial_{\tau}^{2}Z(\tau,p,N,{n_{bos}})$ $\displaystyle=$ $\displaystyle\int_{0}^{1}{\rm d}s_{2}\int_{0}^{s_{2}}{\rm d}s_{1}\,\mathbb{E}_{G,\eta}[{\rm tr}\Bigl{(}\exp\Bigl{\\{}(1-s_{2})\tau H(G)\Bigr{\\}}H(\eta)\exp\Bigl{\\{}(s_{2}-s_{1})\tau H(G)\Bigr{\\}}H(\eta)\exp\Bigl{\\{}s_{1}\tau H(G)\Bigr{\\}}\Bigr{)}].$ Using cyclic properties of the trace, this can be simplified to $\partial_{\tau}^{2}Z(\tau,p,N,{n_{bos}})=\frac{1}{2}\int_{0}^{1}{\rm d}s_{1}\,\mathbb{E}_{G,\eta}[{\rm tr}\Bigl{(}\exp\Bigl{\\{}(1-s_{1})\tau H(G)\Bigr{\\}}H(\eta)\exp\Bigl{\\{}s_{1}\tau H(G)\Bigr{\\}}H(\eta)\Bigr{)}].$ (29) We now use a general result. Consider any Hermitian $H$ (we will use $H=\tau H(G)$) and any operator $O$ (we will use $O=H(\eta)=H(\eta)^{\dagger}$) and any $s_{1}\in[0,1]$. We claim that ${\rm tr}\Bigl{(}\exp(H)O^{\dagger}O\Bigr{)}+O\leftrightarrow O^{\dagger}\geq{\rm tr}\Bigl{(}\exp((1-s_{1})H)O^{\dagger}\exp(s_{1}H)O\Bigr{)}+O\leftrightarrow O^{\dagger},$ where $O\leftrightarrow O^{\dagger}$ indicates the previous term with $O,O^{\dagger}$ interchanged. Proof of claim: work in an eigenbasis of $H$. It suffices to consider the case that $O=|b\rangle\langle a|$ where $|a\rangle,|b\rangle$ are eigenvectors of $H$ with eigenvalues $E_{a},E_{b}$. Then the the right-hand side is equal to $\exp(s_{1}E_{a}+(1-s_{1})E_{b})+\exp(s_{1}E_{b}+(1-s_{1})E_{a})=2\exp((E_{a}+E_{b})/2)\cosh((s_{1}-1/2)(E_{b}-E_{a})$. The cosh function is maximized on the interval $[0,1]$ at $s_{1}=0,1$ when the right-hand side becomes equal to the left-hand side. So, $\partial_{\tau}^{2}Z(\tau,p,N,{n_{bos}})\leq\frac{1}{2}\mathbb{E}_{G,\eta}[{\rm tr}\Bigl{(}\exp(\tau H(G))H(\eta)^{2}\Bigr{)}].$ (30) For the real ensemble without symmetrization, we have $\mathbb{E}[H(\eta)^{2}]=\sum_{\mu_{1},\ldots,\mu_{p/2}}\sum_{\nu_{1},\ldots,\nu_{p/2}}\prod_{i=1}^{p/2}\Bigl{(}a^{\dagger}_{\mu_{i}}a_{\nu_{i}}\Bigr{)}\prod_{i=1}^{p/2}\Bigl{(}a^{\dagger}_{\nu_{i}}a_{\mu_{i}}\Bigr{)}\Bigr{.}$ To leading order in $N$, we may approximate $\sum_{\nu}a_{\nu}a^{\dagger}_{\nu}=N$ so on the given Hilbert space, $\mathbb{E}[H(\eta)^{2}]$ is a scalar equal to $(p/2)!{{n_{bos}}\choose p/2}N^{p/2}+O(N^{p/2-1})$. In general, for the complex or real ensemble, symmetrized or not, we find that $\mathbb{E}[H(\eta)^{2}]$ is a scalar equal to $J(p/2)!{{n_{bos}}\choose p/2}N^{p/2}$, where $J$ obeys the claims of the lemma. To verify that $\mathbb{E}[H(\eta)^{2}]$ is a scalar and to compute the scalar to all orders in $N$, commute the annihilation operators $a_{\nu}$ to the right. The result is some linear combination of operators with all annihilation operators to the right of creation operators which can be written in the form $\sum_{\mu_{1},\ldots,\mu_{k}}(\prod_{i=1}^{k}a^{\dagger}_{\mu_{k}})(\prod_{i=1}^{k}a_{\mu_{k}})$, and each such operator is equal to $k!{{n_{bos}}\choose k}$. Hence, from Eq. (30), $\partial_{\tau}^{2}\log(Z(\tau,p,N,{n_{bos}})\leq(J/2){n_{bos}}^{p/2}N^{p/2}$. ∎ Remark: this result is clearly not tight in the regime where random matrix theory is accurate (${n_{bos}}=p/2$). It is interesting to see what happens there. The correlation function $\mathbb{E}_{G,\eta}[{\rm tr}\Bigl{(}\exp\\{(1-s_{1})\tau H(G)\\}H(\eta)\exp\\{s_{1}\tau H(G)\\}H(\eta)\Bigr{)}]$ is not independent of $s_{1}$, but rather decays as a function of $s_{1}$ for $s_{1}\leq 1/2$ (of course, it increases again as $s_{1}$ becomes larger than $1/2$). Considering the regime $s_{1}\ll 1/2$, using the square-root singularity at the edge of the Wigner semi-circle we can estimate that it decays as $s_{1}^{-3/2}$. This means that the integral of this correlation function over $s_{1}$ is dominated by its value for small $s_{1}$ of order $1/\tau$ so that for $\tau$ large compared to the inverse width of the semi-circle (though of course $\tau$ not too large) the integral becomes of order $1/\tau$. This is stronger than the upper bounds here where we have bounded by the integral by something independent of $\tau$. We may guess that a tighter analysis will show that a similar effect will happen in the case of ${n_{bos}}\gg p/2$; however, an important difference occurs. If we take an eigenstate of $H(G)$ with some eigenvalue $\lambda_{0}$, and apply $H(\eta)$ for random $\eta$, this only changes $p/2$ out of the ${n_{bos}}$ qudits in the state. So, one might guess that the resulting state will have expectation value of $H(G)$ that is $\lambda_{0}(1-p/(2{n_{bos}}))$ rather than an (as in the random matrix case) expectation value of $H(G)$ which is zero. So, we may guess that the correlation function will be non-negligible for $s_{1}\lesssim({n_{bos}}/p)\tau^{-1}$. A heuristic estimate in this fashion suggests that the lemma below for the eigenvalue is tight up to logarithmic factors. From lemma 4, the following lemma is an immediate corollary: ###### Lemma 5. Let $\lambda_{1}$ be the largest eigenvalue of $G$. Let $E_{max}=\sqrt{2J\log(N)}{n_{bos}}^{p/4+1/2}N^{p/4}.$ Then, for any $x$, ${\rm Pr}[\lambda_{1}\geq x]\leq\exp\Bigl{(}-\frac{x-E_{max}}{\xi}\Bigr{)},$ (31) with $\xi=\frac{\sqrt{J}{n_{bos}}^{p/4-1/2}N^{p/4}}{\sqrt{2\log(N)}}$ (32) So, for any $E^{\prime}$ which is $\omega(\xi)$, with high probability $\lambda_{1}\leq E_{max}+E^{\prime}$. ###### Proof. We have ${\rm tr}(\exp\\{\tau H(G)\\})\geq\exp(\tau\lambda_{1})$. Hence, for any $x$, ${\rm Pr}[\lambda_{1}\geq x]\leq Z(\tau,p,N,{n_{bos}})/\exp(\tau x)$. Since $D(N,{n_{bos}})\leq N^{n_{bos}}$, minimizing over $\tau$, we find that ${\rm Pr}[\lambda_{1}\geq x]\leq\exp\Bigl{(}{n_{bos}}\log(N)-\frac{x^{2}}{2J{n_{bos}}^{p/2}N^{p/2}}\Bigr{)}.$ For $x=E_{max}$, the right-hand is equal to $1$ and for $x>E_{max}$ the right- hand side decays exponentially. ∎ ### 4.2 Odd $p$ We now consider the case of odd $p$. Let $Z(\tau,p,N,{n_{bos}})$ denote $\mathbb{E}[{\rm tr}(\exp(\tau H(G)))]$ for a tensor $G$ of rank $p$, with entries of $G$ chosen from the Gaussian distribution, with given $N,{n_{bos}}$. In this subsection, $G$ is complex and not symmetrized. The Hamiltonian $H(G)$ is given by Eq. (6) for even $p$ and by Eq. (3.1) for odd $p$. We will reduce the calculation for odd $p$ to the case for even $p$, up to a bounded error, showing the following ###### Lemma 6. For odd $p$, for ${n_{bos}}^{p-1}\tau N^{1/3}=o(1)$, $Z(\sqrt{2N}\tau,2(p-1),N,{n_{bos}})\leq Z(\tau,p,N,{n_{bos}})\leq Z(\sqrt{2N}\tau,2(p-1),N,{n_{bos}})\exp(o(N^{(p-1)/2})).$ (33) All occurrences of $Z(\cdot)$ in the above equation refer to the complex ensemble without symmetrizing. Remark: the assumptions require that $\tau$ is $o(1)$; however, that is still large enough $\tau$ to be useful later in bounding the spectrum since the largest eigenvalues of $H(G)$ are typically large compared to $1$. ###### Proof. For arbitrary $H(G)$, the exponential $\exp\\{\tau H(G)\\}$ can be expanded as a series $1+\tau H(G)+\frac{\tau^{2}}{2}H(G)^{2}+\ldots.$ We bring the expectation value (over $G$) inside the trace, and compute the expectation value of this series. This expectation value can be computed using Wick’s theorem to compute the expectation value of a moment of a Gaussian distribution. For odd $p$, each term in $H$ is a sum of two terms, one depending on the product of two tensors $G$ and one depending on the product of two tensors $\overline{G}$, where the overline denotes complex conjugation. Hence, the $m$-th order term $\frac{\tau^{m}}{m!}H(G)^{m}$ is a sum of $2^{m}$ terms, corresponding to $m$ different choices of $G$ or $\overline{G}$ in each $H(G)$); we call each such choice a history. Each term is a product of creation and annihilation operators depending on a product of $2m$ tensors, some of which are $G$ and some of which are $\overline{G}$. This expectation value of such a term is non-vanishing only if there are $m$ tensors $G$ and $m$ tensors $\overline{G}$. In that case, the expectation value is given by summing over ways of pairing each tensor $G$ with a distinct tensor $\overline{G}$. There are $m!$ such pairings. Then, for each such pairing, one computes an operator by taking the operator $\frac{\tau^{m}}{m!}H(G)^{m}$ and replacing, for every pair, the two tensors $G_{\mu_{1},\ldots,\mu_{p}}\overline{G}_{\nu_{1},\ldots,\nu_{p}}$ in that pair with a product of $\delta$-functions $\prod_{a=1}^{p}\delta_{\mu_{a},\nu_{a}}=\mathbb{E}[G_{\mu_{1},\ldots,\mu_{p}}\overline{G}_{\nu_{1},\ldots,\nu_{p}}].$ Summing this operator over pairings gives the expectation value. Note also that here we have not symmetrized $G$. If we were to symmetrize $G$, then $\mathbb{E}[G_{\mu_{1},\ldots,\mu_{p}}\overline{G}_{\nu_{1},\ldots,\nu_{p}}]$ is given by $1/p!$ times a sum of $p!$ different products of $\delta$-functions, i.e. $\sum_{\pi}\delta_{\mu_{a},\nu_{\pi(a)}}$, where $\pi$ is a permutation. This would lead to additional terms that we need to compute and would make the analysis more difficult (though in practice may lead to better performance). For given $m$, given history and given pairing, let us define a cluster as follows: define a graph with $2m$ vertices, each vertex corresponding to a single tensor, either $G$ or $\overline{G}$. Let there be an edge between any two tensors which both appear in the same term in the Hamiltonian. Let there also be an edge between any two tensors which are in a pair. Hence, this is a graph of degree $2$ (we allow the possibility of multiple edges connecting two vertices if we pair two tensor which both appear in the same term in the Hamiltonian). A cluster is a connected component of this graph. We refer to a cluster containing four vertices as a minimal cluster. A cluster with six or more vertices is called a non-minimal cluster. Note that there are no clusters containing only two terms because each term in $H(G)$ depends on a product of two tensors $G$ or two tensors $\overline{G}$; if we had not taken the complex ensemble and instead taken $G$ to be real, we would instead have these clusters with two terms. We discuss the case of the real ensemble further after the proof of the lemma. The minimal clusters will turn out to give the dominant order contribution in an expansion in $N$. The non-minimal clusters will be subleading order. We have expressed $\frac{\tau^{m}}{m!}H(G)^{m}$ as a sum over histories and pairings and so $\mathbb{E}[\exp(\tau H(G))]$ is a sum over $m$, histories, and pairings. Each term in the sum over $m$, histories, and pairings for $\mathbb{E}[\exp(\tau H(G))]$ is an operator. Hence, $Z(\tau,p,N,{n_{bos}})$ is also given by a sum over $m$, histories, and pairings as one may take the trace of each term in $\mathbb{E}[\exp(\tau H(G))]$. Note that each $m$, history, and pairing gives a non-negative contribution to $Z(\tau,p,N,{n_{bos}})$, i.e., it has a non-negative trace. Lower Bound— Now we prove the first inequality in Eq. (33), namely that $Z(\sqrt{2N}\tau,2(p-1),N,{n_{bos}})\leq Z(\tau,p,N,{n_{bos}})$. To do this, we first define a similar summation over $m$, histories, and pairings to compute $Z(\tau,q,N,{n_{bos}})$ for even $q=2(p-1)$ and then we compare the two summations. For even $q$, each term in $H$ is a sum of two terms, one depending linearly on a tensor $G$ and one depending on tensor $\overline{G}$. As before, the $m$-th order term $\frac{\tau^{m}}{m!}H(G)^{m}$ is a sum of $2^{m}$ histories, with each history corresponding to $m$ different choices of $G$ or $\overline{G}$ in each $H(G)$. Each term is a product of creation and annihilation operators depending on a product of $m$ tensors, some of which are $G$ and some of which are $\overline{G}$; note the difference here from the odd case as now there are only $m$ tensors, rather than $2m$. This expectation value of such a term is non-vanishing only if there are $m/2$ tensors $G$ and $m/2$ tensors $\overline{G}$. In that case, the expectation value is given as before by summing over pairings and replacing the two tensors in the pair with $\delta$-functions. We claim that if we consider $Z(\tau,p,N,{n_{bos}})$ for odd $p$ and consider only the sum of pairings for which all clusters are minimal, this gives precisely the sum of terms for $Z(\sqrt{2N}\tau,2(p-1),N,{n_{bos}})$. Since all terms contribute non-negatively to the trace, this will prove the first inequality. To show this claim, let $L_{2(p-1)}$ label a given choice of $m$, history, and pairing for $Z(\sqrt{2N}\tau,2(p-1),N,{n_{bos}})$ and let $L_{p}$ label a given choice of $m$, history, and clusters for $Z(\tau,p,N,{n_{bos}})$ for which all clusters are minimal. There are $2^{m/2}$ different pairings for the given choice of clusters labelled by $L_{p}$. We will construct a one-to-one correspondence between $L_{p}$ and $L_{2(p-1)}$ and show that the sum of the terms labelled by $L_{p}$ is equal to the term labelled by $L_{2(p-1)}$, i.e., that they give the same operator, and hence have the same trace. Consider given $L_{p}$. Then, $L_{2(p-1)}$ is as follows. The value of $m$ is the same. The history is also the same: for a given sequence of choices of $G$ or $\overline{G}$ for $L_{p}$, we use the same sequence for $L_{2(p-1)}$. Note that in $L_{p}$, each of the $m$ choices of $G$ or $\overline{G}$ denotes a choice that both tensors in $H(G)$ are equal to $G$ or both are equal to $\overline{G}$, while in $L_{2(p-1)}$ each of the $m$ choices is only a choice about a single tensor in $H(G)$. The pairing is as follows. We introduce notation to label the tensors in $H(G)^{m}$. For even $p$, we label the tensors by an integer $i\in\\{1,2,\ldots,m\\}$ depending which of the $m$ factors of $H(G)$ it is in. Here, we mean that the tensors appear in the product $H(G)^{m}=H(G)\cdot H(G)\cdot\ldots\cdot H(G)$; we label each of the different factors in the product $1,2,\ldots,m$ in sequence. For odd $p$, we label the tensors by a pair $(i,w)$ for $i\in\\{1,2,\ldots,m\\}$ and $w\in\\{1,2\\}$. The integer $i$ labels which of the $m$ factors of $H(G)$ it is in and $w$ labels whether it is the first or second of the two tensors in $H(G)$. For a pairing for $Z(\tau,p,N,{n_{bos}})$ with all clusters minimal, then each cluster is of the form that for some $i,j$ that we pair $(i,1)$ with $(j,1)$ and $(i,2)$ with $(j,2)$ or of the form that we pair $(i,1)$ with $(j,2)$ and $(i,2)$ with $(j,1)$. For a given choice of clusters, the corresponding pairing for $L_{2(p-1)}$ pairs $i$ with $j$. We sketch the claim about the terms. For a term in $L_{p}$, for each cluster we have two tensors $T$ and two tensors $\overline{T}$. We replace these tensors with the expectation value $\mathbb{E}[\sum_{\mu_{1},\ldots,\mu_{p-1}}\sum_{\nu_{1},\ldots,\nu_{p-1}}\sum_{\sigma}T_{\mu_{1},\mu_{2},\ldots,\mu_{p-1},\sigma}T_{\nu_{1},\nu_{2},\ldots,\nu_{p-1},\sigma}\sum_{\alpha_{1},\ldots,\alpha_{p-1}}\sum_{\beta_{1},\ldots,\beta_{p-1}}\sum_{\overline{\sigma}}\overline{T}_{\alpha_{1},\alpha_{2},\ldots,\alpha_{p-1},\overline{\sigma}}\overline{T}_{\beta_{1},\beta_{2},\ldots,\beta_{p-1},\overline{\sigma}}],$ which is equal to some product of $\delta$-functions. This expectation value then multiplies the operators $\Bigl{(}\prod_{i=1}^{(p-1)/2}a^{\dagger}_{\mu_{i}}a^{\dagger}_{\nu_{i}}\Bigr{)}\Bigl{(}\prod_{i=(p-1)/2+1}^{p-1}a_{\mu_{i}}a_{\nu_{i}}\Bigr{)}$ and $\Bigl{(}\prod_{i=1}^{(p-1)/2}a^{\dagger}_{\alpha_{i}}a^{\dagger}_{\beta_{i}}\Bigr{)}\Bigl{(}\prod_{i=(p-1)/2+1}^{p-1}a_{\alpha_{i}}a_{\beta_{i}}\Bigr{)}$ inserted at the appropriate places into the product $H(G)^{m}$. The $\delta$-functions constrain $\sigma=\overline{\sigma}$; summing over this gives a factor of $N$ for each cluster, while there are two pairings for each cluster giving another factor of $2$ for each cluster. The number of clusters is equal to $m/2$, giving an overall factor $(2N)^{m/2}$. This proves the first inequality. Upper bound— Now we prove the second inequality in Eq. (33). To do this, we define the following quantity for $q$ which is a multiple of $4$ (note that $q=2(p-1)$ is a multiple of $4$ if $p$ is odd): $Z^{\prime}(\tau,\tau^{\prime},q,N,{n_{bos}})\equiv\mathbb{E}[{\rm tr}(\exp\\{\tau H(G)+\tau^{\prime}H(G^{\prime})^{2}\\})],$ (34) where the expectation value is over tensors $G$ of order $q$ chosen from the complex ensemble and tensors $G^{\prime}$ of order $q/2$ chosen also from the complex ensemble (note that $q/2$ is even). Note that we square $H(G^{\prime})$ in the exponent. We will prove that $Z(\tau,p,N,{n_{bos}})\leq Z^{\prime}(\sqrt{2N}\tau,\tau^{\prime},2(p-1),N,{n_{bos}})$ (35) for $\tau^{\prime}=N^{1/3}\tau$. From this, the second inequality in Eq. (33) follows. To see this, we have, for any $G^{\prime}$, $\|H(G^{\prime})^{2}\|\leq O({n_{bos}}^{p-1})\|G^{\prime}\|^{2}$ where the operator norm $\|H(G^{\prime})^{2}\|$ denotes the largest eigenvalue in absolute value and $\|G^{\prime}\|$ denotes the largest singular value of $G^{\prime}$, regarding $G^{\prime}$ as a matrix of size $N^{q/4}$-by-$N^{q/4}$. So, by the Golden-Thompson inequality, $\displaystyle\mathbb{E}_{G}[{\rm tr}(\exp\\{\tau H(G)+\tau^{\prime}H(G^{\prime})^{2}\\})]$ $\displaystyle\leq$ $\displaystyle\mathbb{E}_{G}[{\rm tr}(\exp\\{\tau H(G)\\})]\mathbb{E}_{G^{\prime}}[\exp(O({n_{bos}}^{p-1})\tau^{\prime}\|G^{\prime}\|^{2})]$ $\displaystyle=$ $\displaystyle Z(\sqrt{2N}\tau,2(p-1),N,{n_{bos}})\mathbb{E}_{G^{\prime}}[\exp(O({n_{bos}}^{p-1})\tau^{\prime}\|G^{\prime}\|^{2})].$ where the subscript $G$ or $G^{\prime}$ denotes the expectation over $G$ or $G^{\prime}$. The matrix $G^{\prime}$ has complex entries; we can bound its operator norm by the sum of the operator norms of its Hermitian and anti- Hermitian parts. For ${n_{bos}}^{p-1}\tau^{\prime}=o(1)$, we have $\mathbb{E}_{G^{\prime}}[\exp(O({n_{bos}}^{p-1})\tau^{\prime}\|G^{\prime}\|^{2})]=\exp[O({n_{bos}}^{p-1})\tau^{\prime}O(N^{q/4})]$ as can be shown using Hermite polynomials [12] to compute the probability distribution of the eigenvalues of a random matrix and using the decay of a Hermite polynomial multiplying a Gaussian (the study of the largest eigenvalue of a random matrix is a rich field if we consider the probability distribution near the edge of the Wigner semi-circle but here we are satisfied to consider the probability distribution for eigenvalues which are some constant factor $>1$ multiplying the edge of the semi-circle). To show Eq. (35), let $L_{p}$ label the set of terms for $Z(\tau,p,N,{n_{bos}})$ with a given choice of $m$, history, clusters (the clusters need not be minimal), and choice of pairing for all tensors in non- minimal clusters (we do not specify the pairing of the tensors in the minimal clusters; if there are $n_{min}$ minimal clusters then there are $2^{n_{min}}$ terms in the set). Let $L^{\prime}_{2(p-1)}$ label a term for $Z^{\prime}(\sqrt{2N}\tau,\tau^{\prime},2(p-1),N,{n_{bos}})$ with given $m$, history, clusters, and pairing. Here, for $Z^{\prime}$, a history corresponds to $m$ choices of either $H(G)$ or $H^{\prime}(G)^{2}$ and also to choices of $G$ or $\overline{G}$, or $G^{\prime}$ or $\overline{G}^{\prime}$, for each $H(G)$ or $H(G^{\prime})$. We define a map from choice of $L_{p}$ to choice of $L^{\prime}_{2(p-1)}$ so that the sum of terms labelled by $L_{p}$ is equal to the term labelled by the corresponding $L^{\prime}_{2(p-1)}$. This map will be one-to-one but it will not be onto. However, since all terms are non-negative, this will establish the needed inequality. The map is as follows. The $m$ will be the same. Using the notation above, if tensor $(i,1)$ is in a minimal cluster (and hence, so is $(i,2)$) in the set labelled by $L_{p}$, then in the history for $L^{\prime}_{2(p-1)}$ for the $i$-th term we choose $\tau H(G)$, while if $(i,1)$ is not in a minimal cluster, then we choose $\tau^{\prime}H(G^{\prime})^{2}$. Given a history labelled by $L_{p}$, corresponding to a choice of tensor or its complex conjugate (i.e., either $G$ or $\overline{G}$) in each of the $m$ different $H(G)$ in a term in the series for $Z(\tau,p,N,{n_{bos}})$, then we make the same choice of tensor or its complex conjugate in each of the $m$ different $H(G)$ or $H(G^{\prime})^{2}$ in the history labelled by $L^{\prime}_{2(p-1)}$. That is, if we choose $G$ in the $i$-th $H(G)$ in some terms in the series for $Z(\tau,p,N,{n_{bos}})$, then we choose $G$ in $H(G)$ or $G^{\prime}$ in both terms in $H(G^{\prime})^{2}$ and similarly if we choose $\overline{G}$ we choose $\overline{G}$ in $H(G)$ and $\overline{G}^{\prime}$ in both terms in $H(G^{\prime})^{2}$. We label the tensors in the expansion for $Z^{\prime}(\sqrt{2N}\tau,\tau^{\prime},2(p-1),N,{n_{bos}})$ either by an integer $i$, if it appears in $H(G)$, or by a pair $(i,w)$, if it appears in $H(G^{\prime})^{2}$, in which case the index $w\in\\{1,2\\}$ labels which of the two $H(G^{\prime})$ it is in. Finally, we define the pairing labelled by $L^{\prime}_{2(p-1)}$. For a minimal cluster labelled by $L_{p}$, pairing $(j,1)$ and $(i,2)$ with $(j,2)$ or $(i,1)$ with $(j,2)$ and $(i,2)$ with $(j,1)$, then in the pairing labelled by $L^{\prime}_{2(p-1)}$ we pair $i$ with $j$.. If a cluster is non-minimal, then we simply use the same pairing for the corresponding tensors in $L^{\prime}_{2(p-1)}$. That is, suppose a cluster pairs $(i_{1},w_{1})$ with $(i_{2},w^{\prime}_{2})$, and pairs $(i_{2},w_{2})$ with $(i_{3},w^{\prime}_{3})$, and so on, where $w^{\prime}_{a}=1$ if $w_{a}=2$ and $w^{\prime}_{a}=2$ if $w_{a}=1$. Then, we also pair $(i_{1},w_{1})$ with $(i_{2},w^{\prime}_{2})$, and pairs $(i_{2},w_{2})$ with $(i_{3},w^{\prime}_{3})$, and so on. The smallest non-minimal cluster has six vertices. In every cluster, minimal or not, there is a sum over some index (for example, the sum over the index $\sigma=\sigma^{\prime}$ in the lower bound calculation above) which gives a factor $N$. Thus, taking $\tau^{\prime}=\tau N^{1/3}$ accounts for this factor. No factor of $2$ occurs for the non-minimal clusters. ∎ Remark: since we have chosen the entries of $G$ from the complex ensemble, we have that $\mathbb{E}[H(G)]=0$ for $p$ odd. If instead we had chosen the entries of $G$ from the real ensemble we would have (considering the specific case $p=3$ for simplicity and not symmetrizing $G$, again for simplicity) a non-vanishing expectation value since $\mathbb{E}[T_{\mu_{1},\mu_{2},\sigma}T_{\nu_{1},\nu_{2},\sigma}]=N\delta_{\mu_{1},\nu_{1}}\delta_{\mu_{2},\nu_{2}},$ (37) so that $\mathbb{E}[H(G)]\propto N\sum_{\mu_{1},\mu_{2}}a^{\dagger}_{\mu_{2}}a^{\dagger}_{\mu_{2}}a_{\mu_{1}}a_{\mu_{1}}$. Such a term is sometimes called a pairing term or a “cooperon” in the study of disordered system in physics. In the case ${n_{bos}}=2$ (the smallest possible for $p=3$), this term has operator norm $N^{2}$. This is much larger than $N^{1/2}$ times the expected operator norm of $H(G)$ for $p=2(p-1)=4$, i.e., that expected operator norm is proportional to $N$ by random matrix theory, and $N^{2}\gg N^{3/2}$. There may be other ways to deal with this non-vanishing expectation value other than using the complex ensemble. One way is to use the real ensemble, but to consider the Hamiltonian $H(T_{0})-M$, where we define $M=N\sum_{\mu_{1},\mu_{2}}a^{\dagger}_{\mu_{2}}a^{\dagger}_{\mu_{2}}a_{\mu_{1}}a_{\mu_{1}}$ for $p=3$. In this case, the added term $-M$ cancels the expectation value of $H(G)$ term-by-term in the perturbation expansion. However, if we do this we still have some additional terms when we consider clusters of four or more vertices. We expect that the clusters of six or more vertices are still negligible, but the structure of the clusters of four vertices becomes more complicated. We leave the analysis of this case for the future, but we expect that it would work and may be practically useful. From lemma 6, the following lemma is an immediate corollary: ###### Lemma 7. Let $\lambda_{1}$ be the largest eigenvalue of $G$. Let $E_{max}=2\sqrt{J\log(N)}{n_{bos}}^{p/2}N^{p/2},$ where $J$ is the $J$ of lemma 4 for $Z(\tau,2(p-1),N,{n_{bos}})$. Then, for any $x$, ${\rm Pr}[\lambda_{1}\geq x]\leq\exp\Bigl{(}-\frac{x-E_{max}}{\xi}\Bigr{)},$ (38) with $\xi=\frac{\sqrt{J}{n_{bos}}^{p/2-1}N^{p/2}}{\sqrt{\log(N)}}.$ (39) So, for any $E^{\prime}$ which is $\omega(\xi)$, with high probability $\lambda_{1}\leq E_{max}+E^{\prime}$. ###### Proof. From lemmas 4,6, for ${n_{bos}}^{p-1}\tau N^{1/3}=o(1)$, we have $Z(\tau,p,N,{n_{bos}})\leq\exp(\tau^{2}J{n_{bos}}^{p-1}N^{p})\exp(o(N^{(p-1)/2})).$ Let $\tau={n_{bos}}^{1-p/2}N^{-p/2}\sqrt{\log(N)}/\sqrt{J}$. For ${n_{bos}}=o(N^{(p/2-1/3)/(1+p/2)}\log(N)^{1/(2+p)}$, the condition ${n_{bos}}^{p-1}\tau N^{1/3}=o(1)$ holds. So, after some algebra, for any $x$, the result follows. ∎ It is worth noting that lemma 6 has the following corollary: ###### Corollary 2. For odd $p$, for ${n_{bos}}=p-1$, with high probability $\lambda_{1}=O(N^{p/2})$. ###### Proof. To prove this we use the existence of tighter bounds for even $q=2(p-1)$ when ${n_{bos}}=q/2$. By Eq. (33), for ${n_{bos}}^{p}\tau N^{1/3}=o(1)$, we have $Z(\tau,p,N,{n_{bos}})\leq Z(\sqrt{2N}\tau,2(p-1),N,{n_{bos}})\exp(o(N^{(p-1)/2}))$. Since we are considering fixed ${n_{bos}}$, this holds for $\tau N^{1/3}=o(1)$. We have that $Z(\sqrt{2N}\tau,2(p-1),N,p-1)=\mathbb{E}_{G}[\exp(\sqrt{2N}\tau H(G)]$, but the Hamiltonian $H(G)$ is a random matrix of size $N^{p-1}$-by-$N^{p-1}$ chosen from the so-called Gaussian unitary ensemble [12]. With high probability, the largest eigenvalue of this matrix is $\Theta(N^{(p-1)/2})$. Further, we can choose $\tau=\omega(\log(N)N^{-p/2})$ so that $Z(\sqrt{2N}\tau,2(p-1),N,p-1)=N\exp(O(\tau N^{p/2}))$; for example, this can be shown by using orthogonal polynomials to bound the probability distribution of the largest eigenvalue as discussed previously. We have ${\rm Pr}[\lambda_{1}>x]\leq Z(\tau,p,N,{n_{bos}})\exp(-\tau x)$. For $x$ sufficiently large compared to $N^{p/2}$, for $\tau=\omega(\log(N)N^{-p/2})$, the right-hand side is $o(1)$. ∎ ## 5 Quantum and Classical Algorithms We now discuss the complexity of classical and quantum algorithms to implement the needed linear algebra. In particular, we need to determine if that largest eigenvalue is larger than $E_{cut}$ and we need to find some vector in the eigenspace of eigenvalue $\geq E_{cut}$. We emphasize that it is not necessary to find the leading eigenvector itself. We will use $\psi_{\rm target}$ to denote this leading eigenvector. Note that if we lower bound the squared overlap of some vector with $\psi_{\rm target}$, this will lower bound the probability of success of phase estimation. We have defined algorithm 2 so that in the detection step it uses a hard cutoff on eigenvalue: if the leading eigenvalue is $\geq E_{cut}$ it reports that $\lambda=\overline{\lambda}$ while otherwise it reports that $\lambda=0$. However, no algorithm, classical or quantum, will be able to compute the leading eigenvalue exactly; there will always be some limit on the precision. Fortunately, the proofs of theorems 2,4 show that if $E_{0}\geq(1+c)E_{max}$ for any $c>0$, then if $\lambda=\overline{\lambda}$ then with high probability $\lambda_{1}\geq(1-\eta)E_{0}+\eta E_{max}$ for any $0<\eta<1$ while if $\lambda=0$ then with high probability $\lambda_{1}\leq(1-\eta^{\prime})E_{0}+\eta^{\prime}E_{max}$ for any $0<\eta^{\prime}<1$. For example, we might take $\eta=1/8$ and $\eta^{\prime}=1/2$. So, it suffices instead to implement some “soft” estimate of the leading eigenvalue which will be very likely to give one result (i.e., reporting that $\lambda=\overline{\lambda}$) if the leading eigenvalue is larger than $(7/8)E_{0}+(1/8)E_{max}$ but very unlikely to give that result if the leading eigenvalue is $\leq E_{cut}$. In the quantum algorithms, to obtain some vector in the eigenspace of eigenvalue $\geq E_{cut}$ and to do this soft estimate, we will implement an approximate projector by phase estimation in the quantum algorithms onto the eigenspace of eigenvalue $\geq(E_{0}+E_{cut})/2=(3/4)E_{0}+(1/4)E_{cut}$. By doing this, the phase estimation error will become negligible when considering the projection of the resulting vector onto the eigenspace with eigenvalue $<E_{cut}$. Similarly, in the classical power method we will take the number of iterations sufficiently large that the vector has negligible projection on the eigenspace with eigenvalue $<E_{cut}$ and further so that it has expectation value for $H(T_{0})$ greater than $E_{cut}$. We begin with a description of some classical algorithms. The time and space requirements for the classical algorithms are of course not intended to represent a lower bound; rather, they represent times that can be achieved using standard algorithms in the literature. We then give quantum algorithms. Finally, we give a further improvement to the quantum algorithm that may be useful in practice. When we refer to “space” in a classical algorithm, if we store a $D(N,{n_{bos}})$-dimensional vector, the space requirement is equal to $D(N,{n_{bos}})$ multiplied by the number of bits to store a single entry of the vector. In the classical algorithms, we will not discuss issues with finite precision arithmetic in detail. Since we will be applying operators of the form $H(T_{0})^{m}$ to vectors in the “path integral” methods, we might need to approximate each entry of the vector to accuracy $D(N,{n_{bos}})^{-1}\|H(T_{0})\|^{-m}=O({\rm poly}(N^{{n_{bos}}m}))^{-1}$. However, the required number of bits is then only $O(m{n_{bos}}\log(N))$ and $m$ will be logarithmic in $N$ so the required number of bits will be only polylogarithmic in $N$. ### 5.1 Classical Algorithms Classically, the most obvious algorithm is to perform an eigendecomposition on $H(T_{0})$. This requires storing matrices of size $D(N,{n_{bos}})$-by-$D(N,{n_{bos}})$ so that the space required is $\tilde{O}(D(N,{n_{bos}})^{2})$ and the time is $D(N,{n_{bos}})^{\omega}$ where $\omega$ is the matrix multiplication exponent [13], though of course in practice the time is closer to $D(N,{n_{bos}})^{3}$. However, there is no need to perform a full eigendecomposition. One can instead initialize a random vector and then apply the power method to extract some eigenvector of $H(T_{0})$ in the eigenspace with eigenvalue $\geq E_{cut}$. The space required is then only $\tilde{O}(D(N,{n_{bos}}))$. The time required for a single iteration of the power method is $\tilde{O}(D(N,{n_{bos}}))$. If $\lambda=\overline{\lambda}$, then in $O(\log(D(N,{n_{bos}}))/\log(E_{0}/E_{max}))$ iterations, the resulting vector will have an $1-o(1)$ projection onto the eigenspace with eigenvalue $\geq(E_{0}+E_{cut})/2$ and a negligible projection onto the eigenspace with eigenvalue $<E_{cut}$. So, after this many iterations, one can compute the expectation value of $H(T_{0})$ on that vector to perform detection, i.e., if $\lambda=\overline{\lambda}$ the expectation will be larger than $E_{cut}$ but if $\lambda=0$ the expectation will be close to $E_{max}$, and one can compute the single particle density matrix of the vector after those iterations. So, the time is $\tilde{O}(D(N,{n_{bos}}))O(1/\log(E_{0}/E_{max}))$. This power method still requires exponential (in ${n_{bos}}$) space. In fact, one can use only polynomial space. One obvious choice is to perform a “path integral”. That is, given a random initial state $\Psi_{\rm random}$, the power method computes the state $H(T_{0})^{m}|\Psi_{\rm random}\rangle$, for some exponent $m$ which gives the number of iterations in the power method. So, we wish to compute $\frac{\langle\Psi_{\rm random}|H(T_{0})^{m}a^{\dagger}_{\mu}a_{\nu}H(T_{0})^{m}|\Psi_{\rm random}\rangle}{\langle\Psi_{\rm random}|H(T_{0})^{2m}|\Psi_{\rm random}\rangle},$ where the denominator is to correctly normalize the state. Let us choose the initial state to be a (normalized) random state from the basis for the symmetric basis given before. We make this choice to make the “path integral” simpler and since this is a complete orthonormal basis of states, a random state from this basis has expected overlap with the largest eigenvector of $H(T_{0})$ equal to $1/N^{{n_{bos}}}$. Then, both the numerator and denominator above can be expressed by a summation over intermediate states from this basis, requiring only space $\tilde{O}(\log(D(N,{n_{bos}}))m)$; this summation is a “path integral”. The time required however is now $\tilde{O}(D(N,{n_{bos}})^{m})$ and so becomes significantly worse if the number of iterations is much more than $1$. An improvement to this time can be obtained by using the algorithm of theorem 4.1 of Ref. [11]. This algorithm is expressed in terms of qubits, but we can translate the algorithm to get an algorithm for a single qudit of dimension $D(N,{n_{bos}})$, or, using the tradeoffs discussed there, ${n_{bos}}$ qudits each of dimension $N$. Let us explain this in detail, following the discussion there. We wish to compute $\langle y|C|x\rangle$, for some fixed basis states $x,y$ from the basis above (in our application $x=y$, where $C$ is a “circuit” of some depth $d$ (for example $C=2m+1$ for the numerator above and $C=m$ for the denominator above). While in Ref. [11], a circuit was assumed to be unitary, there in fact is no need to assume that; for non-unitary circuits, the issues with finite precision arithmetic do become more important but as remarked above this can be handled with only polylogarithmic overhead. For us, a circuit is not necessarily unitary; rather it is built out of a sequence of operations, each of which is either $H(T_{0})$ or $a^{\dagger}_{\mu}a_{\nu}$; more generally a circuit might include any operations built out of some tensor of order $O(1)$ multiplying some number of creation and annihilation operators. For $d=1$, $\langle y|C|x\rangle$ can be computed computed in time ${\rm poly}({n_{bos}})$. For $d>1$, we have $\displaystyle\langle y|C|x\rangle=\sum_{z}\langle y|C_{[d\leftarrow d/2+1]}|z\rangle\cdot\langle z|C_{[d/2\leftarrow 1]}|x\rangle,$ (40) where the summation is over states $z$ in the basis and where $C_{[d\leftarrow d/2+1]}$ and $C_{[d/2\leftarrow 1]}$ are subcircuits from the second and first half of $C$. If $F(d)$ is the runtime at depth $d$, we have $F(d)\leq 2\cdot D(N,{n_{bos}})\cdot F(\lceil d/2\rceil)$. So, $F(d)\leq{\rm poly}({n_{bos}})(2D(N,{n_{bos}}))^{\lceil\log(d)\rceil}.$ So, the runtime is $\tilde{O}((2D(N,{n_{bos}}))^{\lceil\log(2m+1)\rceil})$. This is much faster than the path integral method but potentially much slower than the power method, depending on the required $m$, i.e., for $E_{0}/E_{max}=\Theta(1)$, we need $m$ proportional to $\log(D(N,{n_{bos}}))$ and so the time required is superpolynomially worse than the time required if one stores the full vector. ### 5.2 Quantum Algorithms We now discuss quantum algorithms for the same problem. In contrast to the classical algorithms above, all these algorithms take only polynomial space. First let us describe a simple algorithm, given as algorithm 3. We then describe a sequence of improvements. For algorithm 3 and all subsequent algorithms, we analyze assuming that $\lambda=\overline{\lambda}$, i.e., for purposes of analysis we consider the problem of recovery rather than detection. All these algorithms report success or failure and, if they fail, they are rerun until they succeed. We give bounds on the expected runtime under the assumption that $\lambda=\overline{\lambda}$. If we consider the problem of detection, and if $\lambda=0$, then the algorithm will not report success in the given runtime (since it will be unable to succeed in a certain phase estimation procedure) and so all these algorithms can also be used for detection by running them for some multiple of the given runtime and reporting that $\lambda=0$ if success is not reported in that time. #### 5.2.1 Maximally Entangled (or Maximally Mixed) Input State Algorithms 3,4 in this subsubsection work on a Hilbert space which is the tensor product of two Hilbert spaces, each of which have dimension $D(N,{n_{bos}})$. (In some versions of the Hamiltonian simulation used in the algorithms, it is convenient to embed the symmetric subspace of dimension $D(N,{n_{bos}})$ within the full Hilbert space.) We use a tensor product notation $A\otimes B$ to denote an operator that is a tensor product of two operators $A,B$ on the two different tensor factors. Algorithm 3 Quantum Algorithm (simplest, unamplified version). This and all other quantum algorithms have the same inputs. outputs, and parameter choices as algorithm 2. * 1. Prepare a maximally entangled state between the two qudits. * 2. Apply phase estimation using $H(T_{0})\otimes I$. Let $\psi$ be the resulting state. If the resulting eigenvalue is larger than $(E_{0}+E_{cut})/2$, report “success”. Otherwise, report “failure”. * 3. If success is reported, measure and return $\langle\psi|a^{\dagger}_{\mu}a_{\nu}\otimes I|\psi\rangle.$ Steps $1-2$ are designed to prepare a state whose density matrix on the first qudit has large projection onto the eigenspace of eigenvalue $\geq E_{cut}$. For purposes of analysis, we trace out the second qudit, so that the input state on the first qudit is a maximally mixed state. If success is reported then (ignoring phase estimation error) we have indeed projected onto this eigenspace. Ignoring phase estimation error, the probability of success is $\geq 1/D(N,{n_{bos}})$. Remark: we prepare a maximally entangled state between the two qudits so that the density matrix on the first state is maximally mixed; we could equally well modify the algorithm to use only a single qudit (reducing the space by a factor of $2$) and prepare a random state on the first qudit. This modification however requires some care when we discuss a version of the algorithm that uses amplitude amplification below. In the unamplified version, we can choose a new random state (for example, choosing it uniformly from any orthogonal basis) each time we perform phase estimation; however, in the amplified version, we must choose a fixed random initial state on the first qudit and amplify the algorithm with that choice of initial state. We claim (we omit the proof) that if one picks a tensor product of random single qudit states $|v_{1}\otimes v_{2}\otimes\ldots\otimes v_{{n_{bos}}}\rangle$ where $v_{a}$ are independently chosen from a Haar uniform distribution on the sphere $|v_{a}|=1$, with high probability the state has squared overlap with the leading eigenvector close to $N^{-{n_{bos}}}$; more precisely, “close” means that with high probability the logarithm of the squared overlap is at least $-{n_{bos}}\log(N)\cdot(1+o(1))$. Note that we might not have this property of the overlap if we had chosen the initial state from the computational basis, for example. Note also that we can efficiently prepare states from this distribution. Sketch of proof of claim: consider the logarithm of the probability that the sequence of measurements $|v_{i}\rangle_{i}\langle v_{i}|$ succeeds for $i=1,\ldots,{n_{bos}}$ in sequence. The probability that the $i$-th measurement succeeds, conditioned on the previous measurements succeeding, can be computed from the trace of $|v_{i}\rangle\langle v_{i}|$ with the reduced density matrix on the $i$-th qudit of some state (i.e., the leading eigenvector projected by the previous measurements), and for any such reduced density matrix with high probability the logarithm of the trace is at least $-\log(N)\cdot(1+o(1))$. Step $3$ of the algorithm measures some property of a state in the eigenspace with eigenvalue $\geq E_{cut}$. It is possible that each time the algorithm is run, one obtains a different energy measurement and hence a different state, so that measuring this property of the state gives some expectation value of $a^{\dagger}_{\mu}a_{\nu}$ in a mixed state. This does not matter since theorems 3 or 5 also hold for mixed states. We explain the measurement in step 3 in more detail below. The simplest possibility is to simply measure one matrix element of $a^{\dagger}_{\mu}a_{\nu}$. Since there are $N^{2}$ matrix elements, we then need to repeat the algorithm ${\rm poly}(N,\log(\epsilon))$ times to measure each matrix element to accuracy $\epsilon$. We explain improvements to this in subsection 5.3. We explain the phase estimation in more detail below. First, let us analyze the algorithm in a rough outline. Let the phase estimation be carried out to a precision sufficiently smaller than $E_{0}-E_{max}$. To define this, we work in an eigenbasis of $H(T_{0})$. Let $\tilde{\epsilon}$ be a bound on the error probability of the phase estimation. More precisely, we will say that phase estimation implements an operator $E_{PE}$ which is diagonal in the eigenbasis such that on a normalized eigenvector $v_{i}$ with eigenvalue $\lambda_{i}$ we have $\displaystyle\lambda_{i}<E_{cut}\;\rightarrow\;\langle v_{i}|E_{PE}|v_{i}\rangle\leq\tilde{\epsilon},$ (41) $\displaystyle\lambda_{i}>(7/8)E_{0}+(1/8)E_{max}\;\rightarrow\;\langle v_{i}|E_{PE}|v_{i}\rangle\geq 1-\tilde{\epsilon}.$ The reader should appreciate that there are a few energy scales here, chosen rather arbitrarily. The exact value of the energies are not too important. We have picked $E_{cut}$ to be partway between $E_{0}$ and $E_{max}$ so that it is very unlikely that the largest eigenvalue of $H(G)$ is above $E_{cut}$ and also very unlikely that the $\lambda_{1}<E_{cut}$. We have picked the energy cutoff in step 2 to be $(E_{0}+E_{cut})/2$ simply to pick some energy partway between $E_{0}$ and $E_{cut}$ so that it is very unlikely that phase estimation reports success on an eigenstate with energy $<E_{cut}$; see first line of Eq. (41). In the last line of Eq. (41) we wrote $(7/8)E_{0}+(1/8)E_{max}$ simply to pick some energy scale slightly below $E_{0}$ above which it is very likely that phase estimation reports success; for us later, it would suffices to have any instead a bound for $\lambda_{i}\geq(1-\eta)E_{0}+\eta E_{max}$ for any $\eta>0$. Of course, the two lines in Eq. (41) are not completely symmetric about $(E_{0}+E_{cut})/2$. Then, choose $\tilde{\epsilon}=\epsilon/D(N,{n_{bos}})$, so that the algorithm reports success with probability at least $(1-\tilde{\epsilon})/D(N,{n_{bos}})$ and, given that the algorithm reports success, the resulting state has projection onto the eigenspace with eigenvalue $\geq E_{cut}$ which is greater than or equal to $\frac{(1-\tilde{\epsilon})}{(1-\tilde{\epsilon})+(D(N,{n_{bos}})-1)\tilde{\epsilon}}=1-O(\epsilon),$ where the first term in the denominator is the probability that it reports success on $\psi_{\rm target}$ as input and the second term is the probability of reporting success on a state in the eigenspace with eigenvalue $<E_{cut}$, multiplied by $D(N,{n_{bos}})-1$, i.e., multiplied by an upper bound on the dimensionality of that eigenspace. Taking $\epsilon\ll 1$, we can obtain a large projection onto the eigenspace with eigenvalue $\geq E_{cut}$, so that the cost of phase estimation increases logarithmically with $D(N,{n_{bos}})/\epsilon$. The success probability for $\epsilon\ll 1$ is greater than or equal to $(1/D(N,{n_{bos}}))(1-\epsilon/D(N,{n_{bos}}))$, so for small $\epsilon$ it is very close to $1/D(N,{n_{bos}})$. Hence, repeating the algorithm until it succeeds, the expected runtime to obtain a single measurement of one matrix element of $a^{\dagger}_{\mu}a_{\nu}$ is bounded the time for phase estimation multiplied by $O(D(N,{n_{bos}}))$. To perform phase estimation, we use controlled simulation of the Hamiltonian $H(T_{0})$. There are a large number of quantum simulation algorithms which would work here, such as Refs. [14, 15, 16, 17, 18] to name just a few. There are two broad possibilities. The first possibility is to work in the symmetric subspace of dimension $D(N,{n_{bos}})$. In this case, $H(T_{0})$ is a sparse Hamiltonian, and sparse simulation algorithms apply. The second possibility is to use the Hamiltonian of Eq. (5) and embed the symmetric subspace into the full Hilbert space; in this case, $H(T_{0})$ is a local Hamiltonian, in that each terms acts on a small number of qudits, each of dimension $N$, and local simulation algorithms apply. The cost for these algorithms to simulate for a time $t$ to error $\tilde{\epsilon}$ is ${\rm poly}(t\|H(T_{0})\|,{n_{bos}},N,\log(\tilde{\epsilon}))$. Using the simplest phase estimation algorithm of Ref. [19], the number of bits that we need to phase estimate is $s=O(\log(\|H(T_{0})\|/(E_{0}-E_{max}))$. The most expensive bit to obtain is the least significant bit, since obtaining the $j$-th least significant bit requires simulating for a time proportional to $2^{s-j}(E_{0}-E_{max})^{-1}$. So, we can obtain the least significant bit to error $\tilde{\epsilon}/2$, then obtain the next least significant bit to error $\tilde{\epsilon}/4$, and so on, making the total error $\tilde{\epsilon}$. Of course, a large number of variations of the Kitaev phase estimation algorithm exist in the literature, and any could be used here. With high probability, $\|H(T_{0})\|$ is ${\rm poly}(N)$. Thus, with high probability the time for phase estimation is ${\rm poly}(N,{n_{bos}},1/(E_{0}-E_{max}),\log(D(N,{n_{bos}})/\epsilon))$, giving an algorithm runtime $D(N,{n_{bos}}){\rm poly}(N,{n_{bos}},1/(E_{0}-E_{max}),\log(D(N,{n_{bos}})/\epsilon)).$ We can speed this algorithm up quadratically by applying amplitude amplification [20]. Modify the phase estimation step $2$ of algorithm 3 so that the algorithm phase estimates the eigenvalue, determines if the eigenvalue is larger than $E_{cut}$, then uncomputes the eigenvalue, returning just a single bit of success or failure. See algorithm 4. Then, applying amplitude amplification, with high probability the algorithm succeeds in expected time $D(N,{n_{bos}})^{1/2}{\rm poly}(N,{n_{bos}},1/(E_{0}-E_{max}),\log(D(N,{n_{bos}})/\epsilon)).$ Multiplying by ${\rm poly}(N,1/\epsilon)$ to measure $a^{\dagger}_{\mu}a_{\nu}$ to accuracy $\epsilon$, the the expected time is still $D(N,{n_{bos}})^{1/2}{\rm poly}(N,{n_{bos}},1/(E_{0}-E_{max}),\log(D(N,{n_{bos}})/\epsilon)),$ giving a quadratic time improvement, up to ${\rm poly}(N)$ factors, and an exponential space improvement, over the fastest classical algorithm described above. Algorithm 4 Quantum Algorithm (amplified version) * 1. Apply amplitude amplification to steps $1-2$ of algorithm 3, modifying step $2$ to uncompute the eigenvalue and return only success or failure. * 2. If success is reported, measure and return $\langle\psi|a^{\dagger}_{\mu}a_{\nu}\otimes I|\psi\rangle.$ #### 5.2.2 Chosen Input State: Simple Version We can obtain a further quadratic speedup by modifying the initial state that we phase estimate. In this subsubsection, let us first explain an algorithm that gives the basic idea of the initial state preparation; we will only be able to prove some slightly weaker results for this algorithm (in particular we will prove a lower bound on the average inverse runtime, rather than an upper bound on the average runtime). We expect that this is primarily a technical issue and that some concentration of measure argument should allow us to prove an upper bound on the average runtime. We will then describe in the next subsubsection a modification to the algorithm which avoids this technical difficulty and for which we can prove a quadratic improvement without further assumption. Instead of a maximally entangled state, we can use the tensor $T_{0}$ to prepare a state with a larger projection onto $\psi_{\rm target}$. In this subsubsection, we will work in the $D^{n_{bos}}$-dimensional Hilbert space of Eq. (5), so we will use ${n_{bos}}$ qudits each of dimension $N$. The unamplified version is algorithm 5 and a version with amplitude amplification is algorithm 6. The most important new step that must be explained is the initial state preparation (we discuss some other details at the end of this subsubsection). We use the fact that, given a classical list of amplitudes for some $M$-dimensional vector, with the vector having unit norm, we can prepare a quantum state on an $M$-dimensional qudit with the given amplitudes (up to an ill-defined overall phase, of course) using a quantum circuit of depth $O(M)$ and using $O(M)$ classical computation. For example, labelling the basis states $|0\rangle,|1\rangle,\ldots,|M-1\rangle$, one can start with initial state $|0\rangle$ and apply a sequence of $M-1$ rotations in the two dimensional subspaces spanned by $i\rangle,|i+1\rangle$ for $i=0,\ldots,M-2$. Algorithm 5 Quantum Algorithm (improved input state, unamplified version) * 1. Use $T_{0}$ to prepare the initial state $\Psi_{\rm input}$ of Eq. (42). * 2. If the initial state is not in the symmetric subspace, report “failure”. If the state is in the symmetric subspace, apply phase estimation using $H(T_{0})$. Let $\psi$ be the resulting state. If the resulting eigenvalue is larger than $(E_{0}+E_{cut})/2$, report “success”. Otherwise, report “failure”. * 3. If success is reported, measure and return $\langle\psi|a^{\dagger}_{\mu}a_{\nu}|\psi\rangle.$ Algorithm 6 Quantum Algorithm (amplified version) * 1. Apply amplitude amplification to steps $1-2$ of algorithm 5, modifying step $2$ to uncompute the eigenvalue and uncompute the determination of whether the state is in the symmetric subspace and to return only success or failure. * 2. If success is reported, measure and return $\langle\psi|a^{\dagger}_{\mu}a_{\nu}|\psi\rangle.$ In this subsubsection, for simplicity in analysis, we assume that the error $\epsilon$ in phase estimation is small enough to be negligible. We use the same method to produce the input state for both even and odd $p$. First consider the case that ${n_{bos}}$ is an integer multiple of $p$. For any tensor $T$ of order $p$, let $|T\rangle$ denote the vector on $p$ qudits (each of dimension $N$) with amplitudes given by the entries of the tensor. This vector is correctly normalized if $|T|=1$. We prepare the input state $\Psi_{\rm input}=\frac{1}{|T_{0}|^{{n_{bos}}/p}}|T_{0}\rangle^{\otimes{n_{bos}}/p}.$ (42) Preparing this state takes circuit depth $O(N^{p})$ since we can prepare ${n_{bos}}/p$ copies of the state $\frac{1}{|T_{0}|}|T_{0}\rangle$ in parallel. We want to know the expectation value $\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle$, but to get oriented, let us estimate the overlap $\langle\Psi_{\rm input}|\Psi_{\rm sig}\rangle$. We have $\langle v_{\rm sig}^{\otimes p}|T_{0}\rangle=\lambda N^{p}+\langle v_{\rm sig}^{\otimes p}|G\rangle.$ (43) The probability distribution of $\langle v_{\rm sig}^{\otimes p}|G\rangle$ is a Gaussian with zero mean and unit variance, so with high probability, $\langle v_{\rm sig}^{\otimes p}|G\rangle$ is $o(\lambda N^{p})$; indeed, for any increasing function of $N$ which diverges as $N\rightarrow\infty$, with high probability it is bounded by that function. Hence, with high probability, $\langle v_{\rm sig}^{\otimes{n_{bos}}}|T_{0}^{\otimes{n_{bos}}/p}\rangle=(1-o(1))\cdot\lambda^{{n_{bos}}/p}N^{n_{bos}}.$ (44) At the same time, $\langle T_{0}^{\otimes{n_{bos}}/p}|T_{0}^{\otimes{n_{bos}}/p}\rangle=|T_{0}|^{2{n_{bos}}/p}.$ (45) We have $\mathbb{E}[|G|^{2}]=O(N^{p}),$ where the precise constant in the big-O notation depends on whether we symmetrize $G$ or not and whether we use complex entries or not. Further, $|G|^{2}$ is a sum of squares of independent random variables (the entries of $G$). So, by central limit, with high probability $|G|^{2}$ is bounded by $O(N^{p})$. So, with high probability, $|T_{0}|^{2{n_{bos}}/p}=O(N^{{n_{bos}}})$. So, with high probability, $\frac{\Bigl{|}\langle v_{\rm sig}^{\otimes{n_{bos}}}|T_{0}^{\otimes{n_{bos}}/p}\rangle\Bigr{|}^{2}}{\langle v_{\rm sig}^{\otimes{n_{bos}}}|v_{\rm sig}^{\otimes{n_{bos}}}\rangle\cdot\langle T_{0}^{\otimes{n_{bos}}/p}|T_{0}^{\otimes{n_{bos}}/p}\rangle}\geq(1-o(1))\cdot\lambda^{2{n_{bos}}/p}.$ (46) For $\lambda=CN^{-p/4}$, this is $(1-o(1))C^{2{n_{bos}}/p}N^{-{n_{bos}}/2}$. If $\psi_{\rm target}$ were equal to $\Psi_{\rm sig}=N^{-{n_{bos}}/2}|v_{\rm sig}^{\otimes{n_{bos}}}\rangle$, then for fixed $N^{-p/4}/\lambda$, Eq. (46) would give a lower bound to the squared overlap of the initial state with $\psi_{\rm target}$ which would be quadratically better (in terms of its scaling with $N$) than the squared overlap for the maximally entangled input state. So, after applying amplitude amplification, this would give a quartic improvement over the fastest classical algorithm. However, $\psi_{\rm target}$ is not equal to $\Psi_{\rm sig}=N^{-{n_{bos}}/2}|v_{\rm sig}^{\otimes{n_{bos}}}\rangle$ and so this does not give a lower bound on $\langle\psi_{\rm target}|\Psi_{\rm input}\rangle$. However, we have: ###### Lemma 8. Suppose that $E_{0}\geq(1+c)E_{max}$. Then, the projection of $\Psi_{\rm sig}$ onto the eigenspace with eigenvalue $\geq(7/8)E_{0}+(1/8)E_{max}$ is greater than or equal to $\Omega(1)\cdot c/(1+c)$. ###### Proof. The expectation value $\langle\Psi_{\rm sig}|H(T_{0})|\Psi_{\rm sig}\rangle$ was estimated in subsections 3.2.1,3.2.2. We have that with high probability, this expectation value is $\geq(1-o(1))E_{0}$. With high probability, the largest eigenvalue of $H(T_{0})$ in absolute value is bounded by $(1+o(1))(E_{0}+E_{max})$; for the even case this is just the triangle inequality, while for the odd case this uses lemma 3. Hence, by Markov’s inequality applied to $\lambda_{1}-H(T_{0})$, the projection of $\Psi_{\rm sig}$ onto the eigenspace with eigenvalue $\geq(7/8)E_{0}+(1/8)E_{max}$ is greater than or equal to $(1-o(1))E_{0}-((7/8)E_{0}+(1/8)E_{max})/((1+o(1))(E_{0}+E_{max}))$ which is $\geq\Omega(1)\cdot c/(1+c)$. ∎ So, $\Psi_{\rm sig}$ has some non-negligible projection onto the desired eigenspace. This does not however yet give us a lower bound on $\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle$: we can expand $\Psi_{\rm input}$ as a linear combination of $\Psi_{\rm sig}$ and some orthogonal state but we have not bounded the cross-terms in the expectation value. However, we now give heuristic evidence (not a proof) for a lower bound on $\mathbb{E}[\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle]$. The main assumption we will make is that $\lambda_{1}=E_{0}\cdot(1+o(1/\log(N))$; we expect that that assumption can be proven to hold with high probability. We consider just the case of even $p$ (we expect that odd $p$ can be handled similarly). The main reason that we do not give a full proof is that a lower bound on $\mathbb{E}[\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle]$ will only imply a lower bound on the expected inverse squared runtime, rather than an upper bound on the expected runtime, which is what we really want. Instead in the next subsubsection we give a modified algorithm with an upper bound on the expected runtime. Let us note that we conjecture that algorithm 6 does have a quartic speedup for the expect runtime with high probability. One might guess that this could be proven using the bound on expectation value of the inverse squared runtime and some concentration of measure argument. However, we have not been able to make this precise. Let us clarify some terminology to distinguish two meanings of the word “expected”, corresponding to averages over $G$ or averages over outcomes of a quantum algorithm, i.e., to the “expected runtime”. From here on, when we refer to the “runtime” of a phase estimation algorithm, this is a short way of saying the expected runtime for a given choice of $G$. When we refer to the “expectation value of the runtime”, we mean the expectation value over $G$ of this expected runtime. Applying amplitude amplification, the runtime is bounded by the time for the state preparation and phase estimation multiplied by the inverse square-root of $\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle$. So, lower bounding $\mathbb{E}_{G}[\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle]$ will upper bound the expectation value over $G$ of the inverse squared runtime and we will find a bound on the expectation value of the inverse squared runtime by $\Bigl{(}N^{{n_{bos}}/4}{\rm poly}(N,{n_{bos}},1/(E_{0}-E_{max}),\log(D(N,{n_{bos}})/\epsilon))\Bigr{)}^{-2}\Bigl{(}N^{-p/4}/\lambda\Bigr{)}^{-2{n_{bos}}/p}$ in the case that $E_{0}\geq E_{max}\cdot(1+c)$ for any $c>0$. For fixed $N^{-p/4}/\lambda$, this gives a further quadratic improvement, in terms of the scaling of the runtime with $N$, over algorithm 4. Given the existence of that modified algorithm of subsubsection 5.2.3, we will just sketch an outline of a possible proof of the lower bound on $\mathbb{E}[\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle]$, leaving several details out. The basic idea is to lower bound this expectation value by a tensor network, working in some approximation which amounts to ignoring fluctuations in $\lambda_{1}$, then average the tensor network by a sum of pairings, and use this sum to lower bound the expectation value. Roughly the physical idea is that the terms in $\Psi_{\rm input}$ proportional to $G$ will tend on average to increase the overlap $\langle\Psi_{\rm input}|\psi_{\rm target}\rangle$, rather than decrease it. Consider the operator $\lambda_{1}^{-m}H(T_{0})^{m}$ for large $m$. If we take $m$ sufficiently large compared to ${n_{bos}}\log(N)/\log(E_{0}/E_{max})$, this will operator will lower bound $E_{PE}$ up to some negligible error. That is, we take $m$ large enough that $\lambda_{1}^{-m}H(T_{0})^{m}$ is negligibly small acting on any eigenvector $v_{i}$ with eigenvalue $\lambda_{i}\leq(7/8)E_{0}+(1/8)E_{max}$, and for $\lambda_{i}\geq(7/8)E_{0}+(1/8)E_{max}$, Eq. (41) gives a lower bound on $E_{PE}$ that is equal to $1$ up to some negligible phase estimation error $\tilde{\epsilon}$ while clearly $\lambda_{1}^{-m}H(T_{0})^{m}\leq 1$. For fixed $E_{0}/E_{max}$, it suffices to take $m=O({n_{bos}}\log(N))$. For the range of $m$ that we consider here, by assumption we can ignore the fluctuations in $\lambda_{1}$, i.e., we approximate $\displaystyle|\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle|^{2}$ $\displaystyle\gtrsim$ $\displaystyle(\mathbb{E}[\lambda_{1}])^{-m}\mathbb{E}[\langle\Psi_{\rm input}|H(T_{0})^{m}|\Psi_{\rm input}\rangle]$ (47) $\displaystyle\approx$ $\displaystyle(\mathbb{E}[\lambda_{1}])^{-m}\frac{1}{\mathbb{E}[\langle T_{0}^{\otimes{n_{bos}}/p}|T_{0}^{\otimes{n_{bos}}/p}\rangle]}[\mathbb{E}[\langle T_{0}^{\otimes{n_{bos}}/p}|H(T_{0})^{m}|T_{0}^{\otimes{n_{bos}}}\rangle],$ (48) where in the second line we further approximate that we can ignore fluctuations in the norm $\langle T_{0}^{\otimes{n_{bos}}/p}|T_{0}^{\otimes{n_{bos}}/p}\rangle$ and treat it as a constant (the proof is a standard large deviation argument on the norm of the tensor). The quantity $\langle T_{0}^{\otimes{n_{bos}}/p}|H(T_{0})^{m}|T_{0}^{\otimes{n_{bos}}/p}\rangle$ can be evaluated by a sum of tensor networks, using the full Hilbert space, i.e., for each of $m$ choices of $i_{1},\ldots,i_{p/2}$ in each of the $m$ factors of $H(T_{0})$ we have a tensor network). We can then write each tensor network as a sum of tensor networks, inserting either $\lambda v_{\rm sig}^{\otimes n}$ or $G$ for each tensor, and we can average these tensor networks over $G$ using the methods of appendix A by summing over pairings. Since every term in this sum over networks and pairings is positive, if we restrict to some subset of terms, we get a lower bound. Let us restrict to the terms in which for $\Psi_{\rm input}$, we choose $\lambda v_{\rm sig}^{\otimes p}$ for every tensor. For this set of terms, the tensor network computes precisely $\mathbb{E}[\langle\lambda^{{n_{bos}}/p}v_{\rm sig}^{\otimes{n_{bos}}}|H(T_{0})^{m}|\lambda^{{n_{bos}}/p}v_{\rm sig}^{\otimes{n_{bos}}}\rangle]$. This in turn is $\geq(E_{0}\cdot(1-o(1/m))^{m}|\lambda^{{n_{bos}}/p}v_{\rm sig}^{\otimes{n_{bos}}}|^{2}$, since $\langle\Psi_{\rm sig}|H(T_{0})|\Psi_{\rm sig}\rangle\geq E_{0}\cdot(1-o(1/m))$ for $m=O({n_{bos}}\log(N))$. So the tensor network is lower bounded by $E_{0}^{m}|\lambda^{{n_{bos}}/p}v_{\rm sig}^{\otimes{n_{bos}}}|^{2}$. So, with these approximations we have lower bounded $\mathbb{E}[\langle\Psi_{\rm input}|E_{PE}|\Psi_{\rm input}\rangle]$. We make some implementation remarks on the algorithm. The algorithm as described requires measuring whether we are in the symmetric subspace. Note that the input state $\Psi_{\rm input}$ need not be in the symmetric subspace. Such a projection can be done for example by phase estimating a Hamiltonian which is a sum of permutation operators. One can also omit this projection onto the symmetric subspace since our upper bounds on $H(G)$ holds both in the full Hilbert space and in the symmetric subspace. We have considered the case that ${n_{bos}}$ is an integer multiple of $p$. If ${n_{bos}}=kp+l$ for some integers $k,l$ with $0<l<p$, then one can use $l$ ancilla qudits, and prepare an input state which is equal to $\frac{1}{|T_{0}|^{k}}|T_{0}\rangle^{\otimes k},$ on $kp$ qudits, tensored with a maximally entangled state between the remaining $l$ qudits and the remaining $l$ ancillas. The idea is that we get the additional quadratic improvement in overlap on $kp$ of the qudits, and the remaining $l$ ancilla only cost $1/{\rm poly}(N)$ overlap since $l=O(1)$. #### 5.2.3 Chosen Input State: Modified Version We now modify the algorithm 5 (and its amplified version) to obtain an algorithm for which we can prove the quadratic improvement over algorithm 4 without any assumption. Consider given $T_{0}$. Let $\Delta$ be a $p$-th order tensor, chosen from the same distribution as $G$. Consider the tensor $\displaystyle T_{0}^{\prime}$ $\displaystyle=$ $\displaystyle T_{0}+x\Delta$ $\displaystyle\equiv$ $\displaystyle\lambda v_{\rm sig}^{\otimes p}+G^{\prime}$ for some real scalar $x$, where the tensor $G^{\prime}\equiv G+x\Delta$ has Gaussian entries with variance of the entries equal to $1+x^{2}$. We will assume $x=O(1)$; indeed later we will choose $x=o(1)$. Let us write $\Psi_{\rm input}(T)\equiv|T|^{-{n_{bos}}/p}|T^{\otimes{n_{bos}}/p}\rangle$ and $E_{PE}(T)$ to denote the phase estimation operator $E_{PE}$ for Hamiltonian $H(T)$. We have $\displaystyle G$ $\displaystyle=$ $\displaystyle\frac{1}{1+x^{2}}(G+x\Delta)+\frac{x}{1+x^{2}}(xG-\Delta)$ $\displaystyle=$ $\displaystyle\frac{1}{1+x^{2}}G^{\prime}+\frac{x}{\sqrt{1+x^{2}}}\frac{(xG-\Delta)}{\sqrt{1+x^{2}}}$ $\displaystyle=$ $\displaystyle\frac{1}{1+x^{2}}G^{\prime}+\frac{x}{\sqrt{1+x^{2}}}\delta,$ where $\delta=(1+x^{2})^{-1/2}(xG-\Delta)$. The two random variables $G^{\prime}$ and $\delta$ have vanishing covariance, so Eq. (5.2.3) expresses $G$ as a scalar multiple of $G^{\prime}$ plus an additional Gaussian random variable $\delta$ which is independent of $G^{\prime}$. The variable $\delta$ also has variance $1$. Given $G$ and $G^{\prime}$, let $G(y)=yG^{\prime}+(1-y)G=(y+\frac{1-y}{1+x^{2}})G^{\prime}+\frac{x(1-y)}{\sqrt{1+x^{2}}}\delta.$ (51) so that $G(y)$ linearly interpolates between $G$ and $G^{\prime}$. The idea behind the algorithm is to take a given $G$ as input, randomly perturb to produce $G^{\prime}$, and then consider several input states $\Psi_{\rm input}(\lambda v_{\rm sig}^{\otimes p}+G(y))$ with different choices of $y\in[0,1]$. Let us first recall a property of normalization. We have $\Psi_{\rm input}=\frac{1}{|T_{0}|^{{n_{bos}}/p}}|T_{0}\rangle^{\otimes{n_{bos}}/p}.$ Let us write $Z=|T_{0}|^{2{n_{bos}}/p},$ so $\Psi_{\rm input}=Z^{-1/2}|T_{0}\rangle^{\otimes{n_{bos}}/p}.$ As shown before, with high probability $Z^{1/2}=|T_{0}|^{{n_{bos}}/p}=O(N^{{n_{bos}}/2})$. Further, with high probability the fluctuations of $Z^{1/2}$ are $o(1)$ compared to its expectation value. So, from here on we will treat this normalization factor $Z$ as a constant, i.e., of course the normalization depends on $N,{n_{bos}}$ but we will ignore its dependence on $G$. We emphasize that we are not making any additional assumption here as with high probability the fluctuations are asymptotically negligible; we are simply choosing not to write the normalization explicitly. (Remark: indeed, all we really need is an upper bound on $|T_{0}|^{{n_{bos}}/p}$ that holds with high probability, since the normalization constant $|T_{0}|^{{n_{bos}}/p}$ always appears in the denominator.) Further, we will, without remarking on it further, treat other normalization factors such as $|\Psi_{\rm input}(\lambda v_{\rm sig}^{\otimes p}+G(y))|$ as constants, and we will introduce other notation for those constants. Indeed, because we treat the normalization factors such as $Z$ as constants, we will mostly work with un-normalized states which simplifies some of the calculations. In an abuse of notation, let us define $\Psi_{\rm input}(y)=\Psi_{\rm input}(\lambda v_{\rm sig}^{\otimes p}+G(y))$, and write $Z(y)=|\lambda v_{\rm sig}^{\otimes p}+G(y)|^{2{n_{bos}}/p}$. Let us write $T_{0}(y)=\lambda v_{\rm sig}^{\otimes p}+G(y)$. Let $\Psi(y)$ denote the un-normalized state $|T_{0}(y)\rangle^{\otimes{n_{bos}}/p}$ so that $\Psi_{\rm input}(y)=Z(y)^{-1/2}\Psi(y)$. Let us expand $\Psi(y)$ as a series in $\delta$ and define $\Psi^{0}(y)$ to denote the zeroth order term in $\delta$. As a warmup, let us consider $\langle\Psi^{0}(y)|E_{PE}(T_{0}^{\prime})|\Psi^{0}(y)\rangle$. We consider the higher order terms in $\delta$ later. We expand the state $\Psi^{0}(y)$ as a series in $(y+\frac{1-y}{1+x^{2}})$. Doing this means that we express $\langle\Psi^{0}(y)|E_{PE}(T_{0}^{\prime})|\Psi^{0}(y)\rangle$ as a polynomial of degree $2{n_{bos}}$ which we write as $\sum_{i\geq 0}a_{i}(y+\frac{1-y}{1+x^{2}})^{i}.$ The zero-th order term $a_{0}$ is simply equal to $\langle(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}|E_{PE}(T_{0}^{\prime})|(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}\rangle$, which is lower bounded in lemma 8. Hence, $a_{0}\geq\lambda^{2{n_{bos}}/p}N^{{n_{bos}}}\Omega(1)\cdot c/(1+c)$ (52) for $E_{0}\geq(1+c)E_{max}$. Now we use a result about polynomials: ###### Lemma 9. Let $p(z)$ be a polynomial of degree $2{n_{bos}}$. Let $[a,b]$ be an interval with $0\leq a<b$. Then, ${\rm max}_{z\in[a,b]}|p(z)|\geq\Bigl{|}\frac{a+b}{b-a}\Bigr{|}^{-2{n_{bos}}}\exp(-O({n_{bos}}))|p(0)|.$ (53) ###### Proof. We minimize ${\rm max}_{z\in[a,b]}|p(z)|$ over polynomials of degree $2{n_{bos}}$ with given value $p(0)$. Applying an affine transformation $z\rightarrow(2/(b-a))(z-(a+b)/2)$, this is equivalent to minimizing ${\rm max}_{z\in[-1,1]}|p(z)|$ over polynomials of degree $2$ with given value $p(z_{0})$ for $z_{0}=(a+b)/(a-b)$. We claim that this is minimized by $\frac{p(z_{0})}{T_{2{n_{bos}}}(z_{0})}T_{2{n_{bos}}}(z),$ where $T_{2{n_{bos}}}$ is a Chebyshev polynomial. Proof of claim: suppose some other polynomial $q(z)$ has a smaller maximum absolute value on $[-1,+1]$ with $q(z_{0})=p(z_{0})$. Then the polynomial $p(z)-q(z)$ has a zero at $z_{0}$ but also has at least $2{n_{bos}}$ zeros on the interval $[-1,+1]$; this follows from the intermediate value theorem because $T_{2{n_{bos}}}$ has $2{n_{bos}}+1$ extreme points on the interval which alternate signs. This gives a contradiction since $p(z)-q(z)$ is degree at most $2{n_{bos}}$. So, ${\rm max}_{z\in[a,b]}|p(z)|\geq|p(z_{0})|/|T_{2{n_{bos}}}(z_{0})|$. If $0\not\in[a,b]$ then $|z_{0}|>1$. We can bound $T_{2{n_{bos}}}(z_{0})$ for $|z_{0}|>1$ by $|z_{0}|^{2{n_{bos}}}$ times the sum of absolute values of coefficients of $T_{2{n_{bos}}}$. This sum of coefficients is bounded by $\exp(O({n_{bos}}))$. ∎ We apply this lemma to lower bound ${\rm max}_{y\in[0,1]}|\sum_{i\geq 0}a_{i}(y+\frac{1-y}{1+x^{2}})^{i}|$. For $z=y+\frac{1-y}{1+x^{2}}$, with $b=1$ and $a=(1+x^{2})^{-1}$, this is ${\rm max}_{z\in[a,b]}|p(z)|$ with $p(z)=\sum_{i}a_{i}z^{i}$. So, using that $(a+b)/(b-a)\leq(2/x^{2})$, we have ${\rm max}_{y\in[0,1]}|\sum_{i\geq 0}a_{i}(y+\frac{1-y}{1+x^{2}})^{i}|\geq x^{4{n_{bos}}}\exp(-O({n_{bos}}))|a_{0}|.$ (54) However, for purposes of an algorithm, we need to consider not just the maximum of this polynomial $p(z)$ over an interval, but also consider whether we can come close to this maximum by sampling it at some small number of discrete points. Fortunately we have the following lemma: ###### Lemma 10. Let $p(z)$ be a polynomial of degree $2{n_{bos}}$. Let $[a,b]$ be an interval which does not contain $0$. Then, with probability at least $1/2$, if one selects a random point $z$ in the interval $[a,b]$ from the uniform distribution, we have $|p(z)|\geq\Bigl{|}\frac{a+b}{b-a}\Bigr{|}^{-2{n_{bos}}}\exp(-O({n_{bos}}))|p(0)|$. ###### Proof. Apply an affine transformation $z\rightarrow(2/(b-a))(z-(a+b)/2)$, which maps $a$ to $-1$, $b$ to $+1$ and $0$ to $(a+b)/(a-b)$. Write $p(z)=\prod_{i}(z-z_{i})$ where $z_{i}$ are zeros of the polynomial; we multiply by $p(z)$ by a normalization constant so that the highest order coefficient is equal to $1$. Then $\log(|p(z)|)=\sum_{i}{\rm Re}(\log(z-z_{i}))$. Let $A$ denote the average of this logarithm over the interval $[-1,+1]$; by calculus this is $A=\frac{1}{2}\sum_{i}{\rm Re}\Bigl{(}(z-z_{i})\log(z-z_{i})-(z-z_{i})\Bigr{)}\Bigl{|}_{z=-1}^{z=+1}.$ We claim that $\log(p((a+b)/(a-b)))-A\leq 2{n_{bos}}\cdot\Bigl{(}\log((a+b)/(a-b))+O(1)\Bigr{)}.$ (55) To show this let $T_{i}$ denote a given term in the sum for $A$, i.e. $T_{i}=(1/2){\rm Re}((z-z_{i})\log(z-z_{i})-(z-z_{i}))\Bigl{|}_{z=-1}^{z=+1}$. Let $D_{i}\equiv{\rm Re}(\log((a+b)/(a-b)-z_{i}))-T_{i}$. By considering various cases, we will show that $D_{i}\leq\log((a+b)/(a-b))+O(1)$ which implies Eq. (55). Proof of claim: for $|z_{i}|$ less than or equal to some fixed constant (for example, $|z_{i}|\leq 10$), $T_{i}$ is lower bounded by some absolute constant, so $D_{i}\leq\log((a+b)/(a-b))+O(1)$. For $|z_{i}|$ larger than this fixed constant (for example, $|z_{i}|>10$), $T_{i}$ is at least $\log(|z_{i}|)$ minus some other absolute constant, so $D_{i}\leq{\rm Re}(\log((a+b)/(a-b)-z_{i})-\log(z_{i}))+O(1)={\rm Re}(\log((a+b)/(a-b)))+O(1)$. Further, for each $i$, we see that ${\rm max}_{z\in[-1,+1]}{\rm Re}(\log(z-z_{i}))$ is upper bounded by $T_{i}+O(1)$. Hence, ${\rm max}_{z\in[-1,+1]}\log(|p(z)|)$ is upper bounded by $A+O({n_{bos}})$. Hence, with probability at least $1/2$, for $z$ chosen uniformly in $[-1,+1]$, we have that $\log(|p(z)|)\geq A-O({n_{bos}})$. Hence, with probability at least $1/2$, we have that $\log(|p(z)|)\geq A-2{n_{bos}}\cdot(\log((a+b)/(a-b))-O({n_{bos}})$. ∎ So, noting that for $z=y+\frac{1-y}{1+x^{2}}$, a uniform choice of $y$ on $[0,1]$ is the same as a uniform choice of $z$ on $[1/(1+x^{2}),1]$, we have ###### Lemma 11. For $y$ chosen randomly from the uniform distribution on $[0,1]$, the quantity $\langle\Psi^{0}(y)|E_{PE}(T_{0}^{\prime})|\Psi^{0}(y)\rangle$ is greater than or equal to $\langle(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}|E_{PE}(T_{0}^{\prime})|(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}\rangle\cdot\exp(-O({n_{bos}}))x^{4{n_{bos}}}$. Using Eq. (52), $\langle\Psi^{0}(y)|E_{PE}(T_{0}^{\prime})|\Psi^{0}(y)\rangle\geq\lambda^{2{n_{bos}}/p}N^{{n_{bos}}}\exp(-O({n_{bos}}))x^{4{n_{bos}}}\cdot c/(1+c).$ (56) We later will worry about optimizing over $x$. Increasing $x$ will increase the right-hand side of Eq. (56), but it will also change the normalization constant $Z(y)$ that we must multiply by to correctly normalize the input state and also will worsen the threshold for recovery (since $E_{max}$ will increase). For now, let us note for orientation that if we choose, for example, $x=0.01$, then the tensor $T_{0}^{\prime}$ is chosen from the same distribution as $T_{0}$ up to a multiplication by a factor of $(1+0.01^{2})^{1/2}$ which is very close to $1$. This leads to only a very small reduction in the threshold $\lambda$ at which recovery is possible. At the same time, the change in normalization $(1+0.01^{2})^{{n_{bos}}}$ is asymptotically negligible (for fixed ${n_{bos}}$) compared to the improvement which is polynomial in $N^{n_{bos}}$. We now consider the additional terms depending on $\delta$, i.e., due to the difference $\Psi(y)-\Psi^{0}(y)$. At first glance, we have not gained much by the above trick introducing $T^{\prime}$ since we still have this Gaussian random variable $\delta$ to consider. However, we have the advantage now that in $\langle\Psi(y)|E_{PE}(T_{0}^{\prime})|\Psi(y)\rangle$, the operator $E_{PE}(T_{0}^{\prime})$ does not depend on $\delta$ and the states in the bra and ket depend on $\delta$ only polynomially (treating the overall normalization as a constant). Let us outline our approach. The main worry that we have is that this quantity $\langle\Psi(y)|E_{PE}(T_{0}^{\prime})|\Psi(y)\rangle$ will have a probability distribution that is peaked near some “small” value, i.e., much less than the value of the overlap at $\delta=0$, rather than some “large” value, i.e., roughly comparable to the value of the overlap at $\delta=0$. To deal with this, we will use a trick of adding additional noise to make the expectation value of the overlap large, i.e., roughly comparable to the value of the overlap at $\delta=0$. Now we still need to worry about whether the probability function might have some large probability of having a small value. So, we will then appeal we appeal to a theorem of Carbery-Wright on “anti-concentration” of polynomials of Gaussian random variables, bounding the probability that the probability lies in some small interval. This theorem gives us a useful bound unless the variance of the polynomial is small; however, in that case we can show that the polynomial is likely to be close to its expectation value. The trick of adding additional noise is as follows. We perturb the input state by adding additional Gaussian random noise. Let $\Psi(y,x^{\prime})$ denote the unnormalized state $|\lambda v_{\rm sig}^{\otimes p}+(y+\frac{1-y}{1+x^{2}})G^{\prime}+\frac{x(1-y)}{\sqrt{1+x^{2}}}(\delta+\delta^{\prime})\rangle^{\otimes{n_{bos}}/p},$ where $\delta^{\prime}$ is an additional tensor chosen from a Gaussian distribution with some standard deviation $x^{\prime}$. So, the tensor $\delta+\delta^{\prime}$ is sampled from some distribution with variance $1+x^{\prime 2}$. Let $Z(y,x^{\prime})=|\Psi(y,x^{\prime})|^{2}$. Let $\Psi_{\rm input}(y,x^{\prime})=Z(y,x^{\prime})^{-1/2}\Psi(y,x^{\prime})$ be the “perturbed input state”. Then, we consider the expectation value over $\delta^{\prime}$ of $\langle\Psi(y,x^{\prime})|E_{PE}(T_{0}^{\prime})|\Psi(y,x^{\prime})\rangle$; we consider this expectation value as a series in the variable $1+x^{\prime 2}$. Using the same treatment as above of this expectation value as a polynomial, we find that if $x^{\prime 2}$ is chosen uniformly from the interval $[0,x^{2}]$ then, with probability at least $1/2$, the quantum expectation value $\langle\Psi(y,x^{\prime})|E_{PE}(T_{0}^{\prime})|\Psi(y,x^{\prime})\rangle$ is at least equal to the quantum expectation value with $\delta^{\prime}=0$ multiplied by $\exp(-O({n_{bos}}))x^{8{n_{bos}}}$. Hence, ###### Lemma 12. With probability at least $1/4$, for uniform random choices of $y$ and $x^{\prime 2}$ on $[0,1]$ and $[0,x^{2}]$, the expectation value over $\delta,\delta^{\prime}$ of $\langle\Psi(y,x^{\prime})|E_{PE}(T_{0}^{\prime})|\Psi(y,x^{\prime})\rangle$ is at least $\langle(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}|E_{PE}(T_{0}^{\prime})|(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}\rangle\cdot\exp(-O({n_{bos}}))x^{8{n_{bos}}}$. We emphasize that throughout we are treating $G^{\prime}$ as fixed and considering $\delta,\delta^{\prime}$ as random. Now we use anti-concentration in the Gaussian random variables $\delta,\delta^{\prime}$. The following lemma is a consequence of the Carbery-Wright theorem [21] ###### Lemma 13. Let $p(z_{1},\ldots,z_{n})$ be a polynomial of degree at most $k$ in $n$ independent normally distributed random variables. Then, for any $t\in{\mathbb{R}}$ and any $\delta>0$ with $|t-\mathbb{E}[p]|>\delta$, $\displaystyle{\rm Pr}[|p(z_{1},\ldots,z_{n})-t|\leq\delta]$ $\displaystyle\leq$ $\displaystyle\Bigl{(}\frac{\delta}{|t-\mathbb{E}[p]|-\delta}\Bigr{)}^{2/(2k+1)}O(k)^{2k/(2k+1)}$ $\displaystyle\leq$ $\displaystyle\Bigl{(}\frac{\delta}{|t-\mathbb{E}[p]|-\delta}\Bigr{)}^{2/(2k+1)}O(k),$ where the big-O notation here hides a universal constant. As a corollary, choosing $t=0$, ${\rm Pr}[|p|\leq\delta|\mathbb{E}[p]|]\leq\Bigl{(}\frac{\delta}{1-\delta}\Bigr{)}^{2/(2k+1)}O(k)=O(\delta^{2/(2k+1)})O(k).$ (58) ###### Proof. Let ${\rm Var}(p)$ denote the variance of $p(\cdot)$. As a trivial bound, ${\rm Pr}[|p(z_{1},\ldots,z_{n})-t|\leq\delta]\leq{\rm Var}(p)/\Bigl{(}|t-\mathbb{E}[p]|-\delta\Bigr{)}^{2}.$ (59) By Carbery-Wright, ${\rm Pr}[|p(z_{1},\ldots,z_{n})-t|\leq\delta]\leq O(k)\cdot(\delta/\sqrt{{\rm Var}(p)})^{1/k}.$ (60) Maximizing the bound from Eqs. (59,60) over ${\rm Var}(p)$, Eq. (13) follows. ∎ Hence, from Eq. (58), ###### Lemma 14. Consider a choice of $y,x^{\prime}$ such that the expectation value over $\delta,\delta^{\prime}$ of the quantum expectation value $\langle\Psi(y,x^{\prime})|E_{PE}(T_{0}^{\prime})|\Psi(y,x^{\prime})\rangle$ is at least $\langle(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}|E_{PE}(T_{0}^{\prime})|(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}\rangle\cdot\exp(-O({n_{bos}}))x^{8{n_{bos}}}$. Then, for random choice of $\delta,\delta^{\prime}$, the quantum expectation value $\langle\Psi(y,x^{\prime})|E_{PE}(T_{0}^{\prime})|\Psi(y,x^{\prime})\rangle$ is at least $\langle(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}|E_{PE}(T_{0}^{\prime})|(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}\rangle\cdot\exp(-O({n_{bos}}))x^{8{n_{bos}}}\delta,$ with probability at least $1-O({n_{bos}})O(\delta)^{2/(2{n_{bos}}+1)}$. Lemma 14 considers unnormalized input states. Dividing by normalization constant $Z(y,x^{\prime})$ we can get a lower bound on the expectation value for the normalized input state $\Psi_{\rm input}(y,x^{\prime})$ by (for $E_{0}\geq(1+c)E_{max}$): $\displaystyle\frac{1}{Z(y,x^{\prime})}\langle(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}|E_{PE}(T_{0}^{\prime})|(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}\rangle\cdot\exp(-O({n_{bos}}))x^{8{n_{bos}}}\delta$ $\displaystyle=$ $\displaystyle\lambda^{2{n_{bos}}/p}\exp(-O({n_{bos}}))x^{8{n_{bos}}}\delta\cdot\Omega(1)$ $\displaystyle=$ $\displaystyle\Bigl{(}\frac{\lambda}{N^{-p/4}}\Bigr{)}^{2{n_{bos}}/p}N^{-{n_{bos}}/2}x^{8{n_{bos}}}\delta\cdot\Omega(1)$ $\displaystyle=$ $\displaystyle\Bigl{(}\frac{\lambda}{N^{-p/4}}\Bigr{)}^{2{n_{bos}}/p}D(N,{n_{bos}})^{-1/2}x^{8{n_{bos}}}\delta\cdot\Omega(1).$ Finally, let us pick $x=1/\log(N)$. Then, we consider the following algorithm 7. Algorithm 7 Quantum Algorithm (modified improved input state, unamplified version) * 1. Choose random $y,x^{\prime}$. Sample $T_{0}^{\prime},\delta^{\prime}$ randomly. Prepare input state $\Psi_{\rm input}(y,x^{\prime})$. * 2. If the initial state is not in the symmetric subspace, report “failure”. If the state is in the symmetric subspace, apply phase estimation using $H(T_{0})$ Let $\psi$ be the resulting state. If the resulting eigenvalue is larger than $(E_{0}+E_{cut})/2$, report “success”. Otherwise, report “failure”. * 3. If success is reported, measure and return $\langle\psi|a^{\dagger}_{\mu}a_{\nu}|\psi\rangle.$ We apply amplitude amplification to algorithm 7. We apply amplitude amplification under the assumption that indeed $\langle\Psi(y,x^{\prime})|E_{PE}(T_{0}^{\prime})|\Psi(y,x^{\prime})\rangle$ is at least $\langle(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}|E_{PE}(T_{0}^{\prime})|(\lambda v_{\rm sig}^{\otimes p})^{\otimes{n_{bos}}/p}\rangle\cdot\exp(-O({n_{bos}}))x^{4{n_{bos}}}\delta.$ From the theorems above, with high probability in $\delta^{\prime}$, this happens at least $1/4$ of the time. If the assumption does not hold, we re- sample $y,x^{\prime}$ and try again. Then, we find that for $E_{0}\geq
# Deterministic Algorithms for Low Degree Factors of Constant Depth Circuits Mrinal Kumar Tata Institute of Fundamental Research, Mumbai, India. Email: {mrinal, varun.ramanathan<EMAIL_ADDRESS>Research supported by the Department of Atomic Energy, Government of India, under project 12-R&D-TFR-5.01-0500. Varun Ramanathan11footnotemark: 1 Ramprasad Saptharishi11footnotemark: 1 ###### Abstract For every constant $d$, we design a subexponential time deterministic algorithm that takes as input a multivariate polynomial $f$ given as a constant depth algebraic circuit over the field of rational numbers, and outputs all irreducible factors of $f$ of degree at most $d$ together with their respective multiplicities. Moreover, if $f$ is a sparse polynomial, then the algorithm runs in quasipolynomial time. Our results are based on a more fine-grained connection between polynomial identity testing (PIT) and polynomial factorization in the context of constant degree factors and rely on a clean connection between divisibility testing of polynomials and PIT due to Forbes [For15] and on subexponential time deterministic PIT algorithms for constant depth algebraic circuits from the recent work of Limaye, Srinivasan and Tavenas [LST21]. ## 1 Introduction A long line of research (cf. [vzG83, Kal85, Kal92, Kal03]) on the question of designing efficient algorithms for multivariate polynomial factorization concluded with the influential works of Kaltofen [Kal89] and Kaltofen & Trager [KT88] which gave efficient randomized algorithms for this problem in the whitebox and blackbox settings respectively.111Throughout this paper, we use efficient to mean an algorithm whose time complexity is polynomially bounded in the size, bit-complexity and the degree of the input algebraic circuit. These results and the technical insights discovered in the course of their proofs have since found numerous direct and indirect applications in various areas of complexity theory. This includes applications to the construction of pseudorandom generators for low degree polynomials [Bog05], algebraic algorithms [KY08], hardness-randomness tradeoffs in algebraic complexity [DSY09, GKSS19], algebraic property testing [PS94, AS03, BSCI+20], error correcting codes [BHKS20], deterministic polynomial identity tests for constant depth circuits [LST21, CKS18] among others. Given the fundamental nature of the problem and its many applications, the question of designing efficient deterministic algorithms for multivariate polynomial factorization is of great interest and importance. Shpilka & Volkovich [SV10] observed that this question is at least as hard as PIT in the sense that a deterministic factoring algorithm (in fact, an algorithm to check irreducibility suffices for this) for polynomials given by algebraic circuits implies a deterministic algorithm for PIT for algebraic circuits, a long standing open problem in computer science. In a later work, Kopparty, Saraf & Shpilka [KSS15] showed a connection in the other direction as well. They showed that an efficient deterministic algorithm for PIT for algebraic circuits implies an efficient deterministic algorithm for polynomial factorization for algebraic circuits. Thus, the questions are essentially equivalent to each other. An intriguing aspect of the aforementioned equivalence is that while deterministic algorithms for factoring any rich enough class of circuits (for instance, constant depth circuits) lead to deterministic PIT for the same class (see Observation 1 in [SV10] for a precise statement), the connection in the other direction due to Kopparty, Saraf & Shpilka [KSS15] does not appear to be so fine-grained. In particular, even if we only wish to factor an otherwise simple class of polynomials, e.g. sparse polynomials (polynomials with a small number of non-zero monomials), the PIT required as per the proof in [KSS15] seems to be for significantly more powerful models of algebraic computation like algebraic branching programs. As a consequence, while there has been steady progress on the state of the art of deterministic PIT algorithms in recent years for various interesting sub- classes of algebraic circuits like sparse polynomials [KS01], depth-$3$ circuits with constant top fan-in [SS09, SS10, KS09], read-once algebraic branching programs [FS13, FSS14, For14, GKST15, GKS16] and constant depth circuits [LST21], this progress hasn’t translated to progress on the question of deterministic factoring algorithms for these circuit classes. In particular, deterministic factorization algorithms have remained elusive even for seemingly simple classes of polynomials like sparse polynomials where the corresponding PIT problem is very well understood. There are only a handful of results that make progress towards this and related problems to the best of our knowledge. Shpilka & Volkovich [SV10] showed a close connection between the problems of polynomial identity testing and that of decomposing a polynomial given by a circuit into variable disjoint factors and build on these ideas to give an efficient deterministic algorithm for factoring sparse multilinear polynomials. In subsequent works, Volkovich [Vol15, Vol17] gave an efficient deterministic algorithm to factor sparse polynomials that split into multilinear factors and sparse polynomials with individual degree at most $2$. More recently, a work of Bhargava, Saraf and Volkovich [BSV18] gives a quasipolynomial time deterministic algorithm for factoring sparse polynomials with small individual degree based on some beautiful geometric insights. In general, when the individual degree of a sparse polynomial is not small, no non-trivial deterministic factoring algorithms appear to be known, even when we have the flexibility of describing the output as algebraic circuits. As Forbes & Shpilka note in their recent survey [FS15] on polynomial factorization, we do not even have structural guarantees on the complexity of factors of sparse polynomials even for seemingly coarse measures of complexity like formula complexity. In fact, questions that might be potentially easier than factorization like checking if a given sparse polynomial is a product of constant degree polynomials or checking if a given sparse polynomial is irreducible are not known to have non-trivial deterministic algorithms. Perhaps a little surprisingly, till a recent work of Forbes [For15], we did not even have a non-trivial deterministic algorithm for checking if a given sparse polynomial is divisible by a given constant degree polynomial! Forbes gave a quasipolynomial time deterministic algorithm for this problem by reducing this question to a very structured instance of PIT for depth-$4$ algebraic circuits and then giving a quasipolynomial time deterministic algorithm for these resulting PIT instances. This work is motivated by some of these problems, most notably by the question of designing efficient deterministic algorithms for factoring sparse polynomials. While we do not manage to solve this problem in this generality, we make modest progress towards this: we design a deterministic quasipolynomial time algorithm that outputs all the low degree factors of a sparse polynomial. More generally, we show that constant degree factors of a polynomial given by a constant depth circuit can be computed deterministically in subexponential time. ### 1.1 Our Results ###### Theorem 1.1 (Low degree factors of constant depth circuits). Let $\mathbb{Q}$ be the field of rational numbers and $\varepsilon>0$, $d,k\in\mathbb{N}$ be arbitrary constants. Then, there is a deterministic algorithm that takes as input an algebraic circuit $C$ of size $s$, bit-complexity $t$, degree $D$ and depth $k$ and outputs all the irreducible factors of $C$ of degree at most $d$, along with their respective multiplicities in time $(sDt)^{O((sDt)^{\varepsilon})}$. We note that the bit-complexity of an algebraic circuit/formula is a measure of the bit-complexities of the rational numbers appearing in the circuit. See 2.1 for a formal definition. When the input polynomial is sparse, i.e. has a small depth-$2$ circuit, then the time complexity of the algorithm in Theorem 1.1 can be improved to be quasipolynomially bounded in the input size. This gives us the following theorem. ###### Theorem 1.2 (Low degree factors of sparse polynomials). Let $d\in\mathbb{N}$ be an arbitrary constant. Then, there is a deterministic algorithm that takes as input a polynomial $f\in\mathbb{Q}[\mathbf{x}]$ of sparsity $s$, bit-complexity $t$, degree $D$, and outputs all the irreducible factors of $f$ of degree at most $d$, along with their respective multiplicities in time $(sDt)^{\operatorname{poly}(\log sDt)}$. These results immediately yield an algorithm (with comparable time complexity) to check if the polynomial computed by a given low depth circuit is a product of polynomials of degree at most $d$. More concretely, we have the following corollary that follows by comparison of degrees of the input polynomial and the low-degree factors (with multiplicities) listed by the algorithms in the above theorems. ###### Corollary 1.3. Let $\mathbb{Q}$ be the field of rational numbers and $\varepsilon>0$, $d,k\in\mathbb{N}$ be arbitrary constants. Then, there is a deterministic algorithm that takes as input an algebraic circuit $C$ of size $s$ , bit-complexity $t$, degree $D$ and depth $k$ and decides if $C$ is a product of irreducibles of degree at most $d$ in time $(sDt)^{O((sDt)^{\varepsilon})}$. Moreover, when $f$ is a sparse polynomial with sparsity $s$, then the algorithm runs in $(sDt)^{\operatorname{poly}(\log sDt)}$ time. Note that in the constant depth regime, circuits and formulas are equivalent upto a polynomial blow-up in size. Thus we will use the terms circuits and formulas interchangeably without any loss in our final bounds, and most of our presentation will be for formulas. ### Field dependence of our results We end this section with a remark about the field dependence of our results. The field dependence in our results stems from two reasons. We need an efficient deterministic algorithm for factorization of univariate polynomials over the underlying field $\mathbb{F}$. In addition to this, our proofs also need non-trivial deterministic algorithms for polynomial identity testing (PIT) for constant depth circuits (or very special depth-$4$ circuits for Theorem 1.2) over the underlying field. The field of rational numbers satisfies both these requirements: a classical algorithm of Lenstra, Lenstra and Lovász[LLL82] solves the problem of deterministic univariate factorization efficiently over $\mathbb{Q}$ and a recent work of Limaye, Srinivasan and Tavenas [LST21] gives a subexponential time deterministic algorithm for PIT for constant depth circuits over $\mathbb{Q}$. For Theorem 1.2, the relevant PIT is for special depth-$4$ circuits and was given in a work of Forbes [For15]. In fact, Forbes’ result holds even over finite fields. We restrict our attention to just the field of rational numbers in the presentation although our results work over any large characteristic field that supports the above requirements. ### 1.2 Proof Overview We now give an overview of some of the main ideas in our proofs. In a nutshell, our proofs are based on relatively simple structural observations on top of the existing factoring algorithms. The key is to understand the structure of circuits for which we need a PIT algorithm at every step a little better, and when looking for low degree factors, we observe that these PIT instances are relatively simple and their circuit complexity is comparable to the circuit complexity of the input polynomials themselves. We also crucially use the divisibility testing idea of Forbes [For15] in our algorithm at two stages; this helps us handle factors of large multiplicities and also lets us obtain true factors from the output of Hensel Lifting step of the factorization algorithms. This idea again helps in reducing the complexity of the PIT instance we face in these steps, and in particular, we completely avoid the linear systems solving step in a typical factorization algorithm that naively (e.g. see [KSS15]) seems to require PIT for algebraic branching programs. Once the PIT instances are shown to be relatively simple, we invoke the PIT algorithms of Forbes [For15] and Limaye, Srinivasan & Tavenas [LST21] to solve these deterministically. ##### Typical steps in a polynomial factorisation algorithm: Most factorisation algorithms (and ours, modulo minor deviations) follow this template: 1. 1. Making $f$ monic: Apply a suitable transformation of the form $x_{i}\mapsto x_{i}+\alpha_{i}y$ to ensure that $f$ is monic in $y$. We may now assume that $f\in\mathbb{Q}[\mathbf{x},y]$. 2. 2. Preparing for Hensel lift: Ensure that $f(\mathbf{x},y)$ is square-free, and further that $f({\bm{0}},y)$ is also square-free. 3. 3. Univariate factorisation: Factorise the univariate polynomial $f({\bm{0}},y)$ as a product $g_{0}(y)\cdot h_{0}(y)$ where $\gcd(g_{0},h_{0})=1$. This can be intepreted as a factorisation $f(\mathbf{x},y)=g_{0}\cdot h_{0}\bmod{\mathcal{I}}$ where $\mathcal{I}=\left\langle\mathbf{x}\right\rangle$. 4. 4. Hensel lifting: Compute an iterated lift to obtain $f=g_{\ell}\cdot h_{\ell}\bmod{\mathcal{I}^{2^{\ell}}}$ for a suitably large $\ell$. 5. 5. Reconstruction: From $g_{\ell}$, obtain an honest-to-god factor $g$ of $f$ (unless $f$ is irreducible). The first two steps typically involve the use of randomness for suitable polynomial identity tests. In the first step, we would like ${\bm{\alpha}}$ to be a point that keeps the highest degree homogeneous component of $f$ non- zero, and the second step is handled by translating $f$ by a point ${\bm{\delta}}$ that keeps the “discriminant” of $f$ non-zero. The Hensel lift is a deterministic subroutine that eventually yields small circuits for the lifted factors and the reconstruction step typically involves solving a linear system. It is mostly due to the “discriminant” that we do not have efficient deterministic factorisation algorithm even for constant-depth circuits as the best upper bound for the discriminant we have is an algebraic branching program and we do not have efficient hitting sets for them. (Yet!) For our case, it is instructive to focus on a specific factor $g$ of $f$ and understand what would be required to make the above template yield this factor. The first observation is that the base case of Hensel Lifting does not require $f$ to be square-free but rather that the factor $g$ we intend to reconstruct satisfies $g|f$ and $g^{2}\nmid f$. For now, let us assume this and also that $f$ (and hence $g$ and $h=f/g$ also) is monic in $y$. We have that $\gcd(g,h)=1$ but for the Hensel lift, we also need to find a ${\bm{\delta}}$ that ensures that $\gcd(g_{0},h_{0})=1$ where $g_{0}=g({\bm{\delta}},y)$ and $h_{0}=h({\bm{\delta}},y)$. The set of “good” ${\bm{\delta}}$’s is precisely the points that do not make the resultant $\operatorname{Res}_{y}(g,h)$ zero and thus we want to understand the circuit complexity of this resultant. The resultant $\operatorname{Res}_{y}(g,h)$ is the determinant of a matrix of dimension $\deg_{y}(g)+\deg_{y}(h)$ and its entries are coefficients of $g,h$ when viewed as univariates in $y$. However, we are only given that $f$ is computable by a constant-depth formula and we do not have any good bound on the complexity of $h$. We circumvent this by working with a _pseudo-quotient_ (introduced by Forbes [For15] in the context of divisibility testing) $\tilde{h}$ of $f$ and $g$; we work with $\operatorname{Res}_{y}(g,\tilde{h})$ and show that it is also computable by constant-depth circuits of not-too- large size. Fortunately, the result of Limaye, Srinivasan and Tavenas [LST21] yields sub-exponential sized hitting sets for constant depth formulas and that enables us to avoid the use of randomness to prepare for the Hensel Lifting step. We can then factorise the univariate polynomial $f({\bm{\delta}},y)$ and attempt all possible factors $g_{0}$ of degree at most $d$ to begin the lifting process from $g_{0}\cdot h_{0}$ (where $h_{0}=f({\bm{\delta}},y)/g_{0}$). After an appropriately large lift, we have small circuits (of possibly unbounded depth) computing $g_{\ell}$ and $h_{\ell}$ such that $\tilde{f}=f(\mathbf{x}+{\bm{\delta}},y)=g_{\ell}\cdot h_{\ell}\bmod{\mathcal{I}^{2^{\ell}}}$. If $g_{\ell}$ is guaranteed to be monic, and the initial choice of $g_{0}$ was indeed $g({\bm{\delta}},y)$, the uniqueness of Hensel lifting would ensure that $g_{\ell}$ is indeed equal to $g$ (after truncating higher order terms). We can then use standard interpolation to obtain $g_{\ell}$ explicitly written as a sum of monomials. Finally, to ensure that $g_{\ell}$ is indeed a legitimate factor of $\tilde{f}$, we perform divisibility testing to check if $g_{\ell}\mid\tilde{f}$. ##### Handling factors of large multiplicity: The above overview is all we need to obtain any factor $g$ of degree $O(1)$ that divides $f$ with $g^{2}\nmid f$. In order to handle factors with higher “factor-multiplicity”, we use a simple observation that $g^{a-1}\mid f$ but $g^{a}\nmid f$ if and only if $g$ divides $f,\partial_{y}f,\ldots,\partial_{y^{a-1}}f$ but not $\partial_{y^{a}}f$. We run our algorithm for each of the partial derivatives to collect the list of candidate factors, and eventually prune them via appropriate divisibility tests. ##### The specific case of $\Sigma\Pi$-formulas (or sparse polynomials): The above sketch yields a sub-exponential time algorithm for obtaining $O(1)$-degree factors of constant depth formulas. However, with some additional care, we obtain a quasipolynomial time algorithm in the case when $f$ is a sparse polynomial. The key observation for this is that we do not really need $f$ to be made monic for the above approach, but we only need $g$ to be monic to exploit the uniqueness of Hensel lifts. Since $g$ is a polynomial of degree at most $d=O(1)$, we can find a _low Hamming weight_ vector ${\bm{\alpha}}$ such that $g(\mathbf{x}+y{\bm{\alpha}})$ is monic in $y$. This allows us to control the sparsity increase of $f$ in the process and we show that the relevant resultant is a polynomial of the form $\sum_{i}\text{monomial}_{i}\cdot(\text{$O(1)$-degree})^{e_{i}}.$ Forbes [For15] shows that there are quasipolynomial size hitting sets for such expressions and we use this instead of the more general hitting set of Limaye, Srinivasan and Tavenas [LST21]. ### Organization of the paper The rest of the paper is organized as follows. In the next section, we start with a discussion of some of the preliminaries and known results from algebraic complexity and previous works on polynomial factorization that we use for the design and analysis of our algorithms. In Section 3, we describe and analyze the algorithm for computing low degree factors of multiplicity one of a given constant depth formula. In Section 4, we build upon this algorithm to compute arbitrary constant degree factors and complete the proofs of Theorem 1.1 and Theorem 1.2. Finally, we conclude with some open problems in Section 5. ## 2 Notation and preliminaries This section consists of all the necessary building blocks to describe and analyse (in Section 3) the main algorithm. ##### Fair warning: A large part of this (slightly lengthy) section is standard techniques in algebraic complexity that are relevant to this specific context, and is intended to keep the main analysis as self-contained as possible. A reader with some familiarity with standard algorithmic and structural results in algebraic complexity might be in a position to directly proceed to Section 3 and revisit this section for relevant results as required. #### Notation 1. 1. Throughout this paper, we work over the field $\mathbb{Q}$ of rational numbers. For some of the statements that are used more generally, we use $\mathbb{F}$ to denote an underlying field. 2. 2. We use boldface lower case letters like $\mathbf{x},\mathbf{y},\mathbf{a}$ to denote tuples, e.g. $\mathbf{x}=(x_{1},x_{2},\ldots,x_{n})$. The arity of the tuple is either stated or will be clear from the context. 3. 3. For a polynomial $f$ and a non-negative integer $k$, $\operatorname{Hom}_{k}[f]$ denotes the homogeneous component of $f$ of degree _equal_ to $k$. $\operatorname{Hom}_{\leq k}[f]$ denotes the sum of homogeneous components of $f$ of degree at most $k$, i.e., $\operatorname{Hom}_{\leq k}[f]:=\sum_{i=0}^{k}\operatorname{Hom}_{i}[f].$ 4. 4. The _sparsity_ of a polynomial $f$ is the number of monomials with a non-zero coefficient in $f$. 5. 5. For a parameter $k\in\mathbb{Z}_{\geq 0}$, we will use $(\Sigma\Pi)^{(k)}$ to refer to product-depth $k$ circuits222We emphasize that this notation does _not_ refer to the $k^{\text{th}}$ power of a polynomial computed by a $\Sigma\Pi$ circuit. with the root gate being $+$ and the deepest layer of gates being $\times$. Since any constant depth algebraic circuit of depth $k$ and size $s$ can be converted to a formula of depth $k$ and size $s^{k+1}$ i.e. $\operatorname{poly}(s)$, we will use the terms circuits and formulas interchangeably, without any loss in the final bounds we prove. 6. 6. Let $f$ and $g$ be multivariate polynomials such that $g\mid f$. Then, the _multiplicity_ or _factor multiplicity_ of $g$ in $f$ is defined to be the greatest integer $a$ such that $g^{a}$ divides $f$. ### 2.1 Circuit/formula bit-complexity ###### Definition 2.1 (Bit-complexity of a circuit/formula). The _bit-complexity_ of a circuit/formula $C$, denoted by $\operatorname{bit}(C)$, is defined as the sum of $\operatorname{size}(C)$ and the bit-complexities of all the scalars333For a rational number $r=p/q$, its bit-complexity $\operatorname{bit}(r)$ is defined as $\log(\max(\left|p\right|,\left|q\right|))$ present on edges or leaves. By default, any edge that does not have a scalar on it will be assigned the scalar 1. ###### Lemma 2.2 (Bit-complexity of evaluations of formulas). Let $C$ be a formula of bit-complexity $s$ computing a polynomial $f(\mathbf{x})$. If $\mathbf{a}\in\mathbb{Q}^{n}$ with each entry of $\mathbf{a}$ having bit-complexity $b$, then the bit-complexity of $f(\mathbf{a})$ is at most $s\cdot b$. (Proof deferred to Appendix A) ### 2.2 Relevant subclasses of algebraic circuits We briefly define subclasses of algebraic circuits that we would use often in this paper. ###### Definition 2.3 (Power of low-degree polynomials). For a parameter $d\in\mathbb{Z}_{\geq 0}$, let $\operatorname{Deg}_{d}$ refer to the class of polynomials of degree at most $d$. We use $(\operatorname{Deg}_{d})^{\ast}$ to denote the class of polynomials that are powers of polynomials of degree at most $d$. ###### Definition 2.4 ($\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$-formulas). We will use $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ to denote the subclass of algebraic formulas that compute expressions of the form $\sum_{i}f_{i}\cdot g_{i}^{e_{i}}$ where each $f_{i}$ is a $(\Sigma\Pi)^{(k)}$ formula and each $g_{i}$ is a polynomial of degree at most $d$ and $e_{i}$’s are arbitrary positive integers. The size and bit-complexity of the above expression is defined as its size and bit-complexity when viewed as a general algebraic formula. ###### Observation 2.5. Let $\mathcal{C}$ be the class of $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formulas for fixed parameters $k$ and $d$. Suppose $P_{1},\ldots,P_{t}$ are polynomials computed by $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formulas of size $s$ and bit-complexity $b$ each. Then, * • $\sum_{i}P_{i}$ is computable by an $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formula of size at most $t\cdot s$ and bit-complexity at most $O(t\cdot b)$. * • $\prod_{i}P_{i}$ is computable by an $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formula of size at most $s^{O(t)}$ and bit-complexity at most $b^{O(t)}$. (Proof deferred to Appendix A.) ### 2.3 Standard preliminaries using interpolation ###### Lemma 2.6 (Univariate interpolation (Lemma 5.3 [Sap15])). Let $f(x)=f_{0}+f_{1}x+\cdots+f_{d}x^{d}$ be a univariate polynomial of degree at most $d$. Then, for any $0\leq r\leq d$ and there are444In fact, for any choice of distinct $\alpha_{0},\ldots,\alpha_{d}$, there are appropriate $\beta_{r0},\ldots,\beta_{rd}$ satisfying the equation. If the $\alpha_{i}$’s are chosen to have small bit-complexity, we can obtain a $\operatorname{poly}(d)$ bound on the bit-complexity of the associated $\beta_{ri}$’s. field constants $\alpha_{0},\ldots,\alpha_{d}$ and $\beta_{r0},\ldots,\beta_{rd}$ such that $f_{r}=\beta_{r0}f(\alpha_{0})+\cdots+\beta_{rd}f(\alpha_{d}).$ Furthermore, the bit-complexity of all field constants is bounded by $\operatorname{poly}(d)$. ###### Lemma 2.7 (Computing homogeneous components (Lemma 5.4 [Sap15])). Let $f\in\mathbb{Q}[\mathbf{x}]$ be an $n$-variate degree $d$ polynomial. Then, for an $0\leq i\leq d$, there are field constants $\alpha_{0},\ldots,\alpha_{d}$ and $\beta_{i0},\beta_{id}$ of bit-complexity $\operatorname{poly}(d)$ such that $\operatorname{Hom}_{i}(f)=\beta_{i0}f(\alpha_{0}\cdot\mathbf{x})+\cdots+\beta_{id}f(\alpha_{d}\cdot\mathbf{x}).$ In particular for $\mathcal{C}=(\Sigma\Pi)^{(k)}$ or $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$, if $f$ is computable by $\mathcal{C}$-formulas of size / bit-complexity at most $s$ then $\operatorname{Hom}_{i}(f)$ is computable by $\mathcal{C}$-formulas of size / bit-complexity at most $\operatorname{poly}(s,d)$. ###### Lemma 2.8 (Computing partial derivatives in one variable). Let $f\in\mathbb{Q}[\mathbf{x}]$ be an $n$-variate degree $d$ polynomial. Then, for an $0\leq r\leq d$, there are field elements $\alpha_{i}$’s and $\beta_{ij}$’s in $\mathbb{Q}$ of bit-complexity $\operatorname{poly}(d)$ such that $\frac{\partial^{r}f}{\partial x_{1}^{r}}=\sum_{i=0}^{d}x_{1}^{i}\cdot\left(\beta_{i0}f(\alpha_{0},x_{2},\ldots,x_{n})+\cdots+\beta_{id}f(\alpha_{d},x_{2},\ldots,x_{n})\right)$ In particular for $\mathcal{C}=(\Sigma\Pi)^{(k)}$ or $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$, if $f$ is computable by $\mathcal{C}$-formulas of size / bit-complexity at most $s$ then $\frac{\partial^{r}f}{\partial x_{1}^{r}}$ is computable by $\mathcal{C}$-formulas / bit-complexity of size at most $O(s\cdot d^{3})$. ###### Proof. We may consider the polynomial $f$ as a univariate in $x_{1}$, and extract each coefficient of $x_{1}^{i}$ using 2.6 and recombine them to get the appropriate partial derivative. That justifies the claimed expression. As for the size, note that if $\mathcal{C}$ is $(\Sigma\Pi)^{(k)}$ or $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$, multiplying a size $s$ formula by $x_{1}^{i}$, by using distributivity of the top addition gate, results in a $\mathcal{C}$-formula of size at most $s\cdot d$. Thus, the overall size of the above expression for the partial derivative is at most $O(s\cdot d^{3})$. ∎ We will be making use of the following identity, which can be proved via appropriate interpolation or by the inclusion-exclusion principle (along the lines of Lemma 2.2 [Shp02]). ###### Lemma 2.9 (Fischer’s identity [Fis94, Ell69, Shp02]). If $\mathbb{F}$ is a field of characteristic zero or larger than $D$, then for any positive integers $e_{1},\ldots,e_{n}$ with $\sum e_{i}=D$ and for $r\leq\prod_{i=1}^{n}\left(e_{i}+1\right)$, there are homogeneous linear forms $L_{1},\ldots,L_{r}$ and field constants $\alpha_{1},\ldots,\alpha_{r}$ of bit-complexity $\operatorname{poly}(d,n)$ such that $x_{1}^{e_{1}}\cdots x_{n}^{e_{n}}=\sum_{i=1}^{r}\alpha_{i}L_{i}^{D}.$ ### 2.4 Polynomial identity testing ###### Lemma 2.10 (Polynomial Identity Lemma [Ore22, DL78, Sch80, Zip79]). Let $f\in\mathbb{Q}[\mathbf{x}]$ be a non-zero $n$ variate polynomial of degree at most $d$. Then, for every set $S\subseteq\mathbb{Q}$, the number of zeroes of $f$ in the set $S^{n}=S\times S\times\cdots\times S$ is at most $d|S|^{n-1}$. ###### Definition 2.11 (Low Hamming weight set). Let $n\geq d\geq 0$ be integer parameters. Fix a set $T_{d}\subseteq\mathbb{Q}$ of size $(d+1)$. The set $\mathcal{H}(d,n)$ is defined as $\mathcal{H}(d,n)=\left\\{(a_{1},\ldots,a_{n})\ :\ S\in\binom{[n]}{\leq d}\;,\;a_{i}\in T_{d}\text{ for all }i\in S\text{ and }a_{j}=0\text{ for all }j\notin S\right\\}.$ The size of the above set is at most $\binom{n}{\leq d}\cdot(d+1)^{d}=n^{O(d)}$. Furthermore, choosing $T_{d}$ to consist of elements of $\mathbb{Q}$ of bit-complexity $\operatorname{poly}(d)$, the bit- complexity of the set $\mathcal{H}(d,n)$ is bounded by $n^{O(d)}$ as well. The following lemma is an easy consequence of Lemma 2.10 and will be crucial for parts of our proof. We also include a short proof sketch. ###### Lemma 2.12 (Hitting set for low degree polynomials). Let $f\in\mathbb{Q}[\mathbf{x}]$ be a non-zero $n$ variate polynomial of degree at most $d$. Then, there exists a vector $\mathbf{a}\in\mathcal{H}(d,n)\subseteq\mathbb{Q}^{n}$ such that $f(\mathbf{a})\neq 0$. (Proof deferred to Appendix A.) ###### Theorem 2.13 (PIT for constant depth formulas (modification of Corollary 6 [LST21])). Let $\varepsilon>0$ be a real number and $\mathbb{F}$ be a field of characteristic 0. Let $C$ be an algebraic formula of size and bit-complexity $s\leq\operatorname{poly}(n)$, depth $k=o(\log\log\log n)$ computing a polynomial on $n$ variables, then there is a deterministic algorithm that can check whether the polynomial computed by $C$ is identically zero or not in time $(s^{O(k)}\cdot n)^{O_{\varepsilon}((sD)^{\varepsilon})}$. The original statement of Corollary 6 in [LST21] deals specifically with circuits of size $s=\operatorname{poly}(n)$. The above statement can be readily inferred from their proof. ###### Theorem 2.14 (PIT for $\Sigma\left((\Sigma\Pi)^{(1)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ (Corollary 6.7, [For15])). Let $t\geq 1$. Then, the class $\mathcal{C}=\Sigma\left((\Sigma\Pi)^{(1)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ that computes polynomials of the form $\sum_{i=1}^{s}f_{i}\cdot g_{i}^{d_{i}}$ with each $f_{i}$ being $s$-sparse and each $\deg(g_{i})\leq d$ has a $\operatorname{poly}(n,s,d\log s)$-explicit hitting set of size $\operatorname{poly}(s)^{O(d\log s)}$. We will also crucially use the following lemma that gives an algorithm to obtain the coefficient vector of a polynomial from an algebraic formula computing it. In our setting, we invoke this algorithm only for low degree polynomials, and in that case, we can tolerate the runtime of this algorithm within our budget. ###### Lemma 2.15 (Interpolating a low degree multivariate polynomial). There is a deterministic algorithm that, when given a parameter $d$ and an $n$ variate algebraic formula $C\in\mathbb{Q}[\mathbf{x}]$ of size at most $s$, bit-complexity at most $b$ and degree at most $d$, outputs the coefficient vector of the polynomial computed by $C$. The algorithm runs in time $\operatorname{poly}(s,b,n^{d})$. (Proof deferred to Appendix A.) ### 2.5 Deterministic divisibility testing and PIT ###### Definition 2.16 (Pseudo-quotients). Let $f,g\in\mathbb{Q}[\mathbf{x}]$ be non-zero polynomials with $g({\bm{0}})=\beta\neq 0$. The _pseudo-quotient of $f$ and $g$_ is defined as $\operatorname{Hom}_{\leq d_{f}-d_{g}}\left(\left(\frac{f(\mathbf{x})}{\beta}\right)\cdot(1+\tilde{g}+\tilde{g}^{2}+\cdots+\tilde{g}^{d_{f}-d_{g}})\right)$ where $d_{f}=\deg(f)$, $d_{g}=\deg(g)$ and $\tilde{g}=1-\frac{g}{\beta}$. More generally, if ${\bm{\alpha}}\in\mathbb{Q}^{n}$ is such that $g({\bm{\alpha}})\neq 0$, the _pseudo-quotient of $f$ and $g$ translated by ${\bm{\alpha}}$_ is defined as the pseudo-quotient of $f(\mathbf{x}+{\bm{\alpha}})$ and $g(\mathbf{x}+{\bm{\alpha}})$. The following lemma immediately follows from the above definition and Lemma 2.7. ###### Lemma 2.17 (Complexity of pseudo-quotients). Suppose $k\geq 1$ and $f(\mathbf{x})\in(\Sigma\Pi)^{(k)}$ and $g(\mathbf{x})\in\operatorname{Deg}_{d}$ of sizes at most $s_{1},s_{2}$ respectively, and suppose $g({\bm{0}})\neq 0$. Then, the pseudo-quotient of $f,g$ is computable by the $\mathcal{C}$-formulas of size at most $\operatorname{poly}(s_{1},s_{2})$, where $\mathcal{C}=\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$. ###### Theorem 2.18 (Divisibility testing to PIT [For15]). Let $f(\mathbf{x})$ and $g(\mathbf{x})$ be non-zero $n$-variate polynomials over a field $\mathbb{Q}$ such that $g(\mathbf{0})=\beta\neq 0$. Then, $g$ divides $f$ if and only if the polynomial $R(\mathbf{x})$ defined as $R(\mathbf{x}):=f(\mathbf{x})-g(\mathbf{x})Q(\mathbf{x})$ is identically zero, where $Q(\mathbf{x})$ is the pseudo-quotient of $f$ and $g$. An immediate consequence of this theorem is the following corollary that takes into account the depth of an algebraic formula computing the polynomial $R(\mathbf{x})$ given above, assuming that $f$ and $g$ themselves can be computed by a low depth formula. ###### Corollary 2.19 (Divisibility testing to PIT for constant depth formulas [For15]). Suppose $f(\mathbf{x})$ is a non-zero $n$-variate polynomial computed by a $(\Sigma\Pi)^{(k)}$ formula of size $s$, and suppose $g(\mathbf{x})$ is a polynomial of degree at most $d$ with $g({\bm{0}})=\beta\neq 0$. Then, we can test if $g$ divides $f$ in time $T(k,d,s^{\prime})$ where $s^{\prime}=\operatorname{poly}(s,d)$ and $T(k,d,s)$ is the time required to test polynomial identities of the size $s$ expressions of the form $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right).$ (Proof deferred to Appendix A.) ###### Theorem 2.20 ([For15]). Let $\mathbb{F}$ be any sufficiently large field. Then, there is a deterministic algorithm that takes an input two polynomials $f$ and $g$ and parameters $d,D,n,s$, where $f$ is an $n$-variate polynomial of degree at most $D$ and sparsity $s$; $g$ is an $n$ variate polynomial of degree $d$, and outputs whether $g$ divides $f$ or not in time $\exp(O(d\log^{2}snDd))$. ### 2.6 Resultants ###### Definition 2.21 (The Resultant). Let $\mathcal{R}$ be a commutative ring. Given polynomials $g$ and $h$ in $\mathcal{R}[y]$, where: $\displaystyle g(y)$ $\displaystyle=g_{0}+\cdots+y^{d}\cdot g_{d}$ $\displaystyle h(y)$ $\displaystyle=h_{0}+y\cdot h_{1}+\cdots+y^{D}\cdot h_{D}$ with $g_{d}$ and $h_{D}\neq 0$ the _Resultant_ of $g$ and $h$, denoted by $\operatorname{Res}_{y}(g,h)$, is the determinant of the $(D+d)\times(D+d)$ Sylvester matrix $\Gamma$ of $g$ and $h$, given by: $\Gamma=\begin{bmatrix}h_{0}&h_{1}&\dots&&h_{D}&&\\\ &\ddots&\ddots&&\ddots&\ddots&\\\ &&h_{0}&h_{1}&&\dots&h_{D}\\\ g_{0}&\dots&&g_{d}&&&\\\ &g_{0}&\dots&&g_{d}&&\\\ &&\ddots&\ddots&&\ddots&\\\ &&&g_{0}&\dots&&g_{d}\end{bmatrix}$ ###### Lemma 2.22 (Resultant and $\gcd$ (Corollary 6.20 [vzGG13])). Let $\mathcal{R}$ be a unique factorization domain and $g,h\in\mathcal{R}[y]$ be non-zero polynomials. Then: $\deg_{y}(\gcd(g,h))>0\iff\operatorname{Res}_{y}(g,h)=0$ where $\gcd(g,h)\in\mathcal{R}[y]$ and $\operatorname{Res}_{y}(g,h)\in\mathcal{R}$. In this paper, $\mathcal{R}$ will be $\mathbb{Q}[\mathbf{x}]$ (which is a unique factorization domain), and $\operatorname{Res}_{y}(g,h)$ will denote the resultant of $g,h\in\mathbb{Q}[\mathbf{x}][y]$ when considered as polynomials in $\mathcal{R}[y]$. We might also occasionally refer to it as the _y-resultant_ of $g$ and $h$. For more details about the resultant as well as a proof of the above lemma, we refer the reader to von zur Gathen and Gerhard’s book on computer algebra (Chapter 6, [vzGG13]). We mention a simple observation from the above definition that would be useful for this paper. ###### Observation 2.23 (Resultant under substitutions). Suppose $g(\mathbf{x},y)=g_{0}(\mathbf{x})+g_{1}(\mathbf{x})y+\cdots g_{d}(\mathbf{x})y^{d}$ and $h(\mathbf{x},y)=h_{0}(\mathbf{x})+h_{1}(\mathbf{x})y+\cdots+h_{D}(\mathbf{x})y^{D}$ with $g_{d},h_{D}\neq 0$. Then, for any $\mathbf{a}\in\mathbb{Q}^{\left|\mathbf{x}\right|}$ that ensures $g_{d}(\mathbf{a}),h_{D}(\mathbf{a})\neq 0$, we have $(\operatorname{Res}_{y}(g,h))(\mathbf{a})=\operatorname{Res}_{y}(g(\mathbf{a},y),h(\mathbf{a},y)).$ ### 2.7 Hensel Lifting Now we will state the definition of a _lift_ and the main lemma for Hensel lifting. For more details, one can look up some of the cited papers or the standard references in computational algebra [KSS15, ST20, vzGG13, Sud98]. ###### Definition 2.24 (Hensel lifts). Let $\mathcal{I}\subseteq\mathbb{Q}[\mathbf{x},y]$ be an ideal. Let $f,g,h,u,v\in\mathbb{Q}[\mathbf{x},y]$ such that $f\equiv gh\bmod{\mathcal{I}}$ and $ug+vh\equiv 1\bmod{\mathcal{I}}$. Then, we call $g^{\prime},h^{\prime}\in\mathbb{Q}[\mathbf{x},y]$ a _lift_ of $g$ and $h$ if: 1. 1. $f\equiv g^{\prime}h^{\prime}\bmod{\mathcal{I}^{2}}$, 2. 2. $g^{\prime}\equiv g\bmod{\mathcal{I}}$ and $h^{\prime}\equiv h\bmod{\mathcal{I}}$, and 3. 3. $\exists u^{\prime},v^{\prime}\in\mathbb{Q}[\mathbf{x},y]$ s.t $u^{\prime}g^{\prime}+v^{\prime}h^{\prime}\equiv 1\bmod{\mathcal{I}^{2}}$. For the rest of the section, we define $\mathcal{I}$ to be the ideal $\left\langle x_{1},\dots,x_{n}\right\rangle$ and $\mathcal{I}_{k}:=\mathcal{I}^{2^{k}}$. ###### Lemma 2.25 (Iterated monic Hensel lifting (Lemma 3.4 [KSS15])). Suppose we’re given $f\in\mathbb{Q}[\mathbf{x},y]$ such that $f=gh$, $g$ is monic in $y$ and $\gcd(g,h)=1$. We are also given $g_{0},h_{0},u_{0},v_{0}\in\mathbb{Q}[\mathbf{x},y]$ such that $g_{0}\equiv g\bmod\mathcal{I}$, $h_{0}\equiv h\bmod\mathcal{I}$ and $u_{0}g_{0}+v_{0}h_{0}\equiv 1\bmod\mathcal{I}$. Then, for all $k\in\mathbb{N},k\geq 1$, there exist $g_{k},h_{k},u_{k},v_{k}\in\mathbb{Q}[\mathbf{x},y]$ , with each $g_{k}$ being monic, such that the following conditions hold: 1. 1. The pair $g_{k},h_{k}$ is a lift of $g_{k-1},h_{k-1}$, with $u_{k}g_{k}+v_{k}h_{k}\equiv 1\bmod\mathcal{I}_{k}$; in particular, $f\equiv g_{k}h_{k}\bmod{\mathcal{I}_{k}}$ 2. 2. $g_{k}\equiv g\bmod{\mathcal{I}_{k}}$ and $h_{k}\equiv h\bmod{\mathcal{I}_{k}}$ Moreover, for each $k$, $g_{k}$ and $h_{k}$ are unique polynomials modulo $\mathcal{I}_{k}$ satisfying the above conditions when the $g_{k}$s are monic. For each $k$, we will call $g_{k},h_{k}$ the _$k$ -th iterated lift of $g_{0}$, $h_{0}$_. If $\deg_{\mathbf{x}}(g)=d$, we can choose an integer $k^{*}$ such that $d<2^{k^{*}}\leq 2d$ and use the above Lemma to get $g_{k^{*}}\equiv g\bmod{\mathcal{I}_{k^{*}}}$, which means we can truncate $g_{k^{*}}$ to degree $d$ and retrieve $g$. The next lemma tells us that this can be done with reasonable bounds on the parameters of the underlying circuits. ###### Lemma 2.26 (Small circuit for Hensel lifting (Lemma 3.6 [KSS15])). Let $f$ be a degree $D$ polynomial in $\mathbb{Q}[\mathbf{x},y]$, computable by a $(\Sigma\Pi)^{(k)}$ formula of size and bit-complexity $s$, with a factorization $f=gh$ such that $\gcd(g,h)=1$ and $g$ is monic. Let $g_{0}=g\bmod{\mathcal{I}}$ and $h_{0}=h\bmod{\mathcal{I}}$ be univariates in $\mathbb{Q}[y]$ with $\gcd(g_{0},h_{0})=1$. Then, there are formulas $C_{g},C_{h}$ of size and bit complexity $(sDk)^{O(k\log D)}$ that compute the $k^{\text{th}}$ iterated lift $g_{k}$,$h_{k}$ of $g_{0}$,$h_{0}$, where $g_{k}$ is monic. More generally, if the total degree of $g_{k}$ is at most $d$, then the size and bit complexity of the formula for $g_{k}$ is at most $(sDk)^{O(\log d)}$. Moreover, there is a deterministic algorithm, that when given the formulas for $f$ and $g_{0},h_{0}$ and integer $k$ as input, outputs the formulas for $g_{k}$ and $h_{k}$ in time $(sDk)^{O(k\log D)}$ ( resp. $(sDk)^{O(\log d)}$ if $g_{k}$ has total degree $d$). (Proof sketch deferred to Appendix A.) ### 2.8 Results on polynomial factorization We rely on the following two fundamental results on polynomial factorization for our results. The first theorem is a classical algorithm of Lenstra, Lenstra and Lovász for factoring univariate polynomials over the field of rational numbers. ###### Theorem 2.27 (Factorizing polynomials with rational coefficients [LLL82, vzGG13]). Let $f\in\mathbb{Q}[x]$ be a monic polynomial of degree $d$. Then there is a deterministic algorithm computing all the irreducible factors of $f$ that runs in time $\operatorname{poly}(d,t)$, where $t$ is the maximum bit-complexity of the coefficients of $f$. The second result we need is an easy consequence of the results of Kopparty, Saraf and Shpilka [KSS15]. They showed that an efficient deterministic algorithm for PIT for algebraic circuits implies an efficient deterministic algorithm for polynomial factorization. The formal statement below essentially invokes this for constant degree polynomials. In this case, the PIT instances also happen to be of constant degree and hence can be easily solved in time that is polynomial in the length of the coefficient vector of these polynomials. ###### Theorem 2.28 ([KSS15]). There is a deterministic algorithm that when given as input the coefficient vector of an $n$ variate polynomial $f(\mathbf{x})\in\mathbb{Q}[\mathbf{x}]$ of total degree $d$, runs in time $n^{O(d^{2})}$ and decides if $f$ is irreducible or not. ## 3 Computing candidate low-degree factors of multiplicity one We first present the algorithm for computing candidate low-degree factors of multiplicity one in Algorithm 1 below. In the next section, we use this as a subroutine in Algorithm 2 to compute factors of all multiplicity and also eliminate those candidates that were not actual factors. 1 Input : A $(\Sigma\Pi)^{(k)}$-formula of size $s$, bit-complexity $t$, degree $D$ computing a polynomial $f(\mathbf{x})$. Output : A list of polynomials of degree at most $d$, that include all factors of $f$ with degree at most $d$ and multiplicity $1$. 2 3Set the output list $L=\emptyset$. 4Compute hitting-set $H_{1}=\mathcal{H}(d,n)$ (as defined in Definition 2.11). 5Compute hitting-set $H_{2}$ for the class of $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$-formulas that have size $s^{\prime}\leq(sD)^{O(d)}$. (Lemma 3.2, Theorem 2.14) 6for _${\bm{\alpha}},{\bm{\beta}}\in H_{1}$ and ${\bm{\delta}}\in H_{2}$_ do 7 8 Define $F(\mathbf{x},y)=f(\mathbf{x}+{\bm{\alpha}}\cdot y+{\bm{\beta}}+{\bm{\delta}})=f(x_{1}+\alpha_{1}y+\beta_{1}+\delta_{1},\ldots,x_{n}+\alpha_{n}y+\beta_{1}+\delta_{n})$ 9 Using interpolation on the formula for $F(\mathbf{x},y)$ (via 2.6), compute $F({\bm{0}},y)$ as a sum of monomials. 10 Factorise the polynomial $F({\bm{0}},y)$ into irreducible factors as $F({\bm{0}},y)=\sigma\cdot F_{1}^{e_{1}}\cdots F_{r}^{e_{r}}.$ where $0\neq\sigma\in\mathbb{Q}$ and each $F_{r}$ is monic in $y$. 11 for _$T\subseteq[r]$ of size at most $d$_ do 12 13 Define $g_{0}=\prod_{i\in T}F_{i}^{e_{i}}$ and $h_{0}=\sigma\cdot\prod_{i\notin T}F_{i}^{e_{i}}$, interpretted as polynomials in $\mathbb{Q}[\mathbf{x},y]$ for Lemma 2.25 14 if _$\deg(g_{0}) >d$_ then 15 Continue to the next choice of $T$ in the current loop. 16 17 Compute polynomials $u_{0},v_{0}$ such that $u_{0}g_{0}+v_{0}h_{0}=1$. 18 Use Hensel-Lifting (Lemma 2.26) to lift the factorisation $F(\mathbf{x},y)=g_{0}(\mathbf{x},y)\cdot h_{0}(\mathbf{x},y)\bmod{I}$, where $I=\left\langle\mathbf{x}\right\rangle$, to obtain algebraic circuits for $g_{\ell},h_{\ell}$ satisfying $F(\mathbf{x},y)=g_{\ell}(\mathbf{x},y)\cdot h_{\ell}(\mathbf{x},y)\bmod{I^{2^{\ell}}}$ with $g_{\ell}$ being monic and $d<2^{\ell}<2d$. 19 Using interpolation on the circuit for $g_{\ell}$ (via Lemma 2.15), compute $g_{\ell}$ as a sum of monomials. 20 Add $\tilde{g}=g_{\ell}(\mathbf{x}-{\bm{\delta}}-{\bm{\beta}},0)$ to $L$. 21 return _$L$_ Algorithm 1 Computing candidate degree $d$ factors of factor-multiplicity one Before we discuss the proof of correctness and running time of Algorithm 1, we state two simple observations that we use in the analysis. We defer the proofs of these observations to the end of the section. ###### Observation 3.1 (Size growth under a translation of low Hamming weight). Let $k>0$ be a parameter. Let $f(\mathbf{x})$ be an $n$-variate polynomial of degree at most $D$ with $(\Sigma\Pi)^{k}\operatorname{-size}$ at most $s$. If ${\bm{\alpha}},{\bm{\beta}}\in\mathcal{H}(d,n)$, the polynomial $\tilde{f}(\mathbf{x},y)=f(\mathbf{x}+y{\bm{\alpha}}+{\bm{\beta}})$ has $(\Sigma\Pi)^{k}\operatorname{-size}$ at most $s\cdot D^{O(d)}$. ###### Lemma 3.2. Let $f(\mathbf{x})$ be an $n$-variate polynomial computed by a $(\Sigma\Pi)^{(k)}$ formula of size $s$, and let $g(\mathbf{x})$ be an $n$-variate degree $d$ polynomial with $g({\bm{0}})\neq 0$. Let $Q(\mathbf{x})$ be the pseudo-quotient of $f$ and $g$. Then, for any variable $y\in\mathbf{x}$, the polynomial $\operatorname{Res}_{y}(Q,g)$ is computable by a $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formula of size at most $s^{O(d)}$. ### 3.1 Proof of correctness of the Algorithm 1 ###### Lemma 3.3 (Correctness of Algorithm 1). For every input polynomial $f$ computed by $(\Sigma\Pi)^{(k)}$ formulas of size $s$, bit-complexity $t$, degree $D$ and any factor $g$ of degree at most $d$ with $g\mid f$ and $g^{2}\nmid f$, the polynomial $g$ is included in the output list of Algorithm 1 on input $f$. ###### Proof. Algorithm 1 outputs a list of candidate factors; we would like to prove that every factor of $f$ with degree $\leq d$ and factor-multiplicity one will be contained in this list. Fix any specific factor $g$ of $f$, with $\deg(g)=d^{\prime}\leq d$ and factor-multiplicity one, which ensures that $\gcd(g,f/g)=1$. 1. 1. Make $g$ monic and $g(\bm{0})\neq 0$ The coefficient of $y^{d^{\prime}}$ in $g^{\prime}(\mathbf{x},y):=g(\mathbf{x}+y{\bm{\alpha}}+{\bm{\beta}})$ is the evaluation of $\operatorname{Hom}_{d^{\prime}}(g)$ at $\bm{\alpha}$ and the constant term of $g^{\prime}(\mathbf{x},y)$ is $g^{\prime}({\bm{0}},0)=g({\bm{\beta}})$. Thus by Lemma 2.12, there is some ${\bm{\alpha}},{\bm{\beta}}\in H_{1}$ such that $\operatorname{Hom}_{d^{\prime}}(g)({\bm{\alpha}})\neq 0$ and $g({\bm{\beta}})\neq 0$. Fix this choice of ${\bm{\alpha}},{\bm{\beta}}$. We then have that $g^{\prime}(\mathbf{x},y)$ is monic in $y$, has $\deg_{y}(g^{\prime})=\deg(g)=d^{\prime}$, and has non-zero constant term. 2. 2. Bound the size of $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formula for the resultant With the above properties, the pseudo-quotient $h^{\prime}$ of $f^{\prime}(\mathbf{x},y):=f(\mathbf{x}+y{\bm{\alpha}}+{\bm{\beta}})$ and $g^{\prime}(\mathbf{x},y)$ is well-defined and is a polynomial in $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ (by Lemma 2.17) of size $\operatorname{poly}(s,D,d)\leq\operatorname{poly}(sD)$. By Lemma 3.2, $\operatorname{Res}_{y}(g^{\prime},h^{\prime})\in\mathbb{Q}[\mathbf{x}]$ is a non-zero polynomial computable by $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formulas of size $(sD)^{O(d)}$. 3. 3. Maintain $\gcd(g,h)=1$ condition in the univariate setting by hitting the resultant Let $\deg_{y}(h^{\prime})=r$ and $h^{\prime}(\mathbf{x},y)=h_{0}^{\prime}(\mathbf{x})+\cdots+h_{r}^{\prime}(\mathbf{x})y^{r}$. Since $h^{\prime}$ is computable by size $(sD)^{O(d)}$ formula from $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$, so is the leading term $h_{r}^{\prime}(\mathbf{x})$ by Lemma 2.7. Therefore by 2.5, the polynomial $\Gamma(\mathbf{x})=\operatorname{Res}_{y}(g^{\prime},h^{\prime})\cdot h^{\prime}_{r}(\mathbf{x})$ is also computable by $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formulas of size $s^{\prime}=(sD)^{O(d)}$. Since $H_{2}$ is a hitting set for $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formulas of size $s^{\prime}$, fix a ${\bm{\delta}}\in H_{2}$ such that $\Gamma({\bm{\delta}})\neq 0$ and in particular, the conditions required in 2.23 are true (note that the leading coefficient of $g^{\prime}$ is just 1 by monicness). By Lemma 2.22 and 2.23, we have that $g^{\prime}({\bm{\delta}},y)$ and $h^{\prime}({\bm{\delta}},y)$ are coprime polynomials. Thus, if $g^{\prime\prime}(\mathbf{x},y)=g^{\prime}(\mathbf{x}+{\bm{\delta}},y)$ and $h^{\prime\prime}(\mathbf{x},y)=h^{\prime}(\mathbf{x}+{\bm{\delta}},y)$ ($h^{\prime}$ being the pseudo-quotient), Theorem 2.18 implies that $\displaystyle f(\mathbf{x}+{\bm{\alpha}}y+{\bm{\beta}}+{\bm{\delta}})$ $\displaystyle=g^{\prime\prime}(\mathbf{x},y)\cdot h^{\prime\prime}(\mathbf{x},y)$ $\displaystyle\implies f({\bm{\alpha}}y+{\bm{\beta}}+{\bm{\delta}})$ $\displaystyle=g^{\prime\prime}({\bm{0}},y)\cdot h^{\prime\prime}({\bm{0}},y)$ $\displaystyle\quad\text{with }\gcd(g^{\prime\prime}({\bm{0}},y),h^{\prime\prime}({\bm{0}},y))=1.$ 4. 4. Univariate factorization and Hensel Lifting 10 thus factorises the univariate polynomial $f({\bm{\alpha}}y+{\bm{\beta}}+{\bm{\delta}})$ and one of the sets $T$ in Algorithm 1 must correspond to $g_{0}(y)$ chosen in Algorithm 1 to satisfy $g_{0}(y)=g^{\prime\prime}({\bm{0}},y)$ and $h_{0}(y)=h^{\prime\prime}({\bm{0}},y)$. Thus, we have a factorisation of the form $\displaystyle f({\bm{\alpha}}y+{\bm{\beta}}+{\bm{\delta}})$ $\displaystyle=g^{\prime\prime}({\bm{0}},y)\cdot h^{\prime\prime}({\bm{0}},y)=g_{0}\cdot h_{0}$ $\displaystyle\implies f(\mathbf{x}+{\bm{\alpha}}y+{\bm{\beta}}+{\bm{\delta}})$ $\displaystyle=g_{0}\cdot h_{0}\bmod{\mathcal{I}},\quad\text{where $\mathcal{I}=\left\langle\mathbf{x}\right\rangle$.}$ We are therefore set-up to apply Hensel Lifting (Lemma 2.25) to obtain $g_{\ell},h_{\ell}$ such that $g_{\ell}$ is monic in $y$ and $f(\mathbf{x}+{\bm{\alpha}}y+{\bm{\beta}}+{\bm{\delta}})=g_{\ell}(\mathbf{x},y)\cdot h_{\ell}(\mathbf{x},y)\bmod{\mathcal{I}^{2^{\ell}}}.$ From the uniqueness of Hensel Lifting (which is guaranteed by Lemma 2.25), we must have that $g_{\ell}(\mathbf{x},y)=g^{\prime\prime}(\mathbf{x},y)=g(\mathbf{x}+{\bm{\alpha}}y+{\bm{\beta}}+{\bm{\delta}})$. Thus, for this choice of ${\bm{\alpha}},{\bm{\beta}},{\bm{\delta}}$ and $T$, we would include $g(\mathbf{x})=g^{\prime\prime}(\mathbf{x}-{\bm{\beta}}-{\bm{\delta}},0)$ in the set of candidate factors in Algorithm 1. Finally, since the lift also ensures that there exist $u_{\ell}$ and $v_{\ell}$ such that $u_{\ell}g_{\ell}+v_{\ell}h_{\ell}=1\bmod{\mathcal{I}^{2^{\ell}}}$, we also have that $g_{\ell}^{2}\nmid f$. ∎ ### 3.2 Running time analysis We now bound the time complexity of the algorithm. ###### Lemma 3.4 (Running time of Algorithm 1). Let $\varepsilon>0,d,k\in\mathbb{N}$ be an arbitrary constants and let $f\in\mathbb{Q}[\mathbf{x}]$ be a polynomial computable by a $(\Sigma\Pi)^{(k)}$ formula $C$ of size $s$, degree at most $D$ and bit- complexity $t$. Then, on input $C$, Algorithm 1 terminates in time at most $(sD)^{O_{\varepsilon}(kd(sD)^{\varepsilon d})}\cdot t^{O(d\log d)}$. Moreover, if $k=1$, i.e. $f$ has sparsity at most $s$, then Algorithm 1 terminates in time at most $(sDt)^{(\operatorname{poly}(d)\log sDt)}$. ###### Proof. Let $T_{k}^{(1)}(s,d)$ be the time-complexity to output the hitting set $H_{1}$ in Algorithm 1 and $T_{k}^{(2)}(s,D,d)$ be the time-complexity to output the hitting set $H_{2}$ in Algorithm 1. From Definition 2.11, we immediately have that $T_{k}^{(1)}(s,d)\leq s^{O(d)}$. As for $T_{k}^{(2)}(s,D,d)$, in the case of $k=1$, Theorem 2.14 shows that $T_{k}^{(2)}(s,D,d)\leq(sD)^{(\operatorname{poly}(d)\log sD)}$. For $k$ satisfying $2\leq k=o(\log\log\log s)$, then Theorem 2.13 shows that $T_{k}^{(2)}(s,D,d)\leq(sD)^{O_{\varepsilon}(kd(sD)^{\varepsilon d})}$ for any constant $\varepsilon>0$. Using 2.6, we get that Algorithm 1 takes $\operatorname{poly}(s,D,t)$-time. Now, each of the coefficients of $F({\bm{0}},y)$ has bit-complexity at most $\operatorname{poly}(s,D,t)$. Thus, from Theorem 2.27, we get that $F({\bm{0}},y)$ can be factorized into its irreducible factors in time at most $\operatorname{poly}(s,D,t)$. There are at most $D^{d}$ choices for the set $T$ in Algorithm 1. For each such choice, Algorithms 1 to 1 compute formulas of size $\operatorname{poly}(s,D,t)$ for $g_{0}$, $h_{0}$, $u_{0}$, $v_{0}$ in time $\operatorname{poly}(s,D,t)$. By Lemma 2.26, we have that 18 takes time $(sDt)^{O(\log d)}$ to compute a formula of the same size and bit-complexity for $g_{\ell}$. From 2.15, we get that we can obtain the coefficient vector of $g_{\ell}$ in time at most $(sDt)^{O(d\log d)}$. Therefore, the overall running time of Algorithm 1 is at most $T_{k}^{(1)}(s,d)\cdot T_{k}^{(2)}(s,D,d)\cdot D^{d}\cdot\operatorname{poly}(s,D,t)\cdot(sDt)^{O(d\log d)}\,.$ Plugging in the estimates for $T_{k}^{(1)}(s,d)$, $T_{k}^{(2)}(s,D,d)$, we get the overall bound of $(sD)^{O_{\varepsilon}(kd(sD)^{\varepsilon d})}\cdot t^{O(d\log d)}$ for $k>1$, which is essentially dominated by $T_{k}^{(2)}(s,D,d)$. When $f$ has sparsity $s$, then as discussed in the proof, $T_{k}^{(2)}(s,d)$ is at most $(sD)^{(\operatorname{poly}(d)\log sD)}$. Plugging this back in the above expression, we get that the running time is at most $(sDt)^{(\operatorname{poly}(d)\log sDd)}$. ∎ ### 3.3 Proof of structural lemmas In this subsection, we include the proofs of 3.1 and Lemma 3.2. This completes the analysis of Algorithm 1. ###### Proof of 3.1. By definition of $\mathcal{H}(d,n)$ (Definition 2.11), the transformation $\mathbf{x}\mapsto\mathbf{x}+y{\bm{\alpha}}+{\bm{\beta}}$ takes a monomial $\prod_{i\in[n]}{x_{i}^{e_{i}}}$ to $\left(\prod_{i\in T}{\left(x_{i}+\alpha_{i}y+\beta_{i}\right)^{e_{i}}}\right)\cdot\left(\prod_{i\in[n]\setminus T}x_{i}^{e_{i}}\right)$, for some $T\subseteq[n]$ s.t. $|T|=d$. If we expand $\prod_{i\in T}{\left(x_{i}+\alpha_{i}y+\beta_{i}\right)^{e_{i}}}$ into a sum of monomials, we will get at most $D^{O(d)}$ monomials (when $\sum_{i}e_{i}\leq D$). Expanding each $\prod_{i\in[n]}{\left(x_{i}+\alpha_{i}y+\beta_{i}\right)^{e_{i}}}$ at the bottom layer into a sum of monomials this way, we get the required $(\Sigma\Pi)^{(k)}$ formula with size at most $s\cdot D^{O(d)}$. ∎ ###### Proof of Lemma 3.2. Let $\mathbf{x}^{\prime}=\mathbf{x}\setminus\left\\{y\right\\}$ and let $\mathcal{C}$ be the class $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$. Let us assume that $\deg_{y}(Q)=D\leq s$ and $\deg_{y}(g)=d$. By Lemma 2.17, we have that $Q(\mathbf{x})$ is computable by a $\mathcal{C}$-formula of size at most $\operatorname{poly}(s,d)$. Let us consider the $(D+d)\times(D+d)$ Sylvester matrix $\Gamma$ of $Q$ and $g$ with respect to the variable $y$ whose determinant is $\operatorname{Res}_{y}(Q,g)$. $\displaystyle Q(\mathbf{x})$ $\displaystyle=Q_{0}(\mathbf{x}^{\prime})+y\cdot Q_{1}(\mathbf{x}^{\prime})+\cdots+y^{D}\cdot Q_{D}(\mathbf{x}^{\prime})$ $\displaystyle g(\mathbf{x})$ $\displaystyle=g_{0}(\mathbf{x}^{\prime})+\cdots+y^{d}\cdot g_{d}(\mathbf{x}^{\prime})$ $\displaystyle\Gamma$ $\displaystyle=\begin{bmatrix}Q_{0}&Q_{1}&\dots&&Q_{D}&&\\\ &\ddots&\ddots&&\ddots&\ddots&\\\ &&Q_{0}&Q_{1}&&\dots&Q_{D}\\\ g_{0}&\dots&&g_{d}&&&\\\ &g_{0}&\dots&&g_{d}&&\\\ &&\ddots&\ddots&&\ddots&\\\ &&&g_{0}&\dots&&g_{d}\end{bmatrix}$ Note that, by 2.6, each of the $Q_{i}$’s are computed by a $\mathcal{C}$-formula of size $\operatorname{poly}(s,D)$ and each $g_{i}$ is a polynomial of degree at most $d$. For a subset $S$ of rows and $T$ of columns, we will use $\Gamma(S,T)$ to refer to the submatrix restricted to the rows in $S$ and columns in $T$, and let $\operatorname{Top}=\left\\{1,\ldots,d\right\\}$ and $\operatorname{Bot}=\left\\{d+1,\ldots,d+D\right\\}$. The determinant of $\Gamma$ can then be expressed as $\det(\Gamma)=\operatorname{Res}_{y}(Q,g)=\sum_{T\in\binom{[D+d]}{d}}\det(\Gamma(\operatorname{Top},T))\cdot\det(\Gamma(\operatorname{Bot},\overline{T}))$ For every choice of $T$, the polynomial $\det(\Gamma(\operatorname{Top},T))$ is the determinant of a $d\times d$ matrix each of whose entries are computable by $s^{\prime}=\operatorname{poly}(s,D)$ sized $\mathcal{C}$-formulas. Therefore, using 2.5, the polynomial $\det(\Gamma(\operatorname{Top},T))$ is computable by $\mathcal{C}$-formulas of size at most $(sD)^{O(d)}$. The polynomial $\det(\Gamma(\operatorname{Bot},\overline{T}))$ is a degree $D$ polynomial combination of $g_{0},\ldots,g_{d}$ and can therefore be expressed as $\displaystyle\det(\Gamma(\operatorname{Bot},\overline{T}))$ $\displaystyle=\sum_{i=1}^{D^{d+1}}a_{i}\cdot g_{0}^{e_{i,0}}\cdots g_{d}^{e_{i,d}}$ $\displaystyle=\sum_{i=1}^{D^{d+1}}a_{i}\cdot\left(\sum_{j=1}^{D^{O(d)}}b_{ij}\cdot f_{ij}^{e_{ij}}\right)\quad\text{(using \lx@cref{creftype~refnum}{lem:fischers-trick})}.$ for some polynomials $f_{j}$ of degree at most $d$. Thus, using 2.5 again, we have that $\operatorname{Res}_{y}(Q,g)$ is computable by $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formulas of size at most $D^{O(d)}\cdot(sD)^{O(d)}=(sD)^{O(d)}$. ∎ ## 4 Computing factors of all multiplicity The following lemma essentially shows that the multiplicity of any factor $g$ of a given polynomial $f$ can be reduced by working with appropriate partial derivatives of $f$, with respect to variables that are present in $g$. This naturally yields an algorithm that uses Algorithm 1 as a subroutine, and computes all irreducible factors of $f$. ###### Lemma 4.1 (Reducing factor multiplicity). Let $f(\mathbf{x}),g(\mathbf{x})\in\mathbb{Q}[\mathbf{x}]$ be non-zero polynomials and let $x\in\mathbf{x}$ be such that $\partial_{x}(g)\neq 0$ and $g$ is square-free. Then, the factor-multiplicity of $g$ in $f$ (i.e. the integer $a$ satisfying $g^{a}\mid f$ and $g^{a+1}\nmid f$) is also the smallest non-negative integer $a$ such that $g\nmid\frac{\partial^{a}f}{\partial x^{a}}$. ###### Proof. If the factor-multiplicity of $g$ in $f$ is zero, i.e. $g\nmid f$, then claim is clearly true. Thus let us assume that the factor-multiplicity of $g$ in $f$ is $a\geq 1$. It suffices to show that the factor-multiplicity of $g$ in $\partial_{x}(f)$ is exactly $a-1$. Suppose $f=g^{a}\cdot h$ where $\gcd(g,h)=1$. Then, $\partial_{x}f=\partial_{x}(g^{a})\cdot h+g^{a}\cdot\partial_{x}(h)=g^{a-1}\cdot(a\cdot\partial_{x}(g)\cdot h+g\cdot\partial_{x}(h)).$ Hence, we have that the factor-multiplicity of $g$ in $\partial_{x}(f)$ is at least $(a-1)$. On the other hand, we have that $\partial_{x}(g)\neq 0$ and $g$ is square-free and hence $\gcd(g,\partial_{x}(g))=1$. Therefore $\gcd(g,a\cdot g\cdot\partial_{x}(h)+h\cdot\partial_{x}(g))=\gcd(g,h\cdot\partial_{x}(g))=\gcd(g,h)=1$ and hence $g^{a}\nmid\partial_{x}(f)$ and therefore the factor-multiplicity of $f$ ∎ We are now ready to describe the algorithm. 1 Input : A $(\Sigma\Pi)^{(k)}$-formula of size $s$, bit-complexity $t$, degree $D$ computing a polynomial $f(\mathbf{x})$. Output : A list of all irreducible factors $f$ of degree at most $d$ and their multiplicities. 2 3Set the output list $L=\emptyset$. 4Set the intermediate candidates list $L^{\prime}=\emptyset$. 5Compute hitting-set $H_{1}=\mathcal{H}(d,n)$ (as defined in Definition 2.11). 6for _${\bm{\alpha}}\in H_{1}$_ do 7 Define $F(\mathbf{x},y)=f(\mathbf{x}+{\bm{\alpha}}\cdot y)=f(x_{1}+\alpha_{1}y,\ldots,x_{n}+\alpha_{n}y)$ 8 for _$i=0,1,\ldots,\deg(F)$_ do 9 10 Define $\tilde{F}(\mathbf{x},y)=\frac{\partial^{i}F}{\partial y^{i}}$. 11 Compute the list $\tilde{L}$ of all candidate degree $d$ multiplicity-one factors of $\tilde{F}(\mathbf{x},y)$ using Algorithm 1. 12 foreach _$\tilde{g}(\mathbf{x},y)\in\tilde{L}$_ do 13 14 Add $g(\mathbf{x}):=\tilde{g}(\mathbf{x},0)$ to $L^{\prime}$. 15 16 17 18 19for _$g\in L^{\prime}$_ do 20 21 if $g$ is not irreducible then skip to the next iteration. 22 Let $x$ be a variable that $g$ depends on, so that $\partial_{x}(g)\neq 0$. 23 Find the smallest non-negative integer $e$ such that $g\nmid\frac{\partial^{e}f}{\partial x^{e}}$. 24 if $e>1$ then add $(g,e)$ to the list $L$. return _$L$_ Algorithm 2 Computing list of all degree $d$ irreducible factors and their multiplicities ###### Lemma 4.2 (Correctness of Algorithm 2). For every input polynomial $f$ computed by a $(\Sigma\Pi)^{(k)}$ formula of size $s$, degree $D$, bit-complexity $t$ and $d\in\mathbb{N}$, the list $L$ output by Algorithm 2 is precisely the list of all irreducible factors of $f$ of degree at most $d$ (up to scalar multiplication) along with their multiplicities in $f$. ###### Proof. From Algorithms 2 to 2 and Lemma 4.1, it is clear that any $(g,e)$ in the output list ensures that $g$ is an irreducible polynomial, $g^{e}\mid f$ and $g^{e+1}\nmid f$. Thus, it suffices to show that for every irreducible polynomial $g$ such that $\deg(g)\leq d$ and $g\mid f$, some non-zero scalar multiple of $g$ is under consideration in the list $L^{\prime}$. Fix any such irreducible factor $g$ of degree at most $r\leq d$ and let its factor- multiplicity be $e$ By Lemma 2.12, there is some ${\bm{\alpha}}\in H_{1}$ such that $\operatorname{Hom}_{r}(g)({\bm{\alpha}})\neq 0$, where $r$ is the total degree of $g$. Thus, for this choice of ${\bm{\alpha}}$, we have that $g^{\prime}(\mathbf{x},y)=g(\mathbf{x}+y{\bm{\alpha}})$ is a factor of $F(\mathbf{x},y)=f(\mathbf{x}+y{\bm{\alpha}})$ and $g^{\prime}$ is monic in $y$ and has factor-multiplicity $e$. By Lemma 4.1, we have that $g^{\prime}$ has factor-multiplicity one in $\tilde{F}(\mathbf{x},y):=\frac{\partial^{e-1}F}{\partial y^{e-1}}$. Thus, by the correctness of Algorithm 1 (Lemma 3.3), a non-zero multiple of the polynomial $g^{\prime}(\mathbf{x},y)$ must be included in the list $\tilde{L}$ in Algorithm 2. Therefore, a non-zero multiple of $g(\mathbf{x})=g^{\prime}(\mathbf{x},0)$ will be added to $L^{\prime}$ in Algorithm 2. ∎ ###### Lemma 4.3 (Running time of Algorithm 2). Let $\varepsilon>0,k,d\in\mathbb{N}$ be arbitrary constants. Let $f\in\mathbb{Q}[\mathbf{x}]$ be a polynomial computable by a $(\Sigma\Pi)^{(k)}$ formula $C$ of size $s$, degree at most $D$ and bit- complexity $t$. Then, on input $C$ and $d\in\mathbb{N}$, Algorithm 1 terminates in time at most $(sDt)^{O(kd(sDt)^{\varepsilon d})}$. Moreover, if $k=1$, i.e. $f$ has sparsity at most $s$, then Algorithm 1 terminates in time at most $(snDt)^{O(\operatorname{poly}(d)\cdot\log snDt)}$. ###### Proof. From Definition 2.11, we have the size of the set $H_{1}$ is $n^{O(d)}$. The time complexity of computing a formula for $F$ from the given formula for $f$ is at most $O(sD)$. From Lemma 2.8, we have that $(\Sigma\Pi)^{k+1}$ formulas for all the $y$ derivatives of $F$ can be computed in time at most $\operatorname{poly}(s,D,t)$, which is also a bound on the bit-complexity and the size of these formulas. Algorithm 1 is invoked at most $D$ times. The total time taken to construct the list $L^{\prime}$ is at most $D\cdot T_{1}$, where $T_{1}$ is the time taken by Algorithm 1 on inputs with formula size and bit-complexity $\operatorname{poly}(s,D,t)$, and degree parameter $d$. $D\cdot T_{1}$ is also an upper bound on the size of the list of candidate factors $L^{\prime}$. Now, for each $g\in L^{\prime}$, from Theorem 2.28, we have that the irreducibility test in Algorithm 2 takes at most $(sDt)^{O(d^{2})}$ time. There are at most $D$ instances of divisibility test performed to determine the exact multiplicity in $f$ of each $g\in L^{\prime}$. This requires computing the corresponding derivatives, which as discussed in the previous paragraph, takes time $\operatorname{poly}(s,D,t)$ and outputs a formula of size and bit-complexity $\operatorname{poly}(s,D,t)$ for the derivatives, and then doing a divisibility test, the time complexity of which we denote by $T_{2}$. Therefore, the total time taken by the algorithm is at most $(n^{O(d)}\cdot\operatorname{poly}(s,D,t)\cdot D\cdot T_{1})+(D\cdot T_{1}\cdot(sDt)^{O(d^{2})}\cdot\operatorname{poly}(s,D,t)\cdot T_{2})$. Now, if $f$ is $s$ sparse, i.e. $k=1$, then from 2.11, we have that every vector in $H_{1}$ has at most $d$ non-zero coordinates. Thus, from 3.1, for every ${\bm{\alpha}}\in H_{1}$, $F(\mathbf{x},y)=f(\mathbf{x}+{\bm{\alpha}}\cdot y)$ has sparsity and bit- complexity at most $s^{\prime}\leq s\cdot D^{d}$. Note that the derivatives of arbitrary order of $F$ with respect to any variable also have the same bound on their sparsity and bit-complexity of coefficients. Thus, in this case, from 3.4, $T_{1}\leq(sDt)^{\operatorname{poly}(d)\log sDt}$. From Theorem 2.20, we have that $T_{2}\leq(snD)^{O(d\log^{2}snD)}$. Therefore, the overall running time of the algorithm is at most $(snDt)^{O(\operatorname{poly}(d)\cdot\log snDt)}$. On the other hand, if $k>1$, then from 3.4, $T_{1}\leq(sD)^{O_{\varepsilon}(kd(sD)^{\varepsilon d})}\cdot t^{O(d\log d)}$. To bound $T_{2}$ in this case, we note from 2.19, this divisibility testing instances reduce to PIT instances for $(\Sigma\Pi)^{(k+1)}$ formula of size and bit-complexity at most $\operatorname{poly}(s,D,t)$ and from Theorem 2.13, this can be done in at most $(sDt)^{O(k(sDt)^{\varepsilon})}$ time for the arbitrary constant $\varepsilon$ chosen in the beginning. Thus, the total time taken is at most $(sDt)^{O_{\varepsilon}(kd(sDt)^{\varepsilon d})}$. ∎ Lemma 4.2 and Lemma 4.3 together imply our main theorems Theorem 1.1 and Theorem 1.2. ## 5 Open problems We conclude with some open problems. * • Perhaps the most natural open problem here is to obtain efficient deterministic algorithms that completely factor sparse polynomials or more generally, polynomials with constant depth formulas (and not just obtain low degree factors). In the absence of better structural guarantees for the factors (for instance, if they are sparse or have small constant depth formulas), we can seek algorithms that output general algebraic circuits for these factors. * • Obtaining improved structural guarantees on the factors of polynomials that are sparse or have small constant depth formulas as mentioned in the first open problem is another very interesting open problem. * • A first step towards obtaining deterministic algorithms for general factorization of polynomials with small constant depth formulas could be to design deterministic algorithms for computing _simple_ factors of such polynomials. While the notion of simplicity discussed in this paper is that of low degree factors, there are other natural notions that seem very interesting. For instance, can we design an efficient deterministic algorithm that outputs all the sparse irreducible factors of a constant depth formula ? * • As alluded to in the introduction, polynomial factorization algorithms have found numerous applications in computer science. It would be interesting to understand if there are applications of deterministic factorization algorithms in general, and in particular the algorithms for computing low degree factors described in this paper. ## Acknowledgements A part of this work was done while the first two authors were at the Workshop on Algebraic Complexity organised at the University of Warwick in March 2023 by Christian Ikenmeyer. We thank Christian for the invitation and the delightful and stimulating atmosphere at the workshop. ††footnotetext: git info: , () ## References * [AS03] Sanjeev Arora and Madhu Sudan. Improved Low-Degree Testing and its Applications. Comb., 23(3):365–426, 2003. * [BHKS20] Siddharth Bhandari, Prahladh Harsha, Mrinal Kumar, and Madhu Sudan. Decoding Multivariate Multiplicity Codes on Product Sets. Electron. Colloquium Comput. Complex., TR20-179, 2020. Pre-print available at arXiv:TR20-179. * [Bog05] Andrej Bogdanov. Pseudorandom generators for low degree polynomials. In Proceedings of the 37th Annual ACM Symposium on Theory of Computing, Baltimore, MD, USA, May 22-24, 2005, pages 21–30. ACM, 2005. * [BSCI+20] Eli Ben-Sasson, Dan Carmon, Yuval Ishai, Swastik Kopparty, and Shubhangi Saraf. Proximity Gaps for Reed–Solomon Codes. In 2020 IEEE 61st Annual Symposium on Foundations of Computer Science (FOCS), pages 900–909, 2020. * [BSV18] Vishwas Bhargava, Shubhangi Saraf, and Ilya Volkovich. Deterministic Factorization of Sparse Polynomials with Bounded Individual Degree. In 59th IEEE Annual Symposium on Foundations of Computer Science, FOCS 2018, pages 485–496. IEEE Computer Society, 2018. * [CKS18] Chi-Ning Chou, Mrinal Kumar, and Noam Solomon. Hardness vs Randomness for Bounded Depth Arithmetic Circuits. In 33rd Computational Complexity Conference, CCC 2018, June 22-24, 2018, San Diego, CA, USA, volume 102 of LIPIcs, pages 13:1–13:17. Schloss Dagstuhl - Leibniz-Zentrum für Informatik, 2018. * [DL78] Richard A. DeMillo and Richard J. Lipton. A Probabilistic Remark on Algebraic Program Testing. Information Processing Letters, 7(4):193–195, 1978. * [DSY09] Zeev Dvir, Amir Shpilka, and Amir Yehudayoff. Hardness-Randomness Tradeoffs for Bounded Depth Arithmetic Circuits. SIAM J. Comput., 39(4):1279–1293, 2009. * [Ell69] W. J. Ellison. A ‘Waring’s Problem’ for homogeneous forms. Proceedings of the Cambridge Philosophical Society, 65:663–672, 1969. * [Fis94] Ismor Fischer. Sums of like powers of multivariate linear forms. Mathematics Magazine, 67(1):59–61, 1994. * [For14] Michael Forbes. Polynomial Identity Testing of Read-Once Oblivious Algebraic Branching Programs. PhD thesis, Massachusetts Institute of Technology, 2014. * [For15] Michael A. Forbes. Deterministic Divisibility Testing via Shifted Partial Derivatives. In Proceedings of the 2015 IEEE 56th Annual Symposium on Foundations of Computer Science (FOCS), FOCS ’15, page 451–465, USA, 2015. IEEE Computer Society. * [FS13] Michael A. Forbes and Amir Shpilka. Quasipolynomial-Time Identity Testing of Non-commutative and Read-Once Oblivious Algebraic Branching Programs. In Proceedings of the 54 Annual IEEE Symposium on Foundations of Computer Science (FOCS 2013), pages 243–252, 2013. Full version at arXiv:1209.2408. * [FS15] Michael A. Forbes and Amir Shpilka. Complexity Theory Column 88: Challenges in Polynomial Factorization1. SIGACT News, 46(4):32–49, dec 2015. * [FSS14] Michael A. Forbes, Ramprasad Saptharishi, and Amir Shpilka. Hitting sets for multilinear read-once algebraic branching programs, in any order. In Proceedings of the 46 Annual ACM Symposium on Theory of Computing (STOC 2014), pages 867–875, 2014. * [GKS16] Rohit Gurjar, Arpita Korwar, and Nitin Saxena. Identity Testing for Constant-Width, and Commutative, Read-Once Oblivious ABPs. In Proceedings of the 31 Annual Computational Complexity Conference (CCC 2016), pages 29:1–29:16, 2016. arXiv:1601.08031. * [GKSS19] Zeyu Guo, Mrinal Kumar, Ramprasad Saptharishi, and Noam Solomon. Derandomization from Algebraic Hardness: Treading the Borders. In 60th IEEE Annual Symposium on Foundations of Computer Science, FOCS 2019, Baltimore, Maryland, USA, November 9-12, 2019, pages 147–157. IEEE Computer Society, 2019. * [GKST15] Rohit Gurjar, Arpita Korwar, Nitin Saxena, and Thomas Thierauf. Deterministic Identity Testing for Sum of Read-once Oblivious Arithmetic Branching Programs. In Proceedings of the 30 Annual Computational Complexity Conference (CCC 2015), pages 323–346, 2015. arXiv:1411.7341. * [Kal85] Erich Kaltofen. Polynomial-Time Reductions from Multivariate to Bi- and Univariate Integral Polynomial Factorization. SIAM Journal of Computing, 14(2):469–489, 1985. * [Kal89] Erich Kaltofen. Factorization of Polynomials Given by Straight-Line Programs. In Randomness and Computation, pages 375–412. JAI Press, 1989. * [Kal92] Erich L. Kaltofen. Polynomial Factorization 1987-1991. In LATIN ’92, 1st Latin American Symposium on Theoretical Informatics, São Paulo, Brazil, April 6-10, 1992, Proceedings, volume 583 of Lecture Notes in Computer Science, pages 294–313. Springer, 1992\. * [Kal03] Erich L. Kaltofen. Polynomial factorization: a success story. In Symbolic and Algebraic Computation, International Symposium ISSAC 2003, Drexel University, Philadelphia, Pennsylvania, USA, August 3-6, 2003, Proceedings, pages 3–4. ACM, 2003. * [KS01] Adam Klivans and Daniel A. Spielman. Randomness efficient identity testing of multivariate polynomials. In Proceedings of the 33 Annual ACM Symposium on Theory of Computing (STOC 2001), pages 216–223, 2001. * [KS09] Neeraj Kayal and Shubhangi Saraf. Blackbox polynomial identity testing for depth-$3$ circuits. In Proceedings of the 50 Annual IEEE Symposium on Foundations of Computer Science (FOCS 2009), 2009. * [KSS15] Swastik Kopparty, Shubhangi Saraf, and Amir Shpilka. Equivalence of Polynomial Identity Testing and Polynomial Factorization. Computational Complexity, 24(2):295–331, 2015. Preliminary version in the _29 Annual IEEE Conference on Computational Complexity (CCC 2014)_. * [KT88] Erich L. Kaltofen and Barry M. Trager. Computing with Polynomials Given By Black Boxes for Their Evaluation: Greatest Common Divisors, Factorization, Separation of Numerators and Denominators. In 29th Annual Symposium on Foundations of Computer Science, White Plains, New York, USA, 24-26 October 1988, pages 296–305. IEEE Computer Society, 1988. * [KY08] Swastik Kopparty and Sergey Yekhanin. Detecting Rational Points on Hypersurfaces over Finite Fields. In Proceedings of the 23 Annual IEEE Conference on Computational Complexity (CCC 2008), pages 311–320, 2008. * [LLL82] Arjen K. Lenstra, Hendrik W. Lenstra Jr., and László Lovász. Factoring polynomials with rational coefficients. Mathematische Annalen, 261(4):515–534, 1982. * [LST21] Nutan Limaye, Srikanth Srinivasan, and Sébastien Tavenas. Superpolynomial Lower Bounds Against Low-Depth Algebraic Circuits. In Proceedings of the 62 Annual IEEE Symposium on Foundations of Computer Science (FOCS 2021), pages 804–814. IEEE, 2021. Preliminary version in the Electronic Colloquium on Computational Complexity (ECCC), Technical Report TR21-081. * [Ore22] Øystein Ore. Über höhere Kongruenzen. Norsk Mat. Forenings Skrifter, 1(7):15, 1922. * [PS94] Alexander Polishchuk and Daniel A. Spielman. Nearly-Linear Size Holographic Proofs. In Proceedings of the Twenty-Sixth Annual ACM Symposium on Theory of Computing, STOC ’94, page 194–203, New York, NY, USA, 1994. Association for Computing Machinery. * [Sap15] Ramprasad Saptharishi. A survey of lower bounds in arithmetic circuit complexity. Github survey, 2015. * [Sch80] Jacob T. Schwartz. Fast Probabilistic Algorithms for Verification of Polynomial Identities. Journal of the ACM, 27(4):701–717, 1980. * [Shp02] Amir Shpilka. Affine projections of symmetric polynomials. Journal of Computer and System Sciences, 65(4):639–659, 2002. Special Issue on Complexity 2001. * [SS09] Nitin Saxena and C. Seshadhri. An Almost Optimal Rank Bound for Depth-3 Identities. In Proceedings of the 24 Annual IEEE Conference on Computational Complexity (CCC 2009), pages 137–148, 2009. * [SS10] Nitin Saxena and C. Seshadhri. From Sylvester-Gallai Configurations to Rank Bounds: Improved Black-Box Identity Test for Depth-3 Circuits. In Proceedings of the 51 Annual IEEE Symposium on Foundations of Computer Science (FOCS 2010), pages 21–29, 2010. * [ST20] Amit Sinhababu and Thomas Thierauf. Factorization of Polynomials Given By Arithmetic Branching Programs. In 35th Computational Complexity Conference (CCC 2020), volume 169 of Leibniz International Proceedings in Informatics (LIPIcs), pages 33:1–33:19, Dagstuhl, Germany, 2020. Schloss Dagstuhl–Leibniz-Zentrum für Informatik. * [Sud98] Madhu Sudan. Lecture notes for the course ‘Algebra and Computation’, 1998. Available from http://people.csail.mit.edu/madhu/FT98/. * [SV10] Amir Shpilka and Ilya Volkovich. On the Relation between Polynomial Identity Testing and Finding Variable Disjoint Factors. In Automata, Languages and Programming, 37th International Colloquium, ICALP 2010, Bordeaux, France, July 6-10, 2010, Proceedings, Part I, volume 6198 of Lecture Notes in Computer Science, pages 408–419. Springer, 2010. * [Vol15] Ilya Volkovich. Deterministically Factoring Sparse Polynomials into Multilinear Factors and Sums of Univariate Polynomials. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, APPROX/RANDOM 2015, August 24-26, 2015, Princeton, NJ, USA, volume 40 of LIPIcs, pages 943–958. Schloss Dagstuhl - Leibniz-Zentrum für Informatik, 2015. * [Vol17] Ilya Volkovich. On Some Computations on Sparse Polynomials. volume 81 of LIPIcs, pages 48:1–48:21. Schloss Dagstuhl - Leibniz-Zentrum für Informatik, 2017. * [VSBR83] Leslie G. Valiant, Sven Skyum, S. Berkowitz, and Charles Rackoff. Fast Parallel Computation of Polynomials Using Few Processors. SIAM Journal of Computing, 12(4):641–644, 1983. Preliminary version in the _6 Internationl Symposium on the Mathematical Foundations of Computer Science (MFCS 1981)_. * [vzG83] Joachim von zur Gathen. Factoring Sparse Multivariate Polynomials. In Proceedings of the 24 Annual IEEE Symposium on Foundations of Computer Science (FOCS 1983), pages 172–179, 1983. * [vzGG13] Joachim von zur Gathen and Jürgen Gerhard. Modern Computer Algebra. Cambridge University Press, 3 edition, 2013. * [Zip79] Richard Zippel. Probabilistic algorithms for sparse polynomials. In Symbolic and Algebraic Computation, EUROSAM ’79, An International Symposiumon Symbolic and Algebraic Computation, volume 72 of Lecture Notes in Computer Science, pages 216–226. Springer, 1979. ## Appendix A Deferred proofs #### Circuit/formula bit-complexity ###### Proof of Lemma 2.2. We will prove an equivalent statement: the numerator and denominator of $f(\mathbf{a})$ have absolute value at most $2^{s\cdot b}$. We prove this by induction on the size of the formula. We will use $N(\cdot)$ and $D(\cdot)$ to denote the numerator and denominator of some rational number. Base case: when $\operatorname{size}(C)=1=s$, there is a single leaf node in the formula that reads and outputs a single rational number of bit-complexity $b$, thus $\operatorname{bit}(f(\mathbf{a}))=b\leq s\cdot b$. The induction hypothesis is that for all formulas $C$ with $\operatorname{size}(C)\leq S$ (for some $S\geq 1$), $\operatorname{bit}(f(\mathbf{a}))\leq\operatorname{bit}(C)\cdot b$. For the induction step, we look at formulas $C$ with $\operatorname{size}(C)=S+1$, and we consider two cases: 1. 1. When the top gate is a sum gate: $f=\sum_{i=1}^{k}{\alpha_{i}g_{i}(\mathbf{x})}$, with $C_{i}$ being the formula computing $g_{i}$ and $\alpha_{i}$s being scalars from $\mathbb{Q}$. $\displaystyle f(\mathbf{a})$ $\displaystyle=\sum_{i=1}^{k}{\alpha_{i}g_{i}(\mathbf{a})}$ $\displaystyle\left|D(f(\mathbf{a}))\right|$ $\displaystyle=\left|\prod_{i=1}^{k}{D(\alpha_{i})D(g_{i}(\mathbf{a}))}\right|$ $\displaystyle\leq\prod_{i=1}^{k}{2^{\operatorname{bit}(\alpha_{i})}2^{\operatorname{bit}(g_{i}(\mathbf{a}))}}$ $\displaystyle\leq\prod_{i=1}^{k}{2^{\operatorname{bit}(\alpha_{i})}2^{\operatorname{bit}(C_{i})\cdot b}}$ (induction hypothesis) $\displaystyle=2^{\sum_{i=1}^{k}{\left(\operatorname{bit}(\alpha_{i})+\operatorname{bit}(C_{i})\cdot b\right)}}\leq 2^{\operatorname{bit}(C)\cdot b}$ $\displaystyle\left|N(f(\mathbf{a}))\right|$ $\displaystyle\leq\sum_{i=1}^{k}{\left|N(\alpha_{i})\right|\left|N(g_{i}(\mathbf{a}))\right|\prod_{j\neq i}{\left|D(\alpha_{j})\right|\left|D(g_{j}(\mathbf{a}))\right|}}$ $\displaystyle\leq\sum_{i=1}^{k}{2^{\operatorname{bit}(\alpha_{i})+\operatorname{bit}(g_{i}(\mathbf{a}))}2^{\sum_{j\neq i}{\operatorname{bit}(\alpha_{j})+\operatorname{bit}(g_{j}(\mathbf{a}))}}}$ $\displaystyle\leq\sum_{i=1}^{k}{2^{\operatorname{bit}(\alpha_{i})+\operatorname{bit}(C_{i})\cdot b}2^{\sum_{j\neq i}{\operatorname{bit}(\alpha_{j})+\operatorname{bit}(C_{j})\cdot b}}}$ (induction hypothesis) $\displaystyle\leq\sum_{i=1}^{k}{2^{\sum_{j=1}^{k}{\operatorname{bit}(\alpha_{j})+\operatorname{bit}(C_{j})\cdot b}}}\leq 2^{k+\sum_{j=1}^{k}{\operatorname{bit}(\alpha_{j})+\operatorname{bit}(C_{j})\cdot b}}\leq 2^{\operatorname{bit}(C)\cdot b}$ Thus, $\operatorname{bit}(f(\mathbf{a}))=\max\\{\operatorname{bit}(N(f(\mathbf{a})),D(f(\mathbf{a})))\\}\leq\operatorname{bit}(C)\cdot b$. 2. 2. When the top gate is a product gate: $f=\prod_{i=1}^{k}{\alpha_{i}g_{i}(\mathbf{x})}$. The proof for the denominator in the case of sum gate will work here for both the numerator and the denominator. The required bound follows. ∎ #### Relevant subclasses of algebraic circuits ###### Proof of 2.5. We prove the size upper bounds here; the bit-complexity upper bounds proceed along exactly the same lines. The size upper bound for the sum is immediate and hence we only need to focus on the product. Let the expression for each $P_{r}$ be $\displaystyle P_{r}$ $\displaystyle=\sum_{i}P_{r,i}^{(r)}\cdot g_{r,i}^{a_{r,i}}$ $\displaystyle\implies\prod P_{r}$ $\displaystyle=\sum_{r_{1},\ldots,r_{t}}\left(P_{1,r_{1}}\cdots P_{t,r_{t}}\right)\cdot\left(g_{1,r_{1}}^{a_{1,r_{1}}}\cdots g_{t,r_{t}}^{a_{t,r_{t}}}\right)$ where each $P_{i,j}$ is computed by $(\Sigma\Pi)^{(k)}$ formulas of size at most $s$, and each $g_{i,j}$ is a polynomial of degree at most $d$. Each $\left(P_{1,r_{1}}\cdots P_{t,r_{t}}\right)$ is computed by a $(\Sigma\Pi)^{(k)}$ formula of size at most $s^{t}$. By Lemma 2.9, $g_{1,r_{1}}^{a_{1,r_{1}}}\cdots g_{t,r_{t}}^{a_{t,r_{t}}}$ can be expressed as a sum $\sum_{\ell=1}^{s^{t}}f_{\ell}^{D}$ where $D=\sum_{j}a_{j,r_{j}}$ and each $f_{\ell}$ is a degree polynomial of degree at most $d$. Thus, $\prod_{r}P_{r}$ is computable by a $\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formula of size at most $s^{O(t)}$. ∎ #### Polynomial identity testing ###### Proof of Lemma 2.12. Since $f$ is a non-zero polynomial of degree at most $d$, there is a monomial $\mathbf{x}^{\mathbf{e}}$ of degree at most $d$ with a non-zero coefficient in $f$. Let $S$ be the support of the monomial $\mathbf{x}^{\mathbf{e}}$, i.e., $S=\\{x_{i}:e_{i}\neq 0\\}$. Clearly, $|S|\leq d$. We now consider the polynomial $\tilde{f}$ obtained from $f$ by setting all the variables $x_{j}$ not in the set $S$ to zero. Since $f$ has a non-zero monomial with support contained in the set $S$, $\tilde{f}$ continues to be a non-zero polynomial of degree at most $d$. Moreover, it is a $d$ variate polynomial since it only depends on the variables in $S$. From Lemma 2.10, we get that for any subset $T_{d}$ of $\mathbb{Q}$ of cardinality at least $d+1$, there exists a vector $\mathbf{b}\in{T_{d}}^{d}$ such that $\tilde{f}(\mathbf{b})\neq 0$. Let $\mathbf{a}\in\mathbb{Q}^{n}$ to be such that for every $i\in S$, $a_{i}=b_{i}$ and for every $i\notin S$, $a_{i}=0$. Then, $f(\mathbf{a})=\tilde{f}(\mathbf{b})\neq 0$. Moreover, $\mathbf{a}$ is in $\mathcal{H}(d,n)$. ∎ ###### Proof of Lemma 2.15. Let $T_{d}$ be the set $\\{0,1,2,3,\ldots,d\\}$ and let $\mathcal{H}(d,n)$ be the set of points defined in Definition 2.11, i.e., $\mathcal{H}(d,n)=\left\\{(a_{1},\ldots,a_{n})\ :\ S\in\binom{[n]}{\leq d}\;,\;a_{i}\in T_{d}\text{ for all }i\in S\text{ and }a_{j}=0\text{ for all }j\notin S\right\\}.$ From Lemma 2.12, we know that every non-zero polynomial $f$ of degree at most $d$ must evaluate to zero on some point of $\mathcal{H}(d,n)$. In other words, two distinct degree $d$ polynomials $f$ and $g$ cannot agree on every point of $\mathcal{H}(d,n)$. An immediate consequence of this is that if we are given the evaluations of an unknown polynomial $f$ on all points of $\mathcal{H}(d,n)$, and we view each of these evaluations as a linear constraint on the unknown coefficients of $f$, then this linear system has a unique solution. Based on this observation, a natural algorithm for computing the coefficient vector of $C$ is the following, we evaluate the given formula on every input in $\mathcal{H}(d,n)$, set up the linear system on the coefficients of $C$ obtained from these evaluations, and use any standard linear system solver over $\mathbb{Q}$ to solve this system. Note that the size of this linear system is at most $n^{O(d)}$, and from 2.2, of the constants in this linear system is at most $\operatorname{poly}(s,b,d)$. Thus, this linear system can be solved in time $\operatorname{poly}(s,b,d,n^{d})\leq\operatorname{poly}(s,b,n^{d})$ time as claimed. ∎ #### Deterministic divisibility testing and PIT ###### Proof of Corollary 2.19. The proof essentially follows immediately from Theorem 2.18. From Theorem 2.18, we have that $g$ divides $f$ if and only if $R(\mathbf{x}):=f(\mathbf{x})-g(\mathbf{x})Q(\mathbf{x})\equiv 0$, where $Q$ is the pseudo-quotient of $f$ and $g$. It suffices to show that $R(\mathbf{x})$ has $\mathcal{C}=\Sigma\left((\Sigma\Pi)^{(k)}\cdot(\operatorname{Deg}_{d})^{\ast}\right)$ formulas of size $\operatorname{poly}(s,d)$, and since $f(\mathbf{x})\in(\Sigma\Pi)^{(k)}$, it suffices to bound the size of $\mathcal{C}$-formulas computing $g(\mathbf{x})\cdot Q(\mathbf{x})$. By Lemma 2.17, the pseudo-quotient $Q(\mathbf{x})$ is computable by $\mathcal{C}$-formulas of size $\operatorname{poly}(s,d)$. Let one such computation be of the form $\displaystyle Q(\mathbf{x})$ $\displaystyle=\sum_{i}f_{i}\cdot g_{i}^{e_{i}}\quad\text{where each $f_{i}\in(\Sigma\Pi)^{(k)}$ and $\deg(g_{i})\leq d$ and each $e_{i}\leq s$}$ $\displaystyle\implies g(\mathbf{x})Q(\mathbf{x})$ $\displaystyle=\sum_{i}f_{i}\cdot(g\cdot g_{i}^{e_{i}})$ From Lemma 2.9, note that any term of the form $(g\cdot h^{e})$ can be expressed as $g\cdot h^{e}=\sum_{i=1}^{\operatorname{poly}(e)}\beta_{i}\cdot(g+\alpha_{i}h)^{e+1}$ for field constants $\alpha_{i}$’s and $\beta_{i}$’s. Thus, feeding this in the above expression for $g\cdot Q$, we have $g(\mathbf{x})\cdot Q(\mathbf{x})=\sum_{i}\sum_{j}f_{i}\cdot\tilde{g}_{ij}^{e_{ij}}$ for polynomial $\tilde{g}_{ij}$ of degree at most $d$, and thus is also a $\mathcal{C}$-formula of size at most $\operatorname{poly}(s,d)$. Therefore, $R(\mathbf{x})=f(\mathbf{x})-g(\mathbf{x})Q(\mathbf{x})$ is also computable by $\mathcal{C}$-formulas of size $s^{\prime}=\operatorname{poly}(s,d)$. Thus, we can check if $g$ divides $f$ by checking if $R(\mathbf{x})\equiv 0$ (by Theorem 2.18) which can be done in $T(k,d,s^{\prime})$ time as claimed. ∎ #### Hensel lifting ###### Proof sketch of Lemma 2.26. As indicated earlier, the lemma is almost an immediate consequence of Lemma 3.6 in [KSS15]. The precise statement there gives a circuit $\tilde{C_{k}}$ of size and bit-complexity $\operatorname{poly}(s,D,2^{k})$ for $g_{k},h_{k}$. We notice that without loss of generality, the degree of $g_{k},h_{k}$ and hence of $\tilde{C_{k}}$ can be assumed to be at most $(D+2^{k})$ since the $y$ degree is at most $D$ and the $x$ degree is at most $2^{k}$. This incurs at most a polynomial blow up in the circuit size. Now, to go from circuits for $g_{k},h_{k}$ to formulas computing these polynomials, we just invoke the classic depth reduction result of Valiant, Skyum, Berkowitz and Rackoff [VSBR83], which states that given an $n$-variate degree-$\Delta$ polynomial $f$ with an arithmetic circuit $\Phi$ of size $s$, there is an arithmetic circuit $\Phi^{\prime}$ that computes $f$, has size $\operatorname{poly}(s,n,\Delta)$ and depth $O(\log\Delta)$. Thus we have a formula of size (and bit-complexity) at most $\operatorname{poly}(s,D,2^{k})^{\log(D+2^{k})}\leq(sDk)^{k\log D}$. Note that a better bound of $d$ on the total degree of $g_{k}$ implies that the size and bit-complexity of the formula for $g_{k}$ is at most $(sDk)^{O(\log d)}$. ∎
# UFed-GAN: A Secure Federated Learning Framework with Constrained Computation and Unlabeled Data Achintha Wijesinghe, Songyang Zhang, , Siyu Qi and Zhi Ding A. Wijesinghe, S. Qi, and Z. Ding are with the Department of Electrical and Computer Engineering, University of California, Davis, CA, 95616. (E-mail: <EMAIL_ADDRESS><EMAIL_ADDRESS>and [email protected]). S. Zhang was with the University of California at Davis, Davis, CA, 95616 and now is with the Department of Electrical and Computer Engineering, University of Louisiana at Lafayette, Lafayette, LA 70504 (E-mail: [email protected]). ###### Abstract To satisfy the broad applications and insatiable hunger for deploying low latency multimedia data classification and data privacy in a cloud-based setting, federated learning (FL) has emerged as an important learning paradigm. For the practical cases involving limited computational power and only unlabeled data in many wireless communications applications, this work investigates FL paradigm in a resource-constrained and label-missing environment. Specifically, we propose a novel framework of UFed-GAN: Unsupervised Federated Generative Adversarial Network, which can capture user- side data distribution without local classification training. We also analyze the convergence and privacy of the proposed UFed-GAN. Our experimental results demonstrate the strong potential of UFed-GAN in addressing limited computational resources and unlabeled data while preserving privacy. ###### Index Terms: Federated learning, unlabeled data, data privacy, generative adversarial networks. ## I Introduction The burgeoning rise of deep learning has shown remarkable achievements in learning, based on the often voluminous amounts of data for centralized training. However, in many cases of learning-based wireless connections for collaboration, decentralized learning is vital to handle the heterogeneous data distribution among nodes (users). Importantly, privacy concerns and resource limitations also prevent direct data sharing. To ensure data privacy and communication efficiency, federated learning (FL) [1] has emerged as an important framework to disengage data collection and model training via local computation and global model aggregation. Despite reported successes, existing FL frameworks have certain limitations. One major obstacle of FL in practice is the heterogeneity of data distribution among participating FL users. It is known [2] that the accuracy of classic FL frameworks such as FedAvg [1] could drop by 55% for some datasets showing non- IID, i.e., not identically and independently distributed, data distributions. To combat performance loss against non-IID datasets, the more general approach of FedProx [3] may depend on certain unrealistic dissimilarity assumptions of local functions [4]. Alternatively, generative adversarial network (GAN) [5] provides another approach to address data heterogeneity. In GAN-based FL, GAN models are used as a proxy to share user updates without training the global model on the user side. For example, in [6], a global classifier is trained using a user-shared generator of the user-end-trained conditional GANs (cGANs). Another example is [7], the authors proposed to share the full user- end GAN for generating a synthetic dataset. However, the high communication cost and the potential privacy leakage hinder the performance gain of these known GAN-based FL frameworks as discussed in [8]. Moreover, the training of the entire GAN and other learning models sometimes can be impractical at the user end, especially for nodes with limited computation resources, such as computation-constrained sensors and devices [9]. How to develop a more efficient GAN-sharing strategy to preserve privacy and handle limited computation remains an open question in generative FL. In addition to data heterogeneity and limited computation resources, most of the existing works focus on supervised FL, where the performance depends heavily on the availability of labeled training data. In practical applications such as user clustering and video segmentation [10], the computational and privacy limitations may prevent user-side data labeling. Therefore, learning from unlabeled data is equally important for FL to reach its full potential. Presently, only limited research works have specifically addressed FL with unlabeled data, primarily due to inherent challenges. A typical category of FL dealing with unlabeled data [11, 12] focuses on clustering tasks, which may limit its generalization to other deep learning tasks. Another line of FL focuses on unsupervised representation learning [13], where knowledge distillation and contrast learning are applied to address heterogeneous user data distributions. Other FL works on unlabeled data [14, 15, 16] leverage the efficiency of latent space and may lack a general description of the original data. As aforementioned, GAN-based approaches can be intuitive solutions to capture the data distributions and assist further applications, even without annotated labels. In this work, we develop a novel FL framework, the Unsupervised Federated Generative Adversarial Network (UFed-GAN), for resource-limited distributed users without labeled data. The novelty is an innovative GAN-based FL and data-sharing strategy to significantly reduce the computational cost at the user end and to preserve privacy. Note that, instead of focusing on one specific unsupervised learning task, we provide a general FL scheme to learn the data distributions without labels. Our framework can be easily adapted to handle specific learning tasks, including unsupervised representation learning, user clustering, and semi-supervised classification. Our contributions can be summarized as follows: * • We propose a novel UFed-GAN as an FL framework to learn and characterize the non-IID user data distributions from unlabeled user data. Our UFed-GAN captures the underlying user data distributions without explicitly training a local GAN model for each user, thereby significantly lowering the computational cost on the user side. To our best knowledge, this is the first work to address such constrained computation in GAN-based FL. * • We analyze the convergence of UFed-GAN and prove that privacy leakage can be prevented by our UFed-GAN, in comparison to traditional GAN-based FL. * • Our experimental results in several benchmark datasets demonstrate the performance of UFed-GAN in a semi-supervised classification setup. In terms of organization, we first introduce the architecture of UFed-GAN and a training strategy in Section II. Following the study on model convergence and the privacy analysis in Section III. We present the experimental results of UFed-GAN on several well-known datasets in Section IV. We provide concluding remarks in Section V. ## II Method and Architecture ### II-A Problem Setup Figure 1: UFed-GAN in a distributed learning setup in an untrustworthy communication scenario, where users are assumed with less computational power. An attacker may eavesdrop to understand the user data. As a typical example of a resource-limited FL setup in Fig. 1, a server aims to learn from user nodes, each with limited computational resources whereas attackers may attempt to eavesdrop on the network links. Users may not be able to annotate their raw data. Moreover, since the target tasks may be different among users and the data may also be skewed, a non-IID data distribution shall be considered in this scenario. Different from local users, the server has sufficient computational resources to obtain a model that learns a global data distribution from all the local data, without initial training data. Such a setup is applicable in many distributed learning scenarios. For example, a distributed camera/sensor system placed for object detection can benefit from the collaboration of different cameras for better feature extraction, where each digital camera or sensor may have limited computation power. To demonstrate the privacy protection offered by UFed-GAN, we consider an attacker that has access to the vulnerable communication links between distributed users and the central server. Such attacks try to gain users’ data features based on the information shared through the channels. Note that, we aim to develop a novel data/model-sharing strategy for FL. The strategy could handle unlabeled data and capture user data distributions in the scenario with limited local computation resources. The proposed framework should offer flexible integration with various unsupervised and semi- supervised learning tasks, such as latent representation learning, user clustering, and semi-supervised classification [13]. ### II-B UFed-GAN Figure 2: Communication rounds till the convergence of UFed-GAN. We use the inception score (IS) as the measure of convergence. We now explain the framework and training process of UFed-GAN, which allows private and secure learning from unlabeled data in a distributed learning setup. Our proposed UFed-GAN aims to train a GAN model on the server to capture the underlying user data distributions without implementing complex GAN training on the user side. First, for each user $u$, we initiate a GAN model on the server, including a generator ($G_{u}$) and a discriminator ($D_{u}$). Due to label inaccessibility, we select deep convolutional GANs (DCGANs) [17] as the backbone. The details of choices for GAN architecture will be elaborated in Section II-D. The GAN training comes in three steps, which contains two steps of $D$ training and one step of $G$ training. We split the dual-step $D$ training into the server side and the user side. On one side, the server initiates a discriminator $D_{u}$ and then shares the initiated model with the corresponding user $u$. Subsequently, on the other end, the user performs a forward pass (FP) on $D_{u}$ using a single batch of real local data $I_{u}$. The gradients and loss are calculated and then shared with the server. Compared with training a full GAN on the user side, this step can significantly reduce the computational cost and be easily deployed in computation-constrained devices. Upon receiving the updated information of $D_{u}$ from the user $u$, the server completes the training of $D_{u}$. It generates a noise vector z to pass through $G_{u}$. Finally, synthetic data $G_{u}(\textbf{z})$ is generated from the generator. $G_{u}(\textbf{z})$ is then sent to $D_{u}$ to calculate the corresponding gradients and loss. These received gradient updates are combined with the gradient updates of the previous step and then back-propagated through $D_{u}$ to update the parameters. Similar to conventional GAN training, starting with a noise vector z’, we train $G_{u}$ with forward- and backward-propagation to update its parameters. In each communication round, the above process repeats until GAN convergence to a favorable point. We illustrate this training process further in Fig. 2. For model convergence monitoring and the stopping criterion, we use inception score (IS) [18], which measures the characteristics of the generated images. After convergence, we create a synthetic dataset using the trained $D_{u}$. With the generated synthetic dataset, we can design corresponding unsupervised or semi-supervised algorithms to implement specific learning tasks. For example, in a semi-supervised classification task, we could apply the MoCo [19], which uses dictionary lookups in contrastive learning to obtain the global classifier. The pseudocode of the proposed training strategy is presented in Algorithm 1. Algorithm 1 UFed-GAN: Training Algorithm for each user $u$ do Server initialization of $G_{u}$ and $D_{u}$ end for for each communication round $r$, until the GAN convergence do for each step $t$ do Share $D_{u}$ with user $u$. Perform FP in $D_{u}$ with user data $\mathcal{I}_{u}$ and share the gradient updates $\nabla W_{u}$. Perform FP in $D_{u}$ with fake data $G(z_{t})$ where $z_{t}$ is a random noise vector and get the gradient updates $\nabla W_{z}$. Update $D_{u}$ with ($\nabla W_{u}+\nabla W_{z}$). Train $G_{u}$ with trained $D_{u}$ and random noise vectors. end for Generate an unlabeled dataset using each $G_{u}$. end for ### II-C Attacker Model We now introduce the attacker model to quantify the privacy leakage of UFed- GAN. It is highly challenging, if not impossible, to obtain access to the global generator $G$ given a secure server or to guess the exact architecture of $G$, regardless of the computation prowess of the attacker. Therefore, as suggested in [8], we focus on a reconstruction attack [20] in this work, where an attacker attempts to reconstruct the training data. Let $\theta$ be the parameters released to the communication channel by the user. We denote the releasing mechanism by $\mathcal{M}$ and the information that an attacker $\mathcal{A}$ could obtain by $\mathcal{M}(\theta)$. According to the UFed-GAN framework, $\mathcal{M}(\theta)$ is the discriminator gradients and the loss values. Let $\mathcal{I}$ represent the reconstructed data, we have $\mathcal{A}:\mathcal{M}(\theta)\mapsto\mathcal{I}.$ (1) Suppose that the generator $G_{\mathcal{A}}$ of attacker $\mathcal{A}$ has the same architecture as $G$ but with initial weights $W_{\mathcal{A}}$ different from those of $G$, represented by $W_{G}$. In parallel to the server training, we train $G_{\mathcal{A}}$ at the attacker’s end. ### II-D GAN Architecture Due to label inaccessibility and resource constraints, we adopt the unsupervised DCGANs [17] over cGANs [8] which saves extra computation needs for any pseudo-labeling. The generator architecture follows five transposed convolutional layers. The first layer takes an input with 100 channels and maps it to 1024 channels. Every subsequent layer reduces the number of channels by half. Every layer uses ReLU activation except for the Tanh activation for the final layer. All the layers in both the generator and the discriminator use $4\times 4$ kernels and batch normalization for each layer before the final layer. For the discriminator model, we use four convolutional layers. The first layer accepts similar channel sizes of the data samples and maps to 256 channels. Every subsequent layer doubles the number of channels except the final layer, which outputs a single channel. The final layer uses a Sigmoid activation whereas all other layers use LeakyReLU activation. ## III Convergece and Privacy Analysis of UFed-GAN In this section, we present the convergence and privacy analysis of the proposed UFed-GAN. We introduce the major proof steps and refer those interested to our corresponding references for more details. ### III-A Convergence of the discriminator Let $\mathcal{G}$ and $\mathcal{D}$ be the generator and the discriminator of a GAN, respectively. Assume $\mathcal{G}$ is capable of capturing a distribution $p_{GS}$ on the server and we are interested in learning a user distribution $p_{data}(x)$. Proposition 1. Any $\mathcal{D}$ initiated on a server and trained in accordance to Algorithm 1 with $\mathcal{G}$ and $p_{data}(x)$ converges to a unique $\mathcal{D}^{*}$ for the given $\mathcal{G}$ as presented in [5], i.e., $\mathcal{D}^{*}=\frac{p_{data(x)}}{p_{data(x)}+p_{GS}}$ (2) Since UFed-GAN merely splits the GAN training, the proof of Proposition 1 shall directly follow that in [5]. This proposition serves as a guarantee of the convergence of the server-side discriminator. It shows that the discriminator is still capable of capturing the user side’s data distribution. This helps the generator on the server-side to generate user-like data as illustrated in the following proposition. ### III-B Convergence of the generator Proposition 2. Any $\mathcal{G}$ initiated on a server and trained in accordance to Algorithm 1 with $\mathcal{D}$ and $p_{data}(x)$ converges to a unique $\mathcal{G}^{*}$ which captures $p_{data}(x)$. i.e. $p_{GS}=p_{data}(x)$. Since UFed-GAN does not alter the training of $\mathcal{G}$, we can imitate and adopt the proof steps in [5]. Proposition 2 suggests that the server-side generator converges to the same generator that could have been trained locally. Therefore, on the server, we are able to regenerate synthetic samples which resemble the user data in terms of data distribution. ### III-C Divergence of any discriminator other than $\mathcal{G}$ Proposition 3. Any generator $G$, other than $\mathcal{G}$ trained in accordance to the Algorithm 1 with $\mathcal{D}$ diverges from unique $\mathcal{G}^{*}$. i.e. $p_{GS}\neq p_{data}(x)$. To prove Proposition 3, we adopt a similar proof process as provided in [8]. Following Proposition 1 and Proposition 2, the pair of $\mathcal{G}$ and $\mathcal{D}$ is unique. Therefore, at each training step $\mathcal{G}^{{}^{\prime}}=G^{{}^{\prime}}$ must be asserted. Hence, any $G$ difference from $\mathcal{G}$ at any step fails to capture $p_{data}(x)$. ## IV Results and Discussion In this section, we present the experimental results on both utility and privacy. ### IV-A Evaluation of the Utility In a common semi-supervised setting as [13], let $N$ be the number of users, $\beta$ be the concentration parameter of Dirichlet distribution ($Dir_{N}(\beta)$), and $s_{ij}$ be a sample taken from $Dir_{N}(\beta)$. We assign $s_{ij}$ in proportion to the $i$-th class size of the user $j$. We pick $N=10$ and $\beta=0.5$ in accordance with [13]. All model comparisons are based on the linear evaluation protocol, which trains a linear classifier on top of representations or regenerated fake data [21]. TABLE I: Classification accuracy comparison of different FL approaches over three datasets. Part of the results in this table are reported from [13]. Method | CIFAR 10 | SVHN | FashionMNIST ---|---|---|--- FedSimCLR | 52.88 | 76.50 | 79.44 \+ FedX | 57.95 | 77.70 | 82.47 FedMoCo | 57.82 | 70.99 | 83.58 \+ FedX | 59.43 | 73.92 | 84.65 FedBYOL | 53.14 | 67.32 | 82.37 \+ FedX | 57.79 | 69.05 | 84.30 FedProtoCL | 52.12 | 50.19 | 83.57 \+ FedX | 56.76 | 69.75 | 83.34 FedU | 50.79 | 66.22 | 82.03 \+ FedX | 57.26 | 68.39 | 84.12 Full GAN | 68.77 | 80.17 | 86.25 UFed-GAN | 67.0 | 80.109 | 86.33 We consider three well-known datasets: CIFAR10 [22], SVHN [23] and FashionMNIST [24]. Comparative results against five other FL algorithms are presented in Table I. These algorithms are: FedSimCLR [25], FedMoco [19], FedBYOL [26], FedProtoCL [27] and FedU [28]. For each method, we present their accuracy on the respective dataset, together with their accuracy when further applying FedX [13]. We also compare the results with “full GAN” sharing. From the results, UFed-GAN outperforms all other FL methods in the benchmark group, with an improvement of around 8% in CIFAR10, 3% in SVHN, and 2% in FashionMNIST. The accuracy gain arises from the power of GANs to understand the underlying data distribution of users and to generate synthetic data by preserving essential features. Another observation is that UFed-GAN is dataset-agnostic in terms of performance, delivering the best outputs in all tested datasets. This observation promotes the generalizability of the proposed method. In fact, UFed-GAN achieves similar performance as with full GAN sharing. However, full GAN sharing is prone to severe privacy leakage as shown in [8] and additionally requires heavy computation at each user node, which is in conflict with the FL objective of privacy preservation and conserving computation resources for users. ### IV-B Evaluation of Privacy We now evaluate the privacy leakage of the proposed UFed-GAN with respect to the attacker $\mathcal{A}$ as described in Section II-C. Suppose that $\mathcal{A}$ has access to each communication round. We initialize the generator of $\mathcal{A}$ as $G_{\mathcal{A}}$, with some random weights and eavesdrop on user uplink and access $\mathcal{M}(\theta)$. $\mathcal{A}$ trains $G_{\mathcal{A}}$ similarly as with the server-side training. The design of $G_{\mathcal{A}}$ is constrained by two parameters, the exact generator architecture and the initial weights $W_{G}$ of the generator at the server. Therefore, any attacker selecting the accurate generator architecture and initial weights is practically unachievable. However, in our experiments, we assume $\mathcal{A}$ knows the exact architecture of $G$, but with different random initial weights $W_{\mathcal{A}}\neq W_{G}$. We compare the generated images of the $G$ and $G_{\mathcal{A}}$ trained on the FashionMNIST dataset in Fig. 3. It can be clearly seen that the generator $G$ on the cloud server captures the user’s underlying data distribution, whereas $G_{\mathcal{A}}$ converges to a trivial point. Moreover, almost no useful visualization information is gained by the attacker as shown in Fig. 3. The main reason is the uniqueness of the generator and the discriminator pair as presented in Proposition 1. (a) (b) Figure 3: Generated images from (a) cloud-server’s model. (b) attacker’s model. TABLE II: FID score, IS score, and SSIM of the $\mathcal{A}$ and the cloud server after 100 communication rounds on the FashionMNIST dataset. Metric | Attaker | Cloud Server ---|---|--- FID | 566.83 | 172.06 IS | 1.01 | 3.15 SSIM | 0.0067 | 0.8191 To examine privacy leakage quantitatively, we use the Frechet Inception Distance (FID) score [29], IS, and structural similarity index measure (SSIM) [30]. As discussed in the paper [31], the similarity between the generated images and real data quantifies the privacy leakage. In table IV-B, we record average FID, IS, and SSIM scores for the data generated by $\mathcal{A}$ and the cloud server. Lower FID values, higher SSIM, and larger IS values represent better reconstruction quality. The FID score for $\mathcal{A}$ is higher than the cloud server by a great margin. This shows that, compared to the cloud server, $\mathcal{A}$ generated data carries less information about the training data. We further corroborate this observation by comparing the SSIM and IS values as well. The experimental results demonstrate the privacy preservation of UFed-GAN against full-GAN sharing. ## V Conclusion In this work, we develop a novel framework of UFed-GAN to address the challenges imposed by the lack of labeled data and by limited local computation resources in federated learning. Moreover, we propose a separate training strategy and a sharing scheme based on DCGANs. We provide an analysis of the convergence and privacy leakage of the proposed framework. Our empirical results demonstrate the superior performance of UFed-GAN in comparison against benchmark FL methods. We plan to investigate the communication overhead reduction for GAN-based FL and the application of semantic learning in distributed learning in future works. ## References * [1] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Arcas, “Communication-efficient learning of deep networks from decentralized data,” in _Proceedings of the 20th International Conference on Artificial Intelligence and Statistics_ , Lauderdale, FL, USA, Apr. 2017, pp. 1273–1282. * [2] Y. Zhao, M. Li, L. Lai, N. Suda, D. Civin, and V. Chandra, “Federated learning with non-iid data,” _arXiv preprint arXiv:1806.00582_ , 2018. * [3] T. Li, A. K. Sahu, M. Sanjabi, M. Zaheer, A. Talwalkar, and V. Smith, “On the convergence of federated optimization in heterogeneous networks,” _arXiv preprint arXiv:1812.06127_ , 2018. * [4] X. Yuan and P. Li, “On convergence of fedprox: Local dissimilarity invariant bounds, non-smoothness and beyond,” in _Advances in Neural Information Processing Systems_ , New Orleans, LA, USA, Jul. 2022, pp. 10 752–10 765. * [5] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” in _Advances in Neural Information Processing Systems_ , vol. 27, Montreal, Canada, Dec. 2014. * [6] H. Zhang, Z. Zhang, A. Odena, and H. Lee, “Consistency regularization for generative adversarial networks,” in _International Conference on Learning Representations_ , Addis Ababa, Ethiopia, Apr. 2020. * [7] X. Cao, G. Sun, H. Yu, and M. Guizani, “Perfed-gan: personalized federated learning via generative adversarial networks,” _IEEE Internet of Things Journal_ , vol. 10, no. 5, pp. 3749–3762, Mar. 2023. * [8] A. Wijesinghe, S. Zhang, and Z. Ding, “Ps-fedgan: an efficient federated learning framework based on partially shared generative adversarial networks for data privacy,” _arXiv preprint arXiv:2305.11437_ , 2023. * [9] J. Mu, D. Xie, H. Huang, and X. Jing, “Computation-constrained spectrum sensing in iot-based scenarios,” _IET Communications_ , vol. 14, no. 20, pp. 3631–3638, Nov. 2020. * [10] J. Bekker and J. Davis, “Learning from positive and unlabeled data: a survey,” _Machine Learning_ , vol. 109, pp. 719–760, Apr. 2020. * [11] D. K. Dennis, T. Li, and V. Smith, “Heterogeneity for the win: One-shot federated clustering,” in _Proceedings of the 38th International Conference on Machine Learning_ , Jul. 2021, pp. 2611–2620. * [12] A. Ghosh, J. Chung, D. Yin, and K. Ramchandran, “An efficient framework for clustered federated learning,” in _Advances in Neural Information Processing Systems_ , Dec. 2020, pp. 19 586–19 597. * [13] S. Han, S. Park, F. Wu, S. Kim, C. Wu, X. Xie, and M. Cha, “Fedx: Unsupervised federated learning with cross knowledge distillation,” in _European Conference on Computer Vision_ , Tel Aviv, Israel, Nov. 2022, pp. 691–707. * [14] M. Servetnyk, C. C. Fung, and Z. Han, “Unsupervised federated learning for unbalanced data,” in _GLOBECOM 2020 - 2020 IEEE Global Communications Conference_ , Taipei, Taiwan, Dec. 2020, pp. 1–6. * [15] N. Lu, Z. Wang, X. Li, G. Niu, Q. Dou, and M. Sugiyama, “Federated learning from only unlabeled data with class-conditional-sharing clients,” in _International Conference on Learning Representations_ , Apr. 2022. * [16] E. S. Lubana, C. I. Tang, F. Kawsar, R. P. Dick, and A. Mathur, “Orchestra: unsupervised federated learning via globally consistent clustering,” _arXiv preprint arXiv:2205.11506_ , 2022. * [17] A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,” _arXiv preprint arXiv:1511.06434_ , 2015. * [18] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,” in _Advances in neural information processing systems_ , Barcelona, Spain, Dec. 2016. * [19] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , Jun. 2020, pp. 9729–9738. * [20] J. Hayes, L. Melis, G. Danezis, and E. De Cristofaro, “Logan: evaluating privacy leakage of generative models using generative adversarial networks,” _arXiv preprint arXiv:1705.07663_ , pp. 506–519, 2017. * [21] F. Zhang, K. Kuang, Z. You, T. Shen, J. Xiao, Y. Zhang, C. Wu, Y. Zhuang, and X. Li, “Federated unsupervised representation learning,” _arXiv preprint arXiv:2010.08982_ , 2020. * [22] A. Krizhevsky, “Learning multiple layers of features from tiny images,” pp. 32–33, 2009. [Online]. Available: https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf * [23] Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y. Ng, “Reading digits in natural images with unsupervised feature learning,” in _NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011_ , Granada, Spain, Dec. 2011. * [24] H. Xiao, K. Rasul, and R. Vollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,” _arXiv preprint arXiv:1708.07747_ , 2017. * [25] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in _Proceedings of the 37th International Conference on Machine Learning_ , vol. 119, Vienna, Austria, Jul. 2020, pp. 1597–1607. * [26] J.-B. Grill, F. Strub, F. Altché, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azar _et al._ , “Bootstrap your own latent-a new approach to self-supervised learning,” _Advances in Neural Information Processing Systems_ , vol. 33, pp. 21 271–21 284, Dec. 2020. * [27] J. Li, P. Zhou, C. Xiong, and S. C. Hoi, “Prototypical contrastive learning of unsupervised representations,” _arXiv preprint arXiv:2005.04966_ , 2020. * [28] W. Zhuang, X. Gan, Y. Wen, S. Zhang, and S. Yi, “Collaborative unsupervised visual representation learning from decentralized data,” in _Proceedings of the IEEE/CVF International Conference on Computer Vision_ , Oct. 2021, pp. 4912–4921. * [29] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” in _Advances in Neural Information Processing Systems_ , vol. 30, Long Beach, CA, USA, Dec. 2017. * [30] Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” _IEEE Transactions on Image Processing_ , vol. 13, no. 4, pp. 600–612, 2004. * [31] D. Dangwal, V. T. Lee, H. J. Kim, T. Shen, M. Cowan, R. Shah, C. Trippel, B. Reagen, T. Sherwood, V. Balntas _et al._ , “Analysis and mitigations of reverse engineering attacks on local feature descriptors,” _arXiv preprint arXiv:2105.03812_ , 2021.
Choice Disjunctive Queries in Logic Programming Keehang Kwon Dept. of Computer Engineering, DongA University Busan 604-714, Korea <EMAIL_ADDRESS> Abstract: One of the long-standing research problems on logic programming is to treat the cut predicate in a logical, high-level way. We argue that this problem can be solved by adopting linear logic and choice-disjunctive goal formulas of the form $G_{0}\oplus G_{1}$ where $G_{0},G_{1}$ are goals. These goals have the following intended semantics: $choose$ the true disjunct $G_{i}$ and execute $G_{i}$ where $i(=0\ {\rm or}\ 1)$, while $discarding$ the unchosen disjunct. Note that only one goal can remain alive during execution. These goals thus allow us to specify mutually exclusive tasks in a high-level way. keywords: Prolog, mutual exclusion, cut, linear logic, computability logic ## 1 Introduction One of the long-standing research problems on logic programming is to treat the extra-logical primitive in a high-level way. The advances of logic programming have enriched Horn clauses with additional programming primitives in a high-level way (higher-order programming, modules, local constants, etc). Nevertheless some key constructs could not be dealt with in a high-level way, in particular when we are concerned with mutual exclusion (and the cut predicate). Consequently, much attention [10, 11, 6] has been given to finding a semantics that captures the cut predicate. However, these proposals are not $logical$ in that a well-defined yet simple declarative meaning as well as its proof theory are still missing, exposing low-level operational details. In this paper, inspired by the work in [4], we propose a purely logical solution to this problem. It involves the direct employment of linear logic [2] to allow for choice-disjunctive goals. A choice-disjunctive goal is of the form $G_{0}\oplus G_{1}$ where $G_{0},G_{1}$ are goals. (A more intuitive name would be $choose(G_{0},G_{1})$.) Executing this goal with respect to a program $D$ – $ex(D,G_{0}\oplus G_{1})$ – has the following intended semantics: $\mbox{\rm choose a true one between}\ ex(D,G_{0}),ex(D,G_{1}).$ An illustration of this aspect is provided by the following definition of the relation $son(X,Y)$ which holds if $Y$ is a son of $X$.: $son(X,Y)$ ${\rm:-}$ | $(male(X)\land father(Y,X))\ \oplus$ ---|--- | $(female(X)\land mother(Y,X)).$ The body of the definition above contains a mutually exclusive goal, denoted by $\oplus$. As a particular example, solving the query $son(tom,Y)$ would result in selecting and executing the first goal $male(tom)\land father(tom,Y)$, while discarding the second one. The given goal will succeed, producing solutions for $Y$. Of course, we can specify mutually exclusive goals using cut in Prolog, but it is well-known that cuts complicates the declarative meaning of the program [1]. Our language makes it possible to formulate mutually exclusive goals in a high-level way. The class of choice disjunctive goals is, in a sense, a high-level abstraction for the cut predicate. As seen from the example above, choice-disjunctive goals can be used to perform mutually exclusive tasks. There are several linear logic languages [3, 12] in which goals of the form $G_{0}\oplus G_{1}$ are present. A common yet problematic aspect of these works is their treatment of the $\oplus$-goals: these goals are treated as inclusive-OR (or classical disjunctive) goals rather than exclusive-OR ones: $ex(D,G_{0}\oplus G_{1})\ {\rm if}\ ex(D,G_{0})\ \lor ex(D,G_{1})$ where $\lor$ represents classical disjunction. Hence, the declarative reading of $\oplus$ – known as the machine’s choice – is $violated$ in these languages. A satisfactory solution can be obtained by adding the choice action, as discussed above, to their execution model of $\oplus$ so that the execution respects the declarative reading of $\oplus$, while maintaining provability. Hence, the main difference is that, once a goal is chosen, the unchosen goal will be discarded in our language, while it will remain alive (typically through a creation of a choicepoint) in those languages. This paper proposes Prolog⊕, an extension of Prolog with choice-disjunctive operators in goal formulas. The remainder of this paper is structured as follows. We describe Prolog⊕ in the next section. In Section 3, we present some examples of Prolog⊕. Section 4 concludes the paper. ## 2 The Language The language is a version of Horn clauses with choice-disjunctive goals. Note that we disallow linear clauses here, thus allowing only reusable clauses. It is described by $G$\- and $D$-formulas given by the syntax rules below: | $G::=$ | $A\;|\;t=s\;|\;G\land G\;|\;\exists x\ G\;|\;G\oplus G$ ---|---|--- | $D::=$ | $A\;|\;G\supset A\ \;|\;\forall x\ D\;|\;D\land D$ In the rules above, $t,s$ represent terms, and $A$ represents an atomic formula. A $D$-formula is called a Horn clause with choice-disjunctive goals. The logic programming paradigm such as Prolog was originally founded on the resolution method. But this approach was difficult to extend to richer logics. The use of sequent calculus allows us to overcome this limit. Furthermore, uniform proofs [9] allows us to execute logic programs in an efficient way by integrating two separate phases – the proof phase and the execution phase – into a single phase. We adopt this approach below. We will present a machine’s strategy for this language as a set of rules. These rules in fact depend on the top-level constructor in the expression, a property known as uniform provability[8, 9, 7]. Note that execution alternates between two phases: the goal-reduction phase and the backchaining phase. In the goal-reduction phase (denoted by $ex(D,G)$), the machine tries to solve a goal $G$ from a clause $D$ by simplifying $G$ . The rule (6) – (9) are related to this phase. If $G$ becomes an atom, the machine switches to the backchaining mode. This is encoded in the rule (5). In the backchaining mode (denoted by $bc(D_{1},D,A)$), the machine tries to solve an atomic goal $A$ by first reducing a Horn clause $D_{1}$ to simpler forms (via rule (3) and (4)) and then backchaining on the resulting clause (via rule (1) and (2)). Definition 1. Let $G$ be a goal and let $D$ be a program. Then the notion of executing $\langle D,G\rangle$ – $ex(D,G)$ – is defined as follows: * (1) $bc(A,D,A)$. % This is a success. * (2) $bc((G_{0}\supset A),D,A)$ if $ex(D,G_{0})$. * (3) $bc(D_{1}\land D_{2},D,A)$ if $bc(D_{1},D,A)$ or $bc(D_{2},D,A)$. * (4) $bc(\forall xD_{1},D,A)$ if $bc([t/x]D_{1},D,A)$. * (5) $ex(D,A)$ if $bc(D,D,A)$. * (6) $ex(D,t=s)$ if $unify(t,s)$ % $t,s$ are terms. * (7) $ex(D,G_{0}\land G_{1})$ if $ex(D,G_{0})$ $and$ $ex(D,G_{1})$. * (8) $ex(D,\exists xG_{0})$ if $ex(D,[t/x]G_{0})$. * (9) $ex(D,G_{0}\oplus G_{1})$ if select a successful disjunct between $ex(D,G_{0})$ and $ex(D,G_{1})$. % This goal behaves as exclusive-OR. In the above rules, only the rule (9) is a novel feature. To be specific, this goal first attempts to execute $G_{0}$, discarding $G_{1}$. If it succeeds, then do nothing (and do not leave any choice point for $G_{1}$ ). If it fails, then $G_{1}$ is attempted. Implementing $G_{0}\oplus G_{1}$ poses no difficulties. For example, it can be done by translating it to a Prolog disjunctive goal of the form $(G_{0},!);G_{1}$ where ; denotes a Prolog disjuction. The cut then destroys the choice point created for $G_{1}$ if $G_{0}$ succeeds. On the contrary, the same goal $G_{0}\oplus G_{1}$ will be translated to $G_{0};G_{1}$ in other linear logic languages. The following theorem connects our language to linear logic. Its proof is obtained from [3] and from the simple observation that our modified execution rule preserves provability. ###### Theorem 1 Let $D$ be a program and let $G$ be a goal. Then, $ex(D,G)$ terminates with a success if and only if $G$ follows from $D$ in intuitionistic linear logic. Furthermore, it respects the declarative reading of the operator $\oplus$. ## 3 Examples Let us first consider the relation $f(X,Y)$ specified by two rules: * (1) if $X<2$, then $Y=0$. * (2) if $X\geq 2$, then $Y=3$. The two conditions are mutually exclusive which is expressed by using the cut in traditional logic programming as shown below: $f(X,0):-\ X<2,!.$ $f(X,3):-\ X\geq 2.$ Using cut, we can specify mutually exclusive goals, but cuts affect the declarative meaning of the program. Our language makes it possible to formulate mutually exclusive goals through the choice-disjunctive goals as shown below: $f(X,Y)$ ${\rm:-}$ | $(X\geq 2\land Y=3)\ \oplus$ ---|--- | $(X<2\land Y=0)$ The new program, equipped with $\oplus$-goals, is more readable than the original version with cuts, while preserving the same efficiency. A similar example is provided by the following “max” program that finds the larger of two numbers. $max(X,Y,Max)$ ${\rm:-}$ | $(X\geq Y\land Max=X)\ \oplus$ ---|--- | $(X<Y\land Max=Y)$ These two goals in the body of the above clause are mutually exclusive. Hence, only one of these two goals can succeed. For example, consider a goal $max(3,9,Max)$. Solving this goal has the effect of choosing and executing the second goal $(3<9)\land Max=9$, producing the result $Max=9$. As another example, we consider the relation $member(X,L)$ for establishing whether $X$ is in the list $L$. A typical Prolog definition of $member(X,L)$ is shown below: $member(X,[Y|L])$ ${\rm:-}$ | $(Y=X)\ \lor member(X,L)$ ---|--- This definition is nondeterministic in the sense that it can find any occurrence of $X$. Our language in Section 2 makes it possible to change $member$ to be deterministic and more efficient: only one occurrence can be found. An example of this is provided by the following program. $member(X,[Y|L])$ ${\rm:-}$ | $(Y=X)\ \oplus member(X,L)$ ---|--- As a final example, we consider the relation $rprime$ for establishing whether the keyboard input data $X$ is prime or not. An example of this is provided by the following program. $rprime$ ${\rm:-}$ | $read(X)\land$ ---|--- | $(prime(X)\land\ write(`prime^{\prime}))\oplus$ | $(composite(X)\land\ write(`composite^{\prime}))$ ## 4 Conclusion In this paper, we have considered an extension to Prolog with choice- disjunctive goals. This extension allows goals of the form $G_{0}\oplus G_{1}$ where $G_{0},G_{1}$ are goals. These goals are particularly useful for replacing the cut in Prolog, making Prolog more concise and more readable. In the near future, we plan to investigate the connection between Prolog⊕and Japaridze’s Computability Logic(CL)[4, 5]. CL is a new semantic platform for reinterpreting logic as a theory of tasks. Formulas in CL stand for instructions that can carry out some tasks. We plan to investigate whether our operational semantics is sound and complete with respect to the semantics of CL. ## References * [1] I. Bratko, “Prolog:programming for AI ”, Addison Wesley, 2001 (3rd edition). * [2] J.Y. Girard, “Linear Logic”, Theoretical Computer Science, vol.50, pp.1–102, 1987\. * [3] J. Hodas and D. Miller, “Logic Programming in a Fragment of Intuitionistic Linear Logic”, Information and Computation, vol.110, pp.327–365, 1994. * [4] G. Japaridze, “Introduction to computability logic”, Annals of Pure and Applied Logic, vol.123, pp.1–99, 2003. * [5] G. Japaridze, “Sequential operators in computability logic”, Information and Computation, vol.206, No.12, pp.1443-1475, 2008. * [6] J. Kriener and A. King, “RedAlert: Determinacy Inference for Prolog”, Theory and Practice of Logic Programming, vol.11, no.4-5. pp.182–196. * [7] E. Komendantskaya and V. Komendantsky, “On uniform proof-theoretical operational semantics for logic programming”, In J.-Y. Beziau and A.Costa-Leite, editors, Perspectives on Universal Logic, pages 379–394. Polimetrica Publisher, 2007. * [8] D. Miller, “A logical analysis of modules in logic programming”, Journal of Logic Programming, vol.6, pp.79–108, 1989. * [9] D. Miller, G. Nadathur, F. Pfenning, and A. Scedrov, “Uniform proofs as a foundation for logic programming”, Annals of Pure and Applied Logic, vol.51, pp.125–157, 1991. * [10] A. Porto, “A structured alternative to Prolog with simple compositional semantics”, Theory and Practice of Logic Programming, vol.11, No.4-5, pp.611-627, 2011. * [11] A. Saurin, “Towards Ludics Programming: Interactive Proof Search”, International Conference on Logic Programming, pages 253–268. 2008. * [12] M. D. Winikoff, “Logic Programming with Linear Logic”, PhD. Thesis, Univ. Melbourne, 1997\.
# Spectral approach to Korteweg-de Vries equations on the compactified real line Christian Klein Institut de Mathématiques de Bourgogne, UMR 5584 Université de Bourgogne-Franche-Comté, 9 avenue Alain Savary, 21078 Dijon Cedex, France E-mail<EMAIL_ADDRESS>and Nikola Stoilov Institut de Mathématiques de Bourgogne, UMR 5584 Université de Bourgogne-Franche-Comté, 9 avenue Alain Savary, 21078 Dijon Cedex, France E-mail<EMAIL_ADDRESS> ###### Abstract. We present a numerical approach for generalised Korteweg-de Vries (KdV) equations on the real line. In the spatial dimension we compactify the real line and apply a Chebyshev collocation method. The time integration is performed with an implicit Runge-Kutta method of fourth order. Several examples are discussed: initial data bounded but not vanishing at infinity as well as data not satisfying the Faddeev condition, i.e. with a slow decay towards infinity. This work is partially supported by the ANR-FWF project ANuI - ANR-17-CE40-0035, the isite BFC project NAANoD, the ANR-17-EURE-0002 EIPHI and by the European Union Horizon 2020 research and innovation program under the Marie Sklodowska-Curie RISE 2017 grant agreement no. 778010 IPaDEGAN ## 1\. Introduction Generalised Korteweg-de Vries (gKdV) equations, (1) $u_{t}(x,t)+u_{xxx}(x,t)+u(x,t)^{p-1}u_{x}(x,t)=0,$ where $p\in\mathbb{N}$, $u:\mathbb{R}\times\mathbb{R}^{+}\mapsto\mathbb{R}$, appear as asymptotic models in hydrodynamics, nonlinear optics, plasma physics, Bose-Einstein condensates, and essentially in most situations where predominantly one-dimensional phenomena are discussed and where dispersion dominates dissipation. Whereas this applies in particular to the case of the classical Korteweg-de Vries (KdV) equation ($p=2$), there are applications for instance in electrodynamics for the modified KdV equation ($p=3$), see [21]. Because of their importance in applications, there has been a considerable activity in developing numerical approaches for the gKdV equations. For initial data which are periodic or rapidly decreasing, numerical approaches based on the approximation of $u$ in (1) via truncated Fourier series, i.e., trigonometric polynomials, are very efficient, see for instance [14, 15] and references therein. The Fourier approach, that is restricting the data to an interval $L[-\pi,\pi]$ ($L=const$, $L\in\mathbb{R}^{+}$) and continuing them periodically with period $2\pi L$ on the whole real line works very well for periodic functions and exhibits _spectral convergence_ , namely an exponential decrease of the numerical error with the number of Fourier modes. Schwartz class functions can be treated as periodic as one works with a finite precision and since $L$ can be chosen large enough such that all necessary derivatives of $u$ vanish at the domain boundaries with numerical precision. However, the same approach for initial data which do not tend to zero or are only slowly decreasing to zero for $x\to\infty$, would in general imply a Gibbs phenomenon at the domain boundaries. The resulting method would therefore be only of first order in the number of Fourier modes, making it impossible to reach the high resolution necessary to treat e.g. rapid oscillations in the solution. Therefore, such situations are typically treated on a finite interval. This leads to the problem of how to impose boundary conditions, so that inside the computational domain the solution is the same as if the computation was done on the whole real line. Such boundary conditions are called _transparent_. Bérenger [2] introduced _perfectly matched layers_ (PML) in electrodynamics to address this problem by extending the computational domain to layers glued to the domain boundaries. Inside the layers, the equation under consideration is deformed to a dissipative one, which is chosen to efficiently dissipate the solution. Whereas this works well for linear equations, examples for the nonlinear Schrödinger equation, see [26, 3], showed that in a nonlinear setting there will be back reflections from the layers to the computational domain. For integrable equations exact transparent boundary conditions (TBC) can be given, e.g. for the case of modified KdV see [27] based on [9]. The problem with both PML and TBC is that they in general require initial data with compact support within the computational domain, thus limiting the class of solutions that can be studied. The goal of the present paper is to establish a spectral numerical approach for generalised KdV equations (1) for initial data that are analytic on the whole real line, that are slowly decreasing towards infinity, or are bounded there. This numerical approach exhibits spectral convergence on the whole real line with a technique similar to [12]. Both the classical KdV equation and the modified KdV equation are completely integrable, which means they have an infinite number of conserved quantities (for a comprehensive review on integrability see [25]). In all other cases, the generalised KdV equations have only three conserved quantities, $\int_{\mathbb{R}}udx$ and the $L^{2}$ norm of $u$ and the energy (2) $E[u]=\int_{\mathbb{R}}\left(\frac{u^{p+1}}{p(p+1)}-\frac{1}{2}u_{x}^{2}\right)dx.$ The complete integrability of the classical KdV equation made it one of the best studied non-linear dispersive equations with a rather complete understanding of its solutions. However, even in this case there are open questions which motivate us to provide numerical tools to complement analytical studies in this context. The standard inverse scattering approach to KdV is only applicable if the Faddeev decay condition [8], (3) $\int_{\mathbb{R}}(1+|x|)|u_{0}(x)|dx<\infty,$ holds for the initial data $u(x,0)=u_{0}(x)$. The direct scattering approach to KdV involves the determination of the spectrum of the Schrödinger equation for the potential $u_{0}(x)$, $\psi_{xx}+u_{0}(x)\psi=E\psi,$ see [25]. It is known that this discrete spectrum is finite if the Faddeev condition (3) is satisfied, see [17]. But except for the periodic case, see for instance [1], there is no complete understanding of the case of initial data not satisfying (3). The goal of the present paper is to provide numerical tools to study such cases. Of course, numerically one will be only able to study finite times and thus will not be able to address the question whether the time evolution of such data can lead to an infinite number of solitons. The paper is organized as follows: In section 2 we summarize a few theoretical facts on generalised KdV equations. In section 3 we choose a compactification of the real line and describe the numerical approach for the generalised KdV equations. In section 4 we discuss several examples. Concluding remarks are added in section 5. ## 2\. Theoretical preliminaries In this section we summarize basic facts about generalised KdV equations needed in the following. Though the KdV equations (1) are only completely integrable for $p=2$ and $p=3$, they have for all integer values of $p\geq 2$ a solitary travelling wave solution which is explicitly given by $u=Q_{c}(x-x_{0}-ct)$ with $x_{0},c=const$ and with (4) $Q_{c}(z)=\left(\frac{p(p+1)c}{2}\,\mbox{sech}^{2}\frac{\sqrt{c}(p-1)}{2}z\right)^{1/(p-1)}.$ Thus, we have $Q_{c}(z)=c^{1/(p-1)}Q(\sqrt{c}z)$, where we have put $Q:=Q_{1}$. This simple scaling property of the _solitons_ allows to concentrate on the case $c=1$. If we refer in the following to the generalised KdV soliton, it is always implied that $c=1$. It was shown in [4] that these solitons are linearly unstable for $p>4$. Note that the energy of the soliton vanishes for $p=5$. The generalised KdV equation has the following scaling invariance: $x\mapsto x/\lambda$, $t\mapsto t/\lambda^{3}$ and $u\mapsto\lambda^{2/(p-1)}u$ with $\lambda=const$. For $p=5$, the $L^{2}$ norm of $u$ is invariant under this rescaling, and this case is consequently called $L^{2}$-critical. It is shown in [18] that solutions to the $L^{2}$ critical generalised KdV equation can have a blow-up in finite time for smooth initial data. The mechanism of the blow-up for initial data close to the soliton is discussed in [20]. In the present paper, we will only study the sub-critical cases, $p\leq 4$. A convenient way to treat solutions varying on length scales of order $1/\epsilon$ for times of order $1/\epsilon$ for $0<\epsilon\ll 1$ is to consider the map $x\mapsto\epsilon x$, $t\mapsto\epsilon t$. This leads for (1) to (once more we use the same symbol $u$ for the transformed and the original solution) (5) $u_{t}+\epsilon^{2}u_{xxx}+u^{p-1}u_{x}=0.$ The formal limit $\epsilon\to 0$ of this equation yields a generalised Hopf equation, $u_{t}+u^{p-1}u_{x}=0$. It is known that such equations can have a _hyperbolic blow-up_ , i.e., shocks for general smooth initial data, for instance for data with a single hump. The generic _break-up_ of such solutions at a point $(x_{c},t_{c},u_{c})$, see for instance the discussion in [7] and references therein, is characterized by the equations $\displaystyle a(u_{c})t_{c}+\Phi(u_{c})$ $\displaystyle=x_{c},$ (6) $\displaystyle a^{\prime}(u_{c})t_{c}+\Phi^{\prime}(u_{c})$ $\displaystyle=0,$ $\displaystyle a^{\prime\prime}(u_{c})t_{c}+\Phi^{\prime\prime}(u_{c})$ $\displaystyle=0,$ where $a(u)=u^{p-1}$. It is known that dispersive regularizations of dispersionless equations as (5) in our case will lead to _dispersive shock waves_ (DSWs), i.e., rapid modulated oscillations near the shock of the solution of the corresponding dispersionless equation for the same initial data. Dubrovin [5, 6] presented a conjecture that the onset of a DSW is _universal_ for a large class of dispersive equations and of initial data, and that it is given by a special solution of the so-called Painlevé $P_{I}^{2}$ equation, see for instance [11]. This conjecture was numerically shown to apply to the generalised KdV equations in [7]. In this paper we provide the numerical tools to do so for larger classes of initial data than in [7], however we leave addressing the universality conjecture in the present context for a future work. ## 3\. Numerical approach In this section we present the numerical approach to treat generalised KdV equations on the compactified real line. We first introduce the compactification which allows to study the equation on the interval $[-1,1]$ where we use a Chebyshev collocation method. The resulting system of ordinary differential equations is then integrated in time with an implicit Runge-Kutta scheme. ### 3.1. Compactification To numerically treat the KdV equation, we first map the real line via the well known map (this is the standard compactification used for Minkowski spacetime) (7) $x=c\tan\frac{\pi l}{2},\quad l\in[-1,1],\quad c=const,$ to the interval $[-1,1]$. This implies (8) $\partial_{x}=\frac{2}{\pi c}\cos^{2}\frac{l\pi}{2}\partial_{l}.$ The role of the constant $c$ is to control the numerical resolution in various parts of the real line within certain limits. This is illustrated in Fig. 1 where we have introduced _Chebyshev collocation points_ for $l\in[-1,1]$, i.e., (9) $l_{n}=\cos(n\pi/N),\quad n=0,1,\ldots,N,\quad N\in\mathbb{N}.$ It can be seen in Fig. 1 that the density of the points near $x=0$ is higher for smaller $c$ (numerically the function $\tilde{v}$ is studied as a function of $l$, but since in applications the dependence on $x$ is more important, we plot its $x$-dependence which can be obtained via (7)). Note, however, that the spectral methods we apply in this paper are global. This means that a high resolution in part of the studied domain is not necessarily beneficial, only the overall resolution on the whole interval is important. Therefore we generally apply values of $c$ close to 1. Figure 1. Distribution of the Chebyshev collocation points (9) on a Lorentzian profile under the map (7) for $N=800$; on the left for $c=0.1$, in the middle for $c=1$, on the right for $c=10$. ### 3.2. Boundary conditions The map (7) transforms the KdV equation (1) on the real line to an equation on the interval $[-1,1]$, which is singular at the points $l=\pm 1$. Because of this singular behavior, no boundary conditions need to be imposed there. However, in practice it is useful to give boundary conditions at these points to stabilize the numerical approaches. We apply a vanishing condition for $u$ at these points, and a _clamped boundary condition_ for $l=-1$, i.e., the three conditions (in an abuse of notation, we denote $u(x,t)$ and $u(l,t)$ with the same letter) (10) $u(l,t)\biggr{\rvert}_{l=1}=0,\quad u(l,t)\biggr{\rvert}_{l=-1}=0,\quad u_{l}(l,t)\biggr{\rvert}_{l=-1}=0.$ The approach we present here can be generalised to functions $u$ which do not tend to 0 for $|x|\to\infty$, but which are bounded there. In this case we write (11) $u=v+V,\quad V=A\frac{1+l}{2}+B\frac{(1+l)^{2}}{4}+C\frac{1-l}{2},$ where $A$, $B$, $C$ are constants such that $v(l=\pm 1)=v_{l}(l=-1)=0$. To treat the clamped boundary condition for $l=-1$, we use as in [22] the ansatz (12) $v=(1+l)\tilde{v}.$ This leads for (1) to the equation (13) $\tilde{v}_{t}+\frac{1}{1+l}[(1+l)\tilde{v}+V]_{xxx}+\frac{1}{1+l}[(1+l)\tilde{v}+V]^{p}[(1+l)\tilde{v}+V]_{x}=0$ which has to be solved for all $t$ with the condition $\tilde{v}(\pm 1)=0$. ### 3.3. Chebyshev differentiation matrices The dependence of $\tilde{v}$ in (13) on $l$ will be treated in standard way via Lagrange interpolation of $\tilde{v}$ on Chebyshev collocation points (9) as discussed in [22]. A derivative of $\tilde{v}$ with respect to $l$ is then approximated via the derivative of the Lagrange polynomial. This leads to the action of a matrix on the vector $\tilde{v}=(\tilde{v}(l_{0}),\ldots,\tilde{v}_{N})$ (again we use the same symbol for the function $\tilde{v}$ and the vector $\tilde{v}$), the well known _Chebyshev differentiation matrices_ $D$, see e.g., [23, 24]. This means with (8) that the derivatives $\partial_{x}$ are approximated by (14) $\partial_{x}\approx\frac{2}{\pi c}\mbox{diag}\left(\cos^{2}\frac{l\pi}{2}\right)D,$ where the diagonal matrix has the components $(\cos^{2}\frac{l_{0}\pi}{2},\ldots,\cos^{2}\frac{l_{N}\pi}{2})$. The Lagrange interpolation of a function on Chebyshev collocation points is closely related to an expansion of the function in terms of Chebyshev polynomials $T_{n}$, $n=0,1,\ldots$ (15) $\tilde{v}(l)\approx\sum_{n=0}^{N}v_{n}T_{n}(l),$ see the discussion in [22]. As shown again in [22], the Chebyshev coefficients $v_{n}$, $n=0,1,\ldots,N$ can be computed efficiently via a fast cosine transform which is closely related to the fast Fourier transform. It is also known that the Chebyshev coefficients for a function analytic on $[-1,1]$ decrease exponentially, and that the numerical error in approximating a function $\tilde{v}$ via (15) is of the order of the first omitted coefficients $v_{n}$ in the Chebyshev series. ### 3.4. Time integration After the discretisation in space equation (13) becomes an $N+1$-dimensional system of ordinary differential equations (ODEs) in $t$ of the form $\tilde{v}_{t}=f(\tilde{v})$ which can be numerically integrated in time with standard techniques. The discussion in [22] shows that Chebyshev differentiation matrices have a conditioning of the order $\mbox{cond}(D^{3})=O(N^{6})$. Thus explicit time integration schemes are problematic since stability conditions would necessitate prohibitively small time steps. Therefore, here we apply an implicit method, and since we are interested in capturing rapid oscillations in the expected DSWs, we use a fourth order method. Concretely, we apply a fourth order Runge-Kutta (IRK4) scheme, also called the Hammer-Hollingsworth method, a 2-stage Gauss scheme. The general formulation of an $s$-stage Runge–Kutta method for the initial value problem $\tilde{v}^{\prime}=f(\tilde{v},t),\,\,\,\,\tilde{v}(t_{0})=\tilde{v}_{0}$ reads: (16) $\displaystyle\tilde{v}_{n+1}=\tilde{v}_{n}+h\underset{i=1}{\overset{s}{\sum}}\,b_{i}K_{i},$ (17) $\displaystyle K_{i}=f\left(t_{n}+c_{i}h,\,\tilde{v}_{n}+h\underset{j=1}{\overset{s}{\sum}}\,a_{ij}K_{j}\right),$ where $b_{i},\,a_{ij},\,\,i,j=1,...,s$ are real numbers and $c_{i}=\underset{j=1}{\overset{s}{\sum}}\,a_{ij}$. For the IRK4 method used here, one has $c_{1}=\frac{1}{2}-\frac{\sqrt{3}}{6}$, $c_{2}=\frac{1}{2}+\frac{\sqrt{3}}{6}$, $a_{11}=a_{22}=1/4$, $a_{12}=\frac{1}{4}-\frac{\sqrt{3}}{6}$, $a_{21}=\frac{1}{4}+\frac{\sqrt{3}}{6}$ and $b_{1}=b_{2}=1/2$. Applying IRK4 to (13) we get the following system, $\displaystyle\mathcal{L}K_{1}$ $\displaystyle=-\frac{1}{1+l}[(1+l)(\tilde{v}+ha_{12}K_{2})+V]_{xxx}$ $\displaystyle-\frac{1}{1+l}[(1+l)(\tilde{v}+ha_{11}K_{1}+ha_{12}K_{2})+V]^{p}[(1+l)(\tilde{v}+ha_{11}K_{1}+ha_{12}K_{2})+V]_{x},$ $\displaystyle\mathcal{L}K_{2}$ $\displaystyle=-\frac{1}{1+l}[(1+l)(\tilde{v}+ha_{21}K_{1})+V]_{xxx}$ (18) $\displaystyle-\frac{1}{1+l}[(1+l)(\tilde{v}+ha_{21}K_{1}+ha_{22}K_{2})+V]^{p}[(1+l)(\tilde{v}+ha_{21}K_{1}+ha_{22}K_{2})+V]_{x},$ where $\mathcal{L}=\hat{1}+ha_{11}\frac{1}{1+l}\partial_{xxx}(1+l).$ Recall that $\partial_{x}$ is approximated in (18) via (14), and the third derivative with respect to $x$ is approximated as the cubic power of this. The system (18) will be solved iteratively with a simplified Newton iteration. This means that in each step of the iteration the new $K_{1}$ and $K_{2}$ are obtained by inverting the operator $\mathcal{L}$ only instead of the full Jacobian. The vanishing boundary conditions for $\tilde{v}$ and thus for $K_{1}$, $K_{2}$ are imposed as in [22]: the equations $\mathcal{L}K_{i}=F_{i}$, $i=1,2$, are solved by considering only the components $1,\ldots,N-1$ of the $K_{i}$; this means that we consider the reduced equations $\sum_{m=1}^{N-1}\mathcal{L}_{nm}K_{i,m}=F_{i,n},\quad n=0,\ldots,N-1,\quad i=1,2,$ and solve for $K_{i,1},\ldots,K_{i,N-1}$ only since the values $K_{i,0}=K_{i,N}=0$ are imposed. The resolution in time can be controlled via the conserved quantities of the generalised KdV equation. We consider in general the energy for functions vanishing at infinity. For functions which are just bounded at infinity, we consider a linear combination of the energy and the $L^{2}$ norm of the solution, (19) $\tilde{E}=\int_{\mathbb{R}}\left(\frac{u^{p+1}-\lambda u^{2}}{p(p+1)}-\frac{1}{2}u_{x}^{2}\right)dx,$ where the constant $\lambda$ is chosen such that the integrand is bounded at infinity. The conserved quantities will in actual computations depend on time due to numerical errors. But as discussed for instance in [14], the conservation of these quantities controls the numerical error which is in general overestimated by up to two orders of magnitude. ## 4\. Examples In this section we study two types of examples which illustrate the potential of the presented numerical approach. First we consider initial data in the form of a mollified (smoothed out) step function. Then we study initial data not satisfying the Faddeev condition $\int^{+\infty}_{-\infty}(1+|x|)|u(x)|dx<\infty$, see [25], but nevertheless vanishing for $|x|\to\infty$. These examples are studied for two types of nonlinearity, for $p=2$, the completely integrable KdV equation, and for $p=4$, a non-integrable generalised KdV equation discussed for instance in [19]. The latter is still sub-critical which means that there is no blow-up for sufficiently regular initial data. ### 4.1. Mollified step initial data We consider initial data of the form (20) $u(x,0)=\begin{cases}1&x<0\\\ \exp(-x^{2n})&x\geq 0\end{cases},$ where $n\in\mathbb{N}$. In Fig. 2 we show these data for $n=4$ on the left, and the corresponding function $\tilde{v}$ on the right (one has $A=-B=C=1$). ###### Remark 4.1. The function (20) is analytic everywhere except at zero, where it is $C^{2n-1}$, thus the convergence rate of a spectral method is expected to be algebraic. Nevertheless, in practice this does not have a detrimental effect on our approach and, as we can see below in figure 5 on the right, the behavior of Chebyshev coefficients is, due to the finite precision used, virtually the same as in the analytic case. Figure 2. Initial data (20) for $n=4$ on the left, and the corresponding $\tilde{v}=\tilde{v}\left(l(x)\right)$ the right. For the time evolution of these data, we use $c=2$, $N=600$ and $N_{t}=1000$ time steps for $t\in[0,0.01]$. The resulting solution to the KdV equation can be seen in Fig. 3. The formation of a dispersive shock wave is clearly visible. Figure 3. Solution to the KdV equation (1) with $p=2$ for the initial data (20) for $n=4$ in dependence of time. The slow decrease of the amplitude of the oscillations towards infinity, similar to the behaviour of the Airy function, is challenging for any numerical method. The Chebyshev coefficients $v_{n}$ (15) of the solution are shown in Fig. 4, on the left for $t=0$, on the right for $t=0.01$. They decrease exponentially to the order of the rounding error for the initial data indicating that an analytic (within numerical precision, see remark 4.1) function is numerically well resolved. The algebraic decay of the spectral coefficients for $t=0.01$ indicates an oscillatory singularity at infinity as for the Airy function. The spatial resolution is thus of the order of $10^{-4}$. The relative conservation of the modified energy (19) is during the whole computation of the order of $10^{-4}$. This means the solution is obtained to better than plotting accuracy. Figure 4. The Chebyshev coefficients (15), on the left for the mollified step initial data (20) on the right for the solution shown in Fig. 3 for $t=0.01$ Note that the DSW is not the same as in the case of an exact step, the classical Gurevitch-Pitaevski problem [13]. But one can for instance verify that this is the correct solution by considering a finite step smoothed out at both sides, (21) $u(x,0)=\begin{cases}1&x_{0}<x<0\\\ \exp(-x^{2n})&x\geq 0\\\ \exp(-(x-x_{0})^{2n})&x\leq x_{0}\end{cases}$ which can be conveniently treated with Fourier methods as in [14] to which the reader is referred for details and references. We use $N=2^{12}$ Fourier modes for $x\in 10[-\pi,\pi]$ and $N_{t}=1000$ time steps for a fourth order exponential time differencing method. In Fig. 5 we show on the left the solution of Fig. 3 for $t=0.01$, and for the initial data (21) with $n=4$ and $x_{0}=-5\pi$ at the same time. Figure 5. Solution to the KdV equation (1) with $p=2$ for the initial data (20) on the left, and for the initial data (21) on the right, both for $n=4$ and $t=0.01$. The solution to the generalised KdV equation with $p=4$ for the same initial data as in Fig. 2 can be seen in Fig. 6. We have used the same numerical parameters as for the case $p=2$, and we obtain the same numerical resolution. The form of the DSW is very similar to one for the standard KdV equation. Figure 6. Solution to the generalised KdV equation (1) with $p=4$ for the initial data (20) for $n=4$ in dependence of time. ### 4.2. Slowly decaying initial data In this subsection we consider initial data not satisfying the Faddeev condition, and we are interested in the long time behavior of the corresponding KdV solutions which is done by introducing a small parameter $\epsilon$ in (5). Concretely, we study initial data of the form (22) $u(x,0)=\frac{1}{(1+x^{2})^{a}},\quad a=\frac{1}{2},1.$ We use $c=2$, $N=800$, and $N_{t}=10^{4}$ time steps for $t\in[0,10]$. In Fig. 7 we show the KdV solution ($p=2$) in (5) for the initial data (22) for $\epsilon=10^{-1}$ and $a=1$ in dependence of time. It can be seen that several solitons appear. Figure 7. Solution to the KdV equation (1) with $p=2$ for the initial data (22) for $a=1$ in dependence of time. The corresponding KdV solution for the initial data (22) with $a=1/2$ and $\epsilon=10^{-1}$ can be seen in Fig. 8. Note that in contrast to the case $a=1$, the initial data do not satisfy the clamped boundary conditions for $x\to-\infty$, one has $A=-B=\pi/2$ and $C=0$ in (11). Figure 8. Solution to the KdV equation (1) with $p=2$ for the initial data (22) for $a=1/2$ in dependence of time. The solutions at the final time of Fig. 7 and 8 can be seen in Fig. 9, on the left for $a=1/2$, on the right for $a=1$. The slower decay towards infinity of the initial data with $a=1/2$ can be recognized. But at the time $t=10$, one observes the same number of solitons in both cases. The peaks in the solutions have been fitted to the solitons (4) which are shown in green in the same figure. It can be seen that the solitons are not yet fully separated from the background, but that they can be already clearly identified at this early stage. Figure 9. Solution to the KdV equation (1) with $p=2$ for the initial data (22) for $t=10$, on the left for $a=1/2$, on the right for $a=1$; in green fitted solitons (4). The relative computed energy is in both cases conserved to the order of $10^{-10}$. The Chebyshev coefficients for $t=10$ are shown in Fig. 10, on the left for $a=1/2$, on the right for $a=1$. It can be seen that the coefficients decrease as expected exponentially, and that the solutions are well resolved in space as well. Figure 10. Solution to the KdV equation (1) with $p=2$ for the initial data (22) for $t=10$, on the left for $a=1/2$, on the right for $a=1$. If the same initial data as in Fig. 9 are considered for the generalised KdV equation (5) with $p=4$, one obtains for $t=10$ the solutions shown in Fig. 11. The solitons are here much more peaked than in the KdV case of Fig. 9 which is also illustrated by the fit to the solitons. Consequently the same numerical parameters as there lead for the generalised KdV solution to a lower resolution: the relative conservation of the energy is of the order of $10^{-4}$, and Chebyshev coefficients decrease still exponentially, but only to the order of $10^{-6}$ in this case. Figure 11. Solution to the KdV equation (1) with $p=4$ for the initial data (22) for $t=10$, on the left for $a=1/2$, on the right for $a=1$. ## 5\. Outlook In this paper we have presented a numerical approach for generalised KdV equations on the compactified real line which allows to approximate functions which are smooth on $\mathbb{R}\cup\\{\infty\\}$ with spectral accuracy, i.e., with a numerical error decreasing exponentially with the number of collocation points. The time integration is performed with an implicit fourth order method. One direction of further research will be to improve the efficiency of the time integration, ideally an explicit approach, for instance similar to the approach of [16] in the context of the Schrödinger equation. Of special interest is, however, the application of the techniques of the present paper to the numerical study of blow-up in the context of generalised KdV equations, i.e., for (1) with $p\geq 5$. The current approach would allow to study the dynamically rescaled generalised KdV equations, see for instance [15] without the problems there related to the use of Fourier methods. In this context it would be beneficial to apply a multidomain spectral method as in [3] for Schrödinger equations where the compactified real axis is divided into several domains each of which is mapped to the interval $[-1,1]$. This allows for a more efficient allocation of numerical resolution than via the choice of the parameter $c$ in (7), in particular a higher resolution near the expected blow-up. To this end, matching conditions at the domain boundaries (the solution $u$ has to be $C^{2}$ there) have to imposed which is numerically problematic. This will be a subject of further research. ## References * [1] Belokolos E, Bobenko A, Enolskii V, Its A and Matveev V 1994 Algebro-Geometric Approach to Nonlinear Integrable Equations, Springer Series in Nonlinear Dynamics (Berlin: Springer) * [2] J. Bérenger. A perfectly matched layer for the absorption of electromagnetic waves. J. Comput. Phys. 114, 185-200, 1994. * [3] M. Birem and C. Klein, Multidomain spectral method for Schrödinger equations, Adv. Comp. Math., 42(2), 395-423 DOI 10.1007/s10444-015-9429-9 (2016) * [4] J. L. Bona, P. E. Souganidis, W. A. Strauss, Stability and instability of solitary waves of Korteweg-de Vries type, Proc. R. Soc. Lond. A 411 (1987) 395 - 412. * [5] B. Dubrovin, On Hamiltonian perturbations of hyperbolic systems of conservation laws, II: universality of critical behaviour, Comm. Math. Phys. 267 (2006) 117 - 139. * [6] B. Dubrovin, On universality of critical behaviour in Hamiltonian PDEs, Amer. Math. Soc. Transl. 224 (2008) 59-109. * [7] B. Dubrovin, T. Grava and C. Klein, Numerical Study of breakup in generalised Korteweg-de Vries and Kawahara equations, SIAM J. Appl. Math., Vol 71, 983-1008 (2011). * [8] L.D. Faddeev, On the relation between the S-matrix and potential for the one dimensional Schrödinger operator, Dok. Akad.Nauk SSSR 121 (1958), 63-66, see also Amer. Math. Soc. Transl; (2) 65 (1967), 139-166. * [9] A.S. Fokas: The generalised Dirichlet-to-Neumann map for certain non linear evolution PDEs. Comm. Pure Appl. Math. 58(5), 639-670 (2005) * [10] T. Grava and C. Klein, Numerical study of the small dispersion limit of the Korteweg-de Vries equation and asymptotic solutions, Physica D, 10.1016/j.physd.2012.04.001 (2012). * [11] T. Grava, A. Kapaev, and C. Klein, On the tritronquée solutions of $P^{2}_{I}$ , Constr. Approx. 41, 425 -466, DOI 10.1007/s00365-015-9285-3 (2015). * [12] C.E. Grosch and S.A. Orszag, Numerical solution of problems in unbounded regions: coordinate transforms, J. Comput. Phys. 25, 273-296 (1977). * [13] A. V. Gurevich, L. P. Pitaevskii, Nonstationary structure of a collisionless shock wave. Eksper. Teoret. Fiz. 65 (1973), no. 8, 5901-7604; translation in Soviet Physics JETP 38 (1974), no. 2, 2911-7297. * [14] C. Klein, Fourth order time-stepping for low dispersion Korteweg-de Vries and nonlinear Schrödinger equation, ETNA Vol. 29 116-135 (2008). * [15] C. Klein and R. Peter, Numerical study of blow-up in solutions to generalised Korteweg-de Vries equations, Physica D 304-305 (2015), 52-78 DOI 10.1016/j.physd.2015.04.003. * [16] C. Klein, N. Stoilov, Numerical study of the transverse stability of the Peregrine solution, Stud Appl Math. 145 (2020) 36-51. https://doi.org/10.1111/sapm.12306 * [17] V.A. Marchenko, Sturm-Liouville operators and applications. Revised edition, AMS Chelsea Publishing vol. 373 (2001). * [18] Y. Martel and F. Merle, Stability of Blow-Up Profile and Lower Bounds for Blow-Up Rate for the Critical generalised KdV Equation, Ann. Math., 155, (2002), 235-280. * [19] Y. Martel, F. Merle, On the nonexistence of pure multi-solitons for the quartic gKdV equation. Int Math Res Notices (2015) (3): 688-739. * [20] Y. Martel, F. Merle, P. Raphaél, Blow up for the critical gKdV equation I: dynamics near the soliton. Acta Math. 212 (2014), no. 1, 59-140. * [21] T. L. Perel’man, A. Kh. Fridman, and M. M. EI’yashevich, A modified Korteweg-de Vries equation in electrohydrodynamics, Sov. Phys. JETP, Vol. 39, No.4, October 1974 * [22] L.N. Trefethen, Spectral Methods in Matlab. SIAM, Philadelphia, PA (2000) * [23] www.comlab.ox.ac.uk/oucl/work/nick.trefethen * [24] J.A.C. Weideman, S.C. Reddy: A Matlab differentiation matrix suite. ACM TOMS 26, 465-1-7519 (2000) * [25] V.E. Zakharov ed., . What is integrability?. Berlin: Springer-Verlag, (1991). * [26] C. Zheng, A perfectly matched layer approach to the nonlinear Schrödinger wave equations. J. Comput. Phys. 227, 537-556 (2007) * [27] C. Zheng, Numerical simulation of a modified KdV equation on the whole real axis, Numer. Math. (2006) 105:315-335 DOI 10.1007/s00211-006-0044-z
# Role of electronic correlations in the Kagome lattice superconductor LaRh3B2 Savita Chaudhary Department of Physical Sciences, Indian Institute of Science Education and Research (IISER) Mohali, Knowledge City, Sector 81, Mohali 140306, India. Shama Department of Physical Sciences, Indian Institute of Science Education and Research (IISER) Mohali, Knowledge City, Sector 81, Mohali 140306, India. Jaskaran Singh Department of Physical Sciences, Indian Institute of Science Education and Research (IISER) Mohali, Knowledge City, Sector 81, Mohali 140306, India. Department of Physics, Punjabi University, Patiala, 147002, India. Armando Consiglio Institut für Theoretische Physik und Astrophysik and Würzburg-Dresden Cluster of Excellence ct.qmat, Universität Würzburg, 97074 Würzburg, Germany Domenico Di Sante Department of Physics and Astronomy, Alma Mater Studiorum, University of Bologna, 40127 Bologna, Italy Center for Computational Quantum Physics, Flatiron Institute, 162 5th Avenue, New York, New York 10010, USA Ronny Thomale Institut für Theoretische Physik und Astrophysik and Würzburg-Dresden Cluster of Excellence ct.qmat, Universität Würzburg, 97074 Würzburg, Germany Yogesh Singh Department of Physical Sciences, Indian Institute of Science Education and Research (IISER) Mohali, Knowledge City, Sector 81, Mohali 140306, India. ###### Abstract LaRh3B2 crystallizes in a layered structure where Rh atoms form a perfect Kagome lattice. The material shows superconductivity at $T_{c}\approx 2.6$ K and no signature for density wave instabilities. We report our measurements of electronic transport, magnetization, and heat capacity in the normal and superconducting state, and derive normal and superconducting parameters. From first principles calculations of the electronic band structure, we identify all features of Kagome bands predominantly formed by the Rh $d$ orbitals: a flat band, Dirac cones, and van Hove singularities. The calculation of the phonon dispersions and electron-phonon coupling suggests a strong similarity between LaRh3B2 and AV3Sb5 (A=K,Cs,Rb). For LaRh3B2, it matches quantitatively with the observed $T_{c}$, supporting a conventional phonon mediated pairing mechanism. By comparison to the $A$V3Sb5 family, we conjecture a reduced importance of electron correlations in LaRh3B2. ## I Introduction The kagome lattice has long been a playground for novel physics in condensed matter. Insulating kagome lattice realizations with localized magnetic moments are platforms to explore the effects of geometric magnetic frustration. The quantum spin liquid (QSL) ground state in the mineral Herbertsmithite ZnCu3(OH)Cl2 is a prime example of this behaviour Fu2015 ; Han2012 . Insulating quantum magnets with a kagome network in higher dimensions have also shown novel frustrated magnetism and QSL behaviour Balz ; Okamoto2007 ; Singh2013 . More recently metallic kagome lattice materials have been brought into focus due to the prediction that the electronic structure of electrons on a kagome lattice might allow to access correlated Dirac cones or van Hove singularities near the Fermi energyKiesel2013 ; Wang2013 ; Mazin2014 . The two-dimensional kagome lattice has features in its band structure which provide, even still in the itinerant limit, the opportunity of marrying non- trivial topology and strong electron correlations. Search for realizations of a material with an ideal isolated kagome lattice is therefore a fundamentally important quest. In recent years a few families of metallic materials possessing a kagome lattice have indeed been reported or theoretically predicted. These include the Herbertsmithite related material Ga/ScCu3(OH)6Cl2 Mazin2014 , the magnetic kagome metals CoSn and FeSn Mingu2020 ; Ming2020 , and the ferromagnetic kagome metal YMn6Sn6 Li2021 . Most recently, the $A$V3Sb5 ($A=~{}$K, Rb, Cs) family of materials have been discovered and shown to host a perfect kagome network of V ions Ortiz2019 ; Neupert2022 . Evidence for electron correlations and non trivial topology in these materials emerges from the discovery of charge density waves, superconductivity, anomalous Hall effect, and multiple van Hove singularities nearby the Fermi energy Ortiz2019 ; Ortiz2020 ; Ortiz2021 ; Yang2020 . Another family of materials possessing the kagome lattice is $R$T${}_{3}X_{2}$ ($R=$ Lanthanide, $T=$ $4$d or $5$d transition metal, $X=$ Si, B). These materials were discovered in the 1980s Ku1980 ; Barz ; Vandenberg and several of them were reported to show superconductivity with $T_{\rm c}$s between $1$ K to $\sim 7$ K Ku1980 ; Barz ; Vandenberg ; Malik ; Athreya ; Rauchschwalbe . However, most of these studies were not made in the context of the connection of properties with the underlying kagome lattice. Only recently LaRu3Si2, which has the highest $T_{\rm c}=7$ K in this family of materials, has been studied in relation to the kagome lattice, and several unconventional properties have been reported possibly arising from electron correlations from the flat bands Li2011 ; Li2012 ; Li2016 ; Mielke2021 . Materials in the $R$T${}_{3}X_{2}$ family thus form another promising platform to study the kagome related features in the band structure, and their interplay with superconductivity. We report on the electronic structure, phonon profile, and superconducting properties of LaRh3B2 which has previously been reported to show superconductivity at low temperatures, where the reported superconducting $T_{\rm c}$ ranges from $<1.2$ to $2.8$ K Ku1980 ; Malik . Our electronic band structure calculations reveal a flat band above the Fermi energy, and van Hove singularities and Dirac cones at several locations in the Brillouin zone including close to the Fermi energy $E_{\rm F}$. We find that the $E_{\rm F}$ is located at the top of a sharp peak in the density of states (DOS). We use this to address the extreme sample dependence of the superconducting $T_{\rm c}$. The superconductivity is found to be of conventional weak coupling type. This is supported by estimations of the $T_{c}$ from phonon calculations and the estimate of electron-phonon coupling. The van Hove singularities in the band structure are found to be located a few eVs away from $E_{\rm F}$, which is large against the characteristic ordering scales and thus explains why these materials do not show signals of correlation-induced phenomena such as charge density waves or other instabilities. This is also supported by the phonon calculations which stress the absence of any imaginary frequency mode. In addition, we observe anomalous temperature dependencies of the magnetic susceptibility and heat capacity, and a slightly enhanced Sommerfeld coefficient, which we argue to arise from the narrow band which is part of the DOS. In comparison with the $A$V3Sb5 materials, our results point to a reduced importance of electronic-correlations in the LaRh3B2 material. ## II Methods Polycrystalline samples of LaRh3B2 were synthesized by arc-melting stoichiometric ratios of La (3N, Alfa Aesar), Rh (5N, Alfa Aesar) and B (6N, Alfa Aesar). The melted buttons were flipped over and melted $5$–$10$ times to promote homogeneity. Powder X-ray diffraction (PXRD) on a Bruker D8 Advance diffractometer system with Cu-K$\alpha$ radiation was used to determine the phase purity of the arc-melted LaRh3B2 sample. The relative stoichiometry of La and Rh was confirmed using energy dispersive spectroscopy using a scanning electron microscope. The dc magnetic susceptibility $\chi$, heat capacity $C$, and electrical transport were measured using a Quantum Design Physical Property Measurement System equipped with a He3 insert. To theoretically simulate the electronic structure of LaRh3B2, we performed first-principles density functional theory (DFT) calculations using the Vienna Ab initio simulation package (VASP) Kresse93 ; Kresse94 ; Kresse96a ; Kresse96b . We considered the projector-augmented wave (PAW) pseudo potential with exchange- correlation functional of generalized gradient approximation (GGA) of Perdew- Burke-Ernzerhof Kresse ; Perdew . Starting with the experimental structure, the lattice relaxation was performed to optimize the crystal structure by using variable cell relaxation. We adopted a $12\times 12\times 12$ k mesh for the first Brillouin zone. We have used an energy cut-off of $450$ eV for the plane wave basis. The convergence criteria for energy and force are set to $10^{-6}$ eV and $0.02$ eV/Å, respectively. In the DFT calculation, spin-orbit coupling was not included. However, we have used scalar relativistic potential which takes scalar relativistic effects into account. Phonon calculations have been performed using density functional perturbation theory, as implemented in Quantum Espresso Giannozzi2020 ; Giannozzi2009 ; Giannozzi . Exchange and correlation effects were included using the generalized gradient approximation (GGA) with the Perdew-Burke-Ernzerhof (PBE) functional Perdew ; the pseudopotentials are norm-conserving, with core correction, and scalar relativistic Hamann . Self-consistent calculations of the previously relaxed unit cell have been performed with a 8$\times$8$\times$12 $k$-grid. The kinetic energy cutoff for the wavefunctions is equal to 100 Ry, while the cutoff for charge density is 400 Ry. Convergence threshold for ionic minimization and electronic self-consistency are set to be 1.0D-15. The self-consistency threshold for phonon calculations is 1.0D-15 as well, with a $q$-grid of 4$\times$4$\times$2\. Non-self consistent calculations for the density of states have been performed with a 60$\times$60$\times$48 $k$-grid. Finally, the electron-phonon interaction is computed via an interpolation over the Brillouin Zone Wierzbowska . Figure 1: (Color online) Powder x-ray diffraction and results of refinement. Figure 2: (Color online) (a) A schematic of the crystal structure of LaRh3B2 viewed perpendicular to the crystallographic $c$-axis showing the layered nature of the structure with Rh atomic planes separated along the $c$-axis by planes made up of La and B atoms. (b) Viewed along the $c$-axis, the Rh atoms form an undistorted kagome lattice. ## III Structure LaRh3B2 crystallizes in a honeycomb structure with space group $P6/mmm$. There are no variable parameters in the structure apart from the unit cell size. The powder x-ray diffraction (PXRD) is shown in Fig. 1. The PXRD pattern confirmed that the synthesized material is single phase and a refinement, shown in Fig. 1, of the powder pattern gave lattice parameters $a=5.486$Å and $c=3.136$Å. In literature, a range of values for the lattice parameters have been reported and our values fall within this range of values Ku1980 . We will make a connection of unit cell parameters with the electronic properties later. A schematic of the crystal structure of LaRh3B2 is shown in Fig. 2. The structure is made up of layers of Rh planes separated by planes of La and B stacked along the $c$-axis, as shown in Fig. 2(a). The arrangement of the Rh atoms within the Rh-planes is a perfect kagome lattice as shown in Fig. 2(b). These materials therefore have the structural ingredients to show electronic structure features expected for a kagome metal. It must be noted however, that the short $c$-axis necessarily means that coupling between kagome planes may be significant. ## IV Results ### IV.1 Electronic Band Structure Figure 3: (Color online) (a) The electronic structure of LaRh3B2 along high symmetry directions in $k$-space. The boxes are to highlight some interesting features as discussed in the main text (b) The total and partial density of states as a function of energy measured from the Fermi energy. Figure 3(a) shows the electronic band structure for LaRh3B2 along some high symmetry directions in the Brillouin zone. It is evident that several bands cross the Fermi level, confirming that LaRh3B2 is a metal. The total and partial density of states (DOS) are shown in Fig. 3(b). The Fermi level ($E_{F}$) is situated near the top of a very narrow band resulting in a fairly large DOS at $E_{F}$ of $6.6$ states/eV. From the partial DOS it is clear that the majority contribution to the total DOS comes from Rh $4$d orbitals and both La and B contribute very small amounts to the total DOS at $E_{F}$. The narrow band at $E_{F}$ leads to a strong sensitivity of the superconducting $T_{c}$ on the unit cell size and to other anomalous physical properties as we will discuss later. We now turn to the novel features of the band structure arising from the kagome Rh planes. As can be seen in Fig. 3(a), we observe a flat band (FB) in the $\Gamma-M-K-\Gamma$ direction about $0.4$ eV above $E_{F}$. This flat band is separate from any other bands. Another series of disconnected flat bands are observed along the $\Gamma-A$ direction about $0.75$ eV above $E_{F}$. In addition to these flat portions of the electronic dispersion, Dirac cones (DC) are observed at several locations in the band structure. There are Dirac bands $140$ meV below and $2.75$ eV above $E_{F}$ at $H$ and a Dirac cone about $1$ eV below $E_{F}$ along the $M-K$ direction in the BZ. We also identify van Hove (VH) singularities located symmetrically above and below the Dirac cone at 2.75 eV. Thus the band structure of LaRh3B2 possesses the predicted features of the kagome lattice band structure near $E_{F}$ with modifications arising most likely from the three-dimensional nature of the material. ### IV.2 Physical Properties Figure 4 shows the electrical, magnetic, and thermal properties of LaRh3B2 in the normal and superconducting states. Figure 4 (a) shows the magnetic susceptibility $\chi$ versus temperature $T$ between $2$ K and $300$ K in an applied magnetic field of $H=2$ T. At low temperatures, small amounts of magnetic impurities lead to a Curie like upturn. The $\chi$ is found to be temperature dependent in the whole temperature range. This is not what is expected for a Pauli paramagnetic metal where a $T$ independent $\chi$ is expected. This $T$ dependent $\chi$ arises due to the $E_{F}$ being situated on a narrow peak in the DOS. The change in temperature results in a change in the DOS at $E_{F}$ leading to a $T$ dependent Pauli paramagnetic susceptibility. To support this idea the $\chi(T)$ in the full temperature range was fit with the expression $\chi(T)=\chi_{o}[1-(T/T_{E})^{2}]+C/(T-\theta)$, where the first term represents the $T$ dependent Pauli paramagnetic susceptibility and the second term represents the contribution from the small amounts of magnetic impurities which give rise to the Curie like upturn in $\chi(T)$ at the lowest temperatures. The fitting parameters are $\chi_{o}$ the temperature independent average Pauli paramagnetic susceptibility, $T_{E}$ which is a phenomenological parameter related to the Fermi energy, $C$ which is the Curie constant of the impurities, and $\theta$ which is the Weiss temperature representing any interactions between the magnetic impurities. A very good fit with the above expression was obtained and is shown as the solid curve through the data in Fig. 4 (a). The fit parameters obtained were $\chi_{o}=11.8(2)\times 10^{-5}$ G cm3/mol, $T_{E}=860(7)$ K, $C=0.0010(4)$ G cm3 K/mol, and $\theta=-5.5(1)$ K. This value of $C$ is equivalent to $0.25\%$ of $S=1/2$ impurities, which is quite small. Figure 4: (Color online) (a) The normal state magnetic susceptibility at $H=2$ T, (b) Resistivity versus temperature at zero field. Inset shows the superconducting transition at various fields, (c) Dimensionless magnetic susceptibility ($4\pi\chi$) in the superconducting state at various fields, and (d) the electronic contribution to the zero field specific heat $C_{\rm el}/T$ versus $T$. Figure 4 (b) shows the electrical resistivity $\rho$ in zero field between $2$ K and $300$ K. We observe metallic behaviour with a residual resistivity ratio RRR $=\rho(300{\rm K})/\rho(2{\rm K})\approx 5$. The inset in Fig. 4 (b) shows the $\rho(T)$ data below $T=4$ K measured in various applied fields. The sharp drop to zero resistance below $T_{c}\sim 2.6$ K in zero field signals the onset of superconductivity in LaRh3B2. Further evidence of a superconducting state is obtained from the diamagnetism observed in magnetic measurements shown in Fig. 4 (c) and (d). From Fig. 4 (c) we can see that the value of $4\pi\chi$ in the superconducting state is greater than $-1$ suggesting demagnetization factors due to the irregular shape of the sample. The magnetization data in Fig. 4 (d) show behaviour typical of a Type-II superconductor. The bulk nature of superconductivity is confirmed from heat capacity measurements. Figure 4 (e) shows the electronic heat capacity $C_{\rm el}$ divided by $T$ versus $T$. The $C_{\rm el}$ is obtained by subtracting a lattice term ($\sim T^{3}$) from the total heat capacity. It is of interest to note that the $C_{\rm el}/T$ in the normal state would be expected to be $T$ independent. However, this is true only for $T\leq 4.5$ K while there is a strong $T$ dependence of $C_{\rm el}/T$ above these temperatures. This suggests that the lattice contribution may not just have a $T^{3}$ term but an-harmonic terms may also contribute to the lattice heat capacity. A sharp anomaly at the onset of the superconducting transition can clearly be seen in Fig. 4 (e). To evaluate the magnitude of the jump in heat capacity at the transition and to obtain an alternate estimate of the bulk superconducting $T_{c}$, we use an equal entropy construction. The result is shown as the solid curve through the data near $T_{c}$. This gives a value $T_{c}\approx 2.5$ K. The normal state data above $T_{c}$ can be extrapolated to $T=0$ to give an estimate of the Sommerfeld coefficient $\gamma_{\rm n}=11.8$ mJ/mol K2. With this we obtain the jump height at $T_{c}$ to be $\Delta C_{\rm el}/\gamma_{\rm n}T_{c}\approx 1.2$, which is smaller than the value $1.43$ expected for a weak coupling single gap superconductor. The $C_{\rm el}$ data at the lowest temperatures seem to extrapolate to a finite value suggesting some residual contribution from normal electrons. The $C_{\rm el}$ data below $T=0.8$ K were fit by the expression $C_{\rm el}=\gamma_{\rm res}T+$ Aexp${}^{-\Delta/k_{B}T}$ assuming a fully gapped superconducting state. The $\gamma_{\rm res}T$ term represents the contribution from any non- superconducting fraction of electrons. An excellent fit, shown in Fig. 4 (e), was obtained with the following values for the parameters $\gamma_{\rm res}=1.7$ mJ/mol K2 and $\Delta=6$ K. This value of $\gamma_{\rm res}$ suggests that $\approx 14\%$ electrons do not participate in the superconductivity. So we must revise our estimate of $\Delta C_{\rm el}/\gamma T_{c}$ using $\gamma=\gamma_{n}-\gamma_{\rm res}$. This gives the value $\Delta C_{\rm el}/\gamma T_{c}\approx 1.44$ which is close to the value expected from a weak coupling single gap superconductor. These results suggest that LaRh3B2 is a weak coupling single-gap Type-II superconductor. From various measurements in finite magnetic field we can track the $T_{c}$ as a function of the field $H$. The $H$-$T$ diagram obtained from the various measurements is shown in Fig. 4 (f) where both the lower critical field $H_{C1}$ and the upper critical field $H_{C2}$ are shown. We observe that the $H_{C2}$ data from all measurements except the resistivity measurements agree with each other, while the critical field measured from $\rho$ are consistently higher than values measured from other bulk probes like magnetization and heat capacity. Such observations have been reported previously in some materials and have been linked to surface superconductivity Zeinali ; Pradip . It has been shown that the critical field for surface superconductivity is $\approx 1.69H_{c2}$, where $H_{c2}$ is the bulk critical field. We also plot in Fig. 4 (f) the $H_{C2}$ obtained from resistivity measurements divided by $1.69$. The critical field so obtained matches the critical field values obtained from other bulk measurements. So we will treat the scaled critical field from resistivity measurements as the true bulk critical field $H_{c2}$. The $H_{c2}$ vs $T$ plot shows an upward curvature in the whole temperature range. This is unusual and inconsistent with observations for most conventional superconductors. To learn about the strength of the electron-phonon coupling we make an estimate of the electron-phonon coupling constant $\lambda_{ep}$ using McMillan’s formula, which relates the superconducting transition temperature $T_{c}$ to $\lambda_{ep}$, the Debye temperature $\theta_{D}$, and the Coulomb repulsion constant $\mu^{*}$ $T_{c}=\frac{\Theta_{D}}{1.45}\exp\left[-\frac{1.04\left(1+\lambda_{\mathrm{ep}}\right)}{\lambda_{\mathrm{ep}}-\mu^{*}\left(1+0.62\lambda_{\mathrm{ep}}\right)}\right]$ which can be inverted to give $\lambda_{ep}$ in terms of $T_{c}$, $\theta_{D}$ and $\mu^{*}$ as $\lambda_{\mathrm{ep}}=\frac{1.04+\mu^{*}\ln\left(\frac{\Theta_{D}}{1.45T_{c}}\right)}{\left(1-0.62\mu^{*}\right)\ln\left(\frac{\Theta_{D}}{1.45T_{c}}\right)-1.04}$ From the heat capacity measurements, we had obtained $\theta_{D}$ = 518 K and using $T_{c}$ = 2.5 K, we get $\lambda_{ep}=0.43$ and 0.52 for $\mu^{*}=0.10$ and 0.15, respectively. These values of $\lambda_{ep}$ suggest moderate electron-phonon coupling in LaRh3B2. This is supported by the estimates of the $\lambda_{ep}$ made from our phonon calculations which will be discussed later. We now present our estimation of various superconducting parameters using expressions previously collected in Refs. Singh2007, ; Singh2010, . An estimate of the $T=0$ upper critical field $H_{c2}(0)$ was made by first making a linear extrapolation of the data near $T_{c}$ to give the slope ${dH_{c2}\over dT}|_{T_{c}}=-511$ Oe/K. This linear slope can then be used to get an estimate of $H_{c2}(0)$ using the Werthamer-Helfand-Hohenberg (WHH) formula for the clean limit $H_{c2}(0)=-0.693T_{c}{dH_{c2}\over dT}|_{T_{c}}=920$Oe. From the value of $H_{c2}$ we can now estimate the value of the coherence length $\xi$ by the expression $H_{c2}=\phi_{0}/2\pi\xi^{2}$, where $\phi_{0}=hc/2e=2.068\times 10^{-7}$ G cm2 is the flux quantum. Using $H_{c2}(0)=920$ Oe obtained above, we estimate $\xi(0)=60$ nm. At $T=2.3$ K near $T_{c}$ where $H_{c2}=250$ Oe we get $\xi(0)=114$ nm. We have collected the various normal and superconducting state parameters in Table 1. Table 1: Normal and superconducting state parameters for LaRh3B2. Here $\gamma$ is the Sommerfeld coefficient, $\beta$ is the coefficient of the $T^{3}$ term in the low temperature heat capacity, $\theta_{\rm D}$ is the Debye temperature, $n$ is the electron density, $\xi$ is the superconducting coherence length, $\lambda$ is the penetration depth, $l$ is the electron mean free path, and $v_{\rm F}$ is the Fermi velocity. RRR | $\approx 5$ ---|--- $\gamma$ (mJ/mol K2) | 11.8 $\beta$ (mJ/mol K4) | 0.084 $\theta_{\rm D}$ (K) | 520 $n$ (cm${}^{-3})$ | $0.97\times 10^{23}$ TC(K) | 2.6 $\xi_{2{\rm K}}$(nm) | 114 $\xi_{0{\rm K}}$(nm) | 60 $\lambda_{0{\rm K}}$(nm) | 5.4 $l_{4{\rm K}}$(nm) | 43 $v_{{\rm F}}$(cm/s) | $3.5\times 10^{8}$ We now address the large variation in the $T_{C}$ of LaRh3B2 which has been reported in the literature Ku1980 ; Malik . Figure 5 shows the resistivity of two samples of LaRh3B2 prepared with the same nominal ratios of starting materials. A refinement of their powder x-ray pattern gave lattice parameters that are slightly different. The lattice parameters for sample 1 (S1) are $a=5.484$ Å and $c=3.139$ Å while those for sample 2 (S2) are $a=5.486$ Å and $c=3.136$ Å. Thus the S1 has slightly smaller in-plane lattice parameters while its $c$-axis is longer indicating that in this sample the kagome planes are shrunk while the separation between the kagome lattice increases. S2 on the other hand has a larger kagome plane but the planes are separated by a smaller distance along the $c$-axis. From Fig. 5 we see that the electrical transport properties are sensitive to these small changes. S1 has a larger residual resistivity ratio RRR $=10$ but has a larger residual resistivity $\rho_{0}=26~{}\mu\Omega$cm. S2 on the other hand has a smaller RRR $=5$ but a smaller residual resistivity $\rho_{0}=15~{}\mu\Omega$cm. From the inset it can be seen that while S2 has a superconductivity onset at $T_{c}=2.6$ K, for S1 the onset is $T_{c}=2.05$ K, more than $0.5$ K lower. We also find this large variation in $T_{\rm c}$ with unit cell size of LaRh3B2 in previous reports Ku1980 ; Malik . For example, superconductivity with a $T_{\rm c}=2.6$ K was reported for a LaRh3B2 sample with lattice parameters $a=5.480$Å and $c=3.137$ Å Ku1980 , a $T_{\rm c}=2.2$ K was reported for a sample with lattice parameters $a=5.483$Å, $c=3.142$ Å while no superconductivity down to $1.2$ K was found for a sample with lattice parameters $a=5.512$Å and $c=3.115$ Å Malik . We can address this variation in $T_{\rm c}$ for different samples using our DFT calculations. Our calculations have shown that the $E_{\rm F}$ lies near the top of a narrow band in the DOS. This sensitivity of $T_{\rm c}$ most likely originates from changes in the DOS at $E_{F}$ due to small changes in $E_{F}$ either by pressure effects (as evidenced by difference in unit cell sizes) or due to a difference in the electron densities (by minute variation in the stoichiometry) in the samples. A slight change in $E_{F}$ will lead to large changes in the DOS at $E_{F}$ because the $E_{F}$ is located on top of a narrow band in the band structure. Figure 5: (Color online) Resistivity versus temperature at zero field for two LaRh3B2 samples. Inset shows the variation in the superconducting transition temperature. Our phonon calculations support the motif of weak-coupling phonon-mediated superconductivity, and also indicate the reason for the absence of a correlation-induced CDW instability for LaRh3B2. The phonon dispersion for LaRh3B2 are shown in Figs. 6 a, b. The phonon dispersion does not exhibit any imaginary frequency mode; this is a sign of dynamical stability, confirming the absence of experimental signatures of charge density wave states. Low- frequency phonon modes mostly originate from La atoms, while intermediate frequencies are essentially due to Rh atoms; the manifold of low-dispersing bands around 120 cm-1 is then due to the kagome network. Finally, B atoms contribute to the high-frequency modes. We computed the electron-phonon coupling to be $\lambda_{\mathrm{e-ph}}\approx 0.55-0.65$. This suggests that LaRh3B2 is a weak-to-moderate coupling superconductor. The McMillan formula was then used to estimate the superconducting critical temperature $T_{c}$ McMillan ; Carbotte : $T_{c}=\frac{\omega_{log}}{1.2}e^{\bigl{[}\frac{-1.04(1+\lambda)}{\lambda(1-0.62\mu^{*})-\mu^{*}}\bigr{]}}$ (1) with $\omega_{log}$ being related to the Eliashberg function: $\omega_{log}=e^{\bigl{[}\frac{2}{\lambda}\int{\frac{d\omega}{\omega}\alpha^{2}F(\omega)\log{\omega}}\bigr{]}}$ (2) While the Coulomb pseudo-potential $\mu^{*}$ lies in the typical range [0.1 - 0.2], we obtain values for $T_{c}$ in fair agreement with the experimental results. Specifically, $T_{c}\approx 2.6$ K is obtained for $\mu^{*}=0.17$ (Fig. 6 d). Figure 6: a) Phonon dispersion along high-symmetry lines; b) Corresponding density of states; c) Eliashberg function; d) Superconducting critical temperature $T_{c}$ as a function of the Coulomb pseudo-potential $\mu^{*}$. ## V Summary and Discussion: We report on the electronic structure, phonon spectrum and physical properties of a kagome lattice superconductor LaRh3B2. The structure of LaRh3B2 is built up of kagome planes of Rh stacked along the $c$-axis with La-B planes separating the kagome planes. The electronic structure contains all features expected for a 2D kagome lattice including a flat band and Dirac bands and van Hove singularities at various positions in the Brillouin zone and in particular near $E_{F}$. This is qualitatively consistent with the band- structure observed for the $A$V3Sb5 kagome metals. In contrast with the $A$V3Sb5 materials, however, we did not observe signatures of strong electronic correlations in LaRh3B2. The van Hove singularities in the electronic band structure are situated further away from $E_{F}$ than for $A$V3Sb5, which is a fermiological reason why one would not expect charge ordering or density wave like instabilities as reported for $A$V3Sb5. The superconductivity in LaRh3B2 seems conventional, and there is no experimental evidence for charge density wave instabilities as reported for the $A$V3Sb5 materials. The majority contribution to the DOS at $E_{F}$ in LaRh3B2 derives from Rh $4$d bands. This suggests that the role of electronic correlations is weakened in LaRh3B2 compared to the family of $A$V3Sb5 kagome metals, since the Rh $4d$ orbitals are less strongly-correlated than the V $3d$ orbitals. Interestingly, the computed $\lambda_{\mathrm{e-ph}}$ for LaRh3B2 is in good agreement with experimentally reported values of $\lambda_{\mathrm{e-ph}}$ for the CsV3Sb5 compound Zhong which, together with the phonon dispersions, suggests a similarity in the principal phonon sector between LaRh3B2 and $A$V3Sb5. It suggests that a central difference between LaRh3B2 and the by now more established kagome metals must be found in terms of differing electronic correlations, which are lower in strength for LaRh3B2. This may explain the absence of CDW in the LaRh3B2 kagome metal, and points towards phonon-mediated $s$-wave superconductivity. Given the large number of materials in the $RT_{3}$B2 and $RT_{3}$Si2 families, their possibility of Fermiology-tuning around $E_{F}$ presents an exciting direction for future work. _Acknowledgments.–_ We thank the X-ray facility at IISER Mohali. JS acknowledges UGC-CSIR India for a fellowship. The phonon-DFT work was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) through Project-ID 258499086-SFB 1170 and by the Würzburg-Dresden Cluster of Excellence on Complexity and Topology in Quantum Matter-ct.qmat Project-ID 390858490-EXC 2147. The research leading to these results has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie Grant Agreement No. 897276. The authors acknowledge the Gauss Centre for Supercomputing e.V. for providing computing time on the GCS Supercomputer SuperMUC-NG at Leibniz Supercomputing Centre. ## References * (1) M. Fu, T. Imai, T.-H. Han, and Y. S. Lee, Science 350, 655 (2015). * (2) T.-H. Han, J. S. Helton, S. Chu, D. G. Nocera, J. A. Rodriguez-Rivera, C. Broholm, and Y. S. Lee, Nature (London) 492, 406 (2012). * (3) C. Balz, B. Lake, J. Reuther, H. Luetkens, R. Schönemann, T. Herrmannsdörfer, Y. Singh, A. T. M. Nazmul Islam, E. M. Wheeler, J. A. Rodriguez-Rivera, T. Guidi, G. G. Simeoni, C. Baines, and H. Ryll, Nature Physics 12, 942 (2015). * (4) Y. Okamoto, M. Nohara, H. Aruga-Katori, and H. Takagi, Phys. Rev. Lett. 99, 137207 (2007). * (5) Y. Singh, Y. Tokiwa, J. Dong, and P. Gegenwart, Phys. Rev. B 88, 220413(R) (2013). * (6) M. Kiesel, C. Platt, and R. Thomale, Phys. Rev. Lett. 110, 126405 (2013). * (7) W. Wang, Z. Li, Y. Xiang, and Q. Wang, Phys. Rev. B 87, 115135 (2013). * (8) I. I. Mazin, H. O. Jeschke, F. Lechermann, H. Lee, M. Fink, R. Thomale, and R. Valenti, Nature Communications 5, 4261 (2014). * (9) M. Kang, S. Fang, L. Ye, H. C. Po, J. Denlinger, C. Jozwiak, A. Bostwick, E. Rotenberg, E. Kaxiras, J. G. Checkelsky, and R. Comin, Nat. Comm. 11, 4004 (2020). * (10) M. Kang, et al. Nat. Mater. 19, 163 (2020). * (11) M. Li, Q. Wang, G. Wang, Z. Yuan, W. Song, R. Lou, Z. Liu, Y. Huang, Z. Liu, H. Lei, Z. Yin, and S. Wang, Nat. Comm. 12, 3129 (2021). * (12) B. R. Ortiz, L. C. Gomes, J. R. Morey, M. Winiarski, M. Bordelon, J. S. Mangum, I. W. H. Oswald, J. A. RodriguezRivera, J. R. Neilson, S. D. Wilson, E. Ertekin, T. M. McQueen, and E. S. Toberer, Phys. Rev. Materials 3, 094407 (2019). * (13) T. Neupert, M. Denner, J. Yin, R. Thomale, and M. Z. Hasan, Nat. Phys. 18 137 (2022). * (14) B. R. Ortiz, S. M. L. Teicher, Y. Hu, J. L. Zuo, P. M. Sarte, E. C. Schueller, A. M. M. Abeykoon, M. J. Krogstad, S. Rosenkranz, R. Osborn, R. Seshadri, L. Balents, J. He, and S. D. Wilson, Phys. Rev. Lett. 125, 247002 (2020). * (15) B. R. Ortiz, P. M. Sarte, E. M. Kenney, M. J. Graf, S. M. L. Teicher , R. Seshadri, and S. D. Wilson, Phys. Rev. Materials 5, 034801 (2021). * (16) S.-Y. Yang, Y. Wang, B. R. Ortiz, D. Liu, J. Gayles, E. Derunova, R. Gonzalez-Hernandez, L. Smejkal, Y. Chen, S. S. Parkin, S. D. Wilson, E. S. Toberer, T. Mcqueen, and M. N. Ali , Sci. Adv. 6, eabb6003 (2020). * (17) H. C. Ku, C. P. Meisner, F. Acker, and D. C. Johnston, Solid State Communications, 35, 91 (1980). * (18) H. Barz, Mater. Res. Bull. PNAS 15, 1489 (1980). * (19) J. M. Vandenberg and H. Barz, Mater. Res. Bull. 15, 1493 (1980). * (20) S. K. Malik, A. M. Umarji, G. K. Shenoy, A. T. Aldred, and D. G. Niarchos, Phys. Rev. B 32, 4742 (1985). * (21) K. S. Athreya, L. S. Hausermann-Berg, R. N. Shelton, S. K. Malik, A. M. Umarji, and G. K. Shenoy, Phys. Lett. A 113, 330 (1985). * (22) U. Rauchschwalbe, W. Lieke, F. Steglich, C. Godart, L. C. Gupta, and R. D. Parks, Phys. Rev. B 30, 444(R) (1984). * (23) S. Li, B. Zeng, X. G. Wan, J. Tao, F. Han, H. Yang, Z. H. Wang, and H.-H. Wen, Phys. Rev. B 84, 214527 (2011). * (24) S. Li, J. Tao, X. G. Wan, X. Ding, H. Yang, and H.-H. Wen Phys. Rev. B 86, 024513 (2012). * (25) B. Li, S. Li, and H.-H. Wen, Phys. Rev. B 94, 094523 (2016) * (26) C. Mielke, Y. Qin, J.-X. Yin, H. Nakamura, D. Das, K. Guo, R. Khasanov, J. Chang, Z. Q. Wang, S. Jia, S. Nakatsuji, A. Amato, H. Luetkens, G. Xu, M. Z. Hasan, and Z. Guguchia, Phys. Rev. Mat. 5, 034803 (2021). * (27) G. Kresse and J. Hafner, Phys. Rev. B 47, 558(R) (1993). * (28) G. Kresse and J. Hafner, Phys. Rev. B 49, 14251 (1994). * (29) G. Kresse and J. Furthmüller, Phys. Rev. B 54, 11169 (1996). * (30) G. Kresse and J. Furthmüller, Comput. Mater. Sci. 6, 15 (1996). * (31) G. Kresse and D. Joubert, Phys. Rev. B 59, 1758 (1999). * (32) J. P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865–3868 (1996). * (33) P. Giannozzi, O. Baseggio, P. Bonfà, D. Brunato, R. Car, I. Carnimeo, C. Cavazzoni, S. de Gironcoli, P. Delugas, F. F. Ruffino, A. Ferretti, N. Marzari, I. Timrov, A. Urru, and S. Baroni, The Journal of Chemical Physics 152, 154105 (2020), * (34) P. Giannozzi, S. Baroni, N. Bonini, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, G. L. Chiarotti, M. Cococcioni, I. Dabo, A. D. Corso, S. de Gironcoli, S. Fabris, G. Fratesi, R. Gebauer, U. Gerstmann, C. Gougoussis, A. Kokalj, M. Lazzeri, L. Martin-Samos, N. Marzari, F. Mauri, R. Mazzarello, S. Paolini, A. Pasquarello, L. Paulatto, C. Sbraccia, S. Scandolo, G. Sclauzero, A. P. Seitsonen, A. Smogunov, P. Umari, and R. M. Wentzcovitch, Journal of Physics: Condensed Matter 21, 395502 (2009). * (35) P. Giannozzi, O. Andreussi, T. Brumme, O. Bunau, M. B. Nardelli, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, M. Cococcioni, N. Colonna, I. Carnimeo, A. D. Corso, S. de Gironcoli, P. Delugas, R. A. DiStasio, A. Ferretti, A. Floris, G. Fratesi, G. Fugallo, R. Gebauer, U. Gerstmann, F. Giustino, T. Gorni, J. Jia, M. Kawamura, H-Y. Ko, A. Kokalj, E. Küçükbenli, M. Lazzeri, M. Marsili, N. Marzari, F. Mauri, N. L. Nguyen, H-V. Nguyen, A. O. de-la Roza, L. Paulatto, S. Poncé, D. Rocca, R. Sabatini, B. Santra, M. Schlipf, A. P. Seitsonen, A. Smogunov, I. Timrov, T. Thonhauser, P. Umari, N. Vast, X. Wu, and S. Baroni, Journal of Physics: Condensed Matter 29, 465901 (2017). * (36) D. R. Hamann, Phys. Rev. B 88, 085117 (2013). * (37) M. Wierzbowska, S. de Gironcoli, and P. Giannozzi, “Origins of low- and high-pressure discontinuities of tc in niobium,” arXiv:cond-mat/0504077 (2005). * (38) A. Zeinali, T. Golod, and V. M. Krasnov, Phys. Rev. B 94, 214506 (2016). * (39) Pradip Das, C. V. Tomy, S. S. Banerjee, H. Takeya, S. Ramakrishnan, and A. K. Grover, Phys. Rev. B 78, 214504 (2008). * (40) Yogesh Singh, A. Niazi, M. D. Vannette, R. Prozorov, and D. C. Johnston, Phys. Rev. B 76, 214510 (2007). * (41) Yogesh Singh, C. Martin, S. L. Bud’ko, A. Ellern, R. Prozorov, and D. C. Johnston, Phys. Rev. B 82, 144532 (2010). * (42) W. L. McMillan, Phys. Rev. 167, 331 (1968). * (43) J. P. Carbotte, Rev. Mod. Phys. 62, 1027 (1990). * (44) Y. Zhong, S. Li, H. Liu, Y. Dong, Y. Arai, H. Li, Y. Shi, Z. Wang, S. Shin, H. N. Lee, H. Miao, T. Kondo, and K. Okazaki, arXiv:2207.02407 (2022).
# Pretraining the Noisy Channel Model for Task-Oriented Dialogue Qi Liu2, Lei Yu1, Laura Rimell1, and Phil Blunsom1,2 1DeepMind, 2University of Oxford <EMAIL_ADDRESS> <EMAIL_ADDRESS> Work completed during an internship at DeepMind. ###### Abstract Direct decoding for task-oriented dialogue is known to suffer from the explaining-away effect, manifested in models that prefer short and generic responses. Here we argue for the use of Bayes’ theorem to factorize the dialogue task into two models, the distribution of the context given the response, and the prior for the response itself. This approach, an instantiation of the noisy channel model, both mitigates the explaining-away effect and allows the principled incorporation of large pretrained models for the response prior. We present extensive experiments showing that a noisy channel model decodes better responses compared to direct decoding and that a two stage pretraining strategy, employing both open-domain and task-oriented dialogue data, improves over randomly initialized models. ## 1 Introduction Task-oriented dialogue agents provide a conversational interface to assist users in accomplishing specific goals, such as finding a restaurant or booking a hotel Seneff and Polifroni (2000); Raux et al. (2005); Budzianowski et al. (2018); Peng et al. (2020a). Increasing demand from industry for natural language assistants and scalable customer service solutions has recently been driving a renaissance in the development of task-oriented dialogue models. In addition, the specification of explicit dialogue agent goals, afforded by the task-oriented paradigm, makes such research easier to ground and evaluate than open-domain chatbots. Current research on task-oriented dialogue is dominated by monolithic sequence-to-sequence models that directly parameterize the conditional distribution of the response given the prior dialogue context. However, this monolithic approach conflates the task-specific and language-general aspects of dialogue, and adversely favors short and generic responses Bao et al. (2020) due to the explaining-away effect Klein and Manning (2002). Here we pursue an alternative to the direct model. Employing Bayes’ rule allows us to factorize the probability of the response given the context $p(\mathcal{R}|\mathcal{C})$ into a language model $p(\mathcal{R})$ and a context model $p(\mathcal{C}|\mathcal{R})$.111Here we abstract away from the prediction of belief states and dialogue acts, which also form part of our generative model; see Section 3 for details. Within natural language processing (NLP), this approach is traditionally known as the noisy channel model Shannon (1948), and has recently seen renewed interest with its successful application to neural machine translation Yu et al. (2017, 2020); Yee et al. (2019). We hypothesize that the noisy channel reformulation is advantageous for dialogue because the factorization enables each sub-module to specialize in a dialogue sub-task. In particular, the context conditional model can help to discount short and generic responses and mitigate the explaining-away effect, while the language model helps ensure that responses are natural. We find that a noisy channel model with the same number of parameters as a direct model achieves better accuracy on three task-oriented dialogue datasets. Moreover, a larger noisy channel model can be trained with the same hardware, by training the sub-modules separately, yielding additional improvements. It has become common in recent years to pretrain dialogue models on large text data, either general text Peng et al. (2020b); Budzianowski and Vulić (2019); Wu et al. (2020a) or dialogue-structured data Roller et al. (2020); Adiwardana et al. (2020), such as tweets and Reddit posts. We utilise a similar strategy with Reddit data and find that the benefits of pretraining to the noisy channel model are similar to those for the direct model. Further, we evaluate transfer across task-oriented dialogue datasets by implementing a second pretraining stage using Taskmaster Byrne et al. (2019) and Schema-Guided Dialogue Rastogi et al. (2020) as training data, before fine-tuning on our final tasks. We evaluate the algorithm on three datasets, MultiWOZ 2.0 Budzianowski et al. (2018), CamRest676 Wen et al. (2017a) and SMCalFlow Andreas et al. (2020), demonstrating that the noisy channel approach is robust to different dialogue schema annotations used across datasets. Further analysis demonstrates that the noisy channel models can decode responses with similar lengths and Zipf scores compared to ground-truth responses and reduce the likelihood of falling into repetition loops Holtzman et al. (2019). ## 2 A Seq-to-Seq Dialogue Model Figure 1: The data flow of one turn in a task-oriented dialogue for train booking from MultiWOZ. In this section, we introduce a discriminative sequence-to-sequence model for task-oriented dialogue. The traditional sequence of steps needed to produce a system turn in a task-directed dialogue is shown in Figure 1, with an example from MultiWOZ 2.0 Budzianowski et al. (2018). Given a dialogue context containing previous user and system utterances, the dialogue system first predicts a belief state, consisting of a set of slot-value pairs (e.g. destination: Cambridge), to capture user intent. To ground the system with external information, the belief state can be converted into a database query in order to retrieve relevant information, such as the number of matches and booking information. Next, the system predicts a set of dialogue acts, representing the abstract meaning of the proposed dialogue response Austin (1975). Finally, a delexicalized dialogue response is generated, where slot values are replaced by generic placeholders, such as value_time for a train departure time, in order to reduce lexical variation. The delexicalized response can be converted to a lexicalized response in post-processing by filling in the slot values based on belief states and database information. We use the MultiWOZ schema for illustration in Section 2 and 3, but our models easily generalize to different schema annotations (e.g. datasets without annotated dialogue acts Andreas et al. (2020)). Since it is well known that pipelined models tend to suffer from error propagation, many NLP tasks have been reformulated in recent years as end-to- end text-to-text transformations Raffel et al. (2020); Brown et al. (2020). State-of-the-art task-oriented dialogue systems have followed this approach Hosseini-Asl et al. (2020); Peng et al. (2020b). We represent the example from Figure 1 as follows, serializing turns and using special start and end tokens to encapsulate each data field: Context: [c] I am looking to … [/u] What is your … [/r] I’ll be leaving … [/u] [/c] Belief: [b] [train] destination Cambridge, day Tuesday, arrive 12:30, departure London [/b] Database: [db] [train] match 1, status not booked [/db] Act: [a] [train] inform arrive, inform leave, offer reservation [/a] Response: [r] There is a train that leaves at [value_time] and arrives at [value_time]. Should I book it? [/r] Given this text representation, the direct discriminative approach models $p(\mathcal{B},\mathcal{A},\mathcal{R}|\mathcal{C})$, where $\mathcal{C}$, $\mathcal{B}$, $\mathcal{A}$, and $\mathcal{R}$ represent dialogue context, belief state, dialogue act, and delexicalized response, respectively.222We do not model the probabilities of database state or lexicalized response, as these are deterministic given the belief state and delexicalized response, respectively. We use the serialized text of the dialogue context as input, and the concatenation of belief state, dialogue act, and response as target output, making the task amenable to the application of an autoregressive sequence-to-sequence model. $\mathcal{B}$, $\mathcal{A}$ and $\mathcal{R}$ can be generated sequentially with direct decoding methods, such as greedy decoding and beam search. We use a sequence-to-sequence Transformer Vaswani et al. (2017) to implement $p(\mathcal{B},\mathcal{A},\mathcal{R}|\mathcal{C})$. This distribution will also be used to build the noisy channel model in Section 3. ## 3 Noisy Channel Model for Dialogue While direct decoding is an effective approach for decoding belief states Hosseini-Asl et al. (2020), it may be sub-optimal for generating responses. First, it favors short and generic responses Bao et al. (2020). As a result, the decoded responses are bland and lack diversity Li et al. (2016). Second, it suffers from the explaining-away effect Klein and Manning (2002), where inputs are “explained-away” by highly predictive output prefixes. For example, if there is one hotel matching the user’s intent as encoded in the belief state, the model is nevertheless prone to decoding “no” given the output prefix “there is”, ignoring the input information. In this work, we propose using the neural noisy channel model Yu et al. (2017) to mitigate the above problems for response generation. Given an input sequence $x$ and output sequence $y$, the noisy channel formulation Shannon (1948) uses Bayes’ rule to rewrite the model $p(y|x)$ as $\frac{p(x|y)p(y)}{p(x)}\ \propto\ p(x|y)p(y)$. It was originally applied to speech recognition, where $p(y|x)$ is a conditional model of the source text given a noisy observation. The channel model $p(x|y)$ estimates the probability of the observation given the source, while $p(y)$ is an unconditional language model (or source model), which can be trained on unpaired data. More recently it has been applied to machine translation, where $y$ is a translation of input text $x$. Abstracting away from belief states and dialogue acts, for task-oriented dialogue we want to estimate $p(\mathcal{R}|\mathcal{C})$, the probability of a response given a context. The channel model $p(\mathcal{C}|\mathcal{R})$, given a response, predicts a distribution over contexts which might have elicited that response. The source model $p(\mathcal{R})$ is an unconditional language model. In this extension of the noisy channel approach to task- oriented dialogue, the “channel” can be understood as connecting dialogue contexts with suitable responses. For the full task, we develop a noisy channel model for $p(\mathcal{B},\mathcal{A},\mathcal{R}|\mathcal{C})$. Using the chain rule, $p(\mathcal{B},\mathcal{A},\mathcal{R}|\mathcal{C})=p(\mathcal{B}|\mathcal{C})\cdot p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})$. Following Hosseini-Asl et al. Hosseini-Asl et al. (2020), we use the direct model described in Section 2 to parameterize $p(\mathcal{B}|\mathcal{C})$ and decode $\mathcal{B}$, which our preliminary experiments confirmed to be advantageous. We use the noisy channel formulation to parameterize $p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})$. Using Bayes’ Rule, $p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})\ \propto\ p(\mathcal{C},\mathcal{B}|\mathcal{A},\mathcal{R})\cdot p(\mathcal{A},\mathcal{R})$. The channel model $p(\mathcal{C},\mathcal{B}|\mathcal{A},\mathcal{R})$ and source model $p(\mathcal{A},\mathcal{R})$ are implemented as Transformers. We choose to use the noisy channel formulation for decoding $\mathcal{A}$ based on preliminary experiments which showed improved overall accuracy over direct decoding, possibly because poor dialogue act prediction by the direct model led to worse quality responses. The serialized text of $\mathcal{A}$ and $\mathcal{R}$ are concatenated during training, and the decoded sequence is split into $\mathcal{A}$ and $\mathcal{R}$ with the special start/end tokens during decoding. We suggest that the noisy channel model has three advantages over the direct model for response generation: (1) The channel model can penalize short and generic responses. Such responses can be mapped to a large number of contexts, resulting in a flat distribution over contexts. This leads to a lower channel model score for short and generic responses Zhang et al. (2020b). (2) The channel model ensures that $(\mathcal{A},\mathcal{R})$ must explain the corresponding $(\mathcal{C},\mathcal{B})$, alleviating the explaining-away effect Yu et al. (2017). (3) The source model, an unconditional distribution over $\mathcal{A}$ and $\mathcal{R}$, can make use of abundant non-dialogue textual data for pretraining, further improving the fluency of generated sequences Brants et al. (2007). We leave exploration of this last advantage for future work, as we pretrain all sub-modules with the same data. ### 3.1 Decoding Since exact decoding from the noisy channel model $\operatorname*{arg\,max}_{\mathcal{A},\mathcal{R}}p(\mathcal{C},\mathcal{B}|\mathcal{A},\mathcal{R})\cdot p(\mathcal{A},\mathcal{R})$333Although exact decoding is also computationally intractable for the direct model, approximating $\operatorname*{arg\,max}_{\mathcal{B}}p(\mathcal{B}|\mathcal{C})$ is well- studied, e.g. beam search. The decoding for $\mathcal{B}$ is therefore omitted here. is computationally intractable, we experiment with two approximation methods, noisy channel reranking and noisy channel online decoding. Since these methods rely on $p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})$ as a proposal distribution for approximation, and both $p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})$ and $p(\mathcal{B}|\mathcal{C})$ are parameterized with the direct model introduced in Section 2, our noisy channel model therefore has three sub- modules: a direct model $p(\mathcal{B},\mathcal{A},\mathcal{R}|\mathcal{C})$, a channel model $p(\mathcal{C},\mathcal{B}|\mathcal{A},\mathcal{R})$, and a source model $p(\mathcal{A},\mathcal{R})$. Input : Context $\mathcal{C}$ Output : Belief, act and response $(\mathcal{B},\mathcal{A},\mathcal{R})$ Decode $\mathcal{B}$ given $\mathcal{C}$ with $p(\mathcal{B}|\mathcal{C})$ Beam: $\mathcal{S}=\\{({\color[rgb]{0,0,0}[a]})\\}$ while _end($\mathcal{S}$) is False_ do $\mathcal{S^{\prime}}=\varnothing$ for _$\mathcal{O}$ in $\mathcal{S}$_ do if _$\mathcal{O}.\texttt{last}$() is [/r] or $\mathcal{|O|}$ $>l$ _ then $\mathcal{S^{\prime}}.\texttt{add}$($\mathcal{O}$) continue end if Get $k_{1}$ tokens $o^{1},...,o^{k_{1}}$ from the direct model $p(O_{|\mathcal{O}|+1}|\mathcal{C},\mathcal{B},\mathcal{O})$ for _$o^{i}$ in $(o^{1},...,o^{k_{1}})$_ do $\mathcal{S^{\prime}}.\texttt{add}$($(\mathcal{O},o^{i})$) end for end for $\mathcal{S}=\operatorname*{top\\_k_{2}}\limits_{\mathcal{O}\in\mathcal{S^{\prime}}}\log p(\mathcal{O}|\mathcal{C},\mathcal{B})+$ $\lambda_{1}\cdot\log p(\mathcal{C},\mathcal{B}|\mathcal{O})+$ $\lambda_{2}\cdot\log p(\mathcal{O})+$ $\lambda_{3}\cdot|\mathcal{O}|$ end while Select $\mathcal{O}\in\mathcal{S}$ with the largest score using Eq. 1 and return $(\mathcal{B},\mathcal{A},\mathcal{R})$ Algorithm 1 Online decoding for the noisy channel. Noisy channel reranking: Noisy channel reranking first decodes $\mathcal{B}$ and then continues decoding a list $\mathcal{S}$ of $(\mathcal{A},\mathcal{R})$ pairs by beam search with the direct model, prior to utilizing the noisy channel model to rerank $(\mathcal{A},\mathcal{R})$ pairs. In particular, during beam search, partial sequences are expanded and pruned with $p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})$ (from the direct model in Section 2). The pairs after decoding are reranked using the following model combination: $\displaystyle(\mathcal{A}^{\prime},\mathcal{R}^{\prime})=\operatorname*{arg\,max}\limits_{(\mathcal{A},\mathcal{R})\in\mathcal{S}}\log p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})$ $\displaystyle+$ (1) $\displaystyle\lambda_{1}\cdot\log p(\mathcal{C},\mathcal{B}|\mathcal{A},\mathcal{R})$ $\displaystyle+$ $\displaystyle\lambda_{2}\cdot\log p(\mathcal{A},\mathcal{R})$ $\displaystyle+$ $\displaystyle\lambda_{3}\cdot|\mathcal{A},\mathcal{R}|$ , where $|\mathcal{A},\mathcal{R}|$ denotes the length of $(\mathcal{A},\mathcal{R})$, and $\lambda_{1}$, $\lambda_{2}$ and $\lambda_{3}$ are hyperparameters. Besides the channel model $p(\mathcal{C},\mathcal{B}|\mathcal{A},\mathcal{R})$ and the source model $p(\mathcal{A},\mathcal{R})$, we additionally use the direct model $p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})$ and a length bias $|\mathcal{A},\mathcal{R}|$ to encourage responses with high direct model likelihood and discourage short responses, respectively. Noisy channel online decoding: In contrast to reranking, online decoding applies the noisy channel model during beam search for pruning partial sequences, thus exploring a larger search space. As shown in Algorithm 1, we first decode the belief state with $p(\mathcal{B}|\mathcal{C})$, which comes from the direct model in Section 2. Then, starting with a beam $\mathcal{S}$ containing a single sequence [a] (the dialogue act start token), we continuously expand the sequences in $\mathcal{S}$ until end($\mathcal{S}$) is met, i.e. all sequences in $\mathcal{S}$ either end with [/r] or have lengths larger than $l$. In each iteration, we first expand the sequences in the beam, then prune the expanded beam. To expand a partial act and response sequence (denoted as $\mathcal{O}$ in Algorithm 1), a naive way is to use the noisy channel model to score $|V|$ (the vocabulary size) possible expansions, which is computationally expensive. Instead, we use the probability of the next token $p(O_{|\mathcal{O}|+1}|\mathcal{C},\mathcal{B},\mathcal{O})$ (where $|\mathcal{O}|$ denotes the length of $\mathcal{O}$) to select $k_{1}$ candidates to be scored by the noisy channel model. This next token probability is from the direct model introduced in Section 2. One straightforward way to select $k_{1}$ expansions from $p(O_{|\mathcal{O}|+1}|\mathcal{C},\mathcal{B},\mathcal{O})$ is using the top-k maximization, but we can also take advantage of the advances in sampling from a categorical distribution for text generation (e.g. top-k sampling Fan et al. (2018) and nucleus sampling Holtzman et al. (2019)). After the expansion, we prune the expanded beam $\mathcal{S}^{\prime}$ to obtain a smaller beam with $k_{2}$ partial sequences based on the model combination in Eq. 1. Compared to noisy channel reranking, online decoding applies the noisy channel model during beam search, which is potentially less biased towards the direct model. In summary, we note that beam search for both the direct model and the online decoding for our noisy channel model decodes ($\mathcal{B},\mathcal{A},\mathcal{R}$) autoregressively. Thus both approaches are end-to-end models for task-oriented dialogue. The key difference is that noisy channel online decoding uses Eq. 1 for pruning, while the direct model uses $p(\mathcal{A},\mathcal{R}|\mathcal{C},\mathcal{B})$. ## 4 Model and Pretraining We use three Transformer Vaswani et al. (2017) networks to parameterize the direct model $p(\mathcal{B},\mathcal{A},\mathcal{R}|\mathcal{C})$, the channel model $p(\mathcal{C},\mathcal{B}|\mathcal{A},\mathcal{R})$ and the source model $p(\mathcal{A},\mathcal{R})$, respectively. The input to each Transformer is the sum of four embeddings: word embeddings, position embeddings, role embeddings (user/system), and turn embeddings (each word corresponds to a turn number). Cross entropy is used as the loss function. Given training samples $(\mathcal{C},\mathcal{B},\mathcal{A},\mathcal{R})$, if we train the channel model using complete $(\mathcal{A},\mathcal{R})$ pairs as input, a significant discrepancy arises between training and decoding for noisy channel online decoding. Since the channel model is used to score partial act and response pairs, i.e. $p(\mathcal{C},\mathcal{B}|\mathcal{O})$ in Algorithm 1, the channel model trained with complete $(\mathcal{A},\mathcal{R})$ pairs is unsuited to scoring partial sequences. In order to manually create partial sequences during training that are better matched for online decoding, we truncate the $(\mathcal{A},\mathcal{R})$ pairs with a truncation length uniformly sampled from 1 to the sequence length (inclusive). The direct model and the source model are trained with complete sequences, as partial sequences occur naturally in their standard autoregressive training procedure. As in-domain dialogue data are usually scarce, we use a two-stage pretraining strategy to enhance the noisy channel model. Although the effectiveness of pretraining with Reddit data has been validated for open-domain dialogue Zhang et al. (2020b); Bao et al. (2019); Adiwardana et al. (2020), relatively little work has applied such data to task-oriented dialogue.444One exception is Henderson et al. Henderson et al. (2019), who use Reddit data to improve response retrieval and selection. We focus on response generation in this work. In the first stage, we explore Reddit pretraining (where the Reddit data is pre-processed into $(\mathcal{C},\mathcal{R})$, i.e. context-response, pairs as described below). In the second stage, we use two task-oriented dialogue datasets, Taskmaster555https://cutt.ly/xkuUHUa Byrne et al. (2019) and Schema-Guided Dialogue666https://cutt.ly/QkuUZUu Rastogi et al. (2020), to specialize the Reddit-pretrained models. Since the Reddit data consists of open-domain-style dialogues (where belief states and dialogue acts are missing), pretraining on these datasets can familiarize the models with the sequence-to-sequence representation of task-oriented dialogue. Three models, a context-to-response model, a response-to-context model and a response language model, are pretrained to initialize the direct model, the channel model and the source model, respectively. Dataset | # Dialog | # Turn | Avg. Turn/Dialog | Avg. Token/Turn | # Domain | Multi-Task | # Unique Slot | # Unique Value ---|---|---|---|---|---|---|---|--- Taskmaster | 17,304 | 341,801 | 19.75 | 7.87 | 7 | ✗ | 281 | 66,659 Schema | 22,825 | 463,284 | 20.3 | 9.86 | 17 | ✓ | 123 | 23,889 CamRest676 | 676 | 5,488 | 8.12 | 10.71 | 1 | ✗ | 4 | 89 MultiWOZ | 10,438 | 143,048 | 13.7 | 15.03 | 7 | ✓ | 46 | 11,828 SMCalFlow | 41,517 | 170,590 | 4.11 | 8.77 | 4 | ✓ | - | - Table 1: Statistics of task-oriented dialogue datasets. We define a multi-task dialogue as a dialogue involving multiple tasks, e.g. hotel and restaurant booking, while its counterpart handles a single task, e.g. hotel booking. Taskmaster and CamRest676 do not contain any multi-task dialogues. ### 4.1 Implementation Details Models: All models are implemented with JAX Bradbury et al. (2018) and Haiku Hennigan et al. (2020). For the direct model introduced in Section 2, we use a Transformer model with hidden size 512, 12 encoder-decoder layers, and 16 self-attention heads. The model has 114M parameters. For the noisy channel model, we use a base setting and a large setting. The base setting reduces the number of layers to 5, hidden size to 384 and self-attention heads to 12. Its sub-modules, a direct model, a reverse model and a language model, have 43M, 43M and 30M parameters, respectively. We employ the base setting for a fair comparison with a single direct model using roughly the same number of parameters (116M vs. 114M). For the large setting, we use the same hyperparameters as the direct model (114M), so that its sub-modules, a direct model, a reverse model and a language model, have 114M, 114M and 64M parameters, respectively. We use this large setting to explore the limits of the noisy channel model. The large noisy channel model (292M) is 2.56 times larger compared to the direct model (114M). This illustrates another advantage of the noisy channel model during training. While training a direct model with 292M parameters will overflow the memory of 16GB TPUs (v3) without using model parallelism, training the sub-modules of the large noisy channel model can easily fit into 16GB TPUs, as these modules are independently trained with no need to load three modules for training. This enables us to train a noisy channel model with more parameters compared to training a direct model using the same hardware. For inference, we still need to load the sub-modules into a TPU. Since gradients are not required during inference, we are able to load the three sub-modules of the large noisy channel model (292M) into a single TPU with 16GB memory for decoding. The large noisy channel model (292M) still consumes more memory than the direct model (114M) during inference. Pretraining settings: The maximum sequence length $l$ is set to 1024, and sequences with longer lengths are truncated. We reuse the vocabulary from GPT-2 Radford et al. (2019), which contains 50,257 BPE tokens. We use PreNorm Nguyen and Salazar (2019) for faster convergence. GELU Hendrycks and Gimpel (2016) is applied as the activation function. Following ALBERT Lan et al. (2020), dropout is disabled during pretraining. We use the normal distribution truncated to the range $[-0.01,0.01]$ to initialize the input embeddings, while other parameters are initialized using the normal distribution with zero mean and standard deviation 0.1. The batch size is set to 256. The LAMB optimizer You et al. (2020) ($b_{1}=0.9$ and $b_{2}=0.999$) is employed for optimization. The initial learning rate is 1e-7, and we apply 4000 warmup steps to increase the learning rate to 1e-3, before utilizing cosine annealing to decay the learning rate. Gradient clipping with clipping value 1 is applied to avoid gradient explosion. We use gradient accumulation with accumulation step 20. Pretraining: For Reddit pretraining, we download a Reddit dump (with Reddit posts ranging from 2005-12 to 2019-09) from PushShift.777https://pushshift.io/ Since the comments of a Reddit post are organized into a tree, we extract paths from a tree as dialogue turns. The last comment of each comment path is regarded as the response, while the others are used as the dialogue context. We pretrain each model for 400,000 steps, consuming 102,400,000 (400,000 $\times$ 256) comment paths in total. For the task-oriented pretraining, we combine the two datasets, Taskmaster and Schema-Guided Dialogue, and pretrain for 1e5 steps. The statistics of the task-oriented dialogue datasets are shown in Table 1. We train each model using 64 TPU chips with 16GB memory each. The pretraining takes around 4 days to complete. ## 5 Experiments We fine-tune and evaluate the pretrained models on three dialogue datasets: MultiWOZ 2.0, CamRest676 and SMCalFlow Andreas et al. (2020). In this section we describe the datasets (Section 5.1), fine-tuning (Section 5.2), decoding (Section 5.3) and evaluation metrics (Section 5.4). Results are presented in Section 6, and analysis and ablation studies in Section 7. Model | Inform $\uparrow$ | Success $\uparrow$ | BLEU $\uparrow$ | Combined $\uparrow$ ---|---|---|---|--- Sequicity Lei et al. (2018) | 66.4 | 45.3 | 15.54 | 71.39 HRED-TS Peng et al. (2019) | 70.0 | 58.0 | 17.50 | 81.50 DSTC8 Track 1 Winner Ham et al. (2020) | 73.0 | 62.4 | 16.00 | 83.50 DAMD Zhang et al. (2020a) | 76.4 | 60.4 | 16.60 | 85.00 SimpleTOD Hosseini-Asl et al. (2020) | 84.4 | 70.1 | 15.01 | 92.26 SOLOIST Peng et al. (2020a) | 85.5 | 72.9 | 16.54 | 95.74 UBAR Yang et al. (2021)† | 88.2 | 79.5 | 16.43 | 100.28 Randomly Initialized Direct decoding (114M) | 81.0 | 54.7 | 15.12 | 82.97 Noisy channel reranking (116M) | 82.7 | 57.1 | 15.29 | 85.19 Noisy channel online decoding (116M) | 82.9 | 58.9 | 15.33 | 86.23 Noisy channel reranking (292M) | 82.1 | 58.1 | 15.37 | 85.47 Noisy channel online decoding (292M) | 83.9 | 60.9 | 15.57 | 87.97 Reddit Pretraining Direct decoding (114M) | 81.0 | 69.2 | 17.06 | 92.16 Noisy channel reranking (116M) | 81.3 | 70.1 | 19.01 | 94.71 Noisy channel online decoding (116M) | 81.6 | 71.1 | 19.31 | 95.66 Noisy channel reranking (292M) | 82.2 | 70.9 | 19.89 | 96.44 Noisy channel online decoding (292M) | 82.4 | 71.7 | 20.49 | 97.54 Task-Oriented Pretraining Direct decoding (114M) | 85.2 | 72.9 | 17.00 | 96.05 Noisy channel reranking (116M) | 85.6 | 73.8 | 19.38 | 99.08 Noisy channel online decoding (116M) | 85.9 | 74.8 | 19.76 | 100.11 Noisy channel reranking (292M) | 86.5 | 74.9 | 20.31 | 101.01 Noisy channel online decoding (292M) | 86.9 | 76.2 | 20.58 | 102.13 Table 2: MultiWOZ test results (end-to-end modeling with generated beliefs) with seq2seq approaches. Results are significant (p < 0.01) comparing noisy channel decoding and direct decoding. $\dagger$ Yang et al. (2021) also report a combined score of 105.1 with an alternative context and evaluation setting, contributions orthogonal to our work and the other benchmarks reported here. ### 5.1 Datasets MultiWOZ888https://cutt.ly/0kuUCRS is a multi-domain dataset consisting of dialogues annotated with $\mathcal{C},\mathcal{B},\mathcal{A},\mathcal{R}$ in the following seven domains: attraction, hotel, hospital, police, restaurant, train, and taxi. Since its release, MultiWOZ has been one of the most commonly used task-oriented dialogue datasets. CamRest676999https://cutt.ly/SkuUNfE is annotated similarly to MultiWOZ and consists of dialogues in a single domain: restaurant reservations. Though CamRest676 is smaller than MultiWOZ and predates it, it still provides a widely used benchmark for evaluating task-oriented dialogue models. SMCalFlow consists of dialogues in four domains: calendar, weather, places, and people. Unlike MultiWOZ and CamRest676, SMCalFlow uses dataflow graphs instead of slot-value pairs to represent belief states and does not annotate dialogue acts. We refer readers to Andreas et al. Andreas et al. (2020) for a detailed description of the dataflow representation. We follow Andreas et al. Andreas et al. (2020) to convert dataflow graphs into sequences to apply seq2seq models. This dataset is newer and offers fewer prior models to compare with, but we use this dataset to study the robustness of the noisy channel model under different annotation schemas. We use the public splits for these datasets, where MultiWOZ, CamRest676 and SMCalFlow are split to 8438/1000/1000, 404/136/136 and 32647/3649/5211 dialogues for training, development and testing, respectively. However, since SMCalFlow’s test set has not been publicly released, we randomly select 500 dialogues from its training set to tune hyperparameters and use its development set for testing. Preprocessing: We use the standard preprocessing procedures for each dataset in order to facilitate fair comparison with previous methods.101010https://cutt.ly/TkuU1oM 111111https://cutt.ly/zkuU0Ht 121212https://cutt.ly/vkuU9bT In particular, for MultiWOZ and CamRest676, delexicalization is used to reduce lexical variation, while SMCalFlow does not use delexicalization. During delexicalization, slot values are replaced by generic placeholders based on a pre-defined dictionary. During decoding, following prior work, our dialogue models generate delexicalized responses. These delexicalized responses are re-lexicalized in post-processing by replacing placeholders with their corresponding slot values based on belief states and database information. Since there is no public code for lexicalization,131313We confirmed this with the dataset authors by email. we implement our own functions for lexicalization with regular expressions, for the purpose of displaying example responses. However, this does not affect reported results, as the standard metrics for MultiWOZ and CamRest676 which we adopt here are calculated using delexicalized responses. Model | Inform $\uparrow$ | Success $\uparrow$ | BLEU $\uparrow$ | Combined $\uparrow$ ---|---|---|---|--- Sequicity Lei et al. (2018) | 92.3 | 85.3 | 21.40 | 110.20 GPT-2 fine-tuned Wu et al. (2019b) | - | 86.2 | 19.20 | - ARDM Wu et al. (2019b) | - | 87.1 | 25.20 | - SOLOIST Peng et al. (2020a) | 94.7 | 87.1 | 25.50 | 116.40 Randomly Initialized Direct decoding (114M) | 78.1 | 83.5 | 21.58 | 102.38 Noisy channel online decoding (116M) | 79.8 | 84.1 | 22.83 | 104.78 Noisy channel online decoding (292M) | 80.9 | 84.9 | 23.19 | 106.09 Reddit Pretraining Direct decoding (114M) | 93.3 | 83.9 | 23.41 | 112.01 Noisy channel online decoding (116M) | 93.7 | 84.5 | 25.14 | 114.24 Noisy channel online decoding (292M) | 93.9 | 84.7 | 25.38 | 114.68 Task-Oriented Pretraining Direct decoding (114M) | 93.4 | 84.3 | 24.92 | 113.77 Noisy channel online decoding (116M) | 94.3 | 85.2 | 25.98 | 115.73 Noisy channel online decoding (292M) | 95.4 | 85.3 | 26.89 | 117.24 Table 3: CamRest676 test results (end-to-end modeling with generated beliefs) with seq2seq approaches. Noisy channel reranking performs comparable with noisy channel online decoding, and the results are not shown. Results are significant (p < 0.01) comparing noisy channel decoding and direct decoding. Model | SacreBLEU $\uparrow$ | TER $\downarrow$ ---|---|--- Randomly Initialized Direct decoding (114M) | 51.30 | 89.13 Online decoding (116M) | 53.66 | 74.18 Online decoding (292M) | 54.39 | 73.18 Reddit Pretraining Direct decoding (114M) | 60.68 | 61.99 Online decoding (116M) | 63.29 | 47.16 Online decoding (292M) | 63.91 | 46.43 Task-Oriented Pretraining Direct decoding (114M) | 61.02 | 59.84 Online decoding (116M) | 63.72 | 46.27 Online decoding (292M) | 64.29 | 45.81 Table 4: SMCalFlow results. Reranking performs worse than online decoding, and the results are not shown. Results are significant (p < 0.01) comparing noisy channel decoding and direct decoding. ### 5.2 Fine-Tuning We apply label smoothing with parameter 0.1. Dropout is used on input embeddings and hidden representations, with dropout rate 0.1. The Adam optimizer Kingma and Ba (2015) ($b_{1}=0.9$ and $b_{2}=0.999$) is adopted. We use a fixed learning rate 1e-4 with gradient clipping for fine-tuning. ### 5.3 Decoding We use direct decoding for belief state. For dialogue act and response, we study three decoding methods: direct decoding, noisy channel reranking and noisy channel online decoding. Since all of these decoding methods require choosing $k_{1}$ tokens from a categorical distribution during expansion, we compare four methods, top-k maximization, sampling without replacement, top-k sampling, and nucleus sampling. Nucleus sampling with cumulative probability 0.98 performs marginally better and is adopted. We perform a range search with the range $[1,20]$ on development sets for the beam sizes $k_{1}$ and $k_{2}$, and we set $k_{1},k_{2}=4$, $k_{1},k_{2}=15$ and $k_{1},k_{2}=4$ for MultiWOZ, CamRest676 and SMCalFlow, respectively. For noisy channel reranking and noisy channel online decoding, a grid search with range $[0,2]$ is performed for $\lambda_{1}$, $\lambda_{2}$ and $\lambda_{3}$. We set ($\lambda_{1}=0.8$, $\lambda_{2}=1$, $\lambda_{3}=0.8$), ($\lambda_{1}=1.2$, $\lambda_{2}=1.2$, $\lambda_{3}=0.8$) and ($\lambda_{1}=0.4$, $\lambda_{2}=1$, $\lambda_{3}=0.2$) for MultiWOZ, CamRest676 and SMCalFlow, respectively. ### 5.4 Evaluation Metrics For MultiWOZ and CamRest676, following previous work, we adopt three automatic evaluation metrics: inform, success and BLEU score. Peng et al. (2020a) showed that these metrics are well correlated to human evaluation. The evaluators141414https://cutt.ly/VkuU3FA 151515https://cutt.ly/MkuU88u provided with the datasets are used for calculating these metrics. To calculate the inform score for a dialogue, the evaluator first checks whether certain placeholders (e.g. [restaurant_name]) appear in decoded responses. If so, decoded belief states are converted to database queries to retrieve database records. These database records are compared with the records retrieved with ground-truth belief states. The inform score is one if these two sets of database records match. The success score takes all the requestable slots (e.g. postcode, phone number and address) from a decoded response and compares these requestable slots with the ones in the ground-truth response. The success score is one if generated requestable slots coincide with the ground- truth ones. BLEU score (BLEU-4) compares the n-grams of generated responses and human responses, and is a widely used metric in NLP for evaluating text quality. Following Budzianowski et al. (2018), we also calculate a combined score, which is (Inform + Success) / 2 + BLEU. For SMCalFlow, inform and success scores are not applicable since calculation of these scores relies on delexicalization placeholders, and this dataset does not use delexicalization. We use SacreBLEU161616https://cutt.ly/BkuU7dL and TER171717https://pypi.org/project/pyter/ to directly measure the quality of responses. As prior work on this dataset has focused on belief tracking rather than end-to-end response generation, we are the first to use these metrics on this dataset. We perform significance tests, where we use t-test for inform, success and TER scores and use permutation test for BLEU. ## 6 Results MultiWOZ: Results on the MultiWOZ test set are shown in Table 2. We observe several trends. First, the base noisy channel model (116M) performs better than direct decoding (114M), despite having a similar number of parameters, showing that the noisy channel factorization is beneficial for task-oriented dialogue. The large noisy channel setting improves further over the base setting. Second, Reddit pretraining provides benefits over random initialization, validating the use of large open-domain dialogue-genre pretraining for task-oriented dialogue, while the models with a second stage of task-oriented pretraining obtain further improvements. This effect is consistent across both direct and noisy channel decoding. Finally, we observe that online decoding consistently outperforms reranking, indicating the benefits of tighter model integration during decoding. Our model performs better on combined score than SOLOIST Peng et al. (2020a), a closely related baseline which pretrains a GPT2-initialized Transformer with Taskmaster and Schema-Guided Dialogue and decodes with nucleus sampling. CamRest676: Results on the CamRest676 test set are shown in Table 3. We observe that the base noisy channel model (116M) obtains better results compared to direct decoding (114M), again demonstrating the effectiveness of the noisy channel model. Reddit pretraining again provides a large benefit over random initialization for both direct decoding and noisy channel decoding, while task-oriented pretraining provides a further boost. Our model again performs better than SOLOIST. SMCalFlow: Results on the SMCalFlow development set are shown in Table 4. As end-to-end models have not previously been tested on this dataset, we use it to demonstrate that the noisy channel model, which we developed primarily on MultiWOZ, continues to be effective on task-oriented dialogue datasets with different annotation schema. The results are consistent with MultiWOZ and CamRest676. The noisy channel model outperforms the direct model by a large margin, demonstrating that dialogue act annotations are not essential for the noisy channel model, and that it remains effective across diverse dialogue representations. Reddit pretraining confers a similar large benefit on SMCalFlow as on the other datasets, but we observe that task-oriented pretraining brings only marginal further improvements. This may be due to differences in domain or format between our pretraining datasets and SMCalFlow. Alternatively, task- oriented pretraining may help more on task-specific metrics, such as inform and success scores, than on text quality metrics such as BLEU and TER scores. This hypothesis is further supported by the MultiWOZ results in Table 2. ## 7 Analysis In this section, we use MultiWOZ and CamRest676 to perform ablation studies on the effects of model combination, large-scale pretraining, and sample efficiency; as well as analyzing the runtime requirements of our model and the reasons for its success. Model | CamRest676 | MultiWOZ ---|---|--- Direct decoding | 115.17 | 96.73 Noisy Channel Online Decoding Direct + Channel | 115.63 | 98.54 Direct + Source | 115.91 | 99.12 Direct + Length | 115.56 | 97.57 Channel + Source | 115.82 | 99.18 Channel + Length | 115.60 | 98.71 Source + Length | 115.62 | 99.19 All - Direct | 115.96 | 100.89 All - Channel | 116.56 | 100.93 All - Source | 116.38 | 99.92 All - Length | 116.52 | 101.11 All | 116.91 | 102.62 Table 5: Ablation results for model combination on development sets (combined score). Results for reranking are similar and are not shown. ‘All’, ‘Direct’ ‘Source’, and ‘Channel’ denote no ablation, direct model, source model and channel model, respectively. Rows with ‘+’ are combinations of two sub- modules, while the rows with ‘-’ are combinations of three sub-modules. (a) Reddit pretraining, CamRest676 (b) Task-oriented pretraining, CamRest676 (c) Reddit pretraining, MultiWOZ (d) Task-oriented pretraining, MultiWOZ Figure 2: Results showing the effect of pretraining scale. (a) Direct decoding, CamRest676 (b) Reranking, CamRest676 (c) Online decoding, CamRest676 (d) Direct decoding, MultiWOZ (e) Reranking, MultiWOZ (f) Online decoding, MultiWOZ Figure 3: Pretraining improves sample efficiency during fine-tuning. Model | CamRest676 | MultiWOZ ---|---|--- Direct decoding | 4.89 | 6.48 Reranking | 5.43 | 6.92 Online decoding | 8.73 | 10.97 Table 6: Average decoding time (in seconds) for each turn with different decoding methods. Model | CamRest676 | MultiWOZ ---|---|--- Ground truth | 14.50 | 16.91 Direct decoding | 12.07 | 12.85 Direct decoding + Length | 15.98 | 17.73 Reranking | 15.09 | 17.47 Online decoding | 15.14 | 17.32 Table 7: The average length of responses with different decoding methods (on test set). The value closest to the ground truth is bold. Model | CamRest676 | MultiWOZ ---|---|--- Ground truth | 1.07 | 1.22 Direct decoding | 0.84 | 0.91 Reranking | 0.87 | 0.99 Online decoding | 0.89 | 1.03 Table 8: The Zipf scores of responses with different decoding methods (on test set). The value closest to the ground truth is bold. Model | CamRest676 | MultiWOZ ---|---|--- Direct decoding | 0.24 | 0.31 Reranking | 0.12 | 0.14 Online decoding | 0.08 | 0.11 Table 9: The likelihood (%) of falling into repetition loops for different decoding methods (on test set). ### 7.1 Ablation on Model Combination Noisy channel decoding involves a combination of four sub-modules, as in Eq. 1: the direct model, channel model, language model, and length bias. We perform an ablation study to determine whether all model components are important to the result, using the large model. Results on the development sets of CamRest676 and MultiWOZ are presented in Table 5. Note that the ablation is performed after applying the direct model to obtain $k_{1}$ expansions at each beam search step for noisy channel online decoding. We find that the combination of all four sub-modules performs the best, followed by combinations of three and then two sub-modules. The results are significant when comparing ‘All’ and the baselines ($p<0.01$). This result demonstrates the effectiveness of the noisy channel factorization, and the importance of each model component. ### 7.2 Effect of Pretraining Scale We investigate the importance of scale for both our pretraining stages. We select different checkpoints for Reddit pretraining, and truncate the two task-oriented dialogue datasets for task-oriented pretraining. We fine-tune these models using the full training data of CamRest676 or MultiWOZ. The results of three decoding methods (with the large noisy channel model) on the development sets are shown in Figure 2. In Figure 2 (a) and (c), the combined scores of all three decoding methods improve with more Reddit pretraining steps, demonstrating the advantage of increasing amounts of data in the open- domain dialogue pretraining stage. In Figure 2 (b) and (d), the combined scores further increase with more task-oriented data, confirming that additional task-oriented pretraining data is useful. Model | Delexicalized Response | SacreBLEU $\uparrow$ ---|---|--- Dialogue: MUL1624, Turn Number: 4 | Ground truth | Sure. The trainid is $[$train_id$]$. It leaves at $[$value_time$]$ and arrives at $[$value_time$]$. Would you like to make a reservation? | - Direct decoding | The travel time is $[$value_count$]$ minutes. | 2.03 Reranking | $[$train_id$]$ leaves $[$value_place$]$ at $[$value_time$]$ and arrives in $[$value_place$]$ at $[$value_time$]$. Would you like me to book that for you? | 46.33 Online decoding | $[$train_id$]$ leaves at $[$value_time$]$ and arrives at $[$value_time$]$. Would you like me to book that? | 60.90 Dialogue: MUL1276, Turn Number: 6 | Ground truth | The $[$restaurant_name$]$ has a postal code of $[$restaurant_postcode$]$. Will this work for you? I can book it if you please. | - Direct decoding | There are $[$value_count$]$ places. What type of food would you like? | 1.63 Reranking | How about $[$restaurant_name$]$ located at $[$restaurant_address$]$, postcode $[$restaurant_postcode$]$. | 24.90 Online decoding | How about $[$restaurant_name$]$ located at $[$restaurant_address$]$, postcode $[$restaurant_postcode$]$? | 22.00 Dialogue: MUL1898, Turn Number: 2 | Ground truth | Yes, $[$attraction_name$]$ is on $[$attraction_address$]$ and is in the $[$value_area$]$ side of town. Is there anything else you need to know? | - Direct decoding | $[$attraction_name$]$ is located in the $[$value_area$]$ part of town and has free admission. | 27.53 Reranking | $[$attraction_name$]$ is located in the $[$value_area$]$ of town at $[$attraction_address$]$. The entrance fee is free. Can I help you with anything else? | 41.66 Online decoding | $[$attraction_name$]$ is located in the $[$value_area$]$ part of town at $[$attraction_address$]$. Can I help you with anything else? | 42.38 Table 10: Case study on the responses decoded by direct decoding, noisy channel reranking and noisy channel online decoding. The large noisy channel model is used. ### 7.3 Sample Efficiency of Fine-Tuning We investigate whether pretraining can improve sample efficiency during fine- tuning. We gradually increase the amount of fine-tuning data and evaluate the randomly-initialized, Reddit pretrained and task-oriented pretrained models. The results on the development sets are shown in Figure 3. Combined scores increase with more training data under all conditions. Crucially, Reddit pretrained models show better performance with a smaller amount of fine-tuning data than randomly initialized models, and task-oriented pretrained models better still. We conclude that both our pretraining stages can improve sample efficiency, which is especially important when the target task has little training data. ### 7.4 Decoding Runtime In Table 6, we report the average clock time for decoding one turn (including its belief state, dialogue act and response). Noisy channel reranking is slightly slower compared to direct decoding, with overhead due to the reranking step in Eq. 1. Noisy channel online decoding is significantly slower, since it needs to apply Eq. 1 at each beam search step. In future work we will investigate ways to improve the efficiency of online decoding. ### 7.5 Decoding Properties In this section we analyze why the noisy channel model performed better than direct decoding. Length: In Table 7 we show the average length of generated responses. Direct decoding produces shorter responses than the ground truth, confirming that the direct model prefers short and generic responses. Adding a length bias to direct decoding (with lambda tuned on the development sets) produces responses longer than the ground truth, which may be a disadvantage. The noisy channel models produce responses with average length closest to the ground truth. Zipf: Table 9 shows the Zipf scores of responses. We find that the word distributions of responses generated by the noisy channel models are closer to the word distribution of ground-truth responses. Repetition: In Table 9 we examine the likelihood of falling into repetition loops Holtzman et al. (2019) for different decoding methods. Repetition loops are rare for all decoding methods, but noisy channel decoding can further decrease their likelihood. The channel model can discount a sequence with a repetition loop, since it conveys less information than a natural sequence of the same length, making it harder to “explain” the context. Examples: Some examples of responses are shown in Table 10. We observe that noisy channel models decode longer responses compared to direct decoding, and that the responses can explain their dialogue contexts well to meet users’ requirements. ## 8 Related Work Task-oriented dialogue models: Most task-oriented dialogue systems break down the task into three components: belief tracking Henderson et al. (2013); Mrkšić et al. (2016); Rastogi et al. (2017); Nouri and Hosseini-Asl (2018); Wu et al. (2019a); Zhang et al. (2019); Zhou and Small (2019); Heck et al. (2020), dialogue act prediction Wen et al. (2017a); Tanaka et al. (2019) and response generation Chen et al. (2019); Budzianowski et al. (2018); Lippe et al. (2020). Traditionally, a modular approach is adopted, where these components are optimized independently (i.e. a pipeline design) or learned via multi-task learning (i.e. some parameters are shared among the components) Wen et al. (2017b); Neelakantan et al. (2019); Zhao et al. (2019); Mehri et al. (2019); Tseng et al. (2020); Lee et al. (2020). However, it is known that improvements in one component do not necessarily lead to overall performance improvements Ham et al. (2020), and the modular approach suffers from error propagation in practice Liu and Lane (2018). These observations gave rise to the sequence-to-sequence approach Lei et al. (2018); Pei et al. (2019); Budzianowski and Vulić (2019); Wu et al. (2019b); Zhang et al. (2020a); Ham et al. (2020); Hosseini-Asl et al. (2020); Peng et al. (2020a); Yang et al. (2021), where dialogue beliefs and acts are represented as text spans, and a sequence-to-sequence model is applied to subsume the three components. Our work is situated within this general approach. In contrast to previous work, however, which uses a direct model for decoding, we introduce the noisy channel model to improve task-oriented dialogue. Pretraining models for dialogue: Recent work has applied pretraining Peters et al. (2018); Devlin et al. (2019); Radford et al. (2019) to dialogue. For open- domain dialogue, DialoGPT Zhang et al. (2020b) and CGRG Wu et al. (2020b) extend GPT-2 Radford et al. (2019) for response generation. PLATO Bao et al. (2019) and PLATO-2 Bao et al. (2020) pretrain a latent variable model with social media data for diversified response generation. Meena Adiwardana et al. (2020) collects a large-scale social media corpus for pretraining and proposes a metric named sensibleness and specificity average for evaluation. Roller et al. Roller et al. (2020) study various strategies for building an open-domain chatbot with Reddit for pretraining. For task-oriented dialogue, ToD-BERT Wu et al. (2020a) fine-tunes BERT Devlin et al. (2019) for four tasks, including intention detection, belief tracking, dialogue act prediction, and response selection. SC-GPT Peng et al. (2020b) fine-tunes GPT-2 for few-shot response generation with given dialogue acts. Ham et al. Ham et al. (2020) fine-tune GPT-2 for belief tracking and context-to-response generation. SimpleTOD Hosseini-Asl et al. (2020) proposes a method to serialize dialogue beliefs and acts into text spans and fine-tunes GPT-2 for end-to-end dialogue modeling. SOLOIST Peng et al. (2020a) uses a series of task-oriented dialogue datasets to further pretrain GPT-2 before fine-tuning it on final tasks for evaluation. Unlike these BERT- or GPT-initialized task-oriented dialogue models, which are essentially pretrained with general text, such as Wikipedia and BookCorpus, we use a Reddit dump to pretrain the models to learn from open-domain dialogues. ## Conclusion We introduced two noisy channel models, noisy channel reranking and noisy channel online decoding, for task-oriented dialogue. Large-scale pretraining was further adopted to tackle data scarcity in downstream tasks. Extensive experiments on MultiWOZ, CamRest676 and SMCalFlow demonstrated that (1) the noisy channel models significantly outperform direct decoding; (2) models with pretraining improve over randomly-initialized models; (3) the models are robust to different dialogue schema annotations; (4) the noisy channel models can decode responses closer to ground-truth responses than direct decoding. ## Acknowledgements We would like to thank the action editors (Maggie, Wenjie Li and Eneko Agirre) and three anonymous reviewers for their insightful comments. We also thank Angeliki Lazaridou, Gábor Melis, Nando de Freitas, Chris Dyer and the DeepMind language team for their helpful discussions. ## References * Adiwardana et al. (2020) Daniel Adiwardana, Minh-Thang Luong, David R So, Jamie Hall, Noah Fiedel, Romal Thoppilan, Zi Yang, Apoorv Kulshreshtha, Gaurav Nemade, Yifeng Lu, et al. 2020\. Towards a human-like open-domain chatbot. _arXiv preprint arXiv:2001.09977_. * Andreas et al. (2020) Jacob Andreas, John Bufe, David Burkett, Charles Chen, Josh Clausman, Jean Crawford, Kate Crim, Jordan DeLoach, Leah Dorner, Jason Eisner, et al. 2020. Task-oriented dialogue as dataflow synthesis. _Transactions of the Association for Computational Linguistics_ , 8:556–571. * Austin (1975) John Langshaw Austin. 1975. _How to do things with words_ , volume 88. Oxford university press. * Bao et al. (2019) Siqi Bao, Huang He, Fan Wang, and Hua Wu. 2019. Plato: Pre-trained dialogue generation model with discrete latent variable. _arXiv preprint arXiv:1910.07931_. * Bao et al. (2020) Siqi Bao, Huang He, Fan Wang, Hua Wu, and Haifeng Wang. 2020. PLATO: pre-trained dialogue generation model with discrete latent variable. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020_ , pages 85–96. Association for Computational Linguistics. * Bradbury et al. (2018) James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, and Skye Wanderman-Milne. 2018. JAX: composable transformations of Python+NumPy programs. * Brants et al. (2007) Thorsten Brants, Ashok C. Popat, Peng Xu, Franz J. Och, and Jeffrey Dean. 2007. Large language models in machine translation. In _Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)_ , pages 858–867, Prague, Czech Republic. Association for Computational Linguistics. * Brown et al. (2020) Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In _Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual_. * Budzianowski and Vulić (2019) Paweł Budzianowski and Ivan Vulić. 2019. Hello, it’s GPT-2 - how can I help you? towards the use of pretrained language models for task-oriented dialogue systems. In _Proceedings of the 3rd Workshop on Neural Generation and Translation_ , pages 15–22, Hong Kong. Association for Computational Linguistics. * Budzianowski et al. (2018) Pawel Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Iñigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gasic. 2018. Multiwoz - A large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling. In _Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018_ , pages 5016–5026. Association for Computational Linguistics. * Byrne et al. (2019) Bill Byrne, Karthik Krishnamoorthi, Chinnadhurai Sankar, Arvind Neelakantan, Ben Goodrich, Daniel Duckworth, Semih Yavuz, Amit Dubey, Kyu-Young Kim, and Andy Cedilnik. 2019. Taskmaster-1: Toward a realistic and diverse dialog dataset. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019_ , pages 4515–4524. Association for Computational Linguistics. * Chen et al. (2019) Wenhu Chen, Jianshu Chen, Pengda Qin, Xifeng Yan, and William Yang Wang. 2019. Semantically conditioned dialog response generation via hierarchical disentangled self-attention. In _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics_ , pages 3696–3709, Florence, Italy. Association for Computational Linguistics. * Devlin et al. (2019) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers)_ , pages 4171–4186. Association for Computational Linguistics. * Fan et al. (2018) Angela Fan, Mike Lewis, and Yann N. Dauphin. 2018. Hierarchical neural story generation. In _Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers_ , pages 889–898. Association for Computational Linguistics. * Ham et al. (2020) Donghoon Ham, Jeong-Gwan Lee, Youngsoo Jang, and Kee-Eung Kim. 2020. End-to-end neural pipeline for goal-oriented dialogue systems using GPT-2. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_ , pages 583–592, Online. Association for Computational Linguistics. * Heck et al. (2020) Michael Heck, Carel van Niekerk, Nurul Lubis, Christian Geishauser, Hsien-Chin Lin, Marco Moresi, and Milica Gasic. 2020. Trippy: A triple copy strategy for value independent neural dialog state tracking. In _Proceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue, SIGdial 2020, 1st virtual meeting, July 1-3, 2020_ , pages 35–44. Association for Computational Linguistics. * Henderson et al. (2013) Matthew Henderson, Blaise Thomson, and Steve Young. 2013. Deep neural network approach for the dialog state tracking challenge. In _Proceedings of the SIGDIAL 2013 Conference_ , pages 467–471. * Henderson et al. (2019) Matthew Henderson, Ivan Vulic, Daniela Gerz, Iñigo Casanueva, Pawel Budzianowski, Sam Coope, Georgios Spithourakis, Tsung-Hsien Wen, Nikola Mrksic, and Pei-Hao Su. 2019. Training neural response selection for task-oriented dialogue systems. In _Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers_ , pages 5392–5404. Association for Computational Linguistics. * Hendrycks and Gimpel (2016) Dan Hendrycks and Kevin Gimpel. 2016. Gaussian error linear units (gelus). _arXiv preprint arXiv:1606.08415_. * Hennigan et al. (2020) Tom Hennigan, Trevor Cai, Tamara Norman, and Igor Babuschkin. 2020. Haiku: Sonnet for JAX. * Holtzman et al. (2019) Ari Holtzman, Jan Buys, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. _CoRR_ , abs/1904.09751. * Hosseini-Asl et al. (2020) Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. 2020. A simple language model for task-oriented dialogue. In _Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual_. * Kingma and Ba (2015) Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In _3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings_. * Klein and Manning (2002) Dan Klein and Christopher D Manning. 2002. Conditional structure versus conditional estimation in nlp models. In _Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing (EMNLP 2002)_ , pages 9–16. * Lan et al. (2020) Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. ALBERT: A lite BERT for self-supervised learning of language representations. In _8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020_. * Lee et al. (2020) Hwaran Lee, Seokhwan Jo, HyungJun Kim, Sangkeun Jung, and Tae-Yoon Kim. 2020. Sumbt+ larl: End-to-end neural task-oriented dialog system with reinforcement learning. _arXiv preprint arXiv:2009.10447_. * Lei et al. (2018) Wenqiang Lei, Xisen Jin, Min-Yen Kan, Zhaochun Ren, Xiangnan He, and Dawei Yin. 2018\. Sequicity: Simplifying task-oriented dialogue systems with single sequence-to-sequence architectures. In _Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , pages 1437–1447. * Li et al. (2016) Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In _NAACL HLT 2016, The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego California, USA, June 12-17, 2016_ , pages 110–119. The Association for Computational Linguistics. * Lippe et al. (2020) Phillip Lippe, Pengjie Ren, Hinda Haned, Bart Voorn, and Maarten de Rijke. 2020\. Diversifying task-oriented dialogue response generation with prototype guided paraphrasing. _CoRR_ , abs/2008.03391. * Liu and Lane (2018) Bing Liu and Ian Lane. 2018. End-to-end learning of task-oriented dialogs. In _Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop_ , pages 67–73. * Mehri et al. (2019) Shikib Mehri, Tejas Srinivasan, and Maxine Eskenazi. 2019. Structured fusion networks for dialog. _arXiv preprint arXiv:1907.10016_. * Mrkšić et al. (2016) Nikola Mrkšić, Diarmuid O Séaghdha, Tsung-Hsien Wen, Blaise Thomson, and Steve Young. 2016. Neural belief tracker: Data-driven dialogue state tracking. _arXiv preprint arXiv:1606.03777_. * Neelakantan et al. (2019) Arvind Neelakantan, Semih Yavuz, Sharan Narang, Vishaal Prasad, Ben Goodrich, Daniel Duckworth, Chinnadhurai Sankar, and Xifeng Yan. 2019. Neural assistant: Joint action prediction, response generation, and latent knowledge reasoning. _arXiv preprint arXiv:1910.14613_. * Nguyen and Salazar (2019) Toan Q Nguyen and Julian Salazar. 2019. Transformers without tears: Improving the normalization of self-attention. _arXiv preprint arXiv:1910.05895_. * Nouri and Hosseini-Asl (2018) Elnaz Nouri and Ehsan Hosseini-Asl. 2018. Toward scalable neural dialogue state tracking model. _arXiv preprint arXiv:1812.00899_. * Pei et al. (2019) Jiahuan Pei, Pengjie Ren, and Maarten de Rijke. 2019. A modular task-oriented dialogue system using a neural mixture-of-experts. _arXiv preprint arXiv:1907.05346_. * Peng et al. (2020a) Baolin Peng, Chunyuan Li, Jinchao Li, Shahin Shayandeh, Lars Liden, and Jianfeng Gao. 2020a. Soloist: Few-shot task-oriented dialog with a single pre-trained auto-regressive model. _arXiv preprint arXiv:2005.05298_. * Peng et al. (2020b) Baolin Peng, Chenguang Zhu, Chunyuan Li, Xiujun Li, Jinchao Li, Michael Zeng, and Jianfeng Gao. 2020b. Few-shot natural language generation for task-oriented dialog. In _Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, EMNLP 2020, Online Event, 16-20 November 2020_ , pages 172–182. Association for Computational Linguistics. * Peng et al. (2019) Shuke Peng, Xinjing Huang, Zehao Lin, Feng Ji, Haiqing Chen, and Yin Zhang. 2019\. Teacher-student framework enhanced multi-domain dialogue generation. _arXiv preprint arXiv:1908.07137_. * Peters et al. (2018) Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In _Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT_ , pages 2227–2237. * Radford et al. (2019) Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. _OpenAI Blog_ , 1(8):9. * Raffel et al. (2020) Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. _J. Mach. Learn. Res._ , 21:140:1–140:67. * Rastogi et al. (2017) Abhinav Rastogi, Dilek Hakkani-Tür, and Larry Heck. 2017. Scalable multi-domain dialogue state tracking. In _2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)_ , pages 561–568. IEEE. * Rastogi et al. (2020) Abhinav Rastogi, Xiaoxue Zang, Srinivas Sunkara, Raghav Gupta, and Pranav Khaitan. 2020. Towards scalable multi-domain conversational agents: The schema-guided dialogue dataset. In _The Thirty-Fourth AAAI Conference on Artificial Intelligence, February 7-12, 2020_ , pages 8689–8696. AAAI Press. * Raux et al. (2005) Antoine Raux, Brian Langner, Dan Bohus, Alan W Black, and Maxine Eskenazi. 2005\. Let’s go public! taking a spoken dialog system to the real world. In _Ninth European conference on speech communication and technology_. * Roller et al. (2020) Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M Smith, et al. 2020. Recipes for building an open-domain chatbot. _arXiv preprint arXiv:2004.13637_. * Seneff and Polifroni (2000) Stephanie Seneff and Joseph Polifroni. 2000. Dialogue management in the mercury flight reservation system. In _ANLP-NAACL 2000 Workshop: Conversational Systems_. * Shannon (1948) Claude Shannon. 1948. A mathematical theory of communication. _Bell System Technical Journal_ , 27:379–423. * Tanaka et al. (2019) Koji Tanaka, Junya Takayama, and Yuki Arase. 2019. Dialogue-act prediction of future responses based on conversation history. In _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop_ , pages 197–202. * Tseng et al. (2020) Bo-Hsiang Tseng, Jianpeng Cheng, Yimai Fang, and David Vandyke. 2020. A generative model for joint natural language understanding and generation. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020_ , pages 1795–1807. Association for Computational Linguistics. * Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In _Advances in neural information processing systems_ , pages 5998–6008. * Wen et al. (2017a) Tsung-Hsien Wen, Yishu Miao, Phil Blunsom, and Steve J. Young. 2017a. Latent intention dialogue models. _CoRR_ , abs/1705.10229. * Wen et al. (2017b) Tsung-Hsien Wen, David Vandyke, Nikola Mrksic, Milica Gasic, Lina Maria Rojas-Barahona, Pei-Hao Su, Stefan Ultes, and Steve J. Young. 2017b. A network-based end-to-end trainable task-oriented dialogue system. In _Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2017, Valencia, Spain, April 3-7, 2017, Volume 1: Long Papers_ , pages 438–449. Association for Computational Linguistics. * Wu et al. (2020a) Chien-Sheng Wu, Steven C. H. Hoi, Richard Socher, and Caiming Xiong. 2020a. TOD-BERT: pre-trained natural language understanding for task-oriented dialogue. In _Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020_ , pages 917–929. Association for Computational Linguistics. * Wu et al. (2019a) Chien-Sheng Wu, Andrea Madotto, Ehsan Hosseini-Asl, Caiming Xiong, Richard Socher, and Pascale Fung. 2019a. Transferable multi-domain state generator for task-oriented dialogue systems. In _Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers_ , pages 808–819. Association for Computational Linguistics. * Wu et al. (2019b) Qingyang Wu, Yichi Zhang, Yu Li, and Zhou Yu. 2019b. Alternating recurrent dialog model with large-scale pre-trained language models. _arXiv preprint arXiv:1910.03756_. * Wu et al. (2020b) Zeqiu Wu, Michel Galley, Chris Brockett, Yizhe Zhang, Xiang Gao, Chris Quirk, Rik Koncel-Kedziorski, Jianfeng Gao, Hannaneh Hajishirzi, Mari Ostendorf, et al. 2020b. A controllable model of grounded response generation. _arXiv preprint arXiv:2005.00613_. * Yang et al. (2021) Yunyi Yang, Yunhao Li, and Xiaojun Quan. 2021. Ubar: Towards fully end-to-end task-oriented dialog systems with gpt-2. _The Thirty-Fifth AAAI Conference on Artificial Intelligence_. * Yee et al. (2019) Kyra Yee, Yann N. Dauphin, and Michael Auli. 2019. Simple and effective noisy channel modeling for neural machine translation. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019_ , pages 5695–5700. Association for Computational Linguistics. * You et al. (2020) Yang You, Jing Li, Sashank J. Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. 2020\. Large batch optimization for deep learning: Training BERT in 76 minutes. In _8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020_. * Yu et al. (2017) Lei Yu, Phil Blunsom, Chris Dyer, Edward Grefenstette, and Tomás Kociský. 2017. The neural noisy channel. In _5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings_. * Yu et al. (2020) Lei Yu, Laurent Sartran, Wojciech Stokowiec, Wang Ling, Lingpeng Kong, Phil Blunsom, and Chris Dyer. 2020. Better document-level machine translation with bayes’ rule. _Transactions of the Association for Computational Linguistics_ , 8:346–360. * Zhang et al. (2019) Jian-Guo Zhang, Kazuma Hashimoto, Chien-Sheng Wu, Yao Wan, Philip S Yu, Richard Socher, and Caiming Xiong. 2019. Find or classify? dual strategy for slot-value predictions on multi-domain dialog state tracking. _arXiv preprint arXiv:1910.03544_. * Zhang et al. (2020a) Yichi Zhang, Zhijian Ou, and Zhou Yu. 2020a. Task-oriented dialog systems that consider multiple appropriate responses under the same context. In _The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020_ , pages 9604–9611. AAAI Press. * Zhang et al. (2020b) Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020b. DIALOGPT : Large-scale generative pre-training for conversational response generation. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, ACL 2020, Online, July 5-10, 2020_ , pages 270–278. Association for Computational Linguistics. * Zhao et al. (2019) Tiancheng Zhao, Kaige Xie, and Maxine Eskénazi. 2019. Rethinking action spaces for reinforcement learning in end-to-end dialog agents with latent variable models. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers)_ , pages 1208–1218. Association for Computational Linguistics. * Zhou and Small (2019) Li Zhou and Kevin Small. 2019. Multi-domain dialogue state tracking as dynamic knowledge graph enhanced question answering. _arXiv preprint arXiv:1911.06192_.
Bochner-Riesz Means Convergence of Prolate Spheroidal Series and Their Extensions Mourad Boulsanea 111 Corresponding author: Mourad Boulsane, Email: <EMAIL_ADDRESS>Ahmed Souabnia a Carthage University, Faculty of Sciences of Bizerte, Department of Mathematics, Jarzouna, 7021, Tunisia. ###### Abstract In this paper, we study the $L^{p}$-Bochner-Riesz mean summability problem related to the spectrum of some particular Sturm-Liouville operators in the weighted $L^{p}([a,b],\omega).$ Our purpose is to establish suitable conditions under which the Bochner-Riesz expansion of a function $f\in L^{p}([a,b],\omega)$,$1<p<\infty$, in two generalisations of Slepian’s basis, converges to $f$ in $L^{p}([a,b],\omega)$. Keywords: Bochner-Riesz mean convergence, eigenfunctions and eigenvalues, prolate spheroidal wave functions. 2010 Mathematics Subject Classification. 42C10, 41A60. ## 1 Introduction The $L^{p}$-Bochner-Riesz mean convergence of orthogonal series has attracted special attention since several decades ago. This kind of convergence is briefly described as follows. Let $1\leq p<\infty$ , $a,b\in\mathbb{R}$ and $\\{\varphi_{n}\\}$ an orthonormal set of eigenfunctions of a positive self- adjoint differential operator $\mathcal{L}$ associated with eigenvalues $\chi_{n}$ on a weighted Hilbert space $L^{2}(I,\omega)$, where $\omega$ is a positive bounded weight function. We define the expansion coefficients of $f\in L^{p}([a,b],\omega)$ by $a_{n}(f)=\int_{a}^{b}f(x)\varphi_{n}(x)\omega(x)dx.$ The orthonormal set $\\{\varphi_{n}\\}$ is said to have the Bochner-Riesz mean convergence of order $p$ over the Banach space $L^{p}(I,\omega)$ if for some suitable $\delta>0$ and for all $f\in L^{p}(I,\omega),$ we have $\lim_{R\to\infty}\int_{a}^{b}|f(x)-\Psi_{R}^{\delta}f(x)|^{p}\omega(x)dx=0,\mbox{ where }\displaystyle\Psi_{R}^{\delta}f=\sum_{n=0}^{\infty}\Big{(}1-\frac{\chi_{n}}{R}\Big{)}^{\delta}_{+}a_{n}(f)\varphi_{n}.$ (1) To the best of our knowledge, M. Riesz was the first, in 1911, to investigate this problem in some special cases. Our problem is a modified summability method of Riesz mean introduced by Salomon Bochner given by (1) . In [7], S.Bochner started by studying this problem for the trigonometric exponential case in higher dimension. Furthermore, in [13], the authors have proved a Bochner-Riesz mean convergence for the orthonormal eigenvectors system of a second order elliptic differential operator on a compact N-dimensional manifold M for $1\leq p\leq 2\frac{N+1}{N+3}$ and $\delta>N\left|\frac{1}{p}-\frac{1}{2}\right|-\frac{1}{2}$. Mauceri and Müller have also studied this problem in [20] and [22] in the framework of the Heisenberg group. This problem has been analysed for Fourier-Bessel expansions series in [10] and [11]. Moreover, in [8], authors have also solved this question in the case of sublaplacien on the sphere $S^{2n-1}$ in the complex n-dimensional space $\mathbb{C}^{n}$,where it has been shown that we have convergence for $\delta>(2n-1)\Big{|}\frac{1}{2}-\frac{1}{p}\Big{|}$. The weak type convergence is investigated in this problem. Indeed, we say that an orthonormal family $\\{\varphi_{n}\\}$ of $L^{p}(I,\omega)$ has a weakly Bochner-Riesz mean convergence if $\Psi_{R}^{\delta}f$ converge to $f$ almost everywhere for every $f\in L^{p}(I,\omega)$. This problem has been solved in some special cases of orthonormal systems like Jacobi and Laguerre polynomials in [21] and for the eigenfunctions of the Hermite operator in higher dimension in [9]. In this work, we extend the $L^{p}$-Bochner-Riesz means convergence to the circular and the generalized (or weighted) prolate spheroidal wave functions denoted by (CPSWFs) and (GPSWFs), respectively. The two last families are defined respectively as the eigenfunctions of the operators $\mathcal{H}_{c}^{\alpha}f(x)=\int_{0}^{1}\sqrt{cxy}J_{\alpha}(cxy)f(y)dy,\quad\mathcal{F}_{c}^{(\alpha)}f(x)=\int_{-1}^{1}e^{icxy}f(y)(1-y^{2})^{\alpha}\,dy,$ where $\alpha>-1/2,\,c>0$ are two real numbers. These two sets of orthonormal functions are characterized as solutions of some Sturm-Liouville problems. The second family we consider is the weighted, some times called generalized, prolate spheroidal wave functions (GPSWFs) introduced by Wang-Zhang [30]. Note that the classical PSWFs correspond to the special case of the GPSWFs with $\alpha=0.$ Our aim in this paper is to prove the $L^{p}$convergence of Bochner-Riesz mean expansion in the GPSWFs and CPSWFs bases. This work is organised as follows. In section 2, we give some mathematical preliminaries on Sturm-Liouville theory and some properties of the CPSWfs and GPSWFs. Note that these functions can be considered as generalizations of the spherical Bessel functions $j_{n}^{(\alpha)}$ and Gegenbauer’s polynomials $\widetilde{P}_{n}^{(\alpha)}$, respectively. In section 3, we state our two main theorems and section 4 and 5 are respectively devoted to the proof of sufficient and necessary conditions of the main results. ## 2 Mathematical preliminaries In this paragraph, we give some mathematical preliminaries that will be frequently used in the proofs of the different results of this work. ### 2.1 Some facts about Sturm-Liouville theory The Sturm-Liouville differential operator is defined as follows, see for example [1], $\mathcal{L}y(x)=\frac{d}{dx}[p(x)y^{\prime}(x)]+q(x)y(x),\quad x\in I=(a,b).$ (2) with $r=\frac{1}{p},q\in L^{1}(I,\mathbb{R}).$ The Sturm-Liouville eigenvalues problem is given by the following differential equation : $\mathcal{L}.u(x)=-\chi\omega(x)u(x),\quad\sigma\in L^{1}(I,\mathbb{R}).$ (3) That is $\frac{d}{dx}\Big{[}p(x)\frac{du}{dx}\Big{]}+q(x)u(x)+\chi\omega(x)u(x)=0,\quad x\in I.$ (4) Note that a Sturm-Liouville operator satisfies the following properties, 1. 1. $u\mathcal{L}v-v\mathcal{L}u=\Big{[}p(uv^{\prime}-vu^{\prime})\Big{]}^{\prime}$ ( Lagrange’s identity ) 2. 2. The eigenvalues of $\mathcal{L}$ are real and form an infinite countable set $\chi_{0}<\chi_{1}<\cdots<\chi_{n}<\cdots$ with $\lim_{n\rightarrow+\infty}\chi_{n}=+\infty.$ 3. 3. For each eigenvalue $\chi_{n}$ there exists an eigenfunction $\phi_{n}$ having n zeros on $[a,b].$ 4. 4. Eigenfunctions corresponding to different eigenvalues are orthogonal with respect to the following inner product ${\left\langle{f,g}\right\rangle}_{\omega}=\int_{a}^{b}f(x)g(x)\omega(x)dx,\quad f,g\in L^{2}(I,\omega).$ In the sequel, we assume that $\omega(x)\geq 0$, for $x\in(a,b).$ ### 2.2 Some facts about GPSWFs and CPSWFs We first recall that, for $c>0$, the prolate spheroidal wave functions PSWFs, denoted $\psi_{n,c}$, have been introduced by D.Slepian as solutions of the following energy maximization problem $\mbox{ Find }f=\arg\max_{f\in B_{c}}\frac{\int_{-1}^{1}|f(t)|^{2}dt}{\int_{\mathbb{R}}|f(t)|^{2}dt},$ where $B_{c}$ is the classical Paley-Wiener space, defined by $B_{c}=\left\\{f\in L^{2}(\mathbb{R}),\,\,\mbox{Support }\widehat{f}\subseteq[-c,c]\right\\}.$ (5) Here, $\widehat{f}$ is the Fourier transform of $f\in L^{2}(\mathbb{R}).$ It has been shown that they are also eigenfunctions of the integral operator with sinc kernel. A breakthrough in the theory of Slepian functions is due to Slepian,Pollard and Landau who have proved that PSWFs are also eigenfunctions of a Sturm-Liouville operator by proving a commutativity property. For more details about Slepian’s functions we refer reader to [25, 26, 27]. In this work we are interested in two generalizations of the PSWFs. The first basis is called circular prolate spheroidal wave functions (CPSWFs) or radial part of the 2d-Slepian, introduced by D.Slepian[27] as solutions of the following problem $\mbox{ Find }f=\arg\max_{f\in HB^{\alpha}_{c}}\frac{\int_{0}^{1}|f(t)|^{2}dt}{\int_{0}^{\infty}|f(t)|^{2}dt},$ where $HB^{\alpha}_{c}$ is the Hankel Paley-Wiener space, defined by $HB^{\alpha}_{c}=\left\\{f\in L^{2}(\mathbb{R}),\,\,\mbox{Support }\mathcal{H}^{\alpha}f\subseteq[-c,c]\right\\}.$ (6) Here the Hankel transform $\mathcal{H}^{\alpha}$ is defined, for $f\in L^{1}(0,\infty)$, by $\mathcal{H}^{\alpha}f(x)=\int_{0}^{\infty}\sqrt{xy}J_{\alpha}(xy)f(y)dy.$ Here $J_{\alpha}(.)$ is the Bessel function and $\alpha>-1/2$. Like Fourier transform, $\mathcal{H}^{\alpha}$ can be extended into a unitary operator on $L^{2}(0,\infty)$. They are also the different band-limited eigenfunctions of the finite Hankel transform $\mathcal{H}_{c}^{\alpha}$ defined on $L^{2}(0,1)$ with kernel $H_{c}^{\alpha}(x,y)=\sqrt{cxy}J_{\alpha}(cxy)$ where $J_{\alpha}$ is the Bessel function of the first type and order $\alpha>-\frac{1}{2}$(see for example [27]). That is $\mathcal{H}_{c}^{\alpha}(\varphi^{\alpha}_{n,c})=\mu_{n,\alpha}(c)\varphi^{\alpha}_{n,c}.$ (7) In his pioneer work [27], D. Slepian has shown that the compact integral operator $\mathcal{H}_{c}^{\alpha}$ commutes with the following Sturm- Liouville differential operator $\mathcal{L}^{\alpha}_{c}$ defined on $C^{2}([0,1])$ by $\mathcal{L}_{c}^{\alpha}(\phi)=-\dfrac{d}{dx}\left[(1-x^{2})\dfrac{d}{dx}\phi\right]+\left(c^{2}x^{2}-\dfrac{\dfrac{1}{4}-\alpha^{2}}{x^{2}}\right)\phi.$ (8) Hence, $\varphi^{\alpha}_{n,c}$ is the $n-$th bounded eigenfunction of the positive self-adjoint operator $\mathcal{L}_{c}^{\alpha}$ associated with the eigenvalue $\chi_{n,\alpha}(c),$ that is $-\dfrac{d}{dx}\left[(1-x^{2})\dfrac{d}{dx}\varphi^{\alpha}_{n,c}(x)\right]+\left(c^{2}x^{2}-\dfrac{\dfrac{1}{4}-\alpha^{2}}{x^{2}}\right)\varphi^{\alpha}_{n,c}(x)=\chi_{n,\alpha}(c)\varphi^{\alpha}_{n,c}(x),\quad x\in[0,1].$ (9) The orthonormal family $\varphi_{n,c}^{\alpha}$ form an orthonormal basis of $L^{2}(0,1)$ and the associated eigenvalues family $\chi_{n,\alpha}(c)$ satisfy the following inequality, (see [27]) $(2n+\alpha+1/2)(2n+\alpha+3/2)\leq\chi_{n,\alpha}(c)\leq(2n+\alpha+1/2)(2n+\alpha+3/2)+c^{2}$ (10) The second family we consider in this work is the weighted, (some times called generalized), prolate spheroidal wave functions introduced by Wang-Zhang [30] as solutions of a Sturm-Liouville problem or equivalently eigenfunctions of an integral operator. GPSWFs are also solutions of the following problem as given in [18] $\mbox{Find }f={\displaystyle arg\max_{f\in B^{\alpha}_{c}}\frac{\|f\|^{2}_{L^{2}_{\omega_{\alpha}}(I)}}{\|\widehat{f}\|^{2}_{L^{2}(\omega_{-\alpha}(\frac{\cdot}{c}))}}},$ where $\omega_{\alpha}(x)=(1-x^{2})^{\alpha}$ and $B^{(\alpha)}_{c}$ is the restricted Paley-Winer space, defined by $B_{c}^{(\alpha)}=\\{f\in L^{2}(\mathbb{R}),\,\,\mbox{Support }\widehat{f}\subseteq[-c,c],\,\,\widehat{f}\in L^{2}\big{(}(-c,c),\omega_{-\alpha}(\frac{\cdot}{c})\big{)}\\}.$ More precisely, the GPSWFs are the eigenfunctions of the weighted finite Fourier transform operator $\mathcal{F}_{c}^{(\alpha)}$ defined by $\mathcal{F}_{c}^{(\alpha)}f(x)=\int_{-1}^{1}e^{icxy}f(y)\,\omega_{\alpha}(y)\,\mathrm{d}y.$ (11) It is well known, (see [18, 30]) that they are also eigenfunctions of the compact and positive operator $\mathcal{Q}_{c}^{(\alpha)}=\frac{c}{2\pi}\mathcal{F}_{c}^{({\alpha})^{*}}\circ\mathcal{F}_{c}^{(\alpha)}$ which is defined on $L^{2}(I,\omega_{\alpha})$ by $\mathcal{Q}_{c}^{(\alpha)}g(x)=\int_{-1}^{1}\frac{c}{2\pi}\mathcal{K}_{\alpha}(c(x-y))g(y)\omega_{\alpha}(y)\d{y}$ (12) Here, $\mathcal{K}_{\alpha}(x)=\sqrt{\pi}2^{\alpha+1/2}\Gamma(\alpha+1)\frac{J_{\alpha+1/2}(x)}{x^{\alpha+1/2}}$ It has been shown in [18, 30] that the last two integral operators commute with the following Sturm-Liouville operator $\mathcal{L}_{c}^{(\alpha)}$ defined on $C^{2}[-1,1]$ by $\mathcal{L}_{c}^{(\alpha)}(f)(x)=-\frac{1}{\omega_{\alpha}(x)}\frac{d}{dx}\left[\omega_{\alpha}(x)(1-x^{2})f^{\prime}(x)\right]+c^{2}x^{2}f(x).$ (13) Also, note that the $(n+1)-$th eigenvalue $\chi_{n,\alpha}(c)$ of $\mathcal{L}_{c}^{(\alpha)}$ satisfies the following classical inequalities, $n(n+2\alpha+1)\leq\chi_{n,\alpha}(c)\leq n(n+2\alpha+1)+c^{2},\quad\forall n\geq 0.$ (14) ## 3 Statement of results In this section, we will state the main results of this paper that we will prove in the following sections. As mentioned before, the main issue studied in this paper is to get a necessary and sufficient conditions of Bochner-Riesz expansion convergence of a function $f$ in the GPSWFs’s and CPSWFs’s basis. Let’s start by studying the case of GPSWFs in the following theorem. ###### Theorem 1. Let $0\leq\alpha<3/2$, $\delta$ and $c$ be two positive number and $(\psi_{n,c}^{(\alpha)})_{n\geq 0}$ be the family of weighted prolate spheroidal wave functions. For a smooth function $f$ on $I=(-1,1)$, we define $\Psi_{R}^{\delta}f=\sum_{n=0}^{\infty}\left(1-\frac{\chi_{n,\alpha}(c)}{R}\right)^{\delta}_{+}{\left\langle{f,\psi_{n,c}^{(\alpha)}}\right\rangle}_{L^{2}(I,\omega_{\alpha})}\psi_{n,c}^{(\alpha)}.$ Then, for every $1\leq p<\infty$, $\Psi^{\delta}_{R}$ can be extended to a bounded operator $L^{p}(I,\omega_{\alpha})\to L^{p}(I,\omega_{\alpha})$. Further, $\Psi^{\delta}_{R}f$ is uniformly bounded if ,and only if, $\delta>\max\\{\frac{\gamma_{\alpha}(p^{\prime})}{2},0\\}$ and $p\not=p_{0}=2-\frac{1}{\alpha+3/2}$ where $\gamma_{\alpha}(p)=\begin{cases}0&\mbox{ if }1<p<p^{\prime}_{0}\\\ \epsilon&\mbox{ if }p=p^{\prime}_{0}\\\ 2(\alpha+1)\left[\frac{1}{2}-\frac{1}{p}\right]-\frac{1}{2}&\mbox{ if }p>p^{\prime}_{0}\\\ \alpha+1&\mbox{ if }p=1\end{cases}.$ and $\epsilon$ is an arbitrary real number. Note that $p^{\prime}$ denote here the dual exponent of $p$. ###### Remark 1. The sufficient condition in both GPSWFs’s and CPSWFs’s case still valid even for all $\alpha>-1/2$. ###### Remark 2 (Two special cases). Recall that $\psi^{(\alpha)}_{n,0}=\widetilde{P}^{(\alpha,\alpha)}_{n}$, then we recover the same result for the case of normalized Geganbauer polynomials. Note that both conditions (A) and (B), defined in the proof of the last theorem, are still valid even for $\widetilde{P}_{n}^{(\alpha,\beta)}$ with exactly the same proof and by noticing that the transferring theorem, which is the key step of the necessary condition, has been proven in [16] , the last result is valid also for Jacobi polynomials for all $\alpha,\beta>-1/2$ . For $\alpha=0$ and $c>0$, $\psi^{0}_{n,c}=\psi_{n,c}$ presents the classical prolate spheroidal wave functions PSWFs which satisfy the Bochner-Riez mean convergence if and only if $\delta>\max\\{0,\frac{\gamma_{0}(p^{\prime})}{2}\\}.$ Let us now focus on the circular case. ###### Theorem 2. Let $\alpha\geq 1/2$, $c>0$ and $(\varphi_{n,c}^{(\alpha)})_{n\geq 0}$ be the family of Hankel prolate spheroidal wave functions. For a smooth function $f$ on $I=(0,1)$, we define $\Psi_{R}^{\delta}f=\sum_{n=0}^{\infty}\left(1-\frac{\chi_{n,\alpha}(c)}{R}\right)^{\delta}_{+}{\left\langle{f,\varphi_{n,c}^{(\alpha)}}\right\rangle}_{L^{2}(0,1)}\varphi_{n,c}^{(\alpha)}.$ Then, for every $1\leq p<\infty$, $\Psi^{\delta}_{R}$ can be extended to a bounded operator $L^{p}(0,1)\to L^{p}(0,1)$. Further $\Psi^{\delta}_{R}f$ is uniformly bounded if ,and only if, $\delta>\max\\{\frac{\gamma(p^{\prime})}{2},0\\}$ where $\gamma(p)=\begin{cases}\frac{1}{p}-\frac{1}{2}&\mbox{ if }1<p<4\\\ \epsilon-\frac{1}{4}&\mbox{ if }p=4\\\ \frac{1}{3}\left[\frac{1}{p}-1\right]&\mbox{ if }p>4\\\ 1&\mbox{ if }p=1\end{cases}.$ . ## 4 Proof of sufficient condition Let $(I,\omega)$ be a measured space such that $\omega$ is a bounded weight function. We denote by $p^{\prime}=\frac{p}{p-1}$ the dual index of $p$. Throughout this section, $\mathcal{L}$ denotes a Sturm-Liouville operator and $\varphi_{n}$ (respectively $\lambda_{n}$) the sequence of the associated eigenfunctions (respectively eigenvalues). The Riesz means of index $\delta>0$ associated with $\mathcal{L}$ of a function $f\in\mathcal{C}^{\infty}(I,\mathbb{R})$ are consequently defined as $\Psi^{\delta}_{R}f=\sum_{n=0}^{\infty}\Big{(}1-\frac{\lambda_{n}}{R}\Big{)}^{\delta}_{+}a_{n}(f)\varphi_{n}\quad\mbox{with}\quad a_{n}(f)=\int_{I}f(y)\varphi_{n}(y)d\mu(y).$ (15) $\Psi^{\delta}_{R}f$ can also be written as $\Psi^{\delta}_{R}.f(x)=\int_{I}K_{R}^{\delta}(x,y)f(y)d\mu(y)\quad\mbox{where}\quad K_{R}^{\delta}(x,y)=\sum_{n=0}^{\infty}\Big{(}1-\frac{\lambda_{n}}{R}\Big{)}^{\delta}_{+}\varphi_{n}(x)\varphi_{n}(y)$ Our aim in this section is to prove at the same time the sufficient conditions of our two main theorems. More precisely, we will define several conditions on $\varphi_{n}$ that will ensure the convergence of $\Psi_{R}^{\delta}.f$ to $f$ in the $L^{p}$norm as $R\to\infty$ and verify that both two families satisfy these conditions. Assume that $\varphi_{n}$ satisfies the following conditions : * $(A)$ For every $1\leq p\leq\infty$, every $n$, $\varphi_{n}\in L^{p}(I,\omega)$. Further, we assume that there is a constant $\gamma(p)\geq 0$ such that ${\left\|{\varphi_{n}}\right\|}_{L^{p}(\mu)}\leq Cn^{\gamma(p)}$. * $(B)$ The sequence $(\lambda_{n})$ of the eigenvalues of the operator $\mathcal{L}$ satisfies the following properties 1. 1. $\displaystyle\sum_{\lambda_{n}\in(m,M)}1\leq C(M-m)$ for all $0\leq m<M$. 2. 2. There exists $\varepsilon>0$ such that $\lambda_{n}\geq Cn^{\varepsilon}.$ First of all, we start by giving sense to $\Psi^{\delta}_{R}.f$ for every $f\in L^{p}(\mu)$. Indeed, ${\left\|{K^{\delta}_{R}}\right\|}_{L^{p}(\mu)\otimes L^{p^{\prime}}(\mu)}\leq\sum_{\lambda_{n}<R}{\left\|{\varphi_{n}}\right\|}_{p}{\left\|{\varphi_{n}}\right\|}_{p^{\prime}}\leq\sum_{\lambda_{n}<R}n^{\gamma(p)+\gamma(p^{\prime})}\leq CR^{\frac{\left(\gamma(p)+\gamma(p^{\prime})\right)}{\varepsilon}+1},$ So that the integral operator $\Psi_{R}^{\delta}$ can be extended to a continuous operator $L^{p}(\mu)\to L^{p}(\mu)$ with ${\left\|{\Psi_{R}^{\delta}}\right\|}_{L^{p}\to L^{p}}\leq{\left\|{K^{\delta}_{R}}\right\|}_{L^{p}(\mu)\otimes L^{p^{\prime}}(\mu)}.$ The following theorem is one of the main results of this paper. ###### Theorem 3. With the above notation and under conditions $(A)$ and $(B)$ with $\delta>\delta(p)=\max\\{\frac{\gamma(p^{\prime})}{\varepsilon},0\\}$, there exists a constant $C>0$ satisfying the following inequality ${\left\|{\Psi_{R}^{\delta}}\right\|}_{(L^{p}(I,w),L^{p}(I,w))}\leq C.$ (16) The following lemma will be used in the proof of the previous theorem. ###### Lemma 1. Let $1\leq p\leq 2$ then for every $f\in L^{p}(I,\omega),$ we have ${\left\|{\sum_{\lambda_{n}\in(m,M)}a_{n}(f)\varphi_{n}}\right\|}_{L^{2}(I,\omega)}\leq C(p)M^{\frac{\gamma(p^{\prime})}{\varepsilon}}(M-m)^{\frac{1}{2}}{\left\|{f}\right\|}_{L^{p}(I,\omega)}.$ (17) ###### Proof. By orthogonality and Hölder’s inequality, we have $\displaystyle{\left\|{\sum_{\lambda_{n}\in(m,M)}a_{n}(f)\varphi_{n}}\right\|}^{2}_{L^{2}(I,\omega)}$ $\displaystyle=$ $\displaystyle\sum_{\lambda_{n}\in(m,M)}a_{n}^{2}(f)\leq\sum_{\lambda_{n}\in(m,M)}{\left\|{\varphi_{n}}\right\|}^{2}_{L^{p^{\prime}}(I,\omega)}{\left\|{f}\right\|}^{2}_{L^{p}(I,\omega)}$ From condition $(A)$, we have ${\left\|{\varphi_{n}}\right\|}_{L^{p^{\prime}}(I,\omega)}\leq n^{\gamma(p^{\prime})}$. We also obtain by using condition ($B1$) $\displaystyle{\left\|{\sum_{\lambda_{n}\in(m,M)}a_{n}(f)\varphi_{n}}\right\|}^{2}_{L^{2}(I,\omega)}$ $\displaystyle\leq$ $\displaystyle\sum_{\lambda_{n}\in(m,M)}n^{2\gamma(p^{\prime})}{\left\|{f}\right\|}^{2}_{L^{p}(I,\omega)}\leq C\sum_{\lambda_{n}\in(m,M)}\lambda_{n}^{\frac{2\gamma(p^{\prime})}{\varepsilon}}{\left\|{f}\right\|}^{2}_{L^{p}(I,\omega)}$ $\displaystyle\leq$ $\displaystyle CM^{\frac{2\gamma(p^{\prime})}{\varepsilon}}\left(\sum_{\lambda_{n}\in(m,M)}1\right){\left\|{f}\right\|}^{2}_{L^{p}(I,\omega)}$ $\displaystyle\leq$ $\displaystyle CM^{(\frac{2\gamma(p^{\prime})}{\varepsilon})}(M-m){\left\|{f}\right\|}^{2}_{L^{p}(I,\omega)}.$ Then one gets ${\left\|{\sum_{\lambda_{n}\in(m,M)}a_{n}(f)\varphi_{n}}\right\|}_{L^{2}(I,\omega)}\leq CM^{(\frac{\gamma(p^{\prime})}{\varepsilon})}(M-m)^{\frac{1}{2}}{\left\|{f}\right\|}_{L^{p}(I,\omega)}.$ ∎ ###### Proof of Theorem1. We should mention here that some parts of the proof of this theorem are inspired from [8]. Without loss of generality, we can consider $1\leq p<2$ and conclude by duality. To prove (16), we are going to have to decompose the multiplier $\Psi_{R}^{\delta}$. In order to do so, let $\phi\in\mathcal{C}^{\infty}_{0}(\mathbb{R})$ with support on $(1/2,2)$ such that $\displaystyle\sum_{k\in\mathbb{Z}}\phi(2^{k}t)=1$ and $\displaystyle\phi_{0}(t)=1-\sum_{k=1}^{+\infty}\phi(2^{k}t)$ for all $t>0$. We define $\phi_{R,k}^{\delta}(t)=\left(1-\frac{t}{R}\right)_{+}^{\delta}\phi\left(2^{k}(1-\frac{t}{R})\right).$ We recall that, from [8], this last function has the following properties : 1. 1. $\mbox{supp}\left(\phi_{R,k}^{\delta}\right)\subseteq(R(1-2^{-k+1}),R(1-2^{-k-1}))$, 2. 2. $\sup_{t\in\mathbb{R}}|\phi_{R,k}^{\delta}(t)|\leq C2^{-k\delta}$, 3. 3. $\forall N\geq 0,$ there exists $C_{N}>0$ such that $|\partial_{t}^{N}\phi_{R,k}^{\delta}(t)|\leq C_{N}\Big{(}\frac{2^{k}}{R}\Big{)}^{N}.$ Furthermore, we denote by $\Psi_{R,k}^{\delta}.f=\sum_{n=0}^{\infty}\phi_{R,k}^{\delta}(\lambda_{n})a_{n}(f)\varphi_{n}\qquad k=1,2,\cdots$ (18) Then, we have $\displaystyle\Psi_{R}^{\delta}f$ $\displaystyle=$ $\displaystyle\sum_{n=0}^{\infty}\left(1-\frac{\lambda_{n}}{R}\right)_{+}^{\delta}a_{n}(f)\varphi_{n}$ $\displaystyle=$ $\displaystyle\sum_{n=0}^{\infty}\phi_{0}(1-\frac{\lambda_{n}}{R})\left(1-\frac{\lambda_{n}}{R}\right)_{+}^{\delta}a_{n}(f)\varphi_{n}+\sum_{n=0}^{\infty}\left(\sum_{k=1}^{+\infty}\phi(2^{k}(1-\frac{\lambda_{n}}{R}))\right)\left(1-\frac{\lambda_{n}}{R}\right)_{+}^{\delta}a_{n}(f)\varphi_{n}$ $\displaystyle=$ $\displaystyle\sum_{n=0}^{\infty}\phi_{0}(1-\frac{\lambda_{n}}{R})\left(1-\frac{\lambda_{n}}{R}\right)_{+}^{\delta}a_{n}(f)\varphi_{n}+\sum_{k=1}^{\left[\frac{\log(R)}{\log(2)}\right]}\sum_{n=0}^{\infty}\phi_{R,k}^{\delta}(\lambda_{n})a_{n}(f)\varphi_{n}+\sum_{k=\left[\frac{\log(R)}{\log(2)}\right]+1}^{\infty}\sum_{n=0}^{\infty}\phi_{R,k}^{\delta}(\lambda_{n})a_{n}(f)\varphi_{n}$ $\displaystyle=$ $\displaystyle\psi_{R,0}^{\delta}f+\sum_{k=1}^{\left[\frac{\log(R)}{\log(2)}\right]}\Psi_{R,k}^{\delta}f+\mathcal{R}_{R}^{\delta}f.$ It is clear that the main term is the second one. With the same approach used in [8], we will prove the following proposition : ###### Proposition 1. Let $1\leq p<2$ and $\delta>\delta(p)=\frac{\gamma(p^{\prime})}{\varepsilon}$. There exists $\beta>0$ such that for every $f\in L^{p}(I,w)$, we have ${\left\|{\Psi_{R,k}^{\delta}f}\right\|}_{L^{p}(I,w)}\leq C2^{-k\beta}{\left\|{f}\right\|}_{L^{p}(I,w)},$ (19) where $C$ is a constant independent of $R$ and $f$. ###### Proof. Let $x_{0}=\frac{a+b}{2}\in(a,b)$ and $r=\frac{b-a}{4}>0$ such that $(x_{0}-r,x_{0}+r)\subseteq(a,b).$ Note that, for every $1\leq k\leq\left[\frac{\log(R)}{\log(2)}\right]=k_{R}$, we have $r_{k}^{\alpha}=\left(\frac{2^{k}}{R}\right)^{\mu(p)}r<r$ where $\mu(p)=\frac{(\frac{\gamma(p^{\prime})}{\varepsilon}+\frac{1}{2})}{(\frac{1}{p}-\frac{1}{2})}$ . So we notice that $I=(a,b)=(x_{0}-r_{k}^{\alpha},x_{0}+r_{k}^{\alpha})\cup\\{y\in(a,b),|y-x_{0}|>r_{k}^{\alpha}\\}=I_{k,1}^{\alpha}\cup I_{k,2}^{\alpha}$. We start by providing an $L^{p}$ bound of ${\left\|{\Psi_{R,k}^{\delta}}\right\|}_{L^{p}(I^{\alpha}_{k,1},\omega)}$. To do so, we proceed in the way to reduce the $L^{p}$ inequality (19) to certain $(L^{p},L^{2})$ inequality using the last lemma. Using Parseval formula and the fact that $\mbox{supp}\left(\phi_{R,k}^{\delta}\right)\subseteq(R_{k,1},R_{k,2})$, where $R_{k,1}=R(1-2^{-k+1})$ and $R_{k,2}=R(1-2^{-k-1}),$ we have $\displaystyle{\left\|{\Psi_{R,k}^{\delta}f}\right\|}^{2}_{L^{2}(I,w)}$ $\displaystyle=$ $\displaystyle{\left\|{\sum_{n=0}^{\infty}\phi_{R,k}^{\delta}(\lambda_{n})a_{n}(f)\varphi_{n}}\right\|}^{2}_{L^{2}(I,w)}$ $\displaystyle=$ $\displaystyle{\left\|{\sum_{R_{k,1}\leq\lambda_{n}\leq R_{k,2}}\phi_{R,k}^{\delta}(\lambda_{n})a_{n}(f)\varphi_{n}}\right\|}^{2}_{L^{2}(I,w)},$ Using the previous lemma with $m=R_{k,1}$, $M=R_{k,2}$ and the fact that $\displaystyle\sup_{t\in\mathbb{R}}|\phi_{R,k}^{\delta}(t)|\leq C2^{-k\delta}$, one gets $\displaystyle{\left\|{\Psi_{R,k}^{\delta}f}\right\|}^{2}_{L^{2}(I,w)}$ $\displaystyle\leq$ $\displaystyle C2^{-2k\delta}{\left\|{\sum_{R_{k,1}\leq\lambda_{n}\leq R_{k,2}}a_{n}(f)\varphi_{n}}\right\|}^{2}_{L^{2}(I,w)}$ $\displaystyle\leq$ $\displaystyle C2^{-2k\delta}R^{(2\frac{\gamma(p^{\prime})}{\varepsilon})}\left(\frac{3R}{2^{k+1}}\right){\left\|{f}\right\|}^{2}_{L^{p}(I,w)}.$ Hence, we have ${\left\|{\Psi_{R,k}^{\delta}f}\right\|}_{L^{2}(I,w)}\leq C2^{-k(\delta+\frac{1}{2})}R^{(\frac{\gamma(p^{\prime})}{\varepsilon})+\frac{1}{2})}{\left\|{f}\right\|}_{L^{p}(I,w)}.$ (20) By combining Hölder inequality and (20), we obtain $\displaystyle{\left\|{\Psi_{R,k}^{\delta}f}\right\|}_{L^{p}(I_{k,1}^{\alpha},w)}$ $\displaystyle\leq$ $\displaystyle(\mu(I_{k,1}))^{\frac{1}{p}-\frac{1}{2}}{\left\|{\Psi_{R,k}^{\delta}f}\right\|}_{L^{2}(I_{k,1}^{\alpha},w)}$ (21) $\displaystyle\leq$ $\displaystyle(2r_{k}^{\alpha})^{\frac{1}{p}-\frac{1}{2}}{\left\|{\Psi_{R,k}^{\delta}f}\right\|}_{L^{2}(I,w)}$ $\displaystyle\leq$ $\displaystyle C2^{-k(\delta-\frac{\gamma(p^{\prime})}{\varepsilon})}{\left\|{f}\right\|}_{L^{p}(I,w)}.$ Let $\displaystyle s_{R,k}^{\delta}(u,v)=\sum_{n=0}^{\infty}\phi_{R,k}^{\delta}(\lambda_{n})\varphi_{n}(x)\varphi_{n}(y)$ be the kernel of $\Psi_{R,k}^{\delta}$. We just have to find an estimate of $||\Psi_{R,k}^{\delta}f||_{L^{p}(I_{k,2}^{\alpha},w)}$, so we will use the Schur test with the symmetric property of $s_{R,k}^{\delta},$ then it suffices to prove the following inequality $\sup_{u\in I_{k,2}^{\alpha}}{\left\|{s_{R,k}^{\delta}(u,.)}\right\|}_{L^{1}(I_{k,2}^{\alpha})}\leq C2^{-k\varepsilon}$ for some $\varepsilon>0$ and $C>0$ depending only on $p.$ We consider $g_{R,k}^{\delta}(\lambda)=\left(1-\frac{\lambda^{2}}{R}\right)_{+}^{\delta}e^{\lambda^{2}/R}\phi(2^{k}(1-\frac{\lambda^{2}}{R}))$ satisfying the following properties, see [8] 1. 1. For every non-negative integer $i$ there exists a constant $C_{i}$ such that for all $s>0$ $\int_{|t|\geq s}|\hat{g}_{R,k}^{\delta}(t)|dt\leq C_{i}s^{-i}R^{-i/2}2^{(i-\delta)k}$ (22) 2. 2. ${\left\|{g_{R,k}^{\delta}(\sqrt{\mathcal{L}})}\right\|}_{(L^{2},L^{2})}\leq C2^{-k\delta}.$ (23) For our purpose, we will consider such a positive self-adjoint operator $\mathcal{L}$ on $L^{2}(\mathbb{R})$ such that the semigroup $e^{-t\mathcal{L}}$, generated by $-\mathcal{L}$, has the kernel $p_{t}(x,y)$ obeying the Gaussian upper bound $|p_{t}(u,v)|\leq\frac{C}{\sqrt{t}}\exp{\left(-\frac{|u-v|^{2}}{Ct}\right)}.$ (24) for a constant $C>0$. (see [14]) For all $u\in\mathbb{R}$ and $t>0$, one gets the following estimate ${\left\|{p_{t}(u,.)}\right\|}_{L^{2}(\mathbb{R})}\leq C.$ (25) On the other hand, there exists $i_{0}\in\mathbb{N}$ such that $2^{i_{o}-1}<R^{\mu(p)}<2^{i_{0}}$ and we can see that $I_{k,2}^{\alpha}\subseteq\displaystyle\cup_{\mu(p)k-i_{0}\leq j\leq 0}D_{j}$ where $D_{j}=\\{y,2^{j}r\leq|y-x_{0}|<2^{j+1}r\\}.$ Since, $\mathcal{L}$ is a positive self-adjoint operator, then it’s clear that $\phi_{R,k}^{\delta}(\mathcal{L})=g_{R,k}^{\delta}(\sqrt{\mathcal{L}})\exp{\left(-\mathcal{L}/R\right)}.$ (26) Hence one gets $\displaystyle s_{R,k}^{\delta}(u,v)$ $\displaystyle=$ $\displaystyle g_{R,k}^{\delta}(\sqrt{\mathcal{L}})\left(p_{1/R}(u,.)\right)(v)$ $\displaystyle=$ $\displaystyle g_{R,k}^{\delta}(\sqrt{\mathcal{L}})\left(p_{1/R}(u,.)\chi_{\\{w,|x_{0}-w|<2^{j-1}r\\}}\right)(v)+g_{R,k}^{\delta}(\sqrt{\mathcal{L}})\left(p_{1/R}(u,.)\chi_{\\{w,|x_{0}-w|\geq 2^{j-1}r\\}}\right)(v)$ $\displaystyle=$ $\displaystyle s_{R,k}^{\delta,1}(u,v)+s_{R,k}^{\delta,2}(u,v).$ Using the fact that $g_{R,k}^{\delta}$ is an even function, with the inversion formula, we have $g_{R,k}^{\delta}(\sqrt{\lambda})=\frac{1}{\sqrt{2\pi}}\int_{\mathbb{R}}\hat{g}_{R,k}^{\delta}(t)\cos{(t\sqrt{\lambda})}dt.$ Hence, we obtain $\displaystyle s_{R,k}^{\delta,1}(u,v)$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2\pi}}\int_{\mathbb{R}}\hat{g}_{R,k}^{\delta}(t)\cos{(t\sqrt{\mathcal{L}})}\left(p_{1/R}(u,.)\chi_{\\{w,|x_{0}-w|<2^{j-1}r\\}}\right)(v)dt.$ Moreover, the operator $\cos{(t\sqrt{\mathcal{L}})}$ is bounded in $L^{2}$ with support kernel $\mathcal{K}_{t}$ satisfying, see [14, 28] $\mbox{Supp}\left(\mathcal{K}_{t}\right)=\\{(u,v)\in\mathbb{R}^{2},|u-v|\leq c_{0}|t|\\}$ From (23), (25) and the previous analysis, one gets $\displaystyle{\left\|{s_{R,k}^{\delta,1}(u,.)}\right\|}_{L^{1}(D_{j})}$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2\pi}}{\left\|{\int_{\mathbb{R}}\hat{g}_{R,k}^{\delta}(t)\cos{(t\sqrt{\mathcal{L}})}\left(p_{1/R}(u,.)\chi_{\\{w,|x_{0}-w|<2^{j-1}r\\}}\right)dt}\right\|}_{L^{1}(D_{j})}$ $\displaystyle=$ $\displaystyle\frac{1}{\sqrt{2\pi}}{\left\|{\int_{|t|>\frac{2^{j-1}r}{c_{0}}}\hat{g}_{R,k}^{\delta}(t)\cos{(t\sqrt{\mathcal{L}})}\left(p_{1/R}(u,.)\chi_{\\{w,|x_{0}-w|<2^{j-1}r\\}}\right)dt}\right\|}_{L^{1}(D_{j})}$ $\displaystyle\leq$ $\displaystyle\frac{\mu^{1/2}(D_{j})}{\sqrt{2\pi}}\int_{|t|>\frac{2^{j-1}r}{c_{0}}}|\hat{g}_{R,k}^{\delta}(t)|{\left\|{p_{1/R}(u,.)}\right\|}_{L^{2}(D_{j})}dt$ $\displaystyle\leq$ $\displaystyle\frac{C}{\sqrt{2\pi}}2^{\frac{j+1}{2}}\int_{|t|>\frac{2^{j-1}r}{c_{0}}}|\hat{g}_{R,k}^{\delta}(t)|dt$ Let $i>\frac{\mu+\frac{1}{2}}{2(\mu+1-\frac{1}{p})}$ where $\mu=\frac{\gamma(p^{\prime})}{\varepsilon}>0.$ Then by (22), there exists a constant $C_{i}>0$ such that $\displaystyle{\left\|{s_{R,k}^{\delta,1}(u,.)}\right\|}_{L^{1}(D_{j})}$ $\displaystyle\leq$ $\displaystyle\frac{C_{i}}{\sqrt{2\pi}}2^{j/2}(\frac{2^{j}}{c_{0}})^{-i}R^{-i/2}2^{(i-\delta)k}$ $\displaystyle\leq$ $\displaystyle\frac{C_{i}}{\sqrt{2\pi}}c_{0}^{i}2^{(i-\delta)k}2^{j(1/2-i)}.$ Then, we obtain $\displaystyle{\left\|{s_{R,k}^{\delta,1}(u,.)}\right\|}_{L^{1}(I_{k,2})}$ $\displaystyle\leq$ $\displaystyle\sum_{\mu(p)k-i_{0}\leq j\leq 0}{\left\|{s_{R,k}^{\delta,1}(u,.)}\right\|}_{L^{1}(D_{j})}$ $\displaystyle\leq$ $\displaystyle\frac{C_{i}}{\sqrt{2\pi}}c_{0}^{i}2^{(i-\delta)k}\sum_{\mu(p)k-i_{0}\leq j\leq 0}2^{j(1/2-i)}$ $\displaystyle\leq$ $\displaystyle\frac{C_{i}}{\sqrt{2\pi}}c_{0}^{i}2^{(i-\delta)k}2^{(i-1/2)(i_{0}-\mu(p)k+1)}$ $\displaystyle\leq$ $\displaystyle C^{\prime}_{i}2^{-k\varepsilon_{1}}.$ From our assumption on $i,$ $\varepsilon_{1}=\delta-i+(i-1/2)(\frac{\mu+\frac{1}{2}}{(\frac{1}{p}-\frac{1}{2})}))>0.$ Then, to have an estimate of the kernel $s_{R,k}^{\delta,1}$ on $L^{1}(I_{k,2}^{\delta})$, it suffices to find an estimate of the kernel $s_{R,k}^{\delta,2}$ on $L^{1}(I_{k,2}^{\delta})$. From (23), (24) and using the fact that $R\leq R^{\mu(p)}$, one gets the following inequality $\displaystyle{\left\|{s_{R,k}^{\delta,2}(u,.)}\right\|}_{L^{1}(D_{j})}$ $\displaystyle=$ $\displaystyle\int_{D_{j}}|g_{R,k}^{\delta}(\sqrt{\mathcal{L}})\left(p_{1/R}(u,.)\chi_{\\{w,|w-x_{0}|>2^{j-1}r\\}}\right)(v)|dv$ $\displaystyle\leq$ $\displaystyle{\left\|{g_{R,k}^{\delta}(\sqrt{\mathcal{L}})}\right\|}_{(L^{2},L^{2})}{\left\|{p_{1/R}(u,.)\chi_{\\{w,|w-x_{0}|>2^{j-1}r\\}}}\right\|}_{L^{2}(D_{j})}$ $\displaystyle\leq$ $\displaystyle C2^{-k\delta}\sqrt{R}e^{(-CR2^{2j-2})}\left(\mu(D_{j})\right)^{1/2}$ $\displaystyle\leq$ $\displaystyle C2^{-k\delta}2^{\frac{i_{0}+j}{2}}e^{-C2^{2(i_{0}+j)}}.$ Hence, we conclude that $\displaystyle{\left\|{s_{R,k}^{\delta,2}(u,.)}\right\|}_{L^{1}(I_{k,2})}$ $\displaystyle\leq$ $\displaystyle\sum_{\mu(p)k-i_{0}\leq j\leq 0}{\left\|{s_{R,k}^{\delta,2}(u,.)}\right\|}_{L^{1}(D_{j})}$ $\displaystyle\leq$ $\displaystyle C2^{-k\delta}\sum_{i=i_{o}+j\geq\mu(p)k}2^{\frac{i}{2}}e^{-C2^{2i}}$ $\displaystyle\leq$ $\displaystyle C^{\prime}2^{-k\delta}.$ ∎ ###### Proposition 2. Let $1\leq p\leq 2$ and $\delta>\delta(p)=\frac{\gamma(p^{\prime})}{\varepsilon}$, then for all $f\in L^{p}(I,w)$, we have ${\left\|{\psi_{R,0}^{\delta}f}\right\|}_{L^{p}(I,w)}\leq C{\left\|{f}\right\|}_{L^{p}(I,w)}.$ (27) where $C$ is a constant independent of $f$ and $R.$ ###### Proof. It suffices to use the same techniques as those used in the previous proof to get an estimate of ${\left\|{\psi_{R,0}^{\delta}f}\right\|}_{L^{p}(I_{1},w)}$ and ${\left\|{\psi_{R,0}^{\delta}f}\right\|}_{L^{p}(I_{2},w)}$ for all $f\in L^{p}(I,w),$ where $I=(a,b)=I_{1}\cup I_{2}$ with $I_{1}=(x_{0}-r^{\alpha}_{0},x_{0}+r^{\alpha}_{0})$ and $I_{2}=\\{y,|y-x_{0}|>r^{\alpha}_{0}\\}$ where $r^{\alpha}_{0}=\frac{r}{R^{\mu(p)}}.$ ∎ To conclude the theorem’s proof it suffices to find a uniform bound of $\mathcal{R}_{R}^{\delta}$. ###### Proposition 3. Let $1\leq p\leq 2$ and $\delta>\delta(p)=\frac{\gamma(p^{\prime})}{\varepsilon}$, then for all $f\in L^{p}(I,w)$, we have ${\left\|{\mathcal{R}_{R}^{\delta}f}\right\|}_{L^{p}(I,w)}\leq C{\left\|{f}\right\|}_{L^{p}(I,w)}.$ (28) where $C$ depends only on $p$. ###### Proof. From Holder’s inequality and the previous lemma, we have $\displaystyle{\left\|{\mathcal{R}_{R}^{\delta}f}\right\|}^{2}_{L^{p}(I,w)}$ $\displaystyle\leq$ $\displaystyle 2^{2(\frac{1}{p}-\frac{1}{2})}{\left\|{\mathcal{R}_{R}^{\delta}f}\right\|}^{2}_{L^{2}(I,w)}$ $\displaystyle\leq$ $\displaystyle 2^{2(\frac{1}{p}-\frac{1}{2})}\sum_{k=K_{R}+1}^{\infty}\sum_{n=0}^{\infty}{\left\|{\phi_{R,k}^{\delta}(\lambda_{n})a_{n}(f)\varphi_{n}}\right\|}_{L^{2}(I,w)}$ $\displaystyle\leq$ $\displaystyle C2^{2(\frac{1}{p}-\frac{1}{2})}\sum_{k=K_{R}+1}^{\infty}2^{-2k\delta}\sum_{R_{k,1}\leq\lambda_{n}\leq R_{k,2}}{\left\|{a_{n}(f)\varphi_{n}}\right\|}^{2}_{L^{2}(I,w)}$ $\displaystyle\leq$ $\displaystyle C2^{2(\frac{1}{p}-\frac{1}{2})}\sum_{k=K_{R}+1}^{\infty}2^{-2k(\delta+\frac{1}{2})}R^{2(\frac{1}{2}+\frac{\gamma(p^{\prime})}{\varepsilon})}{\left\|{f}\right\|}^{2}_{L^{p}(I,\omega)}$ $\displaystyle\leq$ $\displaystyle C2^{2(\frac{1}{p}-\frac{1}{2})}2^{-2(\delta+\frac{1}{2})\big{(}\left[\frac{\log(R)}{\log(2)}\right]+1\big{)}}R^{2(\frac{1}{2}+\frac{\gamma(p^{\prime})}{\varepsilon})}{\left\|{f}\right\|}^{2}_{L^{p}(I,\omega)}$ $\displaystyle\leq$ $\displaystyle C2^{2(\frac{1}{p}-\frac{1}{2})}R^{-2(\delta-(\frac{\gamma(p^{\prime})}{\varepsilon}))}{\left\|{f}\right\|}^{2}_{L^{p}(I,\omega)}$ Finally we obtain $\displaystyle{\left\|{\mathcal{R}_{R}^{\delta}f}\right\|}_{L^{p}(I,w_{\alpha,\beta})}$ $\displaystyle\leq$ $\displaystyle C2^{(\frac{1}{p}-\frac{1}{2})}R^{-(\delta-(\frac{\gamma(p^{\prime})}{\varepsilon}))}{\left\|{f}\right\|}_{L^{p}(I,\omega_{\alpha,\beta})}$ $\displaystyle\leq$ $\displaystyle C(p){\left\|{f}\right\|}_{L^{p}(I,\omega_{\alpha,\beta})}.$ ∎ ∎ ###### Corollary 1. Under the notation and conditions of the previous Theorem, we have for all $f\in L^{p}(I,w)$ $\Psi_{R}^{\delta}f\to f~{}~{}\mbox{as}~{}~{}R\to\infty.$ (29) ###### Proof. Step1: We prove that, for every $f\in\mathcal{C}^{\infty}(I,\mathbb{R})$, $\Psi^{\delta}_{R}f\to f$ in $L^{p}(I,\omega)$. Note that $\displaystyle\Big{|}\Big{(}1-\frac{\lambda_{n}}{R}\Big{)}^{\delta}_{+}{\left\langle{f,\varphi_{n}}\right\rangle}_{L^{2}(I,\omega)}\Big{|}$ $\displaystyle\leq$ $\displaystyle\Big{|}{\left\langle{f,\varphi_{n}}\right\rangle}_{L^{2}(I,\omega)}\Big{|}=\frac{1}{\lambda_{n}}\Big{|}{\left\langle{f,\mathcal{L}\varphi_{n}}\right\rangle}_{L^{2}(I,\omega)}\Big{|}$ (30) $\displaystyle=$ $\displaystyle\frac{1}{\lambda_{n}}\Big{|}{\left\langle{\mathcal{L}f,\varphi_{n}}\right\rangle}_{L^{2}(I,\omega)}\Big{|}=\cdots=\frac{1}{\lambda_{n}^{k}}\Big{|}{\left\langle{\mathcal{L}^{k}.f,\varphi_{n}}\right\rangle}_{L^{2}(I,\omega)}\Big{|}$ $\displaystyle\leq$ $\displaystyle n^{-k\varepsilon}{\left\|{\mathcal{L}^{k}.f}\right\|}_{L^{2}(I,\omega)}.$ Since ${\left\|{\varphi_{n}}\right\|}_{L^{2}(I,\omega)}\leq n^{\gamma(p)}$, it suffices to take $k$ big enough to have $\gamma(p)-k\varepsilon<-1$ and obtain the convergence of the series in $L^{p}(I,\omega)$. Since $\displaystyle{\left\|{\Psi^{\delta}_{R}.f-f}\right\|}_{2}^{2}=\sum_{n=0}^{\infty}\Big{(}(1-\frac{\lambda_{n}}{R})^{\delta}_{+}-1\Big{)}^{2}|a_{n}(f)|^{2}\to 0$ as $R\to\infty$,then the result remains true for $1\leq p<\infty$. Step2: For all $\varepsilon>0$. By density of $\mathcal{C}^{\infty}_{0}(I,\mathbb{R})$ in $L^{p}(I,\omega)$, there exists $g\in\mathcal{C}^{\infty}_{0}(I,\mathbb{R})$ such that ${\left\|{f-g}\right\|}_{L^{p}(I,\omega)}<\varepsilon$ and there exists $R>0$ such that ${\left\|{\Psi^{\delta}_{R}.f-\Psi^{\delta}_{R}.g}\right\|}_{L^{p}(I,\omega)}<\varepsilon$. By writing, ${\left\|{\Psi^{\delta}_{R}.f-f}\right\|}_{L^{p}(I,\omega)}\leq{\left\|{\Psi^{\delta}_{R}.f-\Psi^{\delta}_{R}.g}\right\|}_{L^{p}(I,\omega)}+{\left\|{\Psi^{\delta}_{R}.g-g}\right\|}_{L^{p}(I,\omega)}+{\left\|{f-g}\right\|}_{L^{p}(I,\omega)},$ one gets the desired result. ∎ To conclude for the proof of sufficient conditions of both theorems 1 and 2, it suffices to verify that the two considered bases satisfy conditions (A) and (B).We will prove this result only for the case of GPSWFs. The other case is almost identical. We first recall that from (13), the GPSWFs are the eigenfunctions of the Sturm-Liouville operator $\mathcal{L}_{c}^{(\alpha)}.$ Also, note that the $(n+1)-$th eigenvalue $\chi_{n,\alpha}(c)$ of $\mathcal{L}_{c}^{(\alpha)}$ satisfies the following classical inequalities, $n^{2}\leq n(n+2\alpha+1)\leq\chi_{n,\alpha}(c)\leq n(n+2\alpha+1)+c^{2},\quad\forall n\geq 0.$ Moreover, for every $0\leq m<M$ such that $M-m>1$, we have $\displaystyle\sum_{\chi_{n,\alpha}(c)\in(m,M)}1$ $\displaystyle\leq$ $\displaystyle\sum_{n(n+2\alpha+1)\in(\max(0,m-c^{2}),M)}1$ $\displaystyle\leq$ $\displaystyle\sum_{(n+\alpha+1/2)^{2}-(\alpha+1/2)^{2}\in(\max(0,m-c^{2}),M)}1$ $\displaystyle\leq$ $\displaystyle\sum_{n\in\left((\max(0,m-c^{2})+(\alpha+1/2)^{2})^{\frac{1}{2}}-(1/2+\alpha),(M+(\alpha+1/2)^{2})^{\frac{1}{2}}-(1/2+\alpha)\right)}1$ $\displaystyle\leq$ $\displaystyle C(M-m).$ It follows that condition (B) is satisfied. Form [6] Lemma $2.6$, one can conclude that condition (A) is satisfied for weighted prolate spheroidal wave functions for $1<p<\infty$. Moreover, it has been shown in [18] that ${\left\|{\psi^{(\alpha)}_{n,c}}\right\|}_{\infty}\leq C\Big{(}\chi_{n,\alpha}(c)\Big{)}^{\frac{\alpha+1}{2}}.$ Then, by using (14), we obtain ${\left\|{\psi^{(\alpha)}_{n,c}}\right\|}_{1}\leq C\Big{(}\chi_{n,\alpha}(c)\Big{)}^{\frac{\alpha+1}{2}}\leq Cn^{\alpha+1}.$ ###### Remark 3. The uniform norm of the CPSWFs has been given in [5]. ## 5 Proof of necessary condition The transferring theorem from the uniform boundedness of $\Psi_{R}^{\delta}$ to the uniform boundedness of the Hankel multiplier transform operator $\mathcal{M}_{\alpha}$ defined by $\mathcal{M}_{\alpha}(f)=\mathcal{H}_{\alpha}\left(\phi(.)\mathcal{H}_{\alpha}(f)\right)$ can be used to derive necessary condition. Note here that $\phi$ is a bounded function on $\mathbb{R}$, continuous except on a set of Lebesgue measure zero and $\mathcal{H}_{\alpha}$ is the modified Hankel operator defined by $\mathcal{H}_{\alpha}(f)(x)=\int_{0}^{\infty}\frac{J_{\alpha}(xy)}{(xy)^{\alpha}}f(y)y^{2\alpha+1}dy.$ From [12] and the transferring theorem, the uniform boundedness of $\Psi_{R}^{\delta}$ holds true if and only if $\delta>\max\\{2(\alpha+1)|\frac{1}{p}-\frac{1}{2}|-\frac{1}{2},0\\}.$ It’s easy to check that $\max\\{2(\alpha+1)|\frac{1}{p}-\frac{1}{2}|-\frac{1}{2},0\\}\geq\max\\{\frac{\gamma_{\alpha}(p^{\prime})}{2},0\\}$ for every $p\not=2-\frac{1}{\alpha+3/2}$, then one gets our necessary condition. To be more precise, let’s study each transferring theorem separately. ### 5.1 GPSWFs’s case Let’s recall that the family of weighted prolate spheroidal wave functions $\\{\psi_{n,c}^{(\alpha)}(\cos\theta)\\}_{n}$ form an orthonormal system on $(0,\pi)$ with respect to the measure $(\sin\theta)^{2\alpha+1}d\theta$. For a function $f(\theta)$ integrable on $(0,\pi)$ with respect to the measure defined above, we have formally $f(\theta)=\sum_{n=0}^{\infty}a_{n}(f)\psi_{n,c}^{(\alpha)}(\cos\theta)\qquad a_{n}(f)=\int_{0}^{\pi}f(\theta)\psi_{n,c}^{(\alpha)}(\cos\theta)(\sin\theta)^{2\alpha+1}d\theta$ For $p\geq 1$ and a function $f$ on $(0,\pi)$ we define a norm ${\left\|{f}\right\|}_{p}=\Bigg{(}\int_{0}^{\pi}|f(\theta)|^{p}(\sin\theta)^{2\alpha+1}\Bigg{)}^{1/p}.$ Before stating an adequate transferring theorem, let’s define a GPSWFs- multiplier. ###### Definition 1. Let $\lambda>0$ be a sufficiently large real, the bounded sequence $\\{\phi(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})\\}_{n}$ is called a Weighted prolate multiplier if there exist a constant $C>0$ such that for every $f\in L^{p}(I,\omega_{\alpha})$, we have ${\left\|{\sum_{n=0}^{\infty}\phi(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})a_{n}(f)\psi_{n,c}^{(\alpha)}}\right\|}_{L^{p}(I,\omega_{\alpha})}\leq C{\left\|{f}\right\|}_{L^{p}(I,\omega_{\alpha})}.$ The smallest constant $C$ verifying this last inequality is written ${\left\|{\phi(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})}\right\|}_{p}$. In the same context, the function $\phi$ is called an $\L^{p}$-Hankel transform multiplier if $\mathcal{M}_{\alpha}(f)=\mathcal{H}_{\alpha}(\phi(.)\mathcal{H}_{\alpha}(f))$ is uniformly bounded on $L^{p}\left((0,\infty),\theta^{2\alpha+1}d\theta\right)$. ###### Theorem 4 (Transferring theorem). Let $1<p<\infty$, $0\leq\alpha<3/2$ and $\phi$ be a bounded function on $(0,\infty)$ continuous except on a set of Lebesgue measure zero such that $\\{\phi(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})\\}_{n}$ is a Weighted prolate multiplier for all large $\lambda>0$ and $\displaystyle\liminf_{\lambda\to\infty}{\left\|{\phi(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})}\right\|}_{p}$ is finite then $\phi$ is an $L^{p}$-Hankel transform multiplier and we have ${\left\|{\mathcal{M}_{\alpha}}\right\|}_{p}\leq\liminf_{\lambda\to\infty}{\left\|{\phi\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)}\right\|}_{p}.$ ###### Proof. Let $g$ be an infinitely differentiable function with compact support in $[0,M]$ and put $g_{\lambda}(\theta)=g(\lambda\theta)$. Here $\lambda$ is a positive real so that $supp(g_{\lambda})\subset[0,\pi]$. Recall that we have by assumption ${\left\|{\sum_{n=0}^{\infty}\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)a_{n}(g)\psi_{n,c}^{(\alpha)}(\cos(.))}\right\|}_{p}\leq{\left\|{\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)}\right\|}_{p}{\left\|{g}\right\|}_{p}.$ (31) Via a simple change of variable, one can write $\lim_{\lambda\to\infty}\lambda^{2\alpha+2}{\left\|{g_{\lambda}}\right\|}_{p}^{p}=\lim_{\lambda\to\infty}\int_{0}^{M}|g(\tau)|^{p}\Big{(}\lambda\sin(\tau/\lambda)\Big{)}^{2\alpha+1}d\tau=\int_{0}^{\infty}|g(\tau)|^{p}\tau^{2\alpha+1}d\tau.$ By using (31) together with Fatou’s lemma, one gets $\displaystyle\displaystyle\int_{0}^{\infty}$ $\displaystyle\liminf_{\lambda\to\infty}\Big{|}\chi_{(0,\pi\lambda)}(\tau)\sum_{n=0}^{\infty}\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)a_{n}(g_{\lambda})\psi_{n,c}^{(\alpha)}(\cos\tau/\lambda)\Big{|}^{p}\tau^{2\alpha+1}d\tau$ $\displaystyle=$ $\displaystyle\int_{0}^{\infty}\liminf_{\lambda\to\infty}\Big{|}\chi_{(0,\pi\lambda)}(\tau)\sum_{n=0}^{\infty}\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)a_{n}(g_{\lambda})\psi_{n,c}^{(\alpha)}(\cos\tau/\lambda)\Big{|}^{p}\lambda^{2\alpha+1}\sin(\tau/\lambda)^{2\alpha+1}d\tau$ $\displaystyle\leq$ $\displaystyle\liminf_{\lambda\to\infty}\lambda^{2\alpha+1}\int_{0}^{\infty}\Big{|}\chi_{(0,\pi\lambda)}(\tau)\sum_{n=0}^{\infty}\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)a_{n}(g_{\lambda})\psi_{n,c}^{(\alpha)}(\cos\tau/\lambda)\Big{|}^{p}\sin(\tau/\lambda)^{2\alpha+1}d\tau$ $\displaystyle\leq$ $\displaystyle\liminf_{\lambda\to\infty}\lambda^{2\alpha+2}{\left\|{\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)}\right\|}_{p}{\left\|{g_{\lambda}}\right\|}^{p}_{p}=\liminf_{\lambda\to\infty}{\left\|{\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)}\right\|}_{p}\Bigg{[}\int_{0}^{\infty}|g(\tau)|^{p}\tau^{2\alpha+1}d\tau\Bigg{]}.$ Then there exists a sequence $\lambda_{1}<\lambda_{2}<\cdots<\lambda_{p}\to\infty$ that $G(\tau,\lambda)=\displaystyle\chi_{(0,\pi\lambda)}(\tau)\sum_{n=0}^{\infty}\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)a_{n}(g_{\lambda})\psi_{n,c}^{(\alpha)}(\cos\frac{\tau}{\lambda})$ converges weakly to a function $G(\tau)$. Furthermore, $G$ satisfies $\Bigg{[}\int_{0}^{\infty}|G(\tau)|^{p}\tau^{2\alpha+1}d\tau\Bigg{]}^{1/p}\leq\liminf_{\lambda\to\infty}{\left\|{\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)}\right\|}_{p}\Bigg{[}\int_{0}^{\infty}|g(\tau)|^{p}\tau^{2\alpha+1}d\tau\Bigg{]}^{1/p}.$ Let us now prove that $G=\mathcal{H}_{\alpha}(\phi.\mathcal{H}_{\alpha}(g))$. Let $G(\tau,\lambda)=\chi_{(0,\pi\lambda)}(\tau)\Big{[}\sum_{n=0}^{N[\lambda]}+\sum_{N[\lambda]+1}^{\infty}\Big{]}\phi(\chi^{1/2}_{n,\alpha}(c)/\lambda)a_{n}(g_{\lambda})\psi_{n,c}^{(\alpha)}(\cos\tau/\lambda)=G^{N}(\tau,\lambda)+H^{N}(\tau,\lambda)$ We start by giving the following lemma that will be proved later, ###### Lemma 2. We have $\int_{0}^{\infty}\Big{|}H^{N}(\tau,\lambda)\Big{|}^{2}\tau^{2\alpha+1}d\tau=O(\frac{1}{N})\mbox{ uniformly in }\lambda$ Therefore, by the diagonal argument, there exists a subsequence also noted $\\{\lambda_{j}\\}$ for a sake of clarity, such that $H^{N}(\tau,\lambda_{j})$ converges weakly to a function $H^{N}(\tau)$ and $\displaystyle\int_{0}^{\pi}\Big{|}H^{N}(\tau)\Big{|}^{p}\tau^{2\alpha+1}d\tau=O(\frac{1}{N^{2}}).$ Then, there exists a subsequence $H^{N_{j}}$ denoted for the same reason $H^{N}$ that converges to zero a.e . Since $G^{N}(\tau,\lambda)=G(\tau,\lambda)-H^{N}(\tau,\lambda)$, $G^{N}(\tau,\lambda)$ converge weakly to a limit $G^{N}(\tau)$ and $G(\tau)=G^{N}(\tau)+H^{N}(\tau)$. Thus $G^{N}(\tau)$ converges to $G(\tau)$ almost everywhere. On the other hand, we will prove the following lemma ###### Lemma 3. We have $\lim_{\lambda\to\infty}G^{N}(\tau,\lambda)=\int_{0}^{N}\phi(v)\mathcal{H}_{\alpha}.g(v)\frac{J_{\alpha}(v\tau)}{(v\tau)^{\alpha}}v^{2\alpha+1}dv,$ which implies that $G(\tau)=\int_{0}^{\infty}\phi(v)\mathcal{H}_{\alpha}.g(v)\frac{J_{\alpha}(v\tau)}{(v\tau)^{\alpha}}v^{2\alpha+1}dv,$ and achieves our proof. ∎ ###### Proof of Lemma 2. We have $\displaystyle\int_{0}^{M}|H^{N}(\tau,\lambda)|^{2}\Big{(}\lambda\sin\frac{\tau}{\lambda}\Big{)}^{2\alpha+1}d\tau$ $\displaystyle=$ $\displaystyle\lambda^{2\alpha+2}\int_{0}^{\pi}|H^{N}(\lambda\tau,\lambda)|^{2}(\sin\tau)^{2\alpha+1}d\tau$ (32) $\displaystyle=$ $\displaystyle\lambda^{2\alpha+2}\sum_{N[\lambda]+1}^{\infty}|\phi(\frac{n}{\lambda})|^{2}|a_{n}(g_{\lambda})|^{2}.$ Recall that in [19], authors have given the following uniform approximation of GPSWFs in term of Jacobi polynomials for $0\leq\alpha<3/2$, $\psi_{n,c}^{(\alpha)}(\cos\theta)=A_{n}\widetilde{P}_{n}^{(\alpha)}(\cos\theta)+R_{n,c}^{(\alpha)}(\cos\theta)\qquad{\left\|{R_{n,c}}\right\|}^{(\alpha)}_{\infty}\leq C_{\alpha,c}\frac{1}{2n+2\alpha+1}.$ (33) We also know that (see for example [29]) $n(\sin\theta)^{2\alpha+1}\widetilde{P}_{n}^{(\alpha,\alpha)}(\cos\theta)=2\frac{h^{\alpha+1}_{n-1}}{h^{(\alpha)}_{n}}\frac{d}{d\theta}\Big{[}(\sin\theta)^{2\alpha+2}\widetilde{P}_{n-1}^{(\alpha+1,\alpha+1)}(\cos\theta)\Big{]}$ (34) By combining (33) and (34), one gets $(\sin\theta)^{2\alpha+1}\psi_{n,c}^{(\alpha)}(\cos\theta)=\frac{2}{n}\frac{h^{\alpha+1}_{n-1}}{h^{(\alpha)}_{n}}\frac{d}{d\theta}\Big{[}(\sin\theta)^{2\alpha+2}\widetilde{P}_{n-1}^{(\alpha+1,\alpha+1)}(\cos\theta)\Big{]}+R_{n,c}^{(\alpha)}(\cos\theta).$ Then, integrating by parts one gets $\displaystyle a_{n}(g_{\lambda})$ $\displaystyle=$ $\displaystyle\frac{C}{n}\int_{0}^{\pi}\frac{g^{\prime}(\lambda\theta)}{\sin\theta}\widetilde{P}_{n-1}^{\alpha+1}(\cos\theta)(\sin\theta)^{2\alpha+3}d\theta+\int_{0}^{\pi}R_{n,c}^{(\alpha)}(\cos\theta)g(\lambda\theta)d\theta$ $\displaystyle=$ $\displaystyle a_{n,1}(g_{\lambda})+a_{n,2}(g_{\lambda})$ Let’s come back to (32). We have by Bessel’s inequality $\displaystyle\lambda^{2\alpha+2}\sum_{N[\lambda]+1}^{\infty}|\phi(\frac{n}{\lambda})|^{2}|a_{n,1}(g_{\lambda})|^{2}$ $\displaystyle\leq$ $\displaystyle C\lambda^{2\alpha+2}\Big{[}\frac{\lambda}{N(\lambda-1)}\Big{]}^{2}\sum_{N[\lambda]+1}^{\infty}|\frac{n}{\lambda}a_{n,1}(g_{\lambda})|^{2}$ (35) $\displaystyle\leq$ $\displaystyle\frac{C}{N^{2}}\lambda^{2\alpha+2}\int_{0}^{\pi}\Big{|}\frac{g^{\prime}(\lambda\theta)}{\sin\theta}\Big{|}^{2}(\sin\theta)^{2\alpha+3}d\theta$ $\displaystyle=$ $\displaystyle\frac{C}{N^{2}}\int_{0}^{M}|g^{\prime}(\theta)|^{2}\Big{(}\lambda\sin\frac{\theta}{\lambda}\Big{)}^{2\alpha+1}d\theta$ $\displaystyle=$ $\displaystyle O(\frac{1}{N^{2}})\mbox{ uniformly in }\lambda.$ On the other hand, using Cauchy-Schwarz’s inequality $\displaystyle\lambda^{2\alpha+2}\sum_{N[\lambda]+1}^{\infty}|\phi(\frac{n}{\lambda})|^{2}|a_{n,2}(g_{\lambda})|^{2}$ $\displaystyle\leq$ $\displaystyle C\lambda^{2\alpha+2}\sum_{N[\lambda]+1}^{\infty}{\left\|{R_{n,c}^{\alpha}}\right\|}^{2}_{2}{\left\|{g(\lambda.)}\right\|}_{2}^{2}$ $\displaystyle\leq$ $\displaystyle C\sum_{N[\lambda]+1}^{\infty}\frac{1}{n^{2}}\int_{0}^{M}|g(\theta)|^{2}\big{(}\lambda\sin\frac{\theta}{\lambda}\big{)}^{2\alpha+1}d\theta$ $\displaystyle=$ $\displaystyle O(\frac{1}{N}).$ Then, one conclude that $\int_{0}^{M}|H^{N}(\tau,\lambda)|^{2}\tau^{2\alpha+1}d\tau=O(\frac{1}{N})\mbox{ uniformly in }\lambda.$ ∎ ###### Proof of lemma 3. We use now the following uniform approximation of GPSWFs in term of Bessel function (we refer the reader once again to [19]) $\psi_{n,c}^{(\alpha)}(\cos\frac{\tau}{\lambda})=A_{\alpha}(q)\frac{\chi_{n,c}^{1/4}S(\cos\frac{\tau}{\lambda})^{1/2}J_{\alpha}(\chi_{n,c}^{1/2}S(\cos\frac{\tau}{\lambda}))}{(\sin\frac{\tau}{\lambda})^{\alpha+1/2}(1-q\cos^{2}\frac{\tau}{\lambda})^{1/4}}+E_{n,c}(\cos\frac{\tau}{\lambda}),$ (36) where $\Big{|}E_{n,c}(\cos\theta)\Big{|}\leq\frac{C.A_{\alpha}(q)}{(1-q)}\frac{(\sin\theta)^{1/2}}{(1-q\cos^{2}\theta)^{1/4}}\qquad\forall\theta\in[0,\pi]\quad\mbox{and}\quad S(x)=\int_{x}^{1}\sqrt{\frac{1-qt^{2}}{1-t^{2}}}dt.$ Note that it has also been shown in [4] that $\frac{\sin\theta\sqrt{1-q\cos^{2}\theta}}{S(\cos\theta)}=1+\Big{(}\frac{q}{1-q}+\frac{3}{4}\Big{)}(1-\cos\theta)+o(1-\cos\theta).$ Thus, we can write, for $n\leq N[\lambda]$,and by taking into account that $\sqrt{x}J_{\alpha}(x)$ is bounded then $\displaystyle\frac{\psi_{n,c}^{(\alpha)}(\cos\frac{\tau}{\lambda})}{\lambda^{\alpha}}$ $\displaystyle=$ $\displaystyle n^{1/2}\frac{J_{\alpha}(\frac{n\tau}{\lambda})}{\Big{(}\lambda.\sin\frac{\tau}{\lambda}\Big{)}^{\alpha}}-n^{1/2}\frac{J_{\alpha}(\frac{n\tau}{\lambda})}{\Big{(}\lambda.\sin\frac{\tau}{\lambda}\Big{)}^{\alpha}}\big{(}\frac{q}{1-q}+3/4\big{)}\frac{\tau^{2}}{4\lambda^{\alpha+2}}+O(\frac{1}{n.\lambda^{\alpha+2}})$ (37) $\displaystyle=$ $\displaystyle n^{1/2}J_{\alpha}(\frac{n\tau}{\lambda})\Big{(}\frac{1}{\tau}\Big{)}^{\alpha}+o(\frac{1}{n}).$ On the other hand, $\displaystyle\lambda^{\alpha}a_{n}(g_{\lambda})$ $\displaystyle=$ $\displaystyle\lambda^{\alpha-1}\int_{0}^{M}g(\tau)\psi_{n,c}^{(\alpha)}(\cos\frac{\tau}{\lambda})(\sin\frac{\tau}{\lambda})^{2\alpha+1}d\tau$ $\displaystyle=$ $\displaystyle\frac{1}{\lambda^{2}}\Bigg{[}A_{\alpha}(q)n^{1/2}\int_{0}^{\infty}g(\tau)J_{\alpha}(\frac{n\tau}{\lambda})\Big{(}\lambda\sin\frac{\tau}{\lambda}\Big{)}^{\alpha+1}d\tau\Bigg{]}+o(\frac{1}{\lambda^{2}})$ $\displaystyle=$ $\displaystyle\frac{n^{1/2}}{\lambda^{2}}\int_{0}^{\infty}g(\tau)J_{\alpha}(\frac{n\tau}{\lambda})\tau^{\alpha+1}d\tau+o(\frac{1}{\lambda^{2}}).$ Then, by combining the last two estimates, one gets $G^{N}(\tau,\lambda)=\sum_{n=0}^{N[\lambda]+1}\phi(\frac{n}{\lambda})\mathcal{H}_{\alpha}.g(\frac{n}{\lambda})J_{\alpha}(\frac{n\tau}{\lambda})\frac{1}{\tau^{\alpha}}\frac{n}{\lambda^{2}}+\frac{n}{\lambda^{2}}o(1).$ Therefore, by letting $\lambda\to\infty$, we conclude for the proof of lemma3. ∎ ### 5.2 CPSWF’s case As for the example studied in the previous section, we start by establishing an adequate transferring theorem for the circular case. To do this, we introduce a suitable terminology. ###### Definition 2. Let $\lambda>0$ be a sufficiently large real, a bounded sequence $\\{m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})\\}_{n}$ is called to be a Circular prolate multiplier, if there exists a constant $C>0$ such that for every $f\in L^{p}(0,1)$, we have ${\left\|{\sum_{n=0}^{\infty}m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})a_{n}(f)\varphi^{(\alpha)}_{n,c}}\right\|}_{L^{p}(0,1)}\leq C{\left\|{f}\right\|}_{L^{p}(0,1)}.$ The smallest constant $C$ verifying the last inequality is written ${\left\|{m\Big{(}\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\Big{)}}\right\|}_{p}$. Here $\mathcal{M}:=\mathcal{M}_{0}=\mathcal{H}_{0}\Big{(}m(.)\mathcal{H}_{0}(f)\Big{)}$ is the multiplier related to the Hankel transform operator. ###### Theorem 5 (Circular transferring theorem). Let $1<p<\infty$, $\alpha\geq 1/2$ and $m$ be a bounded function on $(0,\infty)$ continuous except on a set of Lebesgue measure zero such that $\\{m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})\\}_{n}$ is a Circular prolate multiplier for all large $\lambda>0$ and $\displaystyle\liminf_{\lambda\to\infty}{\left\|{m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})}\right\|}_{p}$ is finite then $m$ is an $L^{p}$-Hankel transform multiplier and we have ${\left\|{\mathcal{M}}\right\|}_{p}\leq\liminf_{\lambda\to\infty}{\left\|{m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)}\right\|}_{p}.$ ###### Proof. Let $\lambda>0$ and $g\in C^{\infty}_{c}(0,\infty)$ supported in $(0,M)$ such that $\lambda>\frac{2}{\pi}M$. Let $g_{\lambda}(\tau)=g(\lambda\tau)$ for every $\tau\in(0,1)$ and $G_{\lambda}=g_{\lambda}\circ\arccos.$ By asymption, we have ${\left\|{\sum_{n=0}^{\infty}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a_{n}(G_{\lambda})\varphi_{n}}\right\|}_{L^{p}\left(0,1\right)}\leq{\left\|{m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})}\right\|}_{p}{\left\|{G_{\lambda}}\right\|}_{L^{p}\left(0,1\right)}.$ Then, we get ${\left\|{\chi_{(0,\lambda\frac{\pi}{2})}\sum_{n=0}^{\infty}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a_{n}\left(G_{\lambda}\right)\varphi_{n}(\cos(\frac{.}{\lambda}))}\right\|}^{p}_{L^{p}((0,\infty),\sin(\frac{.}{\lambda}))}\leq{\left\|{m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})}\right\|}^{p}_{p}{\left\|{g}\right\|}^{p}_{L^{p}\left((0,\infty),\sin(\frac{.}{\lambda})\right)}$ We denote by $F_{\lambda}(\theta)=\chi_{(0,\lambda\frac{\pi}{2})}(\theta)\sum_{n=0}^{\infty}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a_{n}\left(G_{\lambda}\right)\varphi_{n}\left(\cos(\frac{\theta}{\lambda})\right),$ hence we have ${\left\|{F_{\lambda}}\right\|}^{p}_{L^{p}((0,\infty),\sin(\frac{.}{\lambda}))}\leq{\left\|{m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})}\right\|}^{p}_{p}{\left\|{g}\right\|}^{p}_{L^{p}((0,\infty),\sin(\frac{.}{\lambda}))}.$ (38) By using (38), Fatou’s Lemma and the fact that $\displaystyle\lim_{\lambda\to\infty}\lambda\sin(\frac{\theta}{\lambda})=\theta$, we obtain ${\left\|{\displaystyle\liminf_{\lambda\to\infty}F_{\lambda}}\right\|}^{p}_{L^{p}((0,\infty),\theta d\theta)}\leq\displaystyle\liminf_{\lambda\to\infty}{\left\|{m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})}\right\|}^{p}_{p}{\left\|{g}\right\|}^{p}_{L^{p}((0,\infty),\theta d\theta)}$ (39) Let $L=\displaystyle\liminf_{\lambda\to\infty}{\left\|{m(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})}\right\|}_{p}$, then there exists a sequence of $(\lambda_{j})_{j\in\mathbb{N}}$ such that $\displaystyle\lim_{j\to\infty}\lambda_{j}=+\infty$ verifying ${\left\|{F_{\lambda_{j}}}\right\|}_{L^{p}((0,\infty),\theta d\theta)}\leq(L+1/j){\left\|{g}\right\|}_{L^{p}((0,\infty),\theta d\theta)}.$ (40) On the other hand, as $m$ is bounded and from Perseval’s formula, we have ${\left\|{F_{\lambda_{j}}}\right\|}_{L^{2}((0,\infty),\theta d\theta)}\leq(L+1/j){\left\|{g}\right\|}_{L^{2}((0,\infty),\theta d\theta)}.$ (41) From (40) and (41) there exists a subsequence of $(\lambda_{j})_{j\in\mathbb{N}}$ denoted also $(\lambda_{j})_{j\in\mathbb{N}}$ such that the sequence $\\{F_{\lambda_{j}}\\}$ converge weakly to a function $F$ in $L^{p}\cap L^{2}((0,\infty),\theta d\theta)$ and satisfying the following inequality ${\left\|{F}\right\|}_{L^{p}((0,\infty),\theta d\theta)}\leq L{\left\|{g}\right\|}_{L^{p}((0,\infty),\theta d\theta)}.$ (42) Our purpose now is to show that $F=\mathcal{H}_{0}\left(m(.)\mathcal{H}_{0}(g)\right)$ almost everywhere on $(0,\infty).$ Let $N\geq 1$ and $\theta\in(0,\infty)$ $\displaystyle F_{\lambda}(\theta)$ $\displaystyle=$ $\displaystyle\chi_{(0,\lambda)}(\theta)\sum_{n=0}^{\infty}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a_{n}(G_{\lambda})\varphi_{n}\left(\cos(\frac{\theta}{\lambda})\right)$ $\displaystyle=$ $\displaystyle\chi_{(0,\lambda)}(\theta)\left[\sum_{n=0}^{N[\lambda]}+\sum_{n=N[\lambda]+1}^{\infty}\right]m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a_{n}(G_{\lambda})\varphi_{n}\left(\cos(\frac{\theta}{\lambda})\right)$ $\displaystyle=$ $\displaystyle F^{N}_{\lambda}(\theta)+K^{N}_{\lambda}(\theta).$ Using (9), the function $F(\theta)=\varphi^{\alpha}_{n,c}\left(\cos(\theta)\right)$ satisfies the following differential equation $\displaystyle\mathcal{L}(F)(\theta)$ $\displaystyle=$ $\displaystyle-F^{\prime\prime}(\theta)-\frac{\cos(\theta)}{\sin(\theta)}F^{\prime}(\theta)+\left(c^{2}\cos^{2}(\theta)-\frac{1/4-\alpha^{2}}{\cos^{2}(\theta)}\right)F(\theta)$ $\displaystyle=$ $\displaystyle\chi_{n,\alpha}(c)F(\theta).$ Using the symmetry of $\mathcal{L}$ on $C^{\infty}_{c}(0,\infty)$, we obtain $\displaystyle a_{n}(G_{\lambda})$ $\displaystyle=$ $\displaystyle{\left\langle{G_{\lambda},\varphi^{\alpha}_{n,c}}\right\rangle}_{L^{2}(0,1)}$ $\displaystyle=$ $\displaystyle\frac{1}{\chi_{n,\alpha}(c)}\int_{0}^{\frac{\pi}{2}}g_{\lambda}(\theta)\chi_{n,\alpha}(c)\varphi^{\alpha}_{n,c}\left(\cos(\theta)\right)\sin(\theta)d\theta$ $\displaystyle=$ $\displaystyle\frac{1}{\chi_{n,\alpha}(c)}\int_{0}^{\frac{\pi}{2}}g_{\lambda}(\theta)\mathcal{L}(F)(\theta)\sin(\theta)d\theta$ $\displaystyle=$ $\displaystyle\frac{\lambda^{2}}{\chi_{n,\alpha}(c)}\int_{0}^{\frac{\pi}{2}}\frac{1}{\lambda^{2}}\mathcal{L}\left(g_{\lambda}\right)(\theta)F(\theta)\sin(\theta)d\theta=\frac{\lambda^{2}}{\chi_{n,\alpha}(c)}a_{n}\left(\frac{1}{\lambda^{2}}\mathcal{L}\left(g_{\lambda}\right)\right).$ Using the previous equality, Perseval’s formula and the fact that $m$ is bounded with the well known inequality $\frac{2}{\pi}\theta\leq\sin(\theta)\leq\theta$, for $0\leq\theta\leq\frac{\pi}{2}$ and (10), we obtain $\displaystyle{\left\|{K_{\lambda}^{N}}\right\|}_{L^{2}((0,\infty),\theta d\theta)}$ $\displaystyle=$ $\displaystyle\left[\int_{0}^{\infty}\chi_{(0,\lambda\frac{\pi}{2})}(\theta)\left|\sum_{n=N[\lambda]+1}^{\infty}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a_{n}(G_{\lambda})\varphi_{n}(\cos(\frac{\theta}{\lambda}))\right|^{2}\theta d\theta\right]^{1/2}$ $\displaystyle=$ $\displaystyle\left[\int_{0}^{\lambda\frac{\pi}{2}}\left|\sum_{n=N[\lambda]+1}^{\infty}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a_{n}(G_{\lambda})\varphi_{n}(\cos\left(\frac{\theta}{\lambda})\right)\right|^{2}\theta d\theta\right]^{1/2}$ $\displaystyle\leq$ $\displaystyle\sqrt{\frac{\pi}{2}}\left[\lambda\int_{0}^{\lambda\frac{\pi}{2}}\left|\sum_{n=N[\lambda]+1}^{\infty}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a_{n}(G_{\lambda})\varphi_{n}\left(\cos(\frac{\theta}{\lambda})\right)\right|^{2}\sin\left(\frac{\theta}{\lambda}\right)d\theta\right]^{1/2}$ $\displaystyle=$ $\displaystyle\sqrt{\frac{\pi}{2}}\left[\lambda^{2}\sum_{n=N[\lambda]+1}^{\infty}m^{2}\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\right)a^{2}_{n}(G_{\lambda})\right]^{1/2}$ $\displaystyle\leq$ $\displaystyle C\,\sqrt{\frac{\pi}{2}}\left[\frac{\lambda^{2}}{N^{4}}\sum_{n=N[\lambda]+1}^{\infty}a^{2}_{n}\left(\frac{1}{\lambda^{2}}\mathcal{L}(g_{\lambda})\right)\right]^{1/2}$ $\displaystyle\leq$ $\displaystyle\frac{C}{N^{2}}\,\sqrt{\frac{\pi}{2}}\left[{\left\|{g^{\prime\prime}+\frac{g^{\prime}}{\theta}}\right\|}_{L^{2}\left((0,\infty),\theta d\theta\right)}+C{\left\|{g}\right\|}_{L^{2}\left((0,\infty),\theta d\theta\right)}\right]$ Then we obtain ${\left\|{K^{N}_{\lambda}}\right\|}_{L^{2}((0,\infty),\theta d\theta)}=O(\frac{1}{N^{2}})$ uniformly in $\lambda.$ Thus by the diagonal argument there exists a subsequence of $\\{\lambda_{j}\\}$ noted again $\\{\lambda_{j}\\}$ such that for every $N\geq 1$, $\\{K^{N}_{\lambda_{j}}\\}_{j\in\mathbb{N}}$ converge weakly to a function $K^{N}$ in $L^{2}((0,\infty),\theta d\theta)$ satisfy ${\left\|{K^{N}}\right\|}_{L^{2}((0,\infty),\theta d\theta)}=O(\frac{1}{N^{2}})$, one conclude that there exists a sequence $\\{N_{k}\\}$ such that $\\{K^{N_{k}}\\}_{k\in\mathbb{N}}$ converge to zero almost everywhere on $(0,\infty)$. Let $F^{N_{k}}=F-K^{N_{k}}$, clearly we have $\\{F_{{\lambda_{j}}}^{N_{k}}\\}_{j\in\mathbb{N}}$ converge weakly to $F^{N_{k}}$ in $L^{2}(0,\infty)$ for every $k\in\mathbb{N}$. Moreover, $\\{F^{N_{k}}\\}$ converge to $F$ almost everywhere on $(0,\infty).$ We prove now the following equality $\lim_{j\to\infty}F_{{\lambda_{j}}}^{N_{k}}(x)=\int_{0}^{N_{k}}m(y)J_{0}(xy)\mathcal{H}_{0}(g)(y)ydy$ (43) for every $x\in(0,\infty)$, the weak convergence of $\\{F_{{\lambda_{j}}}^{N_{k}}\\}_{j\in\mathbb{N}}$ to $F^{N_{k}}$, in particular, ${\left\langle{F_{{\lambda_{j}}}^{N_{k}},\chi_{(r,s)}}\right\rangle}$ converge to ${\left\langle{F^{N_{k}},\chi_{(r,s)}}\right\rangle}$ for every $0<r<s<\infty$ and by using the Lebesgue dominated convergence theorem which give as ${\left\langle{F_{{\lambda_{j}}}^{N_{k}},\chi_{(r,s)}}\right\rangle}$ converge to ${\left\langle{\mathcal{H}_{\alpha}\left(\chi_{(0,N_{k}\pi)}m(.)\mathcal{H}_{\alpha}(g)\right),\chi_{(r,s)}}\right\rangle}$, one conclude that $F^{N_{k}}=\mathcal{H}_{\alpha}\left(\chi_{(0,N_{k}\pi)}m(.)\mathcal{H}_{\alpha}(g)\right)$ almost everywhere on $(0,\infty).$ Finally, as $k\to\infty$, we get our purpose. For the proof of (43), we need the uniform approximation of the family of CPSWFs on $(0,1)$ which is given by the following estimates $\varphi_{n,c}^{\alpha}(\cos(\frac{\theta}{\lambda}))=(-1)^{n}B_{n}\left(\cos(\frac{\theta}{\lambda})\right)^{\alpha+1/2}P_{n}^{(0,\alpha)}\left(\cos(\frac{2\theta}{\lambda})\right)+\gamma_{n}^{\alpha+1/2}O(\frac{c^{2}}{n})$ (44) for every $\theta\in\mathbb{(}\lambda t_{n},\lambda\frac{\pi}{2})$, where $t_{n}=\arccos(\gamma_{n})$ and $\gamma_{n}\sim\frac{\sqrt{\alpha^{2}-1/4}}{\chi_{n,\alpha}^{1/2}(c)}$. $\varphi_{n,c}^{\alpha}(\cos(\frac{\theta}{\lambda}))=A_{n}\,\chi^{1/4}_{n,\alpha}(c)\frac{\sqrt{S(\cos(\frac{\theta}{\lambda}))}J_{0}\left(\chi^{1/2}_{n,\alpha}(c)S(\cos(\frac{\theta}{\lambda}))\right)}{(\sin(\frac{\theta}{\lambda}))^{\frac{1}{2}}r_{n}\left(\cos(\frac{\theta}{\lambda})\right)^{1/4}}+R_{n}(\cos(\frac{\theta}{\lambda}))$ (45) for every $\theta\in\mathbb{(}0,\lambda t_{n})$, where $A_{n}\sim 1,$ $r_{n}(t)=1-qt^{2}+\frac{1/4-\alpha^{2}}{\chi_{n,\alpha}^{1/2}(c)t^{2}}$ and $\displaystyle\sup_{\theta\in\mathbb{(}0,t_{n})}\left|R_{n}(\cos(\theta))\right|\leq\frac{C}{\chi^{1/2}_{n,\alpha}(c)}$, for more details see [17]. By a straightforward computation, we have $\frac{\sqrt{S(\cos(\frac{\theta}{\lambda}))}}{(\sin(\frac{\theta}{\lambda}))^{1/2}r_{n}\left(\cos(\frac{\theta}{\lambda})\right)^{1/4}}=1-\beta(q)(1-\cos(\frac{\theta}{\lambda}))+o(1-\cos(\frac{\theta}{\lambda}))$ then, we can easily check that $\varphi_{n,c}^{\alpha}(\cos(\frac{\theta}{\lambda}))=\chi_{n,\alpha}^{1/4}(c)J_{0}\left(\frac{\chi_{n,\alpha}^{1/2}(c)}{\lambda}\theta\right)+R_{n}(\cos(\frac{\theta}{\lambda})).$ Let $N>0$ and $\lambda>\max\\{\frac{2M}{\pi},N^{3}\\}$. By (45) we have, for every $n\leq N[\lambda]$ $\displaystyle a_{n}(G_{\lambda})$ $\displaystyle=$ $\displaystyle{\left\langle{G_{\lambda},\varphi_{n,c}^{\alpha}}\right\rangle}_{L^{2}(0,1)}$ $\displaystyle=$ $\displaystyle\frac{1}{\lambda}\int_{0}^{\lambda\frac{\pi}{2}}\left(g_{\lambda}\circ\arccos\right)(\cos(\frac{\theta}{\lambda}))\varphi_{n,c}^{\alpha}(\cos(\frac{\theta}{\lambda}))\sin(\frac{\theta}{\lambda})d\theta$ $\displaystyle=$ $\displaystyle\frac{\chi_{n,\alpha}^{1/4}(c)}{\lambda}\int_{0}^{\lambda\frac{\pi}{2}}J_{0}\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\theta\right)g(\theta)\sin(\frac{\theta}{\lambda})d\theta-\frac{\chi_{n,\alpha}^{1/4}(c)}{\lambda}\int_{\lambda t_{n}}^{\lambda\frac{\pi}{2}}J_{0}\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\theta\right)g(\theta)\sin(\frac{\theta}{\lambda})d\theta$ $\displaystyle+$ $\displaystyle\frac{(-1)^{n}B_{n}}{\lambda}\int_{\lambda t_{n}}^{\lambda\frac{\pi}{2}}g(\theta)P_{n}^{(0,\alpha)}\left(\cos(\frac{2\theta}{\lambda})\right)\left(\cos(\frac{\theta}{\lambda})\right)^{\alpha+1/2}\sin(\frac{\theta}{\lambda})d\theta+\frac{1}{n^{a}\chi^{1/4}_{n,\alpha}(c)}O(\frac{1}{\lambda^{b}})$ $\displaystyle=$ $\displaystyle\frac{\chi_{n,\alpha}^{1/4}(c)}{\lambda^{2}}\mathcal{H}_{0}(g)(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})+\frac{1}{n^{a}\chi^{1/4}_{n,\alpha}(c)}O(\frac{1}{\lambda^{b}}).$ where $a>1$ and $b>0.$ Indeed, using the fact that $\sup_{x>0}|\sqrt{x}J_{\alpha}(x)|\leq C_{\alpha}$, see [23], and $\lambda\sin(\frac{\theta}{\lambda})\leq\theta$ we have $\displaystyle\left|\frac{\chi_{n,\alpha}^{1/4}(c)}{\lambda}\int_{\lambda t_{n}}^{\lambda\frac{\pi}{2}}J_{0}\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\theta\right)g(\theta)\sin(\frac{\theta}{\lambda})d\theta\right|$ $\displaystyle\leq$ $\displaystyle\frac{1}{\lambda^{3/2}}\int_{\lambda t_{n}}^{\lambda\frac{\pi}{2}}\left|\frac{\chi_{n,\alpha}^{1/4}(c)}{\lambda^{1/2}}J_{0}\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\theta\right)\right||g(\theta)|\theta d\theta$ $\displaystyle\leq$ $\displaystyle\frac{1}{\lambda^{3/2}}\left[\int_{\lambda t_{n}}^{\lambda\frac{\pi}{2}}\left|\frac{\chi_{n,\alpha}^{1/4}(c)\theta^{1/2}}{\lambda^{1/2}}J_{0}\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda}\theta\right)\right|^{2}\frac{d\theta}{\theta}\right]^{1/2}{\left\|{\theta g}\right\|}_{L^{2}(0,\infty)}$ $\displaystyle\leq$ $\displaystyle\frac{C_{0}}{\lambda^{3/2}}(\ln(\frac{\pi}{2t_{n}}))^{1/2}{\left\|{\theta g}\right\|}_{L^{2}(0,\infty)}$ $\displaystyle\leq$ $\displaystyle\frac{C_{0}}{\lambda^{3/2}\chi_{n,\alpha}^{1/4}(c)}{\left\|{\theta g}\right\|}_{L^{2}(0,\infty)}.$ Moreover, using the fact that $\left|P_{n}^{(0,\alpha)}(\cos(\frac{2\theta}{\lambda}))\right|\leq P_{n}^{(0,\alpha)}(1)=O(n^{\alpha})$ and the deacreasing cosinus function with $|B_{n}|=O(n^{1/2})$, we obtain $\left|\displaystyle\frac{(-1)^{n}B_{n}}{\lambda}\int_{\lambda t_{n}}^{\lambda\frac{\pi}{2}}g(\theta)P_{n}^{(0,\alpha)}\left(\cos(\frac{2\theta}{\lambda})\right)\left(\cos(\frac{\theta}{\lambda})\right)^{\alpha+1/2}\sin(\frac{\theta}{\lambda})d\theta\right|$ $\displaystyle\leq$ $\displaystyle\frac{|B_{n}|}{\lambda^{2}}\int_{\lambda t_{n}}^{\lambda\frac{\pi}{2}}\left|g(\theta)\right|\left|P_{n}^{(0,\alpha)}\left(\cos(\frac{2\theta}{\lambda})\right)\right|\left(\cos(\frac{\theta}{\lambda})\right)^{\alpha+1/2}\theta d\theta$ $\displaystyle\leq$ $\displaystyle\frac{C}{\lambda^{2}\chi^{1/4}_{n,\alpha}(c)}{\left\|{\theta^{3/2}g}\right\|}_{L^{2}(0,\infty)}$ Finally, there exist a constant $a>1$ and $b>0$ such that $a_{n}(G_{\lambda})=\frac{\chi_{n,\alpha}^{1/4}(c)}{\lambda^{2}}\mathcal{H}_{0}(g)(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda})+\frac{1}{n^{a}\chi^{1/4}_{n,\alpha}(c)}O(\frac{1}{\lambda^{b}}).$ (46) Hence, we obtain $F_{{\lambda_{j}}}^{N_{k}}(\theta)=\chi_{(0,\frac{\pi}{2}\lambda_{j})}(\theta)\displaystyle\sum_{n=0}^{N_{k}[\lambda_{j}]}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}}\right)a_{n}(G_{\lambda_{j}})\varphi^{\alpha}_{n,c}(\cos(\frac{\theta}{\lambda_{j}}))$ $\displaystyle=$ $\displaystyle\sum_{n=0}^{N_{k}[\lambda_{j}]}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}}\right)a_{n}(G_{\lambda_{j}})\left(\chi_{(0,\lambda_{j}t_{n})}(\theta)\varphi_{n}(\cos(\frac{\theta}{\lambda_{j}}))+\chi_{(\lambda_{j}t_{n},\frac{\pi}{2}\lambda_{j})}(\theta)\varphi_{n}(\cos(\frac{\theta}{\lambda_{j}}))\right)$ $\displaystyle=$ $\displaystyle\sum_{n=0}^{N_{k}[\lambda_{j}]}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}}\right)\left(\frac{\chi_{n,\alpha}^{1/4}(c)}{\lambda^{2}_{j}}\mathcal{H}_{0}(g)(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}})+\frac{1}{n^{a}\chi^{1/4}_{n,\alpha}(c)}O(\frac{1}{\lambda_{j}^{b}})\right)\chi_{(0,\lambda_{j}\frac{\pi}{2})}(\theta)\chi_{n,\alpha}^{1/4}(c)J_{0}\left(\frac{\chi_{n,\alpha}^{1/2}(c)}{\lambda}\theta\right)$ $\displaystyle+$ $\displaystyle\sum_{n=0}^{N_{k}[\lambda_{j}]}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}}\right)a_{n}(G_{\lambda_{j}})\chi_{(\lambda_{j}t_{n},\frac{\pi}{2}\lambda_{j})}(\theta)\left((-1)^{n}B_{n}\left(\cos(\frac{\theta}{\lambda})\right)^{\alpha+1/2}P_{n}^{(0,\alpha)}\left(\cos(\frac{2\theta}{\lambda})\right)-\chi_{n,\alpha}^{1/4}(c)J_{0}\left(\frac{\chi_{n,\alpha}^{1/2}(c)}{\lambda}\theta\right)\right)$ $\displaystyle+$ $\displaystyle O(\frac{1}{\lambda^{\varepsilon}_{j}})$ $\displaystyle=$ $\displaystyle\chi_{(0,\frac{\pi}{2}\lambda_{j})}(\theta)\frac{1}{\lambda_{j}}\sum_{n=0}^{N_{k}[\lambda_{j}]}m\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}}\right)\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}}J_{0}\left(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}}\theta\right)\mathcal{H}_{0}(g)(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda_{j}})+O(\frac{1}{\lambda^{\varepsilon}_{j}}).$ where $\varepsilon>0.$ Indeed, from [29], we have $\displaystyle(-1)^{n}B_{n}\left(\cos(\frac{\theta}{\lambda})\right)^{\alpha+1/2}P_{n}^{(0,\alpha)}\left(\cos(\frac{2\theta}{\lambda})\right)$ $\displaystyle=$ $\displaystyle(2n+\alpha+1)^{1/2}\left(\cos(\frac{\theta}{\lambda})\right)^{1/2}\left(\frac{\theta/\lambda}{\sin(\theta/\lambda)}\right)^{1/2}J_{0}\left(2(n+\frac{\alpha+1}{2})\frac{\theta}{\lambda}\right)$ $\displaystyle+$ $\displaystyle\frac{1}{\lambda^{1/2}}O(\frac{(2\theta)^{1/2}}{n})$ $\displaystyle=$ $\displaystyle(2n+\alpha+1)^{1/2}J_{0}\left((2n+\alpha+1)\frac{\theta}{\lambda}\right)+O(\frac{1}{n}),$ and by using (10), one gets $\chi_{n,\alpha}(c)\sim(2n+\alpha+1)^{2}$, and concludes that $(-1)^{n}B_{n}\left(\cos(\frac{\theta}{\lambda})\right)^{\alpha+1/2}P_{n}^{(0,\alpha)}\left(\cos(\frac{2\theta}{\lambda})\right)-\chi_{n,\alpha}^{1/4}(c)J_{0}\left(\frac{\chi_{n,\alpha}^{1/2}(c)}{\lambda}\theta\right)=O(\frac{1}{n}).$ Further, from [5], we have ${\left\|{\varphi^{\alpha}_{n,c}}\right\|}_{L^{\infty}(0,1)}=O(\chi^{1/2}_{n,\alpha}(c))$, then we obtain $a_{n}(G_{\lambda_{j}})=O(\frac{\chi^{1/2}_{n,\alpha}(c)}{\lambda^{2}}).$ Finally, as $j\to\infty$, we get $F^{N_{k}}=\mathcal{H}_{0}\left(\chi_{(0,N_{k})}m(.)\mathcal{H}_{0}(g)\right).$ ∎ ## References * [1] W.O. Amrein, A. M. Hinz and D. B. Pearson, Sturm-Liouville Theory: Past and Present, Birkhäuser, Basel-Boston-Berlin, (2005). * [2] A. I. Aptekarev, V. S. Buyarov & I. S. Degeza, Asymptotic behavior of the $L^{p}$-norms and the entropy for general orthogonal polynomials. Russian Acad. Sci. Sb. Math. 82 (1995), 373–395. * [3] J.J. Betankor, K.Stempak , Relating multipliers and transplantation for Fourier Bessel expansions and Hankel transform. Tohoku Math. J. 53 (2001), 109–129. * [4] A.Bonami and A. Karoui , Uniform Approximation and Explicit Estimates for the Prolate Spheroidal Wave Functions. Constr.Approx. 43 (2016), 15–45. * [5] M. Boulsane and A. Karoui, The Finite Hankel Transform Operator: Some Explicit and Local Estimates of the Eigenfunctions and Eigenvalues Decay Rates, J. Four. Anal. Appl, 24 , (2018), 1554-1578. * [6] M. Boulsane, P. Jaming and A. Souabni, Mean convergence of prolate spheroidal series and their extensions, J. Functional Analysis, 277, (2019) * [7] S. Bochner , Summation of multiple Fourier series by spherical means. Trans. Amer. Math. Soc 40 (1936), 175–207. * [8] V. Casarino and M. M. Peloso, $L^{p}$-summability of Riesz means for the sublaplacian on complex spheres, J. London Math. Soc. (2) 83 (2011) 137–152. * [9] P.Chen, X. T. Duong, D. He, S. Lee and L. Yan, Almost everywhere convergence of Bochner-Riesz means for the Hermite operator. Arxiv:2006.05689v3. * [10] Ó. Ciaurri and L. Roncal, The Bochner–Riesz means for Fourier–Bessel expansions Journal of Functional Analysis 228 (2005), 83–113. * [11] Ó.Ciaurri and L.Roncal, The Bochner–Riesz means for Fourier–Bessel expansions: Norm inequalities for the maximal operator and almost everywhere convergence. Journal of Approximation Theory 167 (2013), 121–146. * [12] S.Ciaurri and J. L. Varona, An Uniform Boundedness for Bochner-Riesz Operators Related to the Hankel Transform. J of lnequal and AppL, 2002, 7(6), 759–777. * [13] L. Colzani and G. Travaclini, Estimates for Riesz Kernels of Eigenfunction Expansions of Elliptic Differential Operators on Compact Manifolds.Journal of Functional Analysis 96,( 1991) l-30 * [14] E.B. Davies, Heat Kernels and Spectral Theory, Cambridge University Press, Cambridge, (1989). * [15] J. Horvath, L’oeuvre mathématique de Marcel Riesz I. Cahiers du séminaire d’histoire des mathématiques, tome 3 (1982), 83-121. * [16] S. Igari, On The Multipliers of Hankel Transform. Tohoku Math. Journal. 24(1972), 201-206. * [17] A. Karoui and I. Mehrzi, Asymptotic behaviors and numerical computations of the eigenfunctions and eigenvalues associated with the classical and circular prolate spheroidal wave functions. Journal of Applied Mathematics and Computation 218 (2012), 10871–10888. * [18] A. Karoui and A. Souabni, Generalized Prolate Spheroidal Wave Functions: Spectral Analysis and Approximation of Almost Band-limited Functions. J. Four. Anal. Appl. 22 (2016), 383–412. * [19] A. Karoui and A. Souabni, Weighted finite fourier transform operator:Uniform approximations of the eigenfunctions, eigenvalues decay and behaviour. J.Sci.Comp. 71 (2) (2017), 547–570. * [20] G. Mauceri , Riesz means for the eigenfunction expansions for a class of hypo-elliptic differential operators. Ann. Inst. Fourier 31 (1981), 115–140. * [21] C. Meaney, Divergent Cesaro and Riesz means of Jacobi and Laguerre expansions.Proceedings of the American Mathematical Society. 131 (2003), 3123–3128. * [22] D. Müller, On Riesz means of eigenfunction expansions for the Kohn Laplacian. J. Reine Angew. Math 401 (1989), 113–121. * [23] A. YA. Olenko. Upper bound on $\sqrt{x}J_{\mu}(x)$ and its applications. Integral Transforms and Special Functions. 17 (2006), 455–467. * [24] M. Riesz, Sur les fonctions conjugées. Math. Z. 27 (1927), 218–244. * [25] D. Slepian and H. O. Pollak, Prolate spheroidal wave functions, Fourier analysis and uncertainty I, Bell System Tech. J. 40 (1961), 43–64. * [26] D.Slepian, Prolate spheroidal wave functions, Fourier analysis and uncertainty–IV: Extensions to many dimensions; generalized prolate spheroidal functions, Bell System Tech. J. 43 (1964), 3009–3057. * [27] D. Slepian, Some comments on Fourier analysis, uncertainty and modeling. SIAM Rev. 25 (1983) 379-393. * [28] L. Song , J. Xiao and X. Yan, Preduals of quadratic campanato spaces associated to operators with heat kernel bounds. J. Potential Analysis 41,(2014) 849–867. * [29] G. Szegö, Orthogonal polynomials, Fourth edition, American Mathematical Society, Colloquium Publications, Vol. XXIII. American Mathematical Society, Providence, R.I., 1975. * [30] L. L. Wang and J. Zhang, A new generalization of the PSWFs with applications to spectral approximations on quasi-uniform grids. Appl. Comput. Harmon. Anal. 29 (2010), 303–329.
# Continuous Newton-like Methods featuring Inertia and Variable Mass Camille Castera∗ Department of Mathematics University of Tübingen Germany Hedy Attouch IMAG Université Montpellier CNRS, France Jalal Fadili ENSICAEN Normandie Université CNRS, GREYC, France Peter Ochs Department of Mathematics University of Tübingen Germany ###### Abstract We introduce a new dynamical system, at the interface between second-order dynamics with inertia and Newton’s method. This system extends the class of inertial Newton-like dynamics by featuring a time-dependent parameter in front of the acceleration, called _variable mass_. For strongly convex optimization, we provide guarantees on how the Newtonian and inertial behaviors of the system can be non-asymptotically controlled by means of this variable mass. A connection with the Levenberg–Marquardt (or regularized Newton’s) method is also made. We then show the effect of the variable mass on the asymptotic rate of convergence of the dynamics, and in particular, how it can turn the latter into an accelerated Newton method. We provide numerical experiments supporting our findings. This work represents a significant step towards designing new algorithms that benefit from the best of both first- and second-order optimization methods. 11footnotetext: Corresponding author<EMAIL_ADDRESS> ## 1 Introduction ### 1.1 Problem Statement A major challenge in modern unconstrained convex optimization consists in building fast algorithms while maintaining low computational cost and memory footprint. This plays a central role in many key applications such as large- scale machine learning problems or data processing. The problems we are aiming to study are of the form $\min_{x\in\mathbb{R}^{n}}f(x).$ Large values of $n$ demand for algorithms at the interface of first- and second-order optimization. Limited computational capabilities explain why gradient-based (first-order) algorithms remain prominent in practice. Unfortunately, they often require many iterations which is true even for the provably best algorithms for certain classes of optimization problems; for example that of convex and strongly convex functions with Lipschitz continuous gradient [37, 33, 34]. On the other hand, algorithms using second-order information (the Hessian of $f$)—with Newton’s method as prototype—adapt locally to the geometry of the objective, allowing them to progress much faster towards a solution. However, each iteration comes with high computational and memory costs, which highlights a challenging trade-off. It is therefore essential to develop algorithms that take the best of both worlds. They are commonly referred to as (limited-memory) quasi-Newton methods. Several quasi-Newton algorithms partly address this issue, for example BFGS methods [18, 23, 25, 39, 31], yet, in very large-scale applications, first-order algorithms often remain the preferred choice. In order to reach a new level of efficiency, deep insights into the mechanism and relations between algorithms are required. To that aim, an insightful approach is to see optimization algorithms as discretization of ordinary differential equations (ODEs): for small-enough step-sizes, iterates can be modeled by a continuous-time trajectory [32, 13]. Obtaining a fast algorithm following this strategy depends on two ingredients: choosing an ODE for which rapid convergence to a solution can be proved, and discretizing it with an appropriate scheme that preserves the favorable properties of the ODE. Both steps are highly challenging, our work focuses on the ODE matter. We study the following second-order dynamical system in a general setting: $\varepsilon(t)\ddot{x}(t)+\alpha(t)\dot{x}(t)+\beta\nabla^{2}f(x(t))\dot{x}(t)+\nabla f(x(t))=0,\quad t\geq 0,$ (VM-DIN-AVD) where $f\colon\mathbb{R}^{n}\to\mathbb{R}$ is a smooth convex twice continuously differentiable function, with gradient $\nabla f$ and Hessian $\nabla^{2}f$ defined on $\mathbb{R}^{n}$ equipped with scalar product $\langle\cdot,\cdot\rangle$, and induced norm $\|\cdot\|$. Additionally, $f$ is assumed to be coercive, and strongly convex on bounded subsets of $\mathbb{R}^{P}$. The functions $\varepsilon,\alpha\colon\mathbb{R}_{+}\to\mathbb{R}_{+}$ (where $\mathbb{R}_{+}=[0,+\infty[$) are differentiable, non-increasing, and $\varepsilon(t)>0$ for all $t\geq 0$. Together with $\beta>0$, they are control parameters that define the type of dynamics that drives the trajectory (or solution) $x\colon\mathbb{R}_{+}\to\mathbb{R}^{n}$, whose first- and second-order derivatives are denoted $\dot{x}$ and $\ddot{x}$ respectively. We call the above dynamics (VM-DIN-AVD), which stands for “Variable Mass Dynamical Inertial Newton-like system with Asymptotically Vanishing Damping” since it generalizes a broad class of ODEs whose original member is DIN [2], where $\varepsilon$ and $\alpha$ were constant. DIN was then extended to the case of non-constant _asymptotically vanishing dampings_ (AVD) $\alpha$ [9]. In this work we introduce the non-constant parameter $\varepsilon$ called _variable mass_ (VM) in front of the acceleration $\ddot{x}$, in the same way that $\alpha$ is called (viscous) _damping_ by analogy with classical mechanics. A key feature of these ODEs, that positions them at the interface of first- and second-order optimization, is that they possess equivalent forms involving only $\nabla f$ but not $\nabla^{2}f$, significantly reducing computational costs, hence enabling the design of practical algorithms, see e.g., [20, 10, 21]. The key idea behind this is the relation $\nabla^{2}f(x(t))\dot{x}(t)=\frac{\mathop{}\\!\mathrm{d}}{\mathop{}\\!\mathrm{d}t}\nabla f(x(t))$, see Section 2 for an equivalent formulation of (VM-DIN-AVD) exploiting this. This paper emphasizes the relation between (VM-DIN-AVD) and well-studied special cases. Indeed, taking $\varepsilon=\alpha=0$, one obtains111CN is usually considered with $\beta=1$, we put $\beta$ in the system to ease the discussions below. the Continuous Newton (CN) method [24] $\beta\nabla^{2}f(x_{N}(t))\dot{x}_{N}(t)+\nabla f(x_{N}(t))=0,\quad t\geq 0,$ (CN) known notably for being invariant to affine transformations and yielding fast vanishing of the gradient (see Section 3). In fact, this observation shows that (VM-DIN-AVD) is a singular perturbation of (CN), which also justifies the terminology “Newton-like” in DIN. When $\alpha\neq 0$ but $\varepsilon=0$, we recover the Levenberg–Marquardt (LM) method, $\alpha(t)\dot{x}_{LM}(t)+\beta\nabla^{2}f(x_{LM}(t))\dot{x}_{LM}(t)+\nabla f(x_{LM}(t))=0,\quad t\geq 0,$ (LM) also known as regularized Newton method since it stabilizes (CN). In the rest of the paper, the solutions of (CN) and (LM) will always be denoted by $x_{N}$ and $x_{LM}$ respectively. Alvarez et al. [2] showed that for $\alpha=0$, $\beta=1$, and $\varepsilon$ constant and small, (VM-DIN-AVD) is a “perturbed” Newton method since the distance between the solutions of (VM-DIN-AVD) and (CN) is at most proportional to $\sqrt{\varepsilon}$ at all time. Yet, despite the benefits of this class of ODEs, such as stabilization properties, see e.g., [9, 10], no improvement222DIN-like systems were thought to yield faster vanishing of the gradient compared to inertial first-order dynamics, until recently [4]. of the rate of convergence (in values) has been shown compared to inertial first-order dynamics [37, 41]. This raises the question: “ _are these ODEs really of Newton type?_ ”, which is crucial in view of designing faster algorithms from them. Figure 1: Left: phase diagram on distances from (VM-DIN-AVD) to (CN) and (LM) (see Section 3). For each patch, the color indicates which of the distances $\|x_{N}(t)-x(t)\|$ and $\|x_{LM}(t)-x(t)\|$ is considered, the scaling of a corresponding upper-bound on this distance is written; in white for prior work and in black for our contributions. The green line separates the cases $\varepsilon\geq\alpha$ (above) and $\varepsilon\leq\alpha$ (below). Right: 2D illustration of the trajectories of (VM-DIN-AVD) for several choices of $\varepsilon$ on a quadratic function. Using fast-vanishing $\varepsilon(t)$ (dark-blue solid curves), one can bring the solution of (VM-DIN-AVD) close to that of (CN), making it, for example, more robust to bad conditioning compared to first-order dynamics (such as gradient descent). ### 1.2 Main Contributions We show that the answer to this question is partially positive, and closely related to the choices of $\varepsilon$ and $\alpha$. We provide general results on the role played by these two control parameters and how they can be chosen to control (VM-DIN-AVD), and make it close to (CN) _for all time_ , as illustrated on the right-hand side of Figure 1, but also to obtain fast convergence. This represents a first step towards building new fast practical algorithms. Our main contributions are the following: – We provide a first-order equivalent formulation of (VM-DIN-AVD), and address the questions of existence and uniqueness of the solutions of (VM-DIN-AVD) under mild assumptions. – We generalize the perturbed Newtonian property discussed above to non- constant and possibly vanishing variable masses $\varepsilon$, and “not too large” positive dampings $\alpha$, and derive bounds that (formally) take the form $\|x(t)-x_{N}(t)\|=O(\sqrt{\varepsilon(t)})$. We then extend these results to larger dampings $\alpha$ and make the connection between (VM-DIN- AVD) and (LM). This contribution is summarized in the phase diagram of Figure 1. – Using quadratic functions as a model for strongly convex functions, we shed light on techniques to efficiently approximate solutions of (VM-DIN-AVD). We then show how $\varepsilon$ and $\alpha$ affect the speed of convergence. Depending on their setting, the solutions of (VM-DIN-AVD) may either converge as fast as that of (CN), _faster_ , or rather have a (LM) nature, as summarized in Table 1. – We provide numerical experiments supporting our theoretical findings. Table 1: Informal summary of Section 4. Comparison of (VM-DIN-AVD) with other dynamics Parameters of (VM-DIN-AVD) | | Speed of convergence ---|---|--- Dominant parameter | Integrability in $+\infty$ | | w.r.t. (CN) | w.r.t. (LM) variable mass $\varepsilon$ | yes | | as fast | as fast no | | faster | faster viscous damping $\alpha$ | yes | | as fast | only depends on $\varepsilon$ no | | slower ### 1.3 Related work The system (VM-DIN-AVD) belongs to the class of inertial systems with viscous and geometric (“Hessian-driven”) dampings, initially introduced with constant $\varepsilon=1$ and constant $\alpha$ in [2] and called DIN (for Dynamical Inertial Newton-like system). Except in a few cases [20, 19], most of the follow-up work then considered extensions of DIN with non-constant AVD $\alpha$, with in particular the DIN-AVD system with $\alpha(t)=\alpha_{0}/t$ as introduced in [9]. The reason for this popular choice for $\alpha$ is its link with Nesterov’s method [41]. Non-constant choices for $\beta$ have been considered [10, 1, 29, 11]. We keep it constant here, and rather focus on non- constant $\varepsilon$, unlike prior work that used constant $\varepsilon=1$. The mass $\varepsilon$ was only considered in the original work [2], but only for fixed $\varepsilon$, $\beta=1$ and constant $\alpha=0$. VM-DIN-AVD is however closely related to the IGS system considered in [11] as it is actually equivalent to the latter after dividing both sides of (VM-DIN-AVD) by $\varepsilon(t)$. Our approach—which consists in studying the connections with other second-order dynamics as $\varepsilon$ vanishes asymptotically—is however different from the one followed in [11], which is of independent interest. The literature on DIN is rich, let us mention further connections with Nesterov’s method [40, 1], extensions with Tikhonov regularization [16] and closed-loop damping [12, 29]. The non-smooth and possibly non-convex cases have been considered in [5, 6, 20]. Finally, avoidance of strict saddle points in smooth non-convex optimization has been shown in [19]. The influence of the damping $\alpha$ on the (LM) dynamics has been studied in [7, 8]. Interestingly, the conditions enforced on $\alpha$ in these papers (formally a sub-exponential decay) are very similar to those we make on $\varepsilon$ and $\alpha$ for (VM-DIN-AVD) (see Assumptions 1 and 2). Regarding the second part of our analysis, which deals with the case where $f$ is quadratic, Attouch et al. [10] provided closed-form solutions to (VM-DIN- AVD) for $\varepsilon\equiv 1$ and special choices of $\alpha$. Our work rather deals with approximate solutions which allows considering a wide class of functions $\varepsilon$ and $\alpha$. We rely on the Liouville–Green (LG) method [30, 26] presented in Section 4. Generalizations of LG are also often referred to as WKB methods [44, 28, 17] and seem to be mostly used in physics so far. To the best of the authors’ knowledge, the current work seem to be one of the first to use the LG method in optimization, and the first for DIN-like systems. ### 1.4 Organization The paper is organized as follows. We discuss the existence of solutions in Section 2. Our main results, from a non-asymptotic control perspective, are then presented in Section 3. An analysis of the role played asymptotically by $\varepsilon$ and $\alpha$ is then carried out on quadratic functions in Section 4. Finally, numerical experiments are presented in Section 5, and some conclusions are then drawn. ## 2 Existence and Uniqueness of Solutions In the sequel, we fix $x_{0}\in\mathbb{R}^{n}$ and $\dot{x}_{0}\in\mathbb{R}^{n}$, such that, unless stated otherwise, (VM-DIN- AVD) is always considered with initial condition $(x(0),\dot{x}(0))=(x_{0},\dot{x}_{0})$, and (CN) and (LM) with initial condition $x_{N}(0)=x_{LM}(0)=x_{0}$. We also fix initial values for the control parameters $\varepsilon(0)=\varepsilon_{0}>0$, $\varepsilon^{\prime}(0)=\varepsilon^{\prime}_{0}\leq 0$ and $\alpha(0)=\alpha_{0}\geq 0$. In addition to the definitions of $\varepsilon$ and $\alpha$ in Section 1.1, we assume that $\varepsilon$ is twice differentiable, with bounded second derivative. In order to use the Cauchy–Lipschitz Theorem, we reformulate (VM-DIN-AVD) into a first-order (in time) system by introducing an auxiliary variable $y:\mathbb{R}_{+}\to\mathbb{R}^{n}$. Notably, our reformulation does not involve $\nabla^{2}f$, in the same fashion as [2, 9]. For all $t$, defining $\nu(t)=\alpha(t)-\varepsilon^{\prime}(t)-\frac{1}{\beta}\varepsilon(t)$, we show in Appendix A that (VM-DIN-AVD) is equivalent to $\begin{cases}\varepsilon(t)\dot{x}(t)+\beta\nabla f(x(t))+\nu(t)x(t)+y(t)&=0\\\ \dot{y}(t)+\nu^{\prime}(t)x(t)+\frac{\nu(t)}{\beta}x(t)+\frac{1}{\beta}y(t)&=0\end{cases}$ (gVM-DIN-AVD) with initial conditions $(x(0),y(0))=\left(x_{0},-\varepsilon_{0}\dot{x}_{0}-\beta\nabla f(x_{0})-(\alpha_{0}-\varepsilon^{\prime}_{0}-\frac{1}{\beta}\varepsilon_{0})x_{0}\right)$. One can notice that in the special case where $\varepsilon$ is taken constant and equal to $1$ (that is when (VM-DIN-AVD) is simply the DIN-AVD system [9]), we recover the same first-order formulation as that in [9]. We can then apply the Cauchy–Lipschitz Theorem. For all $t\geq 0$ and $(u,v)\in\mathbb{R}^{n}\times\mathbb{R}^{n}$, define the mapping $G\left(t,(u,v)\right)=\begin{pmatrix}\frac{1}{\varepsilon(t)}(-\beta\nabla f(u)-\nu(t)u-v)\\\ -\nu^{\prime}(t)u-\frac{\nu(t)}{\beta}u-\frac{1}{\beta}v\end{pmatrix},$ so that (gVM-DIN-AVD) rewrites $(\dot{x}(t),\dot{y}(t))=G\left(t,(x(t),y(t))\right)$ for all $t\geq 0$. Since $f$ is twice continuously differentiable, one can see that $G$ is continuously differentiable w.r.t. its second argument $(u,v)$. Consequently $G$ is locally Lipschitz continuous w.r.t. $(u,v)$ and by the Cauchy–Lipschitz Theorem, for each initial condition, there exists a unique local solution to (gVM-DIN-AVD) and thus to (VM-DIN-AVD). We then show that this solution is actually global (in Appendix A) by proving the boundedness of $(x,y)$. We omit the existence and uniqueness of the solutions of (CN) and (LM) since these are standard results, obtained with similar arguments. ## 3 Non-asymptotic Control of (VM-DIN-AVD) The purpose of this section is to understand how close $x$ might be to $x_{N}$ and $x_{LM}$, as a function of $\alpha$ and $\varepsilon$. Since $f$ is coercive and strongly convex on bounded sets, it has a unique minimizer $x^{\star}\in\mathbb{R}^{n}$. Consequently, any two trajectories that converge to $x^{\star}$ will eventually be arbitrarily close to each other. Thus, asymptotic results of the form $\|x(t)-x_{N}(t)\|\xrightarrow[t\to+\infty]{}0$ are not precise enough to claim, for example, that $x$ has a “Newtonian behavior”. Instead, we will derive upper bounds on the distance between trajectories that hold _for all time_ $t\geq 0$, and which typically depend on $\varepsilon$ and/or $\alpha$. We first present the case where $\alpha$ is small relative to $\varepsilon$ and then generalize. ### 3.1 Comparison with (CN) under Moderate Viscous Dampings When the viscous damping $\alpha$ remains moderate w.r.t. the variable mass $\varepsilon$, one can expect the solutions of (VM-DIN-AVD) to be close to that of (CN). We make the following assumptions. ###### Assumption 1. There exists $c_{1},c_{2}\geq 0$ such that for all $t\geq 0$, $|\varepsilon^{\prime}(t)|\leq c_{1}\varepsilon(t)$ and $\alpha(t)\leq c_{2}\varepsilon(t)$. The assumption states that $\alpha$ must decrease at least as fast as $\varepsilon$ (up to a constant).333Assumption 1 can actually hold only after some large-enough $t_{0}\geq 0$, we take $t_{0}=0$ for the sake of simplicity. The reason for assuming $|\varepsilon^{\prime}(t)|\leq c_{1}\varepsilon(t)$ is technical and will appear more clearly in the proofs below. It formally means that $\varepsilon$ can decrease at most exponentially fast.444This is a consequence of Gronwall’s lemma, see e.g., [22]. This is a relatively mild assumption that holds, for example, for any polynomial decay $\varepsilon_{0}/(t+1)^{a}$, $a\in\mathbb{N}$. We start with the main result of this section. ###### Theorem 3.1. Let $x_{N}$ be the solution of (CN), and let $c_{1},c_{2}\geq 0$. There exist $C_{0},C_{1},C_{2}\geq 0$, depending on $c_{1}$, $c_{2}$, such that for all $(\varepsilon,\alpha)$ for which Assumption 1 holds with constants $c_{1}$ and $c_{2}$, the corresponding solution $x$ of (VM-DIN-AVD) is such that for all $t\geq 0$, $\|x(t)-x_{N}(t)\|\leq C_{0}e^{-\frac{t}{\beta}}\varepsilon_{0}\|\dot{x}_{0}\|+C_{1}\sqrt{\varepsilon(t)}+C_{2}\int_{s=0}^{t}e^{\frac{1}{\beta}(s-t)}\sqrt{\varepsilon(s)}\mathop{}\\!\mathrm{d}s.$ (1) This extends a previous result from [2, Proposition 3.1] which states a similar bound for constant $\varepsilon$, $\alpha\equiv 0$ and $\beta=1$. Theorem 3.1 corresponds to the blue parts in the phase diagram of Figure 1 (see also Corollary 3.6 below). ###### Remark 3.2. The strength of the above result comes from the fact that the constants $C_{0},C_{1},C_{2}$ do not depend on $\varepsilon$ and $\alpha$, and that the result is _non-asymptotic_. This allows in particular choosing $(\varepsilon,\alpha)$ to control the distance from $x$ to $x_{N}$, for all time $t\geq 0$. ###### Remark 3.3. Under Assumption 1, the dynamics (VM-DIN-AVD) is dominated by the variable mass. The damping $\alpha$ does not appear in Theorem 3.1. The above theorem and remarks emphasize the “Newtonian nature” of (VM-DIN- AVD). We present two lemmas before proving Theorem 3.1, and then state a simpler bound than (1), see Corollary 3.6. ###### Lemma 3.4. Let $(\varepsilon,\alpha)$, and let $x$ be the corresponding solution of (VM- DIN-AVD). For all $t\geq 0$, define the function, $U(t)=\frac{\varepsilon(t)}{2}\|\dot{x}(t)\|^{2}+f(x(t))-f(x^{\star}).$ Then $U$ is differentiable and for all $t>0$, $\frac{\mathop{}\\!\mathrm{d}U}{\mathop{}\\!\mathrm{d}t}(t)=\frac{\varepsilon^{\prime}(t)}{2}\|\dot{x}(t)\|^{2}-\alpha(t)\|\dot{x}(t)\|^{2}-\beta\langle\nabla^{2}f(x(t))\dot{x}(t),\dot{x}(t)\rangle\leq 0.$ Therefore, in particular, $U$ is non-increasing. ###### Proof. Let $t\geq 0$, since $x$ is twice differentiable, $U$ is differentiable and, $\frac{\mathop{}\\!\mathrm{d}U}{\mathop{}\\!\mathrm{d}t}(t)=\frac{\varepsilon^{\prime}(t)}{2}\|\dot{x}(t)\|^{2}+\varepsilon(t)\langle\dot{x}(t),\ddot{x}(t)\rangle+\langle\dot{x}(t),\nabla f(x(t))\rangle.$ We use the fact that $x$ is solution of (VM-DIN-AVD), to substitute $\varepsilon(t)\ddot{x}(t)$ by its expression, $\frac{\mathop{}\\!\mathrm{d}U}{\mathop{}\\!\mathrm{d}t}(t)=\frac{\varepsilon^{\prime}(t)}{2}\|\dot{x}(t)\|^{2}-\alpha(t)\|\dot{x}(t)\|^{2}-\beta\langle\nabla^{2}f(x(t))\dot{x}(t),\dot{x}(t)\rangle.$ By assumption $\varepsilon$ is non-increasing so for all $t>0$, $\varepsilon^{\prime}(t)\leq 0$. Furthermore $f$ is convex so $\langle\nabla^{2}f(x(t))\dot{x}(t),\dot{x}(t)\rangle\geq 0$. Hence $U$ is non-increasing. ∎ We then state the following bound. ###### Lemma 3.5. There exists $C\geq 0$ such that for any $(\varepsilon,\alpha)$ and the corresponding solution $x$ of (VM-DIN-AVD), for all $t\geq 0$ it holds that, $\varepsilon(t)\|\dot{x}(t)\|\leq C\sqrt{\varepsilon(t)}.$ ###### Proof. Let $t\geq 0$, according to Lemma 3.4, $U$ is non-increasing so $U(t)\leq U(0)$, or equivalently, $\frac{\varepsilon(t)}{2}\|\dot{x}(t)\|^{2}+f(x(t))-f(x^{\star})\leq\frac{\varepsilon_{0}}{2}\|\dot{x}_{0}\|^{2}+f(x_{0})-f(x^{\star}).$ This implies in particular that, $\varepsilon(t)\|\dot{x}(t)\|^{2}\leq\varepsilon_{0}\|\dot{x}_{0}\|^{2}+2(f(x_{0})-f(x^{\star})),$ and hence by multiplying both sides by $\varepsilon(t)$ and composing with the square-root we obtain that, $\varepsilon(t)\|\dot{x}(t)\|\leq C\sqrt{\varepsilon(t)},$ where $C=\sqrt{\varepsilon_{0}\|\dot{x}_{0}\|^{2}+2(f(x_{0})-f(x^{\star}))}$. ∎ ###### Proof of Theorem 3.1. Let $(\varepsilon,\alpha)$ as defined in Sections 1.1 and 2, and let $x$ be the corresponding solution of (VM-DIN-AVD). Then, according to Lemma 3.4, for all $t\geq 0$, $U(t)\leq U(0)$, so in particular $f(x(t))\leq f(x_{0})+\frac{\varepsilon_{0}}{2}\|\dot{x}_{0}\|^{2}.$ Denoting $M_{0}=f(x_{0})+\frac{\varepsilon_{0}}{2}\|\dot{x}_{0}\|^{2}$, the set $\mathsf{K}_{0}=\left\\{y\in\mathbb{R}^{n}\mid f(y)\leq M_{0}\right\\}$ is bounded, since $f$ is coercive ($\lim_{\|y\|\to+\infty}f(y)=+\infty$). So for all $t\geq 0$, $x(t)\in\mathsf{K}_{0}$. Since $M_{0}$ (and hence $\mathsf{K}_{0}$) depends only on $\varepsilon_{0}$, $x_{0}$ and $\dot{x}_{0}$, we have proved that for any choice $(\varepsilon,\alpha)$, the corresponding solution $x$ of (VM-DIN-AVD) is inside $\mathsf{K}_{0}$ at all time. Let $x_{N}$ be the solution of (CN). One can see similarly that for all $t\geq 0$, $f(x_{N}(t))\leq f(x_{N}(0))=f(x_{0})\leq M_{0}$. So we also have $x_{N}(t)\in\mathsf{K}_{0}$ for all $t\geq 0$. Now, fix $c_{1},c_{2}>0$, and let $(\varepsilon,\alpha)$ such that Assumption 1 is satisfied with constants $c_{1},c_{2}$. Let $x$ be the corresponding solution of (VM-DIN-AVD). Since $f$ is strongly convex on bounded sets, it is strongly convex on $\mathsf{K}_{0}$. We denote $\mu>0$ the strong-convexity parameter of $f$ on $\mathsf{K}_{0}$. Equivalently, we have that $\nabla f$ is strongly monotone on $\mathsf{K}_{0}$, that is, $\forall y_{1},y_{2}\in\mathsf{K}_{0}$, $\langle\nabla f(y_{1})-\nabla f(y_{2}),y_{1}-y_{2}\rangle\geq\mu\|y_{1}-y_{2}\|^{2}.$ (2) Let $t\geq 0$, since $x(t)\in\mathsf{K}_{0}$ and $x_{N}(t)\in\mathsf{K}_{0}$, by combining (2) with the Cauchy–Schwarz inequality, we deduce that $\|x(t)-x_{N}(t)\|\leq\frac{1}{\mu}\|\nabla f(x(t))-\nabla f(x_{N}(t))\|.$ (3) Therefore, it is sufficient to bound the difference of gradients in order to bound $\|x(t)-x_{N}(t)\|$. First, remark that (CN) can be rewritten as follows: $\frac{\mathop{}\\!\mathrm{d}}{\mathop{}\\!\mathrm{d}t}\nabla f(x_{N}(t))+\frac{1}{\beta}\nabla f(x_{N}(t))=0$. So we can integrate, for all $t\geq 0$, $\nabla f(x_{N}(t))=e^{-\frac{t}{\beta}}\nabla f(x_{0}).$ (4) We now turn our attention to $\nabla f(x(t))$, for which we cannot find a closed-form solution in general. We rewrite (VM-DIN-AVD) in the following equivalent form $\frac{\mathop{}\\!\mathrm{d}}{\mathop{}\\!\mathrm{d}t}\left[\varepsilon(t)\dot{x}(t)+\beta\nabla f(x(t))\right]+\frac{1}{\beta}\varepsilon(t)\dot{x}(t)+\nabla f(x(t))=\left(\frac{1}{\beta}\varepsilon(t)+\varepsilon^{\prime}(t)-\alpha(t)\right)\dot{x}(t).$ Introducing the variable $\omega(t)=\varepsilon(t)\dot{x}(t)+\beta\nabla f(x(t))$, the latter is thus solution to $\begin{cases}\dot{\omega}(t)+\frac{1}{\beta}\omega(t)=\left(\frac{1}{\beta}\varepsilon(t)+\varepsilon^{\prime}(t)-\alpha(t)\right)\dot{x}(t),\quad t\geq 0,\\\ \omega(0)=\varepsilon_{0}\dot{x}_{0}+\beta\nabla f(x_{0}).\end{cases}$ This is a non-homogeneous first-order ODE in $\omega$, whose solution can be expressed using the integrating factor $\omega(t)=e^{-\frac{t}{\beta}}(\varepsilon_{0}\dot{x}_{0}+\beta\nabla f(x_{0}))+e^{-\frac{t}{\beta}}\int_{0}^{t}e^{\frac{s}{\beta}}\left(\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)-\alpha(s)\right)\dot{x}(s)\mathop{}\\!\mathrm{d}s.$ We thus have the following expression for $\nabla f(x)$, for all $t\geq 0$, $\beta\nabla f(x(t))=\beta e^{-\frac{t}{\beta}}\nabla f(x_{0})+e^{-\frac{t}{\beta}}\varepsilon_{0}\dot{x}_{0}-\varepsilon(t)\dot{x}(t)+e^{-\frac{t}{\beta}}\int_{0}^{t}e^{\frac{s}{\beta}}\left(\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)-\alpha(s)\right)\dot{x}(s)\mathop{}\\!\mathrm{d}s.$ (5) We can now use (4) and (5) in (3) to get $\|x(t)-x_{N}(t)\|\leq\frac{1}{\beta\mu}\left\|e^{-\frac{t}{\beta}}\varepsilon_{0}\dot{x}_{0}-\varepsilon(t)\dot{x}(t)+e^{-\frac{t}{\beta}}\int_{0}^{t}e^{\frac{s}{\beta}}\left(\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)-\alpha(s)\right)\dot{x}(s)\mathop{}\\!\mathrm{d}s\right\|.$ Using the triangle inequality, we obtain, $\|x(t)-x_{N}(t)\|\leq\frac{\varepsilon_{0}\|\dot{x}_{0}\|}{\beta\mu}e^{-\frac{t}{\beta}}+\frac{\varepsilon(t)\|\dot{x}(t)\|}{\beta\mu}+\frac{1}{\beta\mu}\int_{0}^{t}e^{\frac{1}{\beta}(s-t)}\left|\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)-\alpha(s)\right|\|\dot{x}(s)\|\mathop{}\\!\mathrm{d}s.$ (6) The first term in (6) corresponds to the first one in (1) with $C_{0}=1/(\beta\mu)$. As for the second-one, by direct application of Lemma 3.5, there exists $C>0$ such that for all $t\geq 0$, $\frac{\varepsilon(t)\|\dot{x}(t)\|}{\beta\mu}\leq C\sqrt{\varepsilon(t)}$, so we set $C_{1}=C/(\beta\mu)$. Regarding the last term in (6), using Assumption 1 and again Lemma 3.5, it holds that, for all $s\geq 0$, $\left|\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)-\alpha(s)\right|\|\dot{x}(s)\|\leq\left(\frac{1}{\beta}+c_{1}+c_{2}\right)\varepsilon(s)\|\dot{x}(s)\|\leq\left(\frac{1}{\beta}+c_{1}+c_{2}\right)C\sqrt{\varepsilon(s)}.$ This proves the theorem with $C_{2}=\frac{C}{\beta\mu}\left(\frac{1}{\beta}+c_{1}+c_{2}\right)$. ∎ Let us analyze the bound in Theorem 3.1. The first term in (1) decays exponentially fast and can even be zero if the initial speed is $\dot{x}_{0}=0$, the second-one decays like $\sqrt{\varepsilon(t)}$, however, the rate at which the last term decreases is less obvious. The following corollary gives a less-tight but easier-to-understand estimate. ###### Corollary 3.6. Consider the same assumptions and variables as in Theorem 3.1. If furthermore $c_{1}<\frac{2}{\beta}$, then there exists $C_{3}>0$ such that for all $t\geq 0$, $\|x(t)-x_{N}(t)\|\leq C_{0}e^{-\frac{t}{\beta}}\varepsilon_{0}\|\dot{x}_{0}\|+C_{3}\sqrt{\varepsilon(t)}.$ ###### Proof of Corollary 3.6. For all $t\geq 0$, define $J(t)=\int_{0}^{t}e^{\frac{s}{\beta}}\sqrt{\varepsilon(s)}\mathop{}\\!\mathrm{d}s$. An integration by parts yields $J(t)=\left[\beta e^{\frac{s}{\beta}}\sqrt{\varepsilon(s)}\right]_{s=0}^{t}-\int_{s=0}^{t}\beta e^{\frac{s}{\beta}}\frac{\varepsilon^{\prime}(s)}{2\sqrt{\varepsilon(s)}}\mathop{}\\!\mathrm{d}s=\beta e^{\frac{t}{\beta}}\sqrt{\varepsilon(t)}-\beta\varepsilon_{0}+\int_{s=0}^{t}\beta e^{\frac{s}{\beta}}\frac{-\varepsilon^{\prime}(s)}{2\varepsilon(s)}\sqrt{\varepsilon(s)}\mathop{}\\!\mathrm{d}s.$ (7) By assumption, $0\leq c_{1}<\frac{2}{\beta}$ such that for all $s>0$, $|\varepsilon^{\prime}(s)|\leq c_{1}\varepsilon(s)$, which in our setting is equivalent to $\frac{-\varepsilon^{\prime}(s)}{\varepsilon(s)}\leq c_{1}$. So we deduce from (7) that $J(t)\leq\beta e^{\frac{t}{\beta}}\sqrt{\varepsilon(t)}+c_{1}\frac{\beta}{2}\int_{s=0}^{t}e^{\frac{s}{\beta}}\sqrt{\varepsilon(s)}\mathop{}\\!\mathrm{d}s=\beta e^{\frac{t}{\beta}}\sqrt{\varepsilon(t)}+c_{1}\frac{\beta}{2}J(t).$ So, $\left(1-c_{1}\frac{\beta}{2}\right)J(t)\leq\beta e^{t}\sqrt{\varepsilon(t)}$. By assumption $1-c_{1}\frac{\beta}{2}>0$, therefore, $J(t)\leq\frac{2}{2-c_{1}\beta}e^{\frac{t}{\beta}}\sqrt{\varepsilon(t)}$. Finally, using this in (1) and setting $C_{3}=C_{1}+C_{2}\frac{2}{2-c_{1}\beta}$, we obtain the result. ∎ ###### Remark 3.7. The above proofs suggest that an extension to the case where $f$ is non-smooth but strongly convex is possible using regularization techniques. This is left for future work. So far our results only cover the case where $\alpha$ is “not too large” w.r.t. $\varepsilon$, and do not study (LM). We now state a more general result that covers these cases. ### 3.2 Generalization to Arbitrary Viscous Dampings with Sub-exponential Decay This time we do not assume a link between $\varepsilon$ and $\alpha$ but only sub-exponential decays. ###### Assumption 2. There exists $c_{1},c_{2}\geq 0$ such that for all $t\geq 0$, $|\varepsilon^{\prime}(t)|\leq c_{1}\varepsilon(t)$ and $|\alpha^{\prime}(t)|\leq c_{2}\alpha(t)$. We are now in position to state the main result of this section. ###### Theorem 3.8. Let $x_{N}$ and $x_{LM}$ be the solution of (CN) and (LM) respectively, and let $c_{1},c_{2}\geq 0$. There exist constants $C,\tilde{C}\geq 0$, depending on $c_{1}$, $c_{2}$, such that for all $\varepsilon$ and $\alpha$ for which Assumption 2 holds with $c_{1}$ and $c_{2}$, the corresponding solution $x$ of (VM-DIN-AVD) is such that for all $t\geq 0$, $\|x(t)-x_{N}(t)\|\leq C\left[e^{-\frac{t}{\beta}}+\sqrt{\varepsilon(t)}+\alpha(t)+\int_{s=0}^{t}e^{\frac{1}{\beta}(s-t)}(\sqrt{\varepsilon(s)}+\alpha(s))\mathop{}\\!\mathrm{d}s\right],$ (8) and, $\|x(t)-x_{LM}(t)\|\leq\tilde{C}\left[e^{-\frac{t}{\beta}}+\sqrt{\varepsilon(t)}+\alpha(t)+\int_{s=0}^{t}e^{\frac{1}{\beta}(s-t)}(\sqrt{\varepsilon(s)}+\alpha(s))\mathop{}\\!\mathrm{d}s\right].$ (9) The proof is postponed to Appendix B. Although it follows a similar reasoning as that of Theorem 3.1, more involved estimates are needed. Let us comment on these results. The bound (8) generalizes Theorem 3.1, although the constant involved will, in general, be larger than those in (1) (see the proof of Theorem 3.8 in appendix). Theorem 3.8 allows for far more flexibility in the choice of $\varepsilon$ and $\alpha$ in order to control $x$ and make it possibly close to $x_{N}$. The bound in (9) is the same as that in (8) (up to a constant), but this time w.r.t. $x_{LM}$, thus connecting (VM-DIN-AVD) to (LM). We make the following two important remarks. First (9) involves $\alpha$, suggesting that making $x$ close to $x_{LM}$ requires not only $\varepsilon$ but also $\alpha$ to vanish asymptotically. Additionally, Theorem 3.8 does not state to which of $x_{N}$ and $x_{LM}$ the solution of (VM-DIN-AVD) is the closest. It remains an open question to know whether one can make (9) independent of $\alpha$, and to state to which trajectory $x$ is the closest. Yet, the numerical experiments in Section 5 suggest that neither are possible. Indeed, we observe that for some functions $f$, $x$ is _sometimes_ closer to $x_{N}$ than to $x_{LM}$, even when $\varepsilon(t)\leq\alpha(t)$. Nevertheless, Theorem 3.8 answers the question asked in the introduction: yes, (VM-DIN-AVD) is really of second-order nature since it can be brought close to the second-order dynamics (CN) and (LM). Doing so, it benefits from the good properties of these methods, such as the robustness to bad conditioning, as previously illustrated on the right of Figure 1. This concludes the analysis from a control perspective. We will now derive an approximation of the solution $x$ in order to study the impact that $\varepsilon$ and $\alpha$ have on the speed of convergence of $x$ to $x^{\star}$ compared to the speeds of convergence of $x_{N}$ and $x_{LM}$. ## 4 Approximate Solutions and Asymptotic Analysis on Quadratic Functions We consider the particular case where $f$ is a strongly-convex quadratic function in order to study the asymptotic behavior of (VM-DIN-AVD) w.r.t. (CN) and (LM). Quadratic functions are the prototypical example of strongly-convex functions. In particular, any strongly-convex function can be locally approximated by a quadratic one around its minimizer, making the latter a good model for understanding the local behavior of dynamics. In this section, $f$ is quadratic: $f(y)=\frac{1}{2}\|Ay-b\|_{2}^{2}$ for all $y\in\mathbb{R}^{n}$, where $A\in\mathbb{R}^{n\times n}$ is symmetric positive definite and $b\in\mathbb{R}^{n}$. Without loss of generality, we take $b=0$, so that the unique minimum is $x^{\star}=0$. ### 4.1 Setting: the Special Case of Quadratic Functions Quadratic functions are particularly interesting in our setting since DIN-like ODEs take a simpler form (as observed in [10, 40]). Indeed, $\forall y\in\mathbb{R}^{n}$, $\nabla f(y)=A^{T}Ay$ and $\nabla^{2}f(y)=A^{T}A$. Since $\nabla^{2}f(y)$ is independent of $y$ we can rewrite (VM-DIN-AVD) in an eigenspace555This can be generalized to the case where $A^{T}A$ is only semi- definite by considering orthogonal projections on an eigenspace spanned by the positive eigenvalues of $A^{T}A$. of $A^{T}A$. That is, we can study the ODE coordinate-wise by looking at one-dimensional problems of the form $\varepsilon(t)\ddot{x}(t)+(\alpha(t)+\beta\lambda)\dot{x}(t)+\lambda x(t)=0,\quad t\geq 0.$ (Q1-VM-DIN-AVD) Here (and throughout what follows) $\lambda>0$ denotes any eigenvalue of $A^{T}A$ and $x\colon\mathbb{R}_{+}\to\mathbb{R}$ now denotes the corresponding coordinate (function) of the solution of (VM-DIN-AVD) in an eigenspace of $A^{T}A$. The dynamics (Q1-VM-DIN-AVD) is a _linear_ second- order ODE in $x$ with non-constant coefficients. Similarly, (LM) can be rewritten coordinate-wise as $(\alpha(t)+\beta\lambda)\dot{x}_{LM}(t)+\lambda x_{LM}(t)=0,\quad t\geq 0,$ (Q1-LM) where $x_{LM}\colon\mathbb{R}_{+}\to\mathbb{R}$, and (CN) becomes $\beta\dot{x}_{N}(t)+x_{N}(t)=0,\quad t\geq 0,$ (Q1-CN) where again, $x_{N}\colon\mathbb{R}_{+}\to\mathbb{R}$ is one-dimensional. Observe in particular that (CN) and (LM) are now first-order _linear_ ODEs, whose solutions have the closed forms, for all $t\geq 0$, $x_{N}(t)=x_{0}e^{-\frac{t}{\beta}}\quad\text{and}\quad x_{LM}(t)=x_{0}\exp\left(-\int_{0}^{t}\frac{\lambda}{\alpha(s)+\beta\lambda}\mathop{}\\!\mathrm{d}s\right).$ (10) Since the minimizer is $x^{\star}=0$, we directly see that $x_{N}$ converges exponentially fast to $x^{\star}$, with a rate independent of $\lambda$. The rate of $x_{LM}$ depends however on $\lambda$ and how fast $\alpha$ vanishes. Unfortunately, except for some special choices of $\varepsilon$ and $\alpha$ (see [10]), one cannot solve the second-order linear ODE (Q1-VM-DIN-AVD) in closed form in general. Additionally, it is hopeless to circumvent the difficulty by finding a closed form for $\nabla f(x)$, accordingly to what we did in Section 3, since here $\nabla f(x)=\lambda x$. In order to study the speed of convergence of $x$ despite not having access to a closed form, we will approximate it with a controlled error, via a method that we now present. ### 4.2 The Liouville–Green Method In what follows, we rely on the Liouville–Green method [30, 26], a technique for obtaining _non-asymptotic_ approximations to solutions of linear second- order ODEs with non-constant coefficients. First, we give the intuition behind the method, following the presentation of [35]. Consider the differential equation $\ddot{z}(t)-r(t)z(t)=0,\quad t\geq 0,$ (11) where $r$ is real-valued, positive, and twice continuously differentiable. Any linear second-order ODE can be reformulated in the form (11), see Lemma 4.5 below. Since for all $t\geq 0$, $r(t)\neq 0$, we can use the changes of variables $\tau=\int_{0}^{t}\sqrt{r(s)}\mathop{}\\!\mathrm{d}s$ and $w=r^{1/4}z$ and show that $w$ is solution to $\ddot{w}(\tau)-(1+\psi(\tau))w(\tau)=0,\quad t\geq 0,$ (12) where666We express $\psi(\tau)$ using $t$ for the sake of readability, using the one-to-one correspondence between $\tau$ and $t$. $\psi(\tau)=\frac{4r(t)r^{\prime\prime}(t)-5r^{\prime}(t)^{2}}{16r(t)^{3}}$. The LG method consists in neglecting the term $\psi(\tau)$ in (12), which simply yields two approximate solutions $\hat{w}_{1}(\tau)=e^{\tau}$ and $\hat{w}_{2}(\tau)=e^{-\tau}$. Expressing this in terms of $z$ and $t$, we obtain $\hat{z}_{1}(t)=r(t)^{-1/4}\exp\left(\int_{0}^{t}\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right)\quad\text{and}\quad\hat{z}_{2}(t)=r(t)^{-1/4}\exp\left(\int_{0}^{t}-\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right).$ (13) Those are the LG approximations of the solutions of (11). They are formally valid on any interval $[0,T]$, $T>0$ when $\psi$ is “not too large”, provided that $\sqrt{r}$ is integrable on $[0,T]$. ###### Remark 4.1. There exists other (but less intuitive) ways to derive the approximations above, which allow for generalization to higher-order linear ODEs, see e.g., [14, Chapter 10]. The advantage of this approach is the possibility to estimate the error made using (13) w.r.t. the true solutions of (11). This is expressed in the following theorem which gathers results from [15, 36, 42]. ###### Theorem 4.2 (Olver [35]). Let $r\colon\mathbb{R}_{+}\to\mathbb{R}$ be a real, positive, twice continuously differentiable function, and define $\varphi(t)=\frac{4r(t)r^{\prime\prime}(t)-5r^{\prime}(t)^{2}}{16r(t)^{5/2}}$ for all $t\geq 0$. Then for any $T>0$, the differential equation, $\ddot{z}(t)-r(t)z(t)=0,\quad t\in[0,T],$ (14) has two real and twice continuously differentiable solutions defined for all $t\in[0,T]$ by, $z_{1}(t)=\frac{1+\delta_{1}(t)}{r(t)^{1/4}}\exp\left(\int_{0}^{t}\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right)\quad\text{and}\quad z_{2}(t)=\frac{1+\delta_{2}(t)}{r(t)^{1/4}}\exp\left(-\int_{0}^{t}\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right),$ where $\displaystyle|\delta_{1}(t)|\leq\exp\left(\frac{1}{2}\int_{0}^{t}|\varphi(s)|\mathop{}\\!\mathrm{d}s\right)-1$ and $\displaystyle|\delta_{2}(t)|\leq\exp\left(-\frac{1}{2}\int_{t}^{T}|\varphi(s)|\mathop{}\\!\mathrm{d}s\right)-1$. If in addition $\displaystyle\int_{0}^{+\infty}|\varphi(s)|\mathop{}\\!\mathrm{d}s<+\infty$, then the results above also hold for $T=+\infty$. ###### Remark 4.3. We make the following remarks regarding the above result. – Note that $z_{1}$ and $z_{2}$ in Theorem 4.2 are _exact_ solutions to (14). The LG approximations $\hat{z}_{1}$ and $\hat{z}_{2}$ are obtained by neglecting the unknown functions $\delta_{1}$ and $\delta_{2}$ in $z_{1}$ and $z_{2}$. The theorem gives a _non-asymptotic_ bound for the errors $|z_{1}(t)-\hat{z}_{1}(t)|$ and $|z_{2}(t)-\hat{z}_{2}(t)|$, $t\geq 0$. – Since we assumed $r$ to be twice continuously differentiable and positive, $\varphi$ is continuous, so it is integrable except maybe for $t\to+\infty$. – For the sake of simplicity, the formulation of Theorem 4.2 slightly differs from that in [35], the original formulation can be recovered by a change of variable. ### 4.3 Liouville–Green Approximation of (VM-DIN-AVD) We now proceed to make use of the LG method for approximating the solutions of (Q1-VM-DIN-AVD). The reader only interested in the result can jump directly to the Section 4.4. We first make the following assumption. ###### Assumption 3. The functions $\alpha$ and $\varepsilon$ are three times continuously differentiable, and $\varepsilon_{0}$ is such that $\forall t\geq 0$, $\varepsilon_{0}<\frac{(\beta\lambda)^{2}}{2|\alpha^{\prime}(t)|+4\lambda}$. ###### Remark 4.4. The condition on $\varepsilon_{0}$ in Assumption 3 is only technical, so that $r$ defined below is positive. It can be easily satisfied since $|\alpha^{\prime}(t)|$ is uniformly bounded. Indeed, $\alpha$ is non- increasing and non-negative (see Section 1.1), from which one can deduce that $\int_{0}^{+\infty}|\alpha^{\prime}(s)|\mathop{}\\!\mathrm{d}s\leq\alpha_{0}$. We now rewrite (Q1-VM-DIN-AVD) in the form (14). ###### Lemma 4.5. Suppose that Assumption 3 holds, and let $x$ be the solution of (Q1-VM-DIN- AVD). For all $t\geq 0$, define $p(t)=\frac{\alpha(t)+\beta\lambda}{\varepsilon(t)}\quad\text{and}\quad r(t)=\frac{p(t)^{2}}{4}+\frac{p^{\prime}(t)}{2}-\frac{\lambda}{\varepsilon(t)}.$ (15) Then, $p$ and $r$ are twice continuously differentiable, $r$ is positive and the function $y$ defined for all $t\geq 0$ by $y(t)=x(t)\exp\left(\int_{0}^{t}\frac{p(s)}{2}\mathop{}\\!\mathrm{d}s\right)$ is a solution to $\ddot{y}(t)-r(t)y(t)=0,\quad t\geq 0,$ (16) with initial condition $(y(0),\dot{y}(0))=(x_{0},\dot{x}_{0}+\frac{p(0)}{2}x_{0})$. ###### Proof. We first check that for all $t\geq 0$, $r(t)$ is positive. Let $t>0$, $r(t)>0\iff\frac{(\alpha(t)+\beta\lambda)^{2}}{4\varepsilon(t)^{2}}+\frac{\alpha^{\prime}(t)}{2\varepsilon(t)}-\frac{(\alpha(t)+\beta\lambda)\varepsilon^{\prime}(t)}{\varepsilon(t)^{2}}-\frac{\lambda}{\varepsilon(t)}>0.$ (17) Since $\varepsilon^{\prime}(t)\leq 0$ and $\alpha^{\prime}(t)\leq 0$, one can check that a sufficient condition for (17) to hold is, $r(t)>0\impliedby\frac{(\alpha(t)+\beta\lambda)^{2}}{4}>\left(\frac{|\alpha^{\prime}(t)|}{2}+\lambda\right)\varepsilon(t)\impliedby\frac{(\beta\lambda)^{2}}{2|\alpha^{\prime}(t)|+4\lambda}>\varepsilon_{0}.$ So under Assumption 3, for all $t\geq 0$, $r(t)>0$. We then check that $y$ is indeed solution to (16). Let $t>0$, $\dot{y}(t)=\frac{p(t)}{2}x(t)\exp\left(\int_{0}^{t}\frac{p(s)}{2}\mathop{}\\!\mathrm{d}s\right)+\dot{x}(t)\exp\left(\int_{0}^{t}\frac{p(s)}{2}\mathop{}\\!\mathrm{d}s\right),$ and $\ddot{y}(t)=\exp\left(\int_{0}^{t}\frac{p(s)}{2}\mathop{}\\!\mathrm{d}s\right)\left[\left(\frac{p(t)^{2}}{4}+\frac{p^{\prime}(t)}{2}\right)x(t)+p(t)\dot{x}(t)+\ddot{x}(t)\right].$ Since $x$ solves (Q1-VM-DIN-AVD), it holds that $\ddot{x}(t)=-p(t)\dot{x}(t)-\frac{\lambda}{\varepsilon(t)}x(t)$, so, $\ddot{y}(t)=\exp\left(\int_{0}^{t}\frac{p(s)}{2}\mathop{}\\!\mathrm{d}s\right)\left(\frac{p(t)^{2}}{4}+\frac{p^{\prime}(t)}{2}-\frac{\lambda}{\varepsilon(t)}\right)x(t)\\\ =\left(\frac{p(t)^{2}}{4}+\frac{p^{\prime}(t)}{2}-\frac{\lambda}{\varepsilon(t)}\right)y(t)=r(t)y(t).$ ∎ Lemma 4.5 gives a reformulation of (Q1-VM-DIN-AVD) suited to apply Theorem 4.2. To use the theorem for all $t\geq 0$, we need to ensure that $\varphi(t)=\frac{4r(t)r^{\prime\prime}(t)-5r^{\prime}(t)^{2}}{16r(t)^{5/2}}$ is integrable. To this aim we make the following assumption. ###### Assumption 4. The functions $\varepsilon$ and $\alpha$ have first, second and third-order derivatives that are integrable on $[0,+\infty[$. In addition, $\lim\limits_{t\to\infty}{\varepsilon(t)}=0$ and $\varepsilon^{\prime}(t)^{2}/\varepsilon(t)$ is integrable on $[0,+\infty[$. ###### Remark 4.6. Assumption 4 holds for most decays used in practice, with in particular any polynomial decay of the form $\frac{\varepsilon_{0}}{(t+1)^{a}}$ and $\frac{\alpha_{0}}{(t+1)^{b}}$, $a\in\mathbb{N}\setminus\\{0\\}$ and $b\in\mathbb{N}$. Note that $\varepsilon$ and $\alpha$ need not be integrable and $\alpha$ can even be constant. The next lemma states the integrability of $\varphi$ on $[0,+\infty[$. ###### Lemma 4.7. Under Assumption 3 and 4, $\int_{0}^{+\infty}|\varphi(s)|\mathop{}\\!\mathrm{d}s<+\infty$. The proof of this lemma, relies on relatively simple arguments but involves long computations and is thus postponed to Appendix C. We can now use Theorem 4.2 to obtain an exact form for the solution of (Q1-VM-DIN-AVD) based on the LG approximations. ###### Theorem 4.8. Suppose that Assumptions 3 and 4 hold. There exists $A,B\in\mathbb{R}$ such that $x(0)=x_{0}$, $\dot{x}(0)=\dot{x}_{0}$ and for all $t\geq 0$, the solution of (Q1-VM-DIN-AVD) is $\displaystyle\begin{split}&x(t)=A\frac{1+\delta_{1}(t)}{r(t)^{1/4}}\frac{\sqrt{\alpha(t)+\beta\lambda}}{\sqrt{\varepsilon(t)}}\exp\left(\int_{0}^{t}-\frac{\lambda}{\alpha(s)+\beta\lambda}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right)\\\ +&B\frac{1+\delta_{2}(t)}{r(t)^{1/4}}\frac{\sqrt{\varepsilon(t)}}{\sqrt{\alpha(t)+\beta\lambda}}\exp\left(\int_{0}^{t}-\frac{\alpha(s)+\beta\lambda}{\varepsilon(s)}+\frac{\lambda}{\alpha(s)+\beta\lambda}+\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right),\end{split}$ (18) where for all $t\geq 0$, $|\delta_{1}(t)|\leq\exp\left(\frac{1}{2}\int_{0}^{t}|\varphi(s)|\mathop{}\\!\mathrm{d}s\right)-1\quad\text{and}\quad|\delta_{2}(t)|\leq\exp\left(-\frac{1}{2}\int_{t}^{+\infty}|\varphi(s)|\mathop{}\\!\mathrm{d}s\right)-1.$ (19) Thanks to the bounds (19), we now have an approximation of $x$. We will use it in particular to compare $x$ asymptotically to the solutions of (Q1-LM) and (Q1-CN). Before this, we prove Theorem 4.8. ###### Proof of Theorem 4.8. Let $x$ be the solution of (Q1-VM-DIN-AVD) define $p,r$ as in (15). Let us also define $y(t)\stackrel{{\scriptstyle\textrm{def}}}{{=}}x(t)\exp\left(\int_{0}^{t}\frac{p(s)}{2}\mathop{}\\!\mathrm{d}s\right)$. According to Lemma 4.5, $r$ is positive and $y$ is solution to (16). Then, from Lemma 4.7, $\int_{t}^{T}|\varphi(s)|\mathop{}\\!\mathrm{d}s<+\infty$, so we can apply Theorem 4.2 to $y$ on $[0,+\infty[$. Therefore, there exists $A,B\in\mathbb{R}$, such that $\forall t\geq 0$, $y(t)=A\frac{1+\delta_{1}(t)}{r(t)^{1/4}}\exp\left(\int_{0}^{t}\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right)+B\frac{1+\delta_{2}(t)}{r(t)^{1/4}}\exp\left(\int_{0}^{t}-\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right),$ where $A$ and $B$ are determined by the initial conditions, and $\delta_{1}$, $\delta_{2}$ are such that (19) holds. Going back to $x(t)=y(t)\exp\left(\int_{0}^{t}-\frac{p(s)}{2}\mathop{}\\!\mathrm{d}s\right)$, we obtain that for all $t\geq 0$, $x(t)=A\frac{1+\delta_{1}(t)}{r(t)^{1/4}}\exp\left(\int_{0}^{t}-\frac{p(s)}{2}+\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right)+B\frac{1+\delta_{2}(t)}{r(t)^{1/4}}\exp\left(\int_{0}^{t}-\frac{p(s)}{2}-\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right).$ (20) It now remains to expand the terms in the two exponentials in (20) in order to obtain (18). To this aim, we approximate $\sqrt{r(s)}$, let $s\geq 0$, $\displaystyle\begin{split}\sqrt{r(s)}&=\frac{p(s)}{2}\sqrt{1+\frac{2p^{\prime}(s)}{p(s)^{2}}-\frac{4\lambda}{\varepsilon(s)p(s)^{2}}}\\\ &=\frac{p(s)}{2}\left(1+\frac{p^{\prime}(s)}{p(s)^{2}}-\frac{2\lambda}{\varepsilon(s)p(s)^{2}}-\frac{1}{8}\left(\frac{2p^{\prime}(s)}{p(s)^{2}}-\frac{4\lambda}{\varepsilon(s)p(s)^{2}}\right)^{2}+o(\varepsilon(s)^{2})\right)\\\ &=\frac{p(s)}{2}+\frac{p^{\prime}(s)}{2p(s)}-\frac{\lambda}{\varepsilon(s)p(s)}-\frac{1}{16}\left(\frac{2p^{\prime}(s)}{p(s)^{3/2}}-\frac{4\lambda}{\varepsilon(s)p(s)^{3/2}}\right)^{2}+o(\varepsilon(s))\\\ &=\frac{p(s)}{2}+\frac{p^{\prime}(s)\varepsilon(s)}{2(\alpha(s)+\beta\lambda)}-\frac{\lambda}{\alpha(s)+\beta\lambda}-\frac{1}{16}\left(\frac{2p^{\prime}(s)}{p(s)^{3/2}}-\frac{4\lambda\sqrt{\varepsilon(s)}}{(\alpha(s)+\beta\lambda)^{3/2}}\right)^{2}+o(\varepsilon(s))\\\ =\frac{p(s)}{2}&+\frac{\alpha^{\prime}(s)}{2(\alpha(s)+\beta\lambda)}-\frac{\varepsilon^{\prime}(s)}{2\varepsilon(s)}-\frac{\lambda}{\alpha(s)+\beta\lambda}-\frac{1}{16}\left(\frac{2p^{\prime}(s)}{p(s)^{3/2}}-\frac{4\lambda\sqrt{\varepsilon(s)}}{(\alpha(s)+\beta\lambda)^{3/2}}\right)^{2}+o(\varepsilon(s))\end{split}$ (21) Focusing on the first exponential term in (20), we deduce from (21) that for all $t\geq 0$, $\displaystyle\begin{split}&\exp\left(\int_{0}^{t}-\frac{p(s)}{2}+\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right)\\\ &=\exp\left(\int_{0}^{t}\frac{\alpha^{\prime}(s)/2}{\alpha(s)+\beta\lambda}-\frac{\varepsilon^{\prime}(s)}{2\varepsilon(s)}-\frac{\lambda}{\alpha(s)+\beta\lambda}-\frac{1}{16}\left(\frac{2p^{\prime}(s)}{p(s)^{3/2}}-\frac{4\lambda\sqrt{\varepsilon(s)}}{(\alpha(s)+\beta\lambda)^{3/2}}\right)^{2}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right)\\\ &=\frac{\sqrt{\alpha(t)+\beta\lambda})}{\sqrt{\alpha_{0}+\beta\lambda}}\frac{\sqrt{\varepsilon_{0}}}{\sqrt{\varepsilon(t)}}\exp\left(\int_{0}^{t}\frac{-\lambda}{\alpha(s)+\beta\lambda}-\frac{1}{16}\left(\frac{2p^{\prime}(s)}{p(s)^{3/2}}-\frac{4\lambda\sqrt{\varepsilon(s)}}{(\alpha(s)+\beta\lambda)^{3/2}}\right)^{2}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right)\\\ &=\frac{\sqrt{\alpha(t)+\beta\lambda})}{\sqrt{\alpha_{0}+\beta\lambda}}\frac{\sqrt{\varepsilon_{0}}}{\sqrt{\varepsilon(t)}}\exp\left(\int_{0}^{t}\frac{-\lambda}{\alpha(s)+\beta\lambda}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right),\end{split}$ where the last line relies on further computations postponed to Lemma C.1 in Appendix C. Performing the exact same type of computations on $\exp\left(\int_{0}^{t}-\frac{p(s)}{2}-\sqrt{r(s)}\mathop{}\\!\mathrm{d}s\right)$, and up to redefining $A$ and $B$ so as to encompass all the constants, we obtain (18) and the result is proved. ∎ ### 4.4 Comparison of $x$ with $x_{LM}$ and $x_{N}$ We now have an expression for $x$ which is almost explicit: we do not know $\delta_{1}$ and $\delta_{2}$ in closed form, but they are uniformly bounded (by Lemma 4.7). We will now compare the asymptotic behavior of (18) with those of the solutions of (Q1-LM) and (Q1-CN) that we denoted $x_{LM}$ and $x_{N}$ respectively. Our main result of Section 4 is the following, where $\sim_{+\infty}$ denotes the asymptotic equivalence777Two real-valued functions $g_{1}$ and $g_{2}$ are asymptotically equivalent in $+\infty$ if and only if $\lim_{t\to\infty}\frac{g_{1}(t)}{g_{2}(t)}=1$. between two functions as $t\to\infty$. ###### Theorem 4.9. Let $x$ be the solution of (Q1-VM-DIN-AVD), given in (18), and $x_{LM}$ and $x_{N}$ whose closed forms are stated in (10). Under Assumptions 3 and 4, there exists $C>0$ such that the following asymptotic equivalences hold: $\displaystyle\begin{split}x(t)&\sim_{+\infty}x_{LM}(t)C\exp\left(\int_{0}^{t}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right),\quad\text{and}\\\ x(t)&\sim_{+\infty}x_{N}(t)C\exp\left(\int_{0}^{t}\frac{\alpha(s)}{\beta(\alpha(s)+\beta\lambda)}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right).\end{split}$ (22) As a consequence, the convergence of $x$ to $x^{\star}$ is: 1. (i) _Faster_ than that of $x_{LM}$ if $\varepsilon$ is non-integrable and as fast otherwise. 2. (ii) Slower than that of $x_{N}$ if $\alpha$ is non-integrable and as fast if $\alpha$ is integrable, in the case where $\forall t\geq 0$, $\alpha(t)>\varepsilon(t)$. 3. (iii) _Faster_ than that of $x_{N}$ if $\varepsilon$ is non-integrable and as fast if $\varepsilon$ is integrable, in the case where $\forall t\geq 0$, $\alpha(t)<\varepsilon(t)$. While the results of Section 3 were related to the closeness of (VM-DIN-AVD) w.r.t. (CN) and (LM) from a control perspective, Theorem 4.9 provides a different type of insight. First, the results are asymptotic, so they only allow to control (VM-DIN-AVD) for large $t$. They provide however a clear understanding of the nature of the solutions of (VM-DIN-AVD) and their convergence. The conclusions (summarized in Table 1) are in accordance with what we would expect: when the viscous damping is larger than the variable mass, (VM-DIN-AVD) behaves more like the Levenberg–Marquardt method than the Newton one, but it actually becomes an accelerated Levenberg–Marquardt dynamics when $\varepsilon$ is non-integrable but vanishing. However, when the variable mass $\varepsilon$ is larger than $\alpha$, the dynamics is closer to the one of the Newton method, and can actually be an accelerated Newton dynamics, again for non-integrable $\varepsilon$. This is analogous to the necessary condition that $\alpha$ must be non-integrable in order to accelerate first-order methods in convex optimization (see [3]). We conclude this section by proving Theorem 4.9. ###### Proof of Theorem 4.9. Thanks to Assumptions 3 and 4, Theorem 4.8 tells us that $x$ has the form (18). We now analyze the two terms in (18). First, we know from Theorem 4.8 that $\delta_{1}(0)=0$ and $\lim_{t\to+\infty}\delta_{2}(t)=0$. In addition, by Lemma 4.7, $\delta_{1}$ and $\delta_{2}$ are uniformly bounded by some positive constant. Then $r(t)^{-1/4}$ decays asymptotically like $\sqrt{\varepsilon(t)}$ and $\alpha$ is bounded. So $A\frac{1+\delta_{1}(t)}{r(t)^{1/4}}\frac{\sqrt{\alpha(t)+\beta\lambda}}{\sqrt{\varepsilon(t)}}$ is asymptotically equivalent to some constant $c_{1}\in\mathbb{R}$ as $t\to+\infty$. Similarly, $B\frac{1+\delta_{2}(t)}{r(t)^{1/4}}\frac{\sqrt{\varepsilon(t)}}{\sqrt{\alpha(t)+\beta\lambda}}$ is equivalent to $c_{2}\varepsilon(t)$, with $c_{2}\in\mathbb{R}$. We now analyze the “exponential factors” in (18). On the one hand, $\frac{\lambda}{\alpha(s)+\beta\lambda}+\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))$ converges to $\frac{1}{\beta}$ as $s\to\infty$, while $\frac{\alpha(s)+\beta\lambda}{\varepsilon(s)}$ diverges to $+\infty$. Therefore, we deduce that, $\exp\left(\int_{0}^{t}-\frac{\alpha(s)+\beta\lambda}{\varepsilon(s)}+\frac{\lambda}{\alpha(s)+\beta\lambda}+\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right)\\\ =o\left(\exp\left(\int_{0}^{t}-\frac{\lambda}{\alpha(s)+\beta\lambda}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right)\right).$ As a consequence, the second term in (18) will decrease to $0$ faster than the first-one (let alone the additional $\varepsilon(t)$ decay that we have just discussed). The asymptotic behavior of $x$ will thus be governed by the first term in (18). Let us now focus on the first term in (18). Observe that $\exp\left(\int_{0}^{t}-\frac{\lambda}{\alpha(s)+\beta\lambda}\mathop{}\\!\mathrm{d}s\right)$ is exactly the decay of $x_{LM}$ in (10). Thus, we have proved that there exists $C>0$, such that the following asymptotic equivalence holds, $A\frac{1+\delta_{1}(t)}{r(t)^{1/4}}\frac{\sqrt{\alpha(t)+\beta\lambda}}{\sqrt{\varepsilon(t)}}\exp\left(\int_{0}^{t}-\frac{\lambda}{\alpha(s)+\beta\lambda}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right)\\\ \sim_{+\infty}x_{LM}(t)C\exp\left(\int_{0}^{t}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s\right),$ which proves the first part of (22). The second equivalence in (22) is obtained using the following identity, $\int_{0}^{t}-\frac{\lambda}{\alpha(s)+\beta\lambda}\mathop{}\\!\mathrm{d}s=\int_{0}^{t}-\frac{1}{\beta}+\frac{\alpha(s)}{\beta(\alpha(s)+\beta\lambda)}\mathop{}\\!\mathrm{d}s=-\frac{t}{\beta}+\int_{0}^{t}\frac{\alpha(s)}{\beta(\alpha(s)+\beta\lambda)}\mathop{}\\!\mathrm{d}s$ (23) and $e^{-t/\beta}$ is precisely the rate at which $x_{N}$ decreases. So (22) holds. It finally remains to deduce the conclusions of the theorem from (22). – Regarding the comparison with $x_{LM}$, the integral $\int_{0}^{t}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s$ converges if and only if $\varepsilon$ is integrable on $[0,+\infty[$, and diverges to $-\infty$ when $\varepsilon$ is not. So $x$ converges to $0$ at least as fast as $x_{LM}$ and faster when $\varepsilon$ is not integrable. – As for the comparison with $x_{N}$, if $\alpha(s)>\varepsilon(s)\geq 0$ for all $s\geq 0$, then the integral $\int_{0}^{t}\frac{\alpha(s)}{\beta(\alpha(s)+\beta\lambda)}-\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s))\mathop{}\\!\mathrm{d}s$ is convergent in $+\infty$ if and only if $\alpha$ is integrable and diverges to $+\infty$ when $\alpha$ is non-integrable. So when $\alpha$ is integrable, the speed of convergence of $x$ is the same as that of $x_{N}$. When $\alpha$ is not integrable, the convergence to $0$ is slower but still holds. Indeed, for all $s\geq 0$ $\frac{\alpha(s)}{\beta(\alpha(s)+\frac{1}{\beta})}<\frac{\alpha(s)}{\beta\alpha(s)}=\frac{1}{\beta}$. Thus for all $t>0$, $-\frac{t}{\beta}+\int_{0}^{t}\frac{\alpha(s)}{\beta(\alpha(s)+\beta\lambda)}\mathop{}\\!\mathrm{d}s<0$. – Finally, the comparison with $x_{N}$ in the case $\varepsilon(s)>\alpha(s)$ is exactly the same as the comparison with $x_{LM}$ using (23). ∎ ## 5 Numerical Experiments Figure 2: Comparison of the solutions $x_{N}$, $x_{LM}$ and $x$ of (CN), (LM) and (VM-DIN-AVD) respectively, for a strongly-convex function of the form $f(x)=e^{-\|x\|^{2}}+\frac{1}{2}\|Ax\|^{2}$. Left figures: distance $\|x(t)-x_{N}(t)\|$ versus time $t$, each curve corresponds to a different choice of $\varepsilon$; middle figures: distance $\|x(t)-x_{LM}(t)\|$, again for several $\varepsilon$. Right figures: distance to the optimum $x^{\star}$ for reference, $x_{N}$ and $x_{LM}$ are in dotted and dashed lines, other curves correspond to (VM-DIN-AVD) for several choices of $\varepsilon$. The brown curve is often hidden behind the purple (and sometimes the pink) curve. Top and bottom rows show results respectively for non-integrable and integrable viscous dampings $\alpha$. The theoretical bounds from Theorem 3.8 are only displayed on Figure 4 below, for the sake of readability. Figure 3: Similar experiment and figures as those described in Figure 2, but for the function $f(x)=\log\left(\sum_{i=1}^{n}e^{x_{i}}+e^{-x_{i}}\right)+\frac{1}{2}\|Ax\|^{2}$. Figure 4: Similar experiment and figures as those described in Figure 2, but for the function $f(x)=\|x\|^{50}+\frac{1}{2}\|Ax\|^{2}$. The thin “dash dotted” curves represent the theoretical bounds from Theorem 3.8 for each choice of $(\varepsilon,\alpha)$ considered. Figure 5: Numerical validation of Theorem 4.9: distance to the optimum $x^{\star}$ as a function of time. on a quadratic function $f(x)=\frac{1}{2}\|Ax\|^{2}$. Left: speed comparison w.r.t. (CN) for several choices of $\varepsilon$ and $\alpha$. Right: Comparison with LM in for $\alpha$ integrable or not and several choices of $\varepsilon$. Shades of blue represent cases where $\varepsilon(t)>\alpha(t)$ while shades of red represent the opposite setting. We present two set of experiments that illustrate our main results from Sections 3 and 4. We first detail the general methodology. ### 5.1 Methodology We compare the solutions of (CN), (LM) and (VM-DIN-AVD) obtained for strongly- convex functions in dimension $n=100$. Since closed-form solutions are not available, they are estimated via discretization schemes with small step-sizes $\gamma=10^{-1}$. We used Euler semi-explicit schemes, where a linear system is solved at each iteration, for the sake of stability. The resulting algorithms are detailed in Appendix D. ### 5.2 First Experiment: Distance between Trajectories We begin with an empirical validation of the results of Section 3 on the distance between $x$, $x_{LM}$ and $x_{N}$. Each of Figures 2, 3 and 4 (as well as Figures 6 in Appendix D) corresponds to a different strongly-convex function, specified below its corresponding figure. In order to ensure strong convexity, each function contains a quadratic term of the form $\|Ax\|^{2}$, where $A$ is symmetric positive definite. Several observations can be made from the numerical results, but we first note on the right plots of Figures 2 to 4 that $x_{N}$ always converges asymptotically linearly (i.e., exponentially fast). This is also the case for $x$ and $x_{LM}$ in some (but not all) cases. This is important because $\|x(t)-x_{N}(t)\|\leq\|x(t)-x^{\star}\|+\|x_{N}(t)-x^{\star}\|$, so if both $x$ and $x_{N}$ converge linearly, then the bounds of Theorems 3.1 and 3.8 need not be tight asymptotically. That being said, the strength of these results is to be non-asymptotic and this is highlighted by the experiments as we now explain. Looking at the left and middle plots of Figures 2, 3 and 4, we observe that Theorems 3.1 and 3.8 seem empirically validated, since the distances $\|x(t)-x_{N}(t)\|$ and $\|x(t)-x_{LM}(t)\|$ decrease relatively fast to zero. Again, when $x$ converges rapidly to $x^{\star}$ this is not very insightful, however, the main interest of our theorems appears on the left of Figures 3 and 4: the blue and green curves, corresponding to slowly decaying choices of $\varepsilon$, converge more slowly than other trajectories. However, when taking faster decays, we recover fast convergence and closeness to $x_{N}$ (this is particularly true for the purple curve). Very similar observations are made w.r.t. $x_{LM}$ on the middle plots. Despite not being stated in the theorems of Section 3, the experiments match the intuition that when $\varepsilon>\alpha$, $x$ may be closer to $x_{N}$ and when $\varepsilon\leq\alpha$, $x$ would rather be closer to $x_{LM}$. This is more noticeable on the top rows of the figures, where $\alpha$ is not integrable. Figure 4 suggests that the bounds in Theorem 3.8 are rather tight for small $t$, since, for example, the blue and green curves on the left show a relatively slow vanishing of $\|x(t)-x_{N}(t)\|$ for slowly decaying $\varepsilon$. The experiments show that the bounds seem however often too pessimistic for large $t$, for which the second part of our study provides better insights (see Section 4 and below). Interestingly, slow decays of $\varepsilon$ may sometimes result in faster convergence for $x$ than fast decays (and also faster convergence than $x_{LM}$), notably on Figure 2. We also note that $\varepsilon(t)=\varepsilon_{0}/t$ combined either with $\alpha(t)=\alpha_{0}/t$ or $\alpha(t)=\alpha_{0}/t^{2}$ seems to always yield fast convergence on these experiments (and sometimes the fastest of all dynamics). ### 5.3 Second Experiment: Empirical Validation of Theorem 4.9 We now turn our attention to the solutions $x$, $x_{N}$ and $x_{LM}$ for a quadratic function of the form $f(y)=\frac{1}{2}\|Ay\|^{2}$, $y\in\mathbb{R}^{n}$, and for several choices of $\varepsilon$ and $\alpha$. The results in Figure 5 exactly match the expected behavior summarized in Table 1. Indeed, looking first at the right-hand side of Figure 5, $x$ is as fast as the corresponding888That is, the solution of (LM) for the same $\alpha$ as that considered for (VM-DIN-AVD). $x_{LM}$ when $\varepsilon$ is not integrable and regardless of $\alpha$, and $x$ is faster when $\varepsilon$ is non-integrable. Then on the left-hand side, when comparing to $x_{N}$, $x$ is slower in settings where $\alpha$ is larger than $\varepsilon$ and non-integrable (red curves), or almost as fast when $\alpha$ is integrable (pink curve). However, acceleration w.r.t. to $x_{N}$ is indeed achieved for non-integrable $\varepsilon$ regardless of $\alpha$ (first-two blue curves), and the rate is the same as that of $x_{N}$ when $\varepsilon$ is integrable (third blue curve). ## 6 Conclusions and Perspectives We introduced a general ODE (VM-DIN-AVD) featuring variable mass, and provided a deep understanding on the behavior of its solutions w.r.t. time dependent control parameters $\varepsilon$ and $\alpha$, both, asymptotically and non- asymptotically. We can conclude that (VM-DIN-AVD) is indeed of (regularized) Newton type, since it can be controlled to be close to both (CN) and (LM). Yet we also showed that (VM-DIN-AVD) fundamentally differs from the other two dynamics in its nature. In particular, Theorem 4.9 and the numerical experiments emphasized that $\varepsilon$ and $\alpha$ can accelerate (or slow down) (VM-DIN-AVD) w.r.t. (CN) and (LM). We also note that our bounds in Theorems 3.1 and 3.8 seem relatively tight, in particular for functions with large gradients (see Figure 4). Our contribution yields a complete and satisfying picture on the relation between the three systems, which was only partially understood. We believe that our results build a strong foundation for the development of algorithms that combine the best properties of first- and second-order optimization methods. As for future work, we showed that (VM-DIN-AVD) is promising from an optimization perspective. So far we approximated solutions of (VM-DIN-AVD) via schemes that required solving a linear system at each iteration (this is also true for (CN) and (LM)). Our new understanding on $(\varepsilon,\alpha)$ paves the way towards designing new Newton-like algorithms with a significantly reduced computational cost, which is crucial for large-scale optimization. Another open question is whether it is possible to preserve the properties evidenced in this work when $\varepsilon$ is defined in a closed-loop manner (formally depending on $x$ rather than on $t$). Finally, it would be worth investigating how the current work can be extended to general convex and/or non-smooth functions. ## Acknowledgment C. Castera, J. Fadili and P. Ochs are supported by the ANR-DFG joint project TRINOM-DS under number ANR-20-CE92-0037-01. The numerical experiments were made thanks to the development teams of the following libraries: Python [38], Numpy [43] and Matplotlib [27]. Appendices ## Appendix A Equivalent First-order System and Global Existence of Solutions ### A.1 First-order Equivalent Formulation We reformulate (VM-DIN-AVD) as a system of ODE involving only first-order time derivatives and the gradient of $f$. For this purpose, notice that for all $t>0$ (VM-DIN-AVD) can be rewritten as, $\displaystyle\begin{split}\frac{\mathop{}\\!\mathrm{d}}{\mathop{}\\!\mathrm{d}t}\left[\varepsilon(t)\dot{x}(t)\right]+\beta\frac{\mathop{}\\!\mathrm{d}}{\mathop{}\\!\mathrm{d}t}\nabla f(x(t))+\alpha(t)\dot{x}(t)-\varepsilon^{\prime}(t)\dot{x}(t)+\nabla f(x(t))=0,\quad t\geq 0.\end{split}$ (24) We then integrate (24) for all $t\geq 0$, $\varepsilon(t)\dot{x}(t)+\beta\nabla f(x(t))-\varepsilon_{0}\dot{x}_{0}-\beta\nabla f(x_{0})+\int_{0}^{t}(\alpha(s)-\varepsilon^{\prime}(s))\dot{x}(s)+\nabla f(x(s))\mathop{}\\!\mathrm{d}s=0.$ (25) For all $t\geq 0$, we define the variable, $z(t)=\int_{0}^{t}(\alpha(s)-\varepsilon^{\prime}(s))\dot{x}(s)+\nabla f(x(s))\mathop{}\\!\mathrm{d}s-\varepsilon_{0}\dot{x}_{0}-\beta\nabla f(x_{0}).$ We differentiate $z$, for all $t>0$, $\dot{z}(t)=(\alpha(t)-\varepsilon^{\prime}(t))\dot{x}(t)+\nabla f(x(t))$, so that we can rewrite (25) as, $\begin{cases}\varepsilon(t)\dot{x}(t)+\beta\nabla f(x(t))+z(t)=0\\\ \dot{z}(t)-(\alpha(t)-\varepsilon^{\prime}(t))\dot{x}(t)-\nabla f(x(t))=0\end{cases},\quad t\geq 0.$ We substitute the first line in the second-one, $\begin{cases}\varepsilon(t)\dot{x}(t)+\beta\nabla f(x(t))+z(t)=0\\\ \beta\dot{z}(t)-\beta(\alpha(t)-\varepsilon^{\prime}(t)-\frac{1}{\beta}\varepsilon(t))\dot{x}(t)+z(t)=0\end{cases},\quad t\geq 0.$ (26) To ease the readability, we recall the notation $\nu(t)=\alpha(t)-\varepsilon^{\prime}(t)-\frac{1}{\beta}\varepsilon(t)$ from Section 2. Then define for all $t\geq 0$, $y(t)=z(t)-\nu(t)x(t)$, and differentiate, $\dot{y}(t)=\dot{z}(t)-\nu(t)\dot{x}(t)-\nu^{\prime}(t)x(t)$. We finally rewrite (26) as, $\begin{cases}\varepsilon(t)\dot{x}(t)+\beta\nabla f(x(t))&+\nu(t)x(t)+y(t)=0\\\ \dot{y}(t)+\nu^{\prime}(t)x(t)&+\frac{\nu(t)}{\beta}x(t)+\frac{1}{\beta}y(t)=0\end{cases}.{}$ which is (gVM-DIN-AVD). Finally, the initial condition on $y$ is $y(0)=z(0)-\nu(0)x(0)=-\varepsilon_{0}\dot{x}_{0}-\beta\nabla f(x_{0})-(\alpha_{0}-\varepsilon^{\prime}_{0}-\frac{1}{\beta}\varepsilon_{0})x_{0}.$ ###### Remark A.1. Notice that the quantity $\nu(t)=\alpha(t)-\varepsilon^{\prime}(t)-\frac{1}{\beta}\varepsilon(t)$ involved in (gVM-DIN-AVD) also plays a key role in our analysis of Section 3, see e.g., (6). In particular the sign of $\nu(t)$ changes the nature of (VM- DIN-AVD) and is related to Assumption 1. ### A.2 Local Solutions are Global Using the formulation (gVM-DIN-AVD), we proved local existence and uniqueness of solutions of (VM-DIN-AVD) in Section 2. Using the same notations, we justify that the local solution $(x,y)$ actually exists globally. According to Lemma 3.4, the Lyapunov function $U(t)=\frac{\varepsilon(t)}{2}\|\dot{x}(t)\|^{2}+f(x(t))-f(x^{\star})$ is non- negative and decreasing. Thus, it is uniformly bounded on $\mathbb{R}_{+}$ and the same holds for $t\mapsto f(x(t))$ since for all $t\geq 0$, $U(t)\geq f(x(t))$. Then, $f$ is coercive by assumption, so $x$ is uniformly bounded on $\mathbb{R}_{+}$ (otherwise $f(x)$ could not remain bounded). We now prove that $y$ is also uniformly bounded. From (gVM-DIN-AVD), for all $t>0$, $\dot{y}(t)=-\frac{1}{\beta}y(t)-(\frac{\nu(t)}{\beta}+\nu^{\prime}(t))x(t)$ so we can use the following integrating factor, $y(t)=e^{-\frac{t}{\beta}}y_{0}-e^{-\frac{t}{\beta}}\int_{0}^{t}\frac{1}{\beta}e^{\frac{s}{\beta}}(\nu(s)+\beta\nu^{\prime}(s))x(s)\mathop{}\\!\mathrm{d}s.$ Using triangle inequalities, for all $t\geq 0$, $\|y(t)\|\leq e^{-\frac{t}{\beta}}\|y_{0}\|+\sup_{s\geq 0}\|(\nu(s)+\beta\nu^{\prime}(s))x(s)\|e^{-\frac{t}{\beta}}\int_{0}^{t}\frac{1}{\beta}e^{\frac{s}{\beta}}\mathop{}\\!\mathrm{d}s\leq\|y_{0}\|+\sup_{s\geq 0}\|(\nu(s)+\beta\nu^{\prime}(s))x(s)\|.$ (27) Using the definition of $\varepsilon$ and $\alpha$ from Sections 1.1 and 2, observe that $\varepsilon$, $\alpha$, $\varepsilon^{\prime}$ and $\alpha^{\prime}$ are all bounded on $\mathbb{R}_{+}$, and $\varepsilon^{\prime\prime}$ is assumed to be bounded. So $\nu$ and $\nu^{\prime}$ are bounded, and since we also proved that $x$ is uniformly bounded on $\mathbb{R}_{+}$, we deduce from (27) that $y$ is uniformly bounded as well. Hence, the unique local solution $(x,y)$ is global. ## Appendix B Proof of Theorem 3.8 This section is devoted to proving the general result of Section 3. Fix some constants $c_{1},c_{2}>0$ and let $\varepsilon$ and $\alpha$ such that Assumption 2 is satisfied with these constants. Let $x$ be the corresponding solution of (VM-DIN-AVD), $x_{N}$, and $x_{LM}$ that of (CN) and (LM), respectively. Following the same arguments as in the beginning of the proof of Theorem 3.1, for all $t\geq 0$, $x(t)$, $x_{N}(t)$ and $x_{LM}(t)$ belong to the bounded set $\mathsf{K}_{0}$ defined in that proof. Since $f$ is $\mu$-strongly convex on $\mathsf{K}_{0}$, the proof relies again on bounding difference of gradients, indeed, for all $t\geq 0$, $\|x(t)-x_{N}(t)\|\leq\frac{1}{\mu}\|\nabla f(x(t))-\nabla f(x_{N}(t))\|\ \text{and}\ \|x(t)-x_{LM}(t)\|\leq\frac{1}{\mu}\|\nabla f(x(t))-\nabla f(x_{LM}(t))\|.$ (28) Recall also that the closed form of $\nabla f(x_{N})$ is given in (4). #### Expressing $\nabla f(x)$. We follow the exact same steps as in the proof of Theorem 3.1 to obtain the expression of $\nabla f(x)$ given in (5), which we recall, for all $t\geq 0$, $\beta\nabla f(x(t))=\beta e^{-\frac{t}{\beta}}\nabla f(x_{0})+e^{-\frac{t}{\beta}}\varepsilon_{0}\dot{x}_{0}-\varepsilon(t)\dot{x}(t)+\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)-\alpha(s)\right)\dot{x}(s)\mathop{}\\!\mathrm{d}s.$ Here we do not assume any relation between $\varepsilon$ and $\alpha$, and we thus need to find a more suitable expression for $\nabla f(x(t))$. We first expand the terms in the integral, for all $t\geq 0$, $\beta\nabla f(x(t))=\beta e^{-\frac{t}{\beta}}\nabla f(x_{0})+e^{-\frac{t}{\beta}}\varepsilon_{0}\dot{x}_{0}-\varepsilon(t)\dot{x}(t)\\\ +\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)\right)\dot{x}(s)\mathop{}\\!\mathrm{d}s-\int_{0}^{t}e^{\frac{s-t}{\beta}}\alpha(s)\dot{x}(s)\mathop{}\\!\mathrm{d}s.$ (29) Then, for all $s\geq 0$, we have the identity, $e^{\frac{s}{\beta}}\dot{x}(s)=e^{\frac{s}{\beta}}\dot{x}(s)+\frac{1}{\beta}e^{\frac{s}{\beta}}x(s)-\frac{1}{\beta}e^{\frac{s}{\beta}}x(s)=\frac{\mathop{}\\!\mathrm{d}}{\mathop{}\\!\mathrm{d}s}(e^{\frac{s}{\beta}}x(s))-\frac{1}{\beta}e^{\frac{s}{\beta}}x(s),$ (30) which we use to perform an integration by part on the last integral in (29), $\int_{0}^{t}e^{\frac{s}{\beta}}\alpha(s)\dot{x}(s)\mathop{}\\!\mathrm{d}s=\left[\alpha(s)e^{\frac{s}{\beta}}x(s)\right]_{0}^{t}-\int_{0}^{t}\left(\alpha^{\prime}(s)+\frac{\alpha(s)}{\beta}\right)e^{\frac{s}{\beta}}x(s)\mathop{}\\!\mathrm{d}s.$ Therefore, $e^{-\frac{t}{\beta}}\int_{0}^{t}e^{\frac{s}{\beta}}\alpha(s)\dot{x}(s)\mathop{}\\!\mathrm{d}s=\alpha(t)x(t)-e^{-\frac{t}{\beta}}\alpha_{0}x_{0}-\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\alpha^{\prime}(s)+\frac{\alpha(s)}{\beta}\right)x(s)\mathop{}\\!\mathrm{d}s,$ (31) and we can substitute in (29), $\beta\nabla f(x(t))=\beta e^{-\frac{t}{\beta}}\nabla f(x_{0})+e^{-\frac{t}{\beta}}\varepsilon_{0}\dot{x}_{0}-\varepsilon(t)\dot{x}(t)+\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)\right)\dot{x}(s)\mathop{}\\!\mathrm{d}s\\\ -\alpha(t)x(t)+e^{-\frac{t}{\beta}}\alpha_{0}x_{0}+\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\alpha^{\prime}(s)+\frac{\alpha(s)}{\beta}\right)x(s)\mathop{}\\!\mathrm{d}s.$ (32) #### Uniform boundedness. In view of exploiting (32), we recall that for all $(\varepsilon,\alpha)$, $x$ is uniformly bounded. So there exists $R>0$ such that for all $(\varepsilon,\alpha)$, the corresponding solution $x$ of (VM-DIN-AVD) is such that $\sup_{t\geq 0}\|x(t)\|\leq R.$ (33) We are now in position to prove (8). #### Distance from $x$ to $x_{N}$. We first gather (4) and (32). For all $t\geq 0$, $\beta\nabla f(x(t))-\beta\nabla f(x_{N}(t))=e^{-\frac{t}{\beta}}\varepsilon_{0}\dot{x}_{0}-\varepsilon(t)\dot{x}(t)+\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)\right)\dot{x}(s)\mathop{}\\!\mathrm{d}s\\\ +e^{-\frac{t}{\beta}}\alpha_{0}x_{0}-\alpha(t)x(t)+\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\alpha^{\prime}(s)+\frac{\alpha(s)}{\beta}\right)x(s)\mathop{}\\!\mathrm{d}s.$ We then use (28) and triangle inequalities, $\beta\mu\|x(t)-x_{N}(t)\|\leq e^{-\frac{t}{\beta}}\varepsilon_{0}\|\dot{x}_{0}\|+\varepsilon(t)\|\dot{x}(t)\|+\int_{0}^{t}e^{\frac{s-t}{\beta}}\left|\frac{\varepsilon(s)}{\beta}+\varepsilon^{\prime}(s)\right|\|\dot{x}(s)\|\mathop{}\\!\mathrm{d}s\\\ +e^{-\frac{t}{\beta}}\alpha_{0}\|x_{0}\|+\alpha(t)\|x(t)\|+\int_{0}^{t}e^{\frac{s-t}{\beta}}\left|\frac{\alpha(s)}{\beta}+\alpha^{\prime}(s)\right|\|x(s)\|\mathop{}\\!\mathrm{d}s.$ (34) By Assumption 2, for all $s\geq 0$, $|\frac{\varepsilon(s)}{\beta}+\varepsilon^{\prime}(s)|\leq(\frac{1}{\beta}+c_{1})\varepsilon(s)$ and $|\frac{\alpha(s)}{\beta}+\alpha^{\prime}(s)|\leq(\frac{1}{\beta}+c_{2})\alpha(s)$. We then use Lemma 3.5 (denoting by $C>0$ the constant stated in the lemma) on the first line of (34), and we use the boundedness (33) on the second line to obtain, $\beta\mu\|x(t)-x_{N}(t)\|\leq e^{-\frac{t}{\beta}}\varepsilon_{0}\|\dot{x}_{0}\|+C\sqrt{\varepsilon(t)}+C\left(\frac{1}{\beta}+c_{1}\right)\int_{0}^{t}e^{\frac{s-t}{\beta}}\sqrt{\varepsilon(s)}\mathop{}\\!\mathrm{d}s\\\ +e^{-\frac{t}{\beta}}\alpha_{0}\|x_{0}\|+R\alpha(t)+R\left(\frac{1}{\beta}+c_{2}\right)\int_{0}^{t}e^{\frac{s-t}{\beta}}\alpha(s)\mathop{}\\!\mathrm{d}s.$ This proves (8). #### Expressing $\nabla f(x_{LM})$. We now repeat previous arguments but for (LM). First, (LM) is equivalent to $\frac{\mathop{}\\!\mathrm{d}}{\mathop{}\\!\mathrm{d}t}\nabla f(x_{LM}(t))+\frac{1}{\beta}\nabla f(x_{LM}(t))=-\alpha(t)\dot{x}_{LM}(t).$ So using an integrating factor one can check that for all $t\geq 0$, $\nabla f(x_{LM}(t))=e^{-\frac{t}{\beta}}\nabla f(x_{0})-e^{-\frac{t}{\beta}}\int_{0}^{t}\frac{1}{\beta}e^{\frac{s}{\beta}}\alpha(s)\dot{x}_{LM}(s)\mathop{}\\!\mathrm{d}s.$ We can then follow exactly steps (30) to (31) so as to obtain, $e^{-\frac{t}{\beta}}\int_{0}^{t}e^{\frac{s}{\beta}}\alpha(s)\dot{x}_{LM}(s)\mathop{}\\!\mathrm{d}s=\alpha(t)x_{LM}(t)-e^{-\frac{t}{\beta}}\alpha_{0}x_{0}-e^{-\frac{t}{\beta}}\int_{0}^{t}\left(\alpha^{\prime}(s)+\frac{\alpha(s)}{\beta}\right)e^{\frac{s}{\beta}}x_{LM}(s)\mathop{}\\!\mathrm{d}s.$ Finally, remark that for all $t\geq 0$, $\frac{\mathop{}\\!\mathrm{d}}{\mathop{}\\!\mathrm{d}t}f(x_{LM}(t))=-\alpha(t)\|\dot{x}_{LM}(t)\|^{2}-\beta\langle\dot{x}_{LM}(t),\nabla^{2}f(x_{LM}(t))\dot{x}_{LM}(t)\rangle\leq 0.$ So $f(x_{LM}(t))\leq f(x_{0})$ and using the coercivity of $f$ as before we deduce that for all choices $\alpha$, $\sup_{t\geq 0}\|x_{LM}(t)\|\leq R.$ (35) #### Distance from $x$ to $x_{LM}$. We substract gradients, $\beta\nabla f(x(t))-\beta\nabla f(x_{LM}(t)))=e^{-\frac{t}{\beta}}\varepsilon_{0}\dot{x}_{0}-\varepsilon(t)\dot{x}(t)+\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\frac{1}{\beta}\varepsilon(s)+\varepsilon^{\prime}(s)\right)\dot{x}(s)\mathop{}\\!\mathrm{d}s\\\ -\alpha(t)(x(t)-x_{LM}(t))-\int_{0}^{t}e^{\frac{s-t}{\beta}}\left(\alpha^{\prime}(s)+\frac{\alpha(s)}{\beta}\right)(x(s)-x_{LM}(s))\mathop{}\\!\mathrm{d}s,$ and we proceed as before using (28), Assumption 2 and Lemma 3.5. It holds that, $\beta\mu\|x(t)-x_{LM}(t)\|\leq e^{-\frac{t}{\beta}}\varepsilon_{0}\|\dot{x}_{0}\|+C\sqrt{\varepsilon(t)}+C\left(\frac{1}{\beta}+c_{1}\right)\int_{0}^{t}e^{\frac{1}{\beta}(s-t)}\sqrt{\varepsilon(s)}\mathop{}\\!\mathrm{d}s\\\ +\alpha(t)\|x(t)-x_{LM}(t)\|+\left(\frac{1}{\beta}+c_{2}\right)\int_{0}^{t}e^{\frac{1}{\beta}(s-t)}\|x(s)-x_{LM}(s)\|\mathop{}\\!\mathrm{d}s.$ Finally, using (33) and (35), for all $s\geq 0$, $\|x(s)-x_{LM}(s)\|\leq 2R$, which concludes the proof. ## Appendix C Integrability of $\varphi$ and Additional Asymptotic Computations Below we prove Lemma 4.7. ###### Proof. We suppose that Assumptions 3 and 4 hold. As stated in Remark 4.3, since $\varphi$ is continuous, we only need to check its integrability when $t$ tends to $+\infty$. Let $t>0$, we first establish some useful identities, we omit the dependence on $t$ for the sake of readability. $\displaystyle p^{\prime}$ $\displaystyle=\frac{\alpha^{\prime}\varepsilon-(\alpha+\beta\lambda)\varepsilon^{\prime}}{\varepsilon^{2}},$ $\displaystyle p^{\prime\prime}$ $\displaystyle=\frac{\alpha^{\prime\prime}\varepsilon^{2}-2\alpha^{\prime}\varepsilon^{\prime}\varepsilon-(\alpha+\beta\lambda)\varepsilon^{\prime\prime}\varepsilon+2(\alpha+\beta\lambda)(\varepsilon^{\prime})^{2}}{\varepsilon^{3}}.$ Then, $\displaystyle\begin{split}r&=\frac{p^{2}}{4}\left(1+\frac{2p^{\prime}}{p^{2}}-\frac{4\lambda}{\varepsilon p^{2}}\right)=\frac{(\alpha+\beta\lambda)^{2}}{4\varepsilon^{2}}\left(1+\frac{2p^{\prime}\varepsilon^{2}}{(\alpha+\beta\lambda)^{2}}-\frac{4\lambda\varepsilon}{(\alpha+\beta\lambda)^{2}}\right)\\\ &=\frac{(\alpha+\beta\lambda)^{2}}{4\varepsilon^{2}}\left(1+\frac{2\alpha^{\prime}\varepsilon}{(\alpha+\beta\lambda)^{2}}-\frac{2\varepsilon^{\prime}}{(\alpha+\beta\lambda)}-\frac{4\lambda\varepsilon}{(\alpha+\beta\lambda)^{2}}\right).\end{split}$ (36) An important consequence of Assumption 4 is that $|\varepsilon^{\prime}(t)|=o(\varepsilon(t))$, $|\varepsilon^{\prime\prime}(t)|=o(\varepsilon^{\prime}(t))$ (and the same holds for $\alpha$ w.r.t. to its derivatives). Therefore, we deduce from (36) that $r(t)\sim_{+\infty}\frac{(\alpha(t)+\beta\lambda)^{2}}{4\varepsilon(t)^{2}},$ and we note that $1/r$ decays at the same speed as $\varepsilon^{2}$, which will be useful later. In order to study $\varphi$, we now differentiate $r$, $\displaystyle\begin{split}r^{\prime}&=\frac{p^{\prime}p}{2}\left(1+\frac{2p^{\prime}}{p^{2}}-\frac{4\lambda}{\varepsilon p^{2}}\right)+\frac{1}{4}\left(2p^{\prime\prime}-\frac{4(p^{\prime})^{2}}{p}+\frac{8\lambda p^{\prime}}{\varepsilon p}+\frac{4\lambda\varepsilon^{\prime}}{\varepsilon^{2}}\right)\\\ &=\frac{2p^{\prime}}{p}r+\frac{1}{4}\left(2p^{\prime\prime}-\frac{4(p^{\prime})^{2}}{p}+\frac{8\lambda p^{\prime}}{\varepsilon p}+\frac{4\lambda\varepsilon^{\prime}}{\varepsilon^{2}}\right),\end{split}$ and $\displaystyle\begin{split}r^{\prime\prime}=&2\frac{p^{\prime\prime}p-(p^{\prime})^{2}}{p^{2}}r+\frac{2p^{\prime}}{p}r^{\prime}\\\ &+\frac{1}{4}\left(2p^{\prime\prime\prime}+4\frac{(p^{\prime})^{3}-2p^{\prime\prime}p^{\prime}p}{p^{2}}+8\lambda\frac{p^{\prime\prime}p\varepsilon-(p^{\prime})^{2}\varepsilon-p^{\prime}p\varepsilon^{\prime}}{\varepsilon^{2}p^{2}}+\frac{4\lambda\varepsilon^{\prime\prime}}{\varepsilon^{2}}-\frac{8\lambda(\varepsilon^{\prime})^{2}}{\varepsilon^{3}}\right).\end{split}$ (37) Then, to justify that $\varphi$ is integrable, we prove that $\frac{r^{\prime\prime}}{r^{3/2}}$ and $\frac{(r^{\prime})^{2}}{r^{5/2}}$ are integrable. Since we know that $1/r$ decays at the same speed as $\varepsilon^{2}$, we can equivalently show that $\varepsilon^{3}r^{\prime\prime}$ and $\varepsilon^{5}(r^{\prime})^{2}$ are integrable. To this aim we fully expand all the terms in (37), and compute $(r^{\prime})^{2}$, which is extremely long and involved. On the one hand, it holds that, $\displaystyle\small r^{\prime 2}\varepsilon^{5}=$ $\displaystyle\left[-\frac{(\alpha+\beta\lambda)^{2}\left(-\frac{4\lambda\varepsilon}{(\alpha+\beta\lambda)^{2}}+1+\frac{\left(-{2(\alpha+\beta\lambda)\varepsilon^{\prime}}+{2\alpha^{\prime}\varepsilon}\right)}{(\alpha+\beta\lambda)^{2}}\right)\varepsilon^{\prime}}{2\sqrt{\varepsilon}}+\frac{(\alpha+\beta\lambda)^{2}\sqrt{\varepsilon}}{4}\left(-\frac{4\lambda\varepsilon^{\prime}}{(\alpha+\beta\lambda)^{2}}+\frac{8\lambda\alpha^{\prime}\varepsilon}{(\alpha+\beta\lambda)^{3}}\right.\right.$ $\displaystyle\left.\left.+\frac{2\left(-\frac{2(\alpha+\beta\lambda)\varepsilon^{\prime}}{\varepsilon}+{2\alpha^{\prime}\varepsilon}\right)\varepsilon^{\prime}}{(\alpha+\beta\lambda)^{2}}+\frac{\left(\frac{4(\alpha+\beta\lambda)\varepsilon^{\prime 2}}{\varepsilon}-{2(\alpha+\beta\lambda)\varepsilon^{\prime\prime}}-{4\alpha^{\prime}\varepsilon^{\prime}}+{2\alpha^{\prime\prime}\varepsilon}\right)}{(\alpha+\beta\lambda)^{2}}-\frac{2\left(-{2(\alpha+\beta\lambda)\varepsilon^{\prime}}+{2\alpha^{\prime}\varepsilon}\right)\alpha^{\prime}}{(\alpha+\beta\lambda)^{3}}\right)\right.$ $\displaystyle\left.+\frac{(\alpha+\beta\lambda)}{2}{\left(-\frac{4\lambda\varepsilon}{(\alpha+\beta\lambda)^{2}}+1+\frac{\left(-{2(\alpha+\beta\lambda)\varepsilon^{\prime}}+{2\alpha^{\prime}\varepsilon}\right)}{(\alpha+\beta\lambda)^{2}}\right)}\alpha^{\prime}\sqrt{\varepsilon}\phantom{\frac{\frac{\beta}{\beta}}{\beta}}\right]^{2}.$ On the other hand, we have, $\displaystyle r^{\prime\prime}\varepsilon(t)^{3}=$ $\displaystyle-\lambda\varepsilon^{\prime\prime}\varepsilon+\frac{4\lambda\alpha^{\prime}\varepsilon^{\prime}\varepsilon}{\alpha+\beta\lambda}+\frac{2\lambda\alpha^{\prime\prime}\varepsilon^{2}}{\alpha+\beta\lambda}-\frac{6\lambda\alpha^{\prime 2}\varepsilon^{2}}{(\alpha+\beta\lambda)^{2}}+\frac{(\alpha+\beta\lambda)^{2}}{2}\left(-\frac{3\varepsilon^{\prime 2}}{\varepsilon}+\varepsilon^{\prime\prime}\right)\left(\frac{4\lambda\varepsilon}{(\alpha+\beta\lambda)^{2}}-1+\frac{2\left({(\alpha+\beta\lambda)\varepsilon^{\prime}}-\alpha^{\prime}\varepsilon\right)}{(\alpha+\beta\lambda)^{2}}\right)$ $\displaystyle+{2(\alpha+\beta\lambda)\left(\frac{4\lambda\varepsilon}{(\alpha+\beta\lambda)^{2}}-1+\frac{2\left((\alpha+\beta\lambda)\varepsilon^{\prime}-\alpha^{\prime}\varepsilon\right)}{(\alpha+\beta\lambda)^{2}}\right)\alpha^{\prime}\varepsilon^{\prime}}-\frac{\left((\alpha+\beta\lambda)\alpha^{\prime\prime}+\alpha^{\prime 2}\right)}{2}\left(\frac{4\lambda\varepsilon}{(\alpha+\beta\lambda)^{2}}-1+\frac{2\left({(\alpha+\beta\lambda)\varepsilon^{\prime}}-\alpha^{\prime}\varepsilon\right)}{(\alpha+\beta\lambda)^{2}}\right)\varepsilon$ $\displaystyle-{\left(\frac{(\alpha+\beta\lambda)\varepsilon^{\prime}}{\varepsilon}-\alpha^{\prime}\right)\varepsilon^{\prime 2}}-\left({(\alpha+\beta\lambda)\varepsilon^{\prime}}-\alpha^{\prime}\varepsilon\right)\varepsilon^{\prime\prime}-2\left(-\frac{2(\alpha+\beta\lambda)\varepsilon^{\prime 2}}{\varepsilon}+{(\alpha+\beta\lambda)\varepsilon^{\prime\prime}}+{2\alpha^{\prime}\varepsilon^{\prime}}-\alpha^{\prime\prime}\varepsilon\right)\varepsilon^{\prime}$ $\displaystyle+{2\left(2\lambda\varepsilon^{\prime}-\frac{4\lambda\alpha^{\prime}\varepsilon}{\alpha+\beta\lambda}+2\left(\frac{(\alpha+\beta\lambda)\varepsilon^{\prime}}{\varepsilon}-\alpha^{\prime}\right)\varepsilon^{\prime}+\left(-\frac{2(\alpha+\beta\lambda)\varepsilon^{\prime 2}}{\varepsilon}+{(\alpha+\beta\lambda)\varepsilon^{\prime\prime}}+{2\alpha^{\prime}\varepsilon^{\prime}}-\alpha^{\prime\prime}\varepsilon\right)-\frac{2\left({(\alpha+\beta\lambda)\varepsilon^{\prime}}-\alpha^{\prime}\varepsilon\right)\alpha^{\prime}}{\alpha+\beta\lambda}\right)\varepsilon^{\prime}}$ $\displaystyle-\frac{1}{2}\left(\frac{6(\alpha+\beta\lambda)\varepsilon^{\prime 3}}{\varepsilon}-{6(\alpha+\beta\lambda)\varepsilon^{\prime}\varepsilon^{\prime\prime}}+{(\alpha+\beta\lambda)\varepsilon^{\prime\prime\prime}\varepsilon}-{6\alpha^{\prime}\varepsilon^{\prime 2}}+{3\alpha^{\prime}\varepsilon^{\prime\prime}\varepsilon}+{3\alpha^{\prime\prime}\varepsilon^{\prime}\varepsilon}-\alpha^{\prime\prime\prime}\varepsilon^{2}\right)$ $\displaystyle+\frac{1}{\alpha+\beta\lambda}\left({4\left({(\alpha+\beta\lambda)\varepsilon^{\prime}}-\alpha^{\prime}\varepsilon\right)\alpha^{\prime}\varepsilon^{\prime}}+{\left({(\alpha+\beta\lambda)\varepsilon^{\prime}}-\alpha^{\prime}\varepsilon\right)\alpha^{\prime\prime}\varepsilon}+{2\left(-{2(\alpha+\beta\lambda)\varepsilon^{\prime 2}}+{(\alpha+\beta\lambda)\varepsilon^{\prime\prime}\varepsilon}+{2\alpha^{\prime}\varepsilon^{\prime}\varepsilon}-\alpha^{\prime\prime}\varepsilon^{2}\right)\alpha^{\prime}}\right)$ $\displaystyle-\frac{2}{\alpha+\beta\lambda}\left(2\lambda\varepsilon^{\prime}-\frac{4\lambda\alpha^{\prime}\varepsilon}{\alpha+\beta\lambda}+2\left(\frac{(\alpha+\beta\lambda)\varepsilon^{\prime}}{\varepsilon}-\alpha^{\prime}\right)\varepsilon^{\prime}+\left(-\frac{2(\alpha+\beta\lambda)\varepsilon^{\prime 2}}{\varepsilon}+{(\alpha+\beta\lambda)\varepsilon^{\prime\prime}}+{2\alpha^{\prime}\varepsilon^{\prime}}-\alpha^{\prime\prime}\varepsilon\right)-\frac{2\left({(\alpha+\beta\lambda)\varepsilon^{\prime}}-\alpha^{\prime}\varepsilon\right)\alpha^{\prime}}{\alpha+\beta\lambda}\right)\alpha^{\prime}\varepsilon$ $\displaystyle-\frac{3\left({(\alpha+\beta\lambda)\varepsilon^{\prime}\varepsilon}-\alpha^{\prime}\varepsilon^{2}\right)\alpha^{\prime 2}}{(\alpha+\beta\lambda)^{2}}.$ We then analyze the integrability of each of the terms above. By Assumption 4, $\varepsilon^{\prime}$, $\varepsilon^{\prime\prime}$ and $\varepsilon^{\prime\prime\prime}$ are integrable and the same goes for $\alpha^{\prime}$, $\alpha^{\prime\prime}$ and $\alpha^{\prime\prime\prime}$, which is enough to justify the integrability of almost all the terms above. We finally see that we also need $\frac{(\varepsilon^{\prime})^{2}}{\varepsilon}$ and $\frac{(\varepsilon^{\prime})^{3}}{\varepsilon}$ to be integrable, which holds by Assumption 4. Overall, $\varphi$ is integrable on $\mathbb{R}_{+}$. ∎ We now state and prove the following result which was used at the end of the proof of Theorem 4.8. ###### Lemma C.1. Under Assumptions 3 and 4, for all $s\geq 0$, $\frac{1}{16}\left(\frac{2p^{\prime}(s)}{p(s)^{3/2}}-\frac{4\lambda\sqrt{\varepsilon(s)}}{(\alpha(s)+\beta\lambda)^{3/2}}\right)^{2}=\frac{\lambda^{2}\varepsilon(s)}{(\alpha(s)+\beta\lambda)^{3}}+o(\varepsilon(s)).$ ###### Proof. We omit the time dependence on $s\geq 0$ for the sake of readability. Using Assumption 3 we can define and expand the following quantity, $\displaystyle\frac{1}{16}\left(\frac{2p^{\prime}}{p^{3/2}}-\frac{4\lambda\sqrt{\varepsilon}}{(\alpha+\beta\lambda)^{3/2}}\right)^{2}=\frac{\lambda^{2}\varepsilon}{(\alpha+\beta\lambda)^{3}}-\frac{p^{\prime}\lambda\sqrt{\varepsilon}}{p^{3/2}(\alpha+\beta\lambda)^{3/2}}+\frac{(p^{\prime})^{2}}{4p^{3}}$ $\displaystyle=\frac{\lambda^{2}\varepsilon}{(\alpha+\beta\lambda)^{3}}-\lambda(\alpha^{\prime}\varepsilon-(\alpha+\beta\lambda)\varepsilon^{\prime})+\frac{1}{4(\alpha+\beta\lambda)^{3}}\left((\alpha^{\prime})^{2}\varepsilon+\frac{(\varepsilon^{\prime})^{2}}{\varepsilon}(\alpha+\beta\lambda)^{2}-2\alpha^{\prime}\varepsilon^{\prime}(\alpha+\beta\lambda)\right).$ Assumption 4, implies in particular that $|\varepsilon^{\prime}(t)|=o(\varepsilon(t))$ and that $\alpha^{\prime}(t)\to 0$, which we use in the equality above to obtain the desired conclusion. ∎ ## Appendix D Additional Experiments and Details We first detail the discretization that we used for approximating the solutions of the three ODEs considered in Section 5. We use Euler discretization schemes with fixed step-size $\gamma>0$ and approximate the solutions at times $t_{k}=\gamma k$, for all $k\in\mathbb{N}$. For a trajectory $x$, we use the notation $x(t_{k})\stackrel{{\scriptstyle\textrm{def}}}{{=}}x^{(k)}$. The approximation of (CN) is obtained by explicit discretization, so that for all $k\in\mathbb{N}$, we have, $x_{N}^{(k+1)}=x_{N}^{(k)}-\gamma\left[\beta\nabla^{2}f(x_{N}^{(k)})\right]^{-1}\nabla f(x_{N}^{(k)}).$ (38) Then, defining $\varepsilon_{k}=\varepsilon(t_{k})$ and $\alpha_{k}=\alpha(t_{k})$, (LM) and (VM-DIN-AVD) are obtained via Euler semi- implicit discretization. The solution of (LM) is approximated by, $x_{LM}^{(k+1)}=x_{LM}^{(k)}-\gamma\left[\alpha_{k}I_{n}+\beta\nabla^{2}f(x_{LM}^{(k)})\right]^{-1}\nabla f(x_{LM}^{(k)}),$ (39) where $I_{n}$ is the identity matrix on $\mathbb{R}^{n}$. The solution of (VM- DIN-AVD) is obtained similarly, $x^{(k+1)}=x^{(k)}+\left[(\varepsilon_{k}+\gamma\alpha_{k})I_{n}+\gamma\beta\nabla^{2}f(x^{(k)})\right]^{-1}\left(\varepsilon_{k}(x^{(k)}-x^{(k-1)})-\gamma^{2}\nabla f(x^{(k)})\right).$ (40) As safety check, one can see that for $\varepsilon_{k}=0$, (40) is equivalent to (39), which is itself equivalent to (38) when $\alpha_{k}=0$. Figure 6: Similar experiment and figures as those described in Figure 2, but for a poorly conditioned quadratic $f(x)=\frac{1}{2}\|Ax\|^{2}$ (first-two rows) and the function $f(x)=\log\left(\sum_{i=1}^{n}e^{x_{i}}\right)+\frac{1}{2}\|Ax\|^{2}$ (last- two rows). ## References * Alecsa et al. [2021] Cristian Daniel Alecsa, Szilárd Csaba László, and Titus Pinţa. An extension of the second order dynamical system that models Nesterov’s convex gradient method. _Applied Mathematics & Optimization_, 84(2):1687–1716, 2021. * Alvarez et al. [2002] Felipe Alvarez, Hedy Attouch, Jérôme Bolte, and Patrick Redont. A second-order gradient-like dissipative dynamical system with Hessian-driven damping: Application to optimization and mechanics. _Journal de Mathématiques Pures et Appliquées_ , 81(8):747–779, 2002. * Attouch and Cabot [2017] Hedy Attouch and Alexandre Cabot. Asymptotic stabilization of inertial gradient dynamics with time-dependent viscosity. _Journal of Differential Equations_ , 263(9):5412–5458, 2017. * Attouch and Fadili [2022] Hedy Attouch and Jalal Fadili. From the ravine method to the Nesterov method and vice versa: A dynamical system perspective. _SIAM Journal on Optimization_ , 32(3):2074–2101, 2022. * Attouch and László [2020] Hedy Attouch and Szilárd Csaba László. Newton-like inertial dynamics and proximal algorithms governed by maximally monotone operators. _SIAM Journal on Optimization_ , 30(4):3252–3283, 2020. * Attouch and László [2021] Hedy Attouch and Szilárd Csaba László. Continuous Newton-like inertial dynamics for monotone inclusions. _Set-Valued and Variational Analysis_ , 29(3):555–581, 2021. * Attouch and Svaiter [2011] Hedy Attouch and Benar Fux Svaiter. A continuous dynamical Newton-like approach to solving monotone inclusions. _SIAM Journal on Control and Optimization_ , 49(2):574–598, 2011. * Attouch et al. [2013] Hedy Attouch, Patrick Redont, and Benar Fux Svaiter. Global convergence of a closed-loop regularized Newton method for solving monotone inclusions in hilbert spaces. _Journal of Optimization Theory and Applications_ , 157(3):624–650, 2013. * Attouch et al. [2016] Hedy Attouch, Juan Peypouquet, and Patrick Redont. Fast convex optimization via inertial dynamics with Hessian driven damping. _Journal of Differential Equations_ , 261(10):5734–5783, 2016. * Attouch et al. [2020] Hedy Attouch, Zaki Chbani, Jalal Fadili, and Hassan Riahi. First-order optimization algorithms via inertial systems with Hessian driven damping. _Math. Program._ , 194(4):1–43, 2020. * Attouch et al. [2022a] Hedy Attouch, Aïcha Balhag, Zaki Chbani, and Hassan Riahi. Fast convex optimization via inertial dynamics combining viscous and Hessian-driven damping with time rescaling. _Evolution Equations and Control Theory_ , 11(2):487–514, 2022a. * Attouch et al. [2022b] Hedy Attouch, Radu Ioan Boţ, and Ernö Robert Csetnek. Fast optimization via inertial dynamics with closed-loop damping. _Journal of the European Mathematical Society (published online)_ , 2022b. * Benaïm et al. [2005] Michel Benaïm, Josef Hofbauer, and Sylvain Sorin. Stochastic approximations and differential inclusions. _SIAM Journal on Control and Optimization_ , 44(1):328–348, 2005. * Bender and Orszag [1999] Carl M Bender and Steven A Orszag. _Asymptotic methods and perturbation theory_. Springer, 1999. * Blumenthal [1912] Otto Blumenthal. Über asymptotische Integration linearer Differentialgleichungen mit Anwendung auf eine asymptotische Theorie der Kugelfunktionen. _Archiv der Mathematik und Physik_ , 19:136–174, 1912. * Boţ et al. [2021] Radu Ioan Boţ, Ernö Robert Csetnek, and Szilárd Csaba László. Tikhonov regularization of a second order dynamical system with Hessian driven damping. _Math. Program._ , 189(1):151–186, 2021. * Brillouin [1926] Léon Brillouin. Remarques sur la mécanique ondulatoire. _Journal de Physique et Le Radium_ , 7(12):353–368, 1926. * Broyden [1970] Charles George Broyden. The convergence of a class of double-rank minimization algorithms. _IMA Journal of Applied Mathematics_ , 6(1):76–90, 1970. * Castera [2021] Camille Castera. Inertial Newton algorithms avoiding strict saddle points. _arXiv:2111.04596_ , 2021. * Castera et al. [2021] Camille Castera, Jérôme Bolte, Cédric Févotte, and Edouard Pauwels. An inertial Newton algorithm for deep learning. _Journal of Machine Learning Research_ , 22(134):1–31, 2021. * Chen and Luo [2019] Long Chen and Hao Luo. First order optimization methods based on Hessian-driven Nesterov accelerated gradient flow. _arXiv:1912.09276_ , 2019. * Emmrich [1999] Etienne Emmrich. _Discrete versions of Gronwall’s lemma and their application to the numerical analysis of parabolic problems_. TU Fachbereich, 1999. * Fletcher [1970] Roger Fletcher. A new approach to variable metric algorithms. _The computer journal_ , 13(3):317–322, 1970\. * Gavurin [1958] Mark Konstantinovich Gavurin. Nonlinear functional equations and continuous analogues of iteration methods. _Izvestiya Vysshikh Uchebnykh Zavedenii. Matematika_ , 1(5):18–31, 1958. * Goldfarb [1970] Donald Goldfarb. A family of variable-metric methods derived by variational means. _Mathematics of computation_ , 24(109):23–26, 1970. * Green [1838] George Green. On the motion of waves in a variable canal of small depth and width. _Transactions of the Cambridge Philosophical Society_ , 6:457–462, 1838. * Hunter [2007] John D Hunter. Matplotlib: A 2D graphics environment. _Computing in science & engineering_, 9(3):90–95, 2007. * Kramers [1926] Hendrik Anthony Kramers. Wellenmechanik und halbzahlige Quantisierung. _Zeitschrift für Physik_ , 39(10):828–840, 1926. * Lin and Jordan [2022] Tianyi Lin and Michael I Jordan. A control-theoretic perspective on optimal high-order optimization. _Math. Program._ , 195(1):929–975, 2022. * Liouville [1836] Joseph Liouville. Mémoire sur le développement des fonctions ou parties de fonctions en séries dont les divers termes sont assujétis à satisfaire à une même équation différentielle du second ordre, contenant un paramètre variable. _Journal de Mathématiques Pures et Appliquées_ , 2:253–265, 1836. * Liu and Nocedal [1989] Dong C Liu and Jorge Nocedal. On the limited memory BFGS method for large scale optimization. _Math. Program._ , 45(1):503–528, 1989. * Ljung [1977] Lennart Ljung. Analysis of recursive stochastic algorithms. _IEEE transactions on automatic control_ , 22(4):551–575, 1977. * Nesterov [1983] Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of convergence ${O}(1/k^{2})$. In _Doklady an USSR_ , volume 269, pages 543–547, 1983. * Nesterov [2003] Yurii Nesterov. _Introductory lectures on convex optimization: A basic course_ , volume 87. Springer Science & Business Media, 2003. * Olver [1997] Frank Olver. _Asymptotics and special functions_. Academic Press, 1997. * Olver [1961] Frank William John Olver. Error bounds for the Liouville–Green (or WKB) approximation. In _Mathematical Proceedings of the Cambridge Philosophical Society_ , volume 57, pages 790–810. Cambridge University Press, 1961. * Polyak [1964] Boris T Polyak. Some methods of speeding up the convergence of iteration methods. _USSR Computational Mathematics and Mathematical Physics_ , 4(5):1–17, 1964. * Rossum [1995] Guido Rossum. _Python reference manual_. CWI (Centre for Mathematics and Computer Science), 1995. * Shanno [1970] David F Shanno. Conditioning of quasi-Newton methods for function minimization. _Mathematics of computation_ , 24(111):647–656, 1970. * Shi et al. [2022] Bin Shi, Simon S Du, Michael I Jordan, and Weijie J Su. Understanding the acceleration phenomenon via high-resolution differential equations. _Math. Program._ , 195:79–148, 2022. * Su et al. [2014] Weijie Su, Stephen Boyd, and Emmanuel Candes. A differential equation for modeling Nesterov’s accelerated gradient method: Theory and insights. In _Advances in Neural Information Processing Systems (NeurIPS)_ , volume 27, pages 2510–2518, 2014. * Taylor [1982] James G Taylor. Improved error bounds for the Liouville-Green (or WKB) approximation. _Journal of Mathematical Analysis and Applications_ , 85(1):79–89, 1982. * Walt et al. [2011] Stéfan van der Walt, Chris Colbert, and Gael Varoquaux. The NumPy array: a structure for efficient numerical computation. _Computing in Science & Engineering_, 13(2):22–30, 2011. * Wentzel [1926] Gregor Wentzel. Eine Verallgemeinerung der Quantenbedingungen für die Zwecke der Wellenmechanik. _Zeitschrift für Physik_ , 38(6):518–529, 1926.
# Zeckendorf expansion, Dirichlet series and infinite series involving the infinite Fibonacci word Shuo LI<EMAIL_ADDRESS> ###### Abstract Let $\beta=\frac{1+\sqrt{5}}{2}$, $(a_{n})_{n\in\mathbb{N}^{+}}$ be a non- uniform morphic sequence involving the infinite Fibonacci word and $(\delta(n))_{n\in\mathbb{N}^{+}}$ be a positive sequence such that for all positive integers $n$, $\delta(n)=\frac{1}{\sqrt{5}}\sum_{j\geq 0}\epsilon_{j}\beta^{j+2}$ if the unique Zeckendorf expansion of $n$ is $n=\sum_{j\geq 0}\epsilon_{j}F_{j+2}$ with Fibonacci numbers $F_{0},F_{1},F_{2}...$. We define and study some Dirichlet series in the form of $\sum_{n\geq 1}\frac{a_{n}}{(\delta(n))^{s}}$ and relations between them. Moreover, we compute the values of some infinite series involving the infinite Fibonacci word. ## 1 Introduction and main results Several infinite products and series involving sum-of-digits functions as well as block-counting functions were extensively studied in [AC85][AS89][ASS07][ARS18][Hu16]. The major ideal in these articles is that sum-of-digits functions and block-counting functions are all closely related to the notion of the automatic sequence, which can be defined as the images of the fixed points of uniform morphisms under some codings (more background is given in [AS03]). However, we may expect to compute infinite products or series alone with functions in a more general class, for example, the morphic words. Herein, we offer a formal definition of the morphic words: let $\sum$ and $\Delta$ denote two non-empty sets of symbols and let $\sum^{*}$ and $\Delta^{*}$ respectively denote the free monoids on $\sum$ and $\Delta$. A sequence $(a_{n})_{n\in\mathbb{N}^{+}}\in\Delta^{*}$ is called morphic if there exists a sequence $(e_{n})_{n\in\mathbb{N}^{+}}\in\sum^{*}$, a morphism $f:\sum^{*}\to\sum^{*}$ and a morphism called coding $\rho:\sum\to\Delta$ such that $(a_{n})_{n\in\mathbb{N}^{+}}=\rho((e_{n})_{n\in\mathbb{N}^{+}})$ under the condition that the sequence $(e_{n})_{n\in\mathbb{N}^{+}}$ is a fixed point of the morphism $f$. Particularly, if the morphism $f$ is uniform, that is to say the length of $f(a)$ is constant for all elements $a\in\sum$, the sequence $(a_{n})_{n\in\mathbb{N}^{+}}$ is called automatic. A typical example of automatic sequences is the Thue-Morse sequence, which is the fixed point of the morphism $0\to 0,1$ and $1\to 1,0$. In this article, we focus on non- uniform morphic sequences involving the infinite Fibonacci word, which is the fixed point of the non-uniform morphism $0\to 0$ and $1\to 0,1$. The main ideal of this article arises from the fact that the infinite Fibonacci word can be deduced by counting the number of $0$s at the end of the Zeckendrof expansion of each integer. Let the Fibonacci numbers be defined by $F_{0}=0,F_{1}=1$, and $F_{n}=F_{n-1}+F_{n-2}$ for $n\geq 2$. From Zeckendorf’s theorem [Zec72], every positive integer $n$ can be uniquely written as $n=\sum_{j\geq 0}\epsilon_{j}F_{j+2}$ with $\epsilon_{j}\in\left\\{0,1\right\\}$ under the condition that $\epsilon_{j}\epsilon_{j+1}=0$ for all $j\geq 0$. This expansion is called the Zeckendorf expansion. Using this notion, we can define some other sequences which present useful properties in analysis as well as in combinatorics. On the one hand, if we let $\beta$ denote $\frac{1+\sqrt{5}}{2}$, then $F_{n}=\frac{1}{\sqrt{5}}(\beta^{n}-(-\beta)^{-n})$ for all $n\geq 0$. Thus, it is natural to define and study the following two sequences $(\delta(n))_{n\in\mathbb{N}}$ and $(\delta^{\prime}(n))_{n\in\mathbb{N}}$: for all integers $n\geq 0$, if $n=\sum_{j\geq 0}\epsilon_{j}F_{j+2}$ is the Zeckendorf expansion of $n$, then $\delta(n)=\frac{1}{\sqrt{5}}\sum_{j\geq 0}\epsilon_{j}\beta^{j+2}\;\text{and}\;\delta^{\prime}(n)=\frac{1}{\sqrt{5}}\sum_{j\geq 0}\epsilon_{j}(-\beta)^{-j-2}.$ In Section two, we will study the arithmetic properties of the sequences $(\delta(n))_{n\in\mathbb{N^{+}}}$ and $(\delta^{\prime}(n))_{n\in\mathbb{N^{+}}}$, as well as some properties of the Dirichlet series $F(s)=\sum_{n\geq 1}\frac{1}{(\delta(n))^{s}}$. On the other hand, some morphic sequences can also be defined by using the Zeckendorf expansion, namely, the Fibonacci-automatic sequences defined and studied in [MSS16][DMSS16][DMR+17]. Among these sequences, a typical example is the infinite Fibonacci sequence $(f(n))_{n\in\mathbb{N^{+}}}$. In Section three, we study some combinatorial properties of the infinite Fibonacci sequence and other morphic sequences. In Section four, we present a combinatorial proof of the meromorphic continuation of $F$ on the whole complex plane; the main theorem is announced as follows: ###### Theorem 1 The Dirichlet series $F$ converges absolutely on the set $\left\\{s|\Re(s)>1\right\\}$, and has a meromorphic continuation on the whole complex plane. Moreover, its poles are located on the set of zeros of the function $s\to 1-2\beta^{-s}+\beta^{-3s}$. Moreover, we extend this result to other Dirichlet series with morphic coefficients. Similar results can be found in [Sou19]. Finally, in Section five, we use the method introduced in [AC85] to compute some infinite series: ###### Theorem 2 Letting $(f(n))_{n\in\mathbb{N}^{+}}$, $(d(n))_{n\in\mathbb{N}^{+}}$, $(r(n))_{n\in\mathbb{N}^{+}}$, $(s(n))_{n\in\mathbb{N}^{+}}$ and $(t(n))_{n\in\mathbb{N}^{+}}$ be integer sequences defined in the following sections and letting $\beta=\frac{\sqrt{5}+1}{2}$, then we have $\sum_{n\geq 2}r(n)(\frac{\sqrt{5}}{(n-1)\sqrt{5}-\\{\beta(n-1)\\}+1-f(n-1)}-\frac{\sqrt{5}}{n\sqrt{5}-\\{\beta n\\}+1-f(n)})=\frac{\beta-1}{\beta^{2}}\ln(\beta),$ $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}(\frac{\sqrt{5}}{(n-1)\sqrt{5}-\\{\beta(n-1)\\}+1-f(n-1)}-\frac{\sqrt{5}}{n\sqrt{5}-\\{\beta n\\}+1-f(n)})=\beta^{-1}\ln(\beta),$ $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}(\frac{\sqrt{5}}{(n-1)\sqrt{5}-\\{\beta(n-1)\\}+1-f(n-1)}-\frac{\sqrt{5}}{n\sqrt{5}-\\{\beta n\\}+1-f(n)})=\beta^{-2}\ln(\beta),$ $\sum_{n\geq 2}t(n)(\frac{\sqrt{5}}{(n-1)\sqrt{5}-\\{\beta(n-1)\\}+1-f(n-1)}-\frac{\sqrt{5}}{n\sqrt{5}-\\{\beta n\\}+1-f(n)})=\beta^{2}-\frac{\beta-1}{\beta^{2}}\ln(\beta),$ $\sum_{n\geq 2}\frac{s(n)\sqrt{5}}{n\sqrt{5}-\\{\beta n\\}+1-f(n)}=(\frac{3}{2}\beta^{-4}-\beta^{-2})\ln(\beta),$ where $\\{x\\}$ is the fractional part of $x$. ## 2 Arithmetic properties of $(\delta(n))_{n\in\mathbb{N^{+}}}$ ###### Proposition 1 Let $e(n)$ denote the last bit in the Zeckendorf expansion of $n$; that is to say, for a given integer $n$, $e(n)$ is the coefficient $\epsilon_{0}$ in the expansion $n=\sum_{j\geq 0}\epsilon_{j}F_{j+2}$, and we then have the following equations: $\delta(n+1)-\delta(n)=\begin{cases}\frac{1}{\sqrt{5}}\beta^{2}\;\;\;\text{if}\;e(n)=0,\\\ \frac{1}{\sqrt{5}}\beta\;\;\;\;\text{if}\;e(n)=1.\end{cases}$ Consequently, the sequence $(\delta(n))_{n\in\mathbb{N}}$ is an unbounded increasing sequence. * Proof For a given integer $n$, if $e(n)=1$, from the convention that there does not exist the factor $\overline{1,1}$ in the Zeckendorf expansion of $n$, we can suppose that this expansion is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,\underbrace{0,1}_{i\;\text{times}}}$, where $\epsilon_{i}$ are either $0$ or $1$. From the recurrent relations between the Fibonacci numbers and the uniqueness of the Zeckendrof expansion, the expansion of $n+1$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,1,\underbrace{0}_{2i-1\;\text{times}}}$. Therefore, $\delta(n+1)-\delta(n)=\frac{1}{\sqrt{5}}(\beta^{2i+1}-\sum_{k=1}^{i}\beta^{2k}).$ Using the equation $\beta^{2}-\beta=1$ recurrently, we have $\delta(n+1)-\delta(n)=\frac{1}{\sqrt{5}}\beta$. In the same way, if $e(n)=0$ and the Zeckendorf expansion of $n$ ends up with a suffix $\overline{1,0}$, then we can suppose that this expansion is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,\underbrace{0,1}_{i\;\text{times}},0}$. Thus, the expansion of $n+1$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,1,\underbrace{0}_{2i\;\text{times}}}$. Therefore, $\delta(n+1)-\delta(n)=\frac{1}{\sqrt{5}}(\beta^{2i+2}-\sum_{k=1}^{i}\beta^{2k+1})=\frac{1}{\sqrt{5}}\beta^{2}.$ In the last case, if $e(n)=0$ but the Zeckendorf expansion of $n$ ends up with a suffix $\overline{0,0}$, then we can suppose that this expansion is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,0}$. Thus, the expansion of $n+1$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,1}$, so that $\delta(n+1)-\delta(n)=\frac{1}{\sqrt{5}}\beta^{2}.$ ###### Proposition 2 Let us define the sequence $(d(n))_{n\in\mathbb{N^{+}}}$ in the following way: $d(n)=\begin{cases}0\;\text{if the Zeckendorf expansion of $n$ admits the string $\overline{1}$ as a suffix},\\\ 1\;\text{if the Zeckendorf expansion of $n$ admits a string of type $\overline{1,\underbrace{0}_{2i+1;\text{times}}}$ as a suffix},\\\ 2\;\text{if the Zeckendorf expansion of $n$ admits a string of type $\overline{1,\underbrace{0}_{2i+2;\text{times}}}$ as a suffix},\end{cases}$ where $i$ is a positive integer. We then have $|\delta^{\prime}(n)|<\frac{1}{\beta\sqrt{5}}$ for all $n>0$. Moreover, $\begin{cases}\delta^{\prime}(n)>0\;\text{if}\;d(n)=0\;\text{or}\;d(n)=2,\\\ \delta^{\prime}(n)<0\;\text{if}\;d(n)=1.\end{cases}$ * Proof It is clear that $|-\beta^{-1}|<1$. For a given integer $n$, letting $n=\sum_{j\geq 0}\epsilon_{j}F_{j+2}$ with $\epsilon_{j}\in\left\\{0,1\right\\}$ be the Zeckendorf expansion of $n$, then $\displaystyle|\delta^{\prime}(n)|$ $\displaystyle=|\frac{1}{\sqrt{5}}\sum_{i\geq 0}\epsilon_{i}(-\beta)^{-i-2}|$ (1) $\displaystyle\leq\frac{1}{\sqrt{5}}\sum_{i\geq 0}\epsilon_{i}(\beta^{-i-2})$ $\displaystyle\leq\frac{1}{\sqrt{5}}\sum_{i\geq 0}(\beta^{-2i-2})$ $\displaystyle\leq\frac{1}{\sqrt{5}}\beta^{-2}\frac{1}{1-\beta^{-2}}.$ Substituting the equation $\beta^{2}-\beta-1=0$ into Equation 1, we have $|\delta^{\prime}(n)|\leq\frac{1}{\sqrt{5}}\beta^{-1}\frac{\beta^{-1}}{1-\beta^{-2}}=\frac{1}{\sqrt{5}}\beta^{-1}$. For the second part of the proposition, let $n$ be an integer and let $i$ be the smallest index such that $\epsilon_{i}=1$ in the Zeckendorf expansion of $n$. It is easy to verify that $(-\beta)^{-i-2}>0$ if $d(n)=0$ or $2$, and $(-\beta)^{-i-2}<0$ if $d(n)=1$. It follows from Equation 1 that $|\frac{1}{\sqrt{5}}\sum_{k\geq i+1}\epsilon_{k}(-\beta)^{-k-2}|\leq\beta^{-i-3}$, so the sign of $\delta^{\prime}(n)$ is the same as that of $(-\beta)^{-i-2}$. ###### Proposition 3 The function $F(s)=\sum_{n\geq 1}\frac{1}{(\delta(n))^{s}}$ is a Dirichlet series which converges absolutely on $\left\\{s|\Re(s)>1\right\\}$ and has a meromorphic continuation on $\left\\{s|\Re(s)>0\right\\}$. Moreover, $F(s)$ has a single pole at $s=1$ with residue $1$. * Proof From Propositions 1 and 2, the sequence $(\delta(n))_{n\in\mathbb{N}}$ is increasing and $n-1<\delta(n)<n+1$ for all positive integers $n$. Thus, $F(s)$ converges absolutely on $\left\\{s|\Re(s)>1\right\\}$. For the extension, let $\zeta$ be the Riemann zeta function, $\displaystyle\zeta(s)-F(s)$ $\displaystyle=\sum_{n\geq 1}\frac{1}{n^{s}}-\frac{1}{\delta^{s}(n)}$ (2) $\displaystyle=\sum_{n\geq 1}\frac{1}{(\delta(n)-\delta^{\prime}(n))^{s}}-\frac{1}{\delta^{s}(n)}$ $\displaystyle=\sum_{n\geq 1}\frac{1}{\delta^{s}(n)}\frac{1}{(1-\frac{\delta^{\prime}(n)}{\delta(n)})^{s}}-\frac{1}{\delta^{s}(n)}$ $\displaystyle=\sum_{n\geq 1}\frac{1}{\delta^{s}(n)}(\sum_{m\geq 0}\binom{-s}{m}(-\frac{\delta^{\prime}(n)}{\delta(n)})^{m}-1)$ $\displaystyle=\sum_{m\geq 1}\binom{-s}{m}\sum_{n\geq 1}\frac{(-\delta^{\prime}(n))^{m}}{\delta^{s+m}(n)}$ For any given positive integer $m$, $|\sum_{n\geq 1}\frac{(-\delta^{\prime}(n))^{m}}{\delta^{s+m}(n)}|\leq\sum_{n\geq 1}|\frac{(\delta^{\prime}(n))^{m}}{\delta^{s+m}(n)}|\leq\sum_{n\geq 1}\frac{\beta^{-m}}{\delta^{\Re(s)+m}(n)}$. Since the term $\sum_{n\geq 1}\frac{1}{\delta^{\Re(s)+m}(n)}$ is bounded for large $m$, the righthand side of equation (2) converges for all $s$ such that $\Re(s)>0$. Consequently, the function $F(s)$ has a meromorphic continuation on $\left\\{s|\Re(s)>0\right\\}$ and has the same pole with the same residue as the Riemann zeta function on $s=1$. ## 3 Fibonacci sequence and its first differences sequence Let $(f(n))_{n\in\mathbb{N^{+}}}$ be the Fibonacci sequence defined as the fixed point of the morphism $1\to 0,1$ and $0\to 0$ and let us recall the sequence $(d(n))_{n\in\mathbb{N^{+}}}$ defined in Proposition 2: $d(n)=\begin{cases}0\;\text{if the Zeckendorf expansion of $n$ admits the string $\overline{1}$ as a suffix}\\\ 1\;\text{if the Zeckendorf expansion of $n$ admits a string of type $\overline{1,\underbrace{0}_{2i+1;\text{times}}}$ as a suffix}\\\ 2\;\text{if the Zeckendorf expansion of $n$ admits a string of type $\overline{1,\underbrace{0}_{2i+2;\text{times}}}$ as a suffix},\end{cases}$ with some positive integer $i$. In this section, we will show the relations between $(d(n))_{n\in\mathbb{N^{+}}}$, $(f(n))_{n\in\mathbb{N^{+}}}$ and the first differences sequence of $(f(n))_{n\in\mathbb{N^{+}}}$. ###### Proposition 4 The sequence $(d(n))_{n\in\mathbb{N^{+}}}$ satisfies the following properties: 1, if $d(n)=0$, then $d(n+1)=1$; 2, if $d(n)=1$, then $d(n+1)=2$ or $0$; 3, if $d(n)=2$, then $d(n+1)=0$. * Proof If $d(n)=0$, then the Zeckendorf expansion of $n$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,\underbrace{0,1}_{i\;\text{times}}}$ for some $i\geq 1$. From the recurrent relations between the Fibonacci numbers and the uniqueness of the Zeckendrof expansion , the expansion of $n+1$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,1,\underbrace{0}_{2i-1\;\text{times}}}$. Consequently, $d(n+1)=1$. If $d(n)=2$, then the Zeckendorf expansion of $n$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,0}$. Thus, the expansion of $n+1$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,1}$. Therefore, $d(n+1)=0$. If $d(n)=1$, from the definition, the Zeckendorf expansion of $n$ ends up with a suffix $\overline{1,\underbrace{0}_{2i+1\;\text{times}}}$ for some $i\geq 0$. There are two cases: if $i=0$, then the Zeckendorf expansion of $n$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,\underbrace{0,1}_{i\;\text{times}},0}$. Thus, the expansion of $n+1$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,1,\underbrace{0}_{2i\;\text{times}}}$. In this case, $d(n+1)=2$. If $i\geq 1$, then the Zeckendorf expansion of $n$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,0}$. Thus, the expansion of $n+1$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,1}$. Therefore, $d(n+1)=0$. ###### Proposition 5 The sequence $(f(n))_{n\in\mathbb{N^{+}}}$ is the image of $(d(n))_{n\in\mathbb{N^{+}}}$ under the morphism $0\to 0,\;1\to 1$ and $2\to 0$. Moreover, if we define $(h(n))_{n\in\mathbb{N^{+}}}$ as the first differences sequence of $(f(n))_{n\in\mathbb{N^{+}}}$, that is to say, $h(n)=f(n)-f(n+1)$, then $(h(n))_{n\in\mathbb{N^{+}}}$ is the image of $(d(n))_{n\in\mathbb{N^{+}}}$ under the morphism $0\to-1,\;1\to 1$ and $2\to 0$. * Proof It is mentioned in [MSS16] that $f(n)$ is the output of the following direct automaton $0$start$1$010 when the input is the Zeckendorf expansion of $n-1$. In other words, $d(n)=1$ if and only if the Zeckendorf expansion of $n-1$ ends up with a value of $1$ and $d(n)=0$ if and only if the Zeckendorf expansion of $n-1$ ends up with a value of $0$. On the other hand, from Proposition 4, if $d(n)=1$, then $d(n-1)=0$, so that the Zeckendorf expansion of $n-1$ ends up with a value of $1$, and thus, $f(n)=1$; similarly, if $d(n)=0$ or $2$, then $d(n-1)=1$ or $2$, so that the Zeckendorf expansion of $n-1$ ends up with a value of $0$, and thus, $f(n)=0$. The second part of this proposition is a direct consequence of the previous result. ###### Remark 1 From the previous proposition, the sequence $(f(n))_{n\in\mathbb{N^{+}}}$ is the output of the direct automaton $0$start$1$100 when the input is the Zeckendorf expansion of $n$. Moreover, from the descriptions of the sequences A00384, A001468, A014677 and A270788[SI20], the sequence $(d(n))_{n\in\mathbb{N^{+}}}$ is the image of the morphic sequence A270788 under the coding $1\to 0,\;2\to 1,\;3\to 2$. ###### Proposition 6 Let us define two functions $\tau_{0}$, $\tau_{1}:\mathbb{N}\to\mathbb{N}$ in the following way: for every positive integer $n$, letting $n=\sum_{j\geq 0}\epsilon_{j}F_{j+2}$ be the Zeckendorf expansion, then $\tau_{0}(n)=\sum_{j\geq 0}\epsilon_{j}F_{j+3};$ $\tau_{1}(n)=\sum_{j\geq 0}\epsilon_{j}F_{j+3}+F_{2}.$ The following relations exist between the sets: $\left\\{\tau_{0}(n)|n\in\mathbb{N}^{+}\right\\}=\left\\{n|n\in\mathbb{N}^{+},d(n)=1\;\text{or}\;2\right\\},$ $\left\\{\tau_{1}(n)|n\in\mathbb{N}^{+},d(n)=1\;\text{or}\;2\right\\}=\left\\{n|n\in\mathbb{N}^{+},n\geq 2,d(n)=0\right\\},$ $\left\\{\tau_{1}(n)|n\in\mathbb{N}^{+},d(n)=0\right\\}=\left\\{n|n\in\mathbb{N}^{+},d(n)=2\right\\}.$ Consequently, $\left\\{\tau_{0}(n)|n\in\mathbb{N}^{+}\right\\}\cup\left\\{\tau_{1}(n)|n\in\mathbb{N}^{+}\right\\}=\left\\{n|n\in\mathbb{N}^{+},n\geq 2\right\\},$ $\left\\{\tau_{0}(n)|n\in\mathbb{N}^{+}\right\\}\cap\left\\{\tau_{1}(n)|n\in\mathbb{N}^{+}\right\\}=\left\\{n|n\in\mathbb{N}^{+},d(n)=2\right\\}.$ * Proof If $d(n)=1$ or $2$, then the Zeckendorf expansion of $n$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0}$. Thus, the expansion of $\tau_{0}(n)$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,0}$ and the expansion of $\tau_{1}(n)$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,1}$. Therefore, $d(\tau_{0}(n))=1$ or $2$ and $d(\tau_{1}(n))=0$. If $d(n)=0$, then the Zeckendorf expansion of $n$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,\underbrace{1,0}_{i\;\text{times}},1}$ for some $i\geq 0$. Thus, the expansion of $\tau_{0}(n)$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},0,\underbrace{1,0}_{i+1\;\text{times}}}$ and the expansion of $\tau_{1}(n)$ is in the form $\overline{\epsilon_{k},\epsilon_{k-1},\epsilon_{k-2},...,\epsilon_{s},1,\underbrace{0}_{2i+2\;\text{times}}}$. Therefore, $d(\tau_{0}(n))=1$ or $2$ and $d(\tau_{1}(n))=0$. ###### Corollary 1 $\left\\{\beta\delta(n)|n\in\mathbb{N}^{+}\right\\}\cup\left\\{\beta\delta(n)+\frac{1}{\sqrt{5}}\beta^{2}|n\in\mathbb{N}^{+}\right\\}=\left\\{\delta(n)|n\in\mathbb{N}^{+},n\geq 1\right\\}$ $\left\\{\beta\delta(n)|n\in\mathbb{N}^{+}\right\\}\cap\left\\{\beta\delta(n)+\frac{1}{\sqrt{5}}\beta^{2}|n\in\mathbb{N}^{+}\right\\}=\left\\{\delta(n)|n\in\mathbb{N}^{+},d(n)=2\right\\}$ * Proof It is directly from this fact that if $\delta(n)=\frac{1}{\sqrt{5}}\sum_{j\geq 0}\epsilon_{j}\beta^{j+2}$, then $\beta\delta(n)=\frac{1}{\sqrt{5}}\sum_{j\geq 0}\epsilon_{j}\beta^{j+3}$ and $\beta\delta(n)+\frac{1}{\sqrt{5}}\beta^{2}=\frac{1}{\sqrt{5}}(\sum_{j\geq 0}\epsilon_{j}\beta^{j+3}+\beta^{2})$. ## 4 Dirichlet series involving the sequence $(d(n))_{n\in\mathbb{N}^{+}}$ Let us recall the Dirichlet series $F(s)=\sum_{n\geq 1}\frac{1}{(\delta(n))^{s}}$, and define four other Dirichlet series: $G(s)=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{(\delta(n))^{s}};$ $H(s)=\sum_{n\geq 1}\frac{1}{(\beta\delta(n)+\frac{1}{\sqrt{5}}\beta^{2})^{s}};$ $I(s)=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}\frac{1}{(\delta(n))^{s}};$ $J(s)=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}\frac{1}{(\delta(n))^{s}}.$ In this section, we prove the meromorphic continuation of all of these series on the whole complex plane. First, let us prove Theorem 1. * Proof of Theorem 1 From Corollary 1, if $\Re(s)>1$, then $\displaystyle F(s)$ $\displaystyle=\sum_{n\geq 2}\frac{1}{\delta^{s}(n)}+\frac{1}{\delta^{s}(1)}$ (3) $\displaystyle=\sum_{n\geq 1}\frac{1}{(\beta\delta(n))^{s}}+\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{(\delta(n))^{s}}+\frac{1}{\delta^{s}(1)}$ $\displaystyle=\beta^{-s}F(s)+G(s)+\frac{1}{\delta^{s}(1)}$ $\displaystyle G(s)$ $\displaystyle=\sum_{n\geq 1}\frac{1}{(\beta\delta(n)+\frac{1}{\sqrt{5}}\beta^{2})^{s}}-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}\frac{1}{(\delta(n))^{s}}$ (4) $\displaystyle=H(s)-\sum_{m\geq 1}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{(\beta^{2m}\delta(n))^{s}}$ $\displaystyle=H(s)-\frac{\beta^{-2s}}{1-\beta^{-2s}}G(s)$ Substituting (4) into (3), we have $(1-\beta^{-s})F(s)=(1-\beta^{-2s})H(s)+\frac{1}{\delta^{s}(1)}$ (5) Moreover, with the fact that $0<\frac{1}{\sqrt{5}}\beta<1$, we have $\displaystyle H(s)$ $\displaystyle=\sum_{n\geq 1}\frac{1}{(\beta\delta(n)+\frac{1}{\sqrt{5}}\beta^{2})^{s}}$ (6) $\displaystyle=\sum_{n\geq 1}\frac{1}{(\beta\delta(n))^{s}}\frac{1}{(1+\frac{\frac{1}{\sqrt{5}}\beta}{\delta(n)})^{s}}$ $\displaystyle=\sum_{n\geq 1}\frac{1}{(\beta\delta(n))^{s}}\sum_{m\geq 0}\binom{-s}{m}(\frac{\frac{1}{\sqrt{5}}\beta}{\delta(n)})^{m}$ $\displaystyle=\beta^{-s}\sum_{m\geq 0}(\frac{1}{\sqrt{5}}\beta)^{m}\binom{-s}{m}F(s+m)$ From (5) and (6), we can deduce that $(1-2\beta^{-s}+\beta^{-3s})F(s)=(\beta^{-s}-\beta^{-3s})\sum_{m\geq 1}(\frac{1}{\sqrt{5}}\beta)^{m}\binom{-s}{m}F(s+m)+\frac{1}{\delta^{s}(1)}$ (7) For any given complex number $s$ such that $\Re(s)>1$, the sequence $(F(s+k))_{k\in\mathbb{N}}$ is bounded. Thus, the righthand side of equation (7) converges uniformly for $\Re(s)>0$. Hence, $F(s)$ has a meromorphic extension for $0<\Re(s)\leq 1$. Now, if $0<\Re(s)\leq 1$, the righthand side converges, with the exception of those $s$ for which $s$ is a zero of $1-2\beta^{-s}+\beta^{-3s}$. This yields a meromorphic extension of $F$ for $\Re(s)>-1$. Iterating this process shows that $F$ has a meromorphic extension to the whole complex plane. Moreover, the poles of $F$ are located on the set of zeros of the function $s\to 1-2\beta^{-s}+\beta^{-3s}$. ###### Corollary 2 The Dirichlet series $G,H,I,J$ all admit meromorphic continuations on the whole complex plane and have simple poles at $s=1$. Moreover, their residues at $s=1$ are respectively $1-\beta^{-1}$,$\beta^{-1}$,$1-\beta^{-1}$ and $\beta^{-1}-\beta^{-2}$. * Proof The meromorphic continuations of $G$ and $H$ are given respectively by (3) and (5), along with their residues. To see the meromorphic continuation of $I$ and $J$, on the set $\left\\{s|\Re(s)>1\right\\}$, we have $I(s)=\sum_{m\geq 0}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{(\beta^{2m+1}\delta(n))^{s}}=\frac{\beta^{-s}}{1-\beta^{-2s}}G(s);$ (8) and $J(s)=\sum_{m\geq 1}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{(\beta^{2m}\delta(n))^{s}}=\frac{\beta^{-2s}}{1-\beta^{-2s}}G(s).$ (9) For the residues, we can use the fact that the residue of $F$ at $s=1$ is $1$. ###### Proposition 7 Let $a$, $b$ be two real numbers such that $|b|\leq|a|$ and let $i=0,1,2$ or $3$. Letting $K_{a,b}^{(i)}(s)$ be the function $K_{a,b}^{(i)}(s)=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=i\end{subarray}}\frac{1}{(a\delta(n))^{s}}-\frac{1}{(a\delta(n)+b)^{s}}$ for $i=0,1,2$ and letting $K_{a,b}^{(3)}(s)=\sum_{i=0}^{2}K_{a,b}^{(i)}(s)$, then the function $K_{a,b}^{(i)}(s)$ has a meromorphic continuation on the whole complex plane for any $i$. Moreover, it converges absolutely on $\left\\{s|\Re(s)>1\right\\}$, converges pointwisely on $\left\\{s|\Re(s)>0\right\\}$ and $\lim_{s\to 0}K_{a,b}^{(0)}(s)=\frac{b}{a}(1-\beta^{-1})$, $\lim_{s\to 0}K_{a,b}^{(1)}(s)=\frac{b}{a}(1-\beta^{-1})$,$\lim_{s\to 0}K_{a,b}^{(2)}(s)=\frac{b}{a}(\beta^{-1}-\beta^{-2})$ and $\lim_{s\to 0}K_{a,b}^{(3)}(s)=\frac{b}{a}$. * Proof From the hypothesis, we have $|\frac{b}{a\delta(n)}|<1$ for all $n$. Thus, for $i=0,1$ or $2$, on the set $\left\\{s|\Re(s)>1\right\\}$ $\displaystyle K_{a,b}^{(i)}(s)$ $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=i\end{subarray}}(\frac{1}{(a\delta(n))^{s}}-\frac{1}{(a\delta(n)+b)^{s}})$ (10) $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=i\end{subarray}}\frac{1}{(a\delta(n))^{s}}(1-\frac{1}{(1+\frac{b}{a\delta(n)})^{s}})$ $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=i\end{subarray}}\frac{1}{(a\delta(n))^{s}}(1-\sum_{m\geq 0}\binom{-s}{m}(\frac{b}{a\delta(n)})^{m})$ $\displaystyle=-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=i\end{subarray}}\frac{1}{(a\delta(n))^{s}}\sum_{m\geq 1}\binom{-s}{m}(\frac{b}{a\delta(n)})^{m}$ $\displaystyle=-a^{-s}\sum_{m\geq 1}\binom{-s}{m}(\frac{b}{a})^{m}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=i\end{subarray}}\frac{1}{(\delta(n))^{s+m}}$ $\displaystyle=-a^{-s}\sum_{m\geq 2}\binom{-s}{m}(\frac{b}{a})^{m}X^{(i)}(s+m)+sa^{-s}\frac{b}{a}X^{(i)}(s+1),$ where $X^{(i)}$ represents respectively the functions $G$, $I$ and $J$ for $i=0,1$ and $2$. On the other hand, from Proposition 1, $\delta(n)>\delta(1)+(n-1)\frac{1}{\sqrt{5}}\beta=\frac{1}{\sqrt{5}}(1+n\beta)$. Thus, for all $s$ such that $\Re(s)>1$, $\displaystyle|X^{(i)}(s)|$ $\displaystyle\leq\sum_{n\geq 1}\frac{1}{|(\delta(n))^{s}|}$ (11) $\displaystyle\leq\sum_{n\geq 1}\frac{1}{(\frac{1}{\sqrt{5}}(1+n\beta))^{\Re(s)}}$ $\displaystyle\leq(\frac{\sqrt{5}}{\beta^{2}})^{\Re(s)}+\int_{x=1}^{\infty}\frac{1}{(\frac{1}{\sqrt{5}}(1+x\beta))^{\Re(s)}}$ $\displaystyle\leq(\frac{\sqrt{5}}{\beta^{2}})^{\Re(s)}+\frac{\sqrt{5}}{\beta(\Re(s)-1)}(\frac{\sqrt{5}}{\beta^{2}})^{\Re(s)-1}.$ Consequently, $\sum_{m\geq k}\binom{-s}{m}(\frac{b}{a})^{m}X^{(i)}(s+m)$ converges for all $s$ such that $\Re(s)>-k$. Combining the fact that $X^{(i)}$ has a meromorphic continuation on $\mathbb{C}$, we prove the meromorphic continuation of $K^{(i)}_{a,b}$ on $\left\\{s|\Re(s)>-k\right\\}$ for any non- negative number $k$. In particular, for $k=0$, we have the pointwise convergence of $X^{(i)}$ on $\left\\{s|\Re(s)>0\right\\}$. Now, to see the limit at $0$, if $|s|\leq\frac{1}{2}$, then for any positive integer $m$, we have $\displaystyle|\binom{-s}{m}|$ $\displaystyle=|\frac{|s|\times(|s|+1)\times(|s|+2)...\times(|s|+m-1)}{1\times 2\times 3...\times m}|$ (12) $\displaystyle\leq|s|\prod_{n=1}^{m}\frac{|s|+n}{n+1}$ $\displaystyle\leq|s|.$ So that $|-a^{-s}\sum_{m\geq 2}\binom{-s}{m}(\frac{b}{a})^{m}X^{(i)}(s+m)|\leq|s|a^{-\Re(s)}\sum_{m\geq 2}X^{(i)}(\Re(s)+m)$. With the fact that $\sum_{m\geq 2}X^{(i)}(|s|+m)$ is bounded, we have $\lim_{s\to 0}K^{(i)}_{a,b}(s)=\lim_{s\to 0}sa^{-s}\frac{b}{a}X^{(i)}(s+1).$ (13) ## 5 Dirichlet series and infinite series ### 5.1 Infinite series involving $(r(n))_{n\in\mathbb{N}^{+}}$ Let $(r(n))_{n\in\mathbb{N}^{+}}$ be the image of the sequence $(d(n))_{n\in\mathbb{N}^{+}}$ under the map $0\to 0,\;1\to 1,\;2\to-1$. From Remark 1 and the description of the sequence A270788, the sequence $(r(n))_{n\in\mathbb{N}^{+}}$ is the fixed point of the morphism $0\to 0,1$; $1\to-1$ and $-1\to 0,1$. Now let us consider the following functions: $P(s)=\sum_{n\geq 2}r(n)(\frac{1}{\delta(n-1)^{s}}-\frac{1}{\delta(n)^{s}});$ From Proposition 4 and Proposition 6, $r(n)=1$ if and only if $d(n-1)=0$ and $\delta(n)=\delta(n-1)+\frac{\beta}{\sqrt{5}}$; similarly, $r(n)=-1$ if and only if there exists an $m$ such that $d(m)=0$ and $\tau_{0}(m)=n-1$; moreover, $\delta(n)=\delta(n-1)+\frac{\beta^{2}}{\sqrt{5}}$. Thus, for all $s$ such that $\Re(s)>1$, $\displaystyle P(s)$ $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}(\frac{1}{\delta(n-1)^{s}}-\frac{1}{\delta(n)^{s}})-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}(\frac{1}{\delta(n-1)^{s}}-\frac{1}{\delta(n)^{s}})$ (14) $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}(\frac{1}{\delta(n)^{s}}-\frac{1}{\delta(n+1)^{s}})-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}(\frac{1}{\delta(\tau_{0}(n))^{s}}-\frac{1}{(\delta(\tau_{0}(n)+1))^{s}})$ $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}(\frac{1}{\delta(n)^{s}}-\frac{1}{(\delta(n)+\frac{\beta}{\sqrt{5}})^{s}})-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}(\frac{1}{(\beta\delta(n))^{s}}-\frac{1}{(\beta\delta(n)+\frac{\beta^{2}}{\sqrt{5}})^{s}})$ $\displaystyle=(1-\beta^{-s})\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}(\frac{1}{\delta(n)^{s}}-\frac{1}{(\delta(n)+\frac{\beta}{\sqrt{5}})^{s}}).$ This equation yields two consequences. First, from Proposition 7, the infinite sum on the righthand side has a meromorphic continuation on the whole complex plane and $P(0)=0$; secondly, the function $P$ is a derivative on a neighbourhood of $s=0$ and $P^{\prime}(s)=\ln(\beta)\beta^{-s}K^{(0)}_{1,\frac{\beta}{\sqrt{5}}}(s)+(1-\beta^{-s})\frac{d}{ds}K^{(0)}_{1,\frac{\beta}{\sqrt{5}}}(s),$ and thus $P^{\prime}(0)=\ln(\beta)\frac{\beta-1}{\sqrt{5}}.$ (15) Now let us compute an alternative presentation of $P^{\prime}(0)$. From (14), we can compute further $\displaystyle P(s)$ $\displaystyle=(1-\beta^{-s})\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}(\frac{1}{\delta(n)^{s}}-\frac{1}{(\delta(n)+\frac{\beta}{\sqrt{5}})^{s}})$ (16) $\displaystyle=(\beta^{-s}-1)\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}\sum_{m\geq 1}\binom{-s}{m}(\frac{\beta}{\sqrt{5}\delta(n)})^{m}$ $\displaystyle=(\beta^{-s}-1)\sum_{m\geq 1}\binom{-s}{m}(\frac{\beta}{\sqrt{5}})^{m}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{m+s}}$ $\displaystyle=(\beta^{-s}-1)\sum_{m\geq 1}\binom{-s}{m}(\frac{\beta}{\sqrt{5}})^{m}G(m+s).$ On the other hand, for all $s$ such that $\Re(s)>1$, $\displaystyle P(s)$ $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}(\frac{1}{\delta(n-1)^{s}}-\frac{1}{\delta(n)^{s}})-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}(\frac{1}{\delta(n-1)^{s}}-\frac{1}{\delta(n)^{s}})$ (17) $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n)^{s}}-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(\tau_{0}(n))^{s}}+\sum_{\begin{subarray}{c}n\geq 2\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}$ $\displaystyle=(1-\beta^{-1}-\frac{\beta^{-s}}{1-\beta^{-2s}}+\frac{\beta^{-2s}}{1-\beta^{-2s}})\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}$ $\displaystyle=\frac{1-\beta^{-s}-\beta^{-2s}}{1+\beta^{-s}}G(s).$ Substituting (17) into (16), we have $\displaystyle P(s)$ $\displaystyle=(\beta^{-s}-1)\sum_{m\geq 1}\binom{-s}{m}(\frac{\beta}{\sqrt{5}})^{m}G(m+s)$ (18) $\displaystyle=(\beta^{-s}-1)\sum_{m\geq 1}\binom{-s}{m}(\frac{\beta}{\sqrt{5}})^{m}\frac{1+\beta^{-m-s}}{1-\beta^{-m-s}-\beta^{-2m-2s}}P(m+s).$ The infinite sum on the righthand side converges uniformly on $\left\\{s|\Re(s)>0\right\\}$: thus, $P(1)=\sum_{n\geq 2}r(n)(\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)})$. Moreover, from the fact that $P(0)=0$, to compute $P^{\prime}(0)$, it is sufficient to compute $\lim_{s\to 0}\frac{P(s)}{s}$. Thus, $\displaystyle\lim_{s\to 0}\frac{P(s)}{s}$ $\displaystyle=\lim_{s\to 0}(\beta^{-s}-1)\sum_{m\geq 1}\frac{\binom{-s}{m}}{s}(\frac{\beta}{\sqrt{5}})^{m}\frac{1+\beta^{-m-s}}{1-\beta^{-m-s}-\beta^{-2m-2s}}P(m+s)$ (19) $\displaystyle=\lim_{s\to 0}(1-\beta^{-s})(\frac{\beta}{\sqrt{5}})\frac{1+\beta^{-1-s}}{1-\beta^{-1-s}-\beta^{-2-2s}}P(1+s)$ $\displaystyle+\lim_{s\to 0}(\beta^{-s}-1)\sum_{m\geq 2}(-1)^{m}(\frac{\beta}{\sqrt{5}})^{m}\frac{1+\beta^{-m-s}}{1-\beta^{-m-s}-\beta^{-2m-2s}}P(m+s)$ $\displaystyle=(1+\beta^{-1})(\frac{\beta}{\sqrt{5}})P(1)\lim_{s\to 0}\frac{1-\beta^{-s}}{1-\beta^{-1-s}-\beta^{-2-2s}}$ $\displaystyle=(\frac{\beta+1}{\sqrt{5}})P(1)\lim_{s\to 0}\frac{\ln(\beta)\beta^{-s}}{\ln(\beta)\beta^{-1-s}+2\ln(\beta)\beta^{-2-2s}}$ $\displaystyle=(\frac{\beta+1}{\sqrt{5}})P(1).$ Combining Equation 15, Equation 19 and Remark 1, we have ###### Proposition 8 Letting $(r(n))_{n\in\mathbb{N}^{+}}$ be the image of the sequence A270788 in OEIS under the map $1\to 0,2\to 1$ and $3\to-1$, then we have $\sum_{n\geq 2}r(n)(\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)})=\frac{\beta-1}{\beta^{2}}\ln(\beta).$ (20) ###### Corollary 3 Letting $(d(n))_{n\in\mathbb{N}^{+}}$ be the sequence defined as above, then we have $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)}={\beta}^{-1}\ln(\beta).$ (21) $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)}=\beta^{-2}\ln(\beta).$ (22) * Proof First, it is easy to check that the infinite series $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)}$ and $\sum_{\begin{subarray}{c}n\geq 2\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)}$ are both well defined. Second, from (14), $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)}=\beta^{-1}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)}.$ Third, from (20), $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)}-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)}=\frac{\beta-1}{\beta^{2}}\ln(\beta).$ Combining the last two equations, we complete the proof. Moreover, let us recall the fact $\delta(2)=\sum_{n\geq 2}(\delta(n)-\delta(n+1))$ (23) By calculating (23)-(20), we have ###### Proposition 9 Letting $(t(n))_{n\in\mathbb{N}^{+}}$ be the image of the sequence A270788 in OEIS under the map $1\to 1,2\to 0$ and $3\to 2$, then we have $\sum_{n\geq 2}t(n)(\frac{1}{\delta(n-1)}-\frac{1}{\delta(n)})=\beta^{2}-\frac{\beta-1}{\beta^{2}}\ln(\beta).$ (24) ### 5.2 Infinite series involving $(s(n))_{n\in\mathbb{N}^{+}}$ Here let us consider another example. Let $(s(n))_{n\in\mathbb{N}^{+}}$ be a sequence defined in the following way: $s(n)=\begin{cases}-1\;\text{if}\;d(s)=0\;\text{or}\;2\\\ 2\;\text{if the Zeckendorf expansion of $n$ admits the string $\overline{1,0}$ as a suffix}\\\ 1\;\text{otherwise},\end{cases}$ and let us consider the following function: $Q(s)=\sum_{n\geq 1}\frac{s(n)}{\delta(n)^{s}}.$ From Proposition 6 and Corollary 1, $s(n)=2$ if and only if there exists a $m$, such that $d(m)=0$ and $\tau_{0}(m)=n$. Thus, for all $s$ such that $\Re(s)>1$, $\displaystyle Q(s)$ $\displaystyle=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n)^{s}}-\sum_{\begin{subarray}{c}n\geq 0\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=2\end{subarray}}\frac{1}{\delta(n)^{s}}+\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(\tau_{0}(n))^{s}}$ (25) $\displaystyle=\frac{\beta^{-s}}{1-\beta^{-2s}}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}-\sum_{\begin{subarray}{c}n\geq 0\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}-\frac{\beta^{-2s}}{1-\beta^{-2s}}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}+\beta^{-s}\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(n)^{s}}$ $\displaystyle=\frac{\beta^{-s}+\beta^{-2s}-1}{1+\beta^{-s}}G(s).$ Substituting (3) into (25), we have $\displaystyle Q(s)$ $\displaystyle=\frac{\beta^{-s}+\beta^{-2s}-1}{1+\beta^{-s}}G(s)=\frac{(\beta^{-s}+\beta^{-2s}-1)}{1+\beta^{-s}}((1-\beta^{-s})F(s)-\delta(1)^{-s})$ (26) $\displaystyle=\frac{(-\beta^{-3s}+2\beta^{-s}-1)}{1+\beta^{-s}}F(s)-\frac{(\beta^{-s}+\beta^{-2s}-1)\delta(1)^{-s}}{1+\beta^{-s}}.$ On the other hand, $Q(s)=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1,\;\text{or}\;2\end{subarray}}\frac{1}{\delta(n)^{s}}-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0,\;\text{or}\;2\end{subarray}}\frac{1}{\delta(n)^{s}}-(\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1,\;\text{or}\;2\end{subarray}}\frac{1}{\delta(n)^{s}}-\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(\tau_{0}(n))^{s}})+\sum_{\begin{subarray}{c}n\geq 0\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n)^{s}}.$ (27) From Proposition 4 and Proposition 6, $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1,\;\text{or}\;2\end{subarray}}\frac{1}{\delta(n)^{s}}=\sum_{n\geq 1}\frac{1}{\delta(\tau_{0}(n))^{s}}=\sum_{n\geq 1}\frac{1}{\delta(\tau_{0}(\tau_{0}(n)))^{s}}+\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0\end{subarray}}\frac{1}{\delta(\tau_{0}(n))^{s}}.$ $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0,\;\text{or}\;2\end{subarray}}\frac{1}{\delta(n)^{s}}=\sum_{n\geq 1}\frac{1}{\delta(\tau_{0}(n)+1)^{s}}+\frac{1}{\delta(1)^{s}}$ $\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=1\end{subarray}}\frac{1}{\delta(n)^{s}}=\sum_{\begin{subarray}{c}n\geq 1\\\ d(n)=0,\;\text{or}\;2\end{subarray}}\frac{1}{\delta(\tau_{0}(n))^{s}}=\sum_{n\geq 1}\frac{1}{\delta(\tau_{0}(\tau_{0}(n)+1))^{s}}+\frac{1}{\delta(2)^{s}}$ Thus, $\displaystyle Q(s)$ $\displaystyle=\sum_{n\geq 1}\frac{1}{\delta(\tau_{0}(n))^{s}}-(\sum_{n\geq 1}\frac{1}{\delta(\tau_{0}(n)+1)^{s}}+\frac{1}{\delta(1)^{s}})-\sum_{n\geq 1}\frac{1}{\delta(\tau_{0}(\tau_{0}(n)))^{s}}+\sum_{n\geq 1}\frac{1}{\delta(\tau_{0}(\tau_{0}(n)+1))^{s}}+\frac{1}{\delta(2)^{s}}$ (28) $\displaystyle=\sum_{n\geq 1}(\frac{1}{(\beta\delta(n))^{s}}-\frac{1}{(\beta\delta(n)+\frac{\beta^{2}}{\sqrt{5}})^{s}})-\sum_{n\geq 1}(\frac{1}{(\beta^{2}\delta(n))^{s}}-\frac{1}{(\beta^{2}\delta(n)+\frac{\beta^{3}}{\sqrt{5}})^{s}})-\frac{1}{\delta(1)^{s}}+\frac{1}{\delta(2)^{s}}$ $\displaystyle=(1-\beta^{-s})\sum_{m\geq 1}(\frac{1}{(\beta\delta(n))^{s}}-\frac{1}{(\beta\delta(n)+\frac{\beta^{2}}{\sqrt{5}})^{s}}))-\frac{1}{\delta(1)^{s}}+\frac{1}{\delta(2)^{s}}.$ From Proposition 7, for all $s$ such that $\Re(s)>1$, $\displaystyle Q(s)$ $\displaystyle=(1-\beta^{-s})\sum_{m\geq 1}(\frac{1}{(\beta\delta(n))^{s}}-\frac{1}{(\beta\delta(n)+\frac{\beta^{2}}{\sqrt{5}})^{s}}))-\frac{1}{\delta(1)^{s}}+\frac{1}{\delta(2)^{s}}$ (29) $\displaystyle=(\beta^{-2s}-1)\sum_{m\geq 1}\binom{-s}{m}(\frac{\beta}{\sqrt{5}})^{m}F(s+m)-\frac{1}{\delta(1)^{s}}+\frac{1}{\delta(2)^{s}}.$ Substituting (26) into (29), we have $\displaystyle Q(s)$ $\displaystyle=(\beta^{-2s}-1)\sum_{m\geq 1}\binom{-s}{m}(\frac{\beta}{\sqrt{5}})^{m}\frac{1+\beta^{-m-s}}{-\beta^{-3(m+s)}+2\beta^{-(m+s)}-1}(Q(s+m)+\frac{(\beta^{-m-s}+\beta^{-2m-2s}-1)\delta(1)^{-m-s}}{1+\beta^{-m-s}})$ (30) $\displaystyle-\frac{1}{\delta(1)^{s}}+\frac{1}{\delta(2)^{s}}$ From Equation 28 and Proposition 7, we can prove that the function $Q$ has a meromorphic continuation on the whole complex plane. Moreover, we have $Q(0)=0$ and $Q^{\prime}(s)=\ln(\beta)\beta^{-s}K^{(3)}_{\beta,\frac{\beta^{2}}{\sqrt{5}}}(s)+(1-\beta^{-s})\frac{d}{ds}K^{(3)}_{\beta,\frac{\beta^{2}}{\sqrt{5}}}(s)-\ln(\beta),$ thus $Q^{\prime}(0)=\ln(\beta)\frac{\beta}{\sqrt{5}}-\ln(\beta).$ (31) From (30), the infinite sum on the righthand side converges uniformly on $\left\\{s|\Re(s)>0\right\\}$, thus $Q(1)=\sum_{n\geq 1}\frac{s(n)}{\delta(n)}$. To compute $Q^{\prime}(0)$, it is sufficient to compute $\lim_{s\to 0}\frac{Q(s)}{s}$. Thus, $\displaystyle\lim_{s\to 0}\frac{Q(s)}{s}$ $\displaystyle=\lim_{s\to 0}(\beta^{-2s}-1)\sum_{m\geq 1}\binom{-s}{m}(\frac{\beta}{\sqrt{5}})^{m}(\frac{1+\beta^{-m-s}}{-\beta^{-3(m+s)}+2\beta^{-(m+s)}-1}Q(s+m)+\frac{1}{1-\beta^{-m-s}}\delta(1)^{-m-s})$ (32) $\displaystyle+\lim_{s\to 0}\frac{-\frac{1}{\delta(1)^{s}}+\frac{1}{\delta(2)^{s}}}{s}$ $\displaystyle=\lim_{s\to 0}(1-\beta^{-2s})(\frac{\beta}{\sqrt{5}})(\frac{1+\beta^{-m-s}}{-\beta^{-3(1+s)}+2\beta^{-(1+s)}-1}Q(s+1)+\frac{1}{1-\beta^{-1-s}}\delta(1)^{-1-s})$ $\displaystyle-\lim_{s\to 0}\frac{\frac{1}{\delta(1)^{s}}-1}{s}+\lim_{s\to 0}\frac{\frac{1}{\delta(2)^{s}}-1}{s}$ $\displaystyle+\lim_{s\to 0}(\beta^{-2s}-1)\sum_{m\geq 2}\binom{-s}{m}(\frac{1+\beta^{-m-s}}{-\beta^{-3(m+s)}+2\beta^{-(m+s)}-1}Q(s+m)+\frac{1}{1-\beta^{-m-s}}\delta(1)^{-m-s})$ $\displaystyle=(\frac{\beta}{\sqrt{5}})Q(1)\lim_{s\to 0}\frac{1-\beta^{-2s}}{-\beta^{-3(1+s)}+2\beta^{-(1+s)}-1}(1+\beta^{-1-s})+\ln(\delta(1))-\ln(\delta(2))$ $\displaystyle=(\frac{\beta}{\sqrt{5}})Q(1)(1+\beta^{-1})\lim_{s\to 0}\frac{2\ln(\beta)\beta^{-2s}}{3\ln(\beta)\beta^{-3-3s}-2\ln(\beta)\beta^{-1-s}}-\ln(\beta)$ $\displaystyle=(\frac{\beta}{\sqrt{5}})Q(1)\frac{2+2\beta^{-1}}{3\beta^{-3}-2\beta^{-1}}-\ln(\beta).$ Combining (31) and (32), we have ###### Proposition 10 Letting $(s(n))_{n\in\mathbb{N}^{+}}$ be the sequence defined as above, then $\sum_{n\geq 2}\frac{s(n)}{\delta(n)}=(\frac{3}{2}\beta^{-4}-\beta^{-2})\ln(\beta).$ (33) To obtain the Theorem 2, we only need to apply the following proposition: ###### Proposition 11 For any positive integer $n$, $\delta(n)=n-\frac{\\{\beta n\\}-1+f(n)}{\sqrt{5}},$ where $\\{x\\}$ is the fractional part of $x$. * Proof Letting $n$ be a positive integer, then $n=\delta(n)-\delta^{\prime}(n)$ and $\tau_{0}(n)=\beta\delta(n)+\beta^{-1}\delta^{\prime}(n)$. Consequently, $\beta n-\tau_{0}(n)=-(\beta+\beta^{-1})\delta^{\prime}(n)=-\sqrt{5}\delta^{\prime}(n)$. Moreover, from Proposition 2, $|\delta^{\prime}(n)|<\frac{1}{\sqrt{5}\beta}$, we have $|\sqrt{5}\delta^{\prime}(n)|<1$. Thus, $\\{\beta n\\}=1-\sqrt{5}\delta^{\prime}(n)$ if $\delta^{\prime}(n)>0$ and $\\{\beta n\\}=-\sqrt{5}\delta^{\prime}(n)$ if $\delta^{\prime}(n)<0$. Consequently, $\delta(n)=n-\frac{\\{\beta n\\}-1}{\sqrt{5}}$ if $f(n)=0$ and $\delta(n)=n-\frac{\\{\beta n\\}}{\sqrt{5}}$ if $f(n)=1$. ## References * [AC85] J.-P. Allouche and H. Cohen. Dirichlet Series and Curious infinite Products. Bulletin of the London Mathematical Society, 17(6):531–538, 1985\. * [ARS18] J.-P. Allouche, S. Riasat, and J. Shallit. More infinite products: Thue–Morse and the gamma function. The Ramanujan Journal, 49(1):115–128, 2018. * [AS89] J.-P. Allouche and J. Shallit. Infinite Products Associated with Counting Blocks in Binary Strings. Journal of the London Mathematical Society, s2-39(2):193–204, 1989\. * [AS03] J.-P. Allouche and J. Shallit. Automatic Sequences: Theory, Applications, Generalizations. Cambridge University Press, 2003. * [ASS07] J.-P. Allouche, J. Shallit, and J. Sondow. Summation of series defined by counting blocks of digits. Journal of Number Theory, 123(1):133–143, 2007. * [DMR+17] C. F. Du, H. Mousavi, E. Rowland, L. Schaeffer, and J. Shallit. Decision algorithms for Fibonacci-automatic words, II: Related sequences and avoidability. Theoretical Computer Science, 657:146–162, 2017. * [DMSS16] C. F. Du, H. Mousavi, L. Schaeffer, and J. Shallit. Decision Algorithms for Fibonacci-Automatic Words, III: Enumeration and Abelian Properties. International Journal of Foundations of Computer Science, 27(08):943–963, 2016. * [Hu16] Y. Hu. Patterns in numbers and infinite sums and products. Journal of Number Theory, 162:589–600, 2016. * [MSS16] H. Mousavi, L. Schaeffer, and J. Shallit. Decision algorithms for Fibonacci-automatic Words, I: Basic results. RAIRO - Theoretical Informatics and Applications, 50(1):39–66, 2016. * [SI20] Neil J. A. Sloane and The OEIS Foundation Inc. The on-line encyclopedia of integer sequences, 2020. * [Sou19] A. Sourmelidis. On the meromorphic continuation of Beatty Zeta-functions and Sturmian Dirichlet series. Journal of Number Theory, 194:303–318, 2019. * [Zec72] E. Zeckendrof. Représentations des nombres naturels par une somme de nombres de fibonacci on de nombres de lucas. Bulletin de La Society Royale des Sciences de Liege, pages 179–182, 1972.
# Template ###### Abstract We revisit the gravity path integral formalism of JT gravity. We explain how to gauge fix the path integral in the presence of asymptotic boundaries and conical defects, and resolve an ambiguity regarding the dilaton gravity operator that creates a conical defect. Along the way we study JT gravity coupled to matter on surfaces with defects of special opening angles, obtaining expressions for partition and two-point functions of matter fields. The two point function involves a summation over all geodesics on the surface, including self-intersecting geodesics, which we formally manage to include. Revisiting the second order formalism of JT gravity Guanda Lin1, Mykhaylo Usatyuk1,2 1 Center for Theoretical Physics and Department of Physics, Berkeley, CA, 94720, USA 2 Kavli Institute for Theoretical Physics, Santa Barbara, CA 93106, USA <EMAIL_ADDRESS><EMAIL_ADDRESS> ###### Contents 1. 1 Introduction 2. 2 Determinants and correlators on hyperbolic surfaces 1. 2.1 Building hyperbolic geometries 2. 2.2 Two-point functions and determinants on quotient surfaces 3. 2.3 JT examples 1. 2.3.1 Double trumpet 2. 2.3.2 Conical defect 3. 2.3.3 Two conical defects 4. 2.3.4 The handle disk 3. 3 JT gravity path integral 1. 3.1 Gauge fixing the path integral 1. 3.1.1 Integrating over moduli space 2. 3.2 Examples 1. 3.2.1 Disk 2. 3.2.2 Conical defect 3. 3.2.3 Conical defect: $z$ coordinates 4. 3.2.4 General surfaces 4. 4 Discussion 5. A Consistency of disk measure with general surface 6. B Double trumpet gluing measure 7. C The determinant calculation ## 1 Introduction Jackiw-Teitelboim (JT) gravity is a simple two dimensional model that has been central to many recent developments in quantum gravity[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]. In this paper we will be revisiting various unresolved aspects of the second order formalism of the JT gravity path integral. The most common approach to JT gravity is through the first order formalism [7, 12], especially when dealing with the issue of gauge fixing the path integral. There has been some progress on understanding JT gravity from the second order formalism. For compact surfaces the path integral was properly gauge fixed in [7], whereas for surfaces with asymptotically AdS boundaries significant progress was made in [13, 14, 15, 16]. Furthermore, when considering matter coupled to JT it is easiest to use gravitational variables. In this paper we will consider JT gravity on surfaces with asymptotically AdS boundaries with the addition of conical defects[17, 18, 19, 20]. We will calculate determinants of Laplace operators, matter field correlators, and gauge fix the gravity path integral on such surfaces from the perspective of the second order formalism. Determinants of Laplace operators have recently appeared in studies of JT gravity coupled to matter where they correspond to partition functions of matter fields on the geometry[13, 21, 22], and in the evaluation of Lorentzian JT gravity amplitudes with topology changing wormholes [23, 24]. Similarly, correlation functions of matter fields on wormhole geometries have recently appeared in [25, 26, 27, 22, 21], where they probe the length of the wormhole. Conical defect geometries have played an important role in the counting of black hole microstates[10], and matter field correlators on defect geometries were considered in [26, 28]. The procedure for gauge fixing the gravitational path integral is well known from the string perturbation theory literature[29, 30, 31], and requires the evaluation of various determinants on hyperbolic surfaces. This naturally connects the problem of gauge fixing the path integral to understanding matter minimally coupled to JT gravity. We will find that the second order formalism has certain advantages. A proper gauge fixing will resolve an ambiguity regarding the correct definition of the conical defect creation operator. Furthermore, it is straightforward to incorporate matter, and we obtain closed-form expressions for determinants and two-point functions of matter coupled to JT gravity. We now summarize our main results. ### Summary of results In section 2 we begin by explaining how to construct constant negative curvature surfaces by taking a quotient of the upper half-plane by a group $\Gamma\subset\text{PSL}(2,\mathbb{R})$, giving a surface $\Sigma=\mathbb{H}/\Gamma$. Using the quotient construction a variety of hyperbolic surfaces can be constructed including compact surfaces, surfaces with asymptotic AdS boundaries, and surfaces that include conical defects with special opening angles $\theta=\frac{2\pi}{n}$ with integer $n\geq 2$. We then explain how to compute determinants of Laplace type operators on surfaces obtained through the quotient method. The final result is the following. Figure 1: The determinant and two point function are reproduced by a sum over all geodesics on the surface $\Sigma$. On the left we have highlighted three geodesics that contribute to the determinant on the surface with one handle and one defect. On the right we have shown some geodesics that contribute to the two-point function, with the geodesics ending on the operator insertions at the boundary. There are infinitely many additional geodesics on these surfaces that contribute to these quantities. ##### Determinants. Let $\Sigma=\mathbb{H}/\Gamma$ be a hyperbolic surface, either compact or with asymptotic AdS boundaries, with some number of conical defects. Define the scalar Laplacian $\Delta=-g^{ab}\nabla_{a}\nabla_{b}$. The determinant of the Laplace operator on the surface $\Sigma$ is given by $\det\left({\Delta}+s(s-1)\right)_{\Sigma}=\underbrace{Z_{\text{hyp.}}(s)}_{\text{geodesics}}\underbrace{Z_{\text{ell.}}(s)}_{\text{defects}}\left(\text{const}\right),$ (1.1) see section 2.2 and appendix C for additional technical details. The most important contribution to the determinant is the Selberg zeta function111For compact surfaces this form of the determinant is well known [32, 33], and comes from the Selberg trace formula. For surfaces with asymptotically AdS boundaries the techniques for calculating determinants are relatively recent [34, 35, 36], and one of the new results is to incorporate conical defects in the calculation, alongside evaluating the determinant for the vector Laplacian $\det\left(\Delta_{1}+s(s-1)\right)_{\Sigma}$ in appendix C. $Z_{\text{hyp.}}(s)=\prod_{\ell_{\gamma}\in\mathcal{L}_{\Sigma}}\prod_{m=0}^{\infty}\left(1-e^{-(s+m)\ell_{\gamma}}\right),$ (1.2) which is defined to be a sum over all closed geodesics $\mathcal{L}_{\Sigma}$ on the surface $\Sigma$, with lengths given by $\ell_{\gamma}$. See figure 1. The integer $m$ counts the winding of the geodesics. The additional term $Z_{\text{ell.}}(s)$ comes from the presence of conical defects, and does not have an obvious geometric interpretation. We also obtain a similar expression for the determinant of the vector Laplacian $\det\left(\Delta_{1}+s(s-1)\right)_{\Sigma}$, see appendix C. ##### Correlators. As an intermediate step in the determinant calculation we obtain a closed form expression for the two-point function of a scalar field $\langle\phi(x)\phi(y)\rangle$ on an arbitrary surface $\Sigma$. When the points $x,y$ are sent to asymptotic AdS boundaries the geodesic approximation becomes exact, and the correlator reduces to a sum over all geodesics connecting $x$ and $y$. Dressing the operator insertions to the boundary Schwarzian, we find that the correlator for a field of mass $m^{2}=s(s-1)$ on a surface $\Sigma=\mathbb{H}/\Gamma$ is given by $G_{\Sigma}(\tau_{1},\tau_{2})=\sum_{\text{geodesics }\gamma}e^{-s\ell_{\gamma}}~{}=\sum_{T\in\Gamma}\left(\frac{F^{\prime}(\tau_{1})(T\cdot F(\tau_{2}))^{\prime}}{\left(F(\tau_{1})-T\cdot F(\tau_{2})\right)^{2}}\right)^{s},$ (1.3) where the sum over all geodesics connecting the points is given by the sum over the group $\Gamma$ and the Schwarzian fluctuations are given by $F(\tau)$, see equation (2.27).222In this formula we have the action $T\cdot F=\frac{aF+b}{cF+d}$. The formula also applies in the case that the operators are inserted on different asymptotic boundaries with independent Schwarzian fluctuations $F_{i}(\tau)$. See section 2.3 for an example. This formula includes a summation over all geodesics connecting the two operators including self-intersecting geodesics, see figure 1. In section 2.3 we work out a number of examples for both determinants and correlators. We reproduce previous results for the defect and double trumpet geometries, and obtain new expressions for the two defect and handle disk geometries. ##### Gravity path integral. In section 3 we turn to the problem of gauge fixing the gravity path integral for JT gravity. Consider performing the JT gravity path integral on surfaces $\Sigma$ of genus $g$ with asymptotically AdS boundaries of regularized lengths $\vec{\beta}=\left(\beta_{1},\ldots,\beta_{n}\right)$, along with the insertion of $k$ conical defect operators $\mathcal{V}_{\alpha}$ giving rise to conical defects with opening angles $2\pi\left(1-\alpha_{i}\right)$ specified by $\vec{\alpha}=\left(\alpha_{1},\ldots,\alpha_{k}\right)$. We show that the path integral reduces to the Weil-Petersson measure on the moduli space $\mathcal{M}_{g,\vec{\alpha},\vec{\beta}}$ of associated hyperbolic surfaces. More compactly, $Z_{g,{\vec{\alpha}}}(\beta_{1},\ldots,\beta_{n})=\int\frac{\mathcal{D}g\mathcal{D}\Phi}{\text{V}(\text{Diff})}e^{-I_{\text{J}T}[g,\Phi]}\mathcal{V}_{\alpha_{1}}\ldots\mathcal{V}_{\alpha_{k}}=\int_{\mathcal{M}_{g,{\vec{\alpha},\vec{\beta}}}}d\left(\text{Weil- Pet.}\right)e^{-I_{\text{bdy}}},$ (1.4) where on the right-hand side the boundary action is for the Schwarzian on the asymptotic AdS boundaries. In the above we have used determinants computed in section 2.333We are only able to compute the necessary determinants for certain values of $\alpha$, but we argue that the reduction to the Weil- Petersson measure should occur for all $\alpha$. We are now left with the problem of evaluating the integral on the right-hand side. It’s well known from the string perturbation theory literature how to perform the integral over moduli space [29, 31], formally we have $\int_{\mathcal{M}}d(\text{Weil-Pet.})\hskip 2.84544pte^{-I_{\text{bdy}}}=\underbrace{\prod_{n}\int dm_{n}d\overline{m}_{n}}_{\text{coordinates on $\mathcal{M}$}}\underbrace{\frac{\det\langle\mu,\phi\rangle\det\langle\overline{\mu},\overline{\phi}\rangle}{\sqrt{\det\langle\phi,\phi\rangle\det\langle\overline{\phi},\overline{\phi}\rangle}}}_{\text{WP measure}}\hskip 2.84544pte^{-I_{\text{bdy}}[m_{n},\overline{m}_{n}]}.$ (1.5) The above equation can be understood as making a choice of coordinates $m_{n}$ for the space we are integrating over, and computing the Weil-Petersson measure in that set of coordinates. The quantities appearing in the measure are known as quadratic differentials $\phi$ and Beltrami differentials $\mu$, and we explain the technical details of the measure in section 3.1.1. Computing the measure for a particular choice of coordinates is challenging for general surfaces, but in simple examples it is possible. In the case of the disk with and without a conical defect the measure can be evaluated, and in section 3.2 we perform the full integral over moduli space finding agreement with the standard Schwarzian calculation. ##### Conical defect operator. With the above we are ready to determine the dilaton gravity operator $\mathcal{V}_{\alpha}$ that creates a conical defect when inserted into the path integral (1.4). When a defect is added to the surface the real dimension of the moduli space increases by two, which corresponds to the two directions the defect can be moved on the surface. We consider a surface $\Sigma$ with a conical defect, and choose two of the coordinates $m_{n}$ for the moduli space to be the position $x$ of the defect. Using (1.5) we can immediately write down the correct operator to be $\mathcal{V}_{\alpha}=\underbrace{\int_{\Sigma}d^{2}x}_{\text{defect position}}\underbrace{\frac{\langle\mu,\phi_{1}\rangle\langle\overline{\mu},\overline{\phi}_{1}\rangle}{\sqrt{\langle\phi_{1},\phi_{1}\rangle\langle\overline{\phi}_{1},\overline{\phi}_{1}\rangle}}}_{\text{measure for $x$ coords}}e^{-2\pi\left(1-\alpha_{i}\right)\Phi(x)},$ (1.6) in the above we have reintroduced the exponential of the dilaton since in the measure $\eqref{eqn:intro_measure}$ we have already integrated out the dilaton.444When combined with the JT action, integrating out the dilaton creates a delta function source for the Ricci scalar at position $x$. For additional details on the definition of this measure see section 3.1.1 around equation (3.21). This measure cannot be explicitly evaluated for a general surface $\Sigma$. However, in section 3.2.4 we explain that in the special limit that $\alpha\to 1$, with opening angle $\theta=2\pi\alpha$, we can evaluate the measure on an arbitrary surface $\Sigma$. We find that the conical defect operator takes the form $\text{General surface:}\qquad\lim_{\alpha\to 1}\mathcal{V}_{\alpha}=2\pi(1-\alpha)\int_{\Sigma}d^{2}x\sqrt{g(x)}e^{-2\pi(1-\alpha)\Phi(x)}+\mathcal{O}\left((1-\alpha)^{2}\right),$ (1.7) up to corrections in the $\left(1-\alpha\right)$ expansion.555One interesting aspect is that the metric $g(x)$ in (1.7) is for the surface $\Sigma$ with no defect, which is important for our argument for the recursion relation for Weil-Petersson volumes found by [20]. This is consistent with previous expectations for the form of the operator in the blunt defect limit $\alpha\to 1$ [18, 17, 19, 20], and constitutes a gravity path integral argument for the form of the operator. In the case of the disk we can do better, and in section 3.2.3 We evaluate the full measure and find $\text{Disk:}\qquad\mathcal{V}_{\alpha}=\pi(1-\alpha^{2})\int d^{2}x\sqrt{g(x)}e^{-2\pi(1-\alpha)\Phi(x)}.$ (1.8) In appendix A we show how the calculation on the disk can be thought of as performing a resummation of the $\left(1-\alpha\right)$ expansion appearing for general surfaces (1.7). We thus conjecture that the general form of the conical defect operator on an arbitrary surface should take the form $\textbf{Conjecture for general surfaces: }\qquad\mathcal{V}_{\alpha}=\pi(1-\alpha^{2})\int_{\Sigma}d^{2}x\sqrt{g(x)}e^{-2\pi(1-\alpha)\Phi(x)},$ (1.9) after the resummation in (1.7) is performed. Our argument for the above form is that the operator should be independent of the surface it is inserted on, as is the case when these operators are defined as limits of minimal string operators [19, 20], and so the answer we find on the disk should carry over to other surfaces. ##### Applications to dilaton gravity. There are two immediate applications of the conical defect operator. The first is it clarifies the correct dilaton potential that corresponds to JT gravity coupled to a gas of conical defects. In [18, 17, 19, 20] JT gravity coupled to a gas of conical defects was defined by a summation over conical Weil- Petersson volumes with a coupling $\lambda$ weighing each defect. The bulk dilaton gravity action that corresponds to this theory is thus given by $I[g,\Phi]=-\frac{1}{2}\int d^{2}x\sqrt{g}\left(\Phi R+2U(\Phi)\right),$ (1.10) with the dilaton potential following from equation (1.9) $U(\Phi)=\Phi+\pi\lambda(1-\alpha^{2})e^{-2\pi\left(1-\alpha\right)\Phi}.$ (1.11) At each order in the $\lambda$ expansion the path integral will localize onto singular hyperbolic surfaces and reproduce the appropriate Weil-Petersson volume with the required coupling $\lambda^{k}$ for surfaces with $k$ defects. A secondary application is that the conical defect operator can be used to give a gravity path integral argument for a recursion relation of Weil- Petersson volumes found by [20]. Namely, in the limit that one defect on a surface becomes blunt the volume becomes related to the volume without the defect through $\frac{dV_{g,m,n+1}\left(\vec{\alpha}_{n+1},\vec{b}_{m}\right)}{d\alpha_{n+1}}\bigg{\rvert}_{\alpha_{n+1}=1}=4\pi^{2}\chi\left(\Sigma\right)V_{g,m,n}\left(\vec{\alpha}_{n},\vec{b}_{m}\right).$ (1.12) In section 3.2.4 we give a gravity path integral argument for this recursion relation using the conical defect operator. ## 2 Determinants and correlators on hyperbolic surfaces ### 2.1 Building hyperbolic geometries We now explain how to build constant negative curvature geometries with conical defects and asymptotic boundaries by taking quotients of the upper half-plane (UHP) by an appropriate group. For additional details see [36]. The quotient construction is useful for calculating determinants and correlation functions on such surfaces since the method of images can be used. Consider the upper half-plane $\mathbb{H}$ with the standard AdS2 metric $ds^{2}=\frac{dzd\overline{z}}{\left(\operatorname{Im}z\right)^{2}}\,,$ (2.1) The group of isometries is given by PSL$(2,\mathbb{R})$, and elements of the group $T\in\text{PSL}(2,\mathbb{R})$ are either hyperbolic, elliptic, or parabolic. Hyperbolic elements satisfy $\Tr T>2$, elliptic satisfy $\Tr T<2$, while parabolic satisfy $\Tr T=2$. We will primarily be interested in hyperbolic and elliptic elements, and the most general form of these elements is given by $\text{Hyperbolic: }T_{\ell}=\Lambda^{-1}\left(\begin{array}[]{ll}e^{\ell/2}&0\\\ 0&e^{-\ell/2}\end{array}\right)\Lambda,\qquad\text{Elliptic: }T_{\theta}=\Lambda^{-1}\left(\begin{array}[]{ll}\cos(\frac{\theta}{2})&-\sin(\frac{\theta}{2})\\\ \sin(\frac{\theta}{2})&~{}~{}\cos(\frac{\theta}{2})\end{array}\right)\Lambda\,,$ (2.2) where $\Lambda\in\text{PSL}(2,\mathbb{R})$. The action of $T$ on the upper half-plane is defined by $T\cdot z=\frac{az+b}{cz+d}\,.$ (2.3) As an example, setting $\Lambda=1$ the hyperbolic element acts as $T_{\ell}\cdot z=e^{\ell}z$. Elliptic elements leave a fixed point in the interior which will become a bulk conical defect when we consider the quotient geometry.666Parabolic elements leave a fixed point at the asymptotic boundary. The fixed point corresponds to a conical singularity with deficit angle $2\pi$, and is known as a cusp. We will not consider cusps because the quotient method becomes much more complicated since the cusp lives on the asymptotic boundary. However, all of our claims should obviously generalize to surfaces that include cusps. One way to build a hyperbolic surface is to take a subgroup of the isometry group $\Gamma\subset\text{PSL}(2,\mathbb{R})$ and take the quotient of the upper half-plane by this subgroup $\Gamma\backslash\mathbb{H}$. That is, we identify two points as equivalent if $z\cong T\cdot z$ for any $T\in\Gamma$. To build a good hyperbolic surface we must restrict to Fuchsian groups $\Gamma\subset\text{PSL}(2,\mathbb{R})$, which are discrete subgroups of $\text{PSL}(2,\mathbb{R})$. This restricts the elliptic elements we are allowed to consider to $T_{\theta}^{n}=\pm\operatorname{I}$, which only allows opening angles $\theta=2\pi/n$ with integer $n\geq 2$.777Restricting to Fuchsian groups is equivalent to imposing the condition that a sufficiently small ball around the identity element, in the $SL(2,\mathbb{R})$ group manifold, contains no other elements. Intuitively, when quotienting $z\cong T\cdot z$ we do not want there to exist elements $T\in\Gamma$ arbitrarily close to the identity, since then we will be identifying arbitrarily close points $z,T\cdot z$, and the quotient surface will be degenerate. We will take the Fuchsian group to have a finite number of generators $T_{i}$ given by some set set of hyperbolic and elliptic elements (2.2). We are interested in hyperbolic surfaces with both asymptotic boundaries and conical defects. Locally, a conical defect at a point $x_{i}$ is characterized by the fact that we can travel around the point by going through an angle $\theta<2\pi$. For a conical defect of opening angle $\theta=2\pi\alpha$ this translates to a condition on the scalar curvature given by $\frac{1}{2}\sqrt{g}\left(R+2\right)=2\pi\left(1-\alpha\right)\delta^{2}(x-x_{i})\,,$ (2.4) where we restrict to $\alpha\in(0,1)$. A natural question is whether there exists a hyperbolic surface with $k$ defects with angles specified by $\vec{\alpha}=\left(\alpha_{1},\ldots,\alpha_{k}\right)$. It turns out that such a surface always exists provided that the specification of the defects $\alpha_{i}$ does not violate the Gauss-Bonnet theorem [37, 38] $2\pi\chi(\Sigma)=\frac{1}{2}\int_{\Sigma}\sqrt{g}R+\int_{\partial\Sigma}\sqrt{h}K\,,$ (2.5) where $K$ is the extrinsic curvature on the boundary of the surface, $h$ is the induced boundary metric, and $\chi(\Sigma)=2-2g-n$ where $n$ is the number of boundaries and $g$ is the genus. Conical defects fall into two classes: sharp defects with opening angle $\theta\leq\pi$, and blunt defects with opening angle $\theta>\pi$ [19, 20]. This translates to the condition $\alpha\leq\frac{1}{2}$ for sharp and $\alpha>\frac{1}{2}$ for blunt.888One nice property of sharp defects is that for surfaces with an asymptotic boundary and at least two sharp defects, the defects are separated from the boundary by a closed geodesic homotopic to the boundary. This can be seen from the Gauss-Bonnet theorem. The quotient construction only allows us to build surfaces with sharp defects $\alpha\leq\frac{1}{2}$. We now summarize existence theorems on the types of surfaces that can be built using the quotient method. First, consider the moduli space $\mathcal{M}_{g}$ of compact surfaces of genus $g\geq 2$. For each compact hyperbolic surface $\Sigma$ there exists a Fuchsian group $\Gamma$ such that $\Sigma=\Gamma\backslash\mathbb{H}$, and such a representation of the surface is known as a Fuchsian model. As an example, for $g=2$ the Fuchsian group is generated by $4$ hyperbolic elements $T_{i}$ with a non-trivial constraint999The constraint enforces that the closed loop generated by the action of the following group element will be contractible on the surface. on the generators $\Gamma=\langle T_{1},T_{2},T_{3},T_{4}~{}|~{}T_{4}^{-1}T_{3}^{-1}T_{4}T_{3}T_{2}^{-1}T_{1}^{-1}T_{2}T_{1}=1\rangle.$ (2.6) In general, the Fuchsian group of a compact genus $g$ surface is generated by $2g$ distinct hyperbolic elements satisfying non-trivial constraints. Now consider the moduli space $\mathcal{M}_{g,\vec{\alpha}}$ of compact surfaces of genus $g$ with $k$ conical defects with deficit angles specified by $\vec{\alpha}=\left(\alpha_{1},\ldots,\alpha_{k}\right)$. When the opening angles take special values $\theta_{i}=2\pi\alpha_{i}$ with $\alpha_{i}=\frac{1}{n_{i}}$ and integer $n_{i}\geq 2$ then every surface can be obtained by a quotient with a suitable Fuchsian group[39, 40]. The group will contain $k$ elliptic elements that leave a fixed point $z_{i}\in\mathbb{H}$ which becomes the location of the conical defect.101010The order of the elliptic element determines the strength of the opening angle to be $2\pi/n$. Surfaces with other deficit angles cannot be constructed by the quotient method. In JT gravity we typically consider non-compact surfaces with $n$ asymptotic boundaries with regularized lengths $\vec{\beta}=\left(\beta_{1},\ldots,\beta_{n}\right)$. All hyperbolic surfaces with asymptotic boundaries, but without conical defects, $\mathcal{M}_{g,\vec{\beta}}$ can be obtained by a quotient with a Schottky group (see theorem 15.3 in [36]). We are not aware of a similar theorem in the case that we include conical defects $\mathcal{M}_{g,\vec{\alpha},\vec{\beta}}$, but we will assume that if we limit the opening angles $\vec{\alpha}$ to the previously mentioned special values that these surfaces can also be obtained through a quotient. ### 2.2 Two-point functions and determinants on quotient surfaces The quotient construction makes it possible to calculate the determinant and two-point function on the associated surface. In this subsection we explain how to evaluate these quantities. Let $\Sigma$ be a hyperbolic surface with metric $g$ obtained by quotienting $\mathbb{H}$ with a Fuchsian group $\Gamma$. The scalar Laplacian is defined to be $\Delta=-g^{ab}\nabla_{a}\nabla_{b}$. The two-point function on the surface $\Sigma$, also known as the resolvent, is defined through the equation $\left(\Delta+s(s-1)\right)R_{\Sigma}(s;z,w)=\frac{\delta^{2}(z-w)}{\sqrt{g}}\,,$ (2.7) where $z,w$ are two points on the surface $\Sigma$, and $s\geq 1$ is related to the mass of a scalar field through $m^{2}=s(s-1)$. For surfaces obtained through the quotient method we can obtain a formula for the resolvent by using the method of images. The key idea is that the resolvent on the upper half plane $R_{\mathbb{H}}$, which is explicitly known, locally satisfies the desired equation (2.7), but globally we must have that $R_{\Sigma}(s;z,w)=R_{\Sigma}(s;T\cdot z,w)$ since $z\cong T\cdot z$ for all $T\in\Gamma$. Performing a sum over the quotient group gives a function that precisely satisfies all the properties that define the resolvent on $\Sigma$ $R_{\Sigma}(s;z,w)=\sum_{T\in\Gamma}R_{\mathbb{H}}(s;T\cdot z,w).$ (2.8) Once the resolvent is known it is straightforward to calculate the determinant. For smooth compact surfaces it is well known that the determinant of the Laplacian reduces to a product over geodesics due to the Selberg trace formula[32, 33]. However, for non-compact surfaces this method is not applicable, and other techniques need to be used[34, 35, 36]. We will review the technique for surfaces with asymptotic boundaries[36], and extend it to include conical defects. The basic procedure of the calculation and key results will be presented here, leaving technical details to appendix C. To compute the determinant of an operator we must solve the eigenvalue problem $\left(\Delta+s(s-1)\right)\phi_{n}=\lambda_{n}\phi_{n}$ and take the product of all the eigenvalues. However, doing this directly is quite challenging. A simpler technique is to define the determinant through the trace of the resolvent. As an example, consider the case that the eigenvalues are discrete with eigenfunctions $\phi_{n}$. The resolvent is given by $R_{\Sigma}(s;z,w)=\sum_{n}\frac{\phi_{n}(z)\phi_{n}(w)}{\lambda_{n}+s(s-1)}\,,$ (2.9) where the eigenfunctions satisfy $\int_{\Sigma}d^{2}z\sqrt{g}\phi_{n}(z)\phi_{m}(z)=\delta_{nm}$. One can check that the determinant defined as a product of eigenvalues is related to the resolvent trace through $\frac{1}{2s-1}\frac{d}{ds}\log(\det(\Delta+s(s-1)))={\rm tr}R_{\Sigma}(s)\equiv\int_{\Sigma}d^{2}z\sqrt{g}R_{\Sigma}(s;z,z)\,.$ (2.10) Conversely, if a closed form expression for the resolvent can be obtained then the determinant can be obtained by inverting the above formula. For simple surfaces the resolvent trace is easy to evaluate since the resolvent is known on the UHP, and we can perform the sum over the Fuchsian group defining the surface. Consider the double trumpet with geodesic throat of size $\ell$. The Fuchsian group is generated by a single hyperbolic element $\Gamma=\langle T_{\ell}\rangle$, producing the surface $\Sigma=\mathbb{H}/\langle T_{\ell}\rangle$. The resolvent trace is obtained after using equations (2.8) and (2.10) $\displaystyle\sum_{\begin{subarray}{c}m\in\mathbb{Z}_{\neq 0}\end{subarray}}\int_{\mathbb{H}/\langle T_{\ell}\rangle}d^{2}z\sqrt{g}R_{\mathbb{H}}\left(s;T_{\ell}^{m}\cdot z,z\right)=\frac{2\ell}{2s-1}\sum_{m=1}^{\infty}\frac{e^{-sm\ell}}{1-e^{-m\ell}}\,.$ (2.11) The integer $m$ counts the number of windings of the closed geodesic. The integral can be implemented within a fundamental domain given by $1<|z|<e^{\ell}$; as for details of integration, see appendix C. For general surfaces the summation over the group (2.8) will reduce the resolvent trace to simpler building blocks, such as the above calculation on the double trumpet. Let us explain how this works. The sum over the group can be broken up into a sum over conjugacy classes $[\gamma]$ of primitive elements as $\sum_{\Gamma}=\sum_{[\gamma]}\sum_{g\in\Gamma/C_{\langle\gamma\rangle}}$.111111A primitive element $\gamma$ is not a power of any other element in the group. The centralizer is defined as the set of elements that leave the group generated by $\gamma$ to be invariant $C_{\langle\gamma\rangle}=\\{g\langle\gamma\rangle g^{-1}=\langle\gamma\rangle|g\in\Gamma\\}$. The set of left cosets $\Gamma/C_{\langle\gamma\rangle}$ generates a list of distinct elements that we should conjugate $\gamma$ by sum over the full conjugacy class. We must take the centralizer of the group generated by $\gamma$, $C_{\langle\gamma\rangle}$ because we want to break up the sum into conjugacy classes of $[\gamma^{n}]$. It turns out that the centralizer for primitive $\gamma$ is simply given by $C_{\langle\gamma\rangle}=\langle\gamma\rangle$, which we use notationally throughout. To avoid over-counting, we must sum over an element from each coset $g\in\Gamma/C_{\langle\gamma\rangle}$ and conjugate $g\gamma g^{-1}$ to sum over the entire conjugacy class. The primitive elements will either be hyperbolic or elliptic which geometrically corresponds to either closed geodesics or conical defects respectively, and we must also include a sum over conjugacy classes of multiples of the primitive elements $\gamma^{n}$. The final result is that every group element can be reached by a conjugation $g^{-1}\gamma^{n}g$ of some power of a primitive element $\gamma$. Suppose we have a list of primitive elements denoted by $\Pi$, the summation over the group is then given by121212When $\gamma$ is elliptic the summation over powers $\gamma^{m}$ should be cutoff when we reach $\gamma^{m}=I$. $R_{\Sigma}(s;z,w)=R_{\mathbb{H}}\left(s;z,w\right)+\sum_{\gamma\in\Pi}\sum_{g\in\Gamma/\langle\gamma\rangle}\sum_{m\neq 0}^{\infty}R_{\mathbb{H}}\left(s;g^{-1}\gamma^{m}g\cdot z,w\right)\,,$ (2.12) where the first term is the identity contribution. When calculating the trace of the resolvent (2.10) the sum over $\Gamma/\langle T_{\ell}\rangle$ for hyperbolic elements $T_{\ell}$ will transform the domain of integration from the surface $\Sigma$ to the fundamental domain of a double trumpet with a closed geodesic $\ell$, of which we already have the answer in (2.11) $\displaystyle\sum_{\begin{subarray}{c}m\in\mathbb{Z}_{\neq 0}\end{subarray}}\sum_{g\in\Gamma/\langle T_{\ell}\rangle}\int_{\mathbb{H}/\Gamma}d^{2}z\sqrt{g}R_{\mathbb{H}}\left(s;z,T^{m}_{\ell}g\cdot z,g\cdot z\right)$ $\displaystyle=\sum_{\begin{subarray}{c}m\in\mathbb{Z}_{\neq 0}\end{subarray}}\int_{\mathbb{H}/\langle T_{\ell}\rangle}d^{2}z\sqrt{g}R_{\mathbb{H}}\left(s;z,T_{\ell}^{m}\cdot z\right)$ $\displaystyle=\frac{2\ell}{2s-1}\sum_{m=1}^{\infty}\frac{e^{-sm\ell}}{1-e^{-m\ell}}\,,$ (2.13) where the RHS is already given in (2.11). This gives a contribution to the determinant (2.10) for each closed geodesic $\ell$ on the surface $\Sigma$. The same logic applies to elliptic elements $T_{\theta}$ in the sum (2.12) except now the fundamental domain becomes a cone. The full determinant calculation is quite involved and carried out in appendix C. We quote the final result. Consider a hyperbolic surface $\Sigma$ with $k$ conical defects with opening angles $\theta_{i}=\frac{2\pi}{n_{i}}$ with integer $n_{i}\geq 2$. The surface can be compact or with asymptotic AdS boundaries. The determinant is given by $\det\left({\Delta}+s(s-1)\right)=\underbrace{Z_{\text{hyp.}}(s)}_{\text{geodesics}}\underbrace{Z_{\text{ell.}}(s)}_{\text{defects}}\left(\text{const}\right)$ (2.14) The most important contribution is given by $Z_{\text{hyp.}}$, which is the Selberg zeta-function on $\Sigma$. It comes from the summation over hyperbolic elements in (2.12). It is defined by a product over contributions from all closed geodesics on the surface. Define the set of lengths $\ell_{\gamma}$ of primitive closed geodesics to be $\mathcal{L}_{\Sigma}$,131313Excluding exceptional cases, geodesics are endowed with an orientation. For each geodesic $\gamma$ there is a mirror geodesic with opposite orientation and the same length. Both a geodesic and it’s mirror need to be independently included in the Selberg zeta-function. An example of a non-orientable geodesic is given by the two defect surface considered below. and the Selberg zeta-function for the surface $\Sigma$ is defined to be $Z_{\text{hyp.}}(s)=\prod_{\ell_{\gamma}\in\mathcal{L}_{\Sigma}}\prod_{m=0}^{\infty}\left(1-e^{-(s+m)\ell_{\gamma}}\right),$ (2.15) which can be derived from (2.11). In the above equation, the first product is over all primitive geodesics, while the second is over all integer $m$ windings of the geodesic. Primitive geodesics include self-intersecting ones as well as those that touch conical defects. We also mention that geodesics with opposite orientations are counted as distinct in (2.15). The second term comes from the summation over elliptic elements in (2.12). Since elliptic elements give conical defects we label this the “defect” contribution, and it is given by $Z_{\text{ell.}}(s)=\prod_{i=1}^{k}\prod_{r=0}^{n_{i}-1}\Gamma\left(\frac{s+r}{n_{i}}\right)^{\frac{2r+1-n_{i}}{n_{i}}}.$ (2.16) For each conical defect we get a highly non-analytic contribution in the opening angles $\theta_{i}=\frac{2\pi}{n_{i}}$. Unlike the Selberg zeta- function, the contribution due to elliptic elements does not seem to have an obvious geometric interpretation. We emphasize that every term in equation (2.14) depends on the choice of deficit angles in some way. Either implicitly through the types of geodesics that are included in the Selberg zeta-function, or explicitly as some function of the angles as above. The constant term in (2.14) is due to the identity element contribution and given in appendix C. ### 2.3 JT examples Let us briefly explain how these determinants and resolvents arise for JT gravity coupled to matter. Consider a massive scalar field minimally coupled to JT gravity. The path integral is given, up to boundary terms, by $Z_{\text{JT + matter}}=\int\mathcal{D}g\mathcal{D}\Phi\exp\left(\frac{1}{2}\int\sqrt{g}\phi\left(R+2\right)\right)\hskip 1.42271ptZ_{\text{matter}}[g],$ (2.17) where the matter partition function is $Z_{\text{matter}}[g]=\int\mathcal{D}\phi e^{-S[\phi,g]}=\frac{1}{\sqrt{\det\left(\Delta+m^{2}\right)}},\qquad S[\phi,g]=\frac{1}{2}\int_{\Sigma}\sqrt{g}\left(\phi\Delta\phi+m^{2}\phi^{2}\right).$ (2.18) where we define the mass through $m^{2}=s(s-1)$, with $s\geq 1$ the scaling dimension of the boundary operator dual to $\phi$. We see the matter partition function reduces down to a determinant. After integrating out the dilaton we localize onto the moduli space of constant negative curvature geometries,141414By appropriately deforming the JT gravity action we can also localize onto the moduli space of constant negative curvature surfaces with conical points which we will describe in detail in section 3. and so is given by the determinant calculated in section 2.2. Another interesting observable when matter is present is the two-point function/resolvent $\langle\phi(z)\phi(w)\rangle$ given by $R_{\Sigma}(s;z,w)$. On the upper half plane the scalar Laplacian is given by $ds^{2}=\frac{dx^{2}+dy^{2}}{y^{2}},\qquad\Delta=-y^{2}\left(\partial_{x}^{2}+\partial_{y}^{2}\right).$ (2.19) Defining the complex coordinates $z=x+iy$, the resolvent is given by $R_{\mathbb{H}}(s;z,w)=\frac{\Gamma(s)^{2}}{4\pi\Gamma(2s)}\sech^{2s}\left(\frac{\ell(z,w)}{2}\right)\hskip 1.42271pt{}_{2}F_{1}\left(s,s,2s;\sech^{2}\left(\frac{\ell(z,w)}{2}\right)\right).$ (2.20) where $\ell(z,w)$ is the geodesic distance between $z$ and $w$. In JT gravity we want to dress the two-point function to the boundary schwarzian. Since the geodesic distance goes to infinity near the boundary the two-point function simplifies to $R_{\mathbb{H}}(z,w)\mathrel{\mathop{=}\limits_{\ell\to\infty}}\frac{4^{s}\Gamma(s)^{2}}{4\pi\Gamma(2s)}e^{-s\ell(z,w)}\left(1+\mathcal{O}\left(e^{-\ell}\right)\right)\,,$ (2.21) and so the geodesic approximation becomes exact. On an arbitrary surface $\Sigma$ the two-point function is obtained by summing over the group in equation (2.8) so that we obtain $\displaystyle R_{\Sigma}(s;z,w)$ $\displaystyle=\frac{\Gamma(s)^{2}}{4\pi\Gamma(2s)}\sum_{T\in\Gamma}\sech^{2s}\left(\frac{\ell(z,T\cdot w)}{2}\right)\hskip 1.42271pt{}_{2}F_{1}\left(s,s,2s;\sech^{2}\left(\frac{\ell(z,T\cdot w)}{2}\right)\right)$ (2.22) $\displaystyle\hskip-6.0pt\mathrel{\mathop{=}\limits_{\ell\to\infty}}\frac{4^{s}\Gamma(s)^{2}}{4\pi\Gamma(2s)}\sum_{T\in\Gamma}e^{-s\ell(z,T\cdot w)}\left(1+\mathcal{O}\left(e^{-\ell}\right)\right).$ This implements a summation over all possible geodesics connecting the points $z,w$. To dress the operator insertions to the boundary fluctuations we parameterize the boundary by the curve $(x,y)=(F(\tau),\epsilon F^{\prime}(\tau))$ where $\tau$ is an affine time, $\epsilon$ is the boundary cutoff, and we have enforced the condition that the induced metric on the boundary is $\sqrt{h}=1/\epsilon$. We consider inserting operators at two boundary points labelled by times $\tau_{1},\tau_{2}$. The complex coordinates at these times are $z=F(\tau_{1})+i\epsilon F^{\prime}(\tau_{1})$ and $w=F(\tau_{2})+i\epsilon F^{\prime}(\tau_{2})$. The geodesic distance between these points is given by $\ell(z,w)=\log\left(\frac{1}{\epsilon^{2}}\frac{\left(F(\tau_{1})-F(\tau_{2})\right)^{2}}{F^{\prime}(\tau_{1})F^{\prime}(\tau_{2})}+\mathcal{O}\left(\frac{1}{\epsilon}\right)\right)\,.$ (2.23) We will define the two-point function dressed to the boundary fluctuations by $G(\tau_{1},\tau_{2})$, and subtract off the usual divergences to obtain $G(\tau_{1},\tau_{2})=\lim_{\epsilon\rightarrow 0}\frac{4\pi\Gamma(2s)}{4^{s}\Gamma(s)^{2}\epsilon^{2s}}R_{\mathbb{H}}(s;z,w)=\left(\frac{F^{\prime}(\tau_{1})F^{\prime}(\tau_{2})}{\left(F(\tau_{1})-F(\tau_{2})\right)^{2}}\right)^{s}\,.$ (2.24) This is the standard result for the matter two-point function dressed to the Schwarzian boundary at the level of the disk. In terms of the thermal circle reparameterization $f(\tau)$ defined by $F(\tau)\equiv\tan\frac{\pi}{\beta}f(\tau)$ the regularized resolvent is $G(\tau_{1},\tau_{2})=\left(\frac{f^{\prime}(\tau_{1})f^{\prime}(\tau_{2})}{\frac{\beta^{2}}{\pi^{2}}\sin^{2}\frac{\pi}{\beta}\left(f(\tau_{1})-f(\tau_{2})\right)}\right)^{s}\,.$ (2.25) On a more complicated surface obtained from a quotient by $\Gamma$ there are additional geodesics connecting the two boundary points. Since $w\cong T\cdot w$ for $T\in\Gamma$ the geodesic connecting $z$ to $T\cdot w$ on the disk becomes a distinct geodesic connecting $z$ to $w$ on the quotient geometry. We can find the geodesic distance between the point $z$ and $T\cdot w$ to be $\ell\left(z,T\cdot w\right)=\log\left(\frac{1}{\epsilon^{2}}\frac{\left(F(\tau_{1})-T\cdot F(\tau_{2})\right)^{2}}{F^{\prime}(\tau_{1})\left(T\cdot F(\tau_{2})\right)^{\prime}}+\mathcal{O}\left(\frac{1}{\epsilon}\right)\right),\qquad T\cdot F(\tau)=\frac{aF(\tau)+b}{cF(\tau)+d},$ (2.26) The boundary dressed two-point function on a quotient surface can then be immediately obtained from equation (2.22) $G_{\Sigma}(\tau_{1},\tau_{2})=\lim_{\epsilon\rightarrow 0}\frac{4\pi\Gamma(2s)}{4^{s}\Gamma(s)^{2}\epsilon^{2s}}\sum_{T\in\Gamma}e^{-s\ell(z,T\cdot w)}=\sum_{T\in\Gamma}\left(\frac{F^{\prime}(\tau_{1})(T\cdot F(\tau_{2}))^{\prime}}{\left(F(\tau_{1})-T\cdot F(\tau_{2})\right)^{2}}\right)^{s}.$ (2.27) This formula includes all possible geodesics on the surface $\Sigma$ connecting the specified boundary points, and different group elements $T$ lead to geodesics with different windings/self-intersections on $\Sigma$. This includes geodesics that self-intersect any number of times. However, we are still left with the challenge of integrating over the Schwarzian and performing the sum over the group elements. We now explain some simple examples. #### 2.3.1 Double trumpet The metric for the double trumpet can be written as $ds^{2}=dr^{2}+\cosh^{2}\left(r\right)d\theta^{2},\qquad\theta\cong\theta+b.$ (2.28) The two asymptotic boundaries are at $r\to\pm\infty$, and the only primitive geodesic of length $b$ is at $r=0$. The double trumpet is obtained by taking the quotient of the upper half-plane $\mathbb{H}/\Gamma$ by the group $\Gamma=\langle T_{b}\rangle$ generated by the element $T_{b}=\begin{pmatrix}\exp\left(\frac{b}{2}\right)&0\\\ 0&\exp\left(-\frac{b}{2}\right)\end{pmatrix},$ (2.29) which identify points in the UHP plane by $z\cong e^{b}z$. In the fundamental domain the geodesic throat of length $b$ goes along the imaginary axis from $z=i$ to $z=ie^{b}$. The group elements are hyperbolic and take the form $T_{b}^{m}$ for $m\in\mathbb{Z}$, and correspond to the geodesics that winds $m$ times around the throat. The resolvent is given by $R_{\mathbb{H}/\langle T_{b}\rangle}(s;z,w)=\sum_{m}R_{\mathbb{H}}(s;T_{b}\cdot z,w)\,.$ (2.30) The resolvent trace and determinant have been discussed in section 2.2 and we simply repeat the result ${\rm tr}R_{\mathbb{H}/\langle T_{b}\rangle}=\frac{2b}{2s-1}\sum_{m=1}^{\infty}\frac{e^{-smb}}{1-e^{-mb}},\qquad\det\left({\Delta}+s(s-1)\right)_{\text{DT}}=\prod_{m=0}^{\infty}\left(1-e^{-\left(s+m\right)b}\right)^{2}\,.$ (2.31) where the quantity in parenthesis is squared since there are two orientations for the primitive geodesic, and the determinant should be understood to be up to multiplicative constants. ##### One sided two-point function. We first consider the correlator where we insert two matter operators $\langle\phi_{L}(\tau_{1})\phi_{R}(\tau_{2})\rangle$ onto the same asymptotic boundary and dress them to the boundary fluctuations. We take the operator to be located at $z=F_{R}(\tau_{1})+i\epsilon F_{R}^{\prime}(\tau_{1})$ and $w=F_{R}(\tau_{2})+i\epsilon F_{R}^{\prime}(\tau_{2})$. The quotient $z\cong T\cdot z$ enforces a periodicity constraint on the boundary151515This is different from the $\tanh$ one in the JT review, but equivalent The Schwarzian derivative itself is $\operatorname{PSL}(2,\mathbb{R})$ invariant: if $F=\frac{aG+b}{cG+d}$, then $\\{F,t\\}=\\{G,t\\}$. $F_{R}(\tau+\beta)=T_{b}\cdot F_{R}(\tau),\qquad T_{b}^{m}\cdot F(\tau)=e^{mb}F(\tau).$ (2.32) We can introduce the parameterization $F_{R}(\tau)=\exp\left(\frac{b}{\beta}f_{R}(\tau)\right)$ with $f_{R}(\tau+\beta)=f_{R}(\tau)+\beta_{R}$ which respects this identification. We can immediately apply (2.27) to find the two-point function $G^{\text{RR}}_{\text{DT}}(\tau_{1},\tau_{2})=\sum_{m=-\infty}^{\infty}\left(\frac{F_{R}^{\prime}(\tau_{1})(T_{b}^{m}\cdot F_{R}(\tau_{2}))^{\prime}}{(F_{R}(\tau_{1})-T_{b}^{m}\cdot F_{R}(\tau_{2}))^{2}}\right)^{s}=\sum_{m=-\infty}^{\infty}\left(\frac{f_{R}^{\prime}(\tau_{1})f_{R}^{\prime}(\tau_{2})}{\frac{4\beta_{R}^{2}}{b^{2}}\sinh^{2}\left[\frac{b}{2\beta_{R}}\left(f_{R}(\tau_{1})-f_{R}(\tau_{2}+m\beta_{R})\right)\right]}\right)^{s}\,.$ (2.33) We have used that the summation over the group $\Gamma$ is the same as a summation over powers of $T_{b}^{m}$. The sum over $m$ is over windings of the geodesic around the double trumpet as it connects the two points. Positive and negative integers indicate which direction the geodesic winds around the trumpet. The largest contribution is given by the shortest geodesic which doesn’t wind and is given by $m=0$. This reproduces the one sided two-point function on the double trumpet computed in [41]. Figure 2: Two-point function on the double trumpet. We have geodesics connecting the two operator insertions that wind an integer $m$ times around the trumpet. ##### Two sided two-point function. The correlator with matter insertions on opposite boundaries $\langle\phi_{L}(\tau_{1})\phi_{R}(\tau_{2})\rangle$ also immediately follows from the previous result. The operators are now dressed to independent boundary wiggles $z=F_{L}(\tau_{1})+i\epsilon F_{L}^{\prime}(\tau_{1})$ and $w=F_{R}(\tau_{2})+i\epsilon F_{R}^{\prime}(\tau_{2})$. The quotient enforces the periodicity constraint on both boundaries $F_{L,R}(\tau+\beta_{L,R})=T_{b}\cdot F_{L,R}(\tau)$, and we can define $F_{L,R}(\tau)=\mp\exp\left(\frac{b}{\beta_{L,R}}f_{L,R}(\tau)\right)$ which satisfies the constraint as long as $f_{L,R}(\tau+\beta_{L,R})=f_{L,R}(\tau)+\beta_{L,R}$ where the two boundaries can have independent temperatures $\beta_{L,R}$. The relative minus sign is because the left boundary is at $\real(z)<0$, so we must have that $F_{L}<0$ whereas $F_{R}>0$. Performing the sum over the group (2.27) as in the one sided case we obtain $G^{\text{LR}}_{\text{DT}}(\tau_{1},\tau_{2})=\sum_{m=-\infty}^{\infty}\left(\frac{F_{L}^{\prime}(\tau_{1})(T_{b}^{m}\cdot F_{R}(\tau_{2}))^{\prime}}{(F_{L}(\tau_{1})-T_{b}^{m}\cdot F_{R}(\tau_{2}))^{2}}\right)^{s}=\sum_{m=-\infty}^{\infty}\left(\frac{f_{L}^{\prime}(\tau_{1})f_{R}^{\prime}(\tau_{2})}{\frac{4\beta_{L}\beta_{R}}{b^{2}}\cosh^{2}\left[\frac{b}{2}\left(\frac{f_{L}(\tau_{1})}{\beta_{L}}-\frac{f_{R}(\tau_{2}+m\beta_{R})}{\beta_{R}}\right)\right]}\right)^{s}\,.$ (2.34) For the two-sided correlator the geodesic can also wind $m$ times around the trumpet. The dominant contribution is given by the geodesic with smallest length, which is given by $m=0$ winding. #### 2.3.2 Conical defect Figure 3: Two-point function on the conical defect geometry. The geodesics connecting the two operators can wind up to $n-1$ times around the defect. We now consider the geometry with a single conical defect of opening angle $\theta=\frac{2\pi}{n}$ by taking the quotient by the group $\Gamma=\langle T_{\theta}\rangle$. The generator is given by $T_{\theta}=\left(\begin{array}[]{ll}\cos\frac{\pi}{n}&-\sin\frac{\pi}{n}\\\ \sin\frac{\pi}{n}&~{}~{}\cos\frac{\pi}{n}\end{array}\right)\,,$ (2.35) and the conical defect is located at the fixed point $z=i$. This group has $n$ elements corresponding to powers of $T_{\theta}^{m}$ for $m=0,\ldots,n-1$. The determinant on this geometry is not very interesting since the cone has no closed geodesics. Instead we will consider the two-point function dressed to the boundary schwarzian. As in the case of the disk, we insert the two operators at $z=F(\tau_{1})+i\epsilon F^{\prime}(\tau_{1})$ and $w=F(\tau_{2})+i\epsilon F^{\prime}(\tau_{2})$. The identification $z\cong T_{\theta}\cdot z$ enforces the condition $F(\tau+\beta)=T_{\theta}\cdot F(\tau).$ (2.36) We can also use the parameterization $F(\tau)=\tan\frac{\theta}{2\beta}f(\tau)$, with $f(\tau+\beta)=f(\tau)+\beta$. Applying the two-point function formula (2.27) we have $G_{\text{Defect}}(\tau_{1},\tau_{2})=\sum_{m=0}^{n-1}\left(\frac{F^{\prime}(\tau_{1})(T_{\theta}^{m}\cdot F(\tau_{2}))^{\prime}}{\left(F(\tau_{1})-T_{\theta}^{m}\cdot F(\tau_{2})\right)^{2}}\right)^{s}=\sum_{m=0}^{n-1}\left(\frac{f^{\prime}(\tau_{1})f^{\prime}(\tau_{2})}{\frac{4\beta^{2}}{\theta^{2}}\sin^{2}\frac{\theta}{2\beta}\left[f(\tau_{1}+m\beta)-f(\tau_{2})\right]}\right)^{s}\,,$ (2.37) There are $n$ geodesics connecting the two operator insertions on the boundary. These geodesics wind around the defect up to $n-1$ times which is the summation over $m$. Note that for $n=1$ the opening angle becomes $\theta=2\pi$ and we get the answer for the disk. The above is not the complete answer, since we must still integrate over the boundary fluctuations. If we consider the case of a massless field $s=1$ we can perform the summation over winding geodesics to find $G^{s=1}_{\text{Defect}}(\tau_{1},\tau_{2})=\frac{f^{\prime}(\tau_{1})f^{\prime}(\tau_{2})}{\frac{\beta^{2}}{\pi^{2}}\sin^{2}\frac{\pi}{\beta}\left[f(\tau_{1})-f(\tau_{2})\right]},$ (2.38) which is precisely the two-point function for a massless field on the disk (2.25). This is the expected answer for a conformal scalar since the correlator on the defect geometry is related to the correlator on the disk through a conformal rescaling. In this case the rescaling is trivial at the boundary so the answers agree. We see that including all of the winding geodesics is crucial to recover the correct properties of the matter fields.161616We still have to integrate over the schwarzian mode. On the defect geometry the $n=2$ boundary modes are no longer zero modes and must be included in the integral. Even though we have reduced the two point function to the standard disk answer, we must still include these boundary modes when computing the two-point function. #### 2.3.3 Two conical defects To get a disk with two conical defects we need to generate a Fuchsian group with two elliptic generators $\Gamma=\langle T_{\theta_{1}},T_{\theta_{1}}\rangle$. We choose the generators to be Figure 4: The disk with two defects can be represented in the UHP by the pictured fundamental domain, with the two defects represented by “x”. The single primitive closed geodesic is in blue, and travels between the defects. We also have any number of windings of this geodesic. $T_{\theta_{1}}=\begin{pmatrix}\cos\frac{\theta_{1}}{2}&e^{-\ell/2}\sin\frac{\theta_{1}}{2}\\\ e^{\ell/2}\sin\frac{\theta_{1}}{2}&\cos\frac{\theta_{1}}{2}\end{pmatrix}\,,\quad T_{\theta_{2}}=\begin{pmatrix}\cos\frac{\theta_{2}}{2}&e^{\ell/2}\sin\frac{\theta_{2}}{2}\\\ e^{-\ell/2}\sin\frac{\theta_{2}}{2}&\cos\frac{\theta_{2}}{2}\end{pmatrix}\,.$ (2.39) In the UHP the fixed point of $T_{\theta_{1}}$ is at $z=ie^{-\ell/2}$ while the fixed point of $T_{\theta_{2}}$ is at $z=ie^{\ell/2}$, which are the locations of the conical defects with opening angles $\theta_{1},\theta_{2}$ respectively. The simplest example, which captures many features of multi defect surfaces, is to consider the case where both of the angles take the value $\theta_{1,2}=\pi$. In this case the simplest group products take the form $T_{\theta_{1}}^{2}=T_{\theta_{2}}^{2}=\mathbb{I},\qquad T_{2\ell}\equiv T_{\theta_{2}}T_{\theta_{1}}=\begin{pmatrix}e^{\ell}&0\\\ 0&e^{-\ell}\end{pmatrix},$ (2.40) where we see that a product of two elliptic elements becomes a hyperbolic element $T_{2\ell}$. There are thus three primitive elements in the group: two elliptic $T_{\theta_{1}},T_{\theta_{2}}$ and one hyperbolic $T_{2\ell}$. From the general formula for the determinant (2.14) we immediately get171717The careful reader may wonder why the Selberg zeta function does not get a square like the double trumpet case. The reason is that the closed geodesic in this case is not orientable. The closed geodesic with total length $2\ell$ is actually obtained by gluing a copy of a geodesic segement with length $\ell$ back to itself in the opposite direction. $\det(\Delta+s(s-1))=\underbrace{\prod_{m=0}^{\infty}\left(1-e^{-2\ell(s+m)}\right)}_{\text{closed geodesic}}\underbrace{\frac{\Gamma\left(s+\frac{1}{2}\right)}{\Gamma\left(s\right)}.}_{\text{defect contribution}}$ (2.41) The Selberg zeta function contribution is given by the single geodesic traversing between the two defects, which has length $2\ell$. This originates from the single primitive hyperbolic element $T_{2\ell}$. We have also included the contribution coming from the elliptic elements in the resolvent, which correspond to the two defects. One interesting feature of this surface is that the hyperbolic element is generated by products of elliptic elements, and that the resulting closed geodesic touches the conical defects. To calculate the two-point function (2.27) we must sum over the group. By inspection we can find that all elements of the group take the form $T_{2\ell}^{m}=\begin{pmatrix}e^{m\ell}&0\\\ 0&e^{-m\ell}\end{pmatrix}\,,\quad T_{\theta_{1}}T_{2\ell}^{m}=\begin{pmatrix}0&e^{\left(m+\frac{1}{2}\right)\ell}\\\ e^{-\left(m+\frac{1}{2}\right)\ell}&0\end{pmatrix}\,,$ (2.42) with $m$ an integer. From the fundamental domain of the schwarzian we see we must identify $F(\tau+\beta)=e^{2\ell}F(\tau)$, which can be expanded in terms of the identification $F(\tau)=\exp\left(\frac{2\ell}{\beta}f(\tau)\right)$. We can immediately write the answer for the resolvent to be $\displaystyle G_{\text{Defects}}(\tau_{1},\tau_{2})=\sum_{m=-\infty}^{\infty}\left(\frac{F^{\prime}(\tau_{1})(T_{2\ell}^{m}\cdot F(\tau_{2}))^{\prime}}{\left(F(\tau_{1})-T_{\theta}^{m}\cdot F(\tau_{2})\right)^{2}}\right)^{s}+\left(\frac{F^{\prime}(\tau_{1})(T_{\theta_{1}}T_{2\ell}^{m}\cdot F(\tau_{2}))^{\prime}}{\left(F(\tau_{1})-T_{\theta}^{m}\cdot F(\tau_{2})\right)^{2}}\right)^{s}$ $\displaystyle=\sum_{m=-\infty}^{\infty}\left(\frac{f^{\prime}(\tau_{1})f^{\prime}(\tau_{2})}{\frac{\beta^{2}}{b^{2}}\sinh^{2}\left[\frac{b}{\beta}\left(f(\tau_{1})-f(\tau_{2}+m\beta)\right)\right]}\right)^{s}+\left(\frac{f^{\prime}(\tau_{1})f^{\prime}(\tau_{2})}{\frac{\beta^{2}}{b^{2}}\sinh^{2}\left[\frac{b}{\beta}\left(f(\tau_{1})+f(\tau_{2}+m\beta)-\frac{\beta}{2}\right)\right]}\right)^{s}\,.$ (2.43) The first term corresponds to geodesics that wind around the throat, whereas the second term are geodesics that wind around the defects. The above example was for the special case where both defects have deficit angle $\pi$, but the construction immediately extends to more general opening angles smaller than $\pi$. We briefly explain the main differences. * • To ensure the quotient surface has one asymptotic boundary and two defects, the parameter $\ell$ must satisfy $e^{\ell}>\frac{\sin\theta_{1}\left(1+\cos\theta_{2}\right)}{\sin\theta_{2}\left(1-\cos\theta_{1}\right)}\,.$ (2.44) Geometrically, it means that two sharp defects can not be arbitrarily close to each other. * • There are infinitely many closed geodesics on the surface, generated by the primitive “words” of $T_{\theta_{1,2}}$ such as $T_{\theta_{1}}^{2},T_{\theta_{1}}T_{\theta_{2}},T_{\theta_{2}}^{2}T_{\theta_{1}}^{7},\ldots$ as long as the “word” is hyperbolic. * • Such geodesics generically self-intersect. For example, $T_{\theta_{1}}T_{\theta_{2}}^{-1}$ is a geodesic winding around both defects once with a self-intersection. #### 2.3.4 The handle disk We now consider the handle disk which has a non-trivial topology. To build the surface we can cut the handle disk along two geodesics, and flatten it onto the UHP with the identification in figure 5. Therefore, the Fuchsian group is generated by two hyperbolic elements that identify the associated geodesics: the first semicircle with the third and the second semicircle with the fourth.181818Fuchsian groups generated by identifying pairs of semicircles are called the Schottky groups. The generators will be denoted by $T_{1}$ and $T_{2}$,191919The form of the hyperbolic generators is given as follows. A hyperbolic element that identifies semi-circles centered along the real axis at points $c_{1,2}$ with radii $r_{1,2}$ is $T(c_{1},c_{2};r_{1},r_{2})=\left(r_{1}r_{2}(c_{12}-r_{12})(c_{12}+r_{12})\right)^{-\frac{1}{2}}\begin{pmatrix}-c_{2}c_{12}+r_{2}r_{12}&(c_{1}c_{2}+r_{1}r_{2})(c_{12}-r_{12})\\\ -c_{12}&c_{1}c_{12}-r_{1}r_{12}\end{pmatrix}$ where $c_{12}=|c_{1}-c_{2}|$ and $r_{12}=|r_{1}-r_{2}|$. and the fundamental domain is to remove the four half disks from $\mathbb{H}$. Figure 5: The handle disk can be constructed by identifying the orange and blue geodesic semi-circles in the UHP. The fundamental domain is the exterior of all the semi-circles. The two identifications are the generators of the Fuchsian group that generate the surface. The red and purple curves are closed geodesics on the surface. In the right figure we omit the purple closed geodesic because it is broken into four segments with the pictured fundamental domain. As in the case of the two-defect surface any hyperbolic word $T_{1}^{k_{1}}T_{2}^{k_{2}}T_{1}^{k_{3}}\ldots\,$, on the condition that it is primitive, corresponds to a primitive closed geodesic on the handle disk. A generic hyperbolic word corresponds to an integer winding of a primitive geodesic. As an example, the dark red geodesic in figure 5 is associated to $T_{2}$ while the purple one is the closed geodesic separating the torus with the geodesic boundary, generated by $T_{1}T_{2}T_{1}^{-1}T_{2}^{-1}$. More complicated closed geodesics typically have self-intersections. Each closed geodesics (with an associated hyperbolic word) contributes to the determinant, and so it is difficult to write down a simple expression for the full determinant. For the two-point function we must consider boundary anchored geodesics such as the orange and the blue geodesics in figure 5. It is convenient to use a different fundamental domain for the handle disk where the asymptotic boundary is in one connected segment of the UHP. To achieve this, one can first diagonalize $T_{1}T_{2}T_{1}^{-1}T_{2}^{-1}$, which is associated to the geodesic separating the asymptotic boundary from the handle, to be $-\text{diag}(e^{\ell/2},e^{-\ell/2})$. This modifies the generators of the group to be $\tilde{T}_{1},\tilde{T}_{2}$ with generated group $\tilde{\Gamma}$.202020To get $\tilde{T}_{1,2}$ and $\tilde{\Gamma}$, one should conjugate every element in $\Gamma$ with a proper SL(2,$\mathbb{R}$) matrix $V$ so that $V\cdot T_{1}T_{2}T_{1}^{-1}T_{2}^{-1}\cdot V^{-1}=-\text{diag}(e^{\ell/2},e^{-\ell/2})$. Then $\tilde{\Gamma}$ is $V\cdot\Gamma\cdot V^{-1}$, and the generators are $\tilde{T}_{i}=V\cdot T_{i}V^{-1}$ for which an explicit although complicated form can be obtained. The resulting fundamental domain is pictured in figure 6. The fundamental domain looks very similar to the double trumpet, and we can parameterize the Schwarzian boundary by $z=F(\tau)+i\epsilon F^{\prime}(\tau)$ with $F(\tau)=\exp\left(\frac{\ell}{\beta}f(\tau)\right)$. The two point function is given by $G(\tau_{1},\tau_{2})=\sum_{\gamma\in\tilde{\Gamma}}\left(\frac{F^{\prime}(\tau_{1})\left(\gamma\cdot F(\tau_{2})\right)^{\prime}}{\left(F(\tau_{1})-\gamma\cdot F(\tau_{2})\right)^{2}}\right)^{s}\,.$ (2.45) Note that the parameterization is the same as the one-sided two-point function on the double trumpet (2.33), but the handle disk case is much more complicated due to its group $\tilde{\Gamma}$. There are distinct classes of geodesics on the handle disk, and we now explain how the sum over geodesics can be categorized into equivalence classes using large diffeomorphisms. Figure 6: An alternative representation of the handle disk fundamental domain. All the semi-circles are geodesics, and fundamental domain is the bounded region inside all the semi-circles. The identifications are red/pink, blue/light blue/gray, green/green. The purple geodesic winds around the handle. In this representation the asymptotic boundary is in one connected segment. The two green semi-circles are mapped to each other by the group element $\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}^{-1}\tilde{T}_{2}^{-1}.$ The moduli space of the handle disk is composed of schwarzian fluctuation and the moduli of the torus with geodesic boundary denoted by $\mathcal{M}_{1,1}$. The moduli space is obtained from Teichmuller space $\mathcal{T}_{1,1}$ after modding out by the action of large diffs, also known as the mapping class group (MCG). The MCG is generated by three elements $\\{\sigma,P,U\\}$ and has a simple action on the generators of the Fuchsian group [42] $\begin{array}[]{ll}\sigma(\tilde{T}_{1})=\tilde{T}_{1}^{-1},&P(\tilde{T}_{1})=\tilde{T}_{2},\\\ \sigma(\tilde{T}_{2})=\tilde{T}_{2},&P(\tilde{T}_{2})=\tilde{T}_{1},\end{array}\ \begin{aligned} &\tilde{U}(T_{1})=\tilde{T}_{1}\tilde{T}_{2},\\\ &U(\tilde{T}_{2})=\tilde{T}_{2}\,.\end{aligned}$ (2.46) The MCG acts on a group element built from products of the generators by acting on each generator independently. This action can be understood as a map between cycles of the surface. Since each cycle has an associated closed geodesic, we can think of this action as generating a map between closed geodesics on the surface. When considering boundary anchored geodesics, this becomes a map between boundary anchored geodesics. As an example, it can be checked that the MCG preserves the group element $\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}^{-1}\tilde{T}_{2}^{-1}$ up to conjugation,212121For example, $\sigma\left(\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}^{-1}\tilde{T}_{2}^{-1}\right)$ is $\tilde{T}_{1}^{-1}\tilde{T}_{2}\tilde{T}_{1}\tilde{T}_{2}^{-1}$, which is conjugate to $\left(\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}^{-1}\tilde{T}_{2}^{-1}\right)^{-1}$ by $\tilde{T}_{1}$. so the geodesic splitting the handle from the asymptotic boundary is fixed. We can consider a group element $\gamma\in\tilde{\Gamma}$ corresponding to a particular boundary anchored geodesic and act on it with the MCG to generate the equivalence class $\\{\gamma\\}_{\rm MCG}$.222222The geodesic corresponding to $\gamma$ is the geodesic connecting the boundary points $F(\tau_{1})$ and $\gamma\cdot F(\tau_{2})$ if we unwrap the handle disk on the UHP. As an example the element $\tilde{T}_{1}$ corresponds to the purple boundary anchored geodesics that winds the handle in figure 6. We can generate all other geodesics that wind the handle and do not self-intersect by acting with the MCG $\text{non self-intersecting geodesics: }\\{\tilde{T}_{1}\\}_{\rm MCG}=\\{\tilde{T}_{1},\tilde{T}_{2},\tilde{T}_{1}\tilde{T}_{2},\tilde{T}_{1}^{2}\tilde{T}_{2},\tilde{T}_{1}^{3}\tilde{T}_{2},\tilde{T}_{1}^{2}\tilde{T}_{2}^{3},\ldots\\}\,.$ (2.47) The summation over the elements $\\{\tilde{T}_{1}\\}_{\rm MCG}$ in the two point function (2.33) implements a sum over all non-self intersecting geodesics that wind the handle. These contributions were resumed in [25] and proved to give the ramp behavior for the late time two-point function. This method can also be used to partially classify self-intersecting geodesics that wind around the handle. A complete closed form classification appears to be highly non-trivial, but for one self-intersection we can find explicit examples of simple equivalence classes. For boundary anchored once self- intersecting geodesics that wind the handle, two equivalence classes are given by starting with base elements $\tilde{T}_{1}^{2}\tilde{T}_{2}^{2}$, $\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}^{-1}\tilde{T}_{2}^{-1}$ and acting with the MCG one self-intersection: $\displaystyle\\{\tilde{T}_{1}^{2}\tilde{T}_{2}^{2}\\}_{\rm MCG}=\\{\tilde{T}_{1}^{2}\tilde{T}_{2}^{2},\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}\tilde{T}_{2}^{-1},\tilde{T}_{1}^{3}\tilde{T}_{2}\tilde{T}_{1}\tilde{T}_{2},\ldots\\},$ (2.48) $\displaystyle\\{\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}^{-1}\tilde{T}_{2}^{-1}\\}_{\rm MCG}=\\{\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}^{-1}\tilde{T}_{2}^{-1},\tilde{T}_{1}\tilde{T}_{2}\tilde{T}_{1}^{3}\tilde{T}_{2}\tilde{T}_{1}^{-1}\tilde{T}_{2}^{-1},\ldots\\}\,.$ Note that the above is not a complete classification of all once self- intersecting geodesics, but we believe the first equivalence class will always contain the shortest self-intersecting geodesics on the surface.232323The argument is that the length of self-intersecting geodesics increases with additional windings, and the number of windings is controlled by the word- length of the generator. We can write down other equivalence classes systematically using the group theory description.242424Note that geodesics with more self-intersections are generated by words with at least six generators. The greater length of the words implies a greater length of the corresponding geodesic. It would be interesting to perform the integral over the boundary fluctuations in (2.45) for the self-intersecting geodesics falling into the equivalence classes (2.48) to confirm that they do not contribute to the late time two-point function [25]. ## 3 JT gravity path integral ### 3.1 Gauge fixing the path integral We now explain how to gauge fix the gravity path integral for JT gravity. For the case of compact surfaces see [7], while for non-compact surfaces also see [13, 14, 15, 16]. The integral is defined by $Z=\int\frac{\mathcal{D}g\mathcal{D}\Phi}{\text{V}(\text{Diff})}e^{-I_{\text{J}T}[g,\Phi]},$ (3.1) where we divide out by the volume of diffeomorphisms. The JT gravity action is given by $I_{\text{J}T}[g,\Phi]=-S_{0}\chi(\Sigma)-\bigg{[}\frac{1}{2}\int_{\Sigma}\sqrt{g}\Phi\left(R+2\right)+\underbrace{\int_{\partial\Sigma}\sqrt{h}\Phi\left(K-1\right)}_{-I_{\text{bdy}}}\bigg{]},$ (3.2) where the first term is topological and $\chi(\Sigma)=2-2g-b$ for surfaces with $g$ handles and $b$ boundaries. When we include asymptotic boundaries we choose boundary conditions to fix the induced metric $\sqrt{h}=1/\epsilon$ and the proper length of the boundary to be $\beta/\epsilon$, and we fix the dilaton to asymptotically approach $\phi_{b}=\gamma/\epsilon$. To integrate over metrics we must first specify a measure on the moduli space of all metrics, that is we must specify a metric $\langle.\hskip 1.70709pt,.\rangle$ on the tangent space of moduli space. The tangent space is naturally the space of metric deformations, and the standard metric for the space of these deformations is the ultra-local measure[43] $\langle\delta g_{ab},\delta g_{cd}\rangle=\mathcal{N}\int_{\Sigma}\sqrt{g}g^{ac}g^{bd}\delta g_{ab}\delta g_{cd},$ (3.3) which can be extended to other tensor deformations in an obvious way. The measure is defined up to a normalization constant $\mathcal{N}$ that we will fix later. Metric deformation decompose into three orthogonal parts $\delta g_{ab}=\underbrace{\omega g_{ab}}_{\text{weyl}}\hskip 2.84544pt\oplus\hskip 2.84544pt\underbrace{\operatorname{range}P_{1}}_{\text{small diff}}\hskip 2.84544pt\oplus\hskip 2.84544pt\underbrace{\operatorname{ker}P_{1}^{{\dagger}}}_{\text{moduli}},$ (3.4) where $(P_{1}V)_{ab}=\nabla_{a}V_{b}+\nabla_{b}V_{a}-(\nabla_{c}V^{c})g_{ab}$, and $(P_{1}^{\dagger}\delta g)_{a}=-2\nabla^{b}\delta g_{ab}$. A general infinitesimal metric deformation is a combination of a Weyl transformation, a small diffeomorphism, and a deformation of the moduli. Since the metric deformations are orthogonal (3.4) the path integral measure breaks up into integrals over these deformations. We will first consider the case where we are integrating over all metrics on a compact surface of genus $g\geq 2$. To perform the integral we gauge fix to conformal gauge and write the metric as $g=f^{*}\left(e^{2\omega}\hat{g}\right)$ with $f$ a small diffeomorphism, $\omega$ a Weyl factor, and $\hat{g}$ a constant negative curvature metric $\hat{R}=-2$. The measure works out to be[31, 29, 30] $Z_{g}=\int_{\mathcal{M}_{g}}\underbrace{d\left(\text{Weil-Pet.}\right)\hskip 2.84544pt(\det\hat{P}_{1}^{\dagger}\hat{P}_{1})^{1/2}\mathcal{D}\omega}_{\mathcal{D}g/\text{Vol}}\mathcal{D}\Phi e^{-I_{\text{J}T}[e^{2\omega}\hat{g},\Phi]}e^{-26S_{L}[\omega,\hat{g}]}.$ (3.5) The integral is carried out over the moduli space of constant negative curvature surfaces $\mathcal{M}_{g}$ of genus $g$.252525When integrating over small diffeomorphisms we pick up a factor of the volume of small diffs. The ratio of small diffs with all diffs gives the set of large diffs, which reduces the integral $\frac{\text{Vol$($Diff${}_{0})$}}{\text{Vol(Diff)}}\int_{\mathcal{T}_{g}}=\int_{\mathcal{M}_{g}}$ to the moduli space of $R=-2$ surfaces. Here $\mathcal{T}_{g}$ is the set of metrics mod Weyl $[e^{2\omega}g]$ before quotienting by diffeomorphisms, also known as Teichmuller space. In the above $(\det\hat{P}_{1}^{\dagger}\hat{P}_{1})^{1/2}$ appears from the ghost path integral when gauge fixing to conformal gauge, and the notation $\hat{P_{1}}$ implies the operator is defined with respect to $\hat{g}$. We have an integral over the Weyl factor $\mathcal{D}\omega$, and the Liouville action $S_{L}$ appears due to the conformal anomaly arising from defining the integration measures with respect to $\hat{g}$.262626Every integral in (3.5), such as $\mathcal{D}_{\hat{g}}\Phi$, is defined with the appropriate ultra-local measure (3.3) with metric $\hat{g}$. The Liouville action will give no contribution after we perform the $\mathcal{D}\omega$ integral so we discard it from now on. When restricted to constant negative curvature metrics $\hat{g}$, the measure (3.3) is by definition the Weil-Petersson measure [44], which is the origin of $d\left(\text{Weil-Pet.}\right)$. We can write the part of the JT gravity action (3.2) that contains the dilaton as $I[e^{2\omega}\hat{g},\Phi]=-\frac{1}{2}\int_{\Sigma}\sqrt{\hat{g}}\Phi\left(\hat{R}-2\hat{\nabla}^{2}\omega+2e^{2\omega}\right).$ (3.6) We can perform the integral over the dilaton by rotating the integration contour to be along the imaginary axis, giving a delta function constraint that localizes onto constant negative curvature geometries $\int\mathcal{D}\omega\int_{i\mathbb{R}}\mathcal{D}\Phi e^{-I_{\text{J}T}[e^{2\omega}\hat{g},\Phi]}=\int\mathcal{D}\omega\hskip 2.84544pt\delta\left(\hat{R}-2\hat{\nabla}^{2}\omega+2e^{2\omega}\right)=\frac{1}{\det(-2\hat{\nabla}^{2}+4)},$ (3.7) where in the last equality we localize onto $\omega=0$. A straightforward calculation shows that the gauge fixing differential operator takes the form $\hat{P}_{1}^{\dagger}\hat{P}_{1}=2\left(-\hat{\nabla}_{1}^{2}+1\right)$ on a surface with constant negative curvature, with $\hat{\nabla}^{2}_{1}$ the vector laplacian. Putting everything together we find that the JT gravity path integral (3.5) is $\displaystyle Z_{g}$ $\displaystyle=(\text{const})^{\chi}\int_{\mathcal{M}_{g}}d\left(\text{Weil- Pet.}\right)\frac{\det\left(-\hat{\nabla}_{1}^{2}+1\right)^{1/2}}{\det(-\hat{\nabla}^{2}+2)}$ (3.8) $\displaystyle=(\text{const})^{\chi}\int_{\mathcal{M}_{g}}d\left(\text{Weil- Pet.}\right).$ (3.9) In the above we have discarded some multiplicative constants from the determinants, since in zeta function regularization multiplying all the eigenvalues by a constant shifts the determinant by $(\text{const})^{\chi}$.272727All of these constants can be absorbed into a redefinition of $S_{0}$ in the topological term of the action (3.2). Furthermore we also used our result for the vector laplacian determinant $\frac{\det\left(-\hat{\nabla}_{1}^{2}+1\right)^{1/2}}{\det(-\hat{\nabla}^{2}+2)}=2^{\frac{1}{2}\chi},$ (3.10) given in equation (C.10). The above conclusion immediately generalizes beyond compact surfaces. Consider the moduli space of genus $g$ hyperbolic surface with $n$ asymptotic boundaries of regularized lengths $\vec{\beta}=\left(\beta_{i},\ldots,\beta_{n}\right)$ and $k$ conical defects with opening angles specified by $2\pi\alpha_{i}$. We take $\vec{\alpha}=\left(\alpha_{1},\ldots,\alpha_{k}\right)$ with $\alpha_{i}=n_{i}^{-1}$ and integer $n_{i}\geq 2$. To localize onto geometries with such singularities we must insert an operator $\mathcal{V}_{\alpha}$ into the path integral (3.1). We explain the dilaton gravity form of this operator slightly later, but for now in all integrals we assume we have inserted this operator and integrated over the dilaton to localize onto the relevant hyperbolic geometries. When integrating over hyperbolic surfaces with arbitrary conical singularities and asymptotic boundaries we again obtain the Weil-Petersson measure for the associated moduli spaces. This is because the measure (3.3) for metrics with conical singularities is again by definition the Weil-Petersson metric on such spaces[45]. Carrying out the JT gravity path integral in the same way as for a compact surface we arrive at $\displaystyle Z_{g,{\vec{\alpha}}}(\beta_{1},\ldots,\beta_{n})$ $\displaystyle=(\text{const})^{\chi}\int_{\mathcal{M}_{g,{\vec{\alpha},\vec{\beta}}}}d\left(\text{Weil- Pet.}\right)\frac{\det\left(-\hat{\nabla}_{1}^{2}+1\right)^{1/2}}{\det(-\hat{\nabla}^{2}+2)}e^{-I_{\text{bdy}}}$ $\displaystyle=(\text{const})^{\chi}\int_{\mathcal{M}_{g,{\vec{\alpha},\vec{\beta}}}}d(\text{boundary wiggles})d\left(\text{bulk moduli}\right)e^{-I_{\text{bdy}}}.$ (3.11) In the above we have split the integral over the moduli $d\left(\text{Weil- Pet.}\right)$ into the boundary wiggles and the bulk moduli. The Weil- Petersson measure (3.3) will induce a measure for the boundary wiggles, which we will calculate in section 3.2.1 and see is given by the standard schwarzian measure. While we are only able to compute the required determinants for special defect angles, it seems likely that the cancellation between determinants will remain the case for general conical defects. #### 3.1.1 Integrating over moduli space We now explain how to formally carry out the integration over moduli space. For reviews see [29, 31, 46, 47, 48]. From (3.4) the infinitesimal deformations of the moduli correspond to variations of the metric $\delta g\in\operatorname{Ker}P_{1}^{\dagger}$. In conformal gauge where the metric is off-diagonal this translates to the condition $\overline{\partial}\delta g_{zz}=\partial\delta g_{\bar{z}\bar{z}}=0.$ This implies that the moduli deformations are holomorphic/anti-holomorphic deformations of the form $\phi_{n}=\phi_{n}(z)dz^{2},\qquad\overline{\phi_{n}}=\overline{\phi_{n}}(\bar{z})d\bar{z}^{2}.$ (3.12) These are known as quadratic differentials, and they provide a basis for infinitesimal deformations of the moduli. We briefly explain some properties of quadratic differentials. On a compact genus $g\geq 2$ surface there is a basis of $3g-3$ globally defined pairs of holomorphic and anti-holomorphic quadratic differentials giving a moduli space of real dimension $6g-6$. When we consider hyperbolic surfaces with conical defects, the quadratic differentials are allowed to have simple poles at the location of the defect $\phi\sim z^{-1}dz^{2}$ [45].282828For examples of quadratic differential on surfaces with defects see [49]. These deformations correspond to moving the conical defect on the surface, and so the real dimension of the moduli space with $n$ conical defects is $6g-6+2n$. When we consider the inclusion of asymptotic boundaries the boundary fluctuations are also part of the moduli, and there are infinitely many quadratic differentials associated to turning on modes of the boundary fluctuations as we will see in section 3.2.1. An arbitrary infinitesimal moduli deformation, in conformal gauge, can be written in terms of quadratic differentials as292929The normalization factor of (3.14) is our convention. We will fix the normalization factor of the metric perturbation inner product later, and are satisfied with the proportionality here as a motivation. $\delta g=\sum_{n}\delta c_{n}\phi_{n}(z)dz^{2}+\delta\overline{c}_{n}\overline{\phi}_{n}(\bar{z})d\bar{z}^{2},$ (3.13) where $\delta c_{n}$ are deformation parameters. Using the inner product for metric deformations (3.3) we can define an inner product for quadratic differentials, given by $\langle\phi_{n},\phi_{m}\rangle\equiv\int_{\Sigma}d^{2}z\sqrt{g}g^{z\overline{z}}g^{z\overline{z}}\phi_{n}(z)\overline{\phi}_{m}(\bar{z}),$ (3.14) where the second quantity is to be complex conjugated. It is standard to ignore the normalization in (3.3), restoring it only when computing the full path integral measure. To define the measure on moduli space we will need one more quantity known as the Beltrami differential $\mu$. To integrate over moduli space we must first choose coordinates $m_{n},\bar{m}_{n}$ on the space by specifying a set of metrics $g(z,\bar{z};m_{n},\bar{m}_{n})$ that give a slice through the space. We must then compute the Jacobian for this set of coordinates. This can be accomplished as follows. Consider a metric $g(m_{n},\bar{m}_{n})$ on a surface which in some local patch takes form $ds^{2}=e^{2\omega}|dz|^{2}=2g_{z\bar{z}}|dz|^{2}$ for some complex coordinate $z$. As we change the moduli by moving to a nearby metric at $m_{n}+\delta m_{n}$ the new metric can no longer be $\propto|dz|^{2}$ since it is not in the same equivalence class as the original metric. However, there must be complex coordinates $z^{\prime}=z+\delta m_{n}\,v(z,\bar{z})+\ldots$ where the new metric takes the form $e^{2(\omega+\delta\omega)}|dz^{\prime}|^{2}$. We thus find that the deformed metric in the old $z$ coordinates is $ds^{\prime 2}=e^{2(\omega+\delta\omega)}\left(dzd\bar{z}+\delta m_{n}\hskip 1.9919pt\overline{\mu}_{n}dz^{2}+\delta\overline{m}_{n}\hskip 1.9919pt\mu_{n}d\bar{z}^{2}+\ldots\right),\qquad\delta(ds^{2})=\delta m_{n}e^{2\omega}\overline{\mu}_{n}dz^{2}+\text{c.c.}$ (3.15) where we have defined the Beltrami differentials $\mu=\overline{\partial}v$, and $\overline{\mu}=\partial v$,303030The Beltrami differential are commonly viewed as $(-1,1)$ forms $\mu=\mu_{\bar{z}}^{~{}z}dz^{-1}d\bar{z}$ which can be integrated against $(2,0)$ forms $\phi_{zz}dz^{2}$, and it is convention to define the metric deformation by beltrami differentials as $\overline{\mu}dz^{2}$. which capture how the metric infinitesimally changes as we change the $m_{n}$ coordinates. In the second equation we have written the variation to linear order.313131All these statements are up to Weyl rescalings since rescalings do not move the metric in moduli space. The overlap between the quadratic differentials and the change in the metric due to a deformation in $m_{n}$ coordinates can be obtained from the measure (3.3), which is most commonly expressed as an overlap between the Beltrami differentials $\mu$ and the quadratic differentials $\langle\mu_{n},\phi_{m}\rangle\equiv 2\int_{\Sigma}d^{2}z\mu_{n}\hskip 1.42271pt\phi_{m},$ (3.16) where the overlap is defined without a conjugation. Notice that deforming the metric by a quadratic differential $\phi(z)dz^{2}$ (3.13) is equivalent to deforming it by a beltrami differential $\overline{\mu}=\frac{\phi}{2g_{z\bar{z}}}$ to linear order in the deformation. To compute the correct measure we must project the above deformation onto the space of genuine moduli deformations (as opposed to diffeomorphisms and Weyl rescalings). Projecting onto the basis of quadratic differentials we get $\delta g=\delta m_{n}\phi_{j}\langle\overline{\mu}_{n},\overline{\phi}_{i}\rangle\langle\phi_{j},\phi_{i}\rangle^{-1}+\delta\overline{m}_{n}\overline{\phi}_{j}\langle\mu_{n},\phi_{i}\rangle\langle\overline{\phi}_{j},\overline{\phi}_{i}\rangle^{-1},$ (3.17) where the notation $\langle...\rangle^{-1}$ is an inverse matrix. Using (3.3) we can compute the metric $ds^{2}_{\text{Weil- Pet.}}=2\mathcal{N}\langle\overline{\mu}_{n},\overline{\phi}_{j}\rangle\langle\mu_{m},\phi_{i}\rangle\langle\phi_{i},\phi_{j}\rangle^{-1}dm_{n}d\overline{m}_{m}.$ (3.18) Taking the square root of the determinant and treating the metric as a product of matrices, we find that in terms of $m_{n}$ coordinates the integral over moduli space takes the form $\int_{\mathcal{M}}d(\text{Weil-Pet.})\hskip 2.84544pte^{-I}=\prod_{n}\mathcal{N}\int dm_{n}d\overline{m}_{n}\frac{\det\langle\mu,\phi\rangle\det\langle\overline{\mu},\overline{\phi}\rangle}{\sqrt{\det\langle\phi,\phi\rangle\det\langle\overline{\phi},\overline{\phi}\rangle}}\hskip 2.84544pte^{-I[m_{n},\overline{m}_{n}]}.$ (3.19) We have also included a possibility of some action $I$ that depends on the moduli. In JT gravity this action will be given by the boundary term in (3.2). We have also reinstated the overall normalization constant $\mathcal{N}$ appearing in the measure (3.3). In [13], and in appendix B using quadratic differentials, it can be shown that in the second order formalism the gluing measure for geodesic boundaries is given by $4\mathcal{N}\int bdb,$ (3.20) and so to get standard form of the Weil-Petersson measure we must choose the normalization $\mathcal{N}=\frac{1}{4}$. We keep the normalization constant present throughout, inserting it’s numerical value when necessary. #### Conical defect operator From the expression for the Weil-Petersson measure we can extract the dilaton gravity operator that creates a conical defect on the surface. When we have a defect we have two additional moduli that move the defect. We choose the moduli to be parameterized by the location $z_{i}$ of the defect, and we can formally write the measure as $\mathcal{N}\int_{\Sigma}d^{2}z_{i}\frac{\langle\mu,\phi_{1}\rangle\langle\overline{\mu},\overline{\phi}_{1}\rangle}{\sqrt{\langle\phi_{1},\phi_{1}\rangle\langle\overline{\phi}_{1},\overline{\phi}_{1}\rangle}},$ (3.21) where $\mu$ is the beltrami differential that infinitesimally moves the defect to $z_{i}+\delta z_{i}$, and $\phi_{1}$ is the quadratic differential with a simple pole at $z_{i}$. From the JT gravity perspective this is the measure after the dilaton has already been integrated out, and so restoring it we arrive at the dilaton gravity operator that creates a conical defect $\mathcal{V}_{\alpha}=\mathcal{N}\int_{\Sigma}d^{2}z_{i}\frac{\langle\mu,\phi_{1}\rangle\langle\overline{\mu},\overline{\phi}_{1}\rangle}{\sqrt{\langle\phi_{1},\phi_{1}\rangle\langle\overline{\phi}_{1},\overline{\phi}_{1}\rangle}}e^{-2\pi(1-\alpha)\Phi(z_{i})}.$ (3.22) One complication for evaluating the above integral is that quadratic differentials are not generally known, apart from their singular behavior near the defect. However, for the hyperbolic disk everything can be worked out explicitly, and in section 3.2.3 we evaluate the measure on the disk. In section 3.2.4 we argue that for general surfaces in the blunt defect limit $\alpha\to 1$ the measure can be worked out to leading order in the $(1-\alpha)$ expansion by using the universal divergent behavior of the quadratic differential. ### 3.2 Examples #### 3.2.1 Disk In this section we explain how to apply the above formalism for integrating over moduli space to reproduce disk partition function $Z(\beta)$ in JT gravity[14, 15, 16]. The AdS2 disk metric in complex coordinates is given by323232Our conventions for complex coordinates are $z=x+iy$ with $\int d^{2}z=2\int d^{2}x$. The dirac delta function is defined by $\overline{\partial}\partial\log|z|^{2}=2\pi\delta^{2}(z,\bar{z})$, with $\int d^{2}z\delta^{2}(z,\bar{z})=1$. A Weyl rescaled metric $g=e^{2\omega}dzd\bar{z}$ has curvature $R=-8e^{-2\omega}\overline{\partial}\partial\omega$. $ds^{2}=\frac{4dzd\overline{z}}{\left(1-|z|^{2}\right)^{2}}.$ (3.23) This is related to the standard metric on the hyperbolic disk through the coordinate change $ds^{2}=dr^{2}+\sinh^{2}r\hskip 1.70709ptd\tau^{2},\qquad z=e^{i\tau}\tanh\frac{r}{2}.$ (3.24) As explained previously, deformations of the moduli correspond to holomorphic quadratic differentials and their conjugates. On the disk it is simple to write down a basis of holomorphic functions given by $\phi_{n}=\sqrt{\frac{n^{3}-n}{2\pi}}z^{n-2}dz^{2},\qquad\overline{\phi}_{n}=\sqrt{\frac{n^{3}-n}{2\pi}}\overline{z}^{n-2}d\overline{z}^{2},\qquad\langle\phi_{n},\phi_{m}\rangle=\delta_{n,m},$ (3.25) where for holomorphicity we restrict to integers $n\geq 2$. Infinitesimal deformations of the disk are given by deforming by linear combinations of the quadratic differentials $ds^{2}=\frac{4dzd\overline{z}}{\left(1-|z|^{2}\right)^{2}}+c_{n}\sqrt{\frac{n^{3}-n}{2\pi}}z^{n-2}dz^{2}+\overline{c}_{n}\sqrt{\frac{n^{3}-n}{2\pi}}\overline{z}^{n-2}d\overline{z}^{2}.$ (3.26) Note that the above metric is no longer AdS2. We will take our coordinates for moduli space to be given by $c_{n},\overline{c}_{n}$, and so we must work out the measure for these coordinates along with the boundary action of JT gravity (3.2) in terms of $c_{n},\overline{c}_{n}$. The measure is straightforward since for the Beltrami differentials in (3.19) we just use the dual of the quadratic differentials $\overline{\mu}_{n}=\frac{\phi_{n}}{2g_{z\overline{z}}}$. Working out the action is slightly more complicated, which we now do. The finite version of the metric for the above deformation was worked out by [16], and is given by333333We mention that we still mostly work with infinitesimal perturbations of the metric and the following finite metric deformation is not unique. $\displaystyle ds^{2}=e^{C+\overline{C}}\frac{\left(1-|z|^{2}\right)^{2}}{\left(1-|e^{C}z|^{2}\right)^{2}}\left[\left(|1+z\partial C|^{2}+|\overline{z}\overline{\partial}C|^{2}\right)\frac{4dzd\overline{z}}{\left(1-|z|^{2}\right)^{2}}{+}\left(1+z\partial C\right)\varepsilon^{\prime\prime\prime}(z)dz^{2}+\text{c.c.}\right],$ (3.27) $\displaystyle C(z,\overline{z})={{-}}\frac{1}{2z}\left(\varepsilon(z)-z^{2}\overline{\varepsilon}(\overline{z})-z(1-z\overline{z})\overline{\varepsilon}^{\prime}(\overline{z})-\frac{1}{2}(1-z\overline{z})^{2}\overline{\varepsilon}^{\prime\prime}(\overline{z})\right).$ The functions $\varepsilon(z),\overline{\varepsilon}(\overline{z})$ are holomorphic/anti-holomorphic, and the primes $\varepsilon^{\prime}(z),\overline{\varepsilon}^{\prime}(\overline{z})$ denote holomorphic/anti-holomorphic derivatives. The metric (3.27) is AdS2, which can be seen by using coordinates $w=ze^{C}$ in terms of which it takes the standard form (3.23). This coordinate change is a large diffeomorphism and corresponds to a physically distinct configuration. The relation between (3.26) and (3.27) comes from identifying $\varepsilon(z)=\sum_{n\geq 2}\frac{1}{\sqrt{2\pi(n^{3}-n)}}\,c_{n}z^{n+1},\qquad\overline{\varepsilon}(\overline{z})=\sum_{n\geq 2}\frac{1}{\sqrt{2\pi(n^{3}-n)}}\,\overline{c}_{n}\overline{z}^{n+1}.$ (3.28) Expanding the metric (3.27) to leading order in $\varepsilon$ we recover the infinitesimal form (3.26) up to a Weyl factor of order $\varepsilon^{2}$. We now explain how turning on these deformations corresponds to turning on modes of the Schwarzian. In $z$ coordinates (3.27) the boundary cutoff is located at fixed radial distance while in $w=ze^{C}$ coordinates the metric takes the form $ds^{2}=\frac{4dwd\overline{w}}{(1-|w|^{2})^{2}},$ (3.29) and the cutoff surface fluctuates. Taking $z=e^{i\theta}$ we find that at the boundary and linear order of $c_{n}$ $w=e^{if(\theta)},\qquad f(\theta)=\theta+\sum_{n\geq 2}\frac{1}{2i\sqrt{2\pi(n^{3}-n)}}\left(c_{n}e^{in\theta}-\overline{c}_{n}e^{-in\theta}\right).$ (3.30) We would like to find the cutoff surface in terms of the $f(\theta)$ reparameterization. We can do this perturbatively in the cutoff $\delta$ finding $w\left(\theta\right)=e^{if(\theta)}\left(1-\delta f^{\prime}(\theta)+\frac{\delta^{2}f^{\prime}(\theta)^{2}}{2}+\mathcal{O}(\delta^{3})\right)$. Where we have found the radial distance in terms of $f(\theta)$ using that the induced metric along the cutoff surface is fixed to be $\sqrt{h}=1/\delta$. The extrinsic curvature can then be evaluated $\displaystyle K=1+\delta^{2}\left(\frac{f^{\prime\prime\prime}}{f^{\prime}}-\frac{3f^{\prime\prime 2}}{2f^{\prime 2}}+\frac{1}{2}f^{\prime 2}\right)+\mathcal{O}(\delta^{3})=1+\delta^{2}\operatorname{Sch}\left(\tan\frac{f(\theta)}{2},\theta\right),$ (3.31) which is the standard Schwarzian answer. From (3.30) we see that the quadratic differentials $\phi_{n},\overline{\phi}_{n}$ precisely correspond to turning on the modes of the Schwarzian. The boundary action (3.2) can now be evaluated. Reinstating the temperature $\beta$ and the value of the dilaton $\phi=\gamma/\delta$ we find $\displaystyle I_{\text{bdy}}$ $\displaystyle=-\frac{2\pi\gamma}{\beta}\int_{0}^{2\pi}d\theta\left(\frac{f^{\prime\prime\prime}}{f^{\prime}}-\frac{3f^{\prime\prime 2}}{2f^{\prime 2}}+\frac{1}{2}f^{\prime 2}\right),$ (3.32) $\displaystyle=I_{0}+\frac{\pi\gamma}{2\beta}\sum_{n\geq 2}n\hskip 0.28436ptc_{n}\overline{c}_{n},\qquad I_{0}=-\frac{2\pi^{2}\gamma}{\beta}.$ Where in the second line we have used that the Schwarzian path-integral is one-loop exact [50] so that it is sufficient to expand to quadratic order in $c_{n},\overline{c}_{n}$ and carry out the integral[7, 51], and $I_{0}$ is the classical contribution. We can now evaluate the path integral over the moduli space using (3.19) $\displaystyle Z_{\text{disk}}\left(\beta\right)$ $\displaystyle=\prod_{n\geq 2}\mathcal{N}\int dc_{n}d\overline{c}_{n}\left|\frac{\det\langle\mu,\phi\rangle}{\sqrt{\det\langle\phi,\phi\rangle}}\right|^{2}e^{-I_{\text{bdy}}}=e^{\frac{2\pi^{2}\gamma}{\beta}}\prod_{n\geq 2}\mathcal{N}\int dc_{n}d\overline{c}_{n}\exp\left(-\frac{\pi\gamma}{2\beta}\sum_{n\geq 2}n\hskip 0.28436ptc_{n}\overline{c}_{n}\right)$ (3.33) $\displaystyle=e^{\frac{2\pi^{2}\gamma}{\beta}}\prod_{n\geq 2}\mathcal{N}\frac{4\beta}{\gamma n}=\frac{1}{8}\sqrt{\frac{\gamma^{3}}{2\pi\mathcal{N}^{3}\beta^{3}}}\hskip 2.84544pte^{\frac{2\pi^{2}\gamma}{\beta}}.$ In the first line we used our normalization for the quadratic differentials $\langle\phi_{n},\phi_{m}\rangle=\delta_{n,m}$ and that the dual beltrami differentials $\overline{\mu}_{n}=\frac{\phi_{n}}{2g_{z\overline{z}}}$ satisfy $\langle\mu_{n},\phi_{m}\rangle=\delta_{n,m}$. In the last line we used zeta function regularization. This reproduces the JT gravity disk partition function obtained from the first order formalism [7]. #### 3.2.2 Conical defect We now explain how the above calculation is modified in the case of a disk with a single conical defect of opening angle $2\pi\alpha$. If the defect is at the center of the disk then the metric and Ricci scalar are given by $ds^{2}=\frac{4\alpha^{2}|z|^{2(\alpha-1)}}{(1-|z|^{2\alpha})^{2}}dzd\bar{z}\,,\qquad\frac{1}{2}\sqrt{g}\left(R+2\right)=2\pi(1-\alpha)\delta^{(2)}(z).$ (3.34) The quadratic differentials are now given by $\phi_{n}=\sqrt{\frac{n^{3}-n\alpha^{2}}{2\pi}}\hskip 2.84544ptz^{n-2},\qquad\overline{\phi}_{n}=\sqrt{\frac{n^{3}-n\alpha^{2}}{2\pi}}\hskip 2.84544pt\overline{z}^{n-2}d\overline{z}^{2},\qquad\langle\phi_{n},\phi_{m}\rangle=\delta_{n,m}.$ (3.35) When a conical defect is present on the geometry the quadratic differentials are allowed to have simple poles[45] at the location of the defect so we now allow for $n\geq 1$ modes. This increases the real dimension of the moduli space by two, and the new deformation should be interpreted as moving the defect on the surface. The quadratic differentials infinitesimally deform the metric by $ds^{2}=\frac{4\alpha^{2}|z|^{2(\alpha-1)}}{(1-|z|^{2\alpha})^{2}}dzd\bar{z}+c_{n}\sqrt{\frac{n^{3}-n\alpha^{2}}{2\pi}}\hskip 2.84544ptz^{n-2}dz^{2}+\overline{c}_{n}\sqrt{\frac{n^{3}-n\alpha^{2}}{2\pi}}\hskip 2.84544pt\overline{z}^{n-2}d\overline{z}^{2}.$ (3.36) Similar to the case of the disk, there is a vector field $\xi$ that infinitesimally removes these deformations, and we can find it by solving for $2\nabla_{z}\xi_{z}=-c_{n}\sqrt{\frac{n^{3}-n\alpha^{2}}{2\pi}}\hskip 2.84544ptz^{n-2},\qquad 2\nabla_{\bar{z}}\xi_{\bar{z}}=-\overline{c}_{n}\sqrt{\frac{n^{3}-n\alpha^{2}}{2\pi}}\hskip 2.84544pt\overline{z}^{n-2},\qquad\nabla_{z}\xi_{\bar{z}}+\nabla_{\bar{z}}\xi_{z}=0.$ (3.37) The vector field can be found and takes the form $\displaystyle\xi^{\bar{z}}=\frac{\overline{c}_{n}\overline{z}^{n+1}}{2\sqrt{2\pi(n^{3}-n\alpha^{2})}}-\frac{c_{n}\overline{z}(z\overline{z})^{-\alpha}\left(2\alpha^{2}z^{n}(z\overline{z})^{\alpha}+n^{2}z^{n}\left(1-(z\overline{z})^{\alpha}\right)^{2}+n\alpha z^{n}\left(1-(z\overline{z})^{2\alpha}\right)\right)}{4\alpha^{2}\sqrt{2\pi(n^{3}-n\alpha^{2}})},$ (3.38) $\displaystyle\xi^{z}=\frac{c_{n}z^{n+1}}{2\sqrt{2\pi(n^{3}-n\alpha^{2})}}-\frac{c_{n}z(z\overline{z})^{-\alpha}\left(2\alpha^{2}\overline{z}^{n}(z\overline{z})^{\alpha}+n^{2}\overline{z}^{n}\left(1-(z\overline{z})^{\alpha}\right)^{2}+n\alpha\overline{z}^{n}\left(1-(z\overline{z})^{2\alpha}\right)\right)}{4\alpha^{2}\sqrt{2\pi(n^{3}-n\alpha^{2}})}.$ The finite form of the deformed metric can be recovered by choosing coordinates $w=z\exp\left(\frac{1}{z}\xi^{z}\right)$ where in $w$ coordinates we have the metric (3.34) (compare to (3.27)), and we implicitly sum over $c_{n},\overline{c}_{n}$ in $\xi^{z}$. We can now follow the same procedure to find the shape of the cutoff in $w$ coordinates as we followed for the disk. Following the curve $z=e^{i\theta}$ we find in $w$ coordinates $w=e^{if(\theta)},\qquad f(\theta)=\theta+\sum_{n\geq 1}\frac{1}{2i\sqrt{2\pi(n^{3}-n\alpha^{2})}}\left(c_{n}e^{in\theta}-\overline{c}_{n}e^{-in\theta}\right).$ (3.39) Fixing the induced metric along the cutoff surface, we again find the same functional form with a defect present $w\left(\theta\right)=e^{if(\theta)}\left(1-\delta f^{\prime}(\theta)+\frac{\delta^{2}f^{\prime}(\theta)^{2}}{2}+\mathcal{O}(\delta^{3})\right)$. Calculating the extrinsic curvature along this curve we find $\displaystyle K=1+\delta^{2}\left(\frac{f^{\prime\prime\prime}}{f^{\prime}}-\frac{3f^{\prime\prime 2}}{2f^{\prime 2}}+\frac{\alpha^{2}}{2}f^{\prime 2}\right)+\mathcal{O}(\delta^{3})=1+\delta^{2}\operatorname{Sch}\left(\tan\frac{\alpha f(\theta)}{2},\theta\right).$ (3.40) Which is precisely the Schwarzian with a conical defect[52] which is again one-loop exact. Restoring temperature and the dilaton as in the case of the disk we find the action to quadratic order to be $\displaystyle I_{\text{bdy}}$ $\displaystyle=-\frac{2\pi\gamma}{\beta}\int_{0}^{2\pi}d\theta\left(\frac{f^{\prime\prime\prime}}{f^{\prime}}-\frac{3f^{\prime\prime 2}}{2f^{\prime 2}}+\frac{\alpha^{2}}{2}f^{\prime 2}\right),$ (3.41) $\displaystyle=I_{0}+\frac{\pi\gamma}{2\beta}\sum_{n\geq 1}n\hskip 0.28436ptc_{n}\overline{c}_{n},\qquad I_{0}=-\frac{2\pi^{2}\gamma}{\beta}\alpha^{2}.$ Integrating over the moduli space with (3.19) we need to now include the $n=1$ mode compared with the disk calculation $\displaystyle Z_{\text{defect}}\left(\beta,\alpha\right)$ $\displaystyle=\prod_{n\geq 1}\mathcal{N}\int dc_{n}d\overline{c}_{n}\left|\frac{\det\langle\mu,\phi\rangle}{\sqrt{\det\langle\phi,\phi\rangle}}\right|^{2}e^{-I_{\text{b}dy}}=e^{\frac{2\pi^{2}\gamma}{\beta}\alpha^{2}}\prod_{n\geq 1}\mathcal{N}\int dc_{n}d\overline{c}_{n}\exp\left(-\frac{\pi\gamma}{2\beta}\sum_{n\geq 1}n\hskip 0.28436ptc_{n}\overline{c}_{n}\right)$ (3.42) $\displaystyle=e^{\frac{2\pi^{2}\gamma}{\beta}\alpha^{2}}\prod_{n\geq 1}\mathcal{N}\frac{4\beta}{\gamma n}=\sqrt{\frac{\gamma}{8\pi\mathcal{N}\beta}}e^{\frac{2\pi^{2}\gamma}{\beta}\alpha^{2}},$ where again the measure simplifies due to our normalization (3.35) as in the case of the disk (3.33). This again matches the calculation of the partition function with a single defect [52, 7]. #### 3.2.3 Conical defect: $z$ coordinates In the previous section we integrated over moduli space using $c_{n}$ coordinates which correspond to deformations of the metric by quadratic differentials. In this section we will carry out the previous calculation using a coordinate choice that is more physically intuitive, namely the position $z_{i}$ of the defect on the disk. This choice of coordinates is applicable to the $n=1$ mode of the Schwarzian which can be thought of as moving the defect on the disk. To integrate over the moduli space with $z_{i}$ coordinates we must again work out both the measure and the action in (3.19), which we now do. The metric with a conical defect at a general position $z_{i}$ can be obtained by applying an $\text{SL}(2,\mathbb{R})$ transformation to the metric with a defect at the center $ds^{2}=\frac{4\alpha^{2}|w(z)|^{2(\alpha-1)}}{(1-|w(z)|^{2\alpha})^{2}}|w^{\prime}(z)|^{2}dzd\bar{z}\,,\qquad w(z)=\frac{z-z_{i}}{1-z\overline{z}_{i}},$ (3.43) where in $w$ coordinates the defect is at $w=0$, but is at $z=z_{i}$ in $z$ coordinates. To work out the measure in (3.19) we need to know the Beltrami differential $\mu$ which contains information on how the metric infinitesimally changes as we move the defect $z_{i}\to z_{i}+\delta z_{i}$. As explained in section 3.1.1, this can be obtained by finding a set of coordinates $z^{\prime}(z,\bar{z})$ where the deformed metric is flat $\propto|dz^{\prime}|^{2}$ and the defect is mapped to the new point $z^{\prime}(z_{i})=z_{i}+\delta z_{i}$ with the asymptotic boundary undeformed.343434The coordinates $z^{\prime}$ should not affect the boundary of the disk since such a deformation would both move the defect and turn on some additional Schwarzian modes, and we want to isolate the deformation that moves the defect. To extract the Beltrami we only need this coordinate chart to linear order in $\delta z_{i}$, where it is given by $z^{\prime}(z,\bar{z})=\frac{\delta z_{i}\frac{1-z\bar{z}}{1-z_{i}\bar{z_{i}}}+z(1-z_{i}\bar{z})}{\delta z_{i}\bar{z}_{i}\frac{1-z\bar{z}}{1-z_{i}\bar{z_{i}}}+(1-z_{i}\bar{z})}.$ (3.44) $|dz^{\prime}|^{2}\propto|dz+\delta z_{i}\hskip 1.13791pt\mu d\bar{z}+\mathcal{O}((\delta z_{i})^{2})|^{2},\qquad\mu=-\frac{(z-z_{i})(1-z\bar{z}_{i})}{(1-z_{i}\bar{z}_{i})(1-\bar{z}z_{i})^{2}}\,.$ (3.45) Additionally, the quadratic differential that moves the defect when it is located $z=z_{i}$ is given by353535It can be checked that $\langle\mu,\phi_{n}\rangle=0$ for $n\geq 2$, implying that the only action of $\mu$ is to move the defect. $\phi_{1}=\frac{1}{w}(\partial w)^{2}dz^{2}=\frac{(1-z_{i}\bar{z}_{i})^{2}}{(z-z_{i})(1-z\bar{z}_{i})^{3}}dz^{2},\qquad\langle\phi_{1},\phi_{1}\rangle=\frac{2\pi}{1-\alpha^{2}}.$ (3.46) Which is precisely $\phi_{1}=\frac{1}{w}dw^{2}$ in $w$ coordinates, implying the norm relation noted above. Using the Beltrami and quadratic differential we can compute the overlap $\langle\mu,\phi_{1}\rangle=2\int dzd\bar{z}\frac{1-|z_{i}|^{2}}{(1-z\bar{z}_{i})^{2}(1-z_{i}\bar{z})^{2}}=\frac{4\pi}{1-|z_{i}|^{2}}\,.$ (3.47) We can now immediately compute the Weil-Petersson measure (3.19) for the $z_{i}$ coordinates that move the defect $\mathcal{N}\int d^{2}z_{i}\left|\frac{\langle\mu,\phi_{1}\rangle}{\sqrt{\langle\phi_{1},\phi_{1}\rangle}}\right|^{2}=8\pi\mathcal{N}\int d^{2}z_{i}\frac{1-\alpha^{2}}{(1-|z_{i}|^{2})^{2}}=4\pi\mathcal{N}(1-\alpha^{2})\int d^{2}z_{i}\sqrt{g}\,,$ (3.48) where in the last equality we have noticed the measure takes the form of the metric on the disk without a defect. We are now left to compute the action in the $z_{i}$ coordinates. We follow the same procedure as in the previous section, if we take the cutoff to be at $z=e^{i\theta}$ then in $w=e^{if(\theta)}$ coordinates (3.43) we have $w(e^{i\theta})=e^{i\theta}\frac{1-z_{i}e^{-i\theta}}{1-\bar{z}_{i}e^{i\theta}}\ \Rightarrow\ f(\theta)=\theta+i^{-1}\log\left(\frac{1-z_{i}e^{-i\theta}}{1-\bar{z}_{i}e^{i\theta}}\right)\,.$ (3.49) Note that this reparameterization only takes into account the $n=1$ mode of the Schwarzian. We previously worked out the action for $f(\theta)$ in (3.41), giving $\displaystyle I^{n=1}_{\text{bdy}}$ $\displaystyle=-\frac{2\pi\gamma}{\beta}\int_{0}^{2\pi}d\theta\left(\frac{f^{\prime\prime\prime}}{f^{\prime}}-\frac{3f^{\prime\prime 2}}{2f^{\prime 2}}+\frac{\alpha^{2}}{2}f^{\prime 2}\right),$ $\displaystyle=-\frac{2\pi\gamma}{\beta}\int_{0}^{2\pi}d\theta\frac{1}{2}\left[\alpha^{2}\frac{(1-|z_{i}|^{2})^{2}}{|1-z_{i}e^{-i\theta}|^{4}}-\frac{(z_{i}e^{-i\theta}+\bar{z}_{i}e^{i\theta}-2)(z_{i}e^{-i\theta}+\bar{z}_{i}e^{i\theta}-2|z_{i}|^{2})}{|1-z_{i}e^{-i\theta}|^{4}}\right],$ $\displaystyle=-\frac{2\pi^{2}\gamma}{\beta}\left(1-(1-\alpha^{2})\frac{1+|z_{i}|^{2}}{1-|z_{i}|^{2}}\right).$ (3.50) We can now write the full answer using (3.19), (3.48), and (3.2.3). Including all the $n\geq 2$ modes of the Schwarzian we get $Z_{\text{defect}}(\beta,\alpha)=4\pi\mathcal{N}\int d^{2}z_{i}\frac{\left(1-\alpha^{2}\right)}{\left(1-|z_{i}|^{2}\right)^{2}}e^{-I^{n=1}_{\text{bdy}}}\times\prod_{n\geq 2}\mathcal{N}\frac{4\beta}{\gamma n}=e^{\frac{2\pi^{2}\gamma}{\beta}\alpha^{2}}\prod_{n\geq 1}\mathcal{N}\frac{4\beta}{\gamma n},$ (3.51) which is the expected answer (3.42). From the above we see that when integrating over the position of the defect we pick up a measure factor proportional to $(1-\alpha^{2})$ along with an action proportional to the same. In the limit $\alpha\to 1$ the defect becomes very blunt and one expects that it does not backreact strongly on the geometry, so that semiclassical methods can be applied. However, we see that the path integral measure is also becoming important in this limit, and must be taken into account. We can rewrite the integral over $z_{i}$ in a way that connects it to the semiclassical calculation of a single defect on the disk. The contribution of the $n=1$ mode of the Schwarzian is $Z^{n=1}_{\text{defect}}(\beta,\alpha)=4\pi\mathcal{N}(1-\alpha^{2})\int d^{2}z_{i}\sqrt{g(z_{i})}e^{-\pi(1-\alpha^{2})\phi_{\text{cl.}}(z_{i})},\qquad\phi_{\text{cl.}}(z_{i})=\frac{2\pi\gamma}{\beta}\frac{1+|z_{i}|^{2}}{1-|z_{i}|^{2}},$ (3.52) where in the above $\sqrt{g}$ is the metric on the hyperbolic disk without a defect and $\phi_{\text{cl.}}$ is the classical solution for the dilaton on the disk. It has previously been observed that a semiclassical calculation [17, 18] of the disk with a single defect seems to agree with the full Schwarzian calculation if an appropriate measure factor is inserted for the integral over the dilaton. The semiclassical calculation is given by $\lim_{\alpha\to 1}Z_{\text{defect}}^{\text{semiclassical}}\approx 2\pi\left(1-\alpha\right)\int d^{2}z\sqrt{g}e^{-2\pi\left(1-\alpha\right)\phi_{\text{cl.}}(z)},$ (3.53) where we approximate the integral over the dilaton by it’s classical value, and a measure factor of $2\pi\left(1-\alpha\right)$ is inserted by hand to get agreement with the Schwarzian. Comparing with (3.52) we see why this semi- classical calculation works, the integral over the position cancels out the measure in both calculations giving the same answer. #### 3.2.4 General surfaces In this section we will give a general argument that the correct form of the path integral measure for a conical defect in the limit that the angle becomes blunt is given by $\lim_{\alpha\to 1}\mathcal{N}\int d^{2}z\left|\frac{\langle\mu,\phi_{1}\rangle}{\sqrt{\langle\phi_{1},\phi_{1}\rangle}}\right|^{2}=2\pi(1-\alpha)\int d^{2}z\sqrt{g(z)}+\mathcal{O}\left(\left(1-\alpha\right)^{2}\right),$ (3.54) where we have inserted $\mathcal{N}=\frac{1}{4}$ to normalize the operator with respect to the standard gluing measure $\int bdb$. Consider a surface $\Sigma$ with at least one conical defect363636For simplicity, we do not consider geodesic boundary temporarily. located at $z=z_{i}$. The quadratic differential that moves this defect has a simple pole at $z_{i}$ $\phi_{1}=\frac{1}{z-z_{i}}dz^{2}+\ldots\,,$ (3.55) where the subleading terms are holomorphic on $\Sigma$.373737We can also include a residue at the pole, as we did for the disk in (3.46) and (3.47), but it will cancel out of the path integral measure. In principle we need to know the quadratic differential on the entire surface to calculate the measure, but in the blunt defect limit $\alpha\to 1$ we will argue that the singular behavior of the quadratic differential dominates. To calculate the inner products we also need the metric on the surface $ds^{2}=e^{2\omega}dzd\bar{z}$ where we are in local coordinates $z$ near the defect. Demanding that the opening angle is $2\pi\alpha$ at $z_{i}$ the metric is given by solving Liouville’s equation383838We are solving for the constraint $\frac{1}{2}\sqrt{g}\left(R+2\right)=2\pi(1-\alpha)\delta^{2}(z-z_{i})$ and we have used that $R=-2e^{-2\omega}\hat{\nabla}^{2}\omega+e^{-2\omega}\hat{R}$ for a Weyl transformed metric $g=e^{2\omega}\hat{g}$. with a source $-2\overline{\partial}\partial{\omega}+\frac{1}{2}e^{2\omega}=2\pi(1-\alpha)\delta^{2}(z-z_{i})\,.$ (3.56) Solving for the metric perturbatively in $(1-\alpha)$ we find $\omega=\omega_{0}-\frac{1}{2}(1-\alpha)\big{[}\log|z-z_{i}|^{2}+r\big{]}+\mathcal{O}\left(\left(1-\alpha\right)^{2}\right)\,,$ (3.57) where $\omega_{0}$ is the Weyl factor for the surface $\Sigma$ without the defect. The linear order term in $(1-\alpha)$ is split up into a logarithm which gives the delta function singularity, along with a holomorphic function $r$ that satisfies $\overline{\partial}\partial r=e^{2\omega_{0}}\log|z-z_{i}|$. One final part we need for the calculation of the measure is the beltrami differential $\mu$ that corresponds to infinitesimally moving the defect from $z_{i}\to z_{i}+\delta z_{i}$. As explained in the case of the disk, the beltrami differential can be extracted from a coordinate chart $z^{\prime}=z+\delta z_{i}\hskip 1.42271ptv(z,\bar{z})+\mathcal{O}((\delta z_{i})^{2})$ such that $z^{\prime}(z_{i})=z_{i}+\delta z_{i}$, which implies that $v(z_{i},\bar{z}_{i})=1$. We only need this coordinate chart to linear order in $\delta z_{i}$. The beltrami differential is then be given by $\mu=\overline{\partial}v(z,\bar{z})$. We will also demand that $v|_{\text{bdy}}=0$ which is the condition that we are infinitesimally moving the defect but not changing any asymptotically AdS boundaries, which will allow us to integrate by parts. We can now calculate all the components of the path integral measure for moving the defect position in $z_{i}$ coordinates: $\langle\mu,\phi_{1}\rangle=\int_{\Sigma}d^{2}z\hskip 1.42271pt\overline{\partial}v\left(\frac{1}{z-z_{i}}+\text{holo.}\right)=2\pi,$ (3.58) where we have integrated by parts and assumed all boundary terms vanish,393939This can be compared to the calculation performed on the disk where we obtain the same answer by performing the full integral (3.47). used that $\phi_{1}$ is holomorphic away from the defect, and that $v(z_{i},\bar{z}_{i})=1$.404040In our conventions $\overline{\partial}z^{-1}=2\pi\delta^{2}(z)$ and $\int d^{2}z\delta^{2}(z)=1$. Computing the inner product of the quadratic differential we find $\displaystyle\langle\phi_{1},\phi_{1}\rangle$ $\displaystyle=2\int_{\Sigma}d^{2}ze^{-2\omega}\left|\frac{1}{z-z_{i}}+\ldots\right|^{2}=2\int_{\Sigma}d^{2}ze^{-2\omega_{0}}\frac{|z-z_{i}|^{2(1-\alpha)}}{|z-z_{i}|^{2}}+\text{less singular},$ (3.59) $\displaystyle\mathrel{\mathop{=}\limits_{\alpha\to 1}}-\frac{4\pi e^{-2\omega_{0}(z_{i})}}{(1-\alpha)}+\mathcal{O}(1)=-\frac{2\pi}{(1-\alpha)\sqrt{g(z_{i})}}+\mathcal{O}(1).$ where in the last line we take $\alpha\to 1$ and localize onto the most singular part of the integral near $z=z_{i}$,414141The integral must be performed with a radial cutoff around $z_{i}$, after which the leading order answer in $(1-\alpha)$ is cutoff independent. and in the second equation we notice that we have picked up a factor of the metric $\sqrt{g}$ with the defect removed from the surface, evaluated at $z_{i}$. Putting everything together, we find that the Weil-Petersson measure (3.19) for the integral over the defect position in the blunt angle limit is given by $\lim_{\alpha\to 1}\mathcal{N}\int d^{2}z_{i}\left|\frac{\langle\mu,\phi_{1}\rangle}{\sqrt{\langle\phi_{1},\phi_{1}\rangle}}\right|^{2}=2\pi\left(1-\alpha\right)\int d^{2}z_{i}\sqrt{g(z_{i})}+\mathcal{O}\left((1-\alpha)^{2}\right).$ (3.60) Note that on the right $g(z_{i})$ is the metric on the surface without the conical defect. From the above we immediately have that the dilaton-gravity operator that inserts a conical defect in the blunt defect limit is given by $\lim_{\alpha\to 1}\mathcal{V}_{\alpha}=2\pi(1-\alpha)\int d^{2}z_{i}\sqrt{g(z_{i})}e^{-2\pi(1-\alpha)\Phi(z_{i})}+\mathcal{O}\left((1-\alpha)^{2}\right).$ (3.61) #### Recursion Relation For WP Volumes This result also allows us to give a gravitational path integral argument for a recursion relation of Weil-Petersson volumes with blunt defects derived in [20]. Consider the Weil-Petersson volume of surfaces $\Sigma$ of genus $g$ with $m$ geodesic boundaries of lengths $\vec{b}_{m}=(b_{1},\ldots,b_{m})$ and $n+1$ conical defects of opening angles $2\pi\alpha_{i}$ with $\vec{\alpha}_{n+1}=(\alpha_{1},\ldots,\alpha_{n},\alpha_{n+1})$. In the limit where one of the defects becomes blunt $\alpha_{n+1}\to 1$ [20] proved the following relation $\frac{dV_{g,m,n+1}\left(\vec{\alpha}_{n+1},\vec{b}_{m}\right)}{d\alpha_{n+1}}\bigg{\rvert}_{\alpha_{n+1}=1}=-2\pi|\Sigma|V_{g,m,n}\left(\vec{\alpha}_{n},\vec{b}_{m}\right),$ (3.62) where $|\Sigma|$ is the hyperbolic area of the surface with $n$ defects satisfying $|{\Sigma}|=-2\pi\left(2-2g-m-\sum_{i=1}^{n}(1-\alpha_{i})\right)=-\frac{1}{2}\int_{\Sigma/\\{x_{i}\\}}\sqrt{g}R.$ (3.63) We can prove this formula be decomposing the volume into an integral over a coordinate parameterizing the position of the $\alpha_{n+1}$ defect, and all the other moduli of the surface. Working near the blunt defect limit the integral over the defect takes the simplified form (3.60) $V_{g,m,n+1}\left(\vec{\alpha}_{n+1},\vec{b}_{m}\right)=\int d\left(\text{other moduli}\right)\times\left(2\pi(1-\alpha_{n+1})\int_{\Sigma/\\{z_{j}\\}}d^{2}z\sqrt{g}+\mathcal{O}\left((1-\alpha_{n+1})^{2}\right)\right),$ (3.64) where the integral over the “other moduli” take the form of the measure (3.19) with various determinants of beltrami and quadratic differentials. The measure for the other moduli implicitly depends on $\alpha_{n+1}$ through the appearance of the defect metric in the inner products defining the measure. However, as we take $\alpha_{n+1}\to 1$ this dependence goes away since the surface no longer has a defect. Therefore, in the limit that the defect vanishes the integral over the other moduli becomes the Weil-Petersson volume of the surface without the $\alpha_{n+1}$ defect. Another important point as explained around (3.60) is that the metric $\sqrt{g}$ appearing in the above is the metric for the surface without the $\alpha_{n+1}$ defect. We have also excluded the integral over the points $z_{j}$ where the other $\vec{\alpha}_{n}$ defects are located, as these configurations are at the boundary of moduli space.424242This exclusion is necessary to reproduce the relation (3.62) since the Euler characteristic excludes these points. Taking a derivative we immediately find $\frac{dV_{g,m,n+1}\left(\vec{\alpha}_{n+1},\vec{b}_{m}\right)}{d\alpha_{n+1}}\bigg{\rvert}_{\alpha_{n+1}=1}=\left(-2\pi\int_{\Sigma/\\{z_{j}\\}}d^{2}z\sqrt{g}\right)\times V_{g,m,n}\left(\vec{\alpha}_{n},\vec{b}_{m}\right),$ (3.65) which is the desired recursion relation (3.62). Note that this argument also goes through if we replace the measure for the defect with $\pi\left(1-\alpha^{2}\right)\int d^{2}z\sqrt{g}$, as we found for the disk. From the above argument it might be suspected that the volumes also satisfy $\lim_{\alpha_{n}\to 1}V_{g,n,m}\left(\vec{\alpha}_{n},\vec{b}_{m}\right)\stackrel{{\scriptstyle?}}{{=}}0$. In [20] this was shown to be true when there are no geodesic boundaries, but is false when such boundaries are present. We do not have a gravity path integral argument for this. One possibility is that with geodesic boundaries there are additional boundary terms that enter into the measure through (3.58), where we assumed that all boundary terms vanished. ## 4 Discussion In this paper we studied various unresolved aspects of the gravitational path integral of JT gravity. We carried out the gauge fixing of the path integral in second order formalism for general hyperbolic surfaces with asymptotic boundaries and conical singularities. The second order formalism allowed us to clarify the procedure for calculating the proper measure for the conical defect operator, and resolved the question of which dilaton gravity potential should be used for JT gravity coupled to a gas of conical defects [17, 18, 19, 20]. This also allowed us to give a gravity path integral argument for certain recursion relations of Weil-Petersson volumes derived using algebraic geometry techniques [20]. An open problem is to carry out the full computation of the measure for the conical defect operator to all orders in the $(1-\alpha)$ expansion on a general surface to prove the conjectured form given in equation 1.9, which we were only able to fully compute on the disk topology.434343As explained in the introduction, our reasoning for extending the normalization of the operator found on the disk to arbitrary surfaces is that the operator should be surface independent. Along the way we computed determinants of Laplace operators on hyperbolic surfaces with asymptotic boundaries. These determinants are straightforwardly related to partition functions of matter fields minimally coupled to JT gravity. It would be interesting to better understand matter fields coupled to JT with a gas of conical defects, where the bulk geometry would not be pure AdS2. Tangentially, we computed the determinant of the vector Laplacian $\det(\Delta_{1}+s(s-1))$ on a cone geometry. This can be related to the bulk entanglement entropy of a gauge field on one side of the TFD. Since gauge fields in two dimensions have no propagating degrees of freedom a non-trivial entanglement entropy should arise from edge modes, and it would be interesting to understand this better in AdS2.444444We thank Sean Colin-Ellerin for discussion on this point. We obtained an exact expression for the two-point function of matter fields on an arbitrary surface $\Sigma$ obtainable through the quotient method. The correlator takes into account all geodesics on the surface including self- intersecting geodesics, but does not include an integration over the Schwarizan mode. In [53] diagrammatic rules were derived for correlation functions, including the appearance of $6j$ symbols when geodesics intersect, but this has yet to be derived from the second order perspective. It would be interesting to integrate over the boundary fluctuations and get a closed form expression for the full correlator reproducing the expected $6j$ symbols. This would also allow one to incorporate the contribution of self-intersecting geodesics to the late time two-point function calculation on the handle-disk, where it was argued in [25] that such contributions should decay with time. ### Acknowledgments We thank David Borthwick, Sean Colin-Ellerin, Luca Iliesiu, Geoff Penington, and Joaquin Turiaci for discussion. MU is supported in part by the NSF Graduate Research Fellowship Program under grant DGE1752814, by the National Science Foundation under Grants No. NSF PHY-1748958 and PHY-2309135, by the Berkeley Center for Theoretical Physics, by the DOE under award DE-SC0019380 and under the contract DE-AC02-05CH11231, by NSF grant PHY1820912, by the Heising-Simons Foundation, the Simons Foundation, and National Science Foundation Grant No. NSF PHY-1748958. ## Appendix A Consistency of disk measure with general surface In this section we explain how the exact calculation for a disk with a single defect is consistent with the measure derived for a general surface. Recall that we found the measure for the defect to be $\mathcal{N}\int d^{2}z_{i}\left|\frac{\langle\mu,\phi_{1}\rangle}{\sqrt{\langle\phi_{1},\phi_{1}\rangle}}\right|^{2}=\int d^{2}z_{i}~{}\times~{}\begin{cases}\pi\left(1-\alpha^{2}\right)\sqrt{g(z_{i})},&\text{disk,}\\\ 2\pi\left(1-\alpha\right)\sqrt{g(z_{i})}+\mathcal{O}\left(\left(1-\alpha\right)^{2}\right),&\text{general surface}.\end{cases}$ The point is that since the disk calculation is exact we implicitly re-summed the $\left(1-\alpha\right)$ corrections to the measure. We now make this explicit in the case that the defect is at the center. The tower of $\left(1-\alpha\right)$ corrections comes from the inner product for the quadratic differential $\langle\phi_{1},\phi_{1}\rangle_{\text{disk}}=\frac{4\pi}{1-\alpha^{2}}\frac{1}{\sqrt{g(0)}},\qquad\langle\phi_{1},\phi_{1}\rangle_{\text{general}}\mathrel{\mathop{=}\limits_{\alpha\to 1}}\frac{2\pi}{\left(1-\alpha\right)\sqrt{g(0)}}+\mathcal{O}\left(1\right).$ (A.1) We can express the metric for the disk with a conical defect in the same form we used for a general surface $ds^{2}=e^{2\omega}dzd\bar{z},\qquad\omega=\omega_{0}+\log\left(|z|^{\alpha-1}\right)-\log\left(\frac{1-|z|^{2\alpha}}{\alpha^{2}(1-|z|^{2})}\right),\qquad\omega_{0}=\log\left(\frac{2}{1-|z|^{2}}\right),$ (A.2) where $\omega_{0}$ is the Weyl factor for a disk without a defect. We can now calculate the inner product on the disk, doing a series expansion in $(1-\alpha)$ for the last term in the Weyl factor $\displaystyle\langle\phi_{1},\phi_{1}\rangle_{\text{disk}}$ $\displaystyle=\int d^{2}z\left(\frac{e^{2\omega_{0}}}{2}\right)^{-1}\frac{|z|^{2(1-\alpha)}}{|z|^{2}}\times\underbrace{\frac{(1-|z|^{2})^{2}}{\alpha^{2}(1-|z|^{2\alpha})^{2}}}_{\text{expand}},$ (A.3) $\displaystyle=\frac{2\pi}{(1-\alpha)\sqrt{g(0)}}+\frac{\pi}{\sqrt{g(0)}}+\mathcal{O}\left(1-\alpha\right),$ (A.4) $\displaystyle\stackrel{{\scriptstyle\tiny\text{re- sum}}}{{=}}\frac{4\pi}{(1-\alpha^{2})\sqrt{g(0)}}.$ (A.5) Which is in agreement with the expression for general surfaces in the $\left(1-\alpha\right)$ expansion, and the re-summation of the series reproduces the full disk answer (A.1). ## Appendix B Double trumpet gluing measure The Weil-Petersson measure for gluing two geodesic boundaries is famously given by $\int dbd\tau$ [54, 55], where $b\in[0,\infty)$ is the length of the geodesic boundary and $\tau\in[0,b]$ is a relative twist between the two boundaries being glued. In this appendix we will review how this measure can be derived using Beltrami differentials. ##### The Beltrami differential. We will derive the gluing measure for the double trumpet geometry which can be represented by the quotient of the UHP with metric $ds^{2}=\frac{dzd\overline{z}}{\left(\operatorname{Im}z\right)^{2}}$, where the geometry is represented by $\mathbb{H}/\langle T_{b}\rangle$ where $T_{b}\cdot z=e^{b}z$. The fundamental domain $\mathcal{F}$ is the region between the two semicircles $r=1$ and $r=e^{b}$. The main idea is to explicitly find the two quasiconformal deformations that infinitesimally deform the geodesic length $b$ and twist $\tau$, and calculate the measure from equation (3.19). The transformation that infinitesimally deforms the length $b\to b+\epsilon$ is given by the map $z\rightarrow f_{b}(z,\bar{z})=z(z\bar{z})^{\frac{\epsilon}{2b}}\,.$ (B.1) This map satisfies $\left|f_{b}({|z|{=}e^{mb}})\right|=e^{m(b+\epsilon)}$, meaning that semicircles with radius $e^{mb}$ are mapped to semicircles with radius $e^{m(b+\epsilon)}$. Thus this transformation increases the geodesic throat size of the double trumpet. For the twist the appropriate deformation is given by $z\rightarrow f_{\tau}(z,\bar{z})=ze^{\epsilon\Phi(\theta)}\,,\qquad\Phi(0)=0,~{}~{}\Phi(\pi)=1\,.$ (B.2) In the above $\Phi(\theta)$ is an arbitrary smooth function of the angle $z=re^{i\theta}$. The properties of $\Phi$ ensure that the left boundary of the double trumpet is smoothly infinitesimally twisted relative to the right boundary. For example, a point on the left boundary $z=-1$ is sent to $z=-(1+\epsilon)$. These maps are known as quasiconformal transformations.454545Quasiconformal transformations satisfy $|\bar{\partial}f(z,\bar{z})|<\partial f(z,\bar{z})$. A conformal transformation can be viewed as a special class of quasiconformal transformation with $\bar{\partial}f=0$. Geometrically, a conformal map preserves angles and maps small circles to circles; while a quasiconformal map does not preserve angles and maps small circles to ovals. Roughly speaking, the eccentricity of the ovals are given by the ratio $\bar{\partial}f/\partial f$. The transformation is known as an infinitesimaly quasiconformal transformation if this ratio can be made infinitesimal. Such transformations infinitesimally deform the metric in moduli space, since the new metric is no longer Weyl equivalent to the original metric. The Beltrami differential is then defined to be $\mu=\frac{\bar{\partial}f}{\partial f}$.464646Note that the definition of $\mu$ here is the complex conjugate of the definition from most of the math literature. Taking $\epsilon$ to be infinitesimal we can define an infinitesimal beltrami $\hat{\mu}=\lim_{\epsilon\rightarrow 0}\mu/\epsilon$ which through an abuse of notation is the definition of Beltrami differentials used in the main text. As explained in the main text, these infinitesimal beltrami differentials capture the infinitesimal change in the metric as we deform the moduli. We can directly calculate $\hat{\mu}$ from (B.1) and (B.2) and find $\hat{\mu}_{\tau}=\frac{iz}{2\bar{z}}\Phi^{\prime}(\theta)dz^{-1}d\bar{z}\,,\quad\hat{\mu}_{b}=\frac{z}{2b\bar{z}}dz^{-1}d\bar{z}\,,$ (B.3) where we have emphasized that these differentials should be thought of as $(-1,1)$ tensors. It can however be the case that a Beltrami differential does not correspond to a pure deformation of the moduli, for example it can deform the metric by a Weyl transformation. It is thus customary to act on Beltrami differentials by a projection operator that restricts them to pure moduli deformations. This does not affect the calculation of the path integral measure, but ensures that the differentials live in the tangent space to moduli space. The projection is orginally given by [56] $P[\hat{\mu}](z)\equiv\frac{6}{\pi}(\operatorname{Im}z)^{2}\int_{\mathbb{H}}d^{2}\xi\frac{\overline{\hat{\mu}}(\xi,\bar{\xi})}{({\xi}-\bar{z})^{4}}\,.$ (B.4) This gives us the projected differentials474747Note that $P[\hat{\mu}_{\tau}]$ is independent of $\Phi^{\prime}(\theta)$, which shows that the apparent entire function’s worth of degrees of freedom in $\Phi$ do not correspond to genuine moduli deformations. $P[\hat{\mu}_{b}]=\frac{(\text{Im}z)^{2}}{{z}^{2}b}dz^{-1}d\bar{z}\,,\quad P[\hat{\mu}_{\tau}]=\frac{i(\text{Im}z)^{2}}{\pi{z}^{2}}dz^{-1}d\bar{z}.$ (B.5) One remarkable property of this projection is $\langle\hat{\mu},\phi\rangle=\langle P[\hat{\mu}],\phi\rangle$ for any quadratic differential $\phi$, and since we are primarily interested in inner products we use the same notation $\hat{\mu}$ to represent the Beltrami and its projection below. This identity is the statement that the operator projects out the subset of deformations in $\hat{\mu}$ that do not deform the moduli. ##### Quadratic differential. To compute the metric we also need the quadratic differentials, which are thought of as forming the cotangent space to moduli space. These differentials are quite challenging to compute and were found by Wolpert [55], we quote the final result484848Note that our convention here is a bit different from Wolpert’s convention. The differences are (1) our definition of the $\langle\mu,\phi\rangle$ inner product is twice as Wolpert’s and (2) the integral measure on Riemann surface $d^{z}$ is 2 times the Euclidean measure $dxdy$ used by Wolpert. $\phi_{b}=\frac{1}{\pi z^{2}}dz^{2}\,,\qquad\phi_{\tau}=\frac{i}{bz^{2}}dz^{2}.$ (B.6) They have the following inner products with the beltrami differentials $\mu_{b},\mu_{\tau}$494949We slightly change the definition of the inner product by taking the real part, since both $db$ and $d\tau$ are real variables. $\langle\mu_{b},\phi_{b}\rangle\equiv 2\text{Re}\int_{\mathcal{F}}\mu_{b}\bar{\phi}_{b}=\frac{2}{\pi b}\int_{\mathcal{F}}\frac{(\text{Im}\xi)^{2}}{\xi^{2}\bar{\xi}^{2}}d^{2}\xi=2\,,\qquad\langle\mu_{\tau},\phi_{\tau}\rangle=2,$ (B.7) where we have used the inner product in (3.16). It is straight forward to check that the off-diagonal inner products vanish $\langle\mu_{\tau},\phi_{b}\rangle=0$ because of the factor $i$ in the $\mu_{\tau}$ expression. ##### The Weil-Petersson metric. Given these results the Weil-Petersson metric for the moduli space is straightforward to obtain. From the discussion in Section 3, we need to evaluate several determinants. It is straight forward to calculate $\det\langle\mu,\phi\rangle=\det\begin{pmatrix}\langle\mu_{b},\phi_{b}\rangle&\langle\mu_{b},\phi_{\tau}\rangle\\\ \langle\mu_{\tau},\phi_{b}\rangle&\langle\mu_{\tau},\phi_{\tau}\rangle\end{pmatrix}=\det\begin{pmatrix}2&0\\\ 0&2\end{pmatrix}=4\,.$ (B.8) $\det\langle\phi,\phi\rangle=\det\begin{pmatrix}\langle\phi_{b},\phi_{b}\rangle&\langle\phi_{b},\phi_{\tau}\rangle\\\ \langle\phi_{\tau},\phi_{b}\rangle&\langle\phi_{\tau},\phi_{\tau}\rangle\end{pmatrix}=\det\begin{pmatrix}2b&0\\\ 0&2b^{-1}\end{pmatrix}=4\,.$ (B.9) The measure is given by equation (3.19) $d\mu_{\text{DT}}=\mathcal{N}\frac{\det\langle\mu,\phi\rangle\det\langle\bar{\mu},\bar{\phi}\rangle}{\det\langle\phi,\phi\rangle}db\,d\tau=4\mathcal{N}db\,d\tau\,,$ (B.10) where the integral ranges over $\tau\in[0,b]$ and $b\in[0,\infty)$. To get the standard form of the gluing measure $\int bdb$ we must choose the normalization $\mathcal{N}=1/4$ for the measure (3.3) as claimed in the main text. ## Appendix C The determinant calculation ### Notation, conventions, and summary of results We first list the notation and main results of the determinant calculation. We denote the surface obtained by a quotient of the Fuchsian group $\Gamma$ as $\Sigma=\mathbb{H}/\Gamma$. The fundamental domain is denoted by $\mathcal{F}_{\Sigma}$ which is thought of as a subregion of $\mathbb{H}$. We will sometimes not distinguish between $\mathcal{F}_{\Sigma}$ and $\Sigma$ for simplicity. We define the Fuchsian group action to be $T(Z)$ where $T$ is a Fuchsian group element and $Z$ will either be a subregion or a point in $\mathbb{H}$. An important geometrical invariant of the surface $\Sigma$ is the hyperbolic area $|\Sigma|=-\left(\frac{1}{2}\int_{\Sigma/\\{z_{i}\\}}d^{2}z\sqrt{g}R+\int_{\partial\Sigma}\sqrt{h}K\right)=2\pi\bigg{(}2g+n-2+\sum_{j=1}^{k}\left(1-n_{j}^{-1}\right)\bigg{)},$ (C.1) where $g$ is the genus, $n$ is the number of boundaries (including cusps and asymptotic boundaries), and $k$ is the number of defects located at positions $z_{i}$ with deficit angles $\frac{2\pi}{n_{j}}$. For simplicity we exclude surfaces with geodesic boundaries and cusps ($2\pi$ deficit angles) from the determinant calculation. The area element of $\mathbb{H}$ is $d^{2}z\sqrt{g}=y^{-2}dxdy$, and the $u$ variable for hyperbolic distance $u=\cosh^{2}\left(\frac{\ell(z,z^{\prime})}{2}\right)=\frac{1}{2}\left(1+\cosh(\ell(z,z^{\prime}))\right)=\frac{(x-x^{\prime})^{2}+(y+y^{\prime})^{2}}{4yy^{\prime}}\,,$ (C.2) where our complex coordinates are given by $z=x+iy$. We will be interested in the scalar and vector Laplacians. The scalar Laplacian is defined as $-g^{ab}\nabla_{a}\nabla_{b}$ acting on scalar functions. Similarly, the vector Laplacian is $\nabla_{1}^{2}=-g^{ab}\nabla_{a}\nabla_{b}$ acting on two component vectors or covectors, and was introduced in the main text through $\nabla_{1}^{2}=\frac{1}{2}P_{1}^{\dagger}P_{1}-1$. It will be more convenient to conjugate the vector Laplacian and consider $y^{-2}\nabla_{1}^{2}y^{2}+s(s-1)-1=\text{diag}\left(D_{-1}+s(s-1),D_{1}+s(s-1)\right)$ where we define the operators $D_{\pm 1}=-\left(\partial_{x}\pm i\partial_{y}\right)y^{2}\left(\partial_{x}\mp i\partial_{y}\right),$ (C.3) where the conjugated operators acts on vectors with top component $v^{x}+iv^{y}$ and bottom component $v^{x}-iv^{y}$. Conjugated operators have the same spectrum, and so we can equivalently compute the determinant of the above operators. The spectra of $D_{\pm 1}$ are identical, and so we find the result that $\det(\Delta_{1}+1)=\det\left(D_{1}+2\right)^{2}$. The above immediately implies that $\det(\Delta_{1}+s(s-1)-1)=\det\left(D_{1}+s(s-1)\right)^{2}$. We will abuse notation in this appendix and refer to $D_{\pm 1}$ as the vector Laplacian. In the calculation of the determinant we will need the resolvent of the associated operators, also known as the propagator, denoted by $R^{0}_{\Sigma}$ and $R^{1}_{\Sigma}$, and their associated traces over the surface ${\mathrm{t}r}R_{\Sigma}=\int_{\Sigma}d^{2}z\sqrt{g}R_{\Sigma}(z,z)$. Note that the operator, the resolvent, and the trace are defined for a specific surface $\Sigma$. The trace is an integral over the fundamental domain $\mathcal{F}_{\Sigma}\subset\mathbb{H}$. The main result for the determinants are given as follows 505050One can also allow for $n_{c}$ cusps on the surface, which give a parabolic contribution $Z_{\rm par.}=\left(\Gamma(s-1/2)(2s-1)2^{s-1}\right)^{n_{c}}$. With the inclusion of cusps, the scalar determinant would include $Z_{\rm par.}$, namely $\det\left(\ldots\right)\to\det\left(\ldots\right)\times Z_{\rm par.}$ [36]. We believe a similar formula holds for the vector determinant. $\det\left({\Delta}_{0}+s(s-1)\right)={Z_{\text{hyp.}}(s)}{Z_{\text{ell.}}(s)}G_{\infty}(s)^{-\frac{|\Sigma|}{2\pi}}e^{-Bs(s-1)+D}\,.$ (C.4)
PHENIX Collaboration # Nonprompt direct-photon production in Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV U.A. Acharya Georgia State University, Atlanta, Georgia 30303, USA A. Adare University of Colorado, Boulder, Colorado 80309, USA C. Aidala Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA N.N. Ajitanand Deceased Chemistry Department, Stony Brook University, SUNY, Stony Brook, New York 11794-3400, USA Y. Akiba<EMAIL_ADDRESS>RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M. Alfred Department of Physics and Astronomy, Howard University, Washington, DC 20059, USA N. Apadula Iowa State University, Ames, Iowa 50011, USA Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA H. Asano Kyoto University, Kyoto 606-8502, Japan RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan B. Azmoun Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA V. Babintsev IHEP Protvino, State Research Center of Russian Federation, Institute for High Energy Physics, Protvino, 142281, Russia M. Bai Collider-Accelerator Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA N.S. Bandara Department of Physics, University of Massachusetts, Amherst, Massachusetts 01003-9337, USA B. Bannier Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA K.N. Barish University of California-Riverside, Riverside, California 92521, USA S. Bathe Baruch College, City University of New York, New York, New York, 10010 USA RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA A. Bazilevsky Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M. Beaumier University of California-Riverside, Riverside, California 92521, USA S. Beckman University of Colorado, Boulder, Colorado 80309, USA R. Belmont University of Colorado, Boulder, Colorado 80309, USA Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA Physics and Astronomy Department, University of North Carolina at Greensboro, Greensboro, North Carolina 27412, USA A. Berdnikov Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia Y. Berdnikov Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia L. Bichon Vanderbilt University, Nashville, Tennessee 37235, USA B. Blankenship Vanderbilt University, Nashville, Tennessee 37235, USA D.S. Blau National Research Center “Kurchatov Institute”, Moscow, 123098 Russia National Research Nuclear University, MEPhI, Moscow Engineering Physics Institute, Moscow, 115409, Russia J.S. Bok New Mexico State University, Las Cruces, New Mexico 88003, USA V. Borisov Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia K. Boyle RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M.L. Brooks Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA J. Bryslawskyj Baruch College, City University of New York, New York, New York, 10010 USA University of California-Riverside, Riverside, California 92521, USA V. Bumazhnov IHEP Protvino, State Research Center of Russian Federation, Institute for High Energy Physics, Protvino, 142281, Russia S. Campbell Columbia University, New York, New York 10027 and Nevis Laboratories, Irvington, New York 10533, USA Iowa State University, Ames, Iowa 50011, USA V. Canoa Roman Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA C.-H. Chen RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M. Chiu Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA C.Y. Chi Columbia University, New York, New York 10027 and Nevis Laboratories, Irvington, New York 10533, USA I.J. Choi University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA J.B. Choi Deceased Jeonbuk National University, Jeonju, 54896, Korea T. Chujo Tomonaga Center for the History of the Universe, University of Tsukuba, Tsukuba, Ibaraki 305, Japan Z. Citron Weizmann Institute, Rehovot 76100, Israel M. Connors Georgia State University, Atlanta, Georgia 30303, USA R. Corliss Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA Y. Corrales Morales Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA M. Csanád ELTE, Eötvös Loránd University, H-1117 Budapest, Pázmány P. s. 1/A, Hungary T. Csörgő Institute for Particle and Nuclear Physics, Wigner Research Centre for Physics, Hungarian Academy of Sciences (Wigner RCP, RMKI) H-1525 Budapest 114, POBox 49, Budapest, Hungary T.W. Danley Department of Physics and Astronomy, Ohio University, Athens, Ohio 45701, USA A. Datta University of New Mexico, Albuquerque, New Mexico 87131, USA M.S. Daugherity Abilene Christian University, Abilene, Texas 79699, USA G. David Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA C.T. Dean Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA K. DeBlasio University of New Mexico, Albuquerque, New Mexico 87131, USA K. Dehmelt Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA A. Denisov IHEP Protvino, State Research Center of Russian Federation, Institute for High Energy Physics, Protvino, 142281, Russia A. Deshpande RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA E.J. Desmond Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA A. Dion Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA P.B. Diss University of Maryland, College Park, Maryland 20742, USA J.H. Do Yonsei University, IPAP, Seoul 120-749, Korea V. Doomra Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA A. Drees Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA K.A. Drees Collider-Accelerator Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA J.M. Durham Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA A. Durum IHEP Protvino, State Research Center of Russian Federation, Institute for High Energy Physics, Protvino, 142281, Russia A. Enokizono RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Physics Department, Rikkyo University, 3-34-1 Nishi-Ikebukuro, Toshima, Tokyo 171-8501, Japan R. Esha Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA B. Fadem Muhlenberg College, Allentown, Pennsylvania 18104-5586, USA W. Fan Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA N. Feege Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA D.E. Fields University of New Mexico, Albuquerque, New Mexico 87131, USA M. Finger, Jr Charles University, Ovocný trh 5, Praha 1, 116 36, Prague, Czech Republic M. Finger Charles University, Ovocný trh 5, Praha 1, 116 36, Prague, Czech Republic D. Fitzgerald Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA S.L. Fokin National Research Center “Kurchatov Institute”, Moscow, 123098 Russia J.E. Frantz Department of Physics and Astronomy, Ohio University, Athens, Ohio 45701, USA A. Franz Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA A.D. Frawley Florida State University, Tallahassee, Florida 32306, USA P. Gallus Czech Technical University, Zikova 4, 166 36 Prague 6, Czech Republic C. Gal Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA P. Garg Department of Physics, Banaras Hindu University, Varanasi 221005, India Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA H. Ge Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA M. Giles Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA F. Giordano University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA A. Glenn Lawrence Livermore National Laboratory, Livermore, California 94550, USA Y. Goto RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA N. Grau Department of Physics, Augustana University, Sioux Falls, South Dakota 57197, USA S.V. Greene Vanderbilt University, Nashville, Tennessee 37235, USA M. Grosse Perdekamp University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA T. Gunji Center for Nuclear Study, Graduate School of Science, University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-0033, Japan T. Guo Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA T. Hachiya Nara Women’s University, Kita- uoya Nishi-machi Nara 630-8506, Japan RIKEN Nishina Center for Accelerator- Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA J.S. Haggerty Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA K.I. Hahn Ewha Womans University, Seoul 120-750, Korea H. Hamagaki Center for Nuclear Study, Graduate School of Science, University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-0033, Japan H.F. Hamilton Abilene Christian University, Abilene, Texas 79699, USA J. Hanks Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA S.Y. Han Ewha Womans University, Seoul 120-750, Korea Korea University, Seoul 02841, Korea M. Harvey Texas Southern University, Houston, TX 77004, USA S. Hasegawa Advanced Science Research Center, Japan Atomic Energy Agency, 2-4 Shirakata Shirane, Tokai-mura, Naka-gun, Ibaraki-ken 319-1195, Japan T.O.S. Haseler Georgia State University, Atlanta, Georgia 30303, USA K. Hashimoto RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Physics Department, Rikkyo University, 3-34-1 Nishi-Ikebukuro, Toshima, Tokyo 171-8501, Japan T.K. Hemmick Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA X. He Georgia State University, Atlanta, Georgia 30303, USA J.C. Hill Iowa State University, Ames, Iowa 50011, USA A. Hodges Georgia State University, Atlanta, Georgia 30303, USA R.S. Hollis University of California-Riverside, Riverside, California 92521, USA K. Homma Hiroshima University, Kagamiyama, Higashi-Hiroshima 739-8526, Japan B. Hong Korea University, Seoul 02841, Korea T. Hoshino Hiroshima University, Kagamiyama, Higashi-Hiroshima 739-8526, Japan N. Hotvedt Iowa State University, Ames, Iowa 50011, USA J. Huang Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA K. Imai Advanced Science Research Center, Japan Atomic Energy Agency, 2-4 Shirakata Shirane, Tokai-mura, Naka-gun, Ibaraki-ken 319-1195, Japan M. Inaba Tomonaga Center for the History of the Universe, University of Tsukuba, Tsukuba, Ibaraki 305, Japan A. Iordanova University of California-Riverside, Riverside, California 92521, USA D. Isenhower Abilene Christian University, Abilene, Texas 79699, USA D. Ivanishchev PNPI, Petersburg Nuclear Physics Institute, Gatchina, Leningrad region, 188300, Russia B.V. Jacak Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA M. Jezghani Georgia State University, Atlanta, Georgia 30303, USA X. Jiang Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA Z. Ji Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA B.M. Johnson Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Georgia State University, Atlanta, Georgia 30303, USA D. Jouan IPN-Orsay, Univ. Paris-Sud, CNRS/IN2P3, Université Paris-Saclay, BP1, F-91406, Orsay, France D.S. Jumper University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA S. Kanda Center for Nuclear Study, Graduate School of Science, University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-0033, Japan J.H. Kang Yonsei University, IPAP, Seoul 120-749, Korea D. Kawall Department of Physics, University of Massachusetts, Amherst, Massachusetts 01003-9337, USA A.V. Kazantsev National Research Center “Kurchatov Institute”, Moscow, 123098 Russia J.A. Key University of New Mexico, Albuquerque, New Mexico 87131, USA V. Khachatryan Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA A. Khanzadeev PNPI, Petersburg Nuclear Physics Institute, Gatchina, Leningrad region, 188300, Russia A. Khatiwada Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA B. Kimelman Muhlenberg College, Allentown, Pennsylvania 18104-5586, USA C. Kim Korea University, Seoul 02841, Korea D.J. Kim Helsinki Institute of Physics and University of Jyväskylä, P.O.Box 35, FI-40014 Jyväskylä, Finland E.-J. Kim Jeonbuk National University, Jeonju, 54896, Korea G.W. Kim Ewha Womans University, Seoul 120-750, Korea M. Kim Department of Physics and Astronomy, Seoul National University, Seoul 151-742, Korea T. Kim Ewha Womans University, Seoul 120-750, Korea D. Kincses ELTE, Eötvös Loránd University, H-1117 Budapest, Pázmány P. s. 1/A, Hungary A. Kingan Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA E. Kistenev Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA R. Kitamura Center for Nuclear Study, Graduate School of Science, University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-0033, Japan J. Klatsky Florida State University, Tallahassee, Florida 32306, USA D. Kleinjan University of California-Riverside, Riverside, California 92521, USA P. Kline Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA T. Koblesky University of Colorado, Boulder, Colorado 80309, USA B. Komkov PNPI, Petersburg Nuclear Physics Institute, Gatchina, Leningrad region, 188300, Russia D. Kotov PNPI, Petersburg Nuclear Physics Institute, Gatchina, Leningrad region, 188300, Russia Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia L. Kovacs ELTE, Eötvös Loránd University, H-1117 Budapest, Pázmány P. s. 1/A, Hungary K. Kurita Physics Department, Rikkyo University, 3-34-1 Nishi-Ikebukuro, Toshima, Tokyo 171-8501, Japan M. Kurosawa RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Y. Kwon Yonsei University, IPAP, Seoul 120-749, Korea J.G. Lajoie Iowa State University, Ames, Iowa 50011, USA D. Larionova Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia A. Lebedev Iowa State University, Ames, Iowa 50011, USA S. Lee Yonsei University, IPAP, Seoul 120-749, Korea S.H. Lee Iowa State University, Ames, Iowa 50011, USA Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA M.J. Leitch Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA N.A. Lewis Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA S.H. Lim Pusan National University, Pusan 46241, Korea Yonsei University, IPAP, Seoul 120-749, Korea M.X. Liu Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA X. Li Science and Technology on Nuclear Data Laboratory, China Institute of Atomic Energy, Beijing 102413, People’s Republic of China X. Li Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA D.A. Loomis Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA D. Lynch Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA S. Lökös ELTE, Eötvös Loránd University, H-1117 Budapest, Pázmány P. s. 1/A, Hungary T. Majoros Debrecen University, H-4010 Debrecen, Egyetem tér 1, Hungary Y.I. Makdisi Collider-Accelerator Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M. Makek Department of Physics, Faculty of Science, University of Zagreb, Bijenička c. 32 HR-10002 Zagreb, Croatia A. Manion Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA V.I. Manko National Research Center “Kurchatov Institute”, Moscow, 123098 Russia E. Mannel Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M. McCumber Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA P.L. McGaughey Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA D. McGlinchey University of Colorado, Boulder, Colorado 80309, USA Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA C. McKinney University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA A. Meles New Mexico State University, Las Cruces, New Mexico 88003, USA M. Mendoza University of California-Riverside, Riverside, California 92521, USA A.C. Mignerey University of Maryland, College Park, Maryland 20742, USA A. Milov Weizmann Institute, Rehovot 76100, Israel D.K. Mishra Bhabha Atomic Research Centre, Bombay 400 085, India J.T. Mitchell Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M. Mitrankova Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia Iu. Mitrankov Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia S. Miyasaka RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Department of Physics, Tokyo Institute of Technology, Oh-okayama, Meguro, Tokyo 152-8551, Japan S. Mizuno RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Tomonaga Center for the History of the Universe, University of Tsukuba, Tsukuba, Ibaraki 305, Japan A.K. Mohanty Bhabha Atomic Research Centre, Bombay 400 085, India M.M. Mondal Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA P. Montuenga University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA T. Moon Korea University, Seoul 02841, Korea Yonsei University, IPAP, Seoul 120-749, Korea D.P. Morrison Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA T.V. Moukhanova National Research Center “Kurchatov Institute”, Moscow, 123098 Russia B. Mulilo Korea University, Seoul 02841, Korea RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Department of Physics, School of Natural Sciences, University of Zambia, Great East Road Campus, Box 32379, Lusaka, Zambia T. Murakami Kyoto University, Kyoto 606-8502, Japan RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan J. Murata RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Physics Department, Rikkyo University, 3-34-1 Nishi-Ikebukuro, Toshima, Tokyo 171-8501, Japan A. Mwai Chemistry Department, Stony Brook University, SUNY, Stony Brook, New York 11794-3400, USA K. Nagashima Hiroshima University, Kagamiyama, Higashi- Hiroshima 739-8526, Japan J.L. Nagle University of Colorado, Boulder, Colorado 80309, USA M.I. Nagy ELTE, Eötvös Loránd University, H-1117 Budapest, Pázmány P. s. 1/A, Hungary I. Nakagawa RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA H. Nakagomi RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Tomonaga Center for the History of the Universe, University of Tsukuba, Tsukuba, Ibaraki 305, Japan K. Nakano RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Department of Physics, Tokyo Institute of Technology, Oh-okayama, Meguro, Tokyo 152-8551, Japan C. Nattrass University of Tennessee, Knoxville, Tennessee 37996, USA S. Nelson Florida A&M University, Tallahassee, FL 32307, USA P.K. Netrakanti Bhabha Atomic Research Centre, Bombay 400 085, India T. Niida Tomonaga Center for the History of the Universe, University of Tsukuba, Tsukuba, Ibaraki 305, Japan S. Nishimura Center for Nuclear Study, Graduate School of Science, University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-0033, Japan R. Nouicer Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA N. Novitzky Helsinki Institute of Physics and University of Jyväskylä, P.O.Box 35, FI-40014 Jyväskylä, Finland Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA Tomonaga Center for the History of the Universe, University of Tsukuba, Tsukuba, Ibaraki 305, Japan T. Novák Eszterházy Károly University, Károly Róbert Campus, H-3200 Gyöngyös, Mátrai út 36, Hungary Institute for Particle and Nuclear Physics, Wigner Research Centre for Physics, Hungarian Academy of Sciences (Wigner RCP, RMKI) H-1525 Budapest 114, POBox 49, Budapest, Hungary G. Nukazuka RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA A.S. Nyanin National Research Center “Kurchatov Institute”, Moscow, 123098 Russia E. O’Brien Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA C.A. Ogilvie Iowa State University, Ames, Iowa 50011, USA J.D. Orjuela Koop University of Colorado, Boulder, Colorado 80309, USA J.D. Osborn Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831, USA A. Oskarsson Department of Physics, Lund University, Box 118, SE-221 00 Lund, Sweden K. Ozawa KEK, High Energy Accelerator Research Organization, Tsukuba, Ibaraki 305-0801, Japan Tomonaga Center for the History of the Universe, University of Tsukuba, Tsukuba, Ibaraki 305, Japan R. Pak Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA V. Pantuev Institute for Nuclear Research of the Russian Academy of Sciences, prospekt 60-letiya Oktyabrya 7a, Moscow 117312, Russia V. Papavassiliou New Mexico State University, Las Cruces, New Mexico 88003, USA J.S. Park Department of Physics and Astronomy, Seoul National University, Seoul 151-742, Korea S. Park Department of Physics and Astronomy, Seoul National University, Seoul 151-742, Korea Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA M. Patel Iowa State University, Ames, Iowa 50011, USA S.F. Pate New Mexico State University, Las Cruces, New Mexico 88003, USA J.-C. Peng University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA W. Peng Vanderbilt University, Nashville, Tennessee 37235, USA D.V. Perepelitsa Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA University of Colorado, Boulder, Colorado 80309, USA G.D.N. Perera New Mexico State University, Las Cruces, New Mexico 88003, USA D.Yu. Peressounko National Research Center “Kurchatov Institute”, Moscow, 123098 Russia C.E. PerezLara Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA J. Perry Iowa State University, Ames, Iowa 50011, USA R. Petti Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA C. Pinkenburg Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA R. Pinson Abilene Christian University, Abilene, Texas 79699, USA R.P. Pisani Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M. Potekhin Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA A. Pun Department of Physics and Astronomy, Ohio University, Athens, Ohio 45701, USA M.L. Purschke Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA P.V. Radzevich Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia J. Rak Helsinki Institute of Physics and University of Jyväskylä, P.O.Box 35, FI-40014 Jyväskylä, Finland N. Ramasubramanian Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA B.J. Ramson Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA I. Ravinovich Weizmann Institute, Rehovot 76100, Israel K.F. Read Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831, USA University of Tennessee, Knoxville, Tennessee 37996, USA D. Reynolds Chemistry Department, Stony Brook University, SUNY, Stony Brook, New York 11794-3400, USA V. Riabov National Research Nuclear University, MEPhI, Moscow Engineering Physics Institute, Moscow, 115409, Russia PNPI, Petersburg Nuclear Physics Institute, Gatchina, Leningrad region, 188300, Russia Y. Riabov PNPI, Petersburg Nuclear Physics Institute, Gatchina, Leningrad region, 188300, Russia Saint Petersburg State Polytechnic University, St. Petersburg, 195251 Russia D. Richford Baruch College, City University of New York, New York, New York, 10010 USA T. Rinn University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA Iowa State University, Ames, Iowa 50011, USA S.D. Rolnick University of California-Riverside, Riverside, California 92521, USA M. Rosati Iowa State University, Ames, Iowa 50011, USA Z. Rowan Baruch College, City University of New York, New York, New York, 10010 USA J.G. Rubin Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA J. Runchey Iowa State University, Ames, Iowa 50011, USA B. Sahlmueller Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA N. Saito KEK, High Energy Accelerator Research Organization, Tsukuba, Ibaraki 305-0801, Japan T. Sakaguchi Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA H. Sako Advanced Science Research Center, Japan Atomic Energy Agency, 2-4 Shirakata Shirane, Tokai-mura, Naka-gun, Ibaraki-ken 319-1195, Japan V. Samsonov National Research Nuclear University, MEPhI, Moscow Engineering Physics Institute, Moscow, 115409, Russia PNPI, Petersburg Nuclear Physics Institute, Gatchina, Leningrad region, 188300, Russia M. Sarsour Georgia State University, Atlanta, Georgia 30303, USA S. Sato Advanced Science Research Center, Japan Atomic Energy Agency, 2-4 Shirakata Shirane, Tokai-mura, Naka-gun, Ibaraki-ken 319-1195, Japan B. Schaefer Vanderbilt University, Nashville, Tennessee 37235, USA B.K. Schmoll University of Tennessee, Knoxville, Tennessee 37996, USA K. Sedgwick University of California-Riverside, Riverside, California 92521, USA R. Seidl RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA A. Sen Iowa State University, Ames, Iowa 50011, USA University of Tennessee, Knoxville, Tennessee 37996, USA R. Seto University of California-Riverside, Riverside, California 92521, USA P. Sett Bhabha Atomic Research Centre, Bombay 400 085, India A. Sexton University of Maryland, College Park, Maryland 20742, USA D. Sharma Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA I. Shein IHEP Protvino, State Research Center of Russian Federation, Institute for High Energy Physics, Protvino, 142281, Russia T.-A. Shibata RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan Department of Physics, Tokyo Institute of Technology, Oh-okayama, Meguro, Tokyo 152-8551, Japan K. Shigaki Hiroshima University, Kagamiyama, Higashi-Hiroshima 739-8526, Japan M. Shimomura Iowa State University, Ames, Iowa 50011, USA Nara Women’s University, Kita-uoya Nishi-machi Nara 630-8506, Japan P. Shukla Bhabha Atomic Research Centre, Bombay 400 085, India A. Sickles Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA C.L. Silva Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA D. Silvermyr Department of Physics, Lund University, Box 118, SE-221 00 Lund, Sweden Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831, USA B.K. Singh Department of Physics, Banaras Hindu University, Varanasi 221005, India C.P. Singh Department of Physics, Banaras Hindu University, Varanasi 221005, India V. Singh Department of Physics, Banaras Hindu University, Varanasi 221005, India M. Slunečka Charles University, Ovocný trh 5, Praha 1, 116 36, Prague, Czech Republic K.L. Smith Florida State University, Tallahassee, Florida 32306, USA M. Snowball Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA R.A. Soltz Lawrence Livermore National Laboratory, Livermore, California 94550, USA W.E. Sondheim Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA S.P. Sorensen University of Tennessee, Knoxville, Tennessee 37996, USA I.V. Sourikova Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA P.W. Stankus Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831, USA M. Stepanov Deceased Department of Physics, University of Massachusetts, Amherst, Massachusetts 01003-9337, USA S.P. Stoll Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA T. Sugitate Hiroshima University, Kagamiyama, Higashi-Hiroshima 739-8526, Japan A. Sukhanov Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA T. Sumita RIKEN Nishina Center for Accelerator- Based Science, Wako, Saitama 351-0198, Japan J. Sun Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA Z. Sun Debrecen University, H-4010 Debrecen, Egyetem tér 1, Hungary J. Sziklai Institute for Particle and Nuclear Physics, Wigner Research Centre for Physics, Hungarian Academy of Sciences (Wigner RCP, RMKI) H-1525 Budapest 114, POBox 49, Budapest, Hungary A. Taketani RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA K. Tanida Advanced Science Research Center, Japan Atomic Energy Agency, 2-4 Shirakata Shirane, Tokai-mura, Naka-gun, Ibaraki-ken 319-1195, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Department of Physics and Astronomy, Seoul National University, Seoul 151-742, Korea M.J. Tannenbaum Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA S. Tarafdar Vanderbilt University, Nashville, Tennessee 37235, USA Weizmann Institute, Rehovot 76100, Israel A. Taranenko National Research Nuclear University, MEPhI, Moscow Engineering Physics Institute, Moscow, 115409, Russia Chemistry Department, Stony Brook University, SUNY, Stony Brook, New York 11794-3400, USA R. Tieulent Georgia State University, Atlanta, Georgia 30303, USA IPNL, CNRS/IN2P3, Univ Lyon, Université Lyon 1, F-69622, Villeurbanne, France A. Timilsina Iowa State University, Ames, Iowa 50011, USA T. Todoroki RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Tomonaga Center for the History of the Universe, University of Tsukuba, Tsukuba, Ibaraki 305, Japan M. Tomášek Czech Technical University, Zikova 4, 166 36 Prague 6, Czech Republic C.L. Towell Abilene Christian University, Abilene, Texas 79699, USA R. Towell Abilene Christian University, Abilene, Texas 79699, USA R.S. Towell Abilene Christian University, Abilene, Texas 79699, USA I. Tserruya Weizmann Institute, Rehovot 76100, Israel Y. Ueda Hiroshima University, Kagamiyama, Higashi- Hiroshima 739-8526, Japan B. Ujvari Debrecen University, H-4010 Debrecen, Egyetem tér 1, Hungary H.W. van Hecke Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA J. Velkovska Vanderbilt University, Nashville, Tennessee 37235, USA M. Virius Czech Technical University, Zikova 4, 166 36 Prague 6, Czech Republic V. Vrba Czech Technical University, Zikova 4, 166 36 Prague 6, Czech Republic Institute of Physics, Academy of Sciences of the Czech Republic, Na Slovance 2, 182 21 Prague 8, Czech Republic X.R. Wang New Mexico State University, Las Cruces, New Mexico 88003, USA RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Y. Watanabe RIKEN Nishina Center for Accelerator-Based Science, Wako, Saitama 351-0198, Japan RIKEN BNL Research Center, Brookhaven National Laboratory, Upton, New York 11973-5000, USA Y.S. Watanabe Center for Nuclear Study, Graduate School of Science, University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-0033, Japan KEK, High Energy Accelerator Research Organization, Tsukuba, Ibaraki 305-0801, Japan F. Wei New Mexico State University, Las Cruces, New Mexico 88003, USA A.S. White Department of Physics, University of Michigan, Ann Arbor, Michigan 48109-1040, USA C.P. Wong Georgia State University, Atlanta, Georgia 30303, USA Los Alamos National Laboratory, Los Alamos, New Mexico 87545, USA C.L. Woody Physics Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA M. Wysocki Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831, USA B. Xia Department of Physics and Astronomy, Ohio University, Athens, Ohio 45701, USA L. Xue Georgia State University, Atlanta, Georgia 30303, USA S. Yalcin Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA Y.L. Yamaguchi Center for Nuclear Study, Graduate School of Science, University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-0033, Japan Department of Physics and Astronomy, Stony Brook University, SUNY, Stony Brook, New York 11794-3800, USA A. Yanovich IHEP Protvino, State Research Center of Russian Federation, Institute for High Energy Physics, Protvino, 142281, Russia I. Yoon Department of Physics and Astronomy, Seoul National University, Seoul 151-742, Korea J.H. Yoo Korea University, Seoul 02841, Korea I.E. Yushmanov National Research Center “Kurchatov Institute”, Moscow, 123098 Russia H. Yu New Mexico State University, Las Cruces, New Mexico 88003, USA Peking University, Beijing 100871, People’s Republic of China W.A. Zajc Columbia University, New York, New York 10027 and Nevis Laboratories, Irvington, New York 10533, USA A. Zelenski Collider-Accelerator Department, Brookhaven National Laboratory, Upton, New York 11973-5000, USA S. Zhou Science and Technology on Nuclear Data Laboratory, China Institute of Atomic Energy, Beijing 102413, People’s Republic of China L. Zou University of California- Riverside, Riverside, California 92521, USA ###### Abstract The measurement of the direct-photon spectrum from Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV is presented by the PHENIX collaboration using the external-photon-conversion technique for 0%–93% central collisions in a transverse-momentum ($p_{T}$) range of 0.8–10 GeV/$c$. An excess of direct photons, above prompt-photon production from hard-scattering processes, is observed for $p_{T}<6$ GeV/$c$. Nonprompt direct photons are measured by subtracting the prompt component, which is estimated as $N_{\rm coll}$-scaled direct photons from $p$$+$$p$ collisions at 200 GeV, from the direct-photon spectrum. Results are obtained for $0.8<p_{T}<6.0$ GeV/$c$ and suggest that the spectrum has an increasing inverse slope from ${\approx}0.2$ to 0.4 GeV/$c$ with increasing $p_{T}$, which indicates a possible sensitivity of the measurement to photons from earlier stages of the evolution of the collision. In addition, like the direct-photon production, the $p_{T}$-integrated nonprompt direct-photon yields also follow a power-law scaling behavior as a function of collision-system size. The exponent, $\alpha$, for the nonprompt component is found to be consistent with 1.1 with no apparent $p_{T}$ dependence. ## I Introduction Direct photons, defined as those not coming from hadron decays, have long been considered a golden probe towards our understanding of the evolution of relativistic heavy-ion collisions – from the quark-gluon plasma (QGP) phase to the hadron-gas (HG) phase [1]. Unlike strongly interacting probes, such as identified particles and jets, direct photons traverse the medium unmodified due to the small cross section of electromagnetic interaction. These penetrating photons encode information about the environment in which they were created, including the temperature and the collective motion of the medium. While the direct photons at high transverse momentum, $p_{T}$, are dominated by photons created from hard-scattering processes, such as quark- gluon Compton scattering, in the low-$p_{T}$ regime, they were initially predicted to be of a thermal origin, being emitted from the QGP and HG phase (see Ref. [2] for a recent review). The $p_{T}$ spectrum of low-$p_{T}$ direct photons from Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV, first measured by PHENIX [3], shows a clear excess above the hard-scattering contribution estimated from $p$$+$$p$ measurements for $p_{T}$ below 3 GeV/$c$. Followup measurements by PHENIX have established that low-$p_{T}$ direct-photon emission also shows a large anisotropy with respect to the reaction plane [4, 5], and that the yield increases faster than $N_{\rm part}$ or $dN_{\rm ch}/d\eta$ as a function of the centrality of the collision [6]. Low-$p_{T}$ direct photons in Au$+$Au collisions at 200 GeV have also been measured by STAR [7] using the same basic method as [3], but different detection techniques. Quantitatively, STAR results appear to be a factor 3 smaller than those from PHENIX. This tension has not yet been resolved. Furthermore, low $p_{T}$ photons have been measured in Au$+$Au at lower $\sqrt{s_{{}_{NN}}}$ of 39 GeV and 62.9 GeV by PHENIX [8], and in Pb$+$Pb at $\sqrt{s_{{}_{NN}}}=2760$ GeV by ALICE [9]. The excess of direct photons in A$+$A collisions, in the low-$p_{T}$ regime, is usually interpreted as the contribution of thermal radiation emitted from the expanding and cooling QGP and HG phase. Due to the rapid anisotropic expansion of the system, the radiation is Doppler shifted. Over the years, several theoretical models have been developed and refined to describe the production rates and space-time evolution of thermal photons in relativistic heavy-ion collisions [10, 11, 12, 13, 14, 15, 16, 17]. While most of these state-of-the-art models describe the data qualitatively, they fall short of simultaneously describing all the features of the data quantitatively. To describe the large yield, early emission at high temperatures is favored, while sufficient build up of collective motion is required to explain the large anisotropy, thereby favoring late-stage emission. This tension, often termed as the “direct-photon puzzle”, hints at an incomplete understanding of the different sources and mechanisms of direct-photon production. This has triggered more thoughts on other unconventional photon sources, such as emission from the pre-equilibrium stage, strong magnetic field effects, etc. [18, 19, 20, 21, 22, 23, 24, 10]. For that very reason this paper refers to the low-$p_{T}$-excess direct photons as “nonprompt” instead of “thermal”. To provide new insights and further understandings, the PHENIX collaboration presents results from the high-statistics 2014 Au$+$Au data at $\sqrt{s_{{}_{NN}}}=200$ GeV. With a 10-fold increase in statistics compared to previously published results, differential direct-photon measurements as functions of $p_{T}$ and system size over a broad $p_{T}$ range from 0.8–10 GeV/$c$ and in 10% centrality classes are discussed. A new algorithm, which utilizes the silicon-vertex detector (VTX) as the conversion material for photons, is developed for this analysis. The paper is organized as follows: Section II presents the experimental setup relevant to this measurement and the algorithm to reconstruct the conversion photons. Section III describes the double-ratio method to determine the direct-photon excess ratio, $R_{\gamma}$, and gives details of the experimental measurement. Section IV investigates the systematic uncertainties. Section V discusses the results. Section VI presents the summary and conclusions. Finally, there are two appendices: Appendix A discusses the event mixing procedures and their validity, while Appendix B describes the Monte-Carlo (MC)-sampling method used to derive the final systematic uncertainties on the direct-photon yield. ## II Experimental setup and photon measurements ### II.1 PHENIX 2014 Au$+$Au $\sqrt{s_{{}_{NN}}}=200$ GeV data set In 2014, a total of 19 billion Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV were recorded by the PHENIX detector at the Relativistic Heavy Ion Collider (RHIC) with a minimum-bias (MB) trigger, based on the response of two beam-beam counters (BBC) [25]. The BBCs are located on either side of the interaction point along the beam axis at $z={\pm}1.44$ m with a pseudorapidity coverage of $3.1<|\eta|<3.9$ and full $2\pi$ azimuthal acceptance. The MB trigger requires a coincident signal in both BBCs. Each BBC, comprising 64 Čerenkov counters, measures the total number of charged particles produced during the collision within its acceptance. The charged-particle multiplicity is used to divide the MB events into different centrality classes; 0%–10% corresponds to the most central collisions which produces the largest number of charged particles, while 80%–93% corresponds to peripheral collisions with only a small number of charged particles. The BBCs also utilize the arrival time of the produced particles on each side to determine the collision vertex along the beam direction. Figure 1: (a) The beam view of the PHENIX central-arm spectrometer for the year 2014. (b) A magnified view of the silicon-vertex detector. The solid curves correspond to the electron and positron tracks from photon conversion. The direct-photon measurement, presented here, is based on the tracking and identification of electrons and positrons from photon conversions in the detector material and the direct calorimetric measurement of photons in the two PHENIX central arm spectrometers shown in Fig. 1 [26]. The VTX [27] comprises four silicon layers at nominal radii of 2.6, 5.1, 11.8, and 16.7 cm. In the beam direction, the active area covers approximately $\pm 11$ cm for the innermost layer and $\pm 19$ cm for the outer layer. The VTX is not used as an active detector in the measurement. However, it acts as the photon converter, which is critical for this analysis. The total material thickness of the VTX in terms of radiation length, $X_{0}$, is ${\approx}13\%X_{0}$. Events are selected with a $z$ vertex within $\pm 10$ cm of the nominal interaction point. After applying quality assurance criteria, a total of $1.25\times 10^{10}$ events are analyzed. The central-arm spectrometers have three major parts: A charged-particle tracking system [28, 29], particle-identification detectors [30], and electromagnetic calorimeters (EMCal) [31]. Each arm covers 90∘ in the azimuthal direction with $|\eta|<0.35$. The tracking system is located $\approx$2.2 m from the beam axis outside of an axial magnetic field. The main tracking detectors are drift chambers (DC) and pad chambers (PC1). The DC provides a precise measurement of the transverse momentum for charged particles with $\mbox{$p_{T}$}>0.2$ GeV/$c$. The PC1 measures the momentum along beam direction, $p_{z}$. The effective momentum resolution of the central-arm tracking system, for this analysis, is $\sigma_{p}/p=0.8\%{\oplus}2\%\,p$ [GeV/$c$], where $p$ is the transverse momentum of the track. Charged tracks are identified as electrons or positrons with a ring-imaging Čerenkov detector (RICH). The RICH has a $CO_{2}$ gas radiator with a low radiation threshold for electrons (0.018 GeV/$c$) and a relatively high threshold for charged pions ($>4.87$ GeV/$c$). Requiring a signal in at least two phototubes in the focal plane of the RICH at the expected ring location effectively separates electrons below 5 GeV/$c$ from charged hadrons. A further matching of the momentum, $p$, of the charged track to the energy, $E$, as measured in the EMCal within $-2\sigma_{E/p}<E/p<5{\sigma_{E/p}}$ removes most hadrons remaining in the sample. Here $\sigma_{E/p}$ is the momentum-dependent resolution of the energy to momentum ratio, $E/p$. For the calorimetric identification of photons, two types of calorimeters are used, lead-scintillator (PbSc) and lead-glass (PbGl). The PbSc EMCal, which covers 3/4 of the acceptance, is a sandwich sampling detector, also referred to as a Shashlik type calorimeter. Based on the widths of reconstructed $\pi^{0}$ mass through the $\mbox{$\pi^{0}$}\rightarrow\gamma\gamma$ decay, the effective photon-energy resolution in this analysis is $\sigma_{E}/E=8.1\%/\sqrt{E~{}{\rm[GeV]}}{\oplus}5.0\%$. The remaining 1/4 of the acceptance is covered by the PbGl EMCal, which is a homogeneous Čerenkov- type detector with an effective resolution of $\sigma_{E}/E=8.7\%/\sqrt{E~{}{\rm\,[GeV]}}{\oplus}5.8\%$. Nominal cuts on the energy threshold ($E>500$ MeV) and shower shape ($\chi^{2}<3$) are applied to identify photons. ### II.2 External photon conversions in the VTX Earlier measurements of direct photons from PHENIX are based on three different strategies to measure photons in A$+$A collisions. The calorimeter method is used to measure photons with $p_{T}$ of several GeV/$c$ via their energy deposited in the EMCal [4]. To access lower $p_{T}$, $e^{+}e^{-}$ pairs from photon conversions are reconstructed with the tracking system. These $e^{+}e^{-}$ pairs are either from “internal” conversions of virtual photons emitted from the collision [3] or “external” conversions of photons in the detector material [6]. Figure 2: Artificial $e^{+}e^{-}$ pair mass for external photon conversions. Each curve corresponds to a different radius region, which roughly maps to the locations of beam-pipe, layers 1 (B0) through 4 (B3) of the VTX, and the VTX (CF) carbon-fiber enclosure. Here, external photon conversions at the VTX detector are reconstructed from $e^{+}e^{-}$ pairs. The VTX material is distributed between 2 and 25 cm along the radial direction. Depending on the conversion point, a different amount of magnetic field is traversed by the $e^{+}e^{-}$ pair. In the standard PHENIX track-reconstruction algorithm, the tracking system measures a part of the trajectory outside of the magnetic field at a radial position of $\approx$2.2 m. The momentum vector is determined by assuming that the particle originates at the event vertex. This assumption is incorrect for the $e^{+}e^{-}$ pairs from conversions in the VTX material. Both $e^{+}$ and $e^{-}$ traverse a smaller $\int{Bdl}$ than tracks from the vertex and thus the azimuthal component of the momentum vector is mismeasured in opposing directions, leading to an artificial opening angle and mismeasured mass of the $e^{+}e^{-}$ pair. Because the magnetic field in the region of the VTX detector is approximately constant at 0.9 Tesla, the artificial mass acquired is proportional to the radial location of the conversion point. Fig. 2 shows the mass of $e^{+}e^{-}$ pairs simulated with the geant3 PHENIX-detector simulation [32], different curves represent photon conversions in different VTX layers. The $m_{e^{+}e^{-}}$ is larger for conversions at larger radii with most conversions occurring in the third and fourth layers of the VTX, where the material budget is the largest. To correctly reconstruct and identify photon conversions at different VTX layers, a new track-reconstruction algorithm is developed. The new algorithm relies on the fact that the $e^{+}$ and $e^{-}$ from a conversion have the same origin and that their momenta were initially parallel in radial direction. This additional constraint eliminates the need to assume the origin of the track. Figure 3: Schematic view of the conversion-reconstruction algorithm. The two tracks are reconstructed to the same radius r. $\delta\phi$ is the azimuthal- angular difference between the two tracks for a given reconstruction radius. $\delta\phi$ is zero at the conversion point. The algorithm is illustrated in Fig. 3. For all radii between 0 and 30 cm, all possible momenta of the $e^{+}$ and $e^{-}$ are scanned to identify the azimuthal location $\phi_{\pm}$ at which the track is perpendicular to the circle of the given radius, or in other words points back radially to the event vertex. The conversion point is determined by finding the radius for which the difference of the azimuthal angles of the $e^{+}e^{-}$ pair, $\delta\phi=\phi_{+}{-}\phi_{-}$, becomes zero. If such radius exists, the pair is identified as a conversion candidate at the location $(\phi_{\rm conv},r_{\rm conv})$, where $\phi_{\rm conv}$ is the azimuthal angle of the conversion point, reconstructed with a resolution of $\approx$4 mrad, and $r_{\rm conv}$ is the radial position reconstructed with a resolution of $\approx$2 cm. ## III Data Analysis ### III.1 Double-ratio tagging method The number of direct photons emitted in a Au$+$Au collision is small compared to the number of photons from hadron decays. To make a precise measurement of the direct-photon yield, a tagging method is employed [6], which measures the ratio, $R_{\gamma}$, of all photons, referred to as inclusive photons, $\gamma^{\rm incl}$, to the photons from hadron decays, $\gamma^{\rm hadr}$. The ratio $R_{\gamma}$ is evaluated as double ratio, such that most systematic uncertainties cancel explicitly. The $R_{\gamma}$ given in Eq. 1 features three main terms: $R_{\gamma}=\frac{\gamma^{\rm incl}}{\gamma^{\rm hadr}}=\frac{\left(\frac{\gamma^{\rm incl}}{\gamma^{\pi^{0}}}\right)}{\left(\frac{\gamma^{\rm hadr}}{\gamma^{\pi^{0}}}\right)}=\frac{\langle\epsilon_{\gamma}f\rangle\left(\frac{N^{\rm incl}_{\gamma}}{N_{\gamma}^{\pi^{0},\rm tag}}\right)_{\rm Data}}{\left(\frac{\gamma^{\rm hadr}}{\gamma^{\pi^{0}}}\right)_{\rm Sim}}~{},$ (1) * (i) The ratio of measured photon yields $N_{\gamma}^{\rm{incl}}$/$N_{\gamma}^{\rm{\pi^{0},tag}}$ is the number of measured conversion photons in a given $p_{T}$ bin, divided by the sub-sample of those conversion photons that are tagged by a second photon as resulting from a $\mbox{$\pi^{0}$}\rightarrow\gamma\gamma$ decay. This quantity is measured in bins of fixed conversion photon $p_{T}$. * (ii) The conditional acceptance and efficiency $\langle\epsilon_{\gamma}f\rangle$ is the conditional probability to detect and reconstruct the second $\pi^{0}$ decay photon with the EMCal, given that the first decay photon was reconstructed as $e^{+}e^{-}$ pair from a photon conversion. The probability is averaged over all parent $\pi^{0}$ $p_{T}$ that can contribute to the given conversion photon $p_{T}$. * (iii) The cocktail ratio $\gamma^{\rm hadr}$/$\gamma^{\rm{\pi^{0}}}$ is the ratio of all photons from hadron decays over only those photons from $\pi^{0}$ decays. The following sections discuss how each term is determined. ### III.2 Ratio of the measured photon yields $N_{\gamma}^{\rm{incl}}$/$N_{\gamma}^{\rm{\pi^{0},tag}}$ Electrons and positrons in a given event are combined to $e^{+}e^{-}$ pairs and conversion candidates are selected with appropriate cuts, which results in a foreground sample of $e^{+}e^{-}$ pair ${\rm FG}^{\rm{ee}}$. All conversion candidates in a conversion photon $p_{T}$ bin, are combined with all photon showers in the EMCal above an energy threshold, $E_{cut}$. The invariant mass $m_{ee\gamma}$ is calculated and all combinations that lie in a mass window around the $\pi^{0}$ mass are considered as candidates for tagged photons ${\rm FG}^{ee\gamma}$. Due to the large particle multiplicity in Au$+$Au collisions, there are many false combinations where the electron, positron or photon are not from the same source. These background pairs must be subtracted statistically to obtain the signals of interest ${\rm SG}^{\rm{ee}}$ and ${\rm SG}^{ee\gamma}$. For $e^{+}e^{-}$ pairs, there are two possible combinations, signal pairs of interest ${\rm SG}^{\rm{ee}}$ and uncorrelated background ${\rm BG}^{\rm{ee}}$ pairs where the electron and positron are from different sources. Their sum constitutes the foreground ${\rm FG}^{\rm{ee}}$: $\mbox{${\rm FG}^{\rm{ee}}$}=\mbox{${\rm SG}^{\rm{ee}}$}+\mbox{${\rm BG}^{\rm{ee}}$}.$ (2) When the $e^{+}e^{-}$ pairs are combined with photons to $e^{+}e^{-}\gamma$ combinations, both types of $e^{+}e^{-}$ pairs are combined with photons that are either correlated or uncorrelated with the pair: $\mbox{${\rm FG}^{ee\gamma}$}=\mbox{${\rm SG}^{ee\gamma}$}+\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$}+\mbox{${\rm BG}_{\rm corr}^{ee\gamma}$}.$ (3) Introducing $i,j,k$ as the source of the positron, electron, and photon, respectively, the terms in Eq. 3 are: * (1) The first term is the signal of interest with positron, electron, and photon from the same source ($i=j=k$). * (2) The second term represents the cases where the $e^{+}e^{-}$ pair is combined with uncorrelated photons. This includes the case ($i=j\neq k$), where the $e^{+}e^{-}$ pair is correlated and randomly combined with a $\gamma$ as well as the case ($i\neq j\neq k$) where all three are from different sources. * (3) The third term represents cases ($(i\neq j=k)\vee(j\neq i=k)$), where the $e^{+}e^{-}$ pair is not from the same source but the $\gamma$ is correlated with either the $e^{+}$ or the $e^{-}$. Each of the background terms is determined with different event-mixing procedures, which were developed using the MC method. The event-mixing procedures and their validity are discussed in detail in Appendix A. #### III.2.1 Determination of the inclusive photon yield $N_{\gamma}^{\rm{incl}}$ Photons that convert at the VTX detector are selected by pairing electron and positron tracks to $e^{+}e^{-}$ pairs. All pairs are required to have a valid conversion point at a radial location within the VTX detector, between 1 and 29 cm. In addition, both tracks need to match in the beam direction within $|\Delta z|<4$ cm. The invariant mass distribution of the selected $e^{+}e^{-}$ conversion pairs is shown in Fig. 4 for the $p_{T}$ range $1.0<\mbox{$p_{T}$}<1.2$ GeV/$c$. The four panels correspond to four different centrality selections. Each panel shows the same peak structure, which is characteristic of the multilayer structure of the VTX detector. Figure 4: Mass distribution, $m_{e^{+}e^{-}}$, of the $e^{+}e^{-}$ pairs after conversion selection cuts are applied. All four panels are for the same $p_{T}$ range $1.0<\mbox{$p_{T}$}<1.2$ GeV/$c$ for four different centrality selections (a) 0%–20%, (b) 20%–40%, (c) 40%–60% and (d) 60%–93%. Shown are the foreground ${\rm FG}^{\rm{ee}}$, background ${\rm BG}^{\rm{ee}}$ and signal ${\rm SG}^{\rm{ee}}$. The $e^{+}e^{-}$ pairs passing the conversion selection criteria contain uncorrelated $e^{+}e^{-}$ pairs, where the $e^{+}$ and $e^{-}$ are from different sources. These backgrounds are also shown in Fig. 4. Because of its combinatorial nature, the background to foreground ratio increases towards more central-event selections. An event-mixing technique is used to estimate and subtract this background (see Appendix A for details). In this technique, an $e^{+}$ from event A is paired with an $e^{-}$ from another event B to produce the random $e^{+}e^{-}$ pair sample. To assure the events A and B have similar topological characteristics, it is required that both events: * (a) are from the same 10$\%$ centrality selection, * (b) have their interaction vertex within $\Delta z=2.5$ cm, * (c) have their reaction planes aligned within $\Delta\phi=\pi/6$. After the background is subtracted, $N_{\gamma}^{\rm{incl}}$ is calculated by integrating the counts in the mass range from 0.04 to 0.12 GeV/$c^{2}$, corresponding to layers 3 and 4 of the VTX. The analysis is repeated for bins in $p_{T}$ and in centrality. #### III.2.2 Tagged photon raw yield $N_{\gamma}^{\rm{\pi^{0},tag}}$ Next, the subset of $e^{+}e^{-}$ pairs in the $N_{\gamma}^{\rm{incl}}$ sample that can be tagged as photons from a $\pi^{0}$ decay, $N_{\gamma}^{\rm{\pi^{0},tag}}$, is determined. For a given event, each $e^{+}e^{-}$ conversion candidate, in the mass window in which $N_{\gamma}^{\rm{incl}}$ is counted, is paired with all reconstructed showers in the EMCal with shower shape $\chi^{2}<3$ and energy larger than $E_{cut}=0.5$ GeV, excluding those matched to the $e^{+}e^{-}$ pair itself. The energy cut, together with the $p_{T}$ cut of 0.2 GeV/$c$ on the $e^{+}$ and $e^{-}$, constitutes an implicit asymmetry cut on the $\pi^{0}$ decay photons that depends on the $p_{T}$ of the $\pi^{0}$. For all $e^{+}e^{-}\gamma$ combinations, the invariant mass $m_{ee\gamma}$ is calculated. This constitutes the foreground ${\rm FG}^{ee\gamma}$, for which an example is given in Fig. 6 for the $e^{+}e^{-}$ pair in the $p_{T}$ range $1.0<\mbox{$p_{T}$}<1.2$ GeV/$c$. The four panels (a) to (d) correspond to four centrality selections 0%–20%, 20%–40%, 40%–60%, and 60%–93%, respectively. Despite the large background, the signal, ${\rm SG}^{ee\gamma}$, is clearly visible as a peak around the $\pi^{0}$ mass, even in panel (a), which is the most central event selection. As discussed above, the background ${\rm BG}^{ee\gamma}$ has two components: $\mbox{${\rm BG}^{ee\gamma}$}=\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$}+\mbox{${\rm BG}_{\rm corr}^{ee\gamma}$},$ (4) for which the shape and normalization are obtained from the event-mixing procedures described in Appendix A. The results are also shown in Fig. 6. The uncorrelated background, ${\rm BG}_{\rm uncorr}^{ee\gamma}$, is given in panels (a) to (d). The much smaller correlated background, ${\rm BG}_{\rm corr}^{ee\gamma}$, is only revealed after ${\rm BG}_{\rm uncorr}^{ee\gamma}$ is subtracted from the foreground, ${\rm FG}^{ee\gamma}$. The differences are given in panels (e) to (h) for central to peripheral events, respectively. Figure 6 indicates that the correlated background decreases with centrality from $\mbox{${\rm BG}_{\rm corr}^{ee\gamma}$}/(\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$})=8.6\%$ in central collisions to 0.5% in the most-peripheral collisions. For the 0%–20% centrality selection, Fig. 6 shows the mass distributions $m_{ee\gamma}$ for four different $e^{+}e^{-}$ pair $p_{T}$ ranges. The representation is the same as for Fig. 6. Panels (a) through (d) all show a clear peak around the $\pi^{0}$ mass. The backgrounds are the largest for low $p_{T}$ and the most central events. As $p_{T}$ increases and the event multiplicity decreases, the backgrounds are significantly reduced. Figure 5: Mass distribution, $m_{ee\gamma}$, for $e^{+}e^{-}$ pairs with $p_{T}$ from 1.0 to 1.2 GeV/$c$, for four centrality selection (a,e) 0%–20%, (b,f) 20%–40% (c,g) 40%–60%, and (d,h) 60%–93%. Panels (a) through (d) show the foreground ${\rm FG}^{ee\gamma}$ and the uncorrelated background ${\rm BG}_{\rm uncorr}^{ee\gamma}$. Panels (e) through (h) show the difference $\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$}$, together with the correlated background ${\rm BG}_{\rm corr}^{ee\gamma}$. Figure 6: Mass distribution, $m_{ee\gamma}$, for $e^{+}e^{-}$ pairs from the 0%–20% centrality selection for four different $e^{+}e^{-}$ pair $p_{T}$ regions, (a,e) 0.8 to 1.0, (b,f) 1.4 to 1.6, (c,g) 2.0 to 2.5 and (d,h) 3.5 to 4 GeV/$c$. Panels (a) through (d) show the foreground ${\rm FG}^{ee\gamma}$ and the uncorrelated background ${\rm BG}_{\rm uncorr}^{ee\gamma}$. Panels (e) through (h) show the difference $\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$}$, together with the correlated background ${\rm BG}_{\rm corr}^{ee\gamma}$. Because of the complexity of the particle correlations present in the real Au$+$Au collision events, including effects of collective expansion, jet production, hadron decays, etc., there is a small residual background that is not captured by the event-mixing procedure. To remove this background, a low- order polynomial, $f_{ee\gamma}$, is fitted to the ratio $(\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm BG}^{ee\gamma}$})/\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$}$ in the mass range 0.05–0.08 and 0.23–0.45 GeV/$c^{2}$. This function is used to correct ${\rm BG}_{\rm uncorr}^{ee\gamma}$ before it is finally subtracted. Thus, the final distribution for $N_{\gamma}^{\rm{\pi^{0},tag}}$ is: $\mbox{$N_{\gamma}^{\rm{\pi^{0},tag}}$}=\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm BG}_{\rm corr}^{ee\gamma}$}-(1+f_{ee\gamma})\times\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$}.$ (5) An example of the residual background is given in Fig. 7 for the $e^{+}e^{-}$ pair $p_{T}$ range of 1 to 1.2 GeV/$c$ and 0%–20% centrality selection. In panel (a), ${\rm FG}^{ee\gamma}$ with all the background components are shown. Panel (b) gives a second-order polynomial fit to the ratio $(\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm BG}^{ee\gamma}$})/\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$}$ ratio, $f_{ee\gamma}$, which is used to determine the residual background. Due to the unfavorably small signal-to-background ratio in this case, the residual background in the $\pi^{0}$ mass region is $\approx$9.4%. The residual background quickly drops with $p_{T}$ and centrality bins, for example as $p_{T}$ increases to 3 GeV/$c$, the residual background reduces to 2.7%. For each $p_{T}$-centrality bin combination, $N_{\gamma}^{\rm{\pi^{0},tag}}$ is extracted by integrating the number of entries in a window around the $\pi^{0}$ peak ($0.09<\mbox{$m_{ee\gamma}$}<0.19$) GeV/$c^{2}$ after all background subtractions are applied. Figure 7: (a) An example for ${\rm FG}^{ee\gamma}$ and the various background components after normalization in the indicated regions. (b) The ratio $(\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm BG}^{ee\gamma}$})/\mbox{${\rm BG}_{\rm uncorr}^{ee\gamma}$}$ and the polynomial fit to determine the residual-background correction $f_{ee\gamma}$. ### III.3 Conditional probability $\langle\epsilon_{\gamma}f\rangle$ The probability, $\langle\epsilon_{\gamma}f\rangle$, that the second photon is in the acceptance and is reconstructed, given a conversion $e^{+}e^{-}$ pair from a $\pi^{0}$ decay, is extracted from the single $\pi^{0}$ simulation. In this simulation, individual $\pi^{0}$ are tracked through the PHENIX MC- simulation framework. The $\pi^{0}$ are generated with the published $p_{T}$ spectrum (see Sec. III.4) and uniform in pseudorapidity, $\eta$, and azimuthal angle, $\phi$. The energy scale and resolution of the EMCal in the MC simulation is tuned as closely as possible to resemble the one observed in data by comparing the mean and width of the measured and simulated $\pi^{0}$ mass distribution. The $\pi^{0}$ are reconstructed through the $\mbox{$\pi^{0}$}\rightarrow\gamma\gamma$ decay channel. For this purpose an asymmetry of less than 20% between the energies of the two decay photons was applied to keep the two-photon energies similar. In the single $\pi^{0}$ MC simulation, $e^{+}e^{-}$ pairs in the mass window $0.04<\mbox{$m_{e^{+}e^{-}}$}<0.12$ GeV/$c^{2}$ are counted to determine $N_{ee}^{\rm{\pi^{0}}}$, the number of reconstructed $e^{+}e^{-}$ pairs in a given $e^{+}e^{-}$ pair $p_{T}$ bin. The sub-sample for which the second photon of the $\pi^{0}$ decay is reconstructed as a shower in the EMCal is counted as $N_{ee}^{\rm{\pi^{0}},tag}$. The value of $\langle\epsilon_{\gamma}f\rangle$ is then determined as: $\mbox{$\langle\epsilon_{\gamma}f\rangle$}=\frac{\mbox{$N_{ee}^{\rm{\pi^{0}},tag}$}}{{N_{ee}^{\rm\pi^{0}}}}.$ (6) For the extraction of $N_{ee}^{\rm{\pi^{0}},tag}$ the presence of other showers in the EMCal needs to be taken into account. This is done by embedding the showers from the simulated single $\pi^{0}$ into the EMCal response from Au$+$Au collisions at the tower level. The combined EMCal information is then reclustered to form new showers. All of the showers that contain energy deposited by the embedded singe $\pi^{0}$ (identified by the MC ancestry information) are combined with the $e^{+}e^{-}$ pair. Figure 8: Conditional probability $\langle\epsilon_{\gamma}f\rangle$ as a function of $p_{T}$ in 0%–20%, 20%–40%, 40%–60% and 60%–93% centrality classes. Similar to the $N_{\gamma}^{\rm{\pi^{0},tag}}$ extraction from data, a residual background subtraction is applied. This eliminates any remaining background inside the $\pi^{0}$ counting window. The residual background is estimated by a second order polynomial function fit in the mass range 0.05–0.08 and 0.23–0.45 GeV/$c^{2}$. This residual background mainly comes from events where both decay photon convert to $e^{+}e^{-}$ pairs, and the reconstructed conversion photon gets paired with the EMCal cluster of the $e^{+}$ or $e^{-}$ from the other conversion. The extracted $\langle\epsilon_{\gamma}f\rangle$ is shown in Fig. 8 as a function of the $e^{+}e^{-}$ pair $p_{T}$ for the four centrality selections. The increasing trend of $\langle\epsilon_{\gamma}f\rangle$ with increasing conversion photon $p_{T}$ is partly due to the decrease in the opening angle between the conversion photon and the second photon so that the second photon is more likely to fall into the acceptance of the EMCal. Another important factor is that the average energy of the second photon increases with increasing conversion photon $p_{T}$, and hence, the efficiency of the energy threshold cut increases towards higher $p_{T}$. The difference in $\langle\epsilon_{\gamma}f\rangle$ between different centrality classes is mainly related to the shower shape ($\chi^{2}$) selection, because the showers are more distorted in central Au$+$Au collisions due to the larger detector occupancy, resulting in more accidental overlaps from the underlying event, and the centrality dependent parent $\pi^{0}$ $p_{T}$ distributions. Table 1: Parameters for the modified Hagedorn function Eq. 7 to PHENIX data [33, 34, 35] from Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV. centrality | $A$ | $a$ | $b$ | $p_{0}$ | $n$ ---|---|---|---|---|--- | c(GeV/$c$)-2 | (GeV/$c$)-1 | (GeV/$c$)-2 | GeV/$c$ | min.bias | 504.5 | 0.5169 | 0.1626 | 0.7366 | 8.274 0%–10% | 1331.0 | 0.5654 | 0.1945 | 0.7429 | 8.361 10%–20% | 1001.0 | 0.5260 | 0.1628 | 0.7511 | 8.348 20%–30% | 750.7 | 0.4900 | 0.1506 | 0.7478 | 8.229 30%–40% | 535.3 | 0.4534 | 0.1325 | 0.7525 | 8.333 40%–50% | 364.5 | 0.4333 | 0.1221 | 0.7385 | 8.261 50%–60% | 231.2 | 0.4220 | 0.1027 | 0.7258 | 8.220 60%–70% | 118.1 | 0.4416 | 0.0559 | 0.7230 | 8.163 70%–80% | 69.2 | 0.2850 | 0.0347 | 0.7787 | 8.532 80%–93% | 51.1 | 0.2470 | 0.0619 | 0.7101 | 8.453 ### III.4 Cocktail ratio $\gamma^{\rm hadr}$/$\gamma^{\rm{\pi^{0}}}$ The last ingredient to calculate $R_{\gamma}$ is the cocktail ratio $\gamma^{\rm hadr}$/$\gamma^{\rm{\pi^{0}}}$ of photons from $\pi^{0}$, $\eta$, $\omega$, and $\eta^{\prime}$ decays over those from $\pi^{0}$ decays. The cocktail ratio is obtained using the PHENIX meson decay generator EXODUS, which simulates mesons according to given input $p_{T}$ spectra, decays them based on the known decay kinematics and branching ratios, and aggregates the decay photons in the PHENIX detector acceptance. The photons from $\pi^{0}$ decays are generated from distributions obtained by fitting a modified Hagedorn function (Eq. 7) to charged pion [33] and neutral pion [34, 35] data measured by PHENIX. $E\frac{d^{3}N}{dp^{3}}=A\ \Big{(}e^{-(ap_{T}+bp_{T}^{2})}+\frac{\mbox{$p_{T}$}}{p_{0}}\Big{)}^{-n}.$ (7) The fit parameters are summarized in Table 1 for MB collisions, as well as for nine centrality bins. The $\eta$ meson $p_{T}$ spectrum is obtained by multiplying the $\pi^{0}$ spectrum with the $\eta/\mbox{$\pi^{0}$}$ ratio. The ratio is extracted from an analysis of world data [36], which demonstrates a universal value at high $p_{T}$ consistent with 0.487$\pm$0.024, independent of collision energy, system size or centrality. The work takes into account the fact that at lower $p_{T}$ the ratio deviates from $m_{T}$ scaling, and that there are centrality dependent changes of $\eta/\mbox{$\pi^{0}$}$ due to radial flow. The contribution from $\omega$ and $\eta^{\prime}$ decays are based on $p_{T}$ distributions using the $\pi^{0}$ spectrum and replacing by $f$($\sqrt{p_{T}+m^{2}_{\rm meson}-m^{2}_{\pi^{0}}}$). The normalization of $\omega$ and $\eta^{\prime}$ are fixed at $p_{T}$ = 5 GeV/$c$ to $0.9\pm 0.06$ and $0.25\pm 0.075$, respectively [6]. The cocktail ratio $\gamma^{\rm hadr}$/$\gamma^{\rm{\pi^{0}}}$ is shown in Fig. 9. Figure 9: Cocktail ratio as a function of $p_{T}$ in the most central (0%–20%) and the most peripheral (60%–93%) centrality classes. ## IV Systematic uncertainties This section describes the sources of systematic uncertainties for each of the three components for the calculation of $R_{\gamma}$. The systematic uncertainties are categorized into three types according to the correlation between the measured data points: * • Type A: No (or unknown) correlation between data points – uncertainties on the individual data points can fluctuate independently, in the same way as the statistical uncertainties. * • Type B: The uncertainties are correlated between data points – the fluctuation of each data point can be determined by the fluctuation of the neighboring points. * • Type C: A special form of type B uncertainty – every data point fluctuates with the exact same fraction. In the final results, type A systematic uncertainties are combined with the statistical uncertainties and type B and C are combined to obtain the total systematic uncertainty. The following subsections discuss the major individual sources contributing to the systematic uncertainties on $R_{\gamma}$ and on the direct-photon yield. All contributions are summarized in Table 2 and depicted in Fig. 11 and Fig. 11 as functions of $p_{T}$ for $R_{\gamma}$ and $\gamma^{\rm dir}$. The final systematic uncertainties on $\gamma^{\rm dir}$ and on all quantities derived from $\gamma^{\rm dir}$ are determined using the error-sampling method discussed in Appendix B Figure 10: Systematic uncertainties of $R_{\gamma}$ as a function of conversion photon $p_{T}$ in 0%–20%, 20%–40%, 40%–60% and 60%–93% centrality bins. The black curve corresponds to total uncertainty, and colored curves correspond to individual sources. The lines representing uncertainties from the energy scale and the conversion loss overlap at 3%, so do the lines representing uncertainties from the $\gamma$ reconstruction efficiency, acceptance and input $p_{T}$ spectra. Figure 11: Systematic uncertainties of $\gamma^{\rm hadr}$ as a function of photon $p_{T}$. Table 2: Summary of systematic uncertainties for $R_{\gamma}$ and $\gamma^{\rm dir}$. Uncertainties for which ranges are given vary with $p_{T}$. For details see Figs. 11 and 11. Observable | Factor | Source | correlation | correlation | 0%–20% | 20%–40% | 40%–60% | 60%–93% ---|---|---|---|---|---|---|---|--- | | | in $p_{T}$ | in centrality | | | | $R_{\gamma}$ | $N_{\gamma}^{\rm{incl}}$/$N_{\gamma}^{\rm{\pi^{0},tag}}$ | $N_{\gamma}^{\rm{incl}}$ purity | Type B | Type B | $<1\%$ | $<1\%$ | $<1\%$ | $<1\%$ | | $N_{\gamma}^{\rm{\pi^{0},tag}}$ residual background | Type A | Type A | 1.5%–4.5% | 0.5%–4% | 0.5%–4% | 0.5%–4% | | $N_{\gamma}^{\rm{\pi^{0},tag}}$ event mixing | Type B | Type B | 1.5% | 1.5% | 1.5% | 1.5% | $\langle\epsilon_{\gamma}f\rangle$ | energy scale | Type B | Type B | $3\%$ | $3\%$ | $3\%$ | $3\%$ | | conversion loss | Type C | Type C | $3\%$ | $3\%$ | $3\%$ | $3\%$ | | $\gamma$ efficiency | Type B | Type A | $<1.4\%$ | $<1\%$ | $<1\%$ | $<1\%$ | | active area & acceptance | Type C | Type C | $1\%$ | $1\%$ | $1\%$ | $1\%$ | | input $\pi^{0}$ $p_{T}$ spectra | Type B | Type A | $1\%$ | $1\%$ | $1\%$ | $1\%$ | $\gamma^{hadr}/\gamma^{\pi^{0}}$ | $\eta/\pi^{0}$ | Type B | Type C | 1–1.5% | 1–1.5% | 1–1.5% | 1–1.5% | | $\omega,\eta^{\prime}$ | Type B | Type C | $<1\%$ | $<1\%$ | $<1\%$ | $<1\%$ | $\gamma^{hadr}$ | input $\pi^{0}$ $p_{T}$ spectrum | Type B | Type A | 10%–24% | 10%–24% | 10%–25% | 10%–24% ### IV.1 Systematic uncertainties on $N_{\gamma}^{\rm{incl}}$/$N_{\gamma}^{\rm{\pi^{0},tag}}$ #### IV.1.1 Purity of the conversion photon sample Due to the high multiplicity of photons produced in Au$+$Au collisions, the background in the conversion sample from uncorrelated $e^{+}e^{-}$ pairs can be as large as 10% for the most central collisions and the lowest $p_{T}$ from 0.8 to 1.0 GeV/$c$. This background is subtracted statistically with a certain accuracy. To estimate the effect on the final results, significantly more and less stringent conversion selection cuts were applied, hence, increasing or reducing the purity. The value of $\langle\epsilon_{\gamma}f\rangle$ $N_{\gamma}^{\rm{incl}}$/$N_{\gamma}^{\rm{\pi^{0},tag}}$, obtained from the different cuts, varies by less than 1%. This range is quoted as systematic uncertainty due to the limited purity of the conversion sample. #### IV.1.2 $\pi^{0}$ yield extraction One of the main sources of systematic uncertainty on the $R_{\gamma}$ measurement is the tagged photon or $\pi^{0}$ yield extraction. The uncertainty of $\pi^{0}$ yield extraction arises from two sources: (i) from the residual background subtraction, which is highly correlated with the statistical accuracy of the mixed event background normalization, and (ii) imperfect description of the large backgrounds using event-mixing techniques. To evaluate the size of the uncertainty from the residual background subtraction, different estimates are compared. These include using different functional forms for the fit and different fit ranges to anchor the residual background fit. In addition, the counting window for $\pi^{0}$ signal extraction is varied. This gives a spread of $\langle\epsilon_{\gamma}f\rangle$ $N_{\gamma}^{\rm{incl}}$/$N_{\gamma}^{\rm{\pi^{0},tag}}$ values in each $p_{T}$ and centrality bin. The standard deviation of the spread is quoted as the uncertainty. Due to the correlation with the statistical accuracy of the foreground in the background region, this uncertainty depends on $p_{T}$ and centrality. To test the validity of the event-mixing techniques, an MC simulation with high multiplicity $\pi^{0}$ events is performed. Details are discussed in Appendix A. The simulation shows that $N_{\gamma}^{\rm{\pi^{0},tag}}$/$\langle\epsilon_{\gamma}f\rangle$ can be determined with the event-mixing technique to better than 1.5%. ### IV.2 Systematic uncertainty on $\langle\epsilon_{\gamma}f\rangle$ #### IV.2.1 Energy Scale The accuracy of the energy scale of the EMCal is the main source of systematic uncertainties in the $\langle\epsilon_{\gamma}f\rangle$ evaluation. Because of the energy threshold cut, the second photon is reconstructed only for $\approx$25% of the $e^{+}e^{-}$ pairs with the lowest $p_{T}$, even though the photon was in the EMCal acceptance. Any potential mismatch of the energy scale between the simulation and real data will cause $\langle\epsilon_{\gamma}f\rangle$ to be off; a higher (lower) energy scale in simulation will lead to an underestimate (overestimate) of $\langle\epsilon_{\gamma}f\rangle$. As mentioned earlier, to improve the accuracy, the EMCal response in the simulation is carefully tuned to the data using the $\pi^{0}$ mass measurement in the $\mbox{$\pi^{0}$}\rightarrow\gamma\gamma$ decay channel. The tuning includes scaling the MC energy scale by 0.3% and 2.2% for the PbSc and PbGl calorimeters, respectively. In addition, the nonlinearity of the energy response is adjusted by up to 5% at the lowest measured energies. After the tuning, the $\pi^{0}$ peak positions in data and MC are consistent to better than 1%. Considering the additional uncertainty due to the adjustment of the nonlinearity, the energy scale is known to better than 2%. Changing the energy scale by $\pm 2$% introduces a systematic uncertainty on $\langle\epsilon_{\gamma}f\rangle$ of 3% at low $p_{T}$ and decrease towards high $p_{T}$. The uncertainty on the energy resolution has a negligible effect. #### IV.2.2 Conversion Photon Loss Another important source of systematic uncertainty on $\langle\epsilon_{\gamma}f\rangle$ is related to the probability that the second photon converts to an $e^{+}e^{-}$ pair before reaching the EMCal. Depending on the location of the conversion point, the second photon may not be properly reconstructed, thereby reducing $\langle\epsilon_{\gamma}f\rangle$. To account for the “conversion loss”, the material budget, i.e. thickness and location of material, implemented in the simulation framework must accurately reflect reality. If there is a mismatch, the probability for conversions to occur will be different and, hence, $\langle\epsilon_{\gamma}f\rangle$ will by systematically off. As there is essentially no magnetic field after the DC exit window, the $e^{+}e^{-}$ pair from conversions between the DC and the EMCal will likely merge into one shower in the EMCal. Therefore, the value of $\langle\epsilon_{\gamma}f\rangle$ is most sensitive to differences in the material budget of the VTX. Comparison of the available information about the materials and their thickness for all detector subsystems, reveals that the conversion probability in material within the magnetic field is known to better than 3%, which directly translates into and uncertainty of $3\%$ on $R_{\gamma}$. #### IV.2.3 Photon Efficiency An EMCal shower shape, $\chi^{2}$, cut is used to identify photon candidates among the EMCal energy clusters and to reduce the number of hadrons in the sample. Similar to the energy scale uncertainty, a difference between the shower shape in simulation and the data will translate directly into a systematic shift of $\langle\epsilon_{\gamma}f\rangle$. To evaluate this uncertainty, the $\chi^{2}$ is varied simultaneously in both data and simulation and $\langle\epsilon_{\gamma}f\rangle$ $N_{\gamma}^{\rm{incl}}$/$N_{\gamma}^{\rm{\pi^{0},tag}}$ is recalculated. It changes by $1.4\%$ for 0.8–2 GeV/$c$ in the 0%–20% centrality bin and by less than 1$\%$ for all the other cases. #### IV.2.4 Active area and geometric acceptance Due to the limited geometrical acceptance of EMCal and some inactive areas, the second photon is registered only for $\approx$35% of the $e^{+}e^{-}$ pairs at the lowest $p_{T}$. Therefore, the accuracy with which the acceptance and dead areas are known will contribute to the systematic uncertainties on $\langle\epsilon_{\gamma}f\rangle$. The uncertainty of the acceptance is the result of the accuracy with which the radial location of the EMCal sectors can be determined. The possible remaining offset leads to $<0.3\%$ difference in acceptance along $\phi$ direction and $<0.9\%$ in $z$ direction. The dead areas in the real EMCal are carefully matched to the MC simulation and the accuracy of the dead area determination is found to be better than 0.6$\%$. It is due to the cases when a tower malfunctioned only in a very small number of events, and not masked out in the simulation. Combining all these effects, the systematic uncertainty on $R_{\gamma}$ from the acceptance is set to 1%. #### IV.2.5 Input $\pi^{0}$ distribution Because $\langle\epsilon_{\gamma}f\rangle$ is averaged over all parent $\pi^{0}$ $p_{T}$ that contribute to a given $e^{+}e^{-}$ pair $p_{T}$ bin, the $p_{T}$ dependence of $\langle\epsilon_{\gamma}f\rangle$ is sensitive to the shape of the $\pi^{0}$ distribution. The $\pi^{0}$ parent distribution was determined for each centrality selection by a fit to the best available data from Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV measured by the same experiment [33, 34, 35]. The remaining uncertainty on $\langle\epsilon_{\gamma}f\rangle$ is smaller than 1%. ### IV.3 Systematic uncertainty on $\gamma^{\rm hadr}$/$\gamma^{\rm{\pi^{0}}}$ The cocktail $\gamma^{\rm hadr}$/$\gamma^{\rm{\pi^{0}}}$ accounts for photons from hadron decays, other than those from $\pi^{0}$, which contributes $\approx$23% of the decay photons at high $p_{T}$. Of the additional decay photons more than 80% are from the $\eta\rightarrow\gamma\gamma$ decay, hence the accuracy with which $\eta/\pi^{0}$ is known will determine the systematic uncertainties on $R_{\gamma}$ from this source. The $p_{T}$ and centrality dependent upper and lower bounds on $\eta/\pi^{0}$ for Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV are taken from [36]. Together with the much smaller uncertainty on the contribution from $\omega$ and $\eta^{\prime}$ decays, the systematic uncertainty on $R_{\gamma}$ is below 2% for the entire $p_{T}$ range. ### IV.4 Systematic Uncertainties on $\gamma^{\rm dir}$ Once $R_{\gamma}$ is determined, the direct-photon yield $\gamma^{\rm dir}$ is calculated as: $\mbox{$\gamma^{\rm dir}$}=(\mbox{$R_{\gamma}$}-1)\ \mbox{$\gamma^{\rm hadr}$}.$ (8) In addition to the uncertainties on $R_{\gamma}$, the uncertainty on $\gamma^{\rm hadr}$ needs to be determined. These systematic uncertainties have been studied in detail in [6]. The main sources of uncertainty come from the accuracy with which the $\pi^{0}$ $p_{T}$ spectrum can be determined. These largely cancel in $R_{\gamma}$, but propagate directly to $\gamma^{\rm hadr}$. The input $\pi^{0}$ spectrum is based on measurements of charged pions, and $\pi^{0}$ from different data taking periods (see Sec. III.4). Each data set comes with its own systematic uncertainties, and in addition, the differences between different measurements are of the order of 10% [37]. The latter is the dominant uncertainty. The uncertainty on the spectra of other mesons ($\eta$, $\eta^{\prime}$, $\omega$) also contributes to the uncertainty on $\gamma^{\rm hadr}$, but to a much smaller extent. ## V Results ### V.1 Direct photon $R_{\gamma}$ Figure 12 shows $R_{\gamma}$ as function of photon $p_{T}$ for every 20% centrality class. The vertical error bar on each point corresponds to the statistical uncertainty, while the box gives the systematic uncertainty. The new results are compared with all other published PHENIX results for Au$+$Au at $\sqrt{s_{{}_{NN}}}=200$ GeV; these were obtained with different methods and have largely independent systematic uncertainties. The open circles were determined using the external conversion method deploying the HBD detector as converter [6], the full squares are from a virtual photon internal conversion measurement [4], and the open squares were measured with the EMCal alone [38]. All measurements agree well within their independent systematic uncertainties. The 2014 data presented here have smaller statistical uncertainties than in previous publications at RHIC due to the increased luminosity and significantly larger amount of conversion material. The new results provide a continuous measurement across a wide range of $p_{T}$ from 0.8 to 10 GeV/$c$. This range has previously been covered by measurements done with different techniques with different systematics. Up to 3 to 4 GeV/$c$ internal or external photon conversions to $e^{+}e^{-}$ pairs have been used, while above 4 GeV/$c$ photons were measured in the EMCal. For all centrality selections, $R_{\gamma}$ shows a significant excess that is rather constant below $\approx$3 GeV/$c$. Beyond that, $R_{\gamma}$ increases with $p_{T}$, the increase being most pronounced for central collisions, and $R_{\gamma}$ continuously decreases towards more peripheral collisions. This is expected as phenomena such as jet quenching reduce the number of decay photons from hadron decays in more central collisions, which in turn increases $R_{\gamma}$ [34, 35]. The high statistics of the 2014 data set allows to divide the data sample into nine centrality bins, from 0%–10% to 80%–93%, 10% bins each, except for the last one which is slightly larger. The resulting $R_{\gamma}$ are shown in Fig. 13. Up to 50%–60% centrality, data from the earlier calorimeter measurement [38] are also shown. For most bins the overall shape of $R_{\gamma}$ as a function of $p_{T}$ is similar to what is observed in Fig. 12, with a notable difference for panel (i), which is the most-peripheral centrality 80%–93%. Below 5 GeV/$c$, the most-peripheral Au$+$Au data show no significant excess above unity and are very consistent with the direct-photon result from $p$$+$$p$ collisions, which is also shown in panel (i). Figure 12: The ratio, $\mbox{$R_{\gamma}$}=\mbox{$\gamma^{\rm incl}$}/\mbox{$\gamma^{\rm hadr}$}$, as a function of conversion photon $p_{T}$ in 0%–20%, 20%–40%, 40%–60% and 60%–93% centrality bins. The 2014 Au$+$Au data at $\sqrt{s_{{}_{NN}}}=200$ GeV are compared to results from previous PHENIX publications for the same system and $\sqrt{s_{{}_{NN}}}$. Figure 13: $R_{\gamma}$ of direct photons as a function of conversion photon $p_{T}$ in 0%–10% to 80%–93% centrality bins. The MC sampling method is used to calculate both the statistical and systematic uncertainties on $\gamma^{\rm dir}$ and all quantities derived from the direct photon $p_{T}$ spectra. This method propagates the error correctly in the presence of unphysical values of $\mbox{$R_{\gamma}$}<1$ and $p_{T}$ and centrality dependent correlations of uncertainties; it is discussed in detail in Appendix B. Figure 14: Invariant yield of direct photons as a function of conversion photon $p_{T}$ in 0%–10% to 80%–93% centrality bins. Figure 15: Invariant yield of direct photons as a function of conversion photon $p_{T}$ in (a) 0%–20%, (b) 20%–40%, (c) 40%–60% and (d) 60%–93% centrality bins. ### V.2 Direct-photon invariant yield The direct-photon spectra are calculated from $R_{\gamma}$ and $\gamma^{\rm hadr}$ using Eq. 8. The results for all 10% centrality selections are given in Fig. 14. 111As the yields in the most-peripheral bin, 80%–93%, are mostly upper limits on the measurement, this bin will not be included for estimation of any further derived quantities in every 10% centrality selection. Figure 15 compares the direct-photon spectra with previous measurements, as shown in broader centrality bins (a) 0–20%, (b) 20–40%, (c) 40–60%, and (d) 60–93%. Each panel also presents the $N_{\rm coll}$-scaled pQCD calculation [12] and a fit to direct-photon data from $p$$+$$p$ collisions at $\sqrt{s}=200$ GeV [39, 40, 41]. The $p$$+$$p$ fit is performed with a pQCD-inspired functional form [42]: $\frac{d^{3}N}{d^{2}\mbox{$p_{T}$}dy}=\frac{A_{pp}}{(1+(\frac{p_{T}}{p_{0}})^{2})^{n}},$ (9) where the parameters are $A_{pp}=1.60\\!\cdot\\!10^{-4}$ (GeV/$c$)-2, $p_{0}=1.45$ GeV/$c$ and $n=3.3$. The error band around the central fit function represents the uncertainty propagated from both the data and the unknown true functional form of the spectrum down to very low $p_{T}$. The $p$$+$$p$ fit and the pQCD calculation agree well above 2 GeV/$c$, and can be used as an estimate for the prompt-photon contribution. Figure 15 also shows that the direct-photon yield for $p_{T}$ larger than 5 GeV/$c$ is well described by the $N_{\rm coll}$-scaled $p$$+$$p$ result and pQCD calculations, which confirms that the high-$p_{T}$ direct photons are predominately from initial hard-scattering processes. Below 4–5 GeV/$c$ a clear direct-photon excess develops above the prompt component, gradually becoming larger towards lower $p_{T}$. Figure 16: nonprompt direct-photon yield as a function of conversion photon $p_{T}$ in (a) 0%–20%, (b) 20%–40%, (c) 40%–60%, and (d) 60%–93% centrality bins. ### V.3 Nonprompt direct-photon excess To extract the direct-photon excess above the prompt-photon contribution, the $N_{\rm coll}$ scaled $p$$+$$p$ fit is subtracted from the Au$+$Au data. This excess is thought to be mostly the radiation that is emitted during the collision from the hot-expanding fireball, and will be referred to here as nonprompt direct-photon spectra. Figure 16 compares the nonprompt direct- photon spectra to previously published results from Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV [6], which had significantly lower statistics. The new 2014 data extend the coverage, both in $p_{T}$ and centrality. Table 3: Inverse slopes fitted to the direct-photon spectra in different $p_{T}$ ranges, and for different centrality selections. For each centrality range, $N_{\rm coll}$ and $dN_{\rm ch}/d\eta$ values are quoted, which are taken from previous work [43, 44], except for the $dN_{\rm ch}/d\eta$ values for the two most peripheral bins. Those were extrapolated using a fit of the form $\mbox{$dN_{\rm ch}/d\eta$}=B(\mbox{$N_{\rm coll}$})^{\beta}$. centrality | $dN_{\rm ch}/d\eta$ | $N_{\rm coll}$ | $T_{\rm eff}$ (GeV/$c$) | $T_{\rm eff}$ (GeV/$c$) ---|---|---|---|--- | | | $0.8<\mbox{$p_{T}$}<1.9$ GeV/$c$ | $2<\mbox{$p_{T}$}<4$ 0%–20% | $519.2\pm 26.3$ | $770.6\pm 79.9$ | $0.277\pm 0.017\ ^{+0.036}_{-0.014}$ | $0.428\pm 0.031\ ^{+0.031}_{-0.030}$ 20%–40% | $225.4\pm 13.2$ | $241.1\pm 28.4$ | $0.264\pm 0.010\ ^{+0.014}_{-0.007}$ | $0.354\pm 0.019\ ^{+0.020}_{-0.030}$ 40%–60% | $85.5\pm 8.0$ | $82.6\pm 9.3$ | $0.247\pm 0.007\ ^{+0.005}_{-0.004}$ | $0.392\pm 0.023\ ^{+0.022}_{-0.022}$ 60%–93% | $16.4\pm 2.8$ | $12.1\pm 3.1$ | $0.253\pm 0.011\ ^{+0.012}_{-0.006}$ | $0.331\pm 0.036\ ^{+0.031}_{-0.041}$ 0%–10% | $623.9\pm 32.2$ | $951\pm 98.5$ | $0.268\pm 0.024\ ^{+0.026}_{-0.012}$ | $0.514\pm 0.061\ ^{+0.066}_{-0.039}$ 10%–20% | $414.2\pm 20.2$ | $590.1\pm 61.1$ | $0.303\pm 0.024\ ^{+0.062}_{-0.021}$ | $0.358\pm 0.033\ ^{+0.024}_{-0.035}$ 20%–30% | $274\pm 14.8$ | $357.2\pm 35.5$ | $0.263\pm 0.011\ ^{+0.014}_{-0.007}$ | $0.351\pm 0.024\ ^{+0.020}_{-0.030}$ 30%–40% | $176.8\pm 11.6$ | $207.5\pm 21.2$ | $0.256\pm 0.011\ ^{+0.009}_{-0.005}$ | $0.333\pm 0.024\ ^{+0.020}_{-0.032}$ 40%–50% | $109.4\pm 9.1$ | $111.1\pm 10.8$ | $0.244\pm 0.009\ ^{+0.003}_{-0.005}$ | $0.389\pm 0.029\ ^{+0.020}_{-0.021}$ 50%–60% | $61.6\pm 7.1$ | $54.1\pm 7.9$ | $0.246\pm 0.010\ ^{+0.005}_{-0.005}$ | $0.345\pm 0.031\ ^{+0.019}_{-0.032}$ 60%–70% | 32 $\pm$ 5 | $24\pm 6$ | $0.261\pm 0.015\ ^{+0.020}_{-0.008}$ | $0.319\pm 0.049\ ^{+0.037}_{-0.042}$ 70%–80% | 16 $\pm$ 4 | $10\pm 3$ | $0.263\pm 0.016\ ^{+0.016}_{-0.007}$ | $0.335\pm 0.044\ ^{+0.020}_{-0.035}$ 80%–93% | 7 $\pm$ 2 | $4\pm 1$ | $-$ | $-$ The data are very consistent in the region of overlap. In the range 0.8 to 1.9 GeV/$c$, the data are fitted with an exponential function and the results are also shown on the panels of Fig. 16. The slope values are given in Table 3. All centrality selections are consistent with an average inverse slope, $T_{\rm eff}$, of ${\approx}0.260{\pm}0.011$ GeV/$c$. However, it is evident from Fig. 16 that the nonprompt direct-photon spectra are not described by a single exponential but rather have a continually increasing with $p_{T}$ inverse slope, $T_{\rm eff}$. Figure 17 brings this out more clearly where each nonprompt direct-photon spectrum is divided by a fit with a fixed slope, $T_{\rm eff}$ = 0.260 GeV/$c$. All centrality selections follow the same trend. Over the $p_{T}$ range of up to 2 GeV/$c$ the ratios are consistent with unity, but above 2 GeV/$c$, they start to rise monotonically. Figure 17: Ratio of the yield of nonprompt direct photons over the exponential fit result ($T_{\rm eff}$ fixed to 0.26 GeV/$c$) as a function of photon $p_{T}$. Figure 18: $T_{\rm eff}$ as a function of charged-particle multiplicity at midrapidity. To quantify this changing slope, the nonprompt direct-photon spectra are fitted with a second exponential function in the $p_{T}$ range from 2 to 4 GeV/$c$; the results are also included in Fig. 16. All data are consistent with an average inverse slope of $0.376\pm 0.037$ GeV/$c$, which is significantly larger than the slope observed below $p_{T}$ = 2 GeV/$c$. Above 4 GeV/$c$, the statistical and systematic uncertainties from the prompt-photon subtraction become too large for a detailed analysis. To establish any dependence on the system size, the nonprompt direct photon spectra are determined for each 10% centrality bin, and subsequently fitted by two exponential functions in the $p_{T}$ ranges $0.8<\mbox{$p_{T}$}<1.9$ GeV/$c$ and $2<\mbox{$p_{T}$}<4$ GeV/$c$. The resulting $T_{\rm eff}$ values are tabulated in Table 3 and depicted in Fig. 18 as a function of $dN_{\rm ch}/d\eta$. The figure also shows the average of the inverse slope values from fitting Fig. 16. The $T_{\rm eff}$ values are consistent with a constant value, independent of $dN_{\rm ch}/d\eta$. However, given the uncertainties on the data, a possible increase of $T_{\rm eff}$ with $dN_{\rm ch}/d\eta$ can not be excluded. In addition to investigating the $p_{T}$ and system-size dependence of the shape of the nonprompt direct-photon spectra, one can also look at the dependence of the yield on system size and $p_{T}$. As reported previously, the integrated direct-photon yield scales with $dN_{\rm ch}/d\eta$ to a power $\alpha$ [8]: $\frac{dN_{\gamma}}{dy}=\int_{p_{T,\rm{min}}}^{p_{T,\rm{max}}}\frac{dN_{\gamma}^{\rm dir}}{d\mbox{$p_{T}$}dy}d\mbox{$p_{T}$}=A\times\left(\frac{dN_{\rm{ch}}}{d\eta}\right)^{\alpha},$ (10) where all rapidity densities are densities at midrapidity. The direct-photon spectra shown in Fig. 14 are integrated from $\mbox{$p_{T}$}_{,\rm{min}}=1$ GeV/$c$ to $\mbox{$p_{T}$}_{,\rm{max}}=5$ GeV/$c$ and plotted as a function of $dN_{\rm ch}/d\eta$ in Fig. 19. They are in reasonable agreement with a compilation of other direct-photon results [8, 45], also shown in the figure. All data follow a trend similar to the $N_{\rm coll}$ scaled $p$$+$$p$ fit, shown as band, but at a roughly 10 times larger yield. Scaling with $N_{\rm coll}$ corresponds to $\alpha=1.25$ $\pm$ 0.02 [8]. The current high statistics data allow for finer centrality binning and changes this picture somewhat at the lowest and highest $dN_{\rm ch}/d\eta$. Fitting only the new results in Fig. 19 gives a value of $\alpha=1.11\pm 0.02(\rm{stat})\ _{-0.08}^{+0.09}(\rm{sys})$. This value is lower, but consistent within systematic uncertainties, with $\alpha=1.23$ $\pm$ 0.06 $\pm$ 0.18, found by fitting all previously published PHENIX A$+$A data [45]. Note that the previous PHENIX measurements obtained the $\eta$ spectrum by $m_{T}$-scaling the $\pi^{0}$ spectrum, while in the current measurement the $\eta$ spectrum is obtained from the $\eta/$$\pi^{0}$ ratio using the world data. There are significant differences between the two approaches in the low-$p_{T}$ region [36]. Because the integration range starts at low $p_{T}$ and is wide (1–5 GeV/$c$), the power $\alpha$ is smaller than previously published values, but is consistent within stated systematic uncertainties. However, it is also consistent with unity within uncertainties. Figure 19: Integrated direct-photon yield (1–5 GeV$/c$) versus charged-particle multiplicity at midrapidity. The present data is compared to a previous compilation of data from [8, 45] and the $N_{\rm coll}$ scaled fit to $p$$+$$p$ data. Also given are fits with Eq. 10 to different data; the solid line is a fit to the present data resulting in $\alpha=1.11\pm 0.02(\rm{stat})\ _{-0.08}^{+0.09}(\rm{sys})$, and the dashed line is from fitting previously published PHENIX data [45] that gave $\alpha=1.23$ $\pm$ 0.06 $\pm$ 0.18. Table 4: Scaling power, $\alpha$, of the $dN_{\rm ch}/d\eta$ dependence of nonprompt and direct-photon yields in various integration ranges. $p_{T}$ (GeV/$c$) | $\alpha(\mbox{$\gamma^{\rm{nonprompt}}$})$ | $\alpha(\mbox{$\gamma^{\rm dir}$})$ ---|---|--- 0.8–1.2 | $1.119\pm 0.038\ _{-0.094}^{+0.116}$ | $1.124\pm 0.036\ _{-0.089}^{+0.121}$ 1.2–1.6 | $1.107\pm 0.029\ _{-0.082}^{+0.108}$ | $1.118\pm 0.027\ _{-0.073}^{+0.097}$ 1.6–2.0 | $1.136\pm 0.034\ _{-0.091}^{+0.129}$ | $1.152\pm 0.029\ _{-0.077}^{+0.113}$ 2.0–3.0 | $1.087\pm 0.032\ _{-0.092}^{+0.108}$ | $1.120\pm 0.025\ _{-0.065}^{+0.095}$ 3.0–4.0 | $1.119\pm 0.078\ _{-0.134}^{+0.206}$ | $1.171\pm 0.048\ _{-0.076}^{+0.114}$ 4.0–5.0 | $0.950\pm 0.176\ _{-0.205}^{+0.315}$ | $1.137\pm 0.077\ _{-0.082}^{+0.108}$ 5.0–10.0 | | $1.296\pm 0.078\ _{-0.091}^{+0.129}$ To better understand the behavior of the scaling power, $\alpha$, in more detail, the direct-photon yield and its nonprompt component are integrated for six different nonoverlapping finer $p_{T}$ regions and for 10% centrality classes. The integrated nonprompt yields are shown in Fig. 20. The $\alpha$ values are determined for each $p_{T}$ selection by fitting the data with Eq. 10. The fits are also shown in the figure. All $\alpha$ values, both for the direct photon yield and the nonprompt component, are tabulated in Table 4 and shown in Fig. 21. It is evident that the values for the direct component, for higher $p_{T}$ ranges, are consistent with the prompt component, $\alpha=1.25\pm 0.02$, corresponding to $N_{\rm coll}$ scaling. However, they tend to be smaller, but still consistent within systematic uncertainties, with previous measurements [8] for the lower $p_{T}$ ranges. With increasing $p_{T}$, the $\alpha$ values for the nonprompt component are slightly lower than those from direct photons. The systematic uncertainties are larger due to the subtraction. The values of $\alpha$ for the nonprompt component, as shown in Fig. 21, are remarkably constant with no evident $p_{T}$ dependence. Figure 20: Integrated nonprompt direct-photon yield versus charged particle multiplicity at midrapidity for different $p_{T}$ integration ranges. Figure 21: Scaling factors, $\alpha$, extracted from fitting Eq. 10 to integrated direct and nonprompt-photon yields as a function of $dN_{\rm ch}/d\eta$. Values were obtained for different $p_{T}$ integration ranges tabulated in Table 4. ## VI Concluding discussion of the results The PHENIX collaboration has measured direct-photon production in Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV using photon conversions to $e^{+}e^{-}$ pairs. A large yield of direct photons below a $p_{T}$ of 3 GeV/$c$ is observed for all centrality bins except for the most peripheral bin of 80%–93% with $dN_{\rm ch}/d\eta$ = 7.4, where it seems to be consistent with the prompt-photon production with little or no radiation from a fireball. The next centrality bin from 70%–80% with $dN_{\rm ch}/d\eta$ = 15.5 already shows a significant yield with properties very similar to that of the radiation from the more central bins. The nonprompt direct-photon spectra are isolated by subtracting the prompt- photon contribution, which is estimated through a fit to the direct-photon data from $p$$+$$p$ collisions at $\sqrt{s}=200$ GeV, measured by PHENIX, and scaled by $N_{\rm coll}$. Results are obtained for the $p_{T}$ range from 0.8 to 5 GeV/$c$ and for 0%–93% central collisions, covering a system size spanning two orders of magnitude in $dN_{\rm ch}/d\eta$ from $\approx$7 to 620. The wealth of data enabled PHENIX to carry out double-differential analyses of the shape of the momentum spectra and the rapidity density $dN_{\gamma}/dy$ in $p_{T}$ and $dN_{\rm ch}/d\eta$. For the centrality selections from 0%–10% to 70%–80%, all nonprompt direct- photon spectra are very similar in shape, exhibiting increasing $T_{\rm eff}$ from 0.2 to 0.4 GeV/$c$ over the $p_{T}$ range from 0.8 to 4 GeV/$c$. The changing $T_{\rm eff}$ is not surprising, because the spectra are time integrated over the full evolution of the expanding fireball, from its earliest pre-equilibrium state, through the QGP phase, crossing over to a HG, and further expanding and cooling until hadrons eventually stop interacting. Throughout the evolution the system cools, and thus earlier phases are characterized by higher temperatures. In turn, the contributions from the earliest times of the evolution are likely to dominate the emission at higher $p_{T}$, consistent with the observation of an increasing $T_{\rm eff}$ with $p_{T}$. In the lower $p_{T}$ range from 0.8 to 1.9 GeV/$c$, the spectra are well described by a $T_{\rm eff}$ = 0.26 GeV/$c$. This is consistent with what is expected for radiation from the late QGP stage until freeze-out [14]. During this period of the evolution, the temperature drops from $\approx$170 MeV near the transition to $\approx$110 MeV when the system freezes out. At the same time the system is rapidly expanding and thus, the radiation is blue shifted. This compensates the temperature drop and results in an average $T_{\rm eff}$ = $\approx$0.26 GeV/$c$, with only minor variations with centrality of the collision. In Ref. [14], a moderate increase of $T_{\rm eff}$ with centrality was predicted. While the data favors a $T_{\rm eff}$ independent of centrality, they are not precise enough to exclude a moderate change. Above a $p_{T}$ of 2 GeV/$c$, the inverse slope of the spectra continues to increase with $p_{T}$. Between $p_{T}$ = 2 and 4 GeV/$c$ the average inverse slope is $T_{\rm eff}$ $\approx$0.376 GeV/$c$. This $T_{\rm eff}$ is larger than what can be accommodated by a rapidly expanding HG, thus suggesting that emissions from earlier times in the evolution starts to dominate the spectra. Expected initial temperatures at RHIC are $\approx$375 MeV with maximum $T_{\rm eff}$ in the range of 0.35 to 0.4 GeV/$c$, depending on viscosity [14]. Thus, it is likely that an additional contribution from the pre- equilibrium stage is needed to account for the measured $T_{\rm eff}$. Figure 22: Nonprompt direct-photon yields for (a) 0%–20% and (b) 20%–40% compared with model predictions from Ref. [10, 46]. (c,d) ratios of the yields from data to the sum of yields from thermal and pre-equilibrium contributions. In Fig. 22, the measured nonprompt direct-photon spectra are compared to a recent calculation including contributions from the pre-equilibrium phase [10, 46]. These calculations predicted that the pre-equilibrium radiation becomes the dominant source above a $p_{T}$ of 3 GeV/$c$. In the range $2<\mbox{$p_{T}$}<4$ GeV/$c$, a fit of the thermal contribution with an exponential function results in an inverse slope of $\approx$0.36 GeV/$c$, while for the pre-equilibrium contribution a larger inverse slope of $\approx$0.52 GeV/$c$ is found, for the more central collisions. Fitting the same $p_{T}$ range for the combined thermal and pre-equilibrium spectra from the model gives an inverse slope of $\approx$0.425 GeV/$c$. While the shape is reproduced well, the overall yield predicted by the calculations falls short compared to the data, in particular, below 2 GeV/$c$ where the nonprompt- photon yield appears to be a factor of two to three larger. The integrated nonprompt direct-photon yield exhibits a power-law relation with $(\mbox{$dN_{\rm ch}/d\eta$})^{\alpha}$ [8]. Fitting the power $\alpha$ for multiple nonoverlapping $p_{T}$ ranges results in values consistent with $\alpha=1.12\pm 0.06{\rm(\rm{stat})}\pm 0.14{\rm(\rm{sys})}$ with no apparent dependence on $p_{T}$. The model calculations in [14] predict that the radiation from the HG phase scale with $\alpha$ close to 1.2, while those from the hot and dense QGP phase exhibit closer to a $(\mbox{$dN_{\rm ch}/d\eta$})^{2}$ dependence. Because the QGP phase has a larger relative contribution to the $p_{T}$ spectrum with increasing $p_{T}$, it is expected that $\alpha$ increases with $p_{T}$. However, the $p_{T}$ dependence of $\alpha$ from the pre-equilibrium phase needs further theoretical understanding. In conclusion, the 10-fold increase in statistics compared to previous samples of Au$+$Au collisions recorded by PHENIX enabled detailed measurements of the radiation from the hot and expanding fireball. The experimentally observed inverse slopes of the $p_{T}$ spectra are qualitatively consistent with predictions for thermal and pre-equilibrium radiation. However, there seems to be more photons emitted from Au$+$Au collisions than can be accounted for in model calculations. Furthermore, although this work presents no new data on the azimuthal anisotropy, maximum anisotropy is observed for photons $\approx$2–3 GeV/$c$. In this $p_{T}$ range, the yield is larger than what would be expected from a rapidly but anisotropically expanding hadronic fireball. Finally, the centrality dependence of the nonprompt direct-photon yield, expressed in terms of the scaling power $\alpha(\mbox{$p_{T}$})$, shows no indication of changing with $p_{T}$. ###### Acknowledgements. We thank the staff of the Collider-Accelerator and Physics Departments at Brookhaven National Laboratory and the staff of the other PHENIX participating institutions for their vital contributions. We also thank J.F. Paquet for many fruitful discussions and sharing additional information. We acknowledge support from the Office of Nuclear Physics in the Office of Science of the Department of Energy, the National Science Foundation, Abilene Christian University Research Council, Research Foundation of SUNY, and Dean of the College of Arts and Sciences, Vanderbilt University (USA), Ministry of Education, Culture, Sports, Science, and Technology and the Japan Society for the Promotion of Science (Japan), Natural Science Foundation of China (People’s Republic of China), Croatian Science Foundation and Ministry of Science and Education (Croatia), Ministry of Education, Youth and Sports (Czech Republic), Centre National de la Recherche Scientifique, Commissariat à l’Énergie Atomique, and Institut National de Physique Nucléaire et de Physique des Particules (France), J. Bolyai Research Scholarship, EFOP, the New National Excellence Program (ÚNKP), NKFIH, and OTKA (Hungary), Department of Atomic Energy and Department of Science and Technology (India), Israel Science Foundation (Israel), Basic Science Research and SRC(CENuM) Programs through NRF funded by the Ministry of Education and the Ministry of Science and ICT (Korea). Ministry of Education and Science, Russian Academy of Sciences, Federal Agency of Atomic Energy (Russia), VR and Wallenberg Foundation (Sweden), University of Zambia, the Government of the Republic of Zambia (Zambia), the U.S. Civilian Research and Development Foundation for the Independent States of the Former Soviet Union, the Hungarian American Enterprise Scholarship Fund, the US-Hungarian Fulbright Foundation, and the US-Israel Binational Science Foundation. ## Appendix A Event-mixing procedures and validation In this analysis, $e^{+}e^{-}$ pairs and $e^{+}e^{-}\gamma$ combinations result from combining positrons, electrons, and photons measured in the same event. Given the large multiplicity of produced particles in Au$+$Au collisions, the combinations include a significant background from particles of different physical origin, for example different $\pi^{0}$ decays. For $e^{+}e^{-}$ pairs there are two possible combinations: signal pairs, ${\rm SG}^{\rm{ee}}$, that have the same source and background pairs, ${\rm BG}^{\rm{ee}}$, that have different sources. Both types will be combined with photons to get $e^{+}e^{-}\gamma$ combinations. There are three possibilities: A correlated $e^{+}e^{-}$ pair is combined with a photon from the same source (${\rm SG}^{ee\gamma}$); the $e^{+}e^{-}$ pair is not correlated, but the photon is correlated to the $e^{+}$ or $e^{-}$ (${\rm BG}_{\rm corr}^{ee\gamma}$); or the photon is uncorrelated to the $e^{+}e^{-}$ pair, irrespective whether it is ${\rm SG}^{\rm{ee}}$ or ${\rm BG}^{\rm{ee}}$ (${\rm BG}_{\rm uncorr}^{ee\gamma}$). All backgrounds are determined using event-mixing techniques that were developed and validated with MC studies of high-multiplicity events, for which a large sample of simulated $\pi^{0}$ events was generated. These events serve as pseudodata. The $\pi^{0}$ are generated according to the experimentally observed $p_{T}$ spectrum, uniform in azimuthal angle, and with a constant pseudorapidity density of 280 $\pi^{0}$, which corresponds to the typical $\pi^{0}$ multiplicity in the most central Au$+$Au collisions at $\sqrt{s_{{}_{NN}}}=200$ GeV. From these pseudodata, $N_{\gamma}^{\rm{incl}}$ and $N_{\gamma}^{\rm{\pi^{0},tag}}$ are extracted using the cuts and event-mixing schemes developed for the analysis of real data. They are corrected by $\langle\epsilon_{\gamma}f\rangle$, resulting in $R_{\gamma}$. Because in the pseudodata there are no other hadronic decay channels contributing to $\gamma^{\rm hadr}$ other than $\pi^{0}$, the $R_{\gamma}$ from this pseudodata is given by: $R_{\gamma}^{\rm pseudo}=\frac{N^{\rm incl}_{\gamma}}{N_{\gamma}^{\pi^{0},\rm tag}}\times\mbox{$\langle\epsilon_{\gamma}f\rangle$}$ (11) As there are no direct photons in the pseudodata, the expected result would be $R_{\gamma}$ = 1, within the statistical uncertainties of the simulation. The rest of this sections details each step of the $R_{\gamma}$ determination from the pseudodata. The exact same procedure is also applied to the real data. ### A.1 Determination of the inclusive photon yield $N_{\gamma}^{\rm{incl}}$ Photon conversion candidates are created by combining $e^{+}$ and $e^{-}$ from the same pseudodata event by requiring a valid conversion point within $1<R<29$ cm. This results in a foreground, ${\rm FG}^{\rm{ee}}$, containing a signal, ${\rm SG}^{\rm{ee}}$, that is, conversions of $\pi^{0}$ decay photons, and a background, ${\rm BG}^{\rm{ee}}$, where the $e^{+}$ and $e^{-}$ come from conversion of two different $\pi^{0}$ decay photons. The background is determined by combining electrons and positrons from different pseudodata events, which are paired and subjected to the same cuts and conversion selection criteria. The mixed event background thus obtained, ${\rm MBG}^{\rm{ee}}$, is normalized to the foreground, ${\rm FG}^{\rm{ee}}$, in the mass region $0.16<\mbox{$m_{e^{+}e^{-}}$}<$ 0.3 GeV/$c^{2}$, which does not contain $e^{+}e^{-}$ pairs from conversions (see Fig. 2 for reference). Figure 23: Invariant mass distributions of $e^{+}e^{-}$ pairs reconstructed from the high-multiplicity $\pi^{0}$ pseudodata in the $p_{T}$ range $0.8<\mbox{$p_{T}$}<1.0$ GeV/$c$. The least-restrictive conversion selection cuts are applied, which only require that the reconstruction algorithm has identified the $e^{+}e^{-}$ pair as a conversion candidate. Panel (a) compares foreground, ${\rm FG}^{\rm{ee}}$, the true background, ${\rm BG}^{\rm{ee}}$, and the background determined from the mixed event technique, ${\rm MBG}^{\rm{ee}}$. Panel (b) gives the extracted conversion photon signal. Figure 24: Invariant mass distributions of $e^{+}e^{-}$ pairs reconstructed from the high-multiplicity $\pi^{0}$ pseudodata. Same as Fig. 24, but with an additional constraint that the $e^{+}$ and $e^{-}$ match in beam direction. Panel (a) compares foreground, ${\rm FG}^{\rm{ee}}$, the true background, ${\rm BG}^{\rm{ee}}$, and the background determined from the mixed event technique, ${\rm MBG}^{\rm{ee}}$. Panel (b) gives the extracted conversion photon signal. Figure 25: Extracted $N_{\gamma}^{\rm incl}$ after the background subtraction, as a function of conversion photon $p_{T}$. The diamonds are obtained by subtracting the background from the mixed event technique; they are compared to the open symbols for which the true background was subtracted. Panel (b) shows the ratio of the event-mixing result over the true information result. Figure 26: Invariant mass distributions of $e^{+}e^{-}\gamma$ pairs. Figure 24(a) shows the background, ${\rm MBG}^{\rm{ee}}$, obtained from the mixed event technique together with the true background, ${\rm BG}^{\rm{ee}}$, which was obtained from the MC ancestry information. Figure 24(b) shows the results (solid curve) after subtracting the mixed-event background from the foreground and (open symbols) subtracting the true background. Note that the two are practically indistinguishable, which means that ${\rm BG}^{\rm{ee}}$ is equal to ${\rm MBG}^{\rm{ee}}$. Even though the background can be subtracted accurately with the mixed-event technique to obtain $N_{\gamma}^{\rm{incl}}$, the subtraction can only be done statistically. Thus in the next step, where conversion photons from $\pi^{0}$ decays are tagged, the background pairs also need to be matched with EMCal showers. This substantially increases the background in the $m_{ee\gamma}$ distribution. To reduce this background, additional cuts are applied in the conversion-photon selection. The magnetic field deflects electrons and positrons in a plane perpendicular to the beam direction ($z$). Thus, $e^{+}e^{-}$ pairs from a conversion can be constrained by requiring a match in the beam direction using the PC1 information. A cut of $|\Delta z|<4$ cm is applied. Because the conversion reconstruction algorithm uses the projection of the tracks in the plane perpendicular to the beam axis, the additional match reduces the number of possible random-track combinations significantly. The $z$ cut effectively truncates the mass distribution as the $e^{+}e^{-}$ pairs are required to have the possible conversion point at radii below 29 cm and only the pairs with an opening angle in the beam direction will create larger masses. The background rejection is clearly visible in Fig. 24. The background normalization for the mixed events is given by the less-restrictive cuts shown in Fig. 24, and applied here. For the lowest $p_{T}$ and the highest-multiplicity bin, the background rejection is approximately a factor of eight with a signal efficiency of more than 85%. The background to foreground ratio, $\mbox{${\rm BG}^{\rm{ee}}$}/\mbox{${\rm FG}^{\rm{ee}}$}$, is 12.1%. As $p_{T}$ increases the multiplicity decreases and the $\mbox{${\rm BG}^{\rm{ee}}$}/\mbox{${\rm FG}^{\rm{ee}}$}$ ratio decreases to 0.3% at the $p_{T}$ above 7 GeV/$c$. The analysis is repeated for the entire accessible $p_{T}$ range and $N_{\gamma}^{\rm{incl}}$ is calculated in the mass range from 0.04 to 0.12 GeV/$c^{2}$ by subtracting the background obtained from the mixed-event technique, ${\rm MBG}^{\rm{ee}}$, from the foreground, ${\rm FG}^{\rm{ee}}$. The result is compared to the true number of photon conversions determined from the MC-ancestry information in Fig. 25. Panel (b) shows that the difference is less than 1% for all $p_{T}$. ### A.2 The tagged photon yield $N_{\gamma}^{\rm{\pi^{0},tag}}$ Next, the subset $N_{\gamma}^{\rm{\pi^{0},tag}}$ of $e^{+}e^{-}$ pairs in the $N_{\gamma}^{\rm{incl}}$ sample that can be tagged as photons from a $\pi^{0}$ decay is determined. For a given pseudodata event, each $e^{+}e^{-}$ conversion candidate is paired with all reconstructed showers in the EMCal, excluding the showers matched to the $e^{+}e^{-}$ pair itself. For each combination the invariant mass $m_{ee\gamma}$ is calculated. This constitutes the foreground, ${\rm FG}^{ee\gamma}$, for which an example is given in panel (a) of Fig. 26. Despite the large background the signal $N_{\gamma}^{\rm{\pi^{0},tag}}$ is clearly visible as peak around the $\pi^{0}$ mass. The background has two components: (i) combinations of $e^{+}e^{-}$ pairs with an EMCal shower from another unrelated $\pi^{0}$ decay denoted as ${\rm BG}_{\rm uncorr}^{ee\gamma}$, and (ii) a correlated background, ${\rm BG}_{\rm corr}^{ee\gamma}$, where the shower in the EMCal and the electron or positron are from the same $\pi^{0}$ decay, but the $e^{+}e^{-}$ pair itself is a combination of an $e^{+}$ and $e^{-}$ from different $\pi^{0}$ decay photons. The uncorrelated background can be determined with a similar event mixing technique as used for the extraction of $N_{\gamma}^{\rm{incl}}$; an $e^{+}e^{-}$ pair from one event is mixed with the EMCal showers from a different event resulting in mixed combinations, ${\rm MBG}_{\rm uncorr}^{ee\gamma}$. These are normalized to the foreground, ${\rm FG}^{ee\gamma}$, in the mass region from 0.25 to 0.45 GeV/$c^{2}$, where no signal is expected. Figure 26(a) shows the corresponding distribution. There is almost no visible difference between the mixed-event background, ${\rm MBG}_{\rm uncorr}^{ee\gamma}$, and the true background, ${\rm BG}_{\rm uncorr}^{ee\gamma}$, which is obtained using the MC-ancestry information. Figure 26(b) shows the signal and remaining correlated background after the uncorrelated mixed-event background is subtracted (${\rm FG}^{ee\gamma}$-${\rm MBG}_{\rm uncorr}^{ee\gamma}$), as well as after subtracting the true uncorrelated background (${\rm FG}^{ee\gamma}$-${\rm BG}_{\rm uncorr}^{ee\gamma}$). Again they are indistinguishable. The correlated background, ${\rm BG}_{\rm corr}^{ee\gamma}$, is determined with a second event-mixing scheme. An $e^{+}$ from a given event is combined with an $e^{-}$ from a different event, and the resulting $e^{+}e^{-}$ pair is then combined with the showers in the EMCal from both events; again excluding the showers from the $e^{+}$ and $e^{-}$. The $e^{+}e^{-}\gamma$ combinations contain the correlated background, ${\rm MBG}_{\rm cor}^{ee\gamma}$, plus the random background in which the $e^{+}$, $e^{-}$, and $\gamma$ are from three different $\pi^{0}$ decays, ${\rm MBG}_{\rm comb}^{ee\gamma}$. The normalization is per generated $e^{+}e^{-}$ pair, multiplied by ${\rm FG}^{ee\gamma}$, i.e. the number of background pairs in the $e^{+}e^{-}$ pair foreground. Figure 27: Invariant mass distributions of $e^{+}e^{-}\gamma$ pairs from the same event (FG) and different event-mixing setups. The random background, ${\rm MBG}_{\rm comb}^{ee\gamma}$, can easily be determined in a third event-mixing step, where $e^{+}$, $e^{-}$, and $\gamma$ are from three different events. The ${\rm MBG}_{\rm comb}^{ee\gamma}$ is normalized to (${\rm MBG}_{\rm cor}^{ee\gamma}$ $+$${\rm MBG}_{\rm comb}^{ee\gamma}$) in the mass range from 0.65 to 1.0 GeV/$c^{2}$ and subtracted. Figure 27(a) shows the the result, ${\rm MBG}_{\rm cor}^{ee\gamma}$, together with the foreground and the other background components. Last but not least, to account for any possible mismatch between the true background and the one obtained from our multistep event-mixing procedure, the ratio $(\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm MBG}_{\rm cor}^{ee\gamma}$}-\mbox{${\rm MBG}_{\rm uncorr}^{ee\gamma}$})/\mbox{${\rm MBG}_{\rm uncorr}^{ee\gamma}$}$ is fit with a second-order polynomial, $f_{ee\gamma}$, excluding the $\pi^{0}$ peak regions. The fit result is shown as a line on Fig. 27(b). This fit is used to correct ${\rm MBG}_{\rm uncorr}^{ee\gamma}$ before subtraction. The final distribution for $N_{\gamma}^{\rm{\pi^{0},tag}}$ is thus: $\mbox{$N_{\gamma}^{\rm{\pi^{0},tag}}$}=\mbox{${\rm FG}^{ee\gamma}$}-\mbox{${\rm MBG}_{\rm cor}^{ee\gamma}$}-(1+f_{ee\gamma})\ \mbox{${\rm MBG}_{\rm uncorr}^{ee\gamma}$}$ (12) Figure 28: Extracted $N_{\gamma}^{\rm{\pi^{0},tag}}$ as a function of conversion-photon $p_{T}$ using the (red) true information and (blue) event- mixing technique. The bottom panel shows the ratio of the event-mixing result over the true information result. For each $p_{T}$ bin $N_{\gamma}^{\rm{\pi^{0},tag}}$ is extracted by counting the number of entries in a window around the $\pi^{0}$ peak ($0.09<\mbox{$m_{ee\gamma}$}<0.19$) GeV/$c^{2}$. Figure 28 shows $N_{\gamma}^{\rm{\pi^{0},tag}}$ as function of $p_{T}$ using the true MC- ancestry information and the event-mixing technique. Overall the agreement is very good, however, the result from the event-mixing technique is on average lower. This mismatch is accounted for in the systematic uncertainties on $R_{\gamma}$, which is discussed in more detail in the next section. Figure 29: Average conditional probability $\langle\epsilon_{\gamma}f\rangle$ as a function of conversion photon $p_{T}$. Figure 30: Ratio $R_{\gamma}^{\rm pseudo}$ as a function of conversion photon $p_{T}$. The dashed line gives a constant offset of 1.3% fit to the points, and the dashed band represents a $\pm 1.5\%$ range around unity. ### A.3 Completing the validation by determining $R_{\gamma}$ With $N_{\gamma}^{\rm{incl}}$ and $N_{\gamma}^{\rm{\pi^{0},tag}}$ established from the pseudodata, the conditional probability $\langle\epsilon_{\gamma}f\rangle$ remains to be determined to calculate $R_{\gamma}$ and fully validate the background-subtraction procedure. In the same way as for the data, a single $\pi^{0}$ simulation is embedded into pseudodata events. The $e^{+}e^{-}$ pairs and $e^{+}e^{-}\gamma$ combinations are reconstructed and counted as discussed in Sec. III.3. The extracted $\langle\epsilon_{\gamma}f\rangle$ is shown in Fig. 29 as a function of the conversion photon $p_{T}$. With $N_{\gamma}^{\rm{incl}}$/$N_{\gamma}^{\rm{\pi^{0},tag}}$ from the pseudodata and $\langle\epsilon_{\gamma}f\rangle$ from the embedded single $\pi^{0}$ simulation in hand, $R_{\gamma}$ is calculated using Eq. A1. The result is shown in Fig. 30, all points are close to unity indicating that the analysis procedure is self consistent. There may be a 1.5% enhancement above unity, which is consistent with the slightly lower-than-expected value found for $N_{\gamma}^{\rm{\pi^{0},tag}}$. This difference is taken into account in the estimate of the systematic uncertainty. ## Appendix B Uncertainty propagation with a MC sampling method The uncertainties on $\gamma^{\rm dir}$ and any other quantity derived from $\gamma^{\rm dir}$, such as $T_{\rm eff}$ or $\alpha$, are determined using a MC-sampling method, which allows taking into account the $p_{T}$ and centrality dependent correlations of individual sources of systematic uncertainties, as well as the fact that the region $R_{\gamma}$ $<1$ is unphysical. ### B.1 Systematic uncertainties In the MC-sampling method, for each source of uncertainty, $i$, a variation $\delta_{i}$ of $R_{\gamma}$ or $\gamma^{\rm hadr}$ is sampled from a Gaussian distribution centered at zero with a width corresponding to the associated uncertainty, $\sigma_{i}$. The size of $\delta_{i}$ depends not only on $\sigma_{i}$, but also on whether the adjacent bins in $p_{T}$ and centrality have uncorrelated (Type A) or correlated (Type B/C) uncertainties due to the source $i$. The values of $\sigma_{i}$ and classification of each source is summarized in Table 2. If source $i$ is classified as uncorrelated, $\delta_{i}$ is calculated independently for neighboring bins from Gaussian distributions of width $\sigma_{i}$. For correlated uncertainties of Type C in $p_{T}$ or centrality, $\delta_{i}$ is calculated with one common fraction $w$ so that $\delta_{i}=w\sigma_{i}$ for all points. The fraction $w$ is determined randomly from a Gaussian distribution of width 1. And finally, for Type B uncertainties, $\delta_{i}$ is determined separately for the minimum and maximum of the $p_{T}$ or centrality range using the same procedure as Type C. All intermediate points are varied proportionally to create a smooth transition from the minimum to the maximum of the range. Uncertainties on the input $\pi^{0}$ $p_{T}$ distribution are a special case of Type B uncertainties, as it is known that the systematic uncertainties move simultaneously either up or down. In this case, $\delta_{i}$ at the minimum and maximum of the range are chosen to have the same sign. After applying all variations $\delta_{i}$ to recalculate $R_{\gamma}$ and $\gamma^{\rm hadr}$, new values of $\gamma^{\rm dir}$, $T_{\rm eff}$, and $\alpha$ are determined. This process is repeated multiple times, taking into account the different sources of uncertainties, to obtain a distributions of $\gamma^{\rm dir}$, $T_{\rm eff}$, and $\alpha$. The width of these distribution is quoted as the systematic uncertainty. For individual $\gamma^{\rm dir}$ points, it is possible that $\langle\mbox{$\gamma^{\rm dir}$}\rangle-\sigma$ is less than 0, that is, unphysical. In such cases, an upper limit of 90% confidence level (CL) is quoted based on the part of the probability distribution in the physical region $\int_{0}^{\rm upper}/\int_{0}^{+\infty}=90\%$. ### B.2 Statistical uncertainties The statistical uncertainties on $R_{\gamma}$ are assumed to have a Gaussian probability distribution and for most cases the statistical uncertainty on $\gamma^{\rm dir}$ can be calculated with the usual error propagation. However, there are two cases that need to be treated separately: * • $\mbox{$R_{\gamma}$}<1$: In this case $\gamma^{\rm dir}$ is unphysical, and hence an upper limit at 90% CL is quoted, based on the physical part of the probability distribution $\int_{0}^{\rm upper}/\int_{0}^{+\infty}=90\%$. * • $\mbox{$R_{\gamma}$}-\sigma_{\rm stat}<1$: In this case $\gamma^{\rm dir}$ is in the physical region, but consistent with zero within less than one standard deviation. For these situations the central value is shown, but the uncertainty is given as 90% CL, calculated as above. ## References * Shuryak [1978] E. V. Shuryak, Quark-gluon plasma and hadronic production of leptons, photons and psions, Phys. Lett. B 78, 150 (1978). * David [2020] G. David, Direct real photons in relativistic heavy ion collisions, Rept. Prog. Phys. 83, 046301 (2020). * Adare _et al._ [2010] A. Adare _et al._ (PHENIX Collaboration), Enhanced production of direct photons in Au+Au collisions at $\sqrt{s_{NN}}$=200 GeV and implications for the initial temperature, Phys. Rev. Lett. 104, 132301 (2010). * Adare _et al._ [2012a] A. Adare _et al._ (PHENIX Collaboration), Observation of direct-photon collective flow in $\sqrt{s_{NN}}$=200 GeV Au+Au collisions, Phys. Rev. Lett. 109, 122302 (2012a). * Adare _et al._ [2016a] A. Adare _et al._ (PHENIX Collaboration), Azimuthally anisotropic emission of low-momentum direct photons in Au$+$Au collisions at $\sqrt{s_{NN}}$=200 GeV, Phys. Rev. C 94, 064901 (2016a).
# FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design Yangyang Yu∗, Haohang Li∗, Zhi Chen∗, Yuechen Jiang∗, Yang Li∗ Denghui Zhang, Rong Liu, Jordan W. Suchow, Khaldoun Khashanah† Stevens Institute of Technology Hoboken, NJ, United States {yyu44, hli113, zchen100, yjiang52, yli269, dzhang42, rliu20<EMAIL_ADDRESS> ###### Abstract Recent advancements in Large Language Models (LLMs) have exhibited notable efficacy in question-answering (QA) tasks across diverse domains. Their prowess in integrating extensive web knowledge has fueled interest in developing LLM-based autonomous agents. While LLMs are efficient in decoding human instructions and deriving solutions by holistically processing historical inputs, transitioning to purpose-driven agents requires a supplementary rational architecture to process multi-source information, establish reasoning chains, and prioritize critical tasks. Addressing this, we introduce FinMem, a novel LLM-based agent framework devised for financial decision-making. It encompasses three core modules: Profiling, to customize the agent’s characteristics; Memory, with layered message processing, to aid the agent in assimilating hierarchical financial data; and Decision-making, to convert insights gained from memories into investment decisions. Notably, FinMem’s memory module aligns closely with the cognitive structure of human traders, offering robust interpretability and real-time tuning. Its adjustable cognitive span allows for the retention of critical information beyond human perceptual limits, thereby enhancing trading outcomes. This framework enables the agent to self-evolve its professional knowledge, react agilely to new investment cues, and continuously refine trading decisions in the volatile financial environment. We first compare FinMem with various algorithmic agents on a scalable real-world financial dataset, underscoring its leading trading performance in stocks. We then fine-tuned the agent’s perceptual span and character setting to achieve a significantly enhanced trading performance. Collectively, FinMem presents a cutting-edge LLM agent framework for automated trading, boosting cumulative investment returns. ††footnotetext: Corresponding author. Email: <EMAIL_ADDRESS>Equal contribution, with author order decided by dice roll.22footnotetext: The source code of this project can be found via: FinMem LLM Trading KEYWODS: Financial AI, Large Language Model, Trading Algorithms, Deep Learning, Financial Technology ## 1 Introduction With the influx of diverse financial data streams from the web, traders face a deluge of information from various sources. This requires them rapidly to understand, memorize, and filtrate crucial events for investment decisions. However, innate cognitive limitations restrict human traders from processing information within their perception and memory capacity, a span much narrower than the actual volume of available information [5]. Consequently, insufficiently considering or even dismissing critical events affecting trading decisions becomes increasingly concerning as data availability expands. To overcome the physical limitations in the memory systems of human traders, researchers have been consistently working on designing autonomous trading agent systems. These systems need to thoroughly integrate all available information and possess a sophisticated design in the agent’s backbone algorithm to deliver enhanced trading performance. The evolution of autonomous trading systems has transitioned from the initial rule-based trading strategies [12] to more advanced machine-learning-based algorithms [19]. In recent years, Reinforcement Learning (RL)-based agents [14], especially those employing Deep Reinforcement Learning (DRL) [29] as backbone algorithms, garner joint attention of both academia and industry. Leveraging both RL principles and deep learning, DRL agents effectively handle and learn from scalable and diverse financial data, including stock prices, key financial indicators, and market sentiments. They utilize deep neural networks to extract expressive features from input data, representing the complex financial market environment, which enhances their comprehension ability. Retaining the key features of RL agents, they learn through interaction with a predefined environment to maximize investment gain over time. Research suggests DRLs can meet the crucial needs of trading agents to process and make informed decisions from large volumes of data. However, certain inherent features of DRL algorithms exhibit notable deficiencies in financial applications. Firstly, DRL agents exhibit a lack of interpretability concerning the rationale behind their decisions [4]. They are often described as “black boxes,” where the internal processes and computational layers leading to a specific decision are neither easily understandable nor transparent. Secondly, DRL agents find it challenging to effectively integrate textual data with numerical features. Text data plays a vital role in finance, since the majority of market information is conveyed through news articles and financial reports. However, transforming text data to embeddings considerably increases input space dimensionality, making learning more computationally demanding. Plus, empirical studies have shown that combining textual representations with numerical financial indicators often leads to convergence challenges [15]. Therefore, a backbone algorithm with transparent reasoning and the enhanced ability to capture investment-related textual insights comprehensively is essential. Recent advancements in Large Language Models (LLMs), like Generative Pre- trained Transformers (GPTs) [34], offer viable solutions for developing trading agents, alleviating previous concerns. With carefully designed prompts, the LLM-based agent is able to provide reasons and outcomes in plain text. This allows for immediate observation and prompt adjustment of its reasoning process. Employing LLMs as the backbone algorithms for agents also overcomes the constraint of isolated environments. This is achieved through their vast pre-existing knowledge and the effective integration of valuable insights from a variety of data sources, including both textual and numerical. When equipped with suitable prompt templates, this approach significantly enhances decision-making capabilities [50]. Studies indicate that prompt- guided reasoning significantly improves problem-solving rates across various domains [24]. Notably, a growing body of research has focused on utilizing LLMs to make informed trading decisions for stocks and funds by continuously interacting with financial environment information [55, 52]. However, in currently available approaches, LLMs primarily serve as a QA role rather than functioning as autonomous agents. The potential issue with these approaches is the incapability of fully understanding the varying timeliness associated with different types of financial data. These financial LLM agents, despite outperforming traditional trading benchmarks, generally process information indiscriminately through QA iterations, lacking the ability to memorize influential messages. Furthermore, their method of acknowledging the timeliness of financial data is heavily dependent on the uncertain and laborious LLM fine-tuning process. These insufficiencies undermine their ability to update the knowledge base in a daily manner, meaning they lack a memory component. As a result, they may struggle to prioritize significant and influential memory events effectively. Additionally, current literature on LLM-based trading agents lacks a comparative analysis between these applications and other autonomous trading systems, such as DRL agents. To bridge this gap, we present FinMem, an innovative LLM-based autonomous trading agent with a novel layered memory system and dynamic character design. Unlike previous LLM agents in finance, FinMem encompasses a memory module adept at processing multi-source financial data with varying timeliness and self-adaptive character setting for fitting into volatile market environments. Our concept is initially inspired by the Generative Agents framework by Park et al. [35], aimed at enhancing the efficient retrieval of key events for general-purpose LLM agents. This framework features a unique character design and seed memory, activating the agent upon specific query through prompts. It prioritizes events in a unified memory stream, ranked by a linear combination of recency, relevancy, and importance. The framework outlined in [35] provides a foundational structure for LLM agent design. It includes a profiling module for character definition, a memory module for experience recording and critical information retrieval, and an action module to guide actions based on the retrieved memories. This structure effectively facilitates goal achievement for the agent in a general social environment. However, Park et al.’s framework struggles with comprehending financial data with varying timeliness and importance, like daily news versus quarterly and annual reports. Key challenges involve quantifying the timeliness of different information sources, optimizing information retrieval, and enhancing trading decisions with detailed analysis. To tackle these challenges, we further propose FinMem with the following improvements. FinMem maintains a modular approach similar to Park et al. [35], but features novel design of profiling and memory modules. Its specialized profiling module equips FinMem with a trading-task-specific professional background, enhancing robustness to market fluctuations via offering self-adaptive risk inclination option. FinMem’s memory module innovatively incorporates working memory and layered long-term memory components, ideal for stratified information processing. Its working memory acts as a dynamic “workspace,” enabling operations like summarization, observation, and reflection on multi-source information to facilitate trading decisions. Its long-term memory, structured into shallow, intermediate, and deep layers [9], manages varied decay rates to satisfy the need to retain distinct types of financial information within different time scales based on their corresponding timeliness. For instance, daily news, with its immediate effects on stock markets, is channeled into the shallow processing layer. Meanwhile, annual company reports, exerting a more prolonged impact, are processed in the deep layer by FinMem. Each layer in FinMem prioritizes memory events based on the assemble of recency, relevancy, and importance close to Park et al.’s method. However, it introduces new measurements for recency and importance, specifically tailored to better rank financial data according to their unique time sensitivity. FinMem’s memory mechanism can also transit significantly impactful investment memory events to deeper processing layers, ensuring their retention for extended periods. FinMem’s memory module can mirror the human cognitive system [47] and facilitate agile, real-time decisions [46]. It enables continuous evolution in professional knowledge through structured summarizing, retrospecting past experiences, and reacting to new trading scenarios. Additionally, FinMem includes a decision-making module capable of deriving investment decisions by considering top-ranked memory events and current market conditions. Through experiments, we show that FinMem exhibits an outstanding ability to stratify and leverage the various levels of market insights, significantly improving the quality of trading decisions. We claim that FinMem provides these key contributions: FinMem presents a state-of-the-art LLM-based trading agent with a human- aligned memory mechanism and character design, particularly crafted to capture investment insights from the financial market. In its agent memory module design, FinMem innovatively emulates human working and layered long-term memory mechanisms. This approach effectively harnesses the time-sensitive aspects of financial data, capturing crucial investment insights and thereby boosting trading performance. FinMem’s profiling module includes a dynamic character setting feature, offering seed information about professional backgrounds and adjustable risk inclinations. Additionally, it continuously updates domain knowledge as the trading experience grows, thereby enhancing FinMem’s capabilities. Ablation studies demonstrate that FinMem can learn from past trading experiences and evolve its knowledge base through continuous market interaction, maintaining robustness in complex markets for profitable trading decisions. FinMem can utilize its distinctive features to expand the agent’s perceptual range beyond the human limitation to make well-informed trading decisions. Cognitive research indicates that human working memory can recall only five to nine events at a time [30]. This limitation, while preventing information overload, can yield insufficient insight for precise decision-making. In contrast, FinMem’s memory module transcends this constraint. It allows adjusting cognitive load by selecting a flexible number of top-ranked events from each layer of its hierarchical long-term memory, allowing FinMem to maintain agility and deliver superior trading decisions in data-rich contexts. FinMem achieves impressive trading performance using training data that is limited in volume and spans a short time period. Our experiments indicate that training FinMem with daily collected data over just six months to a year is enough to produce robust and notable trading results. This timeframe is considerably shorter than what other comparable models require. This efficiency is achieved through the optimal utilization of multi-source incoming data and the precise identification of key signals for trading decisions. Furthermore, it’s noteworthy that FinMem’s effectiveness is demonstrated on smaller datasets and with general-purpose LLMs. Its capabilities are anticipated to be further amplified with access to larger, higher-quality financial datasets and LLMs fine-tuned specifically for financial applications. In this paper, we begin by explaining the three core modules of FinMem. Subsequently, we emphasize its superior trading performance compared to a range of representative algorithmic agents. We further explore how FinMem achieves its optimal performance, examining adjustments in three key aspects: backbone algorithms, working memory capacity, and character settings. ## 2 Related Work ### 2.1 Backbone Algorithms of Contemporary Autonomous Trading Agents The development of trading agents has evolved over several decades, influenced by advancements in technology, finance, and computational methodologies. Conventionally, a rule-based algorithm for trading stocks is an automated strategy that operates based on a predefined set of rules[7, 49, 37]. These rules are often derived from historical market patterns and trading experience. Compared with rule-based algorithms that use predefined rules and conditions, Reinforcement Learning provides a way for agents to learn by interacting with an environment and receiving feedback in the form of rewards or penalties. [10, 22]. Deep learning models can be integrated with RL to handle large and complex state spaces, like those in stock markets. Such models are often referred to as Deep Reinforcement Learning (DRL) [53, 54]. For example, Deep Q-Network (DQN) [44], Advantage Actor-Critic (A2C) [56], and Proximal Policy Optimization (PPO) [27] are popular algorithms for such tasks. Using DRL agents as automated financial trading backbones, face two key issues: 1) A lack of interpretability, as their decisions, rooted in complex computations and high-dimensional representations, are challenging to articulate [4]. 2) They struggle to fully leverage textual financial information due to the high-dimensional nature and computational intensity of rich text embeddings [11, 13]. Consequently, DRL agents often rely on extracting textual sentiment [39], sidestepping the direct use of embeddings [6, 3], leading to an incomplete representation of crucial market information embedded in news and macroeconomic policies. ### 2.2 Advancements from LLMs to LLM Autonomous Agents The evolution of LLMs has reshaped artificial intelligence and natural language processing. From foundational embeddings like Word2Vec [16] and GloVe [38], the field advanced with the introduction of sequential modeling like Long Short-Term Memory (LSTM) [18] and early transformer models of Bidirectional Encoder Representations from Transformers (BERT) [11]. Today, the new-generation LLMs, like Generative Pre-trained Transformer series (GPTs) [40, 34] and LLM Meta AI (Llamas) [48], stand out in diverse QA tasks. The trend leans towards LLM agents. While LLM agents for domain-specific tasks have been extensively researched [20, 25, 36], their application in financial trading remains underexplored. Existing studies in this domain, such as [52, 55], often lack open-source availability or have not considered an architecture specifically tailored to fit the unique environment of finance markets. Thus, there’s significant value in further investigating advanced, transparent LLM agents for trading. ### 2.3 Architecture Design of LLM Autonomous Agent As Wang et al.[50] emphasizes, an effective architecture for LLMs serving autonomous agents is essential. Typically, this structure comprises modules like profiling, memory, planning, and actions, though not all may be essential for every application. There are cases of two modules (e.g., planning and action modules) being integrated as one component ([35]). The design variations are numerous. For instance, profiling has been achieved through methods like handcrafting [57], generation by LLMs [51], and alignment with real-world datasets [2]. Among these modules, the memory component is essential. Acting as the operational core, it aligns an agent’s actions with real-world tasks. Research indicates that leveraging insights from cognitive science studies on human memory [50, 45] can enhance this alignment. Thus, a well-structured trading LLM agent, comprising aptly designed modules, can sharply tackle the complexities of financial markets to make informed decisions. ## 3 Architecture of FinMem In this section, we comprehensively detail the three core components of FinMem, namely the profiling, memory, and decision-making modules. The profiling module empowers FinMem to adaptively tailor character setting for specific trading tasks. Memory module leverages diverse time-efficiency attributes of financial data, enhancing its trading efficacy. The decision- making module enables FinMem to synchronize its memory streams with market facts, facilitating high-quality trading decisions. The details and notations associated with these three modules are provided in the subsequent sections. Figure 1: The prompt template for FinMem’s profiling module. It includes two key elements of its character setting: professional background knowledge and three distinct investment risk inclinations. In the self-adaptive risk inclination option, the omitted texts align with the detailed descriptions provided for the risk-seeking and risk-averse inclinations. ### 3.1 Profiling Module The profiling module empowers FinMem to develop a dynamic agent character specifically designed to navigate the complex dynamics of financial markets effectively. The dynamic character of FinMem comprises two principal components, as depicted in Figure 1: firstly, a foundational professional knowledge base akin to a trading expert, and secondly, an agent with three distinct investment risk inclinations. The first component includes two types of information: an introduction to the primary trading sectors relevant to the company stock FinMem will trade in, and a concise overview of the historical financial performance of the specified ticker, spanning from the beginning to the end of the training period. Before initiating trades in a new company’s stock, FinMem accesses and updates this sector-specific and historical financial data from a backend database. This professional background setting narrows down information and memory events pertinent to specific trading tasks. The second component of FinMem’s design, illustrated in Figure 1, encompasses three distinct risk inclination options: risk-seeking, risk-averse, and a self-adaptive risk character. The risk-seeking setting gears FinMem towards an aggressive, high-reward approach, while the risk-averse setting gears it towards a conservative, lower-risk strategy. A distinctive aspect of FinMem is its ability to dynamically alternate between these risk settings in response to current market conditions. Specifically, it shifts risk preferences when the Cumulative Return falls to below zero within a brief period, such as three days, and reversely. This flexible design functions as a protective mechanism, mitigating prolonged downturns in turbulent market environments. During the initial stage of the training phase, FinMem is configured with a chosen risk preference, each supplemented with comprehensive textual explanations through LLM prompts. These guidelines shape how FinMem processes incoming messages and determines its subsequent actions in alignment with its designated risk inclination. The system maintains a catalog of all risk inclinations and their detailed explanations in a backlog, enabling seamless adaptation to different stocks by switching among these risk profiles as needed. The dynamic character setting in FinMe’s profiling module provides subjective and professional background knowledge and flexible choice of risk inclinations. It provides crucial context for filtering and retrieving trading-relevant information and memory events, thus improving accurate inferencing and adaptability to fluctuating market conditions. ### 3.2 Memory Module The memory module of FinMem emulates a human trader’s cognitive system so that it can efficiently process hierarchical financial information and prioritize the critical messages for high-quality investment decisions. Furthermore, it adjusts the memory span flexibly, enabling the agent to operate on a wider range of events over a longer retrieval period. FinMem’s memory module, illustrated in Figure 2, comprises working and long-term memory with layered processing capability and is initiated by a specific investment inquiry. Figure 2: Memory module structure of FinMem with a detailed view of components, operations, and workflow. The cognitive architectures of FinMem’s memory module have two core components – Working Memory and Layered Long-term Memory. #### 3.2.1 Working memory Working memory refers to the human cognitive system’s functions for temporary storage and diverse operations. We incorporate this concept into FinMem’s memory module development, creating a central workspace for informed decision- making. Unlike human working memory, having a maximum capacity of seven plus or minus two memory events [30], FinMem has the ability to expand the capacity based on specific requirements. Tailored for converting financial data into trading actions, FinMem’s working memory encompasses three key operations: summarization, observation, and reflection. The mechanisms by which they interact and operate as an integrated decision-making workflow are detailed in the middle box of Figure 2. Additionally, the LLM prompt template that underpins these processes is thoroughly outlined in 3. Summarization: FinMem leverages external market data to derive critical investment insights and sentiments tailored to specific stock trading queries, such as “Can you make an investment decision on TSLA on 1/24/2023?”. As illustrated in Figure 3 (2), this system condenses the original text into a compact yet informative paragraph, thereby enhancing FinMem’s processing efficiency. It efficiently extracts and summarizes pertinent data and sentiments for stock investment decisions, demonstrated here using Tesla Inc. as an example. Subsequently, FinMem directs these insights to an appropriate layer within its long-term memory architecture, selecting the layer based on the time sensitivity of the information. Observation: Triggered the same inquiry, FinMem initiates an observation operation to gather market facts. The information available to FinMem varies between the training and testing phases. During the training phase, FinMem has access to comprehensive stock price data within the specified period. Upon receiving trading inquiries that specify a stock ticker and date, FinMem focuses on the daily adjusted closing price differences, comparing the following day’s price with the current day’s. These price differences are utilized as market ground labels. Specifically, a decrease in price suggests a “Sell” action, while an increase or no change in price indicates a “Buy” action. During the testing phase, at a specific time point, FinMem loses the ability to access future price data. Its focus shifts to the analysis of historical stock price movements, depending on a retrospective evaluation of the cumulative return from the last $M$ trading days to infer future market trends. This phase, characterized by the absence of foreseen market grounds, serves as a critical assessment of FinMem’s development. It tests whether the system has adequately established logical connections between stock price trends and various financial information sources, such as news, reports, and indicators. This stage is key in evaluating FinMem’s capability of independently evolving its trading strategies for subsequent tasks, leveraging its analysis and interpretation of historical data patterns. Reflection: Two types of reflections exist, immediate and extended reflection. (a) Immediate reflection is activated upon receiving a daily trading inquiry for a specific ticker. Using LLM and specific prompts exemplified in Figure 3 (2), the agent merges market indications and top-$K$-ranked events from each long-term memory layer. Market indications are derived from the outcomes of the observation operation and differ between the training and testing phases. During testing, this process yields three types of outputs: the trading direction (“Buy”, “Sell”, or “Hold”), the underlying rationale for this decision, and the most influential memory events, along with their IDs from each layer that informed the decision. In the training phase, specifying the trading direction is unnecessary, as FinMem is already informed of future stock movement directions. The top-$K$-ranked memory events encapsulate key insights and sentiments derived from critical investment-related incoming messages, all distilled by FinMem’s advanced summarization capabilities. (b) Extended reflection reevaluates immediate reflection outcomes for a ticker over a specified $M$-day trace period. It encompasses data like stock price trends, trading returns, and action rationales from multiple immediate reflections. While immediate reflection enables direct trading execution and records current feedback, extended reflection summarizes market trends and reassesses recent Cumulative Return on investment. Extended reflection is eventually transmitted and stored in the deep processing layer to emphasize its criticality (detailed introduced in Section 3.2.2) of long-term memory. $K$ and $M$ are hyperparameters to adjust FinMem’s working memory capacity and information retrieval ability. FinMem gains the flexibility of integrating comprehensive information into well-informed decisions by fine-tuning them. #### 3.2.2 Layered long-term memory Figure 3: (1) The decision-making module workflow of the FinMem trading agent retrieves critical memory events to inform specific decisions. (2) LLM prompt template used by FinMem to interact with incoming financial information. FinMem’s long-term memory organizes hierarchical financial data insights in a stratified structure, as illustrated in the lower section of Figure 2. Drawing inspiration from the varying decay speeds in the human cognitive system’s information processing layers [9], FinMem employs a layered structure to accommodate the diverse time sensitivities inherent to different types of financial data. This structure categorizes summarized insights by their timeliness and decay rates. Insights are derived by the working memory’s summarization operation. Those directed to deeper layers receive smaller decay rates, indicating longer retention, while those in shallower layers are assigned larger decay rates for shorter retention. $\gamma_{l}^{E}=S_{\text{Recency}_{l}}^{E}+S_{\text{Relevancy}_{l}}^{E}+S_{\text{Importance}_{l}}^{E},$ (1) where each memory event is only associated with one score and can only belong to a single layer. Upon receiving an investment inquiry, FinMem retrieves the top-$K$ pivotal memory events from each layer and channels them to the immediate reflection component of the working memory. These events are chosen according to the descending order of their information retrieval score, denoted as $\gamma_{l}^{E}$, where $l$ belongs to the set shallow, intermediate, deep, as specified in Equation 5. $E$ denotes a given memory event. This score, adapted from Park et al. [35] but with modified recency and importance computations, especially tailoring to handle data with various timelines. It encapsulates three metrics: recency, relevancy, and importance. Individual metric scores exceeding 1.0 are scaled to the [0,1] range before being summed. The modification is to achieve the layered processing function and represent the various periodicity of the financial environment. $\begin{split}&S_{\text{Recency}_{l}}^{E}=e^{-\frac{\delta^{E}}{Q_{l}}},\quad\;\delta^{E}=t_{\text{P}}-t_{E},\\\ \end{split}$ (2) where $\delta^{E}$ refers to the time difference between the memory event occurrence and the trading inquiry arrival. $Q_{\text{shallow}}=14$, $Q_{\text{intermediate}}=90$, and $Q_{\text{deep}}=365$ correspond to day counts of two weeks, a quarter, and a year for shallow, intermediate, and deep processing layers, respectively. Upon a trade inquiry $P$ arrival in processing layer $l$ via LLM prompt, the agent computes the recency score $S_{\text{Recency}_{l}}^{E}$ per Equation.2. $S_{\text{Recency}_{l}}^{E}$ inversely correlates with the time gap between the inquiry and the event’s memory timestamp, mirroring Ebbinghaus’s forgetting curve [33]. The stability term $Q_{l}$ in Equation.2 partially controls memory decay rates across layers, indicating longer memory persistence in the long-term layer with a higher stability value. In the context of trading, company annual reports, such as Form 10-Ks, are considered to have more extended timeliness compared to daily financial news. Therefore, they are assigned a higher stability value and are categorized within the deeper processing layer. This classification reflects their extended relevance and impact in financial decision-making scenarios. $\begin{split}&S_{\text{Relevancy}_{l}}^{E}=\frac{\mathbf{m_{E}}\cdot\mathbf{m_{P}}}{\|\mathbf{m_{E}}\|_{2}\times\|\mathbf{m_{P}}\|_{2}}\end{split}$ (3) The relevancy score, denoted as $S_{\text{relevancy}_{l}}^{E}$, quantifies the cosine similarity between the embedding vectors. These vectors are derived from the textual content of the memory event, $\mathbf{m_{E}}$, and the LLM prompt query, $\mathbf{m_{P}}$, using OpenAI’s “text-embedding-ada-002” model, as depicted in Equation 3. The LLM prompt query incorporates inputs related to trading inquiries and the trading agent’s character setting. The importance score $S_{\text{Importance}_{l}}^{E}$ is computed using the value $v_{l}^{E}$ from a uniform piecewise scoring function (Formula 4), multiplied by a degrading ratio $\theta_{l}$ (Formula 5) as per Equation 6. The likelihood of higher $v_{l}^{E}$ values increases from shallow to deep layers. $\theta_{l}$ measures the diminishing importance of an event over time, which has a close form design of [35]. But our approach tailors $\theta_{l}$ to the stratified structure of long-term memory. It adopts unique exponential functions for each layer. The base $\alpha_{l}$ for each layer is a hyperparameter, set to follow the sequence: $\alpha_{shallow}<\alpha_{intermediate}<\alpha_{deep}$. These values correlate with the rate at which their importance degrades after a certain period, providing another angle to measure importance variances across different memory types. Through experimentation, we set $\alpha_{shallow}=0.9$, $\alpha_{intermediate}=0.967$ and $\alpha_{deep}=0.988$. This ensures that $\theta_{l}$ decreases to a threshold score of $5$ after intervals of $30,90,$ and $365$ days for shallow, intermediate, and deep layers, respectively. The three-piece-wise functions for $S_{\text{Importance}_{l}}^{E}$ and $S_{\text{Recency}_{l}}^{E}$ enable FinMem to have layered processing in the long-term memory component. Memory events are purged when $S_{\text{Recency}_{l}}^{E}$ is below $0.05$ or $S_{\text{Importance}_{l}}^{E}$ is under $5$ (pre-scaling). $\begin{split}&v_{l}^{E}=\begin{cases}40&\text{with probability }p_{1}\\\ 60&\text{with probability }p_{2}\\\ 80&\text{with probability }p_{3}\end{cases}\end{split}$ (4) $\begin{split}\theta_{l}=(\alpha_{l})^{\delta^{E}},\quad&l=\text{shallow},\text{intermediate},\text{deep},\end{split}$ (5) where $p_{1}+p_{2}+p_{3}=1$, but their values vary by shallow, intermediate, and deep processing. when shallow processing ${p_{1},p_{2},p_{3}}=\\{0.8,0.15,0.05\\}$, intermediate processing, ${p_{1},p_{2},p_{3}}=\\{0.05,0.8,0.15\\}$ and deep processing, ${p_{1},p_{2},p_{3}}=\\{0.05,0.15,0.8\\}$. $S_{\text{Importance}_{l}}^{E}=v_{l}^{E}*\theta_{l},$ (6) Furthermore, an access counter function oversees the transfer of memory events among layers, ensuring that significant events influencing trading decisions ascend from shallower to deeper layers for extended retention and recurrent access by FinMem. Conversely, less pertinent events gradually diminish. This process is facilitated by the LLM validation tool Guardrails AI [17], which monitors critical memory IDs across different layers. An event identified as pivotal for investment success receives an additional $5$ points in its importance score $S_{\text{Importance}_{l}}^{E}$. Upon meeting the criteria for upgrading to a deeper layer, an event’s recency score $S_{\text{Recency}_{l}}^{E}$ is reset to $1.0$, emphasizing its importance and preventing rapid decay. By implementing this access counter, FinMem effectively identifies and prioritizes key events, taking into account their nature and frequency of retrieval. ### 3.3 Decision-making Module The decision-making module of FinMem efficiently integrates operational outcomes from the profiling and memory modules to support well-informed investment decisions, as depicted in Figure 3 (1). In its daily trading decisions, FinMem is asked to select from three distinct actions for a single share of a specific stock by Guardrails AI text validation function: “Buy”, “Sell”, or “Hold”. Additionally, the inputs and results required by FinMem’s decision-making module vary between its training and testing phases, with each phase’s specifics detailed as follows: During the training phase, FinMem accesses a wide array of multi-source information relevant to the entire time period. When FinMem is prompted with trading inquiries containing stock ticker and date, as well as trader character-related texts, it concurrently initiates observation and summarization operations in its working memory. FinMem observes the market ground labels mentioned in the description about the observation operation in Section 3.2.1, which involve daily adjusted price differences between consecutive days, indicative of “Buy” or “Sell” actions. Utilizing these price change signals, FinMem identifies and prioritizes the top-$K$ memories, ranking them based on retrieval scores from each long-term memory layer. This procedure enables FinMem to produce comprehensive reflections that provide a well-founded rationale and in-depth inference of the correlation between market ground labels and the memories retrieved. Through repeated trading operations, reflections, and memory events with significant impact, transition to a deeper memory processing layer, getting preserved for guiding future investment decisions during the testing phase. In the testing phase, where FinMem cannot access future price data, it relies on the Cumulative Return over the previous $M$ trading days to anticipate future market trends. To compensate for the absence of future market price information, FinMem utilizes enhanced reflections derived from immediate reflections spanning an $M$-trading-day period as supplementary references. When faced with a specific trading inquiry, FinMem integrates insights from various sources, including historical Cumulative Return, outcomes from extended reflection, and the Top-$K$ retrieved memories. This comprehensive approach enables FinMem to execute well-informed trading decisions. It should be noted that FinMem generates executable actions exclusively in the immediate reflection operation of the testing phase. Since the trading direction is guided by the actual price trend, the training phase of FinMem does not make investment decisions. Instead, this phase is dedicated to accumulating trading experience through comparing market trends with incoming multi-source financial messages. Additionally, during this phase, FinMem develops a memory module enriched with a comprehensive knowledge base, thereby evolving its capability for independent decision-making in future trading activities. ## 4 Experiments Setups We aim to evaluate the trading performance of FinMem. And we further illustrate its unique advantages of requiring significantly less historical trading time window to train and take full use of key financial data time series as well as textual information. Specifically, we conducted several experiments to study the following research questions (RQs): * • RQ1: Is FinMem capable of outperforming contemporary state-of-the-art algorithmic trading agents? * • RQ2: Are there tasks that challenge other trading algorithms but are manageable by FinMem? * • RQ3: Which LLM is best suited to form the backbone framework of FinMem? * • RQ4: Does equipping FinMem with different risk inclination choices truly differentiate its trading performance? * • RQ5: Can FinMem effectively filter and prioritize information to facilitate informed trading decisions? In the rest of the section, we begin by introducing the real-world financial dataset used in our experiments. We then describe the comparative algorithmic agents and list several widely used financial metrics. Our experiments fall into two categories: 1) The comparative experiments of FinMem versus other algorithmic trading agents, and FinMem using different LLMs as backbone algorithms. 2) The ablation studies evaluate the effects of FinMem’s adjustable cognitive span and the role of the trader’s dynamic character settings, particularly the risk inclinations, on its trading performance. Through experiments, FinMem demonstrates to outperform other comparative algorithmic agents. Furthermore, we are able to show that its profiling and memory modules are sophisticated and tailored to effectively address the intricacies of the financial landscape, resulting in superior trading performance. ### 4.1 Datasets And Database Structure: We assessed FinMem’s performance using multi-source financial data from August 15, 2021, to April 25, 2023, sourced from reputable financial databases and APIs like Yahoo Finance (via yfinance) and Alpaca News API, detailed explained in Table 1. The stock tickers used in our comparative experiments are detailed in Figure 4. These were selected because they are among those with the highest volumes of accessible news text data, and they are spread across various trading sectors. This selection provides ample data to evaluate FinMem’s generalization capabilities. Additionally, Tesla, Inc. (TSLA) was specifically chosen for ablation studies due to its association with the largest amount of textual data, offering sufficient information to assess performance differences for the FinMem’s key features like cognitive spans. The raw multi-source input data, initially stored in the “Raw Financial Data Warehouse”, are diverged into FinMem’s “Layered Long-term Memory Data Warehouse” based on timeliness through working memory’s summarization operation in Figure 2. The deep processing layer holds annual reports (Form 10K’s) insights, the intermediate layer contains quarterly reports (Form 10Q’s) insights, and the shallow layer accommodates daily financial news insights. We leveraged the open-source vector database FAISS [23] for constructing the memory warehouse of FinMem, benefiting from its rapid querying in high- dimensional vectors and compatibility with OpenAI for cosine-similarity-based semantic searches on specific tickers. This setup facilitates efficient top- ranked event retrieval. Data categorization and memory module workflow are also illustrated in Figure 2 Raw Financial Data Warehouse --- News data associated with ticker indexes: News data is sourced from the Alpaca News API, which utilizes Benzinga as its backend provider. Corporate quarter filings indexes: Quarterly reports (Form 10-Q) are required by the U.S. Securities and Exchange Commission (SEC). Corporate annual filings indexes: Annual reports (Form 10-K) required by the U.S. Securities and Exchange Commission (SEC). Stock price records: Daily stock open-high-close-volume (OHLCV) data from Yahoo Finance. FinMem’s Layered Long-term Memory Data Warehouse Shallow Processing: Insights of real-time market news extracted by LLM. Updated daily. Intermediate Processing: Insights of 10-Q filings extracted by LLM. Updated quarterly. Deep Processing: 10-K fillings, all summarized by LLM. Updated yearly. Extended reflections for the stock in response to the trading inquiry include FinMem’s cumulative trading returns, decision-making processes, trade volumes, and underlying reasons. Updated daily. Table 1: Raw data and memory warehouses of FinMem Figure 4: The distribution of news in scraped from Alpaca News API for the five stocks in the experiments ### 4.2 Baseline And Comparative models: We assess FinMem’s trading performance in comparison to five advanced algorithmic agents and a commonly accepted baseline trading strategy. Among these, three models employ Deep Reinforcement Learning (DRL) approaches, while the remaining two are based on Large Language LLMs. Brief descriptions of each are provided below: Buy-and-Hold strategy (B&H): A passive investment approach, where an investor purchases stocks and holds onto them for an extended period regardless of market fluctuations, is commonly used as a baseline for comparison of stock trading strategies. DRL trading agents: As the FinMem is practiced and examined on the basis of single stock trading and discrete trading actions, we choose three advanced DRL algorithms fitting into the same scenarios according to the previous and shown expressive performance in the work of Liu et al. [28, 26]. The DRL training agents only take numeric features as inputs. * • Proximal Policy Optimization (PPO): PPO [42] is employed in stock trading due to its stability and efficiency. One salient advantage of PPO is that it maintains a balance between exploration and exploitation by bounding the policy update, preventing drastic policy changes. * • Deep Q-Network (DQN): DQN [32] is an adaptation of Q-learning, that can be used to optimize investment strategies. Unlike traditional Q-learning that relies on a tabular approach for storing Q-values, DQN generalizes Q-value estimation across states using deep learning, making it more scalable for complex trading environments. * • Advantage Actor-Critic (A2C): A2C [31] is applied to optimize trading actions in the financial environment. It operates by simultaneously updating both the policy (actor) and the value (critic) functions, providing a balance between exploration and exploitation. LLM trading agents: We evaluate FinMem against two LLM agents in the context of stock trading. The first LLM agent, known for its proficiency in general-purpose tasks, serves as a baseline. The second agent, a leading-edge LLM in trading, has been acclaimed for its promising performance in stock market operations. * • General-purpose Generative Agents – GA: The generative AI agent by Park et al. [36], originally intended to simulate realistic human behavior and make everyday decisions, has been adapted here for specific stock trading tasks. This agent’s architecture includes a memory module that employs recency, relevance, and importance metrics to extract pivotal memory events for informed decision-making. However, it does not provide a layered memory module to effectively differentiate the time sensitivities unique to various types of financial data. Additionally, although it features a profiling module to define agent attributes like professional background, the model does not include a mechanism for self-adaptive risk preference. In our experiments, we modified the original prompt template created by Park et al., which was intended for general daily tasks, to suit financial investment tasks. The textual elements of this revised template closely align with those of FinMem, with the exception of two components that are absent in this version of general-purpose Generative Agents. * • LLM trading agents – FinGPT: A novel open-source LLM framework specialized for converting incoming textual and numeric information into informed financial decision-making, introduced by Yang et al. [55]. It claims superiority over the traditional buy-and-hold strategy. ### 4.3 Evaluation Metrics: Figure 5: Cumulative return comparison over time between FinMem and other algorithmic agents across five stocks. Figure 6: Cumulative Return comparison over time between FinMem and other algorithmic agents on Coinbase Global, Inc. (COIN). We employ five widely-used metrics in finance to compare the investment rewards of FinMem against other algorithmic trading agents. Here are their introductions: * • Cumulative Return [21]: Cumulative Return is a key trading performance metric because it provides a comprehensive insight into investment performance, especially for strategies that emphasize long-term growth and reinvestment. The effectiveness of different investment strategies is evaluated based on their Cumulative Returns, which reflect the total change in value over time. In this study, we compute Cumulative Returns over the specified period by summing daily logarithmic returns, as outlined in Equation 7. This method is widely accepted in the finance area due to its ability to precisely capture minor price fluctuations and symmetrically address gains and losses. In essence, a higher Cumulative Return typically indicates a more effective strategy. Cumulative Return $\displaystyle=\sum_{t=1}^{n}r_{i}$ $\displaystyle=\sum_{t=1}^{n}\left[\ln\left(\frac{p_{t+1}}{p_{t}}\right)\cdot\text{action}_{t}\right],$ (7) where $r_{i}$ represents the logarithmic return for day $t+1$, $p_{t}$ is the closing price on day $t$, $p_{t+1}$ is the closing price on day $t+1$, and $\text{action}_{t}$ denotes the trading decision made by the model for that day. * • Sharpe Ratio [43]: Sharpe Ratio is another core metric for evaluating investment performance and adjusting returns for risk. It is calculated by dividing the portfolio’s average excess return ($R_{p}$) over the risk-free rate ($R_{f}$) by its volatility ($\sigma_{p}$), as shown in Equation 8. This metric adjusts returns for risk, with a higher ratio indicating better risk- adjusted performance. Essential in comparing different portfolios or strategies, it contextualizes performance against similar investments. Although a Sharpe Ratio above 1 is typically considered favorable and above 2 as excellent, these benchmarks can vary depending on the context of comparison. $\textbf{Sharpe Ratio}=\frac{R_{p}-R_{f}}{\sigma_{p}}$ (8) * • Annualized Volatility and Daily Volatility[8]: Annualized Volatility (Annum- Volatility) is calculated as the Daily Volatility (standard deviation of daily logarithmic returns) multiplied by the square root of the typical number of trading days in a year (252) as outlined in Equation 9, is vital for assessing investment risk. This measure reflects the extent of fluctuation in a security or market index’s returns over a year, indicating potential deviations from average returns. It’s especially relevant for investors with specific risk profiles, such as those who are risk-averse, who may prefer portfolios demonstrating lower annualized volatility. Annum-Volatility $\displaystyle=\textbf{Daily Volatility}\times\sqrt{252}$ (9) * • Max Drawdown [1]: Max Drawdown is a metric for assessing risk. It represents the most significant decrease in a portfolio’s value, from its highest ($P_{\text{peak}}$) to its lowest point ($P_{\text{trough}}$) until a new peak emerges, detailed in Equation 10. Indicative of investment strategy robustness, a smaller Max Drawdown suggests reduced risk. $\displaystyle\textbf{Max Drawdown}=\text{max}(\frac{P_{\text{peak}}-P_{\text{trough}}}{P_{\text{peak}}})$ (10) In our experiments and ablation studies, we recorded the metric outcomes as an average from five repeated trials. Ticker Model Cumulative Return (%) Sharpe Ratio Daily Volatility (%) Annualized Volatility (%) Max Drawdown (%) TSLA Buy and Hold -18.6312 -0.5410 4.4084 69.9818 55.3208 FinMem 61.7758* 2.6789 2.9522 46.8649 10.7996 Generative Agents 13.4636 0.5990 2.8774 45.6774 24.3177 FinGPT -7.4554 -0.2795 3.4145 54.2027 42.3993 A2C 13.7067 0.3979 4.4096 70.0009 52.3308 PPO 1.2877 0.0374 4.4110 70.0232 54.3264 DQN 33.3393 0.9694 4.4027 69.8900 52.0033 NFLX Buy and Hold 35.5111 1.4109 3.1964 50.7410 20.9263 FinMem 36.4485* 2.0168 2.2951 36.4342 15.8495 Generative Agents 32.0058 1.5965 2.5460 40.4168 16.9893 FinGPT 9.0090 0.4266 2.6819 42.5732 28.2705 A2C 14.6155 0.5788 3.2071 50.9112 25.0184 PPO 8.4121 0.3330 3.2086 50.9344 25.0184 DQN -12.2067 -0.4833 3.2078 50.9217 28.7017 AMZN Buy and Hold -10.7739 -0.4980 2.7697 43.9674 33.6828 FinMem 4.8850* 0.2327 2.6872 42.6576 22.9294 Generative Agents -13.9271 -0.9981 1.7864 28.3576 27.7334 FinGPT -29.6781 -2.1756 1.7464 27.7225 28.4838 A2C -6.3591 -0.2938 2.7706 43.9819 26.1275 PPO -8.4194 -0.3891 2.7702 43.9761 33.6828 DQN -29.9820 -1.3906 2.7603 43.8177 38.3740 MSFT Buy and Hold 14.6949 0.8359 2.2326 35.4411 15.0097 FinMem 23.2613* 1.4402 2.0512 32.5617 14.9889 Generative Agents -18.1031 -1.6057 1.4318 22.7285 24.2074 FinGPT 5.7356 0.4430 1.6442 26.1008 12.8459 A2C 0.4598 0.0261 2.2357 35.4913 23.6781 PPO 12.8067 0.7282 2.2333 35.4532 19.5355 DQN 14.7397 0.8385 2.2326 35.4408 25.1845 COIN Buy and Hold -30.0071 -0.5150 6.7517 107.1795 60.5084 FinMem 34.9832* 0.7170 5.6538 89.7515 35.7526 Generative Agents 3.4627 0.0896 4.4783 71.0908 32.0957 FinGPT -88.7805 -1.9507 5.2736 83.7153 73.5774 A2C - - - - - PPO - - - - - DQN - - - - - Table 2: Overall trading performance comparison during testing period between FinMem and other algorithmic agents across five stocks.* indicates that the result of the Wilcoxon signed-rank test is statistically significant.333The bold numbers in this and subsequent tables signify the best performance for the respective metrics. Figure 7: Cumulative Return of FinMem on trading Tesla, Inc. (TSLA) stock Over an Extended Testing Period. ## 5 Experiments: ### 5.1 Implementation Details: In the Trading Agents Comparison, FinMem employs GPT-4-Turbo as its backbone algorithm. The temperature parameter of the model was set at 0.7 to maintain a balance between response content consistency and model creativity. It was trained on financial data from August 17, 2021, to October 05, 2022, and underwent testing with data from October 06, 2022, to April 10, 2023. The training period was chosen to account for the seasonal nature of corporate financial reporting and the duration of data retention in FinMem’s memory module. The selected training duration ensures the inclusion of at least one publication cycle of either Form 10-Q, classified as intermediate memory, or Form 10-K, regarded as deep memory, or in some instances, both. This strategy ensures that the experiences retained in FinMem are still influential during the testing phase for a significant period. Additionally, the training duration allowed FinMem sufficient time to establish inferential links between financial news, market indicators, and stock market trends, thereby accumulating substantial experience. Furthermore, we set the number of top memory events retrieved from each layer of long-term memory at 5. We ran FinMem using each of the three available risk inclination settings. The reported performance outcomes are based on the setting that achieved the highest cumulative return during the testing phase. To maintain consistency in the comparison, the training and testing phases for the other two LLM-based agents were aligned with those of FinMem. For parameters of other LLM-based agents that are not encompassed by FinMem’s configuration, they were kept in accordance with their original settings as specified in their respective source codes. Considering that DRL algorithms need extensive training data for stable and converged results, and given our daily evaluation of trading performance, we extended the DRL agents’ training period to roughly a 10-year span, from January 1, 2012, to October 05, 2022, for a fair comparison. The testing period was kept consistent with the other models. The DRL algorithms were implemented using Stable Baselines 3 [41]. FinMem’s performance was benchmarked against that of the most effective comparative model, using Cumulative Return and Sharpe Ratio as the primary evaluation metrics. The statistical significance of FinMem’s superior performance was ascertained through the non-parametric Wilcoxon signed-rank test, which is particularly apt for the non-Gaussian distributed data. ### 5.2 Algorithmic Trading Agents Comparison (RQ1 & RQ2) In this experiment, we assess the stock trading performance of FinMem against other models, focusing on stocks from five companies in different trading sectors: Tesla, Inc. (TSLA), Netflix, Inc. (NFLX), Amazon.com, Inc. (AMZN), Microsoft Corporation (MSFT), and Coinbase Global, Inc. (COIN). The performance of all algorithmic trading agents across five key metrics is consolidated in Table 3. Given the pivotal role of Cumulative Return in evaluating trading performance over time, we present detailed time series plots in Figure 5 and Figure 6. It’s important to note that the trading performance of FinMem for COIN was exclusively compared with LLM trading agents and the baseline. This is because Coinbase Global, Inc. completed its IPO in April 2021 and, as a result, had not accumulated enough trading data to facilitate stable outcomes with Deep Reinforcement Learning (DRL) algorithms. These plots illustrate the changes in Cumulative Return for each of the five companies throughout the testing phase, offering an in-depth comparison of performance. In response to RQ1, the trading outcomes presented in Table 3 reveal that FinMem outperforms all other algorithmic trading agents and the B&H baseline strategy in terms of Cumulative Return and Sharpe Ratio. FinMem’s superiority is statistically significant when compared to the second-best trading strategy. Specifically, for TSLA and NFLX, FinMem’s strategy achieves Sharpe Ratios exceeding $2.0$ and Cumulative Returns surpassing $0.35$ while maintaining the lowest Volatility and Max Drawdown. These indicators underscore FinMem’s ability to generate higher returns per unit of risk. In the case of MSFT and NFLX, FinMem also records a Sharpe Ratio above $1.0$ and a Cumulative Return over $0.2$, coupled with relatively low Volatility and Max Drawdown, demonstrating its impressive trading performance. For AMZN and COIN, FinMem consistently delivers positive Cumulative Returns and superior Sharpe Ratios, outperforming other strategies that yield negative values for these metrics. Additionally, its Volatility and Max Drawdown are on the lower end. Hence, these results collectively demonstrate FinMem’s robust trading performance across a diverse range of trading sectors. Specifically, FinMem exhibits superior performance compared to the two other LLM agents in our study, FinGPT and the general-purpose generative agent developed by Park et al. This underscores the effectiveness of FinMem’s unique profiling and memory structure, which are particularly tailored for LLM agents dealing with financial data, significantly enhancing their investment decision-making capabilities. Figure 8: The optimal risk inclination for FinMem when trading different stocks. In response to RQ2, the main challenge for DRL trading agents is that they require training data with a large volume and extended time span, which are hard to achieve when operating on stocks with limited historical data. As shown in Table 3, our experiments reveal that FinMem achieves superior trading performance with a much shorter training duration compared to DRL trading agents trained on data spanning nearly a decade. This efficiency makes FinMem particularly useful for newly public companies like Coinbase Global, Inc., which have limited trading histories. DRL agents often face convergence issues due to inadequate training data in such cases. Moreover, even among LLM-based trading agents suited for shorter training periods, FinMem’s performance stands out, as illustrated in Figure 6. To further assess FinMem’s adaptability to limited training data, we narrowed the training period down to an even shorter period, spanning from August 17, 2021, to February 10, 2022. We then extended the testing phase to cover from February 11, 2022, to April 25, 2023. This evaluation focused on the stock of Tesla, Inc., which has the largest volume of news data. The trading performance of FinMem during this period is depicted in Figure 7. Remarkably, using less than six months of daily frequency data for training, which encompassed the publication of one Form 10-K and one Form 10-Q, FinMem consistently ranked high in gains and attained the highest cumulative return after the latter half of December 2022. The consistently strong trading performance of FinMem can be attributed to its innovative profiling and memory module design. This design enables FinMem to effectively integrate, comprehend, and prioritize key information from both textual and numerical data. The flexibility of FinMem’s profiling module in selecting risk inclinations plays a pivotal role in its ability to both exploit rising market trends and safeguard assets during downturns. A prime example is TSLA, which achieved its best trading results under a self-adaptive risk setting in FinMem. This configuration enables FinMem to pursue a conservative and cautious strategy when facing negative short-term cumulative returns. On the other hand, with positive short-term returns, FinMem switches to an optimistic and assertive approach, thus avoiding excessive passivity. This self-adaptive risk inclination proved effective for most stocks, apart from MSFT as shown in Table 8. For MSFT, a risk-seeking inclination was most beneficial, resonating with its general bullish in the stock market. Additionally, the memory module’s core features, including varied retention times for different information types and critical memory transitions, equip FinMem to capture essential information for well-informed investment decisions. ## 6 Ablation Studies We conducted three distinct ablation studies to evaluate key component alternatives in FinMem. These studies concentrated on the backbone algorithm, the memory module’s cognitive capacity, and the character setting in the profiling module, particularly examining the aspect of risk inclination. These studies were done using the stock of Tesla, Inc., with a more compact training period from March 14, 2022, to June 15, 2022, and a testing period from June 16, 2022, to December 28, 2022. This shorter duration was chosen for budgetary efficiency, yet it remains sufficient to differentiate the functionality of each component. ### 6.1 FinMem Backbone Algorithm Comparison (RQ3) Metric B&H GPT 3.5-Turbo GPT4 GPT4-Turbo davinci-003 Llama2-70b-chat Cumulative Return (%) -66.9497 16.1501 62.6180 54.6958 1.6308 -52.7233 Sharpe Ratio -2.0845 2.1589 2.2251 2.4960 0.8515 -2.8532 Daily Volatility (%) 3.8050 0.8862 3.3339 2.5960 0.2269 2.1891 Annualized Volatility (%) 60.4020 14.0683 52.9237 41.2100 3.6018 34.7503 Max Drawdown (%) 67.3269 1.1073 17.4012 12.5734 0.8408 44.7168 Table 3: Comparison of trading performance during the testing period for FinMem using different LLMs as backbone algorithms. In our first study, we evaluated the trading performance of FinMem using various LLMs as its backbone algorithms. The LLMs under consideration included davinci-003, GPT 3.5-Turbo, GPT4, GPT4-Turbo, and Llama2-70b-chat. The parameter settings were consistent with its optimal performance in the comparative experiment detailed in Section 5, and the risk inclination was configured to be self-adaptive. All other model settings were maintained as outlined in Section 5.1. The results of this evaluation are compiled in Table 3. In response to RQ3, Addressing Research Question 3, the findings demonstrate that FinMem, powered by GPT-4 and GPT-4 Turbo, delivered superior trading results during the test phase. Specifically, GPT-4 recorded the highest cumulative return, while GPT-4-Turbo exhibited the most favorable Sharpe Ratio. GPT 3.5-Turbo’s performance was also noteworthy, following closely behind. As depicted in Figure 9, though slightly lower than market baseline (B&H), FinMem with GPT-4-Turbo led in cumulative returns before October 2022. This period was characterized by relative stability and a modest upward trend in TSLA stock. After October 2022, with TSLA undergoing increased volatility and a notable downward trend, the cumulative return trajectory for FinMem with GPT-4-Turbo exhibited significantly lower volatility and sustained stable returns not markedly lower than those of GPT-4. These results indicate that GPT-4 Turbo is the most suitable backbone algorithm for FinMem. FinMem configured with davinci-003 and Llama2-70b-chat exhibited the lowest Annualized Volatility and Max Drawdown, yet their Cumulative Return and Sharpe Ratio were underwhelming. As illustrated in Figure 9, both models defaulted to a “Hold” strategy beyond a certain point during periods of intense fluctuation in TSLA stock. The unsatisfactory performance of davinci-003 may be attributed to its limited capability, as an earlier generation language model, to capture and understand nuanced yet decisive information. We selected Llama2-70b-chat as it was deemed to possess stronger in-context learning and instruction-following capabilities compared to other Llama family models with fewer parameters, as noted in Zhao et al. [58]. Nonetheless, in the context of stock trading, it still demonstrated challenges in adequately comprehending key messages necessary for effective trading decisions. The comparatively poorer performance of Llama2-70b-chat can also be attributed to its shorter context window, especially when compared to the GPT models. When integrated with FinMem, it needs to simplify prompts and shorten the length of retrieved memory insights, which could potentially result in some loss of context. The exceptional trading result demonstrated by GPT-4-Turbo across all models was a main factor in choosing it as the backbone algorithm for FinMem in our earlier comparative analysis with other algorithmic trading agents. Figure 9: Comparison of overall Cumulative Returns over time for FinMem using different LLMs as backbone algorithms. ### 6.2 Influence about varying the FinMem character design (RQ4) In our second study, we focused on evaluating the influence of FinMem’s profiling module on its trading effectiveness. Specifically, our assessment centered on the effects of customizing trader profiles according to specific stock trading, with a particular focus on risk inclination. As depicted in Figure 8, we equipped FinMem with three distinct risk profiles: risk-seeking, risk-averse, and a self-adaptive character. We executed a comparative analysis of FinMem’s performance across these risk profiles, maintaining consistency in all other settings as outlined in Section 5.1. In response to RQ4, Table 4 delineates the varied trading performance across different risk profiles. The self-adaptive profile enabled FinMem to achieve the most favorable trading performance, as it was the only one to secure a positive Cumulative Return and a Sharpe Ratio exceeding 2.0, along with the least Max Drawdown by the end of the testing phase. Figure 10 illustrates FinMem’s capacity to adeptly navigate substantial stock price volatility and to strategically modulate its trading behavior when necessary. In contrast, the risk-seeking profile, while beneficial during a stable or bullish market as evidenced by MSFT’s performance in Figure 5, exhibited increased volatility and a decline in the face of a market downturn. The risk-averse profile, on the other hand, maintained a more conservative stance, often opting to hold positions. This approach resulted in a Cumulative Return trajectory that generally lagged behind the market baseline, reflecting a degree of overcaution that limited trading activity and potential gains, particularly in a bullish market. Metric | B&H | Self Adaptive | Risk Seeking | Risk Averse ---|---|---|---|--- Cumulative Return (%) | -66.9497 | 54.6958 | -19.4132 | -12.4679 Sharpe Ratio | -2.0845 | 2.4960 | -0.7866 | -1.5783 Daily Volatility (%) | 3.9527 | 2.7419 | 3.2722 | 1.7744 Annualized Volatility (%) | 3.8050 | 2.5960 | 2.9236 | 0.9358 Max-Drawdown (%) | 67.3269 | 12.5734 | 45.0001 | 15.9882 Table 4: Comparison of overall trading performance during the testing period with different risk inclinations setting in FinMem’s profiling module. Figure 10: Comparison of Cumulative Return during with different risk inclinations setting in FinMem’s profiling module. ### 6.3 Impact of adjusting the capacity of FinMem working memory (RQ5) In our third study, we explored whether appropriately tuning the memory retrieval bandwidth of FinMem could enhance its trading performance. This bandwidth is tied to the working memory’s capacity within its memory module. As depicted in Figure 2, FinMem retrieves the top-$K$ memory events from its long-term memory in response to a trading inquiry. The working memory capacity is thus set at $3\times K$, mirroring the human cognitive system’s limit of processing immediate messages upon specific stimuli ($3$ refers to the three processing layers in long-term memory). By varying the $K$ hyperparameter, FinMem can expand this capacity far beyond the human cognitive scope. We aimed to determine whether such flexibility in adjusting memory bandwidth translates to improvements in FinMem’s performance. Metric | B&H | Top 1 | Top 3 | Top 5 | Top 10 ---|---|---|---|---|--- Cumulative Return (%) | -66.9497 | 52.0936 | 29.4430 | 54.6958 | 79.4448 Sharpe Ratio | -2.0845 | 1.8642 | 1.1214 | 2.4960 | 2.7469 Daily Volatility (%) | 3.8050 | 3.3105 | 3.1105 | 2.5960 | 3.4262 Annualized Volatility (%) | 60.4020 | 52.5529 | 49.3779 | 41.2100 | 54.3891 Max Drawdown (%) | 67.3269 | 25.2355 | 27.0972 | 12.5734 | 17.1360 Table 5: Comparison of overall trading performance during the testing period with different configurations of working memory capacity. As demonstrated in Table 5, we adjusted the hyperparameter $K$ to alter the number of memory events retrieved from shallow, intermediate, and deep long- term memory layers in FinMem. We tested $K$ values of 1, 3, 5, and 10, exploring FinMem’s working memory capabilities at levels below, near, and above the human cognitive limit. For all these $K$ settings, we maintained a self-adaptive risk inclination, while other settings were consistent with those described in Section 5.1. Across all $K$ configurations, FinMem outperformed the Buy & Hold baseline, indicating the effectiveness of its memory module in processing diverse information and capturing critical events, which subsequently enhanced its trading performance, as evidenced by positive Cumulative Returns and Sharpe Ratios. Notably, higher $K$ values, like 5 and 10, enabled FinMem to achieve the best Cumulative Returns and Sharpe Ratios exceeding 2.0. With $K$ set to 1, FinMem still performed moderately well by capturing the most critical memory events of each layer. An in-depth analysis in Figure 11, which shows the Cumulative Return over time for various $K$ settings, reveals that a $K$ value of 5 is optimal for trading TSLA stock, consistently delivering robust performance with the lowest Volatility and Max-Drawdown. Before mid-October 2022, when the stock market was relatively stable and slightly upward, FinMem’s trading actions aligned well with market trends (referring to B&H) and avoided significant losses. During periods of high volatility and continuous downturns (post-mid-October 2022), it maintained earnings by reducing “Buy” actions and favoring more “Hold” and “Sell” strategies. However, setting $K$ to 10, while effective during market volatility, resulted in significant losses in stable market conditions. The issue may stem from the disproportionately loose capacity constraints on incoming information relative to the volume of incoming data. The broad memory retrieval bandwidth might have mixed trivial messages with critical ones, hampering FinMem’s decision precision. This challenge becomes especially evident in neutral market conditions, where the influx of information includes a mix of varying market sentiments and trends. Addressing RQ5, appropriately tuning the number of memory events (Top-$K$ in the FinMem memory module can significantly enhance its trading performance. The aforementioned study illustrates that FinMem can achieve optimal results by effectively assimilating key signals from a sufficient quantity of filtered memories across each layer. However, the optimal value for $K$ may vary depending on the volume and quality of incoming information. Figure 11: Cumulative Return over time for with different FinMem working memory capacity. ## 7 Conclusion and future work In this paper, we introduce FinMem, an innovative automated trading agent framework featuring an adjustable cognitive memory structure and dynamic character design. Our research demonstrates its capacity to enhance stock trading performance substantially using real-world financial datasets. Additionally, the efficacy of each critical component within FinMem is thoroughly demonstrated in our ablation studies, highlighting their roles in optimizing trading outcomes. Its unique features of human-like cognitive memory modules and dynamic character design enable it to tackle the complexities of financial environments and respond aptly to new situations. Compared to other LLM trading agents, FinMem’s memory module equips it with the capability to better comprehend financial data featured by various timeliness and organize them as a self-evolving long-term memory layer. The dynamic character design endows FinMem with critical professional insights, enabling efficient filtering of impactful messages from incoming financial data for trading actions. Additionally, the integration of multiple risk profiles enhances FinMem’s adaptability to a range of market conditions. FinMem’s exceptional performance underscores its remarkable ability to transform diverse financial data into well-informed investment strategies. Its proficiency in integrating various financial data types is further accentuated by a notably reduced training duration, which offers advantages for trading with newly established companies. In our approach, we utilized a limited range and quality of financial news and reports, employing general-purpose LLMs as the backbone algorithms. However, FinMem is fully compatible with LLMs specifically fine-tuned for financial applications. We anticipate that its trading efficacy will be elevated further with access to a more comprehensive and higher-quality dataset, coupled with LLMs tailored specifically for financial contexts. While primarily designed for financial decision-making, the FinMem framework boasts a versatility that extends to domains such as IT consulting and business reporting, where actions are driven by time-sensitive information. Looking ahead, an intriguing direction for development is the creation of a multi-agent trading system, rooted in the FinMem platform, aimed at enhancing investment portfolio optimization. This system, featuring diverse professional backgrounds in its profiling modules, enables concurrent operations and trading across a variety of financial products. It dynamically adjust trading targets based on a sequential assessment of key trading performance indicators. Simultaneously, it can timely reallocate investment proportions for each financial product category, leveraging peer-to-peer communication and systematic performance analysis. Supported by LLMs, this multi-agent trading system is poised to assemble an optimized investment portfolio through its strategically planned operations. ## References * [1] Andrew Ang and Joseph Chen “Downside Risk” In _Journal of Portfolio Management_ 29.4, 2003, pp. 103–112 * [2] Lisa P Argyle et al. “Out of one, many: Using language models to simulate human samples” In _Political Analysis_ 31.3 Cambridge University Press, 2023, pp. 337–351 * [3] Loukia Avramelou, Paraskevi Nousi, Nikolaos Passalis and Anastasios Tefas “Deep reinforcement learning for financial trading using multi-modal features” In _Expert Systems with Applications_ Elsevier, 2023, pp. 121849 * [4] Surjeet Balhara et al. “A survey on deep reinforcement learning architectures, applications and emerging trends” In _IET Communications_ Wiley Online Library, 2022 * [5] Fischer Black “Noise” In _The journal of finance_ 41.3 Wiley Online Library, 1986, pp. 528–543 * [6] Yu-Fu Chen and Szu-Hao Huang “Sentiment-influenced trading system based on multimodal deep reinforcement learning” In _Applied Soft Computing_ 112 Elsevier, 2021, pp. 107788 * [7] Tai-Liang Chen “Forecasting the Taiwan stock market with a novel momentum-based fuzzy time-series” In _Review of Economics & Finance_ 2 Better Advances Press, Canada, 2012, pp. 38–50 * [8] John H. Cochrane “Volatility Tests and Efficient Markets: A Review Essay” In _Journal of Monetary Economics_ 22.3, 1988, pp. 463–485 * [9] Fergus IM Craik and Robert S Lockhart “Levels of processing: A framework for memory research” In _Journal of verbal learning and verbal behavior_ 11.6 Elsevier, 1972, pp. 671–684 * [10] Quang-Vinh Dang “Reinforcement learning in stock trading” In _International conference on computer science, applied mathematics and applications_ , 2019, pp. 311–322 Springer * [11] Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova “Bert: Pre-training of deep bidirectional transformers for language understanding” In _arXiv preprint arXiv:1810.04805_ , 2018 * [12] Robert D Edwards, John Magee and WH Charles Bassetti “Technical analysis of stock trends” CRC press, 2018 * [13] Kawin Ethayarajh “How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings” In _arXiv preprint arXiv:1909.00512_ , 2019 * [14] Thomas G Fischer “Reinforcement learning in financial markets-a survey”, 2018 * [15] Samuel J Gershman and Bence P Ölveczky “The neurobiology of deep reinforcement learning” In _Current Biology_ 30.11 Elsevier, 2020, pp. R629–R632 * [16] Yoav Goldberg and Omer Levy “word2vec Explained: deriving Mikolov et al.’s negative-sampling word-embedding method” In _arXiv preprint arXiv:1402.3722_ , 2014 * [17] “Guardrails AI” Open source library for interacting with Large Language Models, https://docs.guardrailsai.com * [18] Sepp Hochreiter and Jürgen Schmidhuber “Long short-term memory” In _Neural computation_ 9.8 MIT press, 1997, pp. 1735–1780 * [19] Boming Huang et al. “Automated trading systems statistical and machine learning methods and hardware implementation: a survey” In _Enterprise Information Systems_ 13.1 Taylor & Francis, 2019, pp. 132–144 * [20] Ziheng Huang, Sebastian Gutierrez, Hemanth Kamana and Stephen MacNeil “Memory sandbox: Transparent and interactive memory management for conversational agents” In _arXiv preprint arXiv:2308.01542_ , 2023 * [21] John Hull “Risk Management and Financial Institutions” John Wiley & Sons, 2007 * [22] O Jangmin, Jongwoo Lee, Jae Won Lee and Byoung-Tak Zhang “Adaptive stock trading with dynamic asset allocation using reinforcement learning” In _Information Sciences_ 176.15 Elsevier, 2006, pp. 2121–2147 * [23] Jeff Johnson, Matthijs Douze and Hervé Jégou “Billion-scale similarity search with GPUs” In _IEEE Transactions on Big Data_ 7.3 IEEE, 2019, pp. 535–547 * [24] Yifei Li et al. “Making language models better reasoners with step-aware verifier” In _Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , 2023, pp. 5315–5333 * [25] Mark Liffiton, Brad Sheese, Jaromir Savelka and Paul Denny “Codehelp: Using large language models with guardrails for scalable support in programming classes” In _arXiv preprint arXiv:2308.06921_ , 2023 * [26] Xiao-Yang Liu et al. “FinRL-Meta: Market environments and benchmarks for data-driven financial reinforcement learning” In _Advances in Neural Information Processing Systems_ 35, 2022, pp. 1835–1849 * [27] Xiao-Yang Liu et al. “FinRL: A deep reinforcement learning library for automated stock trading in quantitative finance” In _arXiv preprint arXiv:2011.09607_ , 2020 * [28] Xiao-Yang Liu, Hongyang Yang, Jiechao Gao and Christina Dan Wang “FinRL: Deep reinforcement learning framework to automate trading in quantitative finance” In _ACM International Conference on AI in Finance (ICAIF)_ , 2021 * [29] Adrian Millea “Deep reinforcement learning for trading—A critical survey” In _Data_ 6.11 MDPI, 2021, pp. 119 * [30] George A Miller “The magical number seven, plus or minus two: Some limits on our capacity for processing information.” In _Psychological review_ 63.2 American Psychological Association, 1956, pp. 81 * [31] Volodymyr Mnih et al. “Asynchronous methods for deep reinforcement learning” In _International conference on machine learning_ , 2016, pp. 1928–1937 PMLR * [32] Volodymyr Mnih et al. “Playing atari with deep reinforcement learning” In _arXiv preprint arXiv:1312.5602_ , 2013 * [33] Jaap MJ Murre and Joeri Dros “Replication and analysis of Ebbinghaus’ forgetting curve” In _PloS one_ 10.7 Public Library of Science San Francisco, CA USA, 2015, pp. e0120644 * [34] OpenAI “GPT-4 Technical Report”, 2023 arXiv:2303.08774 [cs.CL] * [35] Joon Sung Park et al. “Generative Agents: Interactive Simulacra of Human Behavior” In _Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology_ , UIST ’23 San Francisco, CA, USA: Association for Computing Machinery, 2023 DOI: 10.1145/3586183.3606763 * [36] Joon Sung Park et al. “Social simulacra: Creating populated prototypes for social computing systems” In _Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology_ , 2022, pp. 1–18 * [37] Eero Pätäri and Mika Vilska “Performance of moving average trading strategies over varying stock market conditions: the Finnish evidence” In _Applied Economics_ 46.24 Taylor & Francis, 2014, pp. 2851–2872 * [38] Jeffrey Pennington, Richard Socher and Christopher D Manning “Glove: Global vectors for word representation” In _Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP)_ , 2014, pp. 1532–1543 * [39] Tidor-Vlad Pricope “Deep reinforcement learning in quantitative algorithmic trading: A review” In _arXiv preprint arXiv:2106.00123_ , 2021 * [40] Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever “Improving language understanding by generative pre-training” OpenAI, 2018 * [41] Antonin Raffin et al. “Stable-Baselines3: Reliable Reinforcement Learning Implementations” In _Journal of Machine Learning Research_ 22.268, 2021, pp. 1–8 URL: http://jmlr.org/papers/v22/20-1364.html * [42] John Schulman et al. “Proximal policy optimization algorithms” In _arXiv preprint arXiv:1707.06347_ , 2017 * [43] William F. Sharpe “The Sharpe Ratio” In _The Journal of Portfolio Management_ 21.1, 1994, pp. 49–58 * [44] Yong Shi et al. “Stock trading rule discovery with double deep Q-network” In _Applied Soft Computing_ 107 Elsevier, 2021, pp. 107320 * [45] Theodore Sumers, Shunyu Yao, Karthik Narasimhan and Thomas L Griffiths “Cognitive architectures for language agents” In _arXiv preprint arXiv:2309.02427_ , 2023 * [46] Ron Sun “Desiderata for cognitive architectures” In _Philosophical psychology_ 17.3 Taylor & Francis, 2004, pp. 341–373 * [47] John Sweller “Human cognitive architecture: Why some instructional procedures work and others do not.” American Psychological Association, 2012 * [48] Hugo Touvron et al. “Llama: Open and efficient foundation language models” In _arXiv preprint arXiv:2302.13971_ , 2023 * [49] Rashesh Vaidya “Moving Average Convergence-Divergence (MACD) Trading Rule: An Application in Nepalese Stock Market" NEPSE"” In _Quantitative Economics and Management Studies_ 1.6, 2020, pp. 366–374 * [50] Lei Wang et al. “A survey on large language model-based autonomous agents” In _arXiv preprint arXiv:2308.11432_ , 2023 * [51] Lei Wang et al. “RecAgent: A Novel Simulation Paradigm for Recommender Systems” In _arXiv preprint arXiv:2306.02552_ , 2023 * [52] Shijie Wu et al. “Bloomberggpt: A large language model for finance” In _arXiv preprint arXiv:2303.17564_ , 2023 * [53] Xing Wu et al. “Adaptive stock trading strategies with deep reinforcement learning methods” In _Information Sciences_ 538 Elsevier, 2020, pp. 142–158 * [54] Zhuoran Xiong et al. “Practical deep reinforcement learning approach for stock trading” In _arXiv preprint arXiv:1811.07522_ , 2018, pp. 1–7 * [55] Hongyang Yang, Xiao-Yang Liu and Christina Dan Wang “FinGPT: Open-Source Financial Large Language Models” In _arXiv preprint arXiv:2306.06031_ , 2023 * [56] Hongyang Yang, Xiao-Yang Liu, Shan Zhong and Anwar Walid “Deep reinforcement learning for automated stock trading: An ensemble strategy” In _Proceedings of the first ACM international conference on AI in finance_ , 2020, pp. 1–8 * [57] Hongxin Zhang et al. “Building cooperative embodied agents modularly with large language models” In _arXiv preprint arXiv:2307.02485_ , 2023 * [58] Wayne Xin Zhao et al. “A survey of large language models” In _arXiv preprint arXiv:2303.18223_ , 2023
# Selection of Time Headway in Connected and Autonomous Vehicle Platoons under Noisy V2V Communication Guoqi Ma, Prabhakar R. Pagilla∗, Swaroop Darbha The authors are with the Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA (e-mail<EMAIL_ADDRESS><EMAIL_ADDRESS>[email protected]). ∗Corresponding author. Part of this work was presented at the 26th IEEE International Conference on Intelligent Transportation Systems (ITSC), Bilbao, Spain, 2023 [1]. ###### Abstract In this paper, we investigate the selection of time headway to ensure robust string stability in connected and autonomous vehicle platoons in the presence of signal noise in Vehicle-to-Vehicle (V2V) communication. In particular, we consider the effect of noise in communicated vehicle acceleration from the predecessor vehicle to the follower vehicle on the selection of the time headway in predecessor-follower type vehicle platooning with a Constant Time Headway Policy (CTHP). Employing a CTHP based control law for each vehicle that utilizes on-board sensors for measurement of position and velocity of the predecessor vehicle and wireless communication network for obtaining the acceleration of the predecessor vehicle, we investigate how time headway is affected by communicated signal noise. We derive constraints on the CTHP controller gains for predecessor acceleration, velocity error and spacing error and a lower bound on the time headway which will ensure robust string stability of the platoon against signal noise. We provide comparative numerical simulations on an example to illustrate the main result. ###### Index Terms: Connected and Autonomous Vehicle Platoons, Cooperative Adaptive Cruise Control (CACC), Constant Time Headway Policy (CTHP), V2V Communication, Signal-to- Noise Ratio (SNR), Robust String Stability. ## I Introduction The deployment of connected and autonomous vehicle platoons has the potential to benefit transportation systems in a profound and comprehensive way [2, 3]. Initially, under the Adaptive Cruise Control (ACC) paradigm facilitated by the use of onboard sensors (such as radars) for measuring the velocities of and distances from adjacent vehicles, the vehicle platoon can maintain a constant inter-vehicular spacing, referred to as Constant Spacing Policy (CSP) [4]. Recently, the benefits of employing advanced vehicular communication technologies and modern communication protocols have been expounded in great detail in the literature; for example, Dedicated Short Range Communication (DSRC) [5], Long Term Evolution (LTE) [6], 5G [7], and V2V communication [8]. In addition to the use of onboard sensors, advanced vehicular wireless communications can lead to a higher-level of connectivity by incorporating the acceleration information of other vehicles which has the potential to significantly improve safety, increase mobility and throughput, and reduce fuel consumption under the paradigm of Cooperative Adaptive Cruise Control (CACC) with Constant Time Headway Policy (CTHP) [9]. In addition to internal stability for each vehicle, the connected and autonomous vehicle platoon should also exhibit robust string stability, i.e., robustness of the platoon to uncertain lag, noise, and external disturbances. In recent decades, there has been extensive research devoted to connected and autonomous vehicle platoons on a wide range of topics including controller design [10, 11, 12, 13, 14, 15], communication mechanisms [16, 17, 18], experimental validation in realistic environments [19, 20], mixed human and autonomous traffic [21, 22, 23], string stability analysis [24, 25]. However, most existing results assume ideal V2V communication which is often not practical. Among many other factors, signal noise in communication channels is a key concern, which can be caused by a variety of factors including limited communication bandwidth [26], cyber-attacks [27], etc. If key control parameters, such as time headway, are chosen without consideration of signal noise in communicated signals, then there is a possibility of the onset of string instability and collisions. Although communication quality has improved significantly under 5G and more advanced wireless communication, signal noise when propagating through the platoon could have a substantial affect on safety and performance. Previous research on imperfect communication focused on how packet drops affected string stability [28]. In [29, 30], the disturbance estimation and control problem was investigated, though without considering the quantitative effect of disturbance on the design parameters. In contrast to prior research with imperfect communication, this paper considers the following problem: Given a Signal-to-Noise Ratio (SNR) with V2V communication, what is the lower bound of the employable time headway for CACC based vehicle platooning? For this purpose, we consider a CTHP based control law where the feedforward acceleration of the predecessor vehicle is communicated and contains signal noise. Since the communication is through an $n$-bit channel, we model the noise in the acceleration signal as a sum of $n$ binary random variables. To address the issue of the vehicle closed-loop governing equation being stochastic due to the noise signal, we develop an equivalent deterministic governing equation using the averaging procedure described in [28]. Based on this, we obtain the spacing error propagation equation and show that the platoon is robustly string stable in the presence of signal noise and parasitic lag and derive conditions on control gains and time headway under which the platoon is robustly string stable. We provide numerical simulation results on a platoon to verify the main result. The main contributions of the work can be summarized as follows. In the context of communication from a single predecessor vehicle, this paper provides an extension to our previous work in [9] by considering signal noise in communicated acceleration from the predecessor vehicle. We derive a lower bound on the time headway that is a function of the parasitic lag in the vehicle dynamics ($\tau_{0}$), acceleration gain ($k_{a}$), and SNR ($\rho$). This lower bound aids in the selection of the time headway in the presence of noise. Further, the lower bound reduces to the ideal communication case given in [9] with no noise in the communicated acceleration signal. In addition, our formulation and analysis allow for systematic selection of the control gains as well as time headway for a given signal to noise ratio. The remainder of the paper is organized as follows. Section II contains preliminaries including vehicle dynamics and relevant definitions. The main theoretical results are provided in Section III. An illustrative numerical example and simulation results are provided in Section IV. Finally, some concluding remarks are given in Section V. ## II Preliminaries Consider a string of autonomous vehicles equipped with V2V communication as illustrated in Fig. 1. Figure 1: Autonomous and connected vehicle platoon with V2V communication. The $i$-th vehicle dynamics is given by the model: $\displaystyle\left\\{\begin{array}[]{*{20}{l}}\ddot{x}_{i}(t)=a_{i}(t),\\\ \tau\dot{a}_{i}(t)+a_{i}(t)=u_{i}(t),\end{array}\right.$ (3) where $x_{i}(t)$, $a_{i}(t)$, $u_{i}(t)$ represent the position, acceleration, and control input of the $i$-th vehicle at time instant $t$, and $i\in\mathcal{N}=\\{1,2,\cdots,N\\}$, where $N$ is the total number of the following vehicles in the platoon, $\tau$ denotes the parasitic actuation lag. It is assumed that $\tau$ is uncertain with $\tau\in\left(0,\tau_{0}\right]$, where $\tau_{0}$ is a positive real constant. ###### Definition II.1 Let $d$ denote the minimum or standstill spacing between adjacent vehicles, $v_{i}(t)$ denote the velocity of the $i$-th vehicle, $h_{w}$ denote the time headway, and $e_{i}(t):=x_{i}(t)-x_{i-1}(t)+d$ be the spacing error for the $i$-th vehicle with respect to the $(i-1)$-th vehicle. Define the velocity dependent inter-vehicular spacing error for the $i$-th vehicle as: $\displaystyle\delta_{i}(t)=e_{i}(t)+h_{w}v_{i}(t).$ (4) ###### Definition II.2 ([9]) Let $\delta_{i}(s)$ denote the Laplace transform of $\delta_{i}(t)$ and let $H(s)$ denote the spacing error propagation transfer function that satisfies $\delta_{i}(s)=H(s)\delta_{i-1}(s)$. The connected and autonomous vehicle platoon is said to be robustly string stable if the following two conditions hold for all $\tau\in(0,\tau_{0}]$: (i) $H(s)$ achieves internal stability and (ii) the platoon is string stable, i.e., it holds that $\|\delta_{i}(t)\|_{\infty}\leq\|\delta_{i-1}(t)\|_{\infty}$, or in the frequency domain $\displaystyle\|H(j\omega)\|_{\infty}\leq 1.$ (5) ###### Definition II.3 Let $s(t)$ and $n(t)$ denote the transmitted signal and the noise signal, respectively, and assume $n(t)=0$ when $s(t)=0$. Then, the Signal-to-Noise Ratio (SNR) for $s(t)\neq 0$ is defined as $\displaystyle{\rm SNR}=\min\limits_{t>0}20\log\frac{|s(t)|}{|n(t)|},$ (6) where $\rm log$ denotes the logarithmic function with base 10. From the ${\rm SNR}$ definition (6), we can obtain that if ${\rm SNR}\coloneqq\varrho$, then $\rho=\min\limits_{t>0}\frac{|s(t)|}{|n(t)|}=10^{\frac{\varrho}{20}}$, and for notational convenience, we will use $\rho=\min\limits_{t>0}\frac{|s(t)|}{|n(t)|}$ instead of ${\rm SNR}$. ## III Main Results ### III-A Model of Noise Let $a_{i-1}(t)$ denote the acceleration signal of the preceding vehicle ($(i-1)$-th) and $w_{i,i-1}(t)>0$ denote the noise factor associated with V2V communication of the acceleration signal of the $(i-1)$-th vehicle to the $i$-th vehicle. Considering the noise affected signal, the communicated acceleration signal available to the $i$-th vehicle is given by $w_{i,i-1}(t)a_{i-1}(t)$. We define $\rho>1$ as the signal to noise ratio factor such that $1-\frac{1}{\rho}\leq w_{i,i-1}(t)\leq 1+\frac{1}{\rho}$, that is, the communicated acceleration signal to the $i$-th vehicle satisfies $\left|\left(1-\frac{1}{\rho}\right)a_{i-1}(t)\right|\leq\left|w_{i,i-1}(t)a_{i-1}(t)\right|\leq\left|\left(1+\frac{1}{\rho}\right)a_{i-1}(t)\right|$. An illustration of the admissible region of the noise signal in terms of the ideal communicated acceleration signal is provided in Fig. 2, and the range of the communicated signal with respect to the actual signal is provided in Fig. 3. Figure 2: An illustration of the admissible region for the signal noise in the communicated acceleration signal ($\rho=5$). Figure 3: An illustration of the admissible region of the communicated acceleration signal with respect to the actual one ($\rho=5$). Considering the communicated acceleration signal is through an $n$-bit communication channel, we model $w_{i,i-1}(t)$ as follows: $\displaystyle w_{i,i-1}(t)=\left(1-\frac{1}{\rho}\right)+\frac{1}{\rho}\left(\sum_{j=0}^{n-1}\frac{z_{i,j}(t)}{2^{j}}\right),$ (7) where $z_{i,j}(t),j=0,\ldots,n-1,$ are independent binary random processes. Let $\gamma_{i,j}:=\mathbb{E}[z_{i,j}(t)]$. We assume that $\gamma_{i,j}$ is independent of time, i.e., the characteristics of the noise processes of the communication channel are time-invariant. However, in order to model wide variety of noise processes, we do not assume that $\gamma_{i,j}$’s are known a priori. Note that $\gamma_{i,j}\in(0,1)$ then this noise modeling procedure will allow us to get a communicated signal that takes any value lying between $\left(1-\frac{1}{\rho}\right)a_{i,i-1}$ and $\left(1+\frac{1}{\rho}\right)a_{i,i-1}$. ### III-B Inter-vehicular Spacing Error Propagation Equation Consider the following CTHP control law for the $i$-th vehicle: $\displaystyle u_{i}(t)=k_{a}w_{i,i-1}(t)a_{i-1}(t)-k_{v}(v_{i}(t)-v_{i-1}(t))-k_{p}\delta_{i}(t),$ (8) then the governing equation for the $i$-th vehicle is given by: $\displaystyle\tau\dddot{x}_{i}(t)+\ddot{x}_{i}(t)$ $\displaystyle=k_{a}w_{i,i-1}(t)a_{i-1}(t)$ $\displaystyle\quad- k_{v}(v_{i}(t)-v_{i-1}(t))-k_{p}\delta_{i}(t).$ (9) The governing equation is stochastic due to the presence of the stochastic noise signal $\omega_{i,i-1}(t)$; correspondingly, the state is also stochastic. To perform robust string stability analysis, we consider an equivalent deterministic governing equation along the lines described in [28]. By defining the augmented state vector as $\hat{X}(t)=[x_{0}(t),v_{0}(t),a_{0}(t),x_{1}(t),v_{1}(t),a_{1}(t),\cdots,x_{N}(t),v_{N}(t),a_{N}(t)]$ and letting $\hat{\omega}(t)$ be the vector of noise signals with its $i$-th component being $\omega_{i,i-1}(t)$, and the acceleration input to the lead vehicle to be $U(t)$, the augmented state equation for the vehicle platoon can be developed as $\displaystyle\dot{\hat{X}}(t)=\hat{A}(\hat{w}(t))\hat{X}(t)+BU(t),$ (10) where the structures of the matrices $\hat{A}$ and $B$ are given in [28]. Through a discretization process for a sufficiently small time interval, it was shown that $\displaystyle\mathbb{E}[\hat{A}]=\bar{A},\quad\mathbb{E}[e^{\hat{A}t}]=e^{\bar{A}t},$ (11) where $\bar{A}=\hat{A}(\bar{w})$ with $\bar{w}$ as the expectation of $\hat{w}(t)$. The second equality is true because of the specific lower triangular and banded structure of the governing equations for the CACC case. Let $\bar{X}=\mathbb{E}[\hat{X}]$; in the following, unless specified otherwise, we will use bar to indicate expected value for a corresponding random variable. The evolution of the vehicle states in the platoon converge to the states of the following deterministic state equation: $\displaystyle\dot{\bar{X}}(t)=\bar{A}\bar{X}(t)+BU(t).$ (12) Utilizing this procedure, the governing equation for the $i$-th vehicle is given by: $\displaystyle\tau\dddot{\bar{x}}_{i}(t)+\ddot{\bar{x}}_{i}(t)$ $\displaystyle=k_{a}\mathbb{E}[w_{i,i-1}(t)]\bar{a}_{i-1}(t)$ $\displaystyle\quad- k_{v}(\bar{v}_{i}(t)-\bar{v}_{i-1}(t))-k_{p}\bar{\delta}_{i}(t).$ (13) Let $\displaystyle\tilde{k}_{a}:=k_{a}\mathbb{E}[w_{i,i-1}(t)]=\left[1-\frac{1}{\rho}+\frac{1}{\rho}\left(\sum_{j=0}^{n-1}\frac{\gamma_{i,j}}{2^{j}}\right)\right]k_{a}.$ (14) The governing equation for the $i$-th vehicle can be re-written as: $\displaystyle\tau\dddot{\bar{x}}_{i}(t)+\ddot{\bar{x}}_{i}(t)$ $\displaystyle=\tilde{k}_{a}\ddot{\bar{x}}_{i-1}(t)$ $\displaystyle\quad- k_{v}(\dot{\bar{x}}_{i}(t)-\dot{\bar{x}}_{i-1}(t))-k_{p}\bar{\delta}_{i}(t).$ (15) Let $\bar{\delta}_{i}(t)=\bar{x}_{i}(t)-\bar{x}_{i-1}(t)+d+h_{w}\dot{\bar{x}}_{i}(t)$, and let $\bar{\delta}(s)$ be the Laplace transform of $\bar{\delta}_{i}(t)$. Then, the inter-vehicular spacing error propagation equation can be derived as $\displaystyle\bar{\delta}_{i}(s)=\tilde{H}(s;\tau)\bar{\delta}_{i-1}(s),$ (16) where $\tilde{H}(s;\tau)=\tilde{\mathcal{N}}(s)/\mathcal{D}(s)$ is the inter- vehicular spacing error propagation transfer function with $\displaystyle\tilde{\mathcal{N}}(s)$ $\displaystyle=\tilde{k}_{a}s^{2}+k_{v}s+k_{p},$ $\displaystyle\mathcal{D}(s)$ $\displaystyle=\tau s^{3}+s^{2}+\gamma s+k_{p},$ where $\gamma:=k_{v}+h_{w}k_{p}$. The platoon error propagation equation given by (III-B) was studied in Theorem 1 of [9] which can be recast for the case considered in this paper as follows: ###### Theorem III.1 The following are true for the platoon given by the inter-vehicular spacing error propagation equation (16): 1. (a) $\|\tilde{H}(j\omega;\tau)\|_{\infty}\leq 1$, $\forall\tau\in(0,\tau_{0}]$, implies that $\tilde{k}_{a}\in(0,1)$; 2. (b) given any $\tilde{k}_{a}\in(0,1)$, $h_{w}$ satisfying the following: $\displaystyle h_{w}\geq\frac{2\tau_{0}}{1+\tilde{k}_{a}},$ (17) there exist $k_{p},k_{v}>0$ such that $\|\tilde{H}(j\omega;\tau)\|_{\infty}\leq 1$ for all $\tau\in(0,\tau_{0}]$. ### III-C Stability Analysis ###### Theorem III.2 The following are true for the platoon given by inter-vehicular spacing error propagation equation (16): 1. (a) $\|\tilde{H}(j\omega;\tau)\|_{\infty}\leq 1$, $\forall\tau\in(0,\tau_{0}]$, implies that $\tilde{k}_{a}\in(0,1)$. 2. (b) given any $\displaystyle k_{a}\in\left(0,\frac{1}{1+\frac{1}{\rho}}\right),$ (18) and $h_{w}$ satisfying the following: $\displaystyle h_{w}>h_{w,lb}(k_{a}):=2\tau_{0}\left(\frac{1-\left(1-\frac{1}{\rho}\right)k_{a}}{1-\left(1+\frac{1}{\rho}\right)^{2}k_{a}^{2}}\right),$ (19) there exist $k_{p},k_{v}>0$ such that $\|\tilde{H}(j\omega;\tau)\|_{\infty}\leq 1$ for all $\tau\in(0,\tau_{0}]$; 3. (c) given any $\rho>1$, the minimizing value, $k_{a}^{*}$, of $k_{a}$ and the corresponding minimum value, $h_{w,lb}^{*}$, of $h_{w,lb}(k_{a})$ are given by: $\displaystyle k_{a}^{*}=\left(\frac{1-\frac{1}{\sqrt{\rho}}}{1+\frac{1}{\sqrt{\rho}}}\right)\frac{1}{\left(1+\frac{1}{\rho}\right)},$ (20) $\displaystyle h_{w,lb}^{*}=\tau_{0}\frac{\left(1+\frac{1}{\sqrt{\rho}}\right)^{2}}{\left(1+\frac{1}{\rho}\right)}.$ (21) Correspondingly, there exist $k_{p},k_{v}>0$ such that $\|\tilde{H}(j\omega;\tau)\|_{\infty}\leq 1$ for all $\tau\in(0,\tau_{0}]$. ###### Remark III.1 Minimizing $h_{w,lb}$ is useful for improving traffic throughput because a lower time headway can be employed while still guaranteeing “robust” string stability. ###### Proof: For internal stability, $\gamma-\tau k_{p}>0$ by Routh-Hurwitz criterion on $\mathcal{D}(s)$. Applying _Theorem III.1_ to the error propagation equation (16), we infer that: 1. (a) $\tilde{k}_{a}\in(0,1)$; 2. (b) $h_{w}\geq\dfrac{2\tau_{0}}{1+\tilde{k}_{a}}$. Since $\gamma_{i,j}$’s are not known, we can bound $\tilde{k}_{a}$ by considering their maximum possible values. Hence, we have $\displaystyle\tilde{k}_{a}\leq\left(1-\frac{1}{\rho}+\frac{2}{\rho}\right)k_{a}=\left(1+\frac{1}{\rho}\right)k_{a}.$ (22) Thus, if $\displaystyle k_{a}<\frac{1}{1+\frac{1}{\rho}}\implies\tilde{k}_{a}<1,$ (23) allowing us to apply part (b) of Theorem _Theorem III.1_. We then have $\displaystyle h_{w}\geq\frac{2\tau_{0}}{1+\tilde{k}_{a}}.$ (24) Again, since $\tilde{k}_{a}$ is dependent on $\gamma_{i,j}$’s which are not known a priori, we will consider the worst possible lower bound, i.e., the lower bound for $h_{w}$ that is a maximum over all possible values of $\gamma_{i,j}$’s. To address this, we require $\displaystyle h_{w}$ $\displaystyle\geq\max\limits_{\gamma_{i,0},\ldots,\gamma_{i,n}}\frac{2\tau_{0}}{1+\tilde{k}_{a}}.$ (25) Hence, if the above condition holds, then (24) will hold. Substituting $\tilde{k}_{a}$ from (14) into (25), the condition (24) will hold if $\displaystyle h_{w}$ $\displaystyle\geq\max\limits_{\gamma_{i,0},\ldots,\gamma_{i,n}}\frac{2\tau_{0}}{1+\left(1-\frac{1}{\rho}+\frac{1}{\rho}\left(\sum\limits_{j=0}^{n-1}\frac{\gamma_{i,j}}{2^{j}}\right)\right)k_{a}}$ $\displaystyle=\frac{2\tau_{0}}{\min\limits_{\gamma_{i,0},\ldots,\gamma_{i,n}}\left(1+\left(1-\frac{1}{\rho}+\frac{1}{\rho}\left(\sum\limits_{j=0}^{n-1}\frac{\gamma_{i,j}}{2^{j}}\right)\right)k_{a}\right)}$ $\displaystyle=\frac{2\tau_{0}}{1+\left(1-\frac{1}{\rho}\right)k_{a}}.$ (26) In [9], the feasible region for the control gains $k_{p}$ and $k_{v}$ is specific to one chosen value of $\tilde{k}_{a}\in(0,1)$. However, in this case, $\tilde{k}_{a}$ could be any value in the interval $\mathcal{I}:=\left[\left(1-\frac{1}{\rho}\right)k_{a},\left(1+\frac{1}{\rho}\right)k_{a}\right]$ for a given $k_{a}\in\left(0,\frac{1}{1+\frac{1}{\rho}}\right)$. Let $\mathcal{F}(\tilde{k}_{a})$ denote the set of all $(k_{p},k_{v})$ that ensure robust string stability for a given value of $\tilde{k}_{a}$. Unlike in [9], we need to show that the set of all feasible $(k_{p},k_{v})$ that work for any $\tilde{k}_{a}\in\mathcal{I}$ denoted by $\mathcal{S}$ is non-empty, i.e., $\mathcal{S}:=\bigcap_{\tilde{k}_{a}\in\mathcal{I}}\mathcal{F}(\tilde{k}_{a})\neq\emptyset.$ In the remainder of the proof, we will focus on showing that $\mathcal{S}\neq\emptyset$ and explicitly construct this set for synthesizing the controller. We do so by considering the robust stability condition, $\|\tilde{H}(j\omega;\tau)\|_{\infty}^{2}\leq 1\;\forall\tau\in(0,\tau_{0}]$ and employing a time headway satisfying (III-C). According to _Definition II.2_, robust string stability is guaranteed when $\|\tilde{H}(j\omega;\tau)\|_{\infty}^{2}\leq 1$, i.e., $|\tilde{\mathcal{N}}(j\omega)|^{2}\leq|\mathcal{D}(j\omega)|^{2},\forall\omega$, which can be rewritten as $\displaystyle\tau^{2}\omega^{4}+(1-\tilde{k}_{a}^{2}-2\tau\gamma)\omega^{2}+\gamma^{2}-2k_{p}-k_{v}^{2}+2\tilde{k}_{a}k_{p}\geq 0.$ (27) The above condition is satisfied if $\forall\tau\in(0,\tau_{0}]$, $\displaystyle 1-\tilde{k}_{a}^{2}-2\tau\gamma\geq 0,$ (28a) $\displaystyle\gamma^{2}-2k_{p}-k_{v}^{2}+2\tilde{k}_{a}k_{p}\geq 0,$ (28b) i.e., $\displaystyle\gamma\leq\frac{1-\tilde{k}_{a}^{2}}{2\tau_{0}},$ (29a) $\displaystyle\gamma\geq\sqrt{2k_{p}(1-\tilde{k}_{a})+k_{v}^{2}}.$ (29b) Thus, by considering the appropriate lower and upper bounds of $\tilde{k}_{a}$ in the inequalities (29a) and (29b), these inequalities on $\gamma$ can be satisfied for all $\tilde{k}_{a}\in\mathcal{I}$ if $\gamma$ satisfies $\displaystyle\gamma\leq\min\limits_{\tilde{k}_{a}\in\mathcal{I}}\frac{1-\tilde{k}_{a}^{2}}{2\tau_{0}}=\frac{1-\left(1+\frac{1}{\rho}\right)^{2}k_{a}^{2}}{2\tau_{0}},$ (30a) $\displaystyle\gamma\geq\max\limits_{\tilde{k}_{a}\in\mathcal{I}}\sqrt{2k_{p}(1-\tilde{k}_{a})+k_{v}^{2}}$ $\displaystyle\qquad=\sqrt{2k_{p}\left(1-\left(1-\frac{1}{\rho}\right)k_{a}\right)+k_{v}^{2}},$ (30b) from which the admissible range of $\gamma$ is given by $\displaystyle\sqrt{2k_{p}\left(1-\left(1-\frac{1}{\rho}\right)k_{a}\right)+k_{v}^{2}}\leq\gamma\leq\frac{1-\left(1+\frac{1}{\rho}\right)^{2}k_{a}^{2}}{2\tau_{0}}.$ (31) Next, we will show that the set $\mathcal{S}\neq\emptyset$. First, since $\gamma=k_{v}+h_{w}k_{p}$, the right inequality in (31) can be rewritten as $\displaystyle k_{v}+h_{w}k_{p}\leq\frac{1-\left(1+\frac{1}{\rho}\right)^{2}k_{a}^{2}}{2\tau_{0}}.$ (32) Thus, an admissible set of $k_{p}$ and $k_{v}$ is given by $\displaystyle\mathcal{S}_{1}:=\left\\{(k_{p},k_{v}):k_{p}>0,k_{v}>0,\frac{k_{v}}{a_{1}}+\frac{k_{p}}{b_{1}}\leq 1\right\\},$ (33) where $\displaystyle\begin{cases}a_{1}=\dfrac{1-\left(1+\frac{1}{\rho}\right)^{2}k_{a}^{2}}{2\tau_{0}},\\\ b_{1}=\dfrac{1-\left(1+\frac{1}{\rho}\right)^{2}k_{a}^{2}}{2\tau_{0}h_{w}}=\frac{1}{h_{w}}a_{1}.\end{cases}$ (34) Second, the left inequality in (31) can be rewritten as $\displaystyle k_{v}+h_{w}k_{p}\geq\sqrt{2k_{p}\left(1-\left(1-\frac{1}{\rho}\right)k_{a}\right)+k_{v}^{2}}.$ (35) Squaring both sides of (35) and simplifying, we obtain $\displaystyle 2h_{w}k_{v}+h_{w}^{2}k_{p}\geq 2\left(1-\left(1-\frac{1}{\rho}\right)k_{a}\right).$ (36) Thus, an admissible set of $k_{p}$ and $k_{v}$ is given by $\displaystyle\mathcal{S}_{2}:=\left\\{(k_{p},k_{v}):k_{p}>0,k_{v}>0,\frac{k_{v}}{a_{2}}+\frac{k_{p}}{b_{2}}\geq 1\right\\},$ (37) where $\displaystyle\begin{cases}{}a_{2}=\dfrac{1-\left(1-\frac{1}{\rho}\right)k_{a}}{h_{w}},\\\ b_{2}=\frac{2\left(1-\left(1-\frac{1}{\rho}\right)k_{a}\right)}{h_{w}^{2}}=\frac{2}{h_{w}}a_{2}.\end{cases}$ (38) Then, combining (33) and (37), the feasible region of $k_{p}$ and $k_{v}$ is given by the set $\displaystyle\mathcal{S}=\mathcal{S}_{1}\cap\mathcal{S}_{2}.$ (39) Note that $\mathcal{S}$ is nonempty if $a_{1}\geq a_{2}$ or $b_{1}\geq b_{2}$. In particular, notice that $\frac{a_{1}}{a_{2}}=\frac{h_{w}}{2\tau_{0}}\left(\frac{1-\left(1+\frac{1}{\rho}\right)^{2}k_{a}^{2}}{1-\left(1-\frac{1}{\rho}\right)k_{a}}\right).$ Substituting $h_{w}$ from (19), we have $\frac{a_{1}}{a_{2}}>1.$ Thus, $\mathcal{S}\neq\emptyset$. This completes the proof of Statement (b) of _Theorem III.2_. In the following we prove Statement (c) of _Theorem III.2_. Let $\bar{h}_{w}(k_{a}):=\dfrac{h_{w,lb}(k_{a})}{2\tau_{0}}$, $m:=1-\frac{1}{\rho}$, and $n:=\left(1+\frac{1}{\rho}\right)^{2}$. Then, $\displaystyle\bar{h}_{w}(k_{a})=\frac{1-mk_{a}}{1-nk_{a}^{2}}$ (40) and $\displaystyle\frac{d\bar{h}_{w}(k_{a})}{dk_{a}}=\frac{-mnk_{a}^{2}+2nk_{a}-m}{(1-nk_{a}^{2})^{2}}.$ (41) Let $f(k_{a}):=-mnk_{a}^{2}+2nk_{a}-m$. Then, the roots of $f(k_{a})=0$ are given by $r_{1,2}:=\frac{n\mp\sqrt{n(n-m^{2})}}{mn}.$ Note that $r_{1}<\frac{1}{1+\frac{1}{\rho}}$ and $r_{2}>\frac{1}{1+\frac{1}{\rho}}$. The function $f(k_{a})$ vs. $k_{a}$ is provided as an illustration for $\rho=5$. Figure 4: $f(k_{a})$ vs. $k_{a}$ when $\rho=5$. Since $k_{a}\in\left(0,\frac{1}{1+\frac{1}{\rho}}\right)$, we need to consider only $r_{1}$. Hence, we have $\displaystyle\begin{cases}\dfrac{d\bar{h}_{w}}{dk_{a}}<0,\mbox{when}~{}k_{a}\in(0,r_{1});\\\ \dfrac{d\bar{h}_{w}}{dk_{a}}>0,\mbox{when}~{}k_{a}\in\left(r_{1},\frac{1}{1+\frac{1}{\rho}}\right).\end{cases}$ (42) Thus, $\displaystyle\min\limits_{k_{a}\in\left(0,\frac{1}{1+\frac{1}{\rho}}\right)}\bar{h}_{w}(k_{a})=\bar{h}_{w}(r_{1}).$ (43) Substituting $m$ and $n$ into $r_{1}$, we obtain $\displaystyle k_{a}^{*}:=r_{1}$ $\displaystyle=\left(\frac{1-\frac{1}{\sqrt{\rho}}}{1+\frac{1}{\sqrt{\rho}}}\right)\left(\frac{1}{1+\frac{1}{\rho}}\right).$ (44) Further, substituting $r_{1}$ into $\bar{h}_{w}(r_{1})$, we obtain $\displaystyle\bar{h}_{w}(r_{1})=\frac{\left(1+\frac{1}{\sqrt{\rho}}\right)^{2}}{2\left(1+\frac{1}{\rho}\right)}.$ (45) Therefore, $\displaystyle h_{w,lb}^{*}=2\tau_{0}\bar{h}_{w}(r_{1})=\tau_{0}\frac{\left(1+\frac{1}{\sqrt{\rho}}\right)^{2}}{1+\frac{1}{\rho}}.$ (46) This completes the proof of Statement (c) of _Theorem III.2_. Finally, we will prove the lower bound of the time headway given by (21) satisfies the internal stability condition. For this purpose, note that $\displaystyle h_{w}k_{p}\geq\tau_{0}k_{p}\frac{\left(1+\frac{1}{\sqrt{\rho}}\right)^{2}}{\left(1+\frac{1}{\rho}\right)}.$ (47) Thus, $\gamma=k_{v}+h_{w}k_{p}>h_{w}k_{p}\geq\tau_{0}k_{p}$. Therefore, this completes the proof of _Theorem III.2_. ∎ ###### Remark III.2 If $\rho\to\infty$ (no signal noise case), then $k_{a}\in(0,1)$ and $h_{w}=\frac{2\tau_{0}}{1+k_{a}}$ as given in in [9]. ###### Remark III.3 In (19), if $\rho\to\infty$, then the lower bound of the time headway reduces to $h_{w}\geq h_{w,lb}(k_{a})=\frac{2\tau_{0}}{1+k_{a}}$ as given in [9]. For the same $k_{a}$ value, the lower bound of the time headway increases as $\rho$ decreases in the presence of signal noise. In addition, according to (18), the upper bound of $k_{a}$ is smaller when compared to the no noise case, which also factors into the increase of the lower bound of the time headway in the presence of noise. In addition, in (21), if $\rho\to\infty$, the minimum lower bound of the time headway reduces to $h^{\ast}_{w,lb}\to\tau_{0}$, and the corresponding optimal $k_{a}$ given by (20) becomes $k_{a}^{\ast}\to 1$. ## IV Numerical Simulations In this section, we present a numerical example and simulation results to corroborate the results in Section III. We consider the following numerical values for the system parameters: $\tau_{0}=0.5$ s, $d=5$ m, $N=12$. In the numerical simulation, $\tau$ was chosen as $\tau=\tau_{0}$. We assume that the lead vehicle experiences an external disturbance which causes a perturbation on its acceleration, denoted as $a_{0}(t)$, as follows: $\displaystyle a_{0}(t)=\begin{cases}0.5\sin(0.1(t-10)),10~{}{\rm s}<t<(10+20\pi)~{}{\rm s},\\\ 0,\mbox{otherwise},\end{cases}$ (48) and under which the performance of the communication and control strategy considered in Section III will be evaluated in the following. Assume $\rho=5$, then the upper bound of $k_{a}$ can be computed as $k_{a}<\frac{1}{1+\frac{1}{\rho}}=0.8333$. Then, by choosing $k_{a}=0.5$ and substituting $\tau_{0},\rho,k_{a}$ values in (19), $h_{w}\geq 0.9375$ s. Choosing $h_{w}=0.95$ s, the feasible region of $k_{p}$ and $k_{v}$ is obtained as shown in Fig. 5, from which we choose $k_{p}=0.009$, $k_{v}=0.63$ for the numerical simulations. With the above chosen values of the time headway and control gains, the frequency response of $\left|\tilde{H}(j\omega;\tau_{0})\right|$ is shown in Fig. 6 which demonstrates string stability of the platoon. Figure 5: The feasible region of $k_{p}$ and $k_{v}$ when $k_{a}=0.5$, $h_{w}=0.95$ s. Figure 6: $|\tilde{H}(j\omega;\tau_{0})|$ when $k_{a}=0.5$, $k_{p}=0.009$, $k_{v}=0.63$, $h_{w}=0.95$ s. Suppose we model noise in (7) by choosing $n=16$ discretizations, and the expectation of the binary random variables $z_{i,j}$ are respectively: $\gamma_{i,0}=0.8055$, $\gamma_{i,1}=0.5767$, $\gamma_{i,2}=0.1829$, $\gamma_{i,3}=0.2399$, $\gamma_{i,4}=0.8865$, $\gamma_{i,5}=0.0287$, $\gamma_{i,6}=0.4899$, $\gamma_{i,7}=0.1679$, $\gamma_{i,8}=0.9787$, $\gamma_{i,9}=0.7127$, $\gamma_{i,10}=0.5005$, $\gamma_{i,11}=0.4711$, $\gamma_{i,12}=0.0596$, $\gamma_{i,13}=0.6820$, $\gamma_{i,14}=0.0424$, $\gamma_{i,15}=0.0714$. The evolutions of the inter-vehicular spacing errors are shown in Fig. 7. Figure 7: The inter-vehicular spacing errors of the following vehicles ($k_{a}=0.5$, $k_{p}=0.009$, $k_{v}=0.63$ and $h_{w}=0.95$ s). In addition, as an example, the noise and the communicated acceleration signal from vehicle 1 to vehicle 2 are shown in Fig. 8. Figure 8: The noise and the communicated acceleration signal from vehicle 1 to vehicle 2 ($k_{a}=0.5$, $k_{p}=0.009$, $k_{v}=0.63$ and $h_{w}=0.95$ s). For comparison, under the same condition as above except for $h_{w}=0.65$ s, then the frequency response of $|\tilde{H}(j\omega;\tau_{0})|$ is shown in Fig. 9 which exhibits string instability. In addition, the inter-vehicular spacing errors of the following vehicles are shown in Fig. 10. Figure 9: $|\tilde{H}(j\omega;\tau_{0})|$ when $k_{a}=0.5$, $k_{p}=0.009$, $k_{v}=0.63$, $h_{w}=0.65$ s. Figure 10: The inter-vehicular spacing errors of the following vehicles ($k_{a}=0.5$, $k_{p}=0.009$, $k_{v}=0.63$ and $h_{w}=0.65$ s). In addition, we have conducted numerical simulations to evaluate the performance of the platoon for $k_{a}^{\ast}$ and $h_{w,lb}^{\ast}$ given in Statement (c) of Theorem III.2. For $\rho=5$, we can obtain $k_{a}^{\ast}=0.3183$ and $h^{\ast}_{w,lb}=0.8727$. Choosing $k_{a}=k_{a}^{\ast}$ and $h_{w}=0.88$, the feasible region of $k_{v}$ and $k_{p}$ is shown in Fig. 11. Figure 11: The feasible region of $k_{p}$ and $k_{v}$ when $k_{a}=k_{a}^{\ast}$, $h_{w}=0.88$ s. Choosing $k_{p}=0.003$, $k_{v}=0.85$ from the feasible region, the evolution of the inter-vehicular spacing errors is provided in Fig. 12. Figure 12: The inter-vehicular spacing errors of the following vehicles ($k_{a}=k_{a}^{\ast}$, $k_{p}=0.003$, $k_{v}=0.85$ and $h_{w}=0.88$ s). Figure 13 provides the comparison of the platoon size ($x_{0}-x_{N}$) for the two time headways, $h_{w}=0.95$ s and $h_{w}=0.88$ s, indicating higher throughput as stated in Remark III.1. Figure 13: The comparison of the length of the platoon between $h_{w}=0.95$ and $h_{w}=0.88\approx h^{\ast}_{w,lb}$. From the numerical simulation results, it can be seen that this vehicle platoon is robustly string stable with the synthesized control gains and time headway according to the analysis and design procedure provided in Section III. ## V Conclusion We have investigated robust string stability of connected and autonomous vehicle platoons with cooperative adaptive cruise control systems subject to noisy V2V communication. We have derived conditions on control gains for predecessor acceleration, relative velocity and spacing errors and a lower bound for time headway that are dependent on the signal to noise ratio due to V2V communication of acceleration, while ensuring that the CACC system is internal and string stable. We have provided a systematic analysis through which one can select control gains and time headway for a given SNR. We have also provided an illustrative example and corresponding simulation results to demonstrate the main results. For future work, we will investigate extension to noisy communication when information from multiple predecessors is employed. ## References * [1] G. Ma, P. R. Pagilla, and S. Darbha, “The effect of signal noise and latency in communicated acceleration on time headway in vehicle platooning,” in _2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC)_. IEEE, 2023, pp. 4532–4537. * [2] A. Matin and H. Dia, “Impacts of connected and automated vehicles on road safety and efficiency: A systematic literature review,” _IEEE Transactions on Intelligent Transportation Systems_ , vol. 24, no. 3, pp. 2705–2736, 2023. * [3] M. Garg and M. Bouroche, “Can connected autonomous vehicles improve mixed traffic safety without compromising efficiency in realistic scenarios?” _IEEE Transactions on Intelligent Transportation Systems_ , vol. 24, no. 6, pp. 6674–6689, 2023. * [4] D. Swaroop and J. K. Hedrick, “Constant Spacing Strategies for Platooning in Automated Highway Systems,” _Journal of Dynamic Systems, Measurement, and Control_ , vol. 121, no. 3, pp. 462–470, 09 1999. * [5] J. B. Kenney, “Dedicated short-range communications (DSRC) standards in the United States,” _Proceedings of the IEEE_ , vol. 99, no. 7, pp. 1162–1182, 2011. * [6] R. Molina-Masegosa and J. Gozalvez, “System level evaluation of lte-v2v mode 4 communications and its distributed scheduling,” in _2017 IEEE 85th Vehicular Technology Conference (VTC Spring)_ , 2017, pp. 1–5. * [7] M. H. C. Garcia, A. Molina-Galan, M. Boban, J. Gozalvez, B. Coll-Perales, T. Şahin, and A. Kousaridas, “A tutorial on 5G NR V2X communications,” _IEEE Communications Surveys & Tutorials_, vol. 23, no. 3, pp. 1972–2026, 2021. * [8] A. Vinel, L. Lan, and N. Lyamin, “Vehicle-to-vehicle communication in c-acc/platooning scenarios,” _IEEE Communications Magazine_ , vol. 53, no. 8, pp. 192–197, 2015. * [9] S. Darbha, S. Konduri, and P. R. Pagilla, “Benefits of V2V communication for autonomous and connected vehicles,” _IEEE Transactions on Intelligent Transportation Systems_ , vol. 20, no. 5, pp. 1954–1963, 2018. * [10] J. Ploeg, D. P. Shukla, N. van de Wouw, and H. Nijmeijer, “Controller synthesis for string stability of vehicle platoons,” _IEEE Transactions on Intelligent Transportation Systems_ , vol. 15, no. 2, pp. 854–865, 2014. * [11] G. F. Silva, A. Donaire, A. McFadyen, and J. J. Ford, “String stable integral control design for vehicle platoons with disturbances,” _Automatica_ , vol. 127, p. 109542, 2021. * [12] R. Rajamani and C. Zhu, “Semi-autonomous adaptive cruise control systems,” _IEEE Transactions on Vehicular Technology_ , vol. 51, no. 5, pp. 1186–1192, 2002. * [13] J. Jiang, A. Astolfi, and T. Parisini, “Robust traffic wave damping via shared control,” _Transportation Research Part C: Emerging Technologies_ , vol. 128, p. 103110, 2021. * [14] W. Scholte, P. Zegelaar, and H. Nijmeijer, “A control strategy for merging a single vehicle into a platoon at highway on-ramps,” _Transportation Research Part C: Emerging Technologies_ , vol. 136, p. 103511, 2022. * [15] J. Shen, E. K. H. Kammara, and L. Du, “Fully distributed optimization-based cav platooning control under linear vehicle dynamics,” _Transportation Science_ , vol. 56, no. 2, pp. 381–403, 2022. * [16] S. Darbha, S. Konduri, and P. R. Pagilla, “Vehicle platooning with constant spacing strategies and multiple vehicle look ahead information,” _IET Intelligent Transport Systems_ , vol. 14, no. 6, pp. 589–600, 2020. * [17] Y. Bian, Y. Zheng, W. Ren, S. E. Li, J. Wang, and K. Li, “Reducing time headway for platooning of connected vehicles via v2v communication,” _Transportation Research Part C: Emerging Technologies_ , vol. 102, pp. 87–105, 2019. * [18] G. Guo, J. Kang, H. Lei, and D. Li, “Finite-time stabilization of a collection of connected vehicles subject to communication interruptions,” _IEEE Transactions on Intelligent Transportation Systems_ , vol. 23, no. 8, pp. 10 627–10 635, 2022. * [19] G. J. L. Naus, R. P. A. Vugts, J. Ploeg, M. J. G. van de Molengraft, and M. Steinbuch, “String-stable cacc design and experimental validation: A frequency-domain approach,” _IEEE Transactions on Vehicular Technology_ , vol. 59, no. 9, pp. 4268–4279, 2010. * [20] J. I. Ge and G. Orosz, “Connected cruise control among human-driven vehicles: Experiment-based parameter estimation and optimal control design,” _Transportation Research Part C: Emerging Technologies_ , vol. 95, pp. 445–459, 2018. * [21] L. Jin, M. Čičić, K. H. Johansson, and S. Amin, “Analysis and design of vehicle platooning operations on mixed-traffic highways,” _IEEE Transactions on Automatic Control_ , vol. 66, no. 10, pp. 4715–4730, 2021. * [22] S. Gong and L. Du, “Cooperative platoon control for a mixed traffic flow including human drive vehicles and connected and autonomous vehicles,” _Transportation Research Part B: Methodological_ , vol. 116, pp. 25–61, 2018\. * [23] J. Lan, D. Zhao, and D. Tian, “Data-driven robust predictive control for mixed vehicle platoons using noisy measurement,” _IEEE Transactions on Intelligent Transportation Systems_ , pp. 1–11, 2021. * [24] J. Ploeg, N. van de Wouw, and H. Nijmeijer, “Lp string stability of cascaded systems: Application to vehicle platooning,” _IEEE Transactions on Control Systems Technology_ , vol. 22, no. 2, pp. 786–793, 2014. * [25] B. Besselink and K. H. Johansson, “String stability and a delay-based spacing policy for vehicle platoons subject to disturbances,” _IEEE Transactions on Automatic Control_ , vol. 62, no. 9, pp. 4376–4391, 2017. * [26] Z. Li, B. Hu, M. Li, and G. Luo, “String stability analysis for vehicle platooning under unreliable communication links with event-triggered strategy,” _IEEE Transactions on Vehicular Technology_ , vol. 68, no. 3, pp. 2152–2164, 2019. * [27] R. A. Biroon, Z. A. Biron, and P. Pisu, “False data injection attack in a platoon of cacc: Real-time detection and isolation with a pde approach,” _IEEE Transactions on Intelligent Transportation Systems_ , vol. 23, no. 7, pp. 8692–8703, 2022. * [28] V. Vegamoor, S. Rathinam, and S. Darbha, “String stability of connected vehicle platoons under lossy V2V communication,” _IEEE Transactions on Intelligent Transportation Systems_ , vol. 23, no. 7, pp. 8834–8845, 2022. * [29] M. Hu, X. Wang, Y. Bian, D. Cao, and H. Wang, “Disturbance observer-based cooperative control of vehicle platoons subject to mismatched disturbance,” _IEEE Transactions on Intelligent Vehicles_ , vol. 8, no. 4, pp. 2748–2758, 2023. * [30] Y. Zhu, Y. Li, H. Zhu, W. Hua, G. Huang, S. Yu, S. E. Li, and X. Gao, “Joint sliding-mode controller and observer for vehicle platoon subject to disturbance and acceleration failure of neighboring vehicles,” _IEEE Transactions on Intelligent Vehicles_ , vol. 8, no. 3, pp. 2345–2357, 2023.
# t-DGR: A Trajectory-Based Deep Generative Replay Method for Continual Learning in Decision Making William Yue The University of Texas at Austin <EMAIL_ADDRESS> &Bo Liu The University of Texas at Austin <EMAIL_ADDRESS> &Peter Stone The University of Texas at Austin, Sony AI <EMAIL_ADDRESS> ###### Abstract Deep generative replay has emerged as a promising approach for continual learning in decision-making tasks. This approach addresses the problem of catastrophic forgetting by leveraging the generation of trajectories from previously encountered tasks to augment the current dataset. However, existing deep generative replay methods for continual learning rely on autoregressive models, which suffer from compounding errors in the generated trajectories. In this paper, we propose a simple, scalable, and non-autoregressive method for continual learning in decision-making tasks using a generative model that generates task samples conditioned on the trajectory timestep. We evaluate our method on Continual World benchmarks and find that our approach achieves state-of-the-art performance on the average success rate metric among continual learning methods. Code is available at https://github.com/WilliamYue37/t-DGR. ## 1 Introduction Continual learning, also known as lifelong learning, is a critical challenge in the advancement of general artificial intelligence, as it enables models to learn from a continuous stream of data encompassing various tasks, rather than having access to all data at once [1]. However, a major challenge in continual learning is the phenomenon of catastrophic forgetting, where previously learned skills are lost when attempting to learn new tasks [2]. To mitigate catastrophic forgetting, replay methods have been proposed, which involve saving data from previous tasks and replaying it to the learner during the learning of future tasks. This approach mimics how humans actively prevent forgetting by reviewing material for tests and replaying memories in dreams. However, storing data from previous tasks requires significant storage space and becomes computationally infeasible as the number of tasks increases. In the field of cognitive neuroscience, the Complementary Learning Systems theory offers insights into how the human brain manages memory. This theory suggests that the brain employs two complementary learning systems: a fast- learning episodic system and a slow-learning semantic system [3, 4, 5]. The hippocampus serves as the episodic system, responsible for storing specific memories of unique events, while the neocortex functions as the semantic system, extracting general knowledge from episodic memories and organizing it into abstract representations [6]. Figure 1: The first row presents a comparison of three generative methods for imitating an agent’s movement in a continuous 2D plane with Gaussian noise. The objective is to replicate the ground truth path, which transitions from darker to lighter colors. The autoregressive method (CRIL) encounters a challenge at the first sharp turn as nearby points move in opposing directions. Once the autoregressive method deviates off course, it never recovers and compromises the remaining trajectory. In contrast, sampling individual state observations i.i.d. without considering the temporal nature of trajectories (DGR) leads to a fragmented path with numerous gaps. Our proposed method t-DGR samples individual state observations conditioned on the trajectory timestep. By doing so, t-DGR successfully avoids the pitfalls of CRIL and DGR, ensuring a more accurate replication of the desired trajectory. The second row illustrates how each method generates trajectory data. CRIL generates the next state observation conditioned on the previous state observation. DGR, in contrast, does not attempt to generate a trajectory but generates individual state observations i.i.d. On the other hand, t-DGR generates state observations conditioned on the trajectory timestep. Drawing inspiration from the human brain, deep generative replay (DGR) addresses the catastrophic forgetting issue in decision-making tasks by using a generative model as the hippocampus to generate trajectories from past tasks and replay them to the learner which acts as the neocortex (Figure 2) [7]. The time-series nature of trajectories in decision-making tasks sets it apart from continual supervised learning, as each timestep of the trajectory requires sufficient replay. In supervised learning, the learner’s performance is not significantly affected if it performs poorly on a small subset of the data. However, in decision-making tasks, poor performance on any part of the trajectory can severely impact the overall performance. Therefore, it is crucial to generate state-action pairs that accurately represent the distribution found in trajectories. Furthermore, the high-dimensional distribution space of trajectories makes it computationally infeasible to generate complete trajectories all at once. Existing DGR methods adopt either the generation of individual state observations i.i.d. without considering the temporal nature of trajectories or autoregressive trajectory generation. Autoregressive approaches generate the next state(s) in a trajectory by modeling the conditional probability of the next state(s) given the previously generated state(s). However, autoregressive methods suffer from compounding errors in the generated trajectories. On the other hand, generating individual state observations i.i.d. leads to a higher sample complexity compared to generating entire trajectories, which becomes significant when replay time is limited (see Section 2.2). To address the issues in current DGR methods, we propose a simple, scalable, and non-autoregressive trajectory-based DGR method. We define a generated trajectory as temporally coherent if the transitions from one state to the next appear realistic (refer to Section 3.4 for a formal definition). Given that current decision-making methods are trained on state-action pairs, we do not require trajectories to exhibit temporal coherence. Instead, our focus is on ensuring an equal number of samples generated at each timestep of the trajectory to accurately represent the distribution found in trajectories. To achieve equal sample coverage at each timestep, we train our generator to produce state observations conditioned on the trajectory timestep, and then sample from the generator conditioned on each timestep of the trajectory. The intuition behind our method is illustrated in Figure 1. To evaluate the effectiveness of our proposed method, t-DGR, we conducted experiments on the Continual World benchmarks CW10 and CW20 [8] using imitation learning. Our results indicate that t-DGR achieves state-of-the-art performance in terms of average success rate when compared to other top continual learning methods. ## 2 Related Work This section provides an overview of existing continual learning methods within the context of “General Continual Learning", with a particular focus on pseudo-rehearsal methods. ### 2.1 Continual Learning in the Real World As the field of continual learning continues to grow, there is an increasing emphasis on developing methods that can be effectively applied in real-world scenarios [9, 10, 11, 12, 13]. The concept of “General Continual Learning" was introduced by Buzzega et al. [14] to address certain properties of the real world that are often overlooked or ignored by existing continual learning methods. Specifically, two important properties, bounded memory and blurry task boundaries, are emphasized in this work. Bounded memory refers to the requirement that the memory footprint of a continual learning method should remain bounded throughout the entire lifespan of the learning agent. This property is crucial to ensure practicality and efficiency in real-world scenarios. Additionally, blurry task boundaries highlight the challenge of training on tasks that are intertwined, without clear delineation of when one task ends and another begins. Many existing methods fail to account for this characteristic, which is common in real-world learning scenarios. While there are other significant properties associated with continual learning in the real world, this study focuses on the often-neglected aspects of bounded memory and blurry task boundaries. By addressing these properties, we aim to develop methods that are more robust and applicable in practical settings. ### 2.2 Continual Learning Methods Continual learning methods for decision-making tasks can be categorized into three main categories. #### Regularization Regularization methods in continual learning focus on incorporating constraints during model training to promote the retention of past knowledge. One simple approach is to include an $L_{2}$ penalty in the loss function. Elastic Weight Consolidation (EWC) builds upon this idea by assigning weights to parameters based on their importance for previous tasks using the Fisher information matrix [15]. MAS measures the sensitivity of parameter changes on the model’s output, prioritizing the retention of parameters with a larger effect [16]. VCL leverages variational inference to minimize the Kullback- Leibler divergence between the current and prior parameter distributions [17]. Progress and Compress learns new tasks using a separate model and subsequently distills this knowledge into the main model while safeguarding the previously acquired knowledge [18]. However, regularization methods may struggle with blurry task boundaries as they rely on knowledge of task endpoints to apply regularization techniques effectively. In our experiments, EWC was chosen as the representative regularization method based on its performance in the original Continual World experiments [8]. #### Architecture-based Methods Architecture-based methods aim to maintain distinct sets of parameters for each task, ensuring that future learning does not interfere with the knowledge acquired from previous tasks. Packnet [19], UCL [20], and AGS-CL [21] all safeguard previous task information in a neural network by identifying important parameters and freeing up less important parameters for future learning. Identification of important parameters can be done through iterative pruning (Packnet), parameter uncertainty (UCL), and activation value (AGS-CL). However, a drawback of parameter isolation methods is that each task requires its own set of parameters, which may eventually exhaust the available parameters for new tasks and necessitate a dynamically expanding network without bounded memory [22]. Additionally, parameter isolation methods require training on a single task at a time to prune and isolate parameters, preventing concurrent learning from multiple interwoven tasks. In our experiments, PackNet was selected as the representative architecture-based method based on its performance in the original Continual World experiments [8]. #### Pseudo-rehearsal Methods Pseudo-rehearsal methods mitigate the forgetting of previous tasks by generating synthetic samples from past tasks and replaying them to the learner. Deep generative replay (DGR) (Figure 2) utilizes a generative model, such as generative adversarial networks [23], variational autoencoders [24], or diffusion models [25, 26], to generate the synthetic samples. Originally, deep generative replay was proposed to address continual supervised learning problems, where the generator only needed to generate single data point samples [7]. However, in decision-making tasks, expert demonstrations consist of trajectories (time-series) with a significantly higher-dimensional distribution space. One existing DGR method generates individual state observations i.i.d. instead of entire trajectories. However, this approach leads to a higher sample complexity compared to generating entire trajectories. The sample complexity of generating enough individual state observations i.i.d. to cover every portion of the trajectory $m$ times can be described using the Double Dixie Cup problem [27]. For trajectories of length $n$, it takes an average of $\Theta(n\log n+mn\log\log n)$ i.i.d. samples to ensure at least $m$ samples for each timestep. In scenarios with limited replay time (small $m$) and long trajectories (large $n$) the sample complexity can be approximated as $\Theta(n\log n)$ using the Coupon Collector’s problem [28]. The additional $\Theta(\log n)$ factor reduces the likelihood of achieving complete sample coverage of the trajectory when the number of replays or replay time is limited, especially considering the computationally expensive nature of current generative methods. Furthermore, there is a risk that the generator assigns different probabilities to each timestep of the trajectory, leading to a selective focus on certain timesteps rather than equal representation across the trajectory. Another existing DGR method is autoregressive trajectory generation. In the existing autoregressive method, CRIL, a generator is used to generate samples of the initial state, and a dynamics model predicts the next state based on the current state and action [29]. However, even with a dynamics model accuracy of 99% and a 1% probability of deviating from the desired trajectory, the probability of an autoregressively generated trajectory going off course is $1-0.99^{n}$, where $n$ denotes the trajectory length. With a trajectory length of $n=200$ (as used in our experiments), the probability of an autoregressively generated trajectory going off course is $1-0.99^{200}=0.87$. This example demonstrates how the issue of compounding error leads to a high probability of failure, even with a highly accurate dynamics model. In our experiments, t-DGR is evaluated against all existing trajectory generation methods in pseudo-rehearsal approaches to assess how well t-DGR addresses the limitations of those methods. Figure 2: The deep generative replay paradigm. The algorithm learns to generate trajectories from past tasks to augment real trajectories from the current task in order to mitigate catastrophic forgetting. Both the generator and policy model are updated with this augmented dataset. ## 3 Background This section introduces notation and the formulation of the continual imitation learning problem that we use in this paper. Additionally, we provide a concise overview of diffusion probabilistic models used in our generative model implementation. ### 3.1 Imitation Learning Imitation learning algorithms aim to learn a policy $\pi_{\theta}$ parameterized by $\theta$ by imitating a set of expert demonstrations $D=\\{\tau_{i}\\}_{i=1\ldots M}$. Each trajectory $\tau_{i}$ consists of a sequence of state-action pairs $\\{(s_{j},a_{j})\\}_{j=1\ldots|\tau_{i}|}$ where $|\tau_{i}|$ is the length of the trajectory. Each trajectory comes from a task $\mathcal{T}$ which is a Markov decision process that can be represented as a tuple $\langle S,A,T,\rho_{0}\rangle$ with state space $S$, action space $A$, transition dynamics $T:S\times A\times S\to[0,1]$, and initial state distribution $\rho_{0}$. Various algorithms exist for imitation learning, including behavioral cloning, GAIL [30], and inverse reinforcement learning [31]. In this work, we use behavioral cloning where the objective can be formulated as minimizing the loss function: $\mathcal{L}(\theta)=\mathbb{E}_{s,a\sim D}\bigg{[}\big{\|}\pi_{\theta}(s)-a\big{\|}^{2}_{2}\bigg{]}$ (1) where the state and action spaces are continuous. ### 3.2 Continual Imitation Learning In the basic formulation most common in the field today, continual imitation learning involves sequentially solving multiple tasks $\mathcal{T}_{1},\mathcal{T}_{2},\ldots,\mathcal{T}_{N}$. When solving for task $\mathcal{T}_{i}$, the learner only gets data from task $\mathcal{T}_{i}$ and can not access data for any other task. In a more general scenario, certain tasks may have overlapping boundaries, allowing the learner to encounter training data from multiple tasks during certain phases of training. The learner receives a continuous stream of training data in the form of trajectories $\tau_{1},\tau_{2},\tau_{3},\ldots$ from the environment, where each trajectory $\tau$ corresponds to one of the $N$ tasks. However, the learner can only access a limited contiguous portion of this stream at any given time. Let $s_{i}$ be the success rate of task $\mathcal{T}_{i}$ after training on all $N$ tasks. The continual imitation learning objective is defined as maximizing the average success rate over all tasks: $S=\frac{1}{N}\sum_{i=1}^{N}s_{i}$ (2) The primary issue that arises from the continual learning problem formulation is the problem of catastrophic forgetting where previously learned skills are forgotten when training on a new task. ### 3.3 Diffusion Probabilistic Models Diffusion probabilistic models [25, 26] generate data through a learned reverse denoising diffusion process $p_{\theta}(x_{t-1}\mid x_{t})$. The forward diffusion process $q(x_{t}\mid x_{t-1})$ gradually adds Gaussian noise to an input $x_{0}$ at each time step $t$, ultimately resulting in pure noise $x_{T}$ at $t=T$. The forward diffusion process is defined as: $q(x_{t}\mid x_{t-1})=\mathcal{N}\left(x_{t};\sqrt{1-\beta_{t}}x_{t-1},\beta_{t}\mathbf{I}\right)$ (3) where $0<\beta_{t}<1$ is defined by a known variance schedule. In our implementation, we adopted the cosine schedule proposed by Nichol et al. [32]. For a sufficiently large time horizon $T$ and a well-behaved variance schedule, $x_{T}$ approximates an isotropic Gaussian distribution. If we had the reverse diffusion process $p(x_{t-1}\mid x_{t})$, we could sample $x_{T}\sim\mathcal{N}(0,\mathbf{I})$ and obtain a sample from $q(x_{0})$ by denoising $x_{T}$ with $p(x_{t-1}\mid x_{t})$. However, computing $p(x_{t-1}\mid x_{t})$ is intractable as it necessitates knowledge of the distribution of all possible $x_{t}$. Instead, we approximate $p(x_{t-1}\mid x_{t})$ using a neural network: $p_{\theta}(x_{t-1}\mid x_{t})=\mathcal{N}\left(x_{t-1};\mu_{\theta}(x_{t},t),\Sigma(x_{t},t)\right)$ (4) Since $q$ and $p_{\theta}$ can be viewed as a variational auto-encoder [24], we can use the variational lower bound to minimize the negative log-likelihood of the reverse process. We can express $\mu_{\theta}(x_{t},t)$ from Equation 4 as: $\mu_{\theta}(x_{t},t)=\frac{1}{\sqrt{\alpha_{t}}}\left(x_{t}-\frac{\beta_{t}}{\sqrt{1-\overline{\alpha}_{t}}}\epsilon_{\theta}(x_{t},t)\right)$ (5) where $\alpha_{t}=1-\beta_{t}$ and $\overline{\alpha}_{t}=\prod_{s=0}^{t}\alpha_{s}$. The training loss can then be defined as: $\mathcal{L}(\theta)=\mathbb{E}_{x_{0},t,\epsilon}\left[\lVert\epsilon-\epsilon_{\theta}(x_{t},t)\rVert^{2}\right]$ (6) Note that the timesteps $t$ in the diffusion process differ from the trajectory timesteps $t$. Henceforth, we will refer only to the trajectory timesteps $t$. ### 3.4 Notation Deep generative replay involves training two models: a generator $G_{\gamma}$ parameterized by $\gamma$ and a learner $\pi_{\theta}$ parameterized by $\theta$. We define $G_{\gamma}^{(i)}$ as the generator trained on tasks $\mathcal{T}_{1}\ldots\mathcal{T}_{i}$ and capable of generating data samples from tasks $\mathcal{T}_{1}\ldots\mathcal{T}_{i}$. Similarly, $\pi_{\theta}^{(i)}$ represents the learner trained on tasks $\mathcal{T}_{1}\ldots\mathcal{T}_{i}$ and able to solve tasks $\mathcal{T}_{1}\ldots\mathcal{T}_{i}$. A sequence of state observations $(s_{1},s_{2},\ldots,s_{n-1},s_{n})$ is temporally coherent if $\forall 1\leq i<n,\exists a\in A:T(s_{i},a,s_{i+1})>\varepsilon$, where $0<\varepsilon<1$ is a small constant representing a threshold for negligible probabilities. ## 4 Method Our proposed method, t-DGR, tackles the challenge of generating long trajectories by training a generator, denoted as $G_{\gamma}(j)$, which is conditioned on the trajectory timestep $j$ to generate state observations. Pseudocode for t-DGR is provided as Algorithm 1. The algorithm begins by initializing the task index, replay ratio, generator model, learner model, and learning rates (Line 1). The replay ratio, denoted as $0\leq r<1$, determines the percentage of training samples seen by the learner that are generated. Upon receiving training data from the environment, t-DGR calculates the number of trajectories to generate based on the replay ratio $r$ (Lines 4-5). The variable $L$ (Line 7) represents the maximum length of trajectories observed so far. To generate a trajectory $\tau$ of length $L$, t-DGR iterates over each timestep $1\leq j\leq L$ (Line 9). At each timestep, t-DGR generates the $j$-th state observation of the trajectory using the previous generator $G_{\gamma}^{(t-1)}$ conditioned on timestep $j$ (Line 10), and then labels it with an action using the previous policy $\pi_{\theta}^{(t-1)}$ (Line 11). After generating all timesteps in the trajectory $\tau$, t-DGR adds it to the existing training dataset (Line 14). Note that the generated state observations within a trajectory do not have temporal coherence, as each state observation is generated independently of other timesteps. This approach is acceptable since our learner is trained on state-action pairs rather than full trajectories. However, unlike generating state observations i.i.d., our method ensures equal coverage of every timestep during the generative process, significantly reducing sample complexity. Once t-DGR has augmented the training samples from the environment with our generated training samples, t-DGR employs backpropagation to update both the generator and learner using the augmented dataset (Lines 16-18). The t-DGR algorithm continues this process of generative replay throughout the agent’s lifetime, which can be infinite (Line 2). Although we perform the generative process of t-DGR at task boundaries for ease of understanding, no part of t-DGR is dependent on clear task boundaries. #### Architecture We employ a U-net [33] trained on the loss specified in Equation 6 to implement the generative diffusion model $G_{\gamma}$. Since we utilize proprioceptive observations in our experiments, $\pi_{\theta}$ is implemented with a multi-layer perceptron trained on the loss specified in Equation 1. Algorithm 1 Trajectory-based Deep Generative Replay (t-DGR) 1:Initialize task index $t=0$, replay ratio $r$, generator $G^{(0)}_{\gamma}$, learner $\pi^{(0)}_{\theta}$, and learning rates $\lambda_{\gamma},\lambda_{\theta}$. 2:while new task available do 3: $t\leftarrow t+1$ 4: Initialize dataset $D$ with trajectories from task $t$. 5: $n\leftarrow\frac{r*|D|}{1-r}$ $\triangleright$ number of trajectories to generate 6: for $i=1$ to $n$ do 7: $L\leftarrow$ maximum trajectory length 8: $\tau\leftarrow\emptyset$ $\triangleright$ initialize trajectory of length $L$ 9: for $j=1$ to $L$ do 10: $S\leftarrow G_{\gamma}^{(t-1)}(j)$ $\triangleright$ generate states 11: $A\leftarrow\pi_{\theta}^{(t-1)}(S)$ $\triangleright$ label with actions 12: $\tau_{j}\leftarrow(S,A)$ $\triangleright$ add to trajectory 13: end for 14: $D\leftarrow D\cup\tau$ $\triangleright$ add generated trajectory to $D$ 15: end for 16: Update generator and learner using $D$ 17: $\gamma^{(t)}\leftarrow\gamma^{(t-1)}-\lambda_{\gamma}\nabla_{\gamma}\mathcal{L}_{G^{(t-1)}}(\gamma^{(t-1)})$ 18: $\theta^{(t)}\leftarrow\theta^{(t-1)}-\lambda_{\theta}\nabla_{\theta}\mathcal{L}_{\pi^{(t-1)}}(\theta^{(t-1)})$ 19:end while ## 5 Experiments In this section, we outline the experimental setup and performance metrics employed to compare t-DGR with representative methods, followed by an analysis of experimental results across different benchmarks and performance metrics. ### 5.1 Experimental Setup We evaluate our method on the Continual World benchmarks CW10 and CW20 [8], along with our own “General Continual Learning" variant of CW10 called GCL10. CW10 consists of a sequence of 10 Meta-World [34] tasks, where each task involves a Sawyer arm manipulating one or two objects in the Mujoco physics simulator. For computational efficiency, we provide the agents with proprioceptive observations. Notably, the observation and action spaces are continuous and remain consistent across all tasks. CW20 is an extension of CW10 with the tasks repeated twice. To our knowledge, Continual World is the only standard continual learning benchmark for decision-making tasks. GCL10 gives data to the learner in 10 sequential buckets $B_{1},\ldots,B_{10}$. Data from task $\mathcal{T}_{i}$ from CW10 is split evenly between buckets $B_{i-1}$, $B_{i}$, and $B_{i+1}$, except for the first and last task. Task $\mathcal{T}_{1}$ is evenly split between buckets $B_{0}$ and $B_{1}$, and task $\mathcal{T}_{10}$ is evenly split between buckets $B_{9}$ and $B_{10}$. In order to ensure bounded memory usage, we adopt a one-hot vector approach to condition the model on the task, rather than maintaining a separate final neural network layer for each individual task. Additionally, we do not allow separate biases for each task, as originally done in EWC [15]. Expert demonstrations for training are acquired by gathering 100 trajectories per task using hand-designed policies from Meta-World, with each trajectory limited to a maximum of 200 steps. Importantly, the learner model remains consistent across different methods and benchmark evaluations. Moreover, we maintain a consistent replay ratio of $r=0.9$ across all pseudo-rehearsal methods. We estimated the success rate $S$ of a model by running each task 100 times. The metrics for each method were computed using 5 seeds to create a 90% confidence interval. Further experimental details, such as hyperparameters, model architecture, random seeds, and computational resources, are included in the appendix. This standardization enables a fair and comprehensive comparison of our proposed approach with other existing methods. ### 5.2 Metrics We evaluate our models using three metrics proposed by the Continual World benchmark [8], with the average success rate being the primary metric. Although the forward transfer and forgetting metrics are not well-defined in a “General Continual Learning" setting, they are informative within the context of Continual World benchmarks. As a reminder from Section 3.2, let $N$ denote the number of tasks, and $s_{i}$ represent the success rate of the learner on task $\mathcal{T}_{i}$. Additionally, let $s_{i}(t)$ denote the success rate of the learner on task $\mathcal{T}_{i}$ after training on tasks $\mathcal{T}_{1}$ to $\mathcal{T}_{t}$. #### Average Success Rate The average success rate, as given by Equation 2, serves as the primary evaluation metric for continual learning methods. #### Average Forward Transfer We introduce a slightly modified metric for forward transfer that applies to a broader range of continual learning problems beyond just continual reinforcement learning in the Continual World benchmark. Let $s_{i}^{\mathrm{ref}}$ represent the reference performance of a single-task experiment on task $\mathcal{T}_{i}$. The forward transfer metric $FT_{i}$ is computed as follows: $\displaystyle FT_{i}=\frac{D_{i}-D_{i}^{\mathrm{ref}}}{1-D_{i}^{\mathrm{ref}}}$ $\displaystyle D_{i}=\frac{s_{i}(i)+s_{i}(i-1)}{2}$ $\displaystyle D_{i}^{\mathrm{ref}}=\frac{s_{i}^{\mathrm{ref}}}{2}$ The expressions for $D_{i}$ and $D^{\mathrm{ref}}_{i}$ serve as approximations of the integral of task $\mathcal{T}_{i}$ performance with respect to the training duration for task $\mathcal{T}_{i}$. The average forward transfer $FT$ is then defined as the mean forward transfer over all tasks, calculated as $FT=\frac{1}{N}\sum_{i=1}^{N}FT_{i}$. #### Average Forgetting We measure forgetting using the metric $F_{i}$, which represents the amount of forgetting for task $i$ after all training has concluded. $F_{i}$ is defined as the difference between the success rate on task $\mathcal{T}_{i}$ immediately after training and the success rate on task $\mathcal{T}_{i}$ at the end of training. $F_{i}=s_{i}(i)-s_{i}(N)$ The average forgetting $F$ is then computed as the mean forgetting over all tasks, given by $F=\frac{1}{N}\sum_{i=1}^{N}F_{i}$. ### 5.3 Baselines We compare the following methods on the Continual World benchmark using average success rate as the primary evaluation metric. Representative methods were chosen based on their success in the original Continual World experiments, while DGR-based methods were selected to evaluate whether t-DGR addresses the limitations of existing pseudo-rehearsal methods. * • Finetune: The policy is trained only on data from the current task. * • Multitask: The policy is trained on data from all tasks simultaneously. * • oEWC [18]: A variation of EWC known as online Elastic Weight Consolidation (oEWC) bounds the memory of EWC by employing a single penalty term for the previous model instead of individual penalty terms for each task. This baseline is the representative regularization-based method. * • PackNet [19]: This baseline is the representative parameter isolation method. Packnet safeguards previous task information in a neural network by iteratively pruning, freezing, and retraining parts of the network. * • DGR [7]: This baseline is a deep generative replay method that only generates individual state observations i.i.d. and not entire trajectories. * • CRIL [29]: This baseline is a deep generative replay method that trains a policy along with a start state generator and a dynamics model that predicts the next state given the current state and action. Trajectories are generated by using the dynamics model and policy to autoregressively generate next states from a start state. * • t-DGR: Our proposed method. Due to the inability of oEWC and PackNet to handle blurry task boundaries, we made several adjustments for CW20 and GCL10. Since PackNet cannot continue training parameters for a task once they have been fixed, we treated the second repetition of tasks in CW20 as distinct from the first iteration, resulting in PackNet being evaluated with $N=20$, while the other methods were evaluated with $N=10$. As for GCL10 and its blurry task boundaries, the best approach we could adopt with oEWC and PackNet was to apply their regularization techniques at regular training intervals rather than strictly at task boundaries. During evaluation, all tasks were assessed using the last fixed set of parameters in the case of PackNet. (a) CW10 Method | Success Rate $\uparrow$ | FT$\uparrow$ | Forgetting$\downarrow$ ---|---|---|--- Finetune | 16.4 $\pm$6.4 | -3.0 $\pm$6.0 | 78.8 $\pm$7.6 Multitask | 97.0 $\pm$1.0 | N/A | N/A oEWC | 18.6 $\pm$5.3 | -6.3 $\pm$5.7 | 74.1 $\pm$6.1 PackNet | 81.4 $\pm$3.7 | -14.8 $\pm$7.8 | -0.1 $\pm$1.2 DGR | 75.0 $\pm$5.8 | -4.3 $\pm$5.1 | 17.8 $\pm$4.1 CRIL | 28.4 $\pm$10.6 | -1.1 $\pm$2.8 | 68.6 $\pm$10.4 t-DGR | 81.9 $\pm$3.3 | -0.3 $\pm$4.9 | 14.4 $\pm$2.5 (b) GCL10 Method | Success Rate $\uparrow$ ---|--- Finetune | 21.7 $\pm$2.6 Multitask | 97.0 $\pm$1.0 oEWC | 21.8 $\pm$1.7 PackNet | 26.9 $\pm$5.6 DGR | 75.3 $\pm$4.4 CRIL | 53.5 $\pm$5.5 t-DGR | 81.7 $\pm$4.0 (c) CW20 Method | Success Rate $\uparrow$ | FT$\uparrow$ | Forgetting$\downarrow$ ---|---|---|--- Finetune | 14.2 $\pm$4.0 | -0.5 $\pm$3.0 | 82.2 $\pm$5.6 Multitask | 97.0 $\pm$1.0 | N/A | N/A oEWC | 19.4 $\pm$5.3 | -2.8 $\pm$4.1 | 75.2 $\pm$7.5 PackNet | 74.1 $\pm$4.1 | -20.4 $\pm$3.4 | -0.2 $\pm$0.9 DGR | 74.1 $\pm$4.1 | 18.9 $\pm$2.9 | 23.3 $\pm$3.3 CRIL | 50.8 $\pm$4.4 | 4.4 $\pm$4.9 | 46.1 $\pm$5.4 t-DGR | 83.9 $\pm$3.0 | 30.6 $\pm$4.5 | 14.6 $\pm$2.9 (d) Replay Ratio Ratio | t-DGR | DGR ---|---|--- 0.5 | 63.2 $\pm$2.6 | 52.8 $\pm$2.9 0.6 | 66.3 $\pm$4.4 | 56.9 $\pm$4.5 0.7 | 70.8 $\pm$4.1 | 62.5 $\pm$3.6 0.8 | 75.0 $\pm$6.9 | 69.2 $\pm$4.9 0.9 | 81.9 $\pm$3.3 | 75.0 $\pm$5.8 Table 1: Tables (a), (b), and (c) present the results for Continual World 10, General Continual Learning 10, and Continual World 20, respectively. The tables display the average success rate, forward transfer, and forgetting (if applicable) with 90% confidence intervals using 5 random seeds. An up arrow indicates that higher values are better and a down arrow indicates that smaller values are better. Table (d) compares the impact of replay amount on the average success rate of t-DGR and DGR on CW10 with 90% confidence intervals obtained using 5 random seeds. The best results are highlighted in bold. ### 5.4 Discussion t-DGR emerges as the leading method, demonstrating the highest success rate on CW10 (Table 1(a)), CW20 (Table 1(c)), and GCL10 (Table 1(b)). Notably, PackNet’s performance on the second iteration of tasks in CW20 diminishes, highlighting its limited capacity for continually accommodating new tasks. This limitation underscores the fact that PackNet falls short of being a true lifelong learner, as it necessitates prior knowledge of the task count for appropriate parameter capacity allocation. On the contrary, pseudo-rehearsal methods, such as t-DGR, exhibit improved performance with the second iteration of tasks in CW20 due to an increased replay time. These findings emphasize the ability of DGR methods to effectively leverage past knowledge, as evidenced by their superior forward transfer in both CW10 and CW20. GCL10 (Table 1(b)) demonstrates that pseudo-rehearsal methods are mostly unaffected by blurry task boundaries, whereas PackNet’s success rate experiences a significant drop-off. This discrepancy arises from the fact that PackNet’s regularization technique does not work effectively with less clearly defined task boundaries. Additionally, the diminishing performance gap between DGR and t-DGR as the replay ratio increases in Table 1(d) indicates that a higher replay ratio reduces the likelihood of any portion of the trajectory being insufficiently covered when sampling individual state observations i.i.d., thereby contributing to improved performance. This trend supports the theoretical sample complexity of DGR derived in Section 2.2, as $\Theta(n\log n+mn\log\log n)$ closely approximates the sample complexity of t-DGR, $\Theta(mn)$, when the replay amount $m\to\infty$. However, while DGR can achieve comparable performance to t-DGR with a high replay ratio, the availability of extensive replay time is often limited in many real-world applications. Overall, t-DGR exhibits promising results, outperforming other methods in terms of success rate in all evaluations. Notably, t-DGR achieves a significant improvement over existing pseudo-rehearsal methods on CW20 using a Welch t-test with a significance level of $\text{p-value}=0.005$. Its ability to handle blurry task boundaries, leverage past knowledge, and make the most of replay opportunities position it as a state-of-the-art method for continual lifelong learning in decision-making. ## 6 Conclusion In conclusion, we have introduced t-DGR, a novel method for continual learning in decision-making tasks, which has demonstrated state-of-the-art performance on the Continual World benchmarks. Our approach stands out due to its simplicity, scalability, and non-autoregressive nature, positioning it as a solid foundation for future research in this domain. Importantly, t-DGR aligns with the concept of “General Continual Learning" by taking into account essential properties of the real world, including bounded memory and blurry task boundaries. These considerations ensure that our method remains applicable and effective in real-world scenarios, enabling its potential integration into practical applications. Looking ahead, one potential avenue for future research is the refinement of the replay mechanism employed in t-DGR. Rather than assigning equal weight to all past trajectories, a more selective approach could be explored. By prioritizing certain memories over others and strategically determining when to replay memories to the learner, akin to human learning processes, we could potentially enhance the performance and adaptability of our method. ## Acknowledgements This work has taken place in the Learning Agents Research Group (LARG) at the Artificial Intelligence Laboratory, The University of Texas at Austin. LARG research is supported in part by the National Science Foundation (FAIN-2019844, NRT-2125858), the Office of Naval Research (N00014-18-2243), Army Research Office (E2061621), Bosch, Lockheed Martin, and Good Systems, a research grand challenge at the University of Texas at Austin. The views and conclusions contained in this document are those of the authors alone. Peter Stone serves as the Executive Director of Sony AI America and receives financial compensation for this work. The terms of this arrangement have been reviewed and approved by the University of Texas at Austin in accordance with its policy on objectivity in research. ## References * [1] Mark Ring. Continual Learning in Reinforcement Environments. PhD thesis, University of Texas at Austin, 1994. * [2] Michael McCloskey and Neal J. Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. volume 24 of Psychology of Learning and Motivation, pages 109–165. Academic Press, 1989. * [3] James L. McClelland, Bruce L. McNaughton, and Randall C. O’Reilly. Complementary learning systems within the hippocampus: A neural network modeling approach to understanding episodic memory consolidation. Psychological Review, 102(3):419–457, 1995. * [4] Dharshan Kumaran, Demis Hassabis, and James L. McClelland. What learning systems do intelligent agents need? complementary learning systems theory updated. Trends in Cognitive Sciences, 20(7):512–534, 2016. * [5] James Mcclelland, Bruce Mcnaughton, and Randall O’Reilly. Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. Psychological review, 102:419–57, 08 1995. * [6] Randall C. O’Reilly and Kenneth A. Norman. Hippocampal and neocortical contributions to memory: Advances in the complementary learning systems framework. Trends in Cognitive Sciences, 6(12):505–510, December 2002. * [7] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay, 2017. * [8] Maciej Wołczyk, Michał Zając, Razvan Pascanu, Łukasz Kuciński, and Piotr Miłoś. Continual world: A robotic benchmark for continual reinforcement learning, 2021. * [9] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application, 2023. * [10] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. * [11] Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples, 2021. * [12] Yen-Chang Hsu, Yen-Cheng Liu, Anita Ramasamy, and Zsolt Kira. Re-evaluating continual learning scenarios: A categorization and case for strong baselines, 2019. * [13] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning, 2019. * [14] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline, 2020. * [15] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521–3526, mar 2017. * [16] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget, 2018. * [17] Cuong V. Nguyen, Yingzhen Li, Thang D. Bui, and Richard E. Turner. Variational continual learning, 2018. * [18] Jonathan Schwarz, Jelena Luketina, Wojciech M. Czarnecki, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning, 2018. * [19] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning, 2018. * [20] Hongjoon Ahn, Sungmin Cha, Donggyu Lee, and Taesup Moon. Uncertainty-based continual learning with adaptive regularization, 2019. * [21] Sangwon Jung, Hongjoon Ahn, Sungmin Cha, and Taesup Moon. Continual learning with node-importance based adaptive group sparse regularization, 2021. * [22] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks, 2018. * [23] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. * [24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. * [25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. * [26] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. CoRR, abs/1503.03585, 2015. * [27] Donald J. Newman. The double dixie cup problem. The American Mathematical Monthly, 67(1):58–61, 1960. * [28] Amy N. Myers and Herbert S. Wilf. Some new aspects of the coupon-collector’s problem, 2003. * [29] Chongkai Gao, Haichuan Gao, Shangqi Guo, Tianren Zhang, and Feng Chen. Cril: Continual robot imitation learning via generative and prediction model, 2021. * [30] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning, 2016. * [31] Andrew Y. Ng and Stuart J. Russell. Inverse reinforcement learning. In Proceedings of the 17th International Conference on Machine Learning (ICML-2000), pages 663–670, 2000. * [32] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models, 2021. * [33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. * [34] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Avnish Narayan, Hayden Shively, Adithya Bellathur, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning, 2021. ## Appendix ## Appendix A Hyperparameters ### A.1 Finetune Hyperparameter | Value | Brief Description ---|---|--- batch size | 32 | number of samples in each training iteration epochs | 250 | number of times the entire dataset is passed through per task learning rate | $10^{-4}$ | learning rate for gradient descent optimization algorithm | Adam | optimization algorithm used $\beta_{1}$ | 0.9 | exponential decay rate for first moment estimates in Adam $\beta_{2}$ | 0.999 | exponential decay rate for second moment estimates in Adam epsilon | $10^{-8}$ | small constant for numerical stability weight decay | 0 | weight regularization ### A.2 Multitask Hyperparameter | Value | Brief Description ---|---|--- batch size | 32 | number of samples in each training iteration epochs | 500 | number of times the entire dataset of all tasks is passed through learning rate | $10^{-4}$ | learning rate for gradient descent optimization algorithm | Adam | optimization algorithm used $\beta_{1}$ | 0.9 | exponential decay rate for first moment estimates in Adam $\beta_{2}$ | 0.999 | exponential decay rate for second moment estimates in Adam epsilon | $10^{-8}$ | small constant for numerical stability weight decay | 0 | weight regularization ### A.3 oEWC Hyperparameter | Value | Brief Description ---|---|--- batch size | 32 | number of samples in each training iteration epochs | 250 | number of times the entire dataset of all tasks is passed through learning rate | $10^{-4}$ | learning rate for gradient descent Fisher multiplier | $10^{2}$ | the Fisher is scaled by this number to form the EWC penalty optimization algorithm | Adam | optimization algorithm used $\beta_{1}$ | 0.9 | exponential decay rate for first moment estimates in Adam $\beta_{2}$ | 0.999 | exponential decay rate for second moment estimates in Adam epsilon | $10^{-8}$ | small constant for numerical stability weight decay | 0 | weight regularization The Fisher multiplier hyperparameter was tuned with the values: $10^{-2},10^{-1},10^{0},10^{1},10^{2},10^{3},10^{4},10^{5},10^{6}$. We selected the value $10^{2}$ based on the success rate metric given by Equation 2. ### A.4 PackNet Hyperparameter | Value | Brief Description ---|---|--- batch size | 32 | number of samples in each training iteration epochs | 250 | number of times the entire dataset of all tasks is passed through retrain epochs | 125 | number of training epochs after pruning learning rate | $10^{-4}$ | learning rate for gradient descent prune percent | 0.75 | percent of free parameters pruned for future tasks optimization algorithm | Adam | optimization algorithm used $\beta_{1}$ | 0.9 | exponential decay rate for first moment estimates in Adam $\beta_{2}$ | 0.999 | exponential decay rate for second moment estimates in Adam epsilon | $10^{-8}$ | small constant for numerical stability weight decay | 0 | weight regularization The retrain epochs and prune percent hyperparameters were chosen following the approach in the original PackNet paper. After training the first task, bias layers are frozen. ### A.5 DGR Hyperparameter | Value | Brief Description ---|---|--- batch size | 32 | number of samples in each training iteration epochs | 250 | number of times the entire dataset of all tasks is passed through learning rate | $10^{-4}$ | learning rate for gradient descent diffusion training steps | $10^{4}$ | number of training steps for the diffusion model per task diffusion warmup steps | $5*10^{4}$ | number of extra training steps for the diffusion model on the first task diffusion timesteps | $10^{3}$ | number of timesteps in the diffusion process replay ratio | 0.9 | percentage of training examples that are generated optimization algorithm | Adam | optimization algorithm used $\beta_{1}$ | 0.9 | exponential decay rate for first moment estimates in Adam $\beta_{2}$ | 0.999 | exponential decay rate for second moment estimates in Adam epsilon | $10^{-8}$ | small constant for numerical stability weight decay | 0 | weight regularization ### A.6 CRIL Hyperparameter | Value | Brief Description ---|---|--- batch size | 32 | number of samples in each training iteration epochs | 300 | number of times the entire dataset of all tasks is passed through learning rate | $10^{-4}$ | learning rate for gradient descent diffusion training steps | $10^{4}$ | number of training steps for the diffusion model per task diffusion warmup steps | $5*10^{4}$ | number of extra training steps for the diffusion model on the first task diffusion timesteps | $10^{3}$ | number of timesteps in the diffusion process replay ratio | 0.9 | percentage of training examples that are generated optimization algorithm | Adam | optimization algorithm used $\beta_{1}$ | 0.9 | exponential decay rate for first moment estimates in Adam $\beta_{2}$ | 0.999 | exponential decay rate for second moment estimates in Adam epsilon | $10^{-8}$ | small constant for numerical stability weight decay | 0 | weight regularization ### A.7 t-DGR Hyperparameter | Value | Brief Description ---|---|--- batch size | 32 | number of samples in each training iteration epochs | 300 | number of times the entire dataset of all tasks is passed through learning rate | $10^{-4}$ | learning rate for gradient descent diffusion training steps | $10^{4}$ | number of training steps for the diffusion model per task diffusion warmup steps | $5*10^{4}$ | number of extra training steps for the diffusion model on the first task diffusion timesteps | $10^{3}$ | number of timesteps in the diffusion process replay ratio | 0.9 | percentage of training examples that are generated optimization algorithm | Adam | optimization algorithm used $\beta_{1}$ | 0.9 | exponential decay rate for first moment estimates in Adam $\beta_{2}$ | 0.999 | exponential decay rate for second moment estimates in Adam epsilon | $10^{-8}$ | small constant for numerical stability weight decay | 0 | weight regularization ## Appendix B Model Architecture Layer (type) | Output Shape | Param # ---|---|--- Linear-1 | [32, 512] | 25,600 Linear-2 | [32, 512] | 262,656 Linear-3 | [32, 512] | 262,656 Linear-4 | [32, 512] | 262,656 Linear-5 | [32, 4] | 2,052 Total params: | 815,620 Trainable params: | 815,620 Non-trainable params: | 0 Table 2: Multi-layer perceptron architecture of the learner shared by all methods ## Appendix C Experiment Details We utilized the following random seeds for the experiments: 1, 2, 3, 4, 5. All experiments were conducted on Nvidia A100 GPUs with 80 GB of memory. The computational node consisted of an Intel Xeon Gold 6342 2.80GHz CPU with 500 GB of memory. For our longest benchmark, CW20, the runtimes were as follows: DGR and t-DGR took 3 days, CRIL took 16 hours, finetune and oEWC took 6 hours, and PackNet took 8 hours.
# Evaluating Models’ Local Decision Boundaries via Contrast Sets Matt Gardner★♢ Yoav ArtziΓ Victoria Basmova♢♣ Jonathan Berant♢♠ Ben Bogin♠ Sihao Chen♡ Pradeep Dasigi♢ Dheeru Dua□ Yanai Elazar♢♣ Ananth Gottumukkala□ Nitish Gupta♡ Hanna Hajishirzi♢△ Gabriel Ilharco△ Daniel Khashabi♢ Kevin Lin+ Jiangming Liu♢† Nelson F. Liu¶ Phoebe Mulcaire△ Qiang Ning♢ Sameer Singh□ Noah A. Smith♢△ Sanjay Subramanian♢ Reut Tsarfaty♢♣ Eric Wallace+ Ally ZhangΓ Ben Zhou♡ ♢Allen Institute for AI ΓCornell University ♣Bar-Ilan University ♠Tel-Aviv University ♡University of Pennsylvania △University of Washington □UC Irvine +UC Berkeley †University of Edinburgh ¶Stanford University <EMAIL_ADDRESS> ###### Abstract Standard test sets for supervised learning evaluate in-distribution generalization. Unfortunately, when a dataset has systematic gaps (e.g., annotation artifacts), these evaluations are misleading: a model can learn simple decision rules that perform well on the test set but do not capture the abilities a dataset is intended to test. We propose a more rigorous annotation paradigm for NLP that helps to close systematic gaps in the test data. In particular, after a dataset is constructed, we recommend that the dataset authors manually perturb the test instances in small but meaningful ways that (typically) change the gold label, creating _contrast sets_. Contrast sets provide a local view of a model’s decision boundary, which can be used to more accurately evaluate a model’s true linguistic capabilities. We demonstrate the efficacy of contrast sets by creating them for 10 diverse NLP datasets (e.g., DROP reading comprehension, UD parsing, and IMDb sentiment analysis). Although our contrast sets are not explicitly adversarial, model performance is significantly lower on them than on the original test sets—up to 25% in some cases. We release our contrast sets as new evaluation benchmarks and encourage future dataset construction efforts to follow similar annotation processes. ## 1 Introduction ## 2 Contrast Sets (a) A two-dimensional dataset that requires a complex decision boundary to achieve high accuracy. (b) If the same data distribution is instead sampled with systematic gaps (e.g., due to annotator bias), a simple decision boundary _can perform well on i.i.d. test data_ (shown outlined in pink). (c) Since filling in all gaps in the distribution is infeasible, a _contrast set_ instead fills in a local ball around a test instance to evaluate the model’s decision boundary. Figure 1: An illustration of how contrast sets provide a more comprehensive model evaluation when datasets have systematic gaps. Dataset | Original Instance | Contrastive Instance (color = edit) ---|---|--- IMDb | Hardly one to be faulted for his ambition or his vision, it is genuinely unexpected, then, to see all Park’s effort add up to so very little. …The premise is promising, gags are copious and offbeat humour abounds but it all fails miserably to create any meaningful connection with the audience. (Label: Negative) | Hardly one to be faulted for his ambition or his vision, here we see all Park’s effort come to fruition. …The premise is perfect, gags are hilarious and offbeat humour abounds, and it creates a deep connection with the audience. (Label: Positive) MATRES | Colonel Collins followed a normal progression once she was picked as a NASA astronaut. (“picked” was before “followed”) | Colonel Collins followed a normal progression before she was picked as a NASA astronaut. (“picked” was after “followed”) UD English | They demanded talks with local US commanders. I attach a paper on gas storage value modeling. I need to get a job at the earliest opportunity. | They demanded talks with great urgency. I attach a paper on my own initiative. I need to get a job at House of Pies. PERSPECTRUM | Claim: Should uniforms be worn at school. Perspective: School uniforms emphasize the socio-economic divisions they are supposed to eliminate. Label: Against | Claim: Should uniforms be banned at school. Perspective: School uniforms emphasize the socio-economic divisions they are supposed to eliminate. Label: For DROP | Context: In the spring of 1625 the Spanish regained Bahia in Brazil and Breda in the Netherlands from the Dutch. In the autumn they repulsed the English at Cadiz. Question: What event happened first, the Spanish repulsed the English at Cadiz or the Spanish regained Bahia? | Context: In the spring of 1625 the Spanish regained Bahia in Brazil and Breda in the Netherlands from the Dutch. In winter the year earlier they had repulsed the English at Cadiz. Question: What event happened first, the Spanish repulsed the English at Cadiz or the Spanish regained Bahia? Quoref | Context: Matt Helm is a secret agent. His assignment is to stop the sinister Tung-Tze, armed with spy gadgets. Helm prevails with Gail by his side as he destroys Tung-Tze. Question: Who is armed with spy gadgets? | Context: Matt Helm is a secret agent. His assignment is to stop the sinister Tung-Tze, even though he is armed with spy gadgets. Helm prevails with Gail by his side as he destroys Tung-Tze. Question: Who is armed with spy gadgets? MC-TACO | Context: She renews in Ranchipur an acquaintance with a former lover, Tom Ransome, now a dissolute alcoholic. Question: How frequently does Tom drink? Candidate Answer: Every other night Label: Likely | Context: She renews in Ranchipur an acquaintance with a former lover, Tom Ransome, who keeps very healthy habits. Question: How frequently does Tom drink? Candidate Answer: Every other night Label: Unlikely Table 1: We create contrast sets for 10 datasets and show instances from seven of them here. ### 2.1 The Problem We first give a sketch of the problem that contrast sets attempt to solve in a toy two-dimensional classification setting as shown in Figure 1. Here, the true underlying data distribution requires a complex decision boundary (Figure 1(a)). However, as is common in practice, our toy dataset is rife with systematic gaps (e.g., due to annotator bias, repeated patterns, etc.). This causes simple decision boundaries to emerge (Figure 1(b)). And, because our biased dataset is split _i.i.d._ into train and test sets, this simple decision boundary will perform well on test data. Ideally, we would like to fill in all of a dataset’s systematic gaps, however, this is usually impossible. Instead, we create a _contrast set_ : a collection of instances tightly clustered in input space around a single test instance, or _pivot_ (Figure 1(c); an $\epsilon$-ball in our toy example). This contrast set allows us to measure how well a model’s decision boundary aligns with the correct decision boundary local to the pivot. In this case, the contrast set demonstrates that the model’s simple decision boundary is incorrect. We repeat this process around numerous pivots to form entire evaluation datasets. When we move from toy settings to complex NLP tasks, the precise nature of a “systematic gap” in the data becomes harder to define. Indeed, the geometric view in our toy examples does not correspond directly to experts’ perception of data; there are many ways to “locally perturb” natural language. We do not expect intuition, even of experts, to exhaustively reveal gaps. Nevertheless, the presence of these gaps is well-documented Gururangan et al. (2018); Poliak et al. (2018); Min et al. (2019), and Niven and Kao (2019) give an initial attempt at formally characterizing them. In particular, one common source is annotator bias from data collection processes Geva et al. (2019). For example, in the SNLI dataset Bowman et al. (2015), Gururangan et al. (2018) show that the words _sleeping_ , _tv_ , and _cat_ almost never appear in an entailment example, either in the training set or the test set, though they often appear in contradiction examples. This is not because these words are particularly important to the phenomenon of entailment; their absence in entailment examples is a _systematic gap_ in the data that can be exploited by models to achieve artificially high test accuracy. This is but one kind of systematic gap; there are also biases due to the writing styles of small groups of annotators Geva et al. (2019), the distributional biases in the data that was chosen for annotation, as well as numerous other biases that are more subtle and harder to discern Shah et al. (2020). Completely removing these gaps in the initial data collection process would be ideal, but is likely impossible—language has too much inherent variability in a very high-dimensional space. Instead, we use contrast sets to fill in gaps in the test data to give more thorough evaluations than what the original data provides. ### 2.2 Definitions We begin by defining a _decision boundary_ as a partition of some space into labels.111In this discussion we are talking about the _true_ decision boundary, not a _model’s_ decision boundary. This partition can be represented by the set of all points in the space with their associated labels: $\\{(x,y)\\}$. This definition differs somewhat from the canonical definition, which is a collection of hypersurfaces that separate labels. There is a bijection between partitions and these sets of hypersurfaces in continuous spaces, however, so they are equivalent definitions. We choose to use the partition to represent the decision boundary as it makes it very easy to define a _local_ decision boundary and to generalize the notion to discrete spaces, which we deal with in NLP. A _local decision boundary_ around some _pivot_ $x$ is the set of all points $x^{\prime}$ and their associated labels $y^{\prime}$ that are within some distance $\epsilon$ of $x$. That is, a local decision boundary around $x$ is the set $\\{(x^{\prime},y^{\prime})~{}|~{}d(x,x^{\prime})<\epsilon\\}$. Note here that even though a “boundary” or “surface” is hard to visualize in a discrete input space, using this partition representation instead of hypersurfaces gives us a uniform definition of a local decision boundary in any input space; all that is needed is a distance function $d$. A _contrast set_ $C(x)$ is any sample of points from a local decision boundary around $x$. In other words, $C(x)$ consists of inputs $x^{\prime}$ that are similar to $x$ according to some distance function $d$. Typically these points are sampled such that $y^{\prime}\neq y$. To evaluate a model using these contrast sets, we define the _contrast consistency_ of a model to be whether it makes correct predictions $\hat{y}$ on every element in the set: $\mathrm{all}(\\{\hat{y}=y^{\prime}~{}\forall(x^{\prime},y^{\prime})\in C(x)\\})$. Since the points $x^{\prime}$ were chosen from the local decision boundary, we expect contrast consistency on expert-built contrast sets to be a significantly more accurate evaluation of whether model predictions match the task definition than a random selection of input / output pairs. ### 2.3 Contrast sets in practice Given these definitions, we now turn to the actual construction of contrast sets in practical NLP settings. There were two things left unspecified in the definitions above: the distance function $d$ to use in discrete input spaces, and the method for sampling from a local decision boundary. While there has been some work trying to formally characterize distances for adversarial robustness in NLP Michel et al. (2019); Jia et al. (2019), we find it more useful in our setting to simply rely on expert judgments to generate a similar but meaningfully different $x^{\prime}$ given $x$, addressing both the distance function and the sampling method. Future work could try to give formal treatments of these issues, but we believe expert judgments are sufficient to make initial progress in improving our evaluation methodologies. And while expert-crafted contrast sets can only give us an upper bound on a model’s local alignment with the true decision boundary, an upper bound on local alignment is often more informative than a potentially biased _i.i.d._ evaluation that permits artificially simple decision boundaries. To give a tighter upper bound, we draw pivots $x$ from some _i.i.d._ test set, and we do not provide _i.i.d._ contrast sets at training time, which could provide additional artificially simple decision boundaries to a model. Figure LABEL:fig:teaser displays an example contrast set for the NLVR2 visual reasoning dataset Suhr and Artzi (2019). Here, both the sentence and the image are modified in small ways (e.g., by changing a word in the sentence or finding a similar but different image) to make the output label change. A contrast set is _not_ a collection of adversarial examples Szegedy et al. (2014). Adversarial examples are almost the methodological opposite of contrast sets: they change the input such that a model’s decision changes but the gold label does not Jia and Liang (2017); Wallace et al. (2019a). On the other hand, contrast sets are model-agnostic, constructed by experts to characterize whether a model’s decision boundary locally aligns to the true decision boundary around some point. Doing this requires input changes that also induce changes to the gold label. We recommend that the original dataset authors—the experts on the linguistic phenomena intended to be reflected in their dataset—construct the contrast sets. This is best done by first identifying a list of phenomena that characterize their dataset. In syntactic parsing, for example, this list might include prepositional phrase attachment ambiguities, coordination scope, clausal attachment, etc. After the standard dataset collection process, the authors should sample pivots from their test set and perturb them according to the listed phenomena. ### 2.4 Design Choices of Contrast Sets Here, we discuss possible alternatives to our approach for constructing contrast sets and our reasons for choosing the process we did. ## 3 How to Create Contrast Sets Here, we walk through our process for creating contrast sets for three datasets. Examples are shown in Figure LABEL:fig:teaser and Table 1. #### DROP DROP Dua et al. (2019) is a reading comprehension dataset that is intended to cover compositional reasoning over numbers in a paragraph, including filtering, sorting, and counting sets, and doing numerical arithmetic. The data has three main sources of paragraphs, all from Wikipedia articles: descriptions of American football games, descriptions of census results, and summaries of wars. There are many common patterns used by the crowd workers that make some questions artificially easy: 2 is the most frequent answer to _How many…?_ questions, questions asking about the ordering of events typically follow the linear order of the paragraph, and a large fraction of the questions do not require compositional reasoning. Our strategy for constructing contrast sets for DROP was three-fold. First, we added more compositional reasoning steps. The questions about American football passages in the original data very often had multiple reasoning steps (e.g., _How many yards difference was there between the Broncos’ first touchdown and their last?_), but the questions about the other passage types did not. We drew from common patterns in the training data and added additional reasoning steps to questions in our contrast sets. Second, we inverted the semantics of various parts of the question. This includes perturbations such as changing _shortest_ to _longest_ , _later_ to _earlier_ , as well as changing questions asking for counts to questions asking for sets (_How many countries…_ to _Which countries…_). Finally, we changed the ordering of events. A large number of questions about war paragraphs ask which of two events happened first. We changed (1) the order the events were asked about in the question, (2) the order that the events showed up in the passage, and (3) the dates associated with each event to swap their temporal order. #### NLVR2 We next consider NLVR2, a dataset where a model is given a sentence about two provided images and must determine whether the sentence is true Suhr et al. (2019). The data collection process encouraged highly compositional language, which was intended to require understanding the relationships between objects, properties of objects, and counting. We constructed NLVR2 contrast sets by modifying the sentence or replacing one of the images with freely-licensed images from web searches. For example, we might change _The left image contains twice the number of dogs as the right image_ to _The left image contains three times the number of dogs as the right image_. Similarly, given an image pair with four dogs in the left and two dogs in the right, we can replace individual images with photos of variably-sized groups of dogs. The textual perturbations were often changes in quantifiers (e.g., _at least one_ to _exactly one_), entities (e.g., _dogs_ to _cats_), or properties thereof (e.g., _orange glass_ to _green glass_). An example contrast set for NLVR2 is shown in Figure LABEL:fig:teaser. #### UD Parsing Finally, we discuss dependency parsing in the universal dependencies (UD) formalism Nivre et al. (2016). We look at dependency parsing to show that contrast sets apply not only to modern “high-level” NLP tasks but also to longstanding linguistic analysis tasks. We first chose a specific type of attachment ambiguity to target: the classic problem of prepositional phrase (PP) attachment Collins and Brooks (1995), e.g. _We ate spaghetti with forks_ versus _We ate spaghetti with meatballs_. We use a subset of the English UD treebanks: GUM Zeldes (2017), the English portion of LinES Ahrenberg (2007), the English portion of ParTUT Sanguinetti and Bosco (2015), and the dependency-annotated English Web Treebank Silveira et al. (2014). We searched these treebanks for sentences that include a potentially structurally ambiguous attachment from the head of a PP to either a noun or a verb. We then perturbed these sentences by altering one of their noun phrases such that the semantics of the perturbed sentence required a different attachment for the PP. We then re-annotated these perturbed sentences to indicate the new attachment(s). Dataset | # Examples | # Sets | Model | Original Test | Contrast | Consistency ---|---|---|---|---|---|--- NLVR2 | 994 | 479 | LXMERT | 76.4 | 61.1 | (–15.3) | 30.1 IMDb | 488 | 488 | BERT | 93.8 | 84.2 | (–9.6) | 77.8 MATRES | 401 | 239 | CogCompTime2.0 | 73.2 | 63.3 | (–9.9) | 40.6 UD English | 150 | 150 | Biaffine + ELMo | 64.7 | 46.0 | (–18.7) | 17.3 PERSPECTRUM | 217 | 217 | RoBERTa | 90.3 | 85.7 | (–4.6) | 78.8 DROP | 947 | 623 | MTMSN | 79.9 | 54.2 | (–25.7) | 39.0 QUOREF | 700 | 415 | XLNet-QA | 70.5 | 55.4 | (–15.1) | 29.9 ROPES | 974 | 974 | RoBERTa | 47.7 | 32.5 | (–15.2) | 17.6 BoolQ | 339 | 70 | RoBERTa | 86.1 | 71.1 | (–15.0) | 59.0 MC-TACO | 646 | 646 | RoBERTa | 38.0 | 14.0 | (–24.0) | 8.0 Table 2: Models struggle on the contrast sets compared to the original test sets. For each dataset, we use a (sometimes near) state-of-the-art model and evaluate it on the “# Examples” examples in the contrast sets (_not_ including the original example). We report percentage accuracy for NLVR2, IMDb, PERSPECTRUM, MATRES, and BoolQ; F1 scores for DROP and Quoref; Exact Match (EM) scores for ROPES and MC-TACO; and unlabeled attachment score on modified attachments for the UD English dataset. We also report _contrast consistency_ : the percentage of the “# Sets” contrast sets for which a model’s predictions are correct for all examples in the set (_including_ the original example). More details on datasets, models, and metrics can be found in §A and §B. #### Summary While the overall process we recommend for constructing contrast sets is simple and unified, its actual instantiation varies for each dataset. Dataset authors should use their best judgment to select which phenomena they are most interested in studying and craft their contrast sets to explicitly test those phenomena. Care should be taken during contrast set construction to ensure that the phenomena present in contrast sets are similar to those present in the original test set; the purpose of a contrast set is not to introduce new challenges, but to more thoroughly evaluate the original intent of the test set. ## 4 Datasets and Experiments ### 4.1 Original Datasets We create contrast sets for 10 NLP datasets (full descriptions are provided in Section A): * • NLVR2 Suhr et al. (2019) * • IMDb sentiment analysis Maas et al. (2011) * • MATRES Temporal RE Ning et al. (2018) * • English UD parsing Nivre et al. (2016) * • PERSPECTRUM Chen et al. (2019) * • DROP Dua et al. (2019) * • Quoref Dasigi et al. (2019) * • ROPES Lin et al. (2019) * • BoolQ Clark et al. (2019) * • MC-TACO Zhou et al. (2019) We choose these datasets because they span a variety of tasks (e.g., reading comprehension, sentiment analysis, visual reasoning) and input-output formats (e.g., classification, span extraction, structured prediction). We include high-level tasks for which dataset artifacts are known to be prevalent, as well as longstanding formalism-based tasks, where data artifacts have been less of an issue (or at least have been less well-studied). ### 4.2 Contrast Set Construction The contrast sets were constructed by NLP researchers who were deeply familiar with the phenomena underlying the annotated dataset ## 5 Related Work The fundamental idea of finding or creating data that is “minimally different” has a very long history. In linguistics, for instance, the term _minimal pair_ is used to denote two words with different meaning that differ by a single sound change, thus demonstrating that the sound change is phonemic in that language Pike (1946). Many people have used this idea in NLP (see below), creating challenge sets or providing training data that is “minimally different” in some sense, and we continue this tradition. Our main contribution to this line of work, in addition to the resources that we have created, is giving a simple and intuitive geometric interpretation of “bias” in dataset collection, and showing that this long-standing idea of minimal data changes can be effectively used to solve this problem on a wide variety of NLP tasks. We additionally generalize the idea of a minimal _pair_ to a _set_ , and use a _consistency_ metric, which we contend more closely aligns with what NLP researchers mean by “language understanding”. #### Training on Perturbed Examples Many previous works have provided minimally contrastive examples on which to train models. Selsam et al. (2019), Tafjord et al. (2019), Lin et al. (2019), and Khashabi et al. (2020) designed their data collection process to include contrastive examples. Data augmentation methods have also been used to mitigate gender Zhao et al. (2018), racial Dixon et al. (2018), and other biases Kaushik et al. (2020) during training, or to introduce useful inductive biases Andreas (2020). #### Challenge Sets The idea of creating challenging contrastive evaluation sets has a long history Levesque et al. (2011); Ettinger et al. (2017); Glockner et al. (2018); Naik et al. (2018); Isabelle et al. (2017). Challenge sets exist for various phenomena, including ones with “minimal” edits similar to our contrast sets, e.g., in image captioning Shekhar et al. (2017), machine translation Sennrich (2017); Burlot and Yvon (2017); Burlot et al. (2018), and language modeling Marvin and Linzen (2018); Warstadt et al. (2019). Minimal pairs of edits that perturb gender or racial attributes are also useful for evaluating social biases Rudinger et al. (2018); Zhao et al. (2018); Lu et al. (2018). Our key contribution over this prior work is in grouping perturbed instances into a contrast set, for measuring local alignment of decision boundaries, along with our new, related resources. Additionally, rather than creating new data from scratch, contrast sets augment existing test examples to fill in systematic gaps. Thus contrast sets often require less effort to create, and they remain grounded in the original data distribution of some training set. Since the initial publication of this paper, Shmidman et al. have further demonstrated the utility of contrast sets by applying these ideas to the evaluation of morphological disambiguation in Hebrew. ## 6 Conclusion We presented a new annotation paradigm, based on long-standing ideas around contrastive examples, for constructing more rigorous test sets for NLP. Our procedure maintains most of the established processes for dataset creation but fills in some of the systematic gaps that are typically present in datasets. By shifting evaluations from accuracy on _i.i.d._ test sets to consistency on contrast sets, we can better examine whether models have learned the desired capabilities or simply captured the idiosyncrasies of a dataset. We created contrast sets for 10 NLP datasets and released this data as new evaluation benchmarks. We recommend that future data collection efforts create contrast sets to provide more comprehensive evaluations for both existing and new NLP datasets. While we have created thousands of new test examples across a wide variety of datasets, we have only taken small steps towards the rigorous evaluations we would like to see in NLP. The last several years have given us dramatic modeling advancements; our evaluation methodologies and datasets need to see similar improvements. ## Acknowledgements We thank the anonymous reviewers for their helpful feedback on this paper, as well as many others who gave constructive comments on a publicly-available preprint. Various authors of this paper were supported in part by ERC grant 677352, NSF grant 1562364, NSF grant IIS-1756023, NSF CAREER 1750499, ONR grant N00014-18-1-2826 and DARPA grant N66001-19-2-403. ## References * Ahrenberg (2007) Lars Ahrenberg. 2007. LinES: an English-Swedish parallel treebank. In _NODALIDA_. * Andreas (2020) Jacob Andreas. 2020. Good-enough compositional data augmentation. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_ , pages 7556–7566, Online. Association for Computational Linguistics. * Ben-David et al. (2010) Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. 2010. A theory of learning from different domains. _Machine Learning_. * Bowman et al. (2015) Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015\. A large annotated corpus for learning natural language inference. In _EMNLP_. * Burlot et al. (2018) Franck Burlot, Yves Scherrer, Vinit Ravishankar, Ondřej Bojar, Stig-Arne Grönroos, Maarit Koponen, Tommi Nieminen, and François Yvon. 2018. The WMT’18 morpheval test suites for English-Czech, English-German, English-Finnish and Turkish-English. In _Proceedings of the Third Conference on Machine Translation: Shared Task Papers_ , Belgium, Brussels. Association for Computational Linguistics. * Burlot and Yvon (2017) Franck Burlot and François Yvon. 2017. Evaluating the morphological competence of machine translation systems. In _Proceedings of the Second Conference on Machine Translation_ , pages 43–55, Copenhagen, Denmark. Association for Computational Linguistics. * Chen et al. (2016) Danqi Chen, Jason Bolton, and Christopher D. Manning. 2016. A thorough examination of the CNN/Daily Mail reading comprehension task. In _ACL_. * Chen et al. (2019) Sihao Chen, Daniel Khashabi, Wenpeng Yin, Chris Callison-Burch, and Dan Roth. 2019\. Seeing things from a different angle: Discovering diverse perspectives about claims. In _NAACL_. * Clark et al. (2019) Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. BoolQ: Exploring the surprising difficulty of natural Yes/No questions. In _NAACL_. * Collins and Brooks (1995) Michael Collins and James Brooks. 1995. Prepositional phrase attachment through a backed-off model. In _Third Workshop on Very Large Corpora_. * Dasigi et al. (2019) Pradeep Dasigi, Nelson F Liu, Ana Marasovic, Noah A Smith, and Matt Gardner. 2019\. Quoref: A reading comprehension dataset with questions requiring coreferential reasoning. In _EMNLP_. * Devlin et al. (2019) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In _NAACL_. * Dixon et al. (2018) Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2018. Measuring and mitigating unintended bias in text classification. In _ACM AIES_. * Dozat and Manning (2017) Timothy Dozat and Christopher D Manning. 2017. Deep biaffine attention for neural dependency parsing. In _ICLR_. * Dua et al. (2019) Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In _NAACL_. * Ettinger et al. (2017) Allyson Ettinger, Sudha Rao, Hal Daumé III, and Emily M. Bender. 2017. Towards linguistically generalizable NLP systems: A workshop and shared task. In _Proceedings of the First Workshop on Building Linguistically Generalizable NLP Systems_ , Copenhagen, Denmark. Association for Computational Linguistics. * Feng et al. (2019) Shi Feng, Eric Wallace, and Jordan Boyd-Graber. 2019. Misleading failures of partial-input baselines. In _ACL_. * Feng et al. (2018) Shi Feng, Eric Wallace, Alvin Grissom II, Mohit Iyyer, Pedro Rodriguez, and Jordan Boyd-Graber. 2018. Pathologies of neural models make interpretations difficult. In _EMNLP_. * Geva et al. (2019) Mor Geva, Yoav Goldberg, and Jonathan Berant. 2019. Are we modeling the task or the annotator? An investigation of annotator bias in natural language understanding datasets. In _EMNLP_. * Glockner et al. (2018) Max Glockner, Vered Shwartz, and Yoav Goldberg. 2018. Breaking NLI systems with sentences that require simple lexical inferences. In _ACL_. * Gururangan et al. (2018) Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R. Bowman, and Noah A. Smith. 2018. Annotation artifacts in natural language inference data. In _NAACL_. * Hu et al. (2019) Minghao Hu, Yuxing Peng, Zhen Huang, and Dongsheng Li. 2019. A multi-type multi-span network for reading comprehension that requires discrete reasoning. In _EMNLP_. * Isabelle et al. (2017) Pierre Isabelle, Colin Cherry, and George Foster. 2017. A challenge set approach to evaluating machine translation. In _EMNLP_. * Jia and Liang (2017) Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. In _EMNLP_. * Jia et al. (2019) Robin Jia, Aditi Raghunathan, Kerem Göksel, and Percy Liang. 2019. Certified robustness to adversarial word substitutions. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_ , Hong Kong, China. Association for Computational Linguistics. * Kaushik et al. (2020) Divyansh Kaushik, Eduard Hovy, and Zachary C Lipton. 2020. Learning the difference that makes a difference with counterfactually-augmented data. In _ICLR_. * Khashabi et al. (2020) Daniel Khashabi, Tushar Khot, and Ashish Sabhwawal. 2020. More bang for your buck: Natural perturbation for robust question answering. In _EMNLP_. * Lai et al. (2017) Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. RACE: Large-scale reading comprehension dataset from examinations. In _EMNLP_. * Levesque et al. (2011) Hector J. Levesque, Ernest Davis, and Leora Morgenstern. 2011. The winograd schema challenge. In _KR_. * Lin et al. (2019) Kevin Lin, Oyvind Tafjord, Peter Clark, and Matt Gardner. 2019. Reasoning over paragraph effects in situations. In _EMNLP MRQA Workshop_. * Lipton et al. (2018) Zachary Lipton, Yu-Xiang Wang, and Alexander Smola. 2018. Detecting and correcting for label shift with black box predictors. In _ICML_. * Liu et al. (2019) Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. _arXiv preprint arXiv:1907.11692_. * Lu et al. (2018) Kaiji Lu, Piotr Mardziel, Fangjing Wu, Preetam Amancharla, and Anupam Datta. 2018\. Gender bias in neural natural language processing. _arXiv preprint arXiv:1807.11714_. * Maas et al. (2011) Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In _ACL_. * Marcus et al. (1993) Mitchell Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of english: The Penn treebank. In _Computational Linguistics_. * Marvin and Linzen (2018) Rebecca Marvin and Tal Linzen. 2018. Targeted syntactic evaluation of language models. In _EMNLP_. * Michel et al. (2019) Paul Michel, Xian Li, Graham Neubig, and Juan Miguel Pino. 2019. On evaluation of adversarial perturbations for sequence-to-sequence models. In _NAACL_. * Min et al. (2019) Sewon Min, Eric Wallace, Sameer Singh, Matt Gardner, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2019. Compositional questions do not necessitate multi-hop reasoning. In _ACL_. * Naik et al. (2018) Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Rose, and Graham Neubig. 2018. Stress test evaluation for natural language inference. In _COLING_. * Nie et al. (2019) Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2019. Adversarial NLI: A new benchmark for natural language understanding. _arXiv preprint arXiv:1910.14599_. * Ning et al. (2019) Qiang Ning, Sanjay Subramanian, and Dan Roth. 2019. An Improved Neural Baseline for Temporal Relation Extraction. In _EMNLP_. * Ning et al. (2018) Qiang Ning, Hao Wu, and Dan Roth. 2018. A Multi-Axis Annotation Scheme for Event Temporal Relations. In _ACL_. * Niven and Kao (2019) Timothy Niven and Hung-Yu Kao. 2019. Probing neural network comprehension of natural language arguments. In _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics_ , Florence, Italy. Association for Computational Linguistics. * Nivre et al. (2016) Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Yoav Goldberg, Jan Hajič, Christopher D. Manning, Ryan McDonald, Slav Petrov, Sampo Pyysalo, Natalia Silveira, Reut Tsarfaty, and Daniel Zeman. 2016. Universal dependencies v1: A multilingual treebank collection. In _LREC_. * Peters et al. (2018) Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In _NAACL_. * Pike (1946) K.L. Pike. 1946. _Phonemics: A Technique for Reducing Languages to Writing_. Number v. 2 in Phonemics: A Technique for Reducing Languages to Writing. Summer Institute of Linguistics. * Poliak et al. (2018) Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. In _*SEM_. * Recht et al. (2019) Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. 2019. Do ImageNet classifiers generalize to ImageNet? In _ICML_. * Ribeiro et al. (2019) Marco Tulio Ribeiro, Carlos Guestrin, and Sameer Singh. 2019. Are red roses red? Evaluating consistency of question-answering models. In _ACL_. * Ribeiro et al. (2018a) Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018a. Semantically equivalent adversarial rules for debugging NLP models. In _Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , pages 856–865, Melbourne, Australia. Association for Computational Linguistics. * Ribeiro et al. (2018b) Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018b. Semantically equivalent adversarial rules for debugging NLP models. In _ACL_. * Rudinger et al. (2018) Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In _NAACL_. * Sanguinetti and Bosco (2015) Manuela Sanguinetti and Cristina Bosco. 2015. PartTUT: The Turin university parallel treebank. In _Harmonization and Development of Resources and Tools for Italian Natural Language Processing within the PARLI Project_. * Selsam et al. (2019) Daniel Selsam, Matthew Lamm, Benedikt Bünz, Percy Liang, Leonardo de Moura, and David L. Dill. 2019. Learning a SAT solver from single-bit supervision. In _International Conference on Learning Representations_. * Sennrich (2017) Rico Sennrich. 2017. How grammatical is character-level neural machine translation? Assessing MT quality with contrastive translation pairs. In _EACL_. * Shah et al. (2020) Deven Shah, H. Andrew Schwartz, and Dirk Hovy. 2020. Predictive biases in natural language processing models: A conceptual framework and overview. In _ACL_. * Shekhar et al. (2017) Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. 2017. Foil it! Find One mismatch between image and language caption. In _ACL_. * Shimodaira (2000) Hidetoshi Shimodaira. 2000. Improving predictive inference under covariate shift by weighting the log-likelihood function. In _Journal of Statistical Planning and Inference_. * (59) Avi Shmidman, Joshua Guedalia, Shaltiel Shmidman, Moshe Koppel, and Reut Tsarfaty. A novel challenge set for hebrew morphological disambiguation and diacritics restoration. In _Findings of EMNLP_. * Silveira et al. (2014) Natalia Silveira, Timothy Dozat, Marie-Catherine de Marneffe, Samuel Bowman, Miriam Connor, John Bauer, and Chris Manning. 2014. A gold standard dependency corpus for English. In _LREC_. * Suhr and Artzi (2019) Alane Suhr and Yoav Artzi. 2019. NLVR2 visual bias analysis. _arXiv preprint arXiv:1909.10411_. * Suhr et al. (2019) Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. 2019\. A corpus for reasoning about natural language grounded in photographs. In _ACL_. * Szegedy et al. (2014) Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. 2014. Intriguing properties of neural networks. In _ICLR_. * Tafjord et al. (2019) Oyvind Tafjord, Matt Gardner, Kevin Lin, and Peter Clark. 2019. QuaRTz: An open-domain dataset of qualitative relationship questions. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_ , pages 5941–5946, Hong Kong, China. Association for Computational Linguistics. * Tan and Bansal (2019) Hao Tan and Mohit Bansal. 2019. LXMERT: Learning cross-modality encoder representations from transformers. In _EMNLP_. * UzZaman et al. (2013) Naushad UzZaman, Hector Llorens, Leon Derczynski, James Allen, Marc Verhagen, and James Pustejovsky. 2013. SemEval-2013 Task 1: TempEval-3: Evaluating time expressions, events, and temporal relations. In _*SEM_. * Wallace et al. (2019a) Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019a. Universal adversarial triggers for attacking and analyzing NLP. In _EMNLP_. * Wallace et al. (2019b) Eric Wallace, Pedro Rodriguez, Shi Feng, Ikuya Yamada, and Jordan Boyd-Graber. 2019b. Trick me if you can: Human-in-the-loop generation of adversarial question answering examples. In _TACL_. * Warstadt et al. (2019) Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng-Fu Wang, and Samuel R. Bowman. 2019. BLiMP: A benchmark of linguistic minimal pairs for english. _arXiv preprint arXiv:1912.00582_. * Wu et al. (2019) Tongshuang Wu, Marco Tulio Ribeiro, Jeffrey Heer, and Daniel Weld. 2019. Errudite: Scalable, reproducible, and testable error analysis. In _ACL_. * Yadav and Bottou (2019) Chhavi Yadav and Léon Bottou. 2019. Cold case: The lost MNIST digits. In _NeurIPS_. * Yang et al. (2019) Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le. 2019. XLNet: Generalized autoregressive pretraining for language understanding. In _NeurIPS_. * Zeldes (2017) Amir Zeldes. 2017. The GUM corpus: Creating multilayer resources in the classroom. In _LREC_. * Zellers et al. (2018) Rowan Zellers, Yonatan Bisk, Roy Schwartz, and Yejin Choi. 2018. SWAG: A large-scale adversarial dataset for grounded commonsense inference. In _EMNLP_. * Zellers et al. (2019) Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In _ACL_. * Zhao et al. (2018) Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018\. Gender bias in coreference resolution: Evaluation and debiasing methods. In _NAACL_. * Zhou et al. (2019) Ben Zhou, Daniel Khashabi, Qiang Ning, and Dan Roth. 2019. “Going on a vacation” takes longer than “going for a walk”: A study of temporal commonsense understanding. In _EMNLP_. ## Appendix A Dataset Details Here, we provide details for the datasets that we build contrast sets for. #### Natural Language Visual Reasoning 2 (NLVR$2$) Given a natural language sentence about two photographs, the task is to determine if the sentence is true Suhr et al. (2019). The dataset has highly compositional language, e.g., _The left image contains twice the number of dogs as the right image, and at least two dogs in total are standing_. To succeed at NLVR2, a model is supposed to be able to detect and count objects, recognize spatial relationships, and understand the natural language that describes these phenomena. #### Internet Movie Database (IMDb) The task is to predict the sentiment (positive or negative) of a movie review Maas et al. (2011). We use the same set of reviews from Kaushik et al. (2020) in order to analyze the differences between crowd-edited reviews and expert-edited reviews. #### Temporal relation extraction (MATRES) The task is to determine what temporal relationship exists between two events, i.e., whether some event happened _before_ or _after_ another event Ning et al. (2018). MATRES has events and temporal relations labeled for approximately 300 news articles. The event annotations are taken from the data provided in the TempEval3 workshop UzZaman et al. (2013) and the temporal relations are re-annotated based on a multi-axis formalism. We assume that the events are given and only need to classify the relation label between them. #### English UD Parsing We use a combination of four English treebanks (GUM, EWT, LinES, ParTUT) in the Universal Dependencies parsing framework, covering a range of genres. We focus on the problem of prepositional phrase attachment: whether the head of a prepositional phrase attaches to a verb or to some other dependent of the verb. We manually selected a small set of sentences from these treebanks that had potentially ambiguous attachments. #### Reasoning about perspectives (PERSPECTRUM) Given a debate-worthy natural language claim, the task is to identify the set of relevant argumentative sentences that represent perspectives for/against the claim Chen et al. (2019). We focus on the stance prediction sub-task: a binary prediction of whether a relevant perspective is for/against the given claim. #### Discrete Reasoning Over Paragraphs (DROP) A reading comprehension dataset that requires numerical reasoning, e.g., adding, sorting, and counting numbers in paragraphs Dua et al. (2019). In order to compute the consistency metric for the span answers of DROP, we report the average number of contrast sets in which $F_{1}$ for all instances is above $0.8$. #### Quoref A reading comprehension task with span selection questions that require coreference resolution Dasigi et al. (2019). In this dataset, most questions can be localized to a single event in the passage, and reference an argument in that event that is typically a pronoun or other anaphoric reference. Correctly answering the question requires resolving the pronoun. We use the same definition for consistency for Quorefas we did for _DROP_. #### Reasoning Over Paragraph Effects in Situations (ROPES) A reading comprehension dataset that requires applying knowledge from a background passage to new situations Lin et al. (2019). This task has background paragraphs drawn mostly from science texts that describe causes and effects (e.g., that brightly colored flowers attract insects), and situations written by crowd workers that instantiate either the cause (e.g., bright colors) or the effect (e.g., attracting insects). Questions are written that query the application of the statements in the background paragraphs to the instantiated situation. Correctly answering the questions is intended to require understanding how free-form causal language can be understood and applied. We use the same consistency metric for ROPES as we did for DROP and Quoref. #### BoolQ A dataset of reading comprehension instances with Boolean (yes or no) answers Clark et al. (2019). These questions were obtained from organic Google search queries and paired with paragraphs from Wikipedia pages that are labeled as sufficient to deduce the answer. As the questions are drawn from a distribution of what people search for on the internet, there is no clear set of “intended phenomena” in this data; it is an eclectic mix of different kinds of questions. #### MC-TACO A dataset of reading comprehension questions about multiple temporal common- sense phenomena Zhou et al. (2019). Given a short paragraph (often a single sentence), a question, and a collection of candidate answers, the task is to determine which of the candidate answers are plausible. For example, the paragraph might describe a storm and the question might ask how long the storm lasted, with candidate answers ranging from seconds to weeks. This dataset is intended to test a system’s knowledge of typical event durations, orderings, and frequency. As the paragraph does not contain the information necessary to answer the question, this dataset is largely a test of background (common sense) knowledge. ## Appendix B Contrast Set Details ### B.1 NLVR2 #### Text Perturbation Strategies We use the following text perturbation strategies for NLVR2: * • Perturbing quantifiers, e.g., _There is at least one dog_ $\to$ _There is exactly one dog_. * • Perturbing numbers, e.g., _There is at least one dog_ $\to$ _There are at least two dogs_. * • Perturbing entities, e.g., _There is at least one dog_ $\to$ _There is at least one cat_. * • Perturbing properties of entities, e.g., _There is at least one yellow dog_ $\to$ _There is at least one green dog_. #### Image Perturbation Strategies For image perturbations, the annotators collected images that are perceptually and/or conceptually close to the hypothesized decision boundary, i.e., they represent a minimal change in some concrete aspect of the image. For example, for an image pair with 2 dogs on the left and 1 dog on the right and the sentence _There are more dogs on the left than the right_ , a reasonable image change would be to replace the right-hand image with an image of two dogs. #### Model We use LXMERT Tan and Bansal (2019) trained on the NLVR2 training dataset. #### Contrast Set Statistics Five annotators created 983 perturbed instances that form 479 contrast sets. Annotation took approximately thirty seconds per textual perturbation and two minutes per image perturbation. ### B.2 IMDb #### Perturbation Strategies We minimally perturb reviews to flip the label while ensuring that the review remains coherent and factually consistent. Here, we provide example revisions: Original (Negative): I had quite high hopes for this film, even though it got a bad review in the paper. I was extremely tolerant, and sat through the entire film. I felt quite sick by the end. New (Positive): I had quite high hopes for this film, even though it got a bad review in the paper. I was extremely amused, and sat through the entire film. I felt quite happy by the end. Original (Positive): This is the greatest film I saw in 2002, whereas I’m used to mainstream movies. It is rich and makes a beautiful artistic act from these 11 short films. From the technical info (the chosen directors), I feared it would have an anti-American basis, but … it’s a kind of (11 times) personal tribute. The weakest point comes from Y. Chahine : he does not manage to “swallow his pride” and considers this event as a well-merited punishment … It is really the weakest part of the movie, but this testifies of a real freedom of speech for the whole piece. New (Negative): This is the most horrendous film I saw in 2002, whereas I’m used to mainstream movies. It is low budgeted and makes a less than beautiful artistic act from these 11 short films. From the technical info (the chosen directors), I feared it would have an anti-American basis, but … it’s a kind of (11 times) the same. One of the weakest point comes from Y. Chahine : he does not manage to “swallow his pride” and considers this event as a well- merited punishment … It is not the weakest part of the movie, but this testifies of a real freedom of speech for the whole piece. #### Model We use the same BERT model setup and training data as Kaushik et al. (2020) which allows us to fairly compare the crowd and expert revisions. #### Contrast Set Statistics We use 100 reviews from the validation set and 488 from the test set of Kaushik et al. (2020). Three annotators used approximately 70 hours to construct and validate the dataset. ### B.3 MATRES MATRES has three sections: TimeBank, AQUAINT, and Platinum, with the Platinum section serving as the test set. We use 239 instances (30% of the dataset) from Platinum. #### Perturbation Strategies The annotators perturb one or more of the following aspects: appearance order in text, tense of verb(s), and temporal conjunction words. Below are example revisions: * • Colonel Collins followed a normal progression once she was picked as a NASA astronaut. (original sentence: “followed” is after “picked”) * • Once Colonel Collins was picked as a NASA astronaut, she followed a normal progression. (appearance order change in text; “followed” is still after “picked”) * • Colonel Collins followed a normal progression before she was picked as a NASA astronaut. (changed the temporal conjunction word from “once” to “before” and “followed” is now before “picked”) * • Volleyball is a popular sport in the area, and more than 200 people were watching the game, the chief said. (original sentence: “watching” is before “said”) * • Volleyball is a popular sport in the area, and more than 200 people would be watching the game, the chief said. (changed the verb tense: “watching” is after “said”) #### Model We use CogCompTime 2.0 Ning et al. (2019). #### Contrast Set Statistics Two annotators created 401 perturbed instances that form 239 contrast sets. The annotators used approximately 25 hours to construct and validate the dataset. #### Analysis We recorded the perturbation strategy used for each example. 49% of the perturbations changed the “appearance order”, 31% changed the “tense”, 24% changed the “temporal conjunction words”, and 10% had other changes. We double count the examples that have multiple perturbations. The model accuracy on the different perturbations is reported in the table below. Perturbation Type | Accuracy ---|--- Overall | 63.3% Appearance Order | 66.5% Tense Change | 61.8% Temporal Conjunction | 60.0% Other Changes | 61.8% Table 3: Accuracy breakdown of the perturbation types for MATRES. ### B.4 Syntactic Parsing #### Perturbation Strategies The annotators perturbed noun phrases adjacent to prepositions (leaving the preposition unchanged). For example, _The clerics demanded talks with local US commanders_ $\to$ _The clerics demanded talks with great urgency_. The different semantic content of the noun phrase changes the syntactic path from the preposition with to the parent word of the parent of the preposition; in the initial example, the parent is commanders and the grandparent is the noun talks; in the perturbed version, the grandparent is now the verb demanded. #### Model We use a biaffine parser following the architecture of Dozat and Manning (2017) with ELMo embeddings Peters et al. (2018), trained on the combination of the training sets for the treebanks that we drew test examples from (GUM, EWT, LinES, and ParTUT). #### Contrast Set Statistics One annotator created 150 perturbed examples that form 150 contrast sets. 75 of the contrast sets consist of a sentence in which a prepositional phrase attaches to a verb, paired with an altered version where it attaches to a noun instead. The other 75 sentences were altered in the opposite direction. #### Analysis The process of creating a perturbation for a syntactic parse is highly time- consuming. Only a small fraction of sentences in the test set could be altered in the desired way, even after filtering to find relevant syntactic structures and eliminate unambiguous prepositions (e.g. of always attaches to a noun modifying a noun, making it impossible to change the attachment without changing the preposition). Further, once a potentially ambiguous sentence was identified, annotators had to come up with an alternative noun phrase that sounded natural and did not require extensive changes to the structure of the sentence. They then had to re-annotate the relevant section of the sentence, which could include new POS tags, new UD word features, and new arc labels. On average, each perturbation took 10–15 minutes. Expanding the scope of this augmented dataset to cover other syntactic features, such as adjective scope, apposition versus conjunction, and other forms of clausal attachment, would allow for a significantly larger dataset but would require a large amount of annotator time. The very poor contrast consistency on our dataset (17.3%) suggests that this would be a worthwhile investment to create a more rigorous parsing evaluation. Notably, the model’s accuracy for predicting the target prepositions’ grandparents in the original, unaltered tree (64.7%) is significantly lower than the model’s accuracy for grandparents of all words (78.41%) and for grandparents of all prepositions (78.95%) in the original data. This indicates that these structures are already difficult for the parser due to structural ambiguity. ### B.5 PERSPECTRUM #### Perturbation Strategies The annotators perturbed examples in multiple steps. First, they created non- trivial negations of the claim, e.g., _Should we live in space?_ $\to$ _Should we drop the ambition to live in space?_. Next, they labeled the perturbed claim with respect to each perspective. For example: Claim: Should we live in space? Perspective: Humanity in many ways defines itself through exploration and space is the next logical frontier. Label: True Claim: Should we drop the ambition to live in space? Perspective: Humanity in many ways defines itself through exploration and space is the next logical frontier. Label: False #### Model We use a RoBERTa model Liu et al. (2019) finetuned on PERSPECTRUM following the training process from Chen et al. (2019). #### Contrast Set Statistics The annotators created 217 perturbed instances that form 217 contrast sets. Each example took approximately three minutes to annotate: one minute for an annotator to negate each claim and one minute each for two separate annotators to adjudicate stance labels for each contrastive claim-perspective pair. ### B.6 DROP #### Perturbation Strategies See Section 3 in the main text for details about our perturbation strategies. #### Model We use MTMSN Hu et al. (2019), a DROP question answering model that is built on top of BERT Large Devlin et al. (2019). #### Contrast Set Statistics The total size of the augmented test set is 947 examples and contains a total of 623 contrast sets. Three annotators used approximately 16 hours to construct and validate the dataset. #### Analysis We bucket $100$ of the perturbed instances into the three categories of perturbations described in Section 3. For each subset, we evaluate MTMSN’s performance and show the results in the Table below. Perturbation Type | Frequency | Accuracy ---|---|--- Adding Compositional Steps | 38% | 67.5 $F_{1}$ Inversion of Semantics | 37% | 53.2 $F_{1}$ Re-ordering Events | 25% | 47.3 $F_{1}$ Table 4: Accuracy breakdown of the perturbation types for DROP. ### B.7 Quoref #### Perturbation Strategies We use the following perturbation strategies for Quoref: * • Perturb questions whose answers are entities to instead make the answers a property of those entities, e.g., _Who hides their identity …_ $\to$ _What is the nationality of the person who hides their identity …_. * • Perturb questions to add compositionality, e.g., _What is the name of the person …_ $\to$ _What is the name of the father of the person …_. * • Add sentences between referring expressions and antecedents to the context paragraphs. * • Replace antecedents with less frequent named entities of the same type in the context paragraphs. #### Model We use XLNet-QA, the best model from Dasigi et al. (2019), which is a span extraction model built on top of XLNet Yang et al. (2019). #### Contrast Set Statistics Four annotators created 700 instances that form 415 contrast sets. The mean contrast set size (including the original example) is $2.7(\pm 1.2)$. The annotators used approximately 35 hours to construct and validate the dataset. ### B.8 ROPES #### Perturbation Strategies We use the following perturbation strategies for ROPES: * • Perturbing the background to have the opposite causes and effects or qualitative relation, e.g., _Gibberellins are hormones that cause the plant to grow_ $\to$ _Gibberellins are hormones that cause the plant to stop growing._ * • Perturbing the situation to associate different entities with different instantiations of a certain cause or effect. For example, _Grey tree frogs live in wooded areas and are difficult to see when on tree trunks. Green tree frogs live in wetlands with lots of grass and tall plants._ $\to$ _Grey tree frogs live in wetlands areas and are difficult to see when on stormy days in the plants. Green tree frogs live in wetlands with lots of leaves to hide on._ * • Perturbing the situation to have more complex reasoning steps, e.g., _Sue put 2 cubes of sugar into her tea. Ann decided to use granulated sugar and added the same amount of sugar to her tea._ $\to$ _Sue has 2 cubes of sugar but Ann has the same amount of granulated sugar. They exchange the sugar to each other and put the sugar to their ice tea._ * • Perturbing the questions to have presuppositions that match the situation and background. #### Model We use the best model from Lin et al. (2019), which is a span extraction model built on top of a RoBERTa model Liu et al. (2019) that is first finetuned on RACE Lai et al. (2017). #### Contrast Set Statistics Two annotators created 974 perturbed instances which form 974 contrast sets. The annotators used approximately 65 hours to construct and validate the dataset. ### B.9 BoolQ #### Perturbation Strategies We use a diverse set of perturbations, including adjective, entity, and event changes. We show three representative examples below: Paragraph: The Fate of the Furious premiered in Berlin on April 4, 2017, and was theatrically released in the United States on April 14, 2017, playing in 3D, IMAX 3D and 4DX internationally…A spinoff film starring Johnson and Statham’s characters is scheduled for release in August 2019, while the ninth and tenth films are scheduled for releases on the years 2020 and 2021. Question: Is “Fate and the Furious” the last movie? Answer: False New Question: Is “Fate and the Furious” the first of multiple movies? New Answer: True Perturbation Strategy: Adjective Change Paragraph: Sanders played football primarily at cornerback, but also as a kick returner, punt returner, and occasionally wide receiver…An outfielder in baseball, he played professionally for the New York Yankees, the Atlanta Braves, the Cincinnati Reds and the San Francisco Giants, and participated in the 1992 World Series with the Braves. Question: Did Deion Sanders ever win a world series? Answer: False New Question: Did Deion Sanders ever play in a world series? New Answer: True Perturbation strategy: Event Change Paragraph: The White House is the official residence and workplace of the President of the United States. It is located at 1600 Pennsylvania Avenue NW in Washington, D.C. and has been the residence of every U.S. President since John Adams in 1800. The term is often used as a metonym for the president and his advisers. Question: Does the president live in the White House? Answer: True New Question: Did George Washington live in the White House? New Answer: False Perturbation Strategy: Entity Change #### Model We use RoBERTa base and follow the standard finetuning process from Liu et al. (2019). #### Contrast Set Statistics The annotators created 339 perturbed questions generated that form 70 contrast sets. One annotator created the dataset and a separate annotator verified it. This entire process took approximately 16 hours. ### B.10 MC-TACO #### Perturbation Strategies The main goal when perturbing MC-TACO questions is to retain a similar question that requires the same temporal knowledge to answer, while there are additional constraints with slightly different related context that changes the answers. We also modified the answers accordingly to make sure the question has a combination of plausible and implausible candidates. #### Model We use the best baseline model from the original paper Zhou et al. (2019) which is based on $\textsc{RoBERTa}_{base}$ Liu et al. (2019). #### Contrast Set Statistics The annotators created 646 perturbed question-answer pairs that form 646 contrast sets. Two annotators used approximately 12 hours to construct and validate the dataset.
# Spectral analysis of the AMXP IGR J17591–2342 during its 2018 outburst A. Manca,1 A. F. Gambino,2 A. Sanna,1,3 G. K. Jaisawal,5 T. Di Salvo,2,3,4 R. Iaria,2 S. M. Mazzola,1 A. Marino,8,9 A. Anitra,2 E. Bozzo,7 A. Riggio,1,4 and L. Burderi,1,3,4 1Dipartimento di Fisica, Università degli Studi di Cagliari, SP Monserrato- Sestu, KM 0.7, Monserrato, 09042 Italy 2Università degli Studi di Palermo, Dipartimento di Fisica e Chimica - Emilio Segrè, via Archirafi 36 - 90123 Palermo, Italy 3INFN, Sezione di Cagliari, Cittadella Universitaria, 09042 Monserrato, CA, Italy 4INAF - Osservatorio Astronomico di Cagliari, via della Scienza 5, 09047 Selargius (CA), Italy 5National Space Institute, Technical University of Denmark, Elektrovej 327-328, 2800 Lyngby, Denmark 6Istituto Nazionale di Astrofisica, IASF Palermo, Via U. La Malfa 153, I-90146 Palermo, Italy 7ISDC, Department of Astronomy, University of Geneva, Chemin d’Écogia 16, 1290 Versoix, Switzerland 8Institute of Space Sciences (ICE, CSIC), Campus UAB, Carrer de Can Magrans s/n, E-08193 Barcelona, Spain 9Institut d’Estudis Espacials de Catalunya (IEEC), E-08034 Barcelona, Spain E-mail<EMAIL_ADDRESS> (Accepted XXX. Received YYY; in original form ZZZ) ###### Abstract The Accreting Millisecond X-ray Pulsar IGR J17591–2342 is a LMXB system that went in outburst on August 2018 and it was monitored by the NICER observatory and partially by other facilities. We aim to study how the spectral emission of this source evolved during the outburst, by exploiting the whole X-ray data repository of simultaneous observations. The continuum emission of the combined broad-band spectra is on average well described by an absorbed Comptonisation component scattering black-body-distributed photons peaking at (0.8$\pm$0.5) keV, by a moderately optically thick corona ($\tau$=2.3$\pm$0.5) with temperature of (34$\pm$9) keV. A black-body component with temperature and radial size of (0.8$\pm$0.2) keV and (3.3$\pm$1.5) km respectively is required by some of the spectra and suggests that part of the central emission, possibly a fraction of the neutron star surface, is not efficiently scattered by the corona. The continuum at low energies is characterised by significant residuals suggesting the presence of an absorption edge of O viii and of emission lines of Ne ix ions. Moreover, broad Fe i and Fe xxv K$\alpha$ emission lines are detected at different times of the outburst, suggesting the presence of reflection in the system. ###### keywords: X-rays: binaries – stars: neutron – stars:individual:IGR J17591–2342 – line: formation – line: profiles ††pubyear: 2015††pagerange: Spectral analysis of the AMXP IGR J17591–2342 during its 2018 outburst–A ## 1 Introduction Accreting Millisecond X-ray Pulsars (AMXPs) are Low Mass X-ray binary systems (LMXBs) hosting neutron stars (NS) that show coherent pulsations with periods lower than 10 ms. Neutron Stars in these systems are characterised by low magnetic fields, generally between $10^{8}$–$10^{9}$ G (see, e.g., Di Salvo & Sanna, 2020). Their relatively small spin periods are now established to be a direct consequence of the mass transfer occurring via Roche lobe overflow from a low mass (< 1 M⊙) companion star onto a slow rotating NS. This property makes AMXPs the progenitors of the rotation-powered millisecond pulsars emitting on a large fraction of the electomagnetic spectrum, i.e. from the radio to the gamma-ray band (Alpar et al., 1982). AMXPs usually show sporadic outbursts during which the X-ray luminosity can attain values between $10^{36}$ and $10^{37}$ erg/s. So far, 24 sources are included in this subclass of objects (see Bult et al., 2022; Ng et al., 2021). Even though the spectral evolution of these sources was rarely monitored in detail during the whole outburst, AMXPs are generally observed in hard spectral states with no hard to soft transitions. For this reason AMXPs are usually referred as hard X-ray transients (Di Salvo & Sanna, 2020). The AMXP system IGR J17591–2342 was discovered in outburst by the INTernational Gamma-Ray Astrophysics Laboratory (INTEGRAL) on August 2018 (Ducci et al., 2018). Nowak et al. (2019) inferred the celestial coordinates of the source from a pointed Chandra/HETG observation, i.e. 17h 59m 02.83s, -23∘ 43’ 10.2" (J2000) with an error box of 0.6" representing the 90% confidence level of Chandra positional accuracy. Observations performed by the Australia Telescope Compact Array (ATCA) detected the radio counterpart of IGR J17591–2342 providing an improvement on the previous coordinates that resulted to be 17h 59m 02.86s $\pm$ 0.04, -23∘ 43’ 08.3" $\pm$ 0.1" (J2000) (Russell et al., 2018). Using NICER and NuSTAR observations, the timing analysis performed by Sanna et al. (2018) allowed to constrain coherent pulsations at a period of 1.9 ms and an orbital period of 8.8 hrs from the Doppler modulation. Moreover, the X-ray data show a peculiar spin-down behaviour that is compatible with a magnetically threaded disc (Sanna et al., 2020). The results reported by Sanna et al. (2018) led to discuss a scenario in which the companion star is a late spectral type star having an age between 8–12 Gyrs and with a mass between 0.85–0.92 M⊙, assuming a mass of 1.4 M⊙ for the NS. This scenario is consistent with a value of the inclination angle between 28∘–30∘. Broad-band spectral analysis was performed on this source, but considering an average spectrum obtained by pointed observations at different times of the outburst. Sanna et al. (2018) analysed a combined averaged spectrum consisting in Swift, INTEGRAL/ISGRI and NuSTAR FPMA and FPMB data. They modelled the spectrum using an absorbed soft black-body plus a Comptonisation component. Their work reports a value of the hydrogen column density $N_{H}$ of (3.6$\pm$1.1)$\times 10^{22}$ cm-2, and a Comptonisation component that is characterised by a photon index $\Gamma$ of about 1.8, a temperature $kT_{e}$ of the corona of 22${}^{+4}_{-3}$ keV, and a temperature of the seed photons (assuming a black-body spectrum) of $kT_{seed}=0.79\pm 0.09$ keV. The soft black-body direct emission is characterised by a temperature that is compatible with that of the seed photons and in line with the emission from a region of a few kilometers. These authors also detected a weak Gaussian line centred at an energy compatible with the iron K$\alpha$ complex, and interpreted as a possible signature of disc reflection. Nowak et al. (2019) modelled a 1–9 keV Chandra-NICER spectrum of the source, taken at about 58353 MJD, using a model consisting in an absorbed black-body and Comptonisation component. They inferred a value of $N_{H}$ of (2.9$\pm$0.5)$\times 10^{22}$ cm-2 and a black-body temperature of 0.06 keV, by fixing the $kT{{}_{e}}$ value to that observed by Sanna et al. (2018). In addition, based on a Si xiii absorption line detected in the Chandra spectrum, they propose the presence of an outflow with a velocity of about 2800 km s-1 in the system. They also found evidences of possible Ca lines in the HETGS spectra and hypothesized that the NS could be formed via accretion induced collapse of a white dwarf in a rare, calcium-rich Type Ib supernova explosion. Kuiper et al. (2020) studied the broad-band 0.3–150 keV averaged spectrum of the source using data from NuSTAR FPMA and FPMB modules, from XMM-Newton RGS, Epic-pn and Epic-MOS2 instruments and from INTEGRAL/ISGRI. They modelled the spectrum with an absorbed Comptonisation component finding a value of $N_{H}$ of (2.09$\pm$0.05)$\times 10^{22}$ cm-2 and a Comptonisation component in which the temperature of the seed photons is about 0.64 keV, the coronal temperature $kT_{e}$ is (38.8$\pm$1.2) keV and in which the optical depth of the corona is $\tau=1.59\pm 0.04$. They excluded the presence of a local absorber into the system, due to the fact that their estimation of $N_{H}$ results in line with the one expected according to the total Galactic absorption (i.e. 2.2$\times 10^{22}$ cm-2) resulting from the optical reddening maps (Russell et al., 2018). They also found evidences of an emission line in the Iron K-$\alpha$ region. However, they considered the detection of this line spurious and related to blending of lines from different Fe ionisation stages, or alternatively to uncertainties in the XMM Epic-pn response (Kuiper et al., 2020). In the same work, these authors estimated an upper limit on the distance to the source of $d=(7.6\pm 0.7)$ kpc according to the analysis of a burst showing clues of Photospheric Radius Expansion (PRE) in one INTEGRAL/JEM-X observation. In this work we perform a spectral study of IGR J17591–2342 based on the entire NICER data set and including a large sample of the available observations in the X-ray archive, with the aim of studying its spectral evolution in detail during the whole outburst. The paper is structured as follows: in section 2 we describe the data selection and reduction, in section 3 we report the data analysis, in section 4 we discuss the obtained results, and in section 5 we summarize these results advancing some conclusions. ## 2 Observations and data reduction Figure 1: NICER light curve of IGR J17591–2342 during the outburst of 2018. Superimposed we show the mid-observation times at which other observatories pointed the source and the shaded areas representing the time windows of the INTEGRAL observations. Figure 2: The 3–6 keV (upper panel) and 6–10 keV (middle panel) NICER light curve of IGR J17591–2342 during the outburst of 2018. In the bottom panel we report the hardness ratio evaluated between these two energy bands. A complete follow up of the outburst of IGR J17591–2342 was performed by the X-ray timing instrument (XTI, Gendreau et al. 2012; Gendreau et al. 2016) on board the Neutron star Interior Composition Explorer (NICER) mission, hosted onto the International Space Station (ISS). This instrument collected data of the source between 2018-08-14T23:59:42 (ObsID 1200310101) and 2018-10-17T05:32:22 (ObsID 1200310139). We reduced the NICER data using the nicerl2111https://heasarc.gsfc.nasa.gov/docs/nicer/analysis_threads/nicerl2/ script available under HEAsoft version 6.28. We processed the data using the latest gain and calibration database files of version 20200722. We applied standard filtering criteria based on elevation angle from the Earth limb, pointing offset, offset from the bright Earth, and South Atlantic Anomaly region in our study. Following the above criteria, we selected the Good Time Intervals using the nimaketime task. We extracted the source spectrum using the filtered clean event file in the xselect environment, while we obtained the background spectra corresponding to each observation using the nibackgen3C50222https://heasarc.gsfc.nasa.gov/docs/nicer/toolsnicer_bkg_est_tools.html tool (Remillard et al., 2021). For the spectral analysis, we considered the response matrix and ancillary response file of version 20200722. We added a systematic error of 1% to the obtained source spectra, as suggested by the NICER team (see, e.g., Jaisawal et al., 2019). From Figure 1, it is possible to follow the complete light curve of the source during the outburst as observed by NICER in the band 0.2–10 keV. The first observation of the source shows an average count-rate of about 17 c/s that increased up to a first peak of 29 c/s at 58362 MJD. After this first maximum, the count-rate decreased until it reached a value of about 17 c/s at 58367 MJD, when it started to increase again up to the absolute maximum of about 42 c/s reached at 58380 MJD. Later on, the source started to decrease its luminosity down to a count-rate of 14 c/s, reached on 58397 MJD. In addition, as highlighted by the hardness ratio reported in Figure 2 and evaluated between the energy band 3–6 keV and 6–10 keV, the spectral state of IGR J17591–2342 maintains quite constant, suggesting that no significant spectral changes occurred during the outburst, in agreement with what usually observed for AMXPs (Di Salvo & Sanna, 2020). The outburst of IGR J17591–2342 was also observed by INTEGRAL between revolutions 1986 and 2009. In addition, further single pointed observation of the Nuclear Spectroscopic Telescope Array (NuSTAR), Chandra, XMM-Newton and ASTROSAT were performed during the first part of the outburst. NuSTAR (Harrison et al., 2013) observed IGR J17591–2342 on 2018-08-13T22:36:09 (ObsID 90401331002) and on 2018-08-17T20:01:09 (ObsID 80301311002). The data have been reprocessed with the nupipeline routine while the source and background spectra were extrapolated with the nuproducts pipeline, using circular extraction regions of 60" of radius centred at the coordinates of the source and far away from the source, respectively. The spectra were modelled in the 2–80 keV energy band. Chandra observed IGR J17591–2342 on 2018-08-23T17:40:05 for 20 ks using the High Energy Transmission Grating spectrometer (HETG, Canizares et al., 2005). The events were reprocessed using the official Chandra analysis environment CIAO v. 4.13 with calibration files updated to the version 4.9.4. The data were reprocessed with the chandra_repro pipeline and the first order High- Energy Grating (HEG) and Medium-Energy Grating (MEG) spectra were combined using the combine_spectra tool. The obtained spectra were grouped to have at least 25 counts per energy bin and analysed in the 1.2–7.8 keV energy band. The XMM-Newton observatory observed the source with both the Reflection Grating Spectrometer (RGS, Den Herder et al., 2001) and the European Imaging cameras (EPIC) on 2018-09-03T18:44:55 during a time window in which the count- rate of the source momentarily decreased, as it can be inferred from Figure 1. The science data reduction was performed using the XMM-Newton Science Analysis System (SAS) version 19.0.0. The (imaging) EPIC-pn instrument (Strüder et al., 2001) operated in Timing Mode, while the MOS-1 and MOS-2 cameras (Turner et al., 2001), acquired data in Small window and Timing uncompressed modes, respectively. The Epic-pn spectra were extracted from the RAWX coordinates between 31–44, while the background spectrum was obtained from RAWX coordinates centred between 5 and 15. The MOS 1 data were collected in Small window mode and, as already noticed by Kuiper et al. (2020), are extremely corrupted by pile-up effects and for this reason they are not considered in our work. On the contrary, MOS 2 data are not affected by pile-up, since the instrument was operating in Timing uncompressed mode, which allows to collect data up to 35 mCrab without severe pile-up distortions. The source and background spectra for the MOS2 were extrapolated in the RAWX range 290–320, while the background spectrum was extracted by a box having a width and a height of 8432.64 and 4285.44 in physical coordinates, respectively. The analysed energy range is 2.2–10 keV. The RGS operated in the standard Spectroscopy HighEventRate mode with Single Event Selection. No soft proton background flares were detected during the observation. Then, we combined the first order data of RGS1 and RGS2 using the rgscombine tool included in the SAS package, and grouping the obtained spectrum to have at least 25 counts per energy bin. However, the data counts are considerably low for these data for energies below 1.2 keV, and for this reason we preferred not to consider these data for our analysis, since the energy range above 1 keV is well covered by the rest of the considered space missions. The ASTROSAT mission observed the source with the Large Area X-ray Proportional Counter (LAXPC) instrument (Yadav et al., 2016; Antia et al., 2017; Agrawal et al., 2017) on 2018-08-23T01:10:15 (ObsID 9000002320) and on 2018-08-27T00:00:00 (ObsID 9000002332) for a net exposure of 30 ks and 37.7 ks, respectively. A thermonuclear type I X-ray burst occurred during the ObsID 9000002320 and has been removed before extracting the persistent spectrum of the source. For both the observations, the LAXPC 30 was not working. On the other hand, the LAXPC 10 was active on a low voltage gain setting333http://astrosat-ssc.iucaa.in/. Since the source is faint, we extracted spectra from the top layer of each detector to avoid unnecessary background. However, the LAXPC 10 spectra appear to be affected by the background and gain change, and for this reason they were not included in the analysis. In addition, the LAXPC 20 spectrum extracted for ObsID 9000002320 shows a significant mismatch with respect to the spectra of the closest available observations from other missions, possibly introduced by a bad calibration. For this reason, this observation was also excluded from the analysis. A systematic error of 2% has been added to the LAXPC 20 spectrum of ObsID 9000002332 (see, e.g., Misra et al., 2017), which was also grouped to have at least 25 counts per energy bin. The analysis was conducted in the 4–17 keV energy range. INTEGRAL observed the outburst from the source during the period spanning from 58340 MJD to 58405 MJD, corresponding to satellite revolutions from 1986 to 2016. We considered data from the two JEM-X units (Lund et al., 2003) and from IBIS/ISGRI (Ubertini et al., 2003; Laurent et al., 2003). We analysed the ISGRI, JEM-X 1 and JEM-X2 data with the Offline Scientific Analysis (OSA) software version 11.0 distributed by the ISDC (Courvoisier et al., 2003). Different source spectra were initially extracted for the two JEM-X and IBIS/ISGRI from each revolution. A grouping of 16 bins has been used for the JEM-X (IBIS/ISGRI) spectra extraction, following the standard practice for similar sources. We excluded the science window (SCW) 50 in revolution 2001 due to the presence of a type I X-ray burst (see also Kuiper et al., 2020). The JEM-X spectra were analysed in the 3–20 keV energy range, while the IBIS/ISGRI spectra were analysed in the 25–200 keV range. Figure 3: Evolution of the main spectral parameters describing the continuum emission of IGR J17591–2342 as a function of time. The associated errors are reported at a level of statistical confidence of 90%. In the upper panel the mean count-rate collected during each NICER observation is reported. In each panel, the star-shaped points represent fixed values for the parameters. Figure 4: Evolution of the iron line spectral parameters and statistical significance of the observed feature. In the plot of the line centroids (second plot from the top) we also reported the rest-frame energy of the Fe i K$\alpha 1$ , Fe xxv (resonance), and Fe xxvi Ly$\alpha 1$ transitions, as reference. In the bottom panel, we highlighted in pink the detection area considering as lower limit $\sigma$=2.5 for a weak detection. The dotted line represents the 3$\sigma$ detection threshold. Figure 5: Best fit model and associated residuals obtained for the broad-band spectrum associated to the NICER ObsID 1200310106 (see Table 2 and Table 3). In black the NICER spectrum, in red, green and blue the INTEGRAL JEM-X1, JEM-X2 and ISGRI spectra of revolution 1992 respectively, and in cyan the ASTROSAT LAXPC20 spectrum of ObsID 9000002332. ## 3 Spectral Analysis The spectral analysis was entirely performed using the X-ray spectral fitting package Xspec v. 12.11.1c (Arnaud, 1996). We chose to fit as many spectra as possible in order to follow the evolution of the spectral parameters during the outburst with the highest possible temporal resolution guaranteed by the available data. For this reason, we paired each NICER spectrum with any quasi- simultaneaous, i.e. taken within two days, data-set from the other observatories, when possible. As widely discussed in Sanna et al. (2018) and Kuiper et al. (2020), after some preliminary tests on the available data, we noted that the model which best describes the continuum was an absorbed Comptonisation model. For this reason, we adopted this model in the following analysis. Some of the performed fits, however, revealed that in some cases a soft excess remained in the residuals and this could be corrected with the addition of a black-body emission from the source, as already reported in Sanna et al. (2018). We used the chemical abundances of Wilms et al. (2000a) and the cross sections reported in Verner et al. (1996). We modelled the continuum spectral emission of IGR J17591–2342 adopting the Tuebingen-Boulder ISM absorption model Tbabs, a black-body component bbodyrad, and the thermal Comptonisation component nthcomp. The nthcomp model is described by the asymptotic power-law photon index $\Gamma$, the electron temperature $kT_{e}$ of the hot electron corona, the seed photon temperature $kT_{seed}$, the $inp\\_type$ parameter, which can assume values 0 or 1 for considering black- body or disc-black-body distributions for the seed photons, respectively, and the $redshift$ parameter. We assumed that the seed photons are distributed accordingly to a black-body law by fixing the $inp\\_type$ parameter to 0. Moreover, we adopted a value of redshift equal to zero for all the subsequent analysis. The NICER spectra were fitted in the energy range 0.6–10 keV. On the other hand, the NuSTAR FPMA and FPMB spectra of ObsID 80301311002 were fitted in the range 2–60 keV, while those of ObsID 90401331002 in the range 2–80 keV. The XMM-Newton spectra of ObsID 0795750101 were fitted in the range 3–10 keV and 2.2–10 keV for the EPIC-pn and EPIC-MOS2 spectra, respectively. The Chandra MEG and HEG spectra of ObsID 20173 were fitted in the range 1.2–6.8 keV and 1.4–7.8 keV, respectively, while the ASTROSAT/LAXPC 20 spectrum of ObsID 9000002332 was fitted in the energy range 4–17 keV. We inspected with xspec all INTEGRAL data of the source, but for our broad- band analysis we finally only made use of the spectra extracted during revolutions from 1992 and 2006 because they are more reasonably close in time to the available NICER data. During revolutions from 1986 to 1989, the source was caught by INTEGRAL during the earliest stages of the outburst evolution and the INTEGRAL data were characterised by a number of counts too low to perform any meaningful spectral analysis. A similar conclusion applies to the data collected during revolutions 2008–2016, corresponding to the rapid decay of the source flux toward the end of the outburst. Depending on the source flux, we limited our spectral analysis for the JEM-X data roughly in the interval 3–20 keV. During revolutions 1989, 1994, 1995, 2002, 2004 and 2006, the JEM-X data did not have a large number of counts, in order to be used in the spectral analysis and thus we do not mention these data any longer in the following sections. IBIS/ISGRI data were used, following the OSA 11.0 recommendations, from an energy of 25 keV up to roughly 200 keV, above which the source fell below the detection threshold of the instrument. The higher energy bound of the IBIS/ISGRI spectrum varies across the different revolutions considered depending on the flux of the source but remains in all cases well above 100 keV. The NICER spectra, in combination with those of other missions, generally ensured a good energy coverage over a large energy range. However, some of the NICER spectra could not be combined with other spectra, keeping hard to constrain the high-energy cut-off $kT_{e}$ for the Comptonisation component. In all these cases, this parameter was fixed to the value obtained by Kuiper et al. (2020) (i.e. $kT_{e}=38.8\;keV$). In Figure 3 we show the evolution of the spectral parameters obtained from the best fit model of each observation as a function of time. In all the plots, the star-shaped points indicate parameters that were frozen in the fit. This plot shows how the $kT_{e}$ parameter has been constrained for many observations, with an average value obtained during the whole outburst of about 34 keV, which is in line with that inferred by Kuiper et al. (2020). Some of the NICER spectra showed the presence of residuals in absorption at about 0.87 keV, which are consistent with the presence of an O viii absorption edge. To fit the low energy residuals, as a first step, we tried to replace the Tbabs ISM absorption component with the Tbfeo and the Tbvarabs components that take into account variable abundances for the chemical species in the ISM. However, even if both these models return perfectly compatible values for the $N_{H}$, they are unable to fit the observed residuals. For this reason we manually described the observed feature by recurring to the edge component in Xspec. The energy of the edge was kept frozen for all the fits. Other localised residuals characterise some of the NICER spectra at about 0.922 keV, in line with the presence of Ne ix ions. This feature was modelled with a Gaussian component in which the energy centroid was fixed to the rest-frame energy of the aforementioned ion. Moreover, the spectral resolution of the NICER/XTI instrument around 1 keV was not sufficient to constrain the width of the modelled line. For this reason, we fixed the parameter $sigma$ to 0.085 keV (i.e. the NICER/XTI spectral resolution at 1 keV). On the other hand, the normalisation of the Gaussian component was left free to vary. Firstly, the neutral column density NH in the Tbabs component was left free to vary during the whole duration of the outburst, but we noticed a correlation with the kTseed parameter of the Comptonisation component, with a consequent scattering of the NH component. Therefore, we decided to keep this parameter frozen at the value of (2.09$\pm$0.05)$\times 10^{22}$ cm-2, in accordance with Kuiper et al. (2020) for the rest of the analysis. The Comptonisation component is on average characterised by stable values of $\Gamma$ and $kT_{e}$, equal to 1.9, with a standard deviation of about 0.2, and 34 keV with a standard deviation of about 9 keV, respectively. On the other hand, during the outburst the temperature of the seed photons, assumed to be injected with a black-body-like spectrum, shows a variation in accordance with the advance of the outburst. The mean value of this parameter during the outburst was of 0.8 keV with an associated standard deviation of 0.5 keV. The bbodyrad component appears to be required by the fit only in some of the observations. We plotted the resulting black-body temperature $kT$ in the fifth panel from the top of Figure 3. This parameter shows an almost constant trend during the outburst with a mean value of 0.8 keV and a standard deviation of about 0.2 keV. The associated errors at 90% of confidence level strongly suffer from the statistics of the fitted spectra. However, for all the reported black-body components the significance of its detection results to be higher than 3$\sigma$ (the magenta points have a significance of more than 5$\sigma$, whilst the dark violet ones have a significance between 3$\sigma$ and 4$\sigma$). Some of the observed combined spectra also showed significant residuals in the energy range at which the Fe xxv line is expected, i.e at about 6.5 keV. Once the residuals have been detected, we tried to model them with a Gaussian component. For each of these observations we evaluated the significance of this local feature, by evaluating how much the normalization of this component deviates from the continuum in units of $\sigma$. In the lower panel of Figure 4, we report the significance of the observed iron line. In the upper and medium panels of the same figure, we show how the parameters of this feature evolve in time during the outburst. In particular, we can observe that, with the exception of the second point for which we have an upper limit on the energy centroid, the rest of the lines have been well constrained, even though it was not possible to obtain lower limits for the associated width. However, these observations suggest that the line is marginally detected and that results to be in agreement with a Fe i K$\alpha 1$ line the first phase of the burst. After this phase, the line starts to be consistent with the presence of more ionised species as Fe xxv. Since this line tends to be generally broad, when its width could be constrained, we tried to fit this line also by using the Relativistic model discline (Fabian et al., 1989) or the most recent shaddisc (La Placa et al., 2020). These models, however, with the exception of the centroid energy, returned totally unconstrained spectral parameters and generally not physically reasonable due to the low number of counts on the line profile. As an example, we show the best fit model obtained for the broad-band spectrum associated to the NICER ObsID 1200310106 in Figure 5. ## 4 Discussion ### 4.1 Spectral evolution We report hereon the spectral analysis of the whole set of NICER observations of the outburst of IGR J17591–2342 occurred in 2018, integrating these observations with all the available quasi-simultaneous observations stored in the X-ray data archive. The NICER light curve during the outburst is characterised by a double peak. The spectral model adopted to fit the spectra of the source suggests that during the first peak, occurring at 58362 MJD, the source showed an unabsorbed bolometric flux of (4.6$\pm$0.2)$\times 10^{-10}$ erg cm-2 s-1 in the range 0.1–100 keV, which increases up to 7.2${}^{+0.7}_{-0.5}\times 10^{-10}$ erg cm-2 s-1 in the same energy band at 58380 MJD, when the second peak of the outburst occurred. The flux errors are reported at 3$\sigma$ c.l., as derived from the cflux convolution model. Comptonisation is the physical process for which we observe the major contribution in terms of flux. Actually, this component, on average, contributes to the total unabsorbed flux for 95% in the case of IGR J17591–2342. According to the obtained results, the Comptonisation component appears to be characterised by an electron corona with a temperature $kT_{e}$ of about 34 keV, which is in line with the value of 38.8 keV reported by Kuiper et al. (2020). Unfortunately, due to the lack of coverage at the higher energies in some points of the outburst, we were unable to constrain this parameter for all the observations performed by NICER. However, we constrained this parameter in 29% of the NICER observations, inferring that at the beginning of the outburst the cloud was characterised by a slightly lower temperature of about 24 keV. The photons that are Comptonised in this corona appear to be mainly with a temperature of about 0.8 keV, with the tendency to slightly increase during the peaks of the outburst, always, however, at the limit of statistical compatibility at the 90% confidence level. This could be due to episodes of increased heating of the neutron star surface/boundary layer, plausibly caused by occasional rises in the mass-accretion rate. The asymptotic power law index $\Gamma$, on the contrary, remains almost stable between 1.7 and 2.1. The black-body component has been tested for each spectrum and only 26% of the best fit models require this component at a level of confidence that is greater than 3$\sigma$. The equivalent radius of emission of the black-body component has been obtained from the normalization as Rbb=$N_{bb}d^{2}_{10}$, where Nbb is the normalization of the bbodyrad component and $d^{2}_{10}$ is the distance to the source in units of 10 kpc. Assuming the distance to the source obtained in this work (i.e., (7.2$\pm$0.8) kpc), we obtained the radii reported in Figure 6. Figure 6: Trend of the equivalent radius of emission of the black-body component (upper panel) and of the Comptonisation seed photons (middle panel) as a function of time. In the lower panel, the trend followed by the optical depth of the Comptonising electron cloud is also reported. The arrows indicate upper limits on the values of the parameters. From the observed trend, it is possible to notice how the radius of the emitting black-body is mainly at a value of about $\overline{R_{bb}}$ = (3.3$\pm$1.5) km, where the associated error in this case coincides with the standard deviation of the distribution of measurements. This component, peaking mainly at $\overline{kT}$ = (0.8$\pm$0.2) keV, could arise from a direct emission by a fraction of the NS surface. It is interesting to notice that the size of this region varies in accordance with the variation of the flux extrapolated in the band 0.1–100 keV during the outburst. In particular, for lower values of the flux, it appears smaller with respect to the case in which the flux is higher. One possible explanation could be represented by the fact that when the mass accretion rate increases, the region responsible for the emission becomes larger, reaching a maximum in proximity of the two peaks of the outburst, where the values are consistent with a radius of about 5 km, i.e half of the NS size, assuming a NS of 10 km of radius. Then, we tried to test if the direct observation of the black-body emission from the NS might be related to a low optical depth of the hot corona, using the relation of Zdziarski et al. (1996): $\Gamma=\left[\dfrac{9}{4}+\dfrac{1}{\tau\left(1+\dfrac{\tau}{3}\right)\left(\dfrac{kT_{e}}{m_{e}c^{2}}\right)}\right]^{1/2}-\dfrac{1}{2},$ (1) with the spectral parameters obtained in the best fit model for each observation. The evolution of this parameter as a function of the observing time is reported in Figure 6 in the lower panel. As shown by this trend, the value of the optical depth remains almost stable during the outburst at a value of about $\tau\sim$2.3 and a standard deviation of 0.5, at least for the parameters that were constrained. For this reason, it is possible that the corona Comptonises the majority of the photons emitted by the NS surface leaving, however, only a small fraction ($\sim$ 10%) of them not significantly Comptonised. The latter could contribute to the observed direct black-body component, peaking approximately at the same temperature of the seed photons, as evidenced in the obtained results. A direct proof of this scenario could be provided by estimating the equivalent radius of the region emitting the seed photons for the Comptonisation, which we expect more or less similar to the radius obtained for the direct black- body emission region. An estimation of this physical parameter can be obtained by using the relation of In ’t Zand et al. (1999), assuming a spherical geometry of the corona, $R_{0}=3\times 10^{4}\;d\left(\dfrac{f_{bol}}{1+y}\right)^{1/2}\left(kT_{seed}\right)^{-2},$ (2) where $d$ is the distance to the source in kpc inferred in this work, $f_{bol}$ is the unabsorbed bolometric flux extrapolated from the Comptonisation component in erg cm-2 s-1, $kT_{seed}$ is the temperature of the seed photons in keV, and $y=4kT_{e}max[\tau,\tau^{2}]/(m_{e}c^{2})$ is the Compton parameter, in which $kT_{e}$ is the electron temperature in keV. The radius $R_{0}$ obtained for each observation is plotted as a function of the observing time in the middle panel of Figure 6. These parameters, at least in the cases in which the statistics of the data allowed to constrain them, are scattered around a mean value of $\overline{R_{0}}$ = (5.0$\pm$2.8) km, which is in agreement with the obtained value for $\overline{R_{bb}}$. ### 4.2 Spectral lines The spectral continuum of the source is characterised by the presence of several local features. On the one hand, almost the totality of the analysed spectra present an absorption edge at 0.871 keV, which suggests the presence of O viii ions. Despite the fact that the energy of this feature needed to be fixed due to the complexity of the residuals at the lowest energies in the NICER spectra, the significance of this feature is always higher than 3$\sigma$ for all the results reported in Table 2. The low energy residuals of some spectra are well fitted by taking into account a Ne ix emission line for all the observations (see Table 3). In accordance with the statistics of the available spectra and with the spectral resolution of NICER/XTI, which is not sufficient to constrain the line width, we fixed the parameter $sigma$ of the Gaussian component to a value of 85 eV, which is the spectral resolution of the instrument at 1 keV. Trying to leave this parameter free to vary during the fit, indeed results in a value of width that is of the order of the spectral resolution of NICER at 1 keV, or slightly higher, even though not constrained. This could probably suggest two possible scenarios: on one hand, the lines could be produced approximately in the same region of an accretion disc where they are relativistically broadened due to the proximity to the NS. This scenario, however, is not supported by evidences of any further black- body or multicolour-disc black-body component needed to fit the spectrum and attributable to emission from the disc. On the other hand, this line could be smeared only as a consequence of Compton broadening, or as an effect of the not sufficient spectral resolution of the XTI instrument to resolve a complex of lines at those energies. The test of this latter scenario deserves observations with high spectral resolution instruments. Indeed, the low energy spectrum of the source is completely dominated by the background flux below 1 keV in the case of the pointed observations performed by the Chandra/HETG and the two MOS instruments on board the XMM-Newton mission. About 19% of the observations show a significant evidence ($>3\sigma$) of Fe emission lines. In Figure 4, we show the evolution followed by the parameters describing the line profile, assumed to be Gaussian. The detection of this line is significant only for the time window including the range spanned by the first peak of the outburst and by part of the rise of flux towards the main peak of the burst, i.e during the periods of higher statistics of the NICER spectra. Our analysis shows how, in the reported cases, it is possible to constrain the centroid energy of the lines, with the exception of the observations occurred on 58357 MJD and on 58395 MJD, for which we find only an upper limit of 6.3 and 6.7 keV, respectively. The remaining sample of line profiles result to be in line with the rest-frame energy corresponding to the Fe xxv K$\alpha$ line, usually observed in bright LMXB systems showing evidences of reflection. The observed sigma for these lines results to be on average of about 0.5 keV and its distribution shows a standard deviation of about 0.5 keV. On the basis of this result, it seems that the line profile is tendentially broad. Kuiper et al. (2020) detected a similar line profile in the average spectrum of the source. The feature they observed is centred at about 6.7 keV and has a width of 0.69 keV, in line with our results. However, they consider this feature to be characterised by values of the centroid energy and sigma that are too far off and too broad. They propose that this feature could be produced by a blend of Fe lines deriving from different ionisation stages, which can not be distinguished individually by the MOS due to its spectral resolution. Moreover, they propose that this detection could reflect uncertainties in the XMM-Newton EPIC-pn response for observations taken in timing mode, also because there is no detection for Chandra/HETG data, having a higher spectral resolution. For this reason they do not consider the detection as real. On the contrary, we detected this line at different times during the outburst, including the case occurring on 58353 MJD for which we observe a detection at about 3$\sigma$, and that was obtained by considering a combined spectrum that also includes a Chandra/HETG spectrum. In accordance with this evidence, and to the fact that also other spectra during the outburst show the same feature, we conclude that this feature is real. The poor statistics offered by the available data on the iron line profile, and the lack of detected lines for close observations, does not allow a deeper investigation on the nature of this feature that could be produced by reflection. A simple modelling with a Relativistic component as discline or shaddisc returns unconstrained values for the physical parameters describing the line profile. Further observations of the source in outburst, performed with future space missions that will be provided with higher effective area and spectral resolution over a wider energy range, will be fundamental to investigate the reflection component in IGR J17591–2342, for a deeper comprehension of the ionisation state of the matter and of the geometry of the system. ### 4.3 A new measurement of the distance The analysis of PRE type I X-ray bursts is largely used in literature for inferring the distance to the sources, the method being often the only possible way to obtain such an estimate. However, this method suffers from systematic uncertainties (see, e.g., Marino et al., 2019, and references therein). For instance, the flux reached by PRE bursts for the same source generally scatters around a mean value with variations of about 15% (see, e.g., Kuulkers et al., 2003; Galloway et al., 2003, 2008). Moreover, the value used for the Eddington luminosity of the source may not be accurate without details on, e.g., the composition of the NS atmosphere and and therefore of its opacity. The latest value for the distance of the source is $d=(7.6\pm 0.7)$ kpc, obtained by Kuiper et al. (2020) from the analysis of a type-I X-ray burst which was suggested to reach the Eddington limit. However, without strong evidences of PRE, the obtained distance could be an upper limit rather than an actual estimate. Furthermore, the aforementioned systematic uncertainties that affect the method, demand to try an alternative way to find the distance and eventually confirm the measurement by Kuiper et al. (2020). From the value of the hydrogen column density, we can derive an estimate of the distance of the system to compare with the existing one by Kuiper et al. (2020), by invoking the 3D extinction map of the radiation in the Ks band for our Galaxy of Chen, B. Q. et al. (2013). The map provides a profile of the radiation extinction in the direction of the Galactic bulge, as a function of the distance to the source. We used the profile at galactic coordinates $l=0.00$, $b=1.00$, reported in Figure 7. Figure 7: Expected profile of the extinction in the Ks band, as a function of the distance to the source in the direction of IGR J17591–2342. The blue vertical lines indicates the best fit value (navy blue line) of the distance to the source inferred by Kuiper et al. (2020) and its relative error (lighter blue lines). The green line represents the best fit parabolic function that fits the profile in the range 5–10 kpc. The value of NH is related to the visual extinction of the source radiation AV through the relation of Güver & Özel (2009): $N_{H}=(2.21\pm 0.09)\times 10^{21}A_{V}.$ (3) The visual extinction is then related to the extinction of the radiation in the Ks band (A${}_{K_{S}}$) through the relation of Nishiyama et al. (2008): $A_{K_{S}}=(0.062\pm 0.005)\;A_{V}\;{\rm mag}.$ (4) We fitted the profile of Figure 7 with a parabolic function in the range 5–10 kpc, i.e. in the region corresponding to values of the distance inferred by Kuiper et al. (2020), that is d = (7.6$\pm$0.7) kpc. Assuming the value of NH found by Kuiper et al. (2020), the value of expected extinction is A${}_{K_{s}}$=0.59$\pm$0.01 mag, corresponding to a distance of $d=(7.2\pm 0.8)$ kpc. The estimation of the distance through the hydrogen column density is strongly affected by uncertainties on this parameter. The accuracy in the evaluation of the column density depends on the quality of the data and on the energy range in which the fit is conducted. A small energy range and a low resolution can lead to correlation effects with other parameters of the model. The use of solar abundances as initial values for the parameters can also lead to differences around 5% in the estimation (see Wilms et al., 2000b). Moreover, the accuracy extinction map itself depends on the quality of the data used, whether they are recent data with higher quality, and the analysed region of the sky, which can be affected by strong intrinsic variations in the interstellar medium. Nonetheless, the two estimates of the distance, by means of the PRE analysis of Kuiper et al. (2020) and through the spectral evolution of this work, are compatible. ## 5 Conclusions We analysed a large sample of the available observations of IGR J17591–2342 in the X-ray archive with the aim of characterising the spectral emission of the source and its evolution during the outburst. The source is well fitted by an absorbed Comptonisation component that on average contributes to 95% of the whole budget of the unabsorbed emitted flux. No significant spectral changes were found for the source during the whole outburst. The estimation of the distance is in line with the same values reported in literature, and equal to $d=(7.2\pm 0.8)$ kpc. The spectral continuum, especially in the time window at which the flux reaches higher values during the outburst, needs to be fitted with a black- body component peaking at a temperature that is correlated with the NICER count-rate. This component is characterised by a temperature of about 0.8 keV and appears to be emitted from a region with a radius of (3.3$\pm$1.5) km that could be compatible with a fraction of the NS surface, or possibly with the boundary layer. No significant variations are observed on the electron temperature of the Comptonising cloud that mainly shows a temperature of about 34 keV. The corona appears to be characterised by an optical depth of about 2.3, which could in part explain the direct black-body component that could arise from a fraction of photons emitted by a small region of the NS surface and that are not significantly scattered in the corona. The spectral continuum appears to be characterised by a Ne ix emission line. This line, however, seems to have a broad profile that is compatible or some times higher than the spectral resolution of NICER at 1 keV. This could be produced by relativistic effects, if this feature is originated in the innermost parts of the accretion disc (the evidence of which is not detectable from the analysed data), or could be an effect of the spectral resolution of NICER, being unable to resolve a complex of lines at those energies. A broad iron emission line has been detected in the spectra of a number of observations. The energy line appears to be correlated to the phase of the outburst. In particular, for lower flux regimes, we observe a broad ($\sim 1$ keV) Fe i K$\alpha$ line, while in proximity of the peak of the outburst we observed a Fe xxv K$\alpha$ line. Observations of future outbursts of IGR J17591–2342 with instruments equipped with larger effective area over a wider energy range, as for example the enhanced X-ray Timing and Polarimetry mission (eXTP, Zhang et al., 2019; In’t Zand et al., 2019), could provide important constraints on the possible reflection component in this system, and then also on the ionization state of the matter and on the inclination angle of the system that can not be inferred by the current statistics. ## Acknowledgements The authors acknowledge financial contribution from the agreement ASI-INAF n.2017-14-H.0 from INAF mainstream (PI: A. De Rosa), and from the HERMES project financed by the Italian Space Agency (ASI) Agreement n. 2016/13 U.O and from the ASI-INAF Accordo Attuativo HERMES Technologic Pathfinder n. 2018-10-H.1-2020. We also acknowledge support from the European Union Horizon 2020 Research and Innovation Framework Programme under grant agreement HERMES- Scientific Pathfinder n. 821896. RI and TDS acknowledge the research grant iPeska (PI: Andrea Possenti) funded under the INAF national call Prin-SKA/CTA approved with the Presidential Decree 70/2016. RI acknowledges financial contribution from the agreement ASI-INAF n.2017-14-H.0, from INAF mainstream (PI: T. Belloni). A. Marino is supported by the H2020 ERC Consolidator Grant "MAGNESIA" under grant agreement No. 817661 (PI: Rea) and National Spanish grant PGC2018-095512-BI00. This work was also partially supported by the program Unidad de Excelencia Maria de Maeztu CEX2020-001058-M, and by the PHAROS COST Action (No. CA16214). ## Data availability The data utilized in this article are publicly available in the Heasarc Data Archive at https://heasarc.gsfc.nasa.gov/cgi-bin/W3Browse/w3browse.pl. The ASTROSAT data are publicly available in the ISRO Science Data Archive at https://webapps.issdc.gov.in/astro_archive/archive/Search.jsp. ## References * Agrawal et al. (2017) Agrawal P. C., et al., 2017, Journal of Astrophysics and Astronomy, 38, 30 * Alpar et al. (1982) Alpar M. A., Cheng A. F., Ruderman M. A., Shaham J., 1982, Nature, 300, 728 * Antia et al. (2017) Antia H. M., et al., 2017, ApJS, 231, 10 * Arnaud (1996) Arnaud K. A., 1996, in Jacoby G. H., Barnes J., eds, Astronomical Society of the Pacific Conference Series Vol. 101, Astronomical Data Analysis Software and Systems V. p. 17 * Bult et al. (2022) Bult P., et al., 2022, The discovery of the 528.6 Hz accreting millisecond X-ray pulsar MAXI J1816-195, doi:10.48550/ARXIV.2208.04721, https://arxiv.org/abs/2208.04721 * Canizares et al. (2005) Canizares C. R., et al., 2005, PASP, 117, 1144 * Chen, B. Q. et al. (2013) Chen, B. Q. Schultheis, M. Jiang, B. W. Gonzalez, O. A. Robin, A. C. Rejkuba, M. Minniti, D. 2013, A&A, 550, A42 * Courvoisier et al. (2003) Courvoisier T. J. L., et al., 2003, A&A, 411, L53 * Den Herder et al. (2001) Den Herder J. W., et al., 2001, A&A, 365, L7 * Di Salvo & Sanna (2020) Di Salvo T., Sanna A., 2020, arXiv e-prints, p. arXiv:2010.09005 * Ducci et al. (2018) Ducci L., Kuulkers E., Grinberg V., Paizis A., Sidoli L., Bozzo E., Ferrigno C., Savchenko V., 2018, The Astronomer’s Telegram, 11941, 1 * Fabian et al. (1989) Fabian A. C., Rees M. J., Stella L., White N. E., 1989, MNRAS, 238, 729 * Galloway et al. (2003) Galloway D. K., Psaltis D., Chakrabarty D., Muno M. P., 2003, ApJ, 590, 999 * Galloway et al. (2008) Galloway D. K., Muno M. P., Hartman J. M., Psaltis D., Chakrabarty D., 2008, ApJS, 179, 360 * Gendreau et al. (2012) Gendreau K. C., Arzoumanian Z., Okajima T., 2012, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series. p. 13, doi:10.1117/12.926396 * Gendreau et al. (2016) Gendreau K. C., et al., 2016, in Space Telescopes and Instrumentation 2016: Ultraviolet to Gamma Ray. p. 99051H, doi:10.1117/12.2231304 * Güver & Özel (2009) Güver T., Özel F., 2009, MNRAS, 400, 2050 * Harrison et al. (2013) Harrison F. A., et al., 2013, ApJ, 770, 103 * In ’t Zand et al. (1999) In ’t Zand J. J. M., et al., 1999, A&A, 345, 100 * In’t Zand et al. (2019) In’t Zand J. J. M., et al., 2019, Science China Physics, Mechanics, and Astronomy, 62, 29506 * Jaisawal et al. (2019) Jaisawal G. K., et al., 2019, ApJ, 885, 18 * Kuiper et al. (2020) Kuiper L., Tsygankov S. S., Falanga M., Mereminskiy I. A., Galloway D. K., Poutanen J., Li Z., 2020, A&A, 641, A37 * Kuulkers et al. (2003) Kuulkers E., den Hartog P. R., in’t Zand J. J. M., Verbunt F. W. M., Harris W. E., Cocchi M., 2003, A&A, 399, 663 * La Placa et al. (2020) La Placa R., Stella L., Papitto A., Bakala P., Di Salvo T., Falanga M., De Falco V., De Rosa A., 2020, ApJ, 893, 129 * Laurent et al. (2003) Laurent P., et al., 2003, A&A, 411, L185 * Lund et al. (2003) Lund N., et al., 2003, A&A, 411, L231 * Marino et al. (2019) Marino A., et al., 2019, MNRAS, 490, 2300 * Misra et al. (2017) Misra R., et al., 2017, The Astrophysical Journal, 835, 195 * Ng et al. (2021) Ng M., et al., 2021, ApJ, 908, L15 * Nishiyama et al. (2008) Nishiyama S., Nagata T., Tamura M., Kandori R., Hatano H., Sato S., Sugitani K., 2008, ApJ, 680, 1174 * Nowak et al. (2019) Nowak M. A., Paizis A., Jaisawal G. K., Chenevez J., Chaty S., Fortin F., Rodriguez J., Wilms J., 2019, ApJ, 874, 69 * Remillard et al. (2021) Remillard R. A., et al., 2021, arXiv e-prints, p. arXiv:2105.09901 * Russell et al. (2018) Russell T. D., Degenaar N., Wijnands R., van den Eijnden J., Gusinskaia N. V., Hessels J. W. T., Miller-Jones J. C. A., 2018, ApJ, 869, L16 * Sanna et al. (2018) Sanna A., et al., 2018, A&A, 617, L8 * Sanna et al. (2020) Sanna A., et al., 2020, MNRAS, 495, 1641 * Strüder et al. (2001) Strüder L., et al., 2001, A&A, 365, L18 * Turner et al. (2001) Turner M. J. L., et al., 2001, A&A, 365, L27 * Ubertini et al. (2003) Ubertini P., et al., 2003, A&A, 411, L131 * Verner et al. (1996) Verner D. A., Ferland G. J., Korista K. T., Yakovlev D. G., 1996, ApJ, 465, 487 * Wilms et al. (2000a) Wilms J., Nowak M. A., Boyd P., Pottschmidt K., Heindl W. A., Begelman M. C., 2000a, in AAS/High Energy Astrophysics Division #5. p. 1247 * Wilms et al. (2000b) Wilms J., Allen A., McCray R., 2000b, ApJ, 542, 914 * Yadav et al. (2016) Yadav J. S., et al., 2016, ApJ, 833, 27 * Zdziarski et al. (1996) Zdziarski A. A., Johnson W. N., Magdziarz P., 1996, MNRAS, 283, 193 * Zhang et al. (2019) Zhang S., et al., 2019, Science China Physics, Mechanics, and Astronomy, 62, 29502 ## Appendix A Tables Table 1: Observations considered in this work. | | | ---|---|---|--- ObsID | Satellite | Start Time (UT) | Stop Time (UT) 90401331002 | NuSTAR | 2018-08-13T22:36:09 | 2018-08-14T14:26:09 1200310101 | NICER | 2018-08-14T23:59:42 | 2018-08-15T14:08:00 80301311002 | NuSTAR | 2018-08-17T20:01:09 | 2018-08-18T13:21:09 Rev. 1989 | INTEGRAL | 2018-8-17 10:45:21 | 2018-8-19 19:44:37 1200310102 | NICER | 2018-08-18T02:07:53 | 2018-08-18T03:56:20 1200310103 | NICER | 2018-08-23T00:58:20 | 2018-08-23T23:01:36 9000002320 | AstroSAT | 2018-08-23T01:10:15 | 2018-08-24T00:41:18 20173 | Chandra | 2018-08-23T17:40:05 | 2018-08-23T23:31:42 1200310104 | NICER | 2018-08-24T01:51:10 | 2018-08-24T14:39:23 Rev. 1992 | INTEGRAL | 2018-8-25 13:08:59 | 2018-8-27 19:11:30 1200310105 | NICER | 2018-08-27T00:49:37 | 2018-08-27T22:39:23 9000002332 | AstroSAT | 2018-08-27T00:05:56 | 2018-08-27T21:59:56 1200310106 | NICER | 2018-08-27T23:55:20 | 2018-08-28T01:44:44 1200310107 | NICER | 2018-08-30T15:18:00 | 2018-08-30T17:35:00 Rev. 1994 | INTEGRAL | 2018-8-30 20:48:49 | 2018-9-2 3:31:23 1200310108 | NICER | 2018-08-31T22:13:38 | 2018-08-31T22:55:20 1200310109 | NICER | 2018-08-31T23:47:15 | 2018-09-01T17:05:18 1200310110 | NICER | 2018-09-02T06:40:10 | 2018-09-02T23:58:36 Rev. 1995 | INTEGRAL | 2018-9-2 12:39:01 | 2018-9-4 19:22:19 795750101 | XMM-Newton | 2018-09-03T18:47:34.08 | 2018-09-04T03:58:54.86 1200310111 | NICER | 2018-09-03T01:12:30 | 2018-09-03T07:41:56 1200310112 | NICER | 2018-09-05T17:59:40 | 2018-09-05T21:55:00 1200310113 | NICER | 2018-09-06T00:24:00 | 2018-09-06T06:49:00 1200310114 | NICER | 2018-09-07T08:42:40 | 2018-09-07T09:26:40 1200310115 | NICER | 2018-09-08T01:42:11 | 2018-09-08T11:42:51 1200310116 | NICER | 2018-09-09T03:58:11 | 2018-09-09T22:59:33 Rev. 1998 | INTEGRAL | 2018-9-10 12:07:30 | 2018-9-12 18:48:26 1200310118 | NICER | 2018-09-11T00:44:36 | 2018-09-11T17:58:50 1200310119 | NICER | 2018-09-11T23:54:45 | 2018-09-12T21:49:14 1200310120 | NICER | 2018-09-13T05:15:23 | 2018-09-13T22:53:29 1200310121 | NICER | 2018-09-14T07:57:04 | 2018-09-14T17:04:00 1200310122 | NICER | 2018-09-15T00:54:21 | 2018-09-15T19:31:40 Rev. 2001 | INTEGRAL | 2018-9-18 11:33:28 | 2018-9-20 18:17:35 1200310124 | NICER | 2018-09-19T06:26:49 | 2018-09-19T22:06:40 1200310125 | NICER | 2018-09-20T05:35:18 | 2018-09-20T15:09:00 Rev. 2002 | INTEGRAL | 2018-9-21 4:47:37 | 2018-9-23 10:06:50 1200310126 | NICER | 2018-09-21T00:07:18 | 2018-09-21T11:14:40 1200310127 | NICER | 2018-09-23T07:45:04 | 2018-09-23T11:11:45 Rev. 2003 | INTEGRAL | 2018-9-23 19:12:26 | 2018-9-26 1:57:06 1200310128 | NICER | 2018-09-24T02:17:04 | 2018-09-24T02:41:20 1200310129 | NICER | 2018-09-25T01:27:04 | 2018-09-25T15:49:20 1200310130 | NICER | 2018-09-26T03:55:00 | 2018-09-26T04:11:20 Rev. 2004 | INTEGRAL | 2018-9-26 11:02:00 | 2018-9-28 17:47:47 1200310131 | NICER | 2018-09-28T00:30:47 | 2018-09-28T05:38:00 Rev. 2005 | INTEGRAL | 2018-9-29 4:14:04 | 2018-10-1 9:38:30 Rev. 2006 | INTEGRAL | 2018-10-1 18:42:44 | 2018-10-4 0:28:03 1200310132 | NICER | 2018-10-04T01:52:19 | 2018-10-04T22:15:00 1200310133 | NICER | 2018-10-05T18:10:40 | 2018-10-05T21:19:40 1200310134 | NICER | 2018-10-10T20:33:44 | 2018-10-10T20:48:40 1200310135 | NICER | 2018-10-11T01:11:02 | 2018-10-11T01:28:20 1200310136 | NICER | 2018-10-12T06:32:14 | 2018-10-12T06:56:18 1200310137 | NICER | 2018-10-13T13:24:10 | 2018-10-13T15:15:40 1200310138 | NICER | 2018-10-16T03:14:20 | 2018-10-16T05:06:20 1200310139 | NICER | 2018-10-17T05:32:22 | 2018-10-17T18:13:20 Table 2: Best fit model continuum for each observation of IGR J17591–2342. The associated errors are reported at 90% confidence level. | | Tbabs | Edge | Bbodyrad | nthComp | $\chi^{2}/dof$ ---|---|---|---|---|---|--- ObsID | MJD | NH | EdgeE | MaxTau | kT | Norm | Gamma | $kT_{e}$ | kTseed | Norm | | | ($\times 10^{22}$) | (keV) | | (keV) | | | (keV) | (keV) | | 1200310101 | 58345.0 | 2.09* | $0.871*$ | $2.0\pm 0.2$ | $1.10\pm 0.06$ | $0.8\pm 0.2$ | $1.76\pm 0.02$ | $22^{+4}_{-3}$ | $0.43\pm 0.03$ | $0.0153^{+0.0010}_{-0.0009}$ | $2444.3/2147$ 1200310102 | 58348.09 | 2.09* | $0.871*$ | $2.5^{+0.6}_{-0.2}$ | – | – | $1.77\pm 0.02$ | $22^{+8}_{-4}$ | $1.03^{+0.06}_{-0.07}$ | $0.0019^{+0.0003}_{-0.0002}$ | $1541.3/1515$ 1200310103 | 58353.04 | 2.09* | $0.871*$ | $2.0\pm 0.1$ | – | – | 1.9* | 38.8* | $0.58\pm 0.01$ | $0.0150\pm 0.0004$ | $1507.4/1641$ 1200310104 | 58354.08 | 2.09* | $0.871*$ | $1.7\pm 0.2$ | – | – | $1.91\pm 0.05$ | 38.8* | $0.63\pm 0.02$ | $0.0139^{+0.0007}_{-0.0006}$ | $1346.2/1563$ 1200310105 | 58357.03 | 2.09* | $0.871*$ | $2.2\pm 0.3$ | $1.02^{+0.08}_{-0.09}$ | $4.4^{+2.1}_{-1.6}$ | $1.65^{+0.07}_{-0.07}$ | 38.8* | $0.44^{+0.06}_{-0.07}$ | $0.016\pm 0.002$ | $760.8/718$ 1200310106 | 58358.0 | 2.09* | $0.871*$ | $2.5^{+0.3}_{-0.7}$ | $1.15^{+0.09}_{-0.07}$ | $3\pm 1$ | $1.71^{+0.07}_{-0.06}$ | $40^{+57}_{-13}$ | $0.42^{+0.10}_{-0.13}$ | $0.019^{+0.009}_{-0.004}$ | $538.0/557$ 1200310107 | 58360.64 | 2.09* | – | – | – | – | $1.82\pm 0.05$ | $31^{+55}_{-10}$ | $0.54\pm 0.02$ | $0.012\pm 0.002$ | $657.4/623$ 1200310108 | 58361.93 | 2.09* | $0.871*$ | $2.1\pm 0.3$ | – | – | $1.84^{+0.07}_{-0.06}$ | $32^{+86}_{-11}$ | $0.55\pm 0.04$ | $0.022\pm 0.002$ | $587.1/540$ 1200310109 | 58361.99 | 2.09* | $0.871*$ | $1.8\pm 0.2$ | – | – | $1.90\pm 0.04$ | $\geq 42$ | $0.60\pm 0.02$ | $0.0207^{+0.0009}_{-0.0008}$ | $798.7/769$ 1200310110 | 58363.28 | 2.09* | $0.871*$ | $1.9\pm 0.2$ | $0.58\pm^{+0.04}_{-0.05}$ | $58^{+10}_{-8}$ | $2.3^{+0.3}_{-0.2}$ | 38.8* | $1.1\pm 0.2$ | $0.0061^{+0.0028}_{-0.0015}$ | $865.2/745$ 1200310111 | 58364.05 | 2.09* | $0.871*$ | $1.6\pm 0.2$ | – | – | $1.79\pm 0.02$ | 38.8* | $0.57\pm 0.02$ | $0.0164\pm 0.0008$ | $1190.3/840$ 1200310112 | 58366.75 | 2.09* | $0.871*$ | $1.4\pm 0.6$ | $0.61^{+0.09}_{-0.10}$ | $36^{+21}_{-8}$ | 1.9* | 38.8* | $1.3^{+0.5}_{-0.4}$ | $\leq 0.0023$ | $533.0/489$ 1200310113 | 58367.02 | 2.09* | $0.871*$ | $2.4\pm 0.4$ | – | – | $1.78^{+0.06}_{-0.05}$ | 38.8* | $0.45\pm 0.05$ | $0.017\pm 0.002$ | $561.6/535$ 1200310114 | 58368.36 | 2.09* | $0.871*$ | $1.4\pm 0.3$ | – | – | 1.9* | 38.8* | $0.60\pm 0.03$ | $0.0132^{+0.0010}_{-0.0009}$ | $492.6/496$ 1200310115 | 58369.07 | 2.09* | $0.871*$ | $1.9\pm 0.2$ | – | – | $1.87^{+0.06}_{-0.05}$ | 38.8* | $0.57\pm 0.03$ | $0.0148^{+0.0010}_{-0.0009}$ | $683.9/643$ 1200310116 | 58370.17 | 2.09* | $0.871*$ | $2.0\pm 0.2$ | – | – | $1.93^{+0.06}_{-0.05}$ | $\geq 41$ | $0.62\pm 0.03$ | $0.0168^{+0.0013}_{-0.0009}$ | $735.3/718$ 1200310118 | 58372.03 | 2.09* | $0.871*$ | $1.3\pm 0.3$ | $0.73^{+0.08}_{-0.09}$ | $45^{+8}_{-40}$ | $1.9\pm 0.2$ | 38.8* | $1.8^{+0.8}_{-0.5}$ | $0.005^{+0.014}_{-0.002}$ | $530.0/499$ 1200310119 | 58373.0 | 2.09* | $0.871*$ | $2.0\pm 0.3$ | $0.65^{+0.06}_{-0.10}$ | $45^{+8}_{-40}$ | $1.8^{+0.5}_{-0.4}$ | $\geq 27$ | $1.2^{+0.4}_{-0.7}$ | $0.005^{+0.014}_{-0.002}$ | $719.1/648$ 1200310120 | 58374.22 | 2.09* | $0.871*$ | $2.2\pm 0.3$ | – | – | $1.84\pm 0.06$ | 38.8* | $0.62\pm 0.03$ | $0.021\pm 0.001$ | $641.6/642$ 1200310121 | 58375.33 | 2.09* | – | – | – | – | $2.19\pm 0.02$ | 38.8* | $0.87\pm 0.04$ | $0.0130\pm 0.005$ | $561.0/426$ 1200310122 | 58376.04 | 2.09* | $0.871*$ | $2.4\pm 0.3$ | – | – | $1.81^{+0.08}_{-0.07}$ | 38.8* | $0.60\pm 0.04$ | $0.025\pm 0.002$ | $601.8/548$ 1200310124 | 58380.27 | 2.09* | $0.871*$ | $2.2\pm 0.3$ | – | – | $1.90^{+0.09}_{-0.07}$ | $58^{+289}_{-23}$ | $0.64\pm 0.04$ | $0.028\pm 0.002$ | $511.7/554$ 1200310125 | 58381.23 | 2.09* | $0.871*$ | $3.1\pm 0.5$ | – | – | $1.77^{+0.09}_{-0.07}$ | $33^{+20}_{-8}$ | $0.53\pm 0.6$ | $0.033^{+0.005}_{-0.004}$ | $401.3/420$ 1200310126 | 58382.01 | 2.09* | $0.871*$ | $2.2\pm 0.2$ | – | – | $1.89^{+0.07}_{-0.06}$ | $40^{+54}_{-13}$ | $0.62\pm 0.03$ | $0.029\pm 0.002$ | $605.5/621$ 1200310127 | 58384.32 | 2.09* | $0.871*$ | $2.1\pm 0.3$ | – | – | $1.91^{+0.09}_{-0.08}$ | $33^{+33}_{-10}$ | $0.63\pm 0.04$ | $0.025\pm 0.002$ | $555.6/561$ 1200310128 | 58385.1 | 2.09* | $0.871*$ | $2.0\pm 0.3$ | – | – | $1.90^{+0.08}_{-0.07}$ | 38.8* | $0.61\pm 0.04$ | $0.023\pm 0.002$ | $525.1/523$ 1200310129 | 58386.06 | 2.09* | $0.871*$ | $2.1\pm 0.3$ | – | – | $1.90\pm 0.06$ | 38.8* | $0.61^{+0.03}_{-0.04}$ | $0.0215^{+0.002}_{-0.001}$ | $688.0/630$ 1200310130 | 58387.16 | 2.09* | $0.871*$ | $2.0\pm 0.4$ | $0.68^{+0.07}_{-0.08}$ | $49^{+14}_{-10}$ | $2.0\pm 0.1$ | 38.8* | $1.4\pm 0.4$ | $0.004^{+0.003}_{-0.001}$ | $468.7/465$ 1200310131 | 58389.02 | 2.09* | $0.871*$ | $2.0\pm 0.3$ | – | – | $1.91\pm 0.06$ | 38.8* | $0.57\pm 0.04$ | $0.019^{+0.002}_{-0.001}$ | $619.8/613$ 1200310132 | 58395.08 | 2.09* | $0.871*$ | $2.1\pm 0.4$ | – | – | $1.87^{+0.21}_{-0.07}$ | $\geq 45$ | $0.45\pm 0.05$ | $0.0168^{+0.003}_{-0.002}$ | $453.4/453$ 1200310133 | 58396.76 | 2.09* | – | – | – | – | $2.0^{+0.3}_{-0.2}$ | 38.8* | $0.62\pm 0.07$ | $0.0075^{+0.0007}_{-0.0006}$ | $134.4/120$ Table 3: Best fit parameters for the emission lines detected in the IGR J17591–2342 spectrum for each observation. The line energy for the Ne ix ion was fixed to the rest-frame energy, while the sigma was fixed to 0.085 keV for each ion transition (i.e. the spectral resolution of NICER/XTI at 1 keV). All the associated errors are reported at 68% confidence level. | | Ne ix | Fe xxv ---|---|---|--- ObsID | MJD | Norm | LineE | Sigma | Norm | | | (keV) | (keV) | 1200310101 | 58345.0 | $0.015\pm 0.002$ | – | – | – 1200310102 | 58348.09 | $0.008^{+0.003}_{-0.002}$ | – | – | – 1200310103 | 58353.04 | $0.021\pm 0.002$ | $6.6^{+0.3}_{-0.4}$ | $0.5\pm 0.3$ | $0.00015^{0.00006}_{-0.00005}$ 1200310104 | 58354.08 | $0.019^{+0.003}_{-0.002}$ | – | – | – 1200310105 | 58357.03 | $0.0108^{+0.0016}_{-0.0014}$ | $\leq 6.3$ | $1.5\pm 0.2$ | $0.0026^{+0.0007}_{-0.0008}$ 1200310106 | 58358.0 | $0.024^{+0.007}_{-0.009}$ | $6.59\pm 0.08$ | $0.1*$ | $0.00018^{+0.00002}_{-0.00005}$ 1200310107 | 58360.64 | $0.034^{+0.007}_{-0.006}$ | – | – | – 1200310108 | 58361.93 | $0.028^{+0.007}_{-0.006}$ | – | – | – 1200310109 | 58361.99 | $0.016\pm 0.002$ | – | – | – 1200310110 | 58363.28 | $0.016\pm 0.002$ | – | – | – 1200310111 | 58364.05 | $0.016^{+0.003}_{-0.002}$ | $6.7\pm 0.2$ | $0.8^{+0.3}_{-0.2}$ | $0.00019^{+0.00005}_{-0.00004}$ 1200310112 | 58366.75 | $0.0068^{+0.0019}_{-0.0015}$ | – | – | – 1200310113 | 58367.02 | $0.017^{+0.005}_{-0.004}$ | – | – | – 1200310114 | 58368.36 | $0.016^{+0.004}_{-0.003}$ | $6.53^{+0.07}_{-0.06}$ | $0.02^{+0.10}_{-0.02}$ | $0.00009\pm 0.00004$ 1200310115 | 58369.07 | $0.017\pm 0.003$ | – | – | – 1200310116 | 58370.17 | $0.019\pm 0.003$ | – | – | – 1200310118 | 58372.03 | $0.016^{+0.005}_{-0.004}$ | – | – | – 1200310119 | 58373.0 | $0.019^{+0.004}_{-0.003}$ | – | – | – 1200310120 | 58374.22 | $0.030^{+0.006}_{-0.005}$ | – | – | – 1200310121 | 58375.33 | $0.048^{+0.020}_{-0.014}$ | – | – | – 1200310122 | 58376.04 | $0.041^{+0.011}_{-0.009}$ | – | – | – 1200310124 | 58380.27 | $0.031^{+0.009}_{-0.007}$ | – | – | – 1200310125 | 58381.23 | $0.08^{+0.04}_{-0.03}$ | – | – | – 1200310126 | 58382.01 | $0.033^{+0.007}_{-0.006}$ | – | – | – 1200310127 | 58384.32 | $0.019^{+0.005}_{-0.004}$ | – | – | – 1200310128 | 58385.1 | $0.024^{+0.007}_{-0.005}$ | – | – | – 1200310129 | 58386.06 | $0.019^{+0.004}_{-0.003}$ | – | – | – 1200310130 | 58387.16 | $0.017^{+0.006}_{-0.005}$ | – | – | – 1200310131 | 58389.02 | $0.018^{+0.004}_{-0.003}$ | – | – | – 1200310132 | 58395.08 | $0.013^{+0.004}_{-0.003}$ | $6.63^{+0.06}_{-0.05}$ | $0.04331^{+0.09}_{-0.04}$ | $0.00013^{+0.00005}_{-0.00004}$ 1200310133 | 58396.76 | – | – | – | –
11institutetext: School of Technology, Pontifícia Universidade Católica do Rio Grande do Sul Av. Ipiranga, 6681, 90619-900, Porto Alegre, RS, Brazil 11email<EMAIL_ADDRESS> 11email<EMAIL_ADDRESS> # Zero-shot performance of the Segment Anything Model (SAM) in 2D medical imaging: A comprehensive evaluation and practical guidelines Christian Mattjie 11 Luis Vinicius de Moura 11 Rafaela Cappelari Ravazio 11 Lucas Silveira Kupssinskü 11 Otávio Parraga 11 Marcelo Mussi Delucis 11 Rodrigo C. Barros 11 ###### Abstract Segmentation in medical imaging is a critical component for the diagnosis, monitoring, and treatment of various diseases and medical conditions. Presently, the medical segmentation landscape is dominated by numerous specialized deep learning models, each fine-tuned for specific segmentation tasks and image modalities. The recently-introduced Segment Anything Model (SAM) employs the ViT neural architecture and harnesses a massive training dataset to segment nearly any object; however, its suitability to the medical domain has not yet been investigated. In this study, we explore the zero-shot performance of SAM in medical imaging by implementing eight distinct prompt strategies across six datasets from four imaging modalities, including X-ray, ultrasound, dermatoscopy, and colonoscopy. Our findings reveal that SAM’s zero-shot performance is not only comparable to, but in certain cases, surpasses the current state-of-the-art. Based on these results, we propose practical guidelines that require minimal interaction while consistently yielding robust outcomes across all assessed contexts. The source code, along with a demonstration of the recommended guidelines, can be accessed at https://github.com/Malta-Lab/SAM-zero-shot-in-Medical-Imaging. ###### Keywords: Medical Imaging Segmentation Segment Anything Model Zero-shot Learning Deep Neural Networks. ## 1 Introduction Medical imaging plays a pivotal role in the diagnosis, monitoring, and treatment of a wide range of diseases and conditions [1]. Accurate segmentation of these images is often critical in extracting valuable information that can aid clinical decision-making. However, traditional segmentation methods primarily rely on labor-intensive, manually-engineered features and error-prone thresholding designed for specific scenarios, resulting in limited generalizability to new images [2]. Large advancements in medical image segmentation have been achieved with the advent of deep learning (DL) techniques, owing to their ability to learn intrinsic features and patterns from large datasets [3, 4, 5]. The DL revolution was ignited by the groundbreaking success of Convolutional Neural Networks (CNNs) in computer vision applications [6]. Recently, a new wave of innovative applications based on the Transformer architecture has emerged [7]. Transformers enhance the training process by harnessing larger datasets while providing smaller induction bias, thereby creating models that can generalize to unseen distributions and even adapt to diverse tasks. Nonetheless, medical image segmentation poses significant challenges for DL due to the substantial cost associated with specialized professionals annotating images, leading to the scarcity of available data. Furthermore, there is limited evidence regarding the ability of DL models trained on natural images to generalize to medical application settings. The Segment Anything Model (SAM) has been recently introduced by Meta [8]. SAM, a state-of-the-art vision transformer (ViT), is capable of generating segmentation masks for virtually any object. It introduces the concept of prompting in image segmentation, whereby the model’s inference process is guided by providing points inside the region of interest (ROI) or by drawing a bounding box around it. In this paper, we rigorously evaluate the zero-shot capabilities of SAM in segmenting 2D medical images. We assess its performance across six datasets encompassing four distinct imaging modalities: X-ray, ultrasound, dermatoscopy, and colonoscopy, using various prompting strategies. Our comprehensive evaluation reveals that SAM demonstrates promising results in those medical imaging modalities, even when we have complex patterns such as hair on skin lesions. We also propose practical guidelines for physicians to utilize SAM in medical image segmentation tasks. This guideline suggests starting with a bounding box prompt, selecting the optimal prediction from the generated outputs, and refining the segmentation using point prompts when necessary. ## 2 Related Work ### 2.1 Medical Image Segmentation Medical image segmentation plays a pivotal role in medical imaging analysis, focusing on the identification and delineation of structures or regions such as organs, tissues, or lesions. Accurate segmentation is crucial for various clinical applications, encompassing diagnosis, treatment, and monitoring of disease progression. This enables essential tasks like measuring tissue volume for tracking growth and outlining radiosensitive organs in radiotherapy treatment. In the current domain of medical image segmentation, specific methods are tailored to the application, imaging modality, and body part under examination [9, 10]. However, automatic segmentation remains a formidable challenge due to the intricacy of medical images and data scarcity. The segmentation algorithm’s output is influenced by multiple factors, including the partial volume effect, intensity inhomogeneity, presence of artifacts, and insufficient contrast between soft regions [11]. Deep learning techniques have garnered considerable attention in medical image segmentation, owing to their capacity for capturing intricate patterns and representations from large-scale datasets. Among the most prevalent DL approaches for medical image segmentation are CNNs. Widely employed models for medical image segmentation include U-Net [3] and its derivatives, which were explicitly developed for biomedical image segmentation. U-Net utilizes a symmetric encoder-decoder architecture, enabling the model to capture both high-level contextual information and fine-grained details, resulting in enhanced segmentation outcomes. In recent years, novel state-of-the-art segmentation techniques have emerged, such as training DL models on polar images [12], integrating textual information with vision-language models [4], and employing attention mechanisms with CNNs in ViTs [13]. ### 2.2 Vision Transformer (ViT) ViTs constitute a class of DL models that leverage the transformer architecture [14]. These models process images by dividing them into fixed- size, non-overlapping patches and linearly embedding these patches into a flat sequence of tokens. Each token is subsequently passed through a series of self-attention layers to learn relevant contextual relationships and spatial information, enabling the model to discern semantically-rich patterns [7]. ViTs do not share some of the inductive biases inherent in CNNs, such as locality and translation equivariance. A reduced inductive bias allows ViTs to be more adaptable even though it necessitates more data for generalization. The data demand may limit the application of ViTs in medical imaging, where data is often scarce. Nevertheless, by capitalizing on pre-training and fine- tuning strategies, ViTs are revolutionizing the computer vision landscape with strong generalization performance [15, 16]. Recently, ViTs have demonstrated strong results in zero-shot learning [17, 18, 19]. This setting presents a challenge since the model must learn to generalize for classes and contexts not encountered during training. In medical imaging, ViT-based models have achieved state-of-the-art results [20, 21, 13], though very few studies address the zero-shot capabilities of the learning models, and whether their performance in zero-shot settings is reasonable or even competitive to fine-tuning [22, 13]. ## 3 Methodology ### 3.1 Segment Anything Model (SAM) SAM [8] is a state-of-the-art ViT model trained on the massive SA-1B dataset (also introduced in [8]). This dataset comprises approximately 11 million images and 1 billion segmentation masks, making it the largest publicly available image segmentation dataset to date. The model’s high accuracy has been demonstrated through its impressive capability of segmenting a wide variety of objects and shapes, thereby validating its effectiveness in segmenting virtually any object within a 2D image. SAM can function in two distinct ways: by segmenting all objects present in an input image or by utilizing prompts that explicitly specify the target region for segmentation. These prompts can take the form of points identifying the region of interest or regions that should be excluded. Additionally, a bounding box may be provided to delineate the area containing the object of interest. While initial results with SAM showcase strong segmentation quality and zero-shot generalization to novel scenes and unseen objects, it is important to note that the model’s training dataset lacks medical images. Consequently, its generalizability to the medical domain remains an open question. To address potential issues arising from ambiguous prompts, SAM generates a set of three masks, each with an accompanying score reflecting a different interpretation of the intended region. The first mask in the output sequence represents the smallest, most conservative interpretation of the intended region according to the given prompt. As the sequence progresses, the subsequent masks increase in size, with each mask encompassing the previous one. The score assigned to each mask is an indicator of SAM’s confidence in that particular prediction. This design enables SAM to accommodate a wider range of potential segmentation outcomes, reflecting the model’s efforts to account for the ambiguity in the target region’s size due to the prompt’s limited information. In practical applications, especially within the medical imaging domain, it is crucial to ensure that the model accurately identifies and segments pertinent structures or regions of interest. Given this requirement, our study focused on investigating input prompt strategies for guiding SAM’s segmentation process. This decision stems from the inherent uncertainties associated with the segment-everything approach, as the model’s comprehension of the segmented objects cannot be guaranteed. By utilizing prompts, we aimed to improve SAM’s segmentation capabilities in medical imaging tasks and provide a more reliable and controlled evaluation of its performance. Furthermore, we did not consider the confidence scores provided by SAM for each mask, as these scores reflect the quality of the segmentation without accounting for the accuracy of the target region relative to the intended object. The ViT architecture employed by SAM consists of three distinct iterations, each with unique trade-offs between computational requirements and model performance: ViT Base (ViT-B), ViT Large (ViT-L), and ViT Huge (ViT-H). The primary differences between these iterations lie in the model’s number of layers and parameters, as illustrated in Table 1. As the number of layers and parameters increases, the model becomes more powerful, enabling the capture of more intricate aspects of the input images. However, larger models necessitate more computing power, which may pose a drawback in certain situations. Nevertheless, even the largest iteration of SAM remains relatively compact. Architecture | Transformer Layers | Parameters | Size (Mb) ---|---|---|--- ViT-B | 12 | 91M | 776 ViT-L | 24 | 308M | 1582 ViT-H | 32 | 636M | 2950 Table 1: Summary of SAM’s ViT architecture variations. ### 3.2 Datasets For evaluating SAM, we used six datasets from four medical imaging modalities: X-ray, Ultrasound, dermatoscopic, and colonoscopy images. Our primary objective is to assess the model’s performance and versatility when prompted with various strategies, simulating a physician’s approach to segmenting specific organs or ROIs in medical images. Fig 1 shows a sample from each dataset. * • ISIC 2018 [23]: this publicly available dataset comprises $2,594$ dermatoscopic images from $2,056$ unique patients, showcasing skin lesions with varying types, sizes, and colors. The images have resolutions ranging from $640\times 480$ to approximately $6,700\times 4,400$ pixels and are provided in JPEG format. Expert dermatologists generated accompanying segmentation masks using a manual annotation tool, and a second expert reviewed each mask for accuracy. * • HAM10000 [24]: this dataset contains $10,015$ dermatoscopic images of skin lesions from $7,388$ unique patients, with varying types, sizes, and colors. All images have a resolution of $640\times 450$ and are provided in JPEG format. Recently, Tschandl, P. et al.[25] supplied expert segmentation masks for all images, with corresponding resolutions. * • Montgomery-Shenzhen [26, 27]: this dataset is a fusion of two publicly available chest X-ray datasets collected from respective hospitals. It comprises $800$ X-ray images, with $704$ accompanying lung segmentation masks manually created by expert radiologists. The dataset is available in PNG format. * • X-ray Images of Hip Joints [28]: this publicly available dataset contains $140$ X-ray images of the lower legs, with an average resolution of $327\times 512$. Corresponding segmentation masks for the femur and ilium are provided separately. The images and masks are available in NII format. * • CVC-ClinicDB [29]: this dataset consists of $612$ images from $31$ colonoscopy sequences, with a resolution of $384\times 288$. The images are provided in PNG format. Expert gastroenterologists have created segmentation masks for the polyps, which are provided for all available images. * • Breast Ultrasound Images [30]: this dataset comprises $780$ ultrasound images of the breast from $600$ patients, with an average size of $500\times 500$ pixels. The images are provided in PNG format and are categorized into normal, benign, and malignant. Segmentation masks for tumors are supplied for both benign and malignant cases. Figure 1: Samples from each of the six datasets used in this study. A: ISIC, B: HAM, C: CXR, D: HJXR, E: CVC, F: BUSI. ### 3.3 Prompt Strategies In the context of interactive segmentation, a physician may guide the procedure using various strategies, such as clicking within the region of interest, clicking outside the region, or drawing a bounding box around the target. To investigate the impact of these plausible prompting strategies on our segmentation models, we conducted a series of experiments with the following approaches: * • Central-point (CP): utilizing only the centroid of the ground-truth mask, which is anticipated to be the most informative single-point prompt; * • Random-point (RP): eroding the ground-truth mask and subsequently selecting a random point within it, representing minimal guidance; * • Distributed random-points (RP3 and RP5): eroding the ground-truth mask, dividing it vertically into sections (three and five, respectively), and selecting a random point within each section to provide a more distributed set of prompts; * • Bounding-box (BB): prompting with the bounding box of the ground-truth mask, offering a more explicit spatial constraint for segmentation; and * • Perturbed bounding-box (BBS5, BBS10, and BBS20): modifying the size and position of the bounding box by $5$%, $10$%, and $20$% of the ground-truth mask size, respectively, simulating variations in the accuracy of a physician’s initial assessment. For the multiple points strategy, we divided the mask into three and five sections, and for the varied bounding box strategy, we randomly altered its size and position up to 5%, 10%, and 20% of the ground-truth mask. Given these variations, we ran a total of eight experiments per model/dataset, which are shown in Fig 2: central-point (CP), random-point (RP), random-points-3 (RP3), random-points-5 (RP5), bounding-box (BB), bounding-box-similar-5 (BBS5), bounding-box-similar-10 (BBS10) and bounding-box-similar-20 (BBS20). Figure 2: Example of all prompt strategies on a skin lesion image and mask. (A): original image, (B): CP, (C): RP, (D): RP3, (E): RP5, (F): BB in green, BBS5 in red, BBS10 in blue, and BBS20 in yellow. The size and position shown are their max variation for BBS methods. In the RP, RP3, and RP5 strategies, we apply an erosion morphological operator to the ground-truth mask before selecting a random point within the resulting region. This process ensures that the selected point is not situated near the region of interest’s edges while preserving the element of randomness expected in real-world scenarios. The erosion value was determined according to the dataset: for the CXR and ISIC datasets, where the regions are larger, we used a $30$-pixel radius; for the CVC dataset, which contains smaller images that could be completely eroded, we employed a $1$-pixel radius; for all other datasets with relatively small regions, we opted for a $10$-pixel radius. The various prompting strategies applied to a skin lesion image and mask are illustrated in Fig 2. This figure demonstrates the original image (A), and the different prompt strategies: CP (B), RP (C), RP3 (D), RP5 (E), and BB (F) in green, BBS5 in red, BBS10 in blue, and BBS20 in yellow. The size and position shown for the BBS methods represent their maximum variation, while in our experiments, they were altered randomly. ### 3.4 Preprocessing Throughout our experimentation process, we encountered various challenges stemming from the characteristics of the datasets under consideration. To address these issues, we implemented a fill-holes technique aimed at rectifying mask information, particularly in the context of the ISIC dataset, wherein some masks solely outlined the relevant lesion. Moreover, in instances where an image contained multiple masks (e.g., dual lungs or skin lesions), we isolated the two most substantial regions and processed them independently, employing the prompt strategies delineated in the previous section. This approach ensured accurate and precise segmentation. A manual inspection of all images was conducted to confirm the absence of any containing three distinct and relevant regions. The model-generated predictions for both regions were combined to form a single prediction. In the case of the HJXR dataset, which uniquely contained images in a format incompatible with SAM, we transformed the images from NII to PNG format, normalizing their values within the range of $0$ to $255$. Given that masks for the femur and ilios were available individually for each image in the dataset, we assessed the predictions separately in HJXR-F and HJXR-I. ### 3.5 Evaluation The Dice Similarity Coefficient (DSC) serves as a widely recognized statistical metric for gauging the accuracy of image segmentation. This coefficient quantifies the similarity between two sets of data, typically represented as binary arrays, by comparing a predicted segmentation mask to the ground-truth mask. The DSC operates on a scale from one to zero, with one signifying a perfect match and zero indicating a complete mismatch. The utility of this metric lies in its ability to discern performance differences between classifiers, rendering it an invaluable instrument for evaluating segmentation algorithms. The DSC can be calculated as follows: $\operatorname{DSC}\left(m\left(x_{i}\right),y_{i}\right)=\frac{\left|2*\left(m\left(x_{i}\right)\cap y_{i}\right)\right|}{\left|\left(m\left(x_{i}\right)\cap y_{i}\right)\right|+\left|m\left(x_{i}\right)\cup y_{i}\right|},$ (1) where $\left|m\left(x_{i}\right)\cap y_{i}\right|$ and $\left|m\left(x_{i}\right)\cup y_{i}\right|$ refer to the area of overlap and area of union, respectively. ## 4 Results and Discussion In this section, we present the results of our comprehensive experiments conducted on six datasets, employing eight prompting strategies, and utilizing three variations of the SAM. The performance of these models is compared to the current state-of-the-art (SOTA) methods, with certain zero-shot results of SAM surpassing established benchmarks. We subsequently engage in a qualitative discussion of the observed results, showcasing select challenging images to elucidate our findings. Finally, we provide a practical implementation guideline for physicians to effectively utilize the SAM, ensuring minimal interaction and delivering robust outcomes. Table 2 shows the DSC of the predictions for ViT-H, the largest SAM model, with results for ViT-B and ViT-L shown in the Supplementary Material. The terms 1st, 2nd, and 3rd correspond to the three predictions generated by SAM, and the table presents the metrics when only one of these predictions is used consistently for all images. Fig 3 showcases an example of these predictions for the Chest X-Ray (CXR) dataset, employing both the RP5 and BBS10 strategies. The RP5 method provides better differentiation between predictions, while the BBS10 approach demonstrates greater uniformity. This observation could potentially be attributed to the bounding box, which simultaneously indicates the target region for segmentation and the areas to be excluded (outside the box). Figure 3: Three returning predictions from SAM using RP5 (A, B, C) and BBS10 (D, E, F) input methods for the CXR dataset. A physician may choose the one that best fits the corresponding region to be segmented. Table 2: DSC of predictions for the ViT-H model for six datasets using the eight proposed prompt strategies considering the 1st, 2nd, and 3rd prediction. Dataset | Pred | CP | RP | RP3 | RP5 | BB | BBS5 | BBS10 | BBS20 ---|---|---|---|---|---|---|---|---|--- ISIC | 1st | $0.538$ | $0.531$ | $0.762$ | $0.774$ | $0.745$ | $0.737$ | $0.715$ | $0.603$ 2nd | $0.718$ | $0.677$ | $0.769$ | $0.788$ | $0.845$ | $0.842$ | $0.833$ | $0.789$ 3rd | $0.375$ | $0.363$ | $0.390$ | $0.483$ | 0.872 | $0.868$ | $0.860$ | $0.816$ HAM | 1st | $0.544$ | $0.527$ | $0.752$ | $0.765$ | $0.732$ | $0.724$ | $0.700$ | $0.589$ 2nd | $0.729$ | $0.686$ | $0.768$ | $0.785$ | $0.838$ | $0.835$ | $0.824$ | $0.778$ 3rd | $0.420$ | $0.406$ | $0.443$ | $0.541$ | 0.865 | $0.861$ | $0.851$ | $0.809$ CXR | 1st | $0.904$ | $0.863$ | $0.923$ | $0.927$ | $0.936$ | $0.934$ | $0.911$ | $0.686$ 2nd | $0.758$ | $0.727$ | $0.766$ | $0.828$ | 0.942 | $0.939$ | $0.929$ | $0.826$ 3rd | $0.471$ | $0.469$ | $0.482$ | $0.514$ | $0.935$ | $0.930$ | $0.913$ | $0.803$ HJXR-F | 1st | $0.876$ | $0.822$ | $0.941$ | $0.948$ | $0.924$ | $0.908$ | $0.848$ | $0.618$ 2nd | $0.743$ | $0.767$ | $0.767$ | $0.776$ | 0.962 | $0.958$ | $0.904$ | $0.746$ 3rd | $0.517$ | $0.543$ | $0.548$ | $0.599$ | $0.949$ | $0.945$ | $0.905$ | $0.723$ HJXR-I | 1st | $0.211$ | $0.742$ | $0.808$ | $0.828$ | 0.875 | $0.866$ | $0.734$ | $0.624$ 2nd | $0.393$ | $0.479$ | $0.449$ | $0.491$ | $0.855$ | $0.849$ | $0.790$ | $0.620$ 3rd | $0.294$ | $0.295$ | $0.316$ | $0.384$ | $0.800$ | $0.796$ | $0.758$ | $0.629$ CVC | 1st | $0.716$ | $0.763$ | $0.861$ | $0.880$ | $0.889$ | $0.881$ | $0.835$ | $0.702$ 2nd | $0.554$ | $0.544$ | $0.642$ | $0.754$ | 0.926 | $0.924$ | $0.916$ | $0.844$ 3rd | $0.232$ | $0.224$ | $0.224$ | $0.245$ | $0.924$ | $0.922$ | $0.918$ | $0.868$ BUSI | 1st | $0.583$ | $0.541$ | $0.736$ | $0.766$ | $0.754$ | $0.744$ | $0.713$ | $0.631$ 2nd | $0.641$ | $0.616$ | $0.688$ | $0.735$ | $0.840$ | $0.837$ | $0.823$ | $0.768$ 3rd | $0.192$ | $0.184$ | $0.196$ | $0.254$ | 0.863 | $0.859$ | $0.848$ | $0.800$ In a real-world clinical setting, a physician may opt to select the most suitable prediction. To simulate this decision-making process, we assessed the highest DSC per image, irrespective of being the 1st, 2nd, or 3rd prediction. The results are presented in Table 3. This approach led to a modest improvement of approximately 1% compared to using only the 1st, 2nd or 3rd prediction in all images, as shown in Fig 4. Even though the overall enhancement is marginal, it holds significance for certain subjects and necessitates minimal input from the physician. Table 3: DSC of predictions for all variations of SAM for six datasets using the eight proposed prompt strategies. For each set of predictions, only the one with the highest DSC was considered. Dataset | Model | CP | RP | RP3 | RP5 | BB | BBS5 | BBS10 | BBS20 ---|---|---|---|---|---|---|---|---|--- ISIC | ViT-H | 0.788 | 0.768 | 0.820 | 0.835 | 0.877 | 0.874 | 0.866 | 0.829 ViT-L | 0.783 | 0.768 | 0.811 | 0.818 | 0.876 | 0.872 | 0.864 | 0.819 ViT-B | 0.764 | 0.733 | 0.804 | 0.815 | 0.879 | 0.876 | 0.864 | 0.822 HAM | ViT-H | 0.782 | 0.764 | 0.812 | 0.824 | 0.870 | 0.866 | 0.857 | 0.820 ViT-L | 0.784 | 0.772 | 0.809 | 0.819 | 0.867 | 0.864 | 0.854 | 0.809 ViT-B | 0.745 | 0.706 | 0.785 | 0.796 | 0.872 | 0.867 | 0.855 | 0.810 CXR | ViT-H | 0.922 | 0.902 | 0.928 | 0.936 | 0.952 | 0.950 | 0.942 | 0.862 ViT-L | 0.929 | 0.917 | 0.932 | 0.930 | 0.954 | 0.952 | 0.943 | 0.849 ViT-B | 0.915 | 0.893 | 0.930 | 0.935 | 0.948 | 0.943 | 0.932 | 0.858 HJXR-F | ViT-H | 0.906 | 0.917 | 0.943 | 0.950 | 0.973 | 0.973 | 0.957 | 0.861 ViT-L | 0.910 | 0.916 | 0.939 | 0.948 | 0.973 | 0.973 | 0.956 | 0.880 ViT-B | 0.927 | 0.882 | 0.910 | 0.907 | 0.971 | 0.969 | 0.950 | 0.870 HJXR-I | ViT-H | 0.483 | 0.786 | 0.808 | 0.828 | 0.889 | 0.886 | 0.843 | 0.719 ViT-L | 0.478 | 0.841 | 0.865 | 0.860 | 0.894 | 0.889 | 0.839 | 0.726 ViT-B | 0.500 | 0.765 | 0.825 | 0.830 | 0.875 | 0.870 | 0.838 | 0.696 CVC | ViT-H | 0.838 | 0.854 | 0.884 | 0.898 | 0.940 | 0.938 | 0.934 | 0.889 ViT-L | 0.815 | 0.823 | 0.848 | 0.847 | 0.934 | 0.931 | 0.920 | 0.869 ViT-B | 0.739 | 0.749 | 0.783 | 0.784 | 0.932 | 0.930 | 0.921 | 0.851 BUSI | ViT-H | 0.732 | 0.706 | 0.791 | 0.816 | 0.870 | 0.868 | 0.855 | 0.813 ViT-L | 0.744 | 0.727 | 0.800 | 0.807 | 0.875 | 0.872 | 0.865 | 0.810 ViT-B | 0.734 | 0.701 | 0.804 | 0.818 | 0.886 | 0.884 | 0.874 | 0.831 Figure 4: Comparison of using always the 1st, 2nd, or 3rd prediction versus choosing the best one per image (max) with the BB strategy for all datasets. The Bounding Box (BB) strategy consistently exhibited superior performance across all datasets, as illustrated in Table 2 and Table 3. Even with variations of $5\%$ or $10$% (BBS5, BBS10), this method outperforms all point prompt strategies, while BBS20 achieved results comparable to RP5. This observation underscores the robustness of the bounding box approach, even in the presence of minor inaccuracies while delineating the desired segmentation region. Regarding point prompt methods (CP, RP, RP3, RP5), an increased number of input points correspond to enhanced model performance. However, these techniques could not outperform the BB, BBS5, and BBS10 strategies. Moreover, RP5 requires greater manual intervention, rendering it more labor-intensive compared to employing a bounding box. Our experiments do not incorporate additional prompt points that can be introduced post-prediction to refine the segmentation. This fine-tuning process can be applied to both encompass regions excluded from the prediction and eliminate regions that should not be part of the segmentation. As a result, physicians can achieve even more precise segmentation masks with minimal additional effort. We also highlight that the ViT-B model attained performance levels comparable to the larger variants of SAM, occasionally even surpassing them. Furthermore, owing to its modest GPU memory requirements, it can be readily utilized with cost-effective hardware, making SAM’s application in medical imaging highly accessible without a significant cost. ### 4.1 Comparison with state-of-the-art (SOTA) segmentation models We employ the intermediate-sized SAM (ViT-L) for a comparative analysis with the current state-of-the-art (SOTA) models. Table 4 presents a performance comparison of SAM with the BB5 strategy (emulating a physician annotating with minimal error, followed by selecting the most accurate among three predictions) against SOTA models employed on each dataset. Notably, no baseline models were found for evaluation in the HJXR dataset. Table 4: Comparison of the results of the BBS5 strategy using the ViT-L model with the current state-of-the-art DL models. Dataset | Model | DSC ---|---|--- ISIC | Rema-net [5] | 0.944 SAM ViT-L BBS5 | 0.872 HAM | Rema-net [5] | 0.936 SAM ViT-L BBS5 | 0.864 CXR | Attention U-Net [31] | 0.982 ReSE-Net [32] | 0.976 SAM ViT-L BBS5 | 0.952 CVC | FSA-Net [33] | 0.947 SAM ViT-L BBS5 | 0.931 BUSI | PODDA, A. et al [34] | 0.826 SAM ViT-L BBS5 | 0.872 HJXR-F | SAM ViT-L BBS5 | 0.973 HJXR-I | SAM ViT-L BBS5 | 0.889 SAM achieved very strong results for a zero-shot (no training/fine-tuning) approach in comparison to the SOTA. In the BUSI dataset, SAM surpassed the SOTA by approximately 5%, sustaining its superior performance even when employing the BBS20 strategy, which accommodates a substantial margin of error in image annotation. In the CVC dataset, SAM’s performance was marginally lower (less than 2%), while in the CXR dataset, the gap was a mere 3%. Although no directly comparable studies exist, SAM exhibited a very high DSC ($0.973$) for femur segmentation. The segmentation of ilios is a more intricate task due to reduced contrast with adjacent regions. Taking that into account, the results for ilios segmentation can also be considered quite strong. For the ISIC and HAM datasets, SAM was outperformed by $\approx 7$%. But here we need to take into account the unique characteristics of these datasets, and a more nuanced analysis is presented in the next section. Moreover, the substantial volume of available data (over $10,000$ images) renders the training of task-specific deep learning (DL) models more viable for those tasks. In contrast, with smaller datasets like BUSI, training an end-to-end DL model becomes strenuous due to the scarcity of data. In such scenarios, employing a model like SAM proves to be the best option, as it benefits from exposure to an extensive range of data across various domains. ### 4.2 Qualitative Analysis The analysis of medical images presents a unique set of challenges due to the complex and diverse nature of datasets. For instance, the CXR dataset, which consists of chest X-rays and their corresponding segmentation masks, contains inconsistencies in the segmented regions, as depicted in Fig 5. Some ground- truth masks include the heart while others exclude it. Still, SAM can rapidly rectify these discrepancies by allowing users to select the most appropriate prediction, as demonstrated in Fig 3, or by refining input points to include or exclude specific regions as needed. Figure 5: Example of inconsistencies within the ground-truth region in the CXR dataset. The standard DICOM format for X-ray images typically features a $12$ or $16$ bit depth, enabling physicians to manipulate the window/level settings for enhanced visualization of tissues and organs. We postulate that optimizing the window/level parameters during conversion to JPEG or PNG formats could improve tissue delineation and subsequently enhance SAM’s performance for this imaging modality. Nevertheless, we did not assess this approach, given that the CXR dataset is provided in PNG format, and the HJXR dataset was normalized and converted to PNG using its maximum and minimum values. For the ISIC dataset, which comprises skin lesion images, we identified numerous instances of inaccurate ground-truth mask annotations, as illustrated in Fig 6. These inaccuracies impacted the DSC results, as the masks generated by SAM appear to exhibit higher precision compared to the original masks. Moreover, the presence of body hair in the ISIC and HAM datasets significantly influences the segmentation process, particularly when employing point prompt strategies. For example, a hair intersecting a lesion may erroneously indicate two distinct regions instead of one. To address this issue, bounding box strategies can be implemented to provide sufficient information to the model. However, SAM’s exclusion of hair from the segmentation negatively affects its performance. Additionally, the skin lesion datasets present challenges due to indistinct lesion boundaries, rendering accurate segmentation of skin lesions a challenging task. Figure 6: Example of inconsistencies in the ground-truth region in the ISIC dataset. Ultrasound images pose considerable difficulties for DL models, attributable to their inhomogeneous intensities and low signal-to-noise ratio, which hinder the accurate delineation of breast tumors in datasets such as BUSI. Furthermore, the absorption and reflection of ultrasound can give rise to artifacts in the image, exacerbating the segmentation task even for well- optimized models. Despite these obstacles, SAM achieved strong results in this dataset. However, it encountered challenges in accurately segmenting the boundaries of breast tumors due to the inherent blurred edges in ultrasound images. ### 4.3 Guidelines In light of our empirical findings, we propose a robust and pragmatic framework for utilizing the Segment Anything Model (SAM) in the realm of medical imaging tasks. This methodology empowers physicians to capitalize on the capabilities of SAM to attain precise segmentation outcomes, while preserving their autonomy in overseeing the process. Our recommendation is to employ the largest SAM variation that is feasible given the constraints of the available hardware; nonetheless, any of the three model variants may be utilized. 1. 1. Initiate with a bounding box prompt: our experimental results consistently indicate that among various prompting strategies, the bounding box technique exhibits superior performance, even in the presence of minor variations. Thus, we advocate that physicians start the segmentation procedure by supplying a bounding box prompt encompassing the region of interest. 2. 2. Evaluate the generated predictions: SAM generates a triplet of segmentation masks in response to an input image and a bounding box, each signifying a distinct interpretation of the intended region’s dimensions. Physicians are advised to visually scrutinize and juxtapose the three produced masks against the source image. If there is a suitable prediction, select it. If none of the predictions correctly segment the intended region, proceed to the next step. 3. 3. Refine the segmentation employing point prompts: in cases where none of the initial predictions adequately segment the intended region, assess the best prediction and identify the areas it incorrectly captures or omits in the segmentation. Utilize input points to include (label $1$) or exclude (label $0$) these areas. SAM will generate three new predictions. Repeat the process of refining the segmentation using point prompts until an adequate segmentation is achieved. Fig. 7 and Fig. 8 demonstrate the application of our proposed framework on images from the BUSI and CVC datasets, including the bounding box prompt and subsequent predictions. Since the intended regions were accurately segmented, the physician merely has to select them. Figure 7: Image from the BUSI dataset with bounding box input accompanied by SAM’s predictions. Both the 2nd and 3rd predictions exhibit accurate segmentation of the intended region. Figure 8: Image from the CVC dataset with bounding box input accompanied by SAM’s predictions. Both the 2nd and 3rd predictions exhibit accurate segmentation of the intended region. Fig. 9 presents the application of our framework on an image from the ISIC dataset, followed by the bounding box prompt and predictions. This represents a more intricate scenario, as discussed earlier. None of the predictions provided satisfactory results; therefore, the physician must evaluate the best one (2nd) and incorporate point prompts to guide the model. Fig 10 displays the original bounding box input in conjunction with the point prompts and the generated predictions. A significant improvement in segmentation is observable in the 2nd prediction due to the additional point prompts. Figure 9: Image from the ISIC dataset with bounding box input accompanied by SAM’s predictions. None of them are adequate and require further prompt points. Figure 10: Image from the ISIC dataset with bounding box and point inputs accompanied by SAM’s predictions. The point prompts guide the model to remove these areas. The 2nd prediction reached an adequate segmentation. This methodology ensures that the model’s output coheres with the physician’s expertise, culminating in accurate and dependable segmentation results across diverse clinical applications and imaging modalities.111A demo of this framework is available at https://github.com/Malta-Lab/SAM-zero-shot-in- Medical-Imaging. ## 5 Conclusion and Future Work In this study, we thoroughly evaluated the zero-shot performance of SAM by employing eight distinct prompting strategies across six datasets from four different 2D medical image modalities. Our comprehensive analysis shed light on the advantages and limitations of these strategies in various scenarios for the three SAM ViT sizes. Remarkably, SAM demonstrated exceptional performance as a zero-shot approach, achieving competitive results in comparison to the state-of-the-art segmentation methods specifically designed or fine-tuned for a particular modality of medical imaging. Notably, SAM outperformed the current best performance on the BUSI dataset by a substantial margin. Taken together, our findings underscore the immense potential of SAM as a powerful tool for low-effort medical image segmentation. Drawing upon our results, we propose pragmatic guidelines that facilitate easy implementation, necessitate minimal user interaction, and yield robust outcomes in medical imaging segmentation with SAM. By incorporating the bounding box method and refining the segmentation using point prompts, medical practitioners can effectively harness SAM’s potential to attain accurate results while maintaining control over the segmentation process. Furthermore, given the comparable performance of the three SAM sizes, practitioners can choose any of them based on their hardware resource constraints. The segmentation results generated by SAM have the potential to exceed the most stringent quality standards with minimal involvement from physicians. Our findings highlight concerns regarding the quality of some manually-annotated ground truth masks, as SAM outcomes appear to delineate the region of interest more accurately in certain instances. This observation holds particular significance for labeling new datasets, as it substantially reduces the time and effort required for this laborious and tedious task. Consequently, SAM- generated segmentation masks offer immense promise for streamlining data annotation processes and enhancing workflow efficiency in the area of medical image analysis. Future research endeavors could focus on further augmenting SAM’s capabilities in this domain, achieving even higher performance while preserving SAM’s extensive refinement options. Additionally, investigating the potential of adapting SAM for 3D medical imaging represents a valuable research direction, as it would extend the model’s applicability to a broader spectrum of medical imaging tasks. ## References * Wang et al. [2017] Shuo Wang, Mu Zhou, Zaiyi Liu, Zhenyu Liu, Dongsheng Gu, Yali Zang, Di Dong, Olivier Gevaert, and Jie Tian. Central focused convolutional neural networks: Developing a data-driven model for lung nodule segmentation. _Medical image analysis_ , 40:172–183, 2017. * Pham et al. [2000] Dzung L Pham, Chenyang Xu, and Jerry L Prince. Current methods in medical image segmentation. _Annual review of biomedical engineering_ , 2(1):315–337, 2000. * Ronneberger et al. [2015] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In _Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18_ , pages 234–241. Springer, 2015. * Liu et al. [2023] Jie Liu, Yixiao Zhang, Jieneng Chen, Junfei Xiao, Yongyi Lu, Bennett A. Landman, Yixuan Yuan, Alan L. Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. _CoRR_ , abs/2301.00785, 2023. doi: 10.48550/arXiv.2301.00785. URL https://doi.org/10.48550/arXiv.2301.00785. * Yang et al. [2023] Litao Yang, Chao Fan, Hao Lin, and Yingying Qiu. Rema-net: An efficient multi-attention convolutional neural network for rapid skin lesion segmentation. _Computers in Biology and Medicine_ , page 106952, 2023. * Litjens et al. [2017] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. _Medical image analysis_ , 42:60–88, 2017. * Dosovitskiy et al. [2020] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. _arXiv preprint arXiv:2010.11929_ , 2020. * Kirillov et al. [2023] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. _arXiv preprint arXiv:2304.02643_ , 2023. * Chowdhary and Acharjya [2020] Chiranji Lal Chowdhary and D Prasanna Acharjya. Segmentation and feature extraction in medical imaging: a systematic review. _Procedia Computer Science_ , 167:26–36, 2020. * Sharma and Aggarwal [2010] Neeraj Sharma and Lalit M Aggarwal. Automated medical image segmentation techniques. _Journal of medical physics_ , 35(1):3–14, 2010\. * Shirly and Ramesh [2019] S Shirly and K Ramesh. Review on 2d and 3d mri image segmentation techniques. _Current Medical Imaging_ , 15(2):150–160, 2019\. * Bencevic et al. [2021] Marin Bencevic, Irena Galic, Marija Habijan, and Danilo Babin. Training on polar image transformations improves biomedical image segmentation. _IEEE Access_ , 9:133365–133375, 2021. doi: 10.1109/ACCESS.2021.3116265. URL https://doi.org/10.1109/ACCESS.2021.3116265. * Henry et al. [2022] Emerald U. Henry, Onyeka Emebob, and Conrad Asotie Omonhinmin. Vision transformers in medical imaging: A review. _CoRR_ , abs/2211.10043, 2022. doi: 10.48550/arXiv.2211.10043. URL https://doi.org/10.48550/arXiv.2211.10043. * Vaswani et al. [2017] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, _Advances in Neural Information Processing Systems_ , volume 30. Curran Associates, Inc., 2017. * Touvron et al. [2021] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In _International conference on machine learning_ , pages 10347–10357. PMLR, 2021. * Wang et al. [2021] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In _Proceedings of the IEEE/CVF international conference on computer vision_ , pages 568–578, 2021. * Guo et al. [2022] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xupeng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. _arXiv preprint arXiv:2209.14169_ , 2022. * Pham et al. [2021] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V Le. Combined scaling for zero-shot transfer learning. _arXiv preprint arXiv:2111.10050_ , 2021. * Radford et al. [2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In _International conference on machine learning_ , pages 8748–8763. PMLR, 2021. * Jang et al. [2022] Jongseong Jang, Daeun Kyung, Seung Hwan Kim, Honglak Lee, Kyunghoon Bae, and Edward Choi. Significantly improving zero-shot x-ray pathology classification via fine-tuning pre-trained image-text encoders. _arXiv preprint arXiv:2212.07050_ , 2022. * Wang et al. [2022] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. _arXiv preprint arXiv:2210.10163_ , 2022. * Guo and Fan [2022] Fu-Ming Guo and Yingfang Fan. Zero-shot and few-shot learning for lung cancer multi-label classification using vision transformer. _arXiv preprint arXiv:2205.15290_ , 2022. * Codella et al. [2019] Noel C. F. Codella, Veronica Rotemberg, Philipp Tschandl, M. Emre Celebi, Stephen W. Dusza, David A. Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael A. Marchetti, Harald Kittler, and Allan Halpern. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (ISIC). _CoRR_ , abs/1902.03368, 2019. URL http://arxiv.org/abs/1902.03368. * Tschandl et al. [2018] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The HAM10000 dataset: A large collection of multi-source dermatoscopic images of common pigmented skin lesions. _CoRR_ , abs/1803.10417, 2018. URL http://arxiv.org/abs/1803.10417. * Tschandl et al. [2020] Philipp Tschandl, Christoph Rinner, Zoe Apalla, Giuseppe Argenziano, Noel Codella, Allan Halpern, Monika Janda, Aimilios Lallas, Caterina Longo, Josep Malvehy, John Paoli, Susana Puig, Cliff Rosendahl, H. Peter Soyer, Iris Zalaudek, and Harald Kittler. Human–computer collaboration for skin cancer recognition. _Nature Medicine_ , 26(8):1229–1234, June 2020\. doi: 10.1038/s41591-020-0942-0. URL https://doi.org/10.1038/s41591-020-0942-0. * Jaeger et al. [2014] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les R. Folio, Jenifer Siegelman, Fiona M. Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K. Singh, Sameer K. Antani, George R. Thoma, Yi-Xiang J. Wang, Pu-Xuan Lu, and Clement J. McDonald. Automatic tuberculosis screening using chest radiographs. _IEEE Trans. Medical Imaging_ , 33(2):233–245, 2014. doi: 10.1109/TMI.2013.2284099. URL https://doi.org/10.1109/TMI.2013.2284099. * Candemir et al. [2014] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P. Musco, Rahul K. Singh, Zhiyun Xue, Alexandros Karargyris, Sameer K. Antani, George R. Thoma, and Clement J. McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. _IEEE Trans. Medical Imaging_ , 33(2):577–590, 2014. doi: 10.1109/TMI.2013.2290491. URL https://doi.org/10.1109/TMI.2013.2290491. * Gut [2021] Daniel Gut. X-ray images of the hip joints, 2021. URL https://data.mendeley.com/datasets/zm6bxzhmfz/1. * Bernal et al. [2015] Jorge Bernal, Francisco Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez de Miguel, and Fernando Vilariño. WM-DOVA maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. _Comput. Medical Imaging Graph._ , 43:99–111, 2015. doi: 10.1016/j.compmedimag.2015.02.007. URL https://doi.org/10.1016/j.compmedimag.2015.02.007. * Al-Dhabyani et al. [2020] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. _Data in Brief_ , 28:104863, 2020. ISSN 2352-3409. doi: https://doi.org/10.1016/j.dib.2019.104863. URL https://www.sciencedirect.com/science/article/pii/S2352340919312181. * Kim and Lee [2021] Minki Kim and Byoung-Dai Lee. Automatic lung segmentation on chest x-rays using self-attention deep neural network. _Sensors_ , 21(2):369, 2021. * [32] Tarun Agrawal and Prakash Choudhary. Rese-net: Enhanced unet architecture for lung segmentation in chest radiography images. _Computational Intelligence_. * Zhan et al. [2023] Bangcheng Zhan, Enmin Song, and Hong Liu. Fsa-net: Rethinking the attention mechanisms in medical image segmentation from releasing global suppressed information. _Computers in Biology and Medicine_ , page 106932, 2023. * Podda et al. [2022] Alessandro Sebastian Podda, Riccardo Balia, Silvio Barra, Salvatore Carta, Gianni Fenu, and Leonardo Piano. Fully-automated deep learning pipeline for segmentation and classification of breast ultrasound images. _Journal of Computational Science_ , 63:101816, 2022. ## 6 Supplementary Material Table 5: DSC of predictions for six datasets using the eight proposed prompt strategies for the three SAM ViT sizes. Dataset | Model | Pred | CP | RP | RP3 | RP5 | BB | BBS5 | BBS10 | BBS20 ---|---|---|---|---|---|---|---|---|---|--- ISIC | ViT-H | 1st | 0.538 | 0.531 | 0.762 | 0.774 | 0.745 | 0.737 | 0.715 | 0.603 2nd | 0.718 | 0.677 | 0.769 | 0.788 | 0.845 | 0.842 | 0.833 | 0.789 3rd | 0.375 | 0.363 | 0.390 | 0.483 | 0.872 | 0.868 | 0.860 | 0.816 ViT-L | 1st | 0.704 | 0.665 | 0.703 | 0.700 | 0.864 | 0.861 | 0.852 | 0.805 2nd | 0.518 | 0.521 | 0.768 | 0.794 | 0.763 | 0.757 | 0.733 | 0.623 3rd | 0.382 | 0.366 | 0.358 | 0.362 | 0.841 | 0.836 | 0.819 | 0.730 ViT-B | 1st | 0.366 | 0.355 | 0.354 | 0.375 | 0.870 | 0.866 | 0.855 | 0.810 2nd | 0.665 | 0.618 | 0.692 | 0.695 | 0.825 | 0.823 | 0.807 | 0.751 3rd | 0.504 | 0.490 | 0.766 | 0.790 | 0.640 | 0.631 | 0.601 | 0.496 HAM | ViT-H | 1st | 0.544 | 0.527 | 0.752 | 0.765 | 0.732 | 0.724 | 0.700 | 0.589 | 2nd | 0.729 | 0.686 | 0.768 | 0.785 | 0.838 | 0.835 | 0.824 | 0.778 | 3rd | 0.420 | 0.406 | 0.443 | 0.541 | 0.865 | 0.861 | 0.851 | 0.809 ViT-L | 1st | 0.731 | 0.689 | 0.723 | 0.721 | 0.859 | 0.856 | 0.846 | 0.799 2nd | 0.522 | 0.518 | 0.764 | 0.793 | 0.766 | 0.761 | 0.740 | 0.626 3rd | 0.435 | 0.413 | 0.406 | 0.408 | 0.830 | 0.824 | 0.805 | 0.699 ViT-B | 1st | 0.414 | 0.403 | 0.401 | 0.425 | 0.863 | 0.859 | 0.846 | 0.799 2nd | 0.659 | 0.607 | 0.681 | 0.683 | 0.810 | 0.807 | 0.795 | 0.740 3rd | 0.478 | 0.431 | 0.749 | 0.772 | 0.619 | 0.610 | 0.578 | 0.466 CXR | ViT-H | 1st | 0.904 | 0.863 | 0.923 | 0.927 | 0.936 | 0.934 | 0.911 | 0.686 2nd | 0.758 | 0.727 | 0.766 | 0.828 | 0.942 | 0.939 | 0.929 | 0.826 3rd | 0.471 | 0.469 | 0.482 | 0.514 | 0.935 | 0.930 | 0.913 | 0.803 ViT-L | 1st | 0.834 | 0.814 | 0.786 | 0.776 | 0.932 | 0.929 | 0.916 | 0.805 2nd | 0.915 | 0.870 | 0.930 | 0.929 | 0.940 | 0.936 | 0.906 | 0.660 3rd | 0.472 | 0.471 | 0.468 | 0.474 | 0.945 | 0.942 | 0.928 | 0.758 ViT-B | 1st | 0.459 | 0.459 | 0.467 | 0.497 | 0.916 | 0.910 | 0.894 | 0.817 2nd | 0.804 | 0.782 | 0.786 | 0.803 | 0.937 | 0.933 | 0.921 | 0.813 3rd | 0.882 | 0.813 | 0.928 | 0.932 | 0.916 | 0.898 | 0.818 | 0.524 HJXR-F | ViT-H | 1st | 0.876 | 0.822 | 0.941 | 0.948 | 0.924 | 0.908 | 0.848 | 0.618 2nd | 0.743 | 0.767 | 0.767 | 0.776 | 0.962 | 0.958 | 0.904 | 0.746 3rd | 0.517 | 0.543 | 0.548 | 0.599 | 0.949 | 0.945 | 0.905 | 0.723 ViT-L | 1st | 0.773 | 0.800 | 0.788 | 0.791 | 0.972 | 0.969 | 0.951 | 0.843 2nd | 0.874 | 0.804 | 0.927 | 0.944 | 0.925 | 0.922 | 0.844 | 0.685 3rd | 0.516 | 0.540 | 0.540 | 0.619 | 0.961 | 0.944 | 0.818 | 0.448 ViT-B | 1st | 0.466 | 0.486 | 0.481 | 0.489 | 0.924 | 0.915 | 0.888 | 0.788 2nd | 0.733 | 0.775 | 0.742 | 0.727 | 0.958 | 0.954 | 0.926 | 0.771 3rd | 0.911 | 0.774 | 0.909 | 0.907 | 0.899 | 0.876 | 0.735 | 0.490 HJXR-I | ViT-H | 1st | 0.211 | 0.742 | 0.808 | 0.828 | 0.875 | 0.866 | 0.734 | 0.624 2nd | 0.393 | 0.479 | 0.449 | 0.491 | 0.855 | 0.849 | 0.790 | 0.620 3rd | 0.294 | 0.295 | 0.316 | 0.384 | 0.800 | 0.796 | 0.758 | 0.629 ViT-L | 1st | 0.363 | 0.540 | 0.448 | 0.451 | 0.824 | 0.817 | 0.748 | 0.594 2nd | 0.165 | 0.758 | 0.864 | 0.860 | 0.887 | 0.877 | 0.762 | 0.555 3rd | 0.301 | 0.306 | 0.292 | 0.330 | 0.862 | 0.841 | 0.733 | 0.580 ViT-B | 1st | 0.259 | 0.303 | 0.328 | 0.368 | 0.772 | 0.767 | 0.734 | 0.591 2nd | 0.403 | 0.502 | 0.467 | 0.478 | 0.849 | 0.843 | 0.802 | 0.615 3rd | 0.314 | 0.717 | 0.823 | 0.830 | 0.838 | 0.838 | 0.779 | 0.622 CVC | ViT-H | 1st | 0.716 | 0.763 | 0.861 | 0.880 | 0.889 | 0.881 | 0.835 | 0.702 2nd | 0.554 | 0.544 | 0.642 | 0.754 | 0.926 | 0.924 | 0.916 | 0.844 3rd | 0.232 | 0.224 | 0.224 | 0.245 | 0.924 | 0.922 | 0.918 | 0.868 ViT-L | 1st | 0.498 | 0.482 | 0.508 | 0.522 | 0.920 | 0.918 | 0.906 | 0.853 2nd | 0.702 | 0.728 | 0.836 | 0.841 | 0.873 | 0.867 | 0.818 | 0.672 3rd | 0.229 | 0.222 | 0.217 | 0.223 | 0.909 | 0.904 | 0.870 | 0.773 ViT-B | 1st | 0.234 | 0.225 | 0.222 | 0.226 | 0.920 | 0.916 | 0.906 | 0.833 2nd | 0.447 | 0.440 | 0.495 | 0.510 | 0.907 | 0.906 | 0.892 | 0.796 3rd | 0.644 | 0.688 | 0.778 | 0.783 | 0.821 | 0.810 | 0.758 | 0.585 BUSI | ViT-H | 1st | 0.583 | 0.541 | 0.736 | 0.766 | 0.754 | 0.744 | 0.713 | 0.631 2nd | 0.641 | 0.616 | 0.688 | 0.735 | 0.840 | 0.837 | 0.823 | 0.768 3rd | 0.192 | 0.184 | 0.196 | 0.254 | 0.863 | 0.859 | 0.848 | 0.800 ViT-L | 1st | 0.656 | 0.649 | 0.674 | 0.663 | 0.866 | 0.862 | 0.855 | 0.794 2nd | 0.567 | 0.536 | 0.748 | 0.779 | 0.782 | 0.777 | 0.754 | 0.649 3rd | 0.228 | 0.205 | 0.202 | 0.252 | 0.849 | 0.847 | 0.830 | 0.741 ViT-B | 1st | 0.202 | 0.192 | 0.181 | 0.213 | 0.884 | 0.881 | 0.869 | 0.823 2nd | 0.634 | 0.604 | 0.682 | 0.691 | 0.832 | 0.830 | 0.818 | 0.766 3rd | 0.562 | 0.522 | 0.773 | 0.797 | 0.725 | 0.722 | 0.689 | 0.582
# OV-DQUO: Open-Vocabulary DETR with Denoising Text Query Training and Open- World Unknown Objects Supervision Junjie Wang1 Bin Chen1, 2, 3 Bin Kang2 Yulin Li1 Yichi Chen2 Weizhi Xian3 Huifeng Chang4 1 Harbin Institute of Technology, Shenzhen 2 University of Chinese Academy of Sciences 3 Harbin Institute of Technology, Chongqing Research Institute 4 CECloud Computing Technology Co., Ltd <EMAIL_ADDRESS><EMAIL_ADDRESS> ###### Abstract Open-Vocabulary Detection (OVD) aims to detect objects from novel categories beyond the base categories on which the detector is trained. However, existing open-vocabulary detectors trained on known category data tend to assign higher confidence to trained categories and confuse novel categories with background. To resolve this, we propose OV-DQUO, an Open-Vocabulary DETR with Denoising text Query training and open-world Unknown Objects supervision. Specifically, we introduce a wildcard matching method that enables the detector to learn from pairs of unknown objects recognized by the open-world detector and text embeddings with general semantics, mitigating the confidence bias between base and novel categories. Additionally, we propose a denoising text query training strategy that synthesizes additional noisy query-box pairs from open-world unknown objects to trains the detector through contrastive learning, enhancing its ability to distinguish novel objects from the background. We conducted extensive experiments on the challenging OV-COCO and OV-LVIS benchmarks, achieving new state-of-the-art results of 45.6 AP50 and 39.3 mAP on novel categories respectively, without the need for additional training data. Models and code are released at https://github.com/xiaomoguhz/OV-DQUO ## 1 Introduction Open-Vocabulary Detection [37] focuses on identifying objects from novel categories not encountered during training. Recently, Vision-Language Models (VLMs)[22, 28, 16] pretrained on large-scale image-text pairs, such as CLIP[22], have demonstrated impressive performance in zero-shot image classification, providing new avenues for open-vocabulary detection. Figure 1: (a) Detector confidence bias is a primary reason for the suboptimal detection performance on novel categories. (b) Existing methods use VLM and RPN to generate pseudo region-text pairs from image-caption datasets. (c) Instead, this work leverages the open-world detector to recognize _novel unknown objects_ within the training set and learns to match them with wildcard text embeddings. ViLD [6] is the first work to distill VLMs’ classification knowledge into an object detector by aligning the detector-generated region embeddings with the corresponding features extracted from VLMs. Subsequent methods [32, 29, 34, 36, 15] have proposed more elaborately designed strategies to improve the efficiency of knowledge distillation, such as BARON [32], which aligns bag-of- regions embeddings with image features extracted by VLMs. However, the context discrepancy limits the effectiveness of knowledge distillation [44]. RegionCLIP [42] is a representative method that utilizes VLMs for pseudo- labeling by generating pseudo region-text pairs from caption datasets[26] using RPN and CLIP to train open-vocabulary detectors. Later works [2, 41, 40] have further extended the implementation of pseudo-labeling, such as SASDet [41], which explores leveraging a self-training paradigm for pseudo-labeling. Nevertheless, these methods suffer from pseudo-label noise. All of the above methods employ indirect utilization of VLMs, thus not unleashing their full potential. Existing state-of-the-art methods [33, 35, 14] typically deploy a frozen CLIP image encoder as the image backbone and perform open-vocabulary detection by extracting region features within the prediction box. Intuitively, the performance ceiling of such methods depends directly on the classification ability of VLMs. Therefore, current works mainly enhance VLM’s region recognition accuracy through fine-tuning [35] or self-distillation [33]. Yet, these methods overlook the fact that detectors trained on known category data tend to assign higher confidence to trained categories and confuse novel categories with background. To verify the impact of confidence bias on novel category detection, we first analyze the confidence assigned by VLMs and detectors to base and novel categories, as shown in Figure 1(a). It is evident that the detector assigns significantly lower confidence to novel category objects (e.g., umbrella) than to known categories (e.g., person). Furthermore, we observed a significant performance gap when using VLM to classify Ground Truth (GT) boxes compared to detector predictions. However, this gap narrows when we manually adjust the prediction confidence of bounding boxes based on their Intersection over Union (IoU) with GT boxes. The experimental results reveal that confidence bias is one of the factors responsible for suboptimal performance in novel category detection. Based on the above findings, we propose OV-DQUO, an open-vocabulary detection framework with denoising text query training and open-world unknown objects supervision. Unlike existing methods that generate pseudo region-text pairs (Figure 1(b)), our framework propose a wildcard matching method and a contrastive denoising training strategy to directly learn from open-world unknown objects, mitigating performance degradation in novel category detection caused by confidence bias. As shown in Figure 1(c), to address the confidence bias between base and novel categories, OV-DQUO first utilizes an open-world detector to recognize novel unknown objects within the training set. It then queries these unknown objects using text embeddings with general semantics (i.e., wildcard matching) and enables the detector to regard them as query-box match. Since the open-world detector cannot identify all novel unknown objects, we designed a denoising text query training strategy to address the detector’s confusion between novel categories and the background. This method synthesizes additional query-box pairs by perturbing bounding boxes of unknown objects and assigning noisy text embeddings, enabling OV-DQUO to leverage contrastive learning to better distinguish novel objects from the background. Finally, to mitigate the impact of confidence bias on region proposal selection, we propose RoQIs Selection, which integrates region-text similarity with confidence scores to select region proposals, achieving a more balanced recall of base and novel category objects. The main contributions of this paper can be summarized as follows: * • Inspired by the open-world detection task of recognizing novel unknown objects, we propose an OV-DQUO framework, which mitigates the detector’s confidence bias on novel category detection. * • We design a wildcard matching method, which enables the detector to learn from pairs of text embeddings with general semantics and novel unknown objects recognized by the open-world detector, thereby alleviating the confidence bias between base and novel categories. * • We introduce the denoising text query training strategy, which allows a detector to perform contrastive learning from synthetic noisy query-box pairs, thus enhancing its ability to distinguish novel objects from the background. * • OV-DQUS consistently outperforms existing state-of-the-art methods on the OV- COCO and OV-LVIS open-vocabulary detection benchmarks and demonstrates excellent performance in cross-dataset detection on COCO and Objects365. ## 2 Related Works Open-Vocabulary Detection (OVD) is a paradigm proposed by OVR-CNN [37], which aims to train models to detect objects from arbitrary categories, including those not seen during training. State-of-the-art methods [14, 35, 33] leverage a frozen VLM image encoder as the backbone to extract features and perform open-vocabulary detection. Compared to pseudo-labeling [1, 43, 42, 40, 21, 27] and knowledge distillation-based methods [32, 29, 34, 36, 15], these approaches directly benefit from the large-scale pretraining knowledge of VLMs and better generalize to novel objects. F-VLM [14] pioneered the discovery that VLMs retain region-sensitive features useful for object detection. It freezes the VLM and uses it as a backbone for feature extraction and region classification. CORA [35] also uses a frozen VLM but fine-tunes it with a lightweight region prompt layer, enhancing region classification accuracy. CLIPself [33] reveals that the ViT version of VLM performs better on image crops than on dense features, and explores aligning dense features with image crop features through self-distillation. However, we identify that these methods suffer from a confidence bias issue, leading to suboptimal performance in novel category detection. Open-World Detection (OWD) is a paradigm proposed by ORE[10], which aims to achieve two goals: (1) recognizing both known category objects and the unknown objects not present in the training set, and (2) enabling incremental object detection learning through newly introduced external knowledge. OW-DETR [8] attempts to identify potential unknown objects based on feature map activation scores, as foreground objects typically exhibit stronger activation responses compared to the background. PROB [45] performs distribution modeling on the model output logits to identify unknown objects and decouples the identification of background, known objects, and unknown objects. Based on the observation that foreground regions exhibit more variability while background regions change monotonously, MEPU [5] employs Weibull modeling on the feature reconstruction error of these regions and proposes the Reconstruction Error- based Weibull (REW) model. REW assigns likelihood scores to region proposals that potentially belong to unknown objects. These methods inspire us to leverage open-world detectors to address the confidence bias issue in OVD. ## 3 Methodology In this section, we present OV-DQUO, a novel OVD framework with denoising text query training and open-world unknown objects supervision. An overview is given in Figure 2. First, we briefly introduce the OVD setup. Then, we detail the open-world pseudo-labeling pipeline and the corresponding wildcard matching strategy, which is our key approach for mitigating the confidence bias between known and novel categories (Sec. 3.1). Subsequently, we elaborate the denoising text query training strategy that enhances a model ability to distinguish novel objects from the background (Sec. 3.2). Finally, we introduce the region of query interests selection module, which achieves a more balanced recall of base and novel category objects (Sec. 3.3). Task Formulation. In our study, we follow the classical open-vocabulary problem setup as in OVR-CNN [37]. In this setup, only partial class annotations of the dataset are available during the training process, commonly referred to as base classes and denoted by the symbol $\mathcal{C}^{\text{base}}$. During the inference stage, the model is required to recognize objects from both the base classes and the novel classes (denoted as $\mathcal{C}^{\text{novel}}$, where $\mathcal{C}^{\text{base}}\cap\mathcal{C}^{\text{novel}}=\varnothing$) that were not seen during training, while the names of the novel classes are provided as clues during inference. Figure 2: Overview of OV-DQUO. (a) Open-world pseudo labeling pipeline, which iteratively trains the detector, generates unknown object proposals, estimates and filters foreground probabilities, and updates the training set. (b) Denoising text query training, which enables contrastive learning with synthetic noisy query-box pairs from unknown objects. (c) RoQIs selection module, which considers both objectness and region-text similarity for selecting regions of query interest. ### 3.1 Open-World Pseudo Labeling & Wildcard Matching Unknown object proposals from the external OLN. Existing works [42, 6, 43, 41, 1, 2] leverage RPNs to mine potential novel objects, but these RPNs are biased towards the base classes they are trained on and perform poorly on novel classes. Unlike these approaches, we leverage the Object Localization Network (OLN) [11] to recognize novel unknown objects from the training set in the OV- DQUO framework, as shown in Figure 2(a). OLN is an open-world detector trained to estimate the objectness of each region purely based on how well the location and shape of a region overlap with any ground-truth object (e.g., centerness and IoU). After training OLN with $\mathcal{C}^{\text{base}}$ annotations from the OVD benchmark, we apply it to the training set to run inference and generate open-world unknown object proposals. Specifically, given an input image $I\in\mathbb{R}^{3\times H\times W}$, OLN outputs a series of tuples $\mathcal{R}=\\{r_{1},r_{2},\ldots,r_{n}\\}$, where each $r_{i}=[b_{i},q_{i}]$. Here, $b_{i}$ represents the coordinates of an unknown proposal, and $q_{i}$ denotes the localization quality estimations. Foreground likelihood estimation for novel unknown objects. Reducing the impact of noisy labels is a key challenge in pseudo-label learning. Inspired by [5], we leverage a probability distribution, which we denote as the Foreground Estimator (FE), to estimate the likelihood that a novel unknown object $r_{i}$ belongs to a foreground region. FE is based on the Weibull distribution and is modeled upon the feature reconstruction error of the foreground and background regions. Specifically, we first train a feature reconstruction network using images from the OVD benchmark in an unsupervised setting. Then, we collect the feature reconstruction errors for foreground and background regions based on the $\mathcal{C}^{\text{base}}$ annotations. Subsequently, we apply maximum likelihood estimation on Equation 1 to model the foreground and background Weibull distributions, denoted as $\mathcal{D}_{\textit{fg}}$ and $\mathcal{D}_{\textit{bg}}$, respectively. $\mathcal{D}(\eta|a,c)=ac\left[1-\exp\left(-\eta^{c}\right)\right]^{a-1}\exp\left(-\eta^{c}\right)\eta^{c-1}$ (1) where symbols $a$ and $c$ represent the shape parameters of the distribution, while $\eta$ represents the feature reconstruction error of the foreground or background region. With $\mathcal{D}_{\textit{fg}}$ and $\mathcal{D}_{\textit{bg}}$, we can estimate the foreground likelihood $w_{i}$ for each novel unknown object $r_{i}=[b_{i},q_{i}]$ in $\mathcal{R}$ using Equation 2, resulting in $\hat{\mathcal{R}}=\\{\hat{r}_{1},\hat{r}_{2},\ldots,\hat{r}_{n}\\}$, where $\hat{r}_{i}=[b_{i},q_{i},w_{i}]$. $w_{i}=\frac{\mathcal{D}_{\textit{fg}}\left(\eta({b_{i}})\right)}{\mathcal{D}_{\textit{fg}}\left(\eta({b_{i}})\right)+\mathcal{D}_{\textit{bg}}\left(\eta({b_{i}})\right))},$ (2) $\hat{\mathcal{R}}$ are used to update the training set annotations $\mathcal{C}^{\text{base}}$ after being filtered by ground truth annotations and some heuristic criteria. Once the training set is updated, it can be used to retrain OLN. Subsequently, the entire process can be iterated to yield more unknown objects, as shown in Figure 2(a). The visualization of unknown proposals and their corresponding foreground likelihood estimations are provided in Appendix A.3, and the details of the heuristic criteria can be found in Appendix A.6. Learning from open-world unknown objects via wildcard matching. The additional supervision signals provided by open-world detectors enable OV-DQUO to avoid treating novel objects as background during training, thereby mitigating the confidence bias between known and novel categories. However, applying an open- vocabulary training framework to open-world pseudo-labels raises the following challenges: open-world unknown objects lack category information. Unlike existing works [42, 41, 2] that re-label each proposal to specific categories using VLMs, we propose to match open-world unknown objects directly using text embeddings with general semantics, thereby avoiding additional label noise. Specifically, let $\mathcal{V}_{t}$ represent the text encoder of the VLM. The query text for unknown objects is "a photo of a $\\{\text{wildcard}\\}$", denoted as $T_{wc}$, where the wildcard can be terms like "object" or "thing." The query text for base classes is "a photo of a $\\{\mathcal{C}^{\text{base}}\\}$", denoted as $T_{base}$. In the learning process of pseudo-labels, if a region proposal $p_{i}$ generated by the OV- DQUO encoder has an IoU with any pseudo-label in $\hat{\mathcal{R}}$ greater than the threshold $\tau$, we assign the proposal with wildcard query embedding $\mathcal{V}_{t}(T_{wc})$; otherwise, we assign it the text embeddings of the base category with the maximum similarity, $\mathcal{V}_{t}(T_{base}^{*})$, as shown in the following equation: $(m_{i},\hat{p}_{i})=\operatorname{Decoder}(q_{i},p_{i}),\ \text{where}\ q_{i}=\begin{cases}\mathcal{V}_{t}(T_{wc})&\text{if }\operatorname{IoU}(p_{i},\hat{\mathcal{R}})>\tau,\\\ \mathcal{V}_{t}(T_{base}^{*})&\text{otherwise}.\end{cases}$ (3) where $\hat{\mathcal{R}}$ represents the set of open-world pseudo-labels. The decoder of OV-DQUO iteratively refines each query with its associated anchor box $(q_{i},p_{i})$ into output $o_{i}=(m_{i},\hat{p}_{i})$, where $m_{i}$ denotes the probability that the input query embedding matches the category of its corresponding bounding box, and $\hat{p}_{i}$ represents the predicted box. To achieve text query conditional matching, OV-DQUO constrains each ground-truth box to match predictions with the same category query embedding, including the pseudo-labels. Specifically, given a prediction set $\mathcal{O}^{wc}=\\{o_{1}^{wc},o_{2}^{wc},\ldots,o_{n}^{wc}\mid q_{i}=\mathcal{V}_{t}(T_{wc})\\}$ that is conditioned on wildcard query embedding, the class-aware Hungarian matching algorithm $\mathcal{H}_{\text{cls}}$ yields the optimal permutation $\mathcal{M}^{wc}=\\{(\hat{r}_{1},o_{1}^{wc}),(\hat{r}_{2},o_{2}^{wc}),\ldots,(\hat{r}_{k},o_{k}^{wc})\\}$ that minimizes the matching cost $\mathcal{L}_{\text{cost}}$ between the open- world pseudo-labels set $\hat{\mathcal{R}}$ and the predicted set $\mathcal{O}^{wc}$ as follows: $\mathcal{M}^{wc}=\mathcal{H}_{\text{cls}}\left(\hat{\mathcal{R}},\ \mathcal{O}^{wc},\ \mathcal{L}_{\text{cost}}\right),\ \text{where}\ \mathcal{L}_{\text{cost}}=\mathcal{L}_{\text{focal }}(m_{i}^{wc})+\mathcal{L}_{\text{bbox}}(\hat{p}_{i}^{wc},\hat{r}_{i})$ (4) $\mathcal{L}_{\text{focal}}$ denotes the binary focal loss [19], while $\mathcal{L}_{\text{bbox }}$ consists of L1 loss and GIoU loss [38]. With the matching results, the loss for unknown objects and base annotations can be expressed as follows: $\displaystyle\mathcal{L}_{pseudo}=\sum_{o_{i}^{wc}\in\mathcal{M}^{wc}}w_{i}\mathcal{L}_{\text{focal}}\left(m_{i}^{wc}\right),\quad\mathcal{L}_{base}=\sum_{c\in\mathcal{C}^{\text{base}}}\sum_{o_{i}^{c}\in\mathcal{M}^{c}}\left(\mathcal{L}_{\text{focal}}\left(m_{i}^{c}\right)+\mathcal{L}_{\text{bbox}}\left(\hat{p}_{i}^{c},y_{i}^{c}\right)\right)$ (5) where $o_{i}^{wc}=(m_{i}^{wc},\hat{p}_{i}^{wc})$ and $o_{i}^{c}=(m_{i}^{c},\hat{p}_{i}^{c})$ are the predictions selected by the Hungarian matching algorithm, whose query embeddings are $\mathcal{V}_{t}(T_{wc})$ and $\mathcal{V}_{t}(T_{c})$, respectively. $y_{i}^{c}$ represents a GT of base category $c$. $w_{i}$ is the foreground probability estimation of unknwn object $\hat{r}_{i}$. We only compute the $\mathcal{L}_{\text{focal}}$ for unknown objects. Additionally, the classification targets for predictions matched by $\mathcal{H}_{\text{cls}}$ are 1; otherwise, the target is 0. We omit them from the Equation 5 for simplicity. ### 3.2 Denoising Text Query Training Since the open-world detector cannot recognize all potential novel objects and provide supervision signals, we propose denoising text query training to enhance a detector’s ability to distinguish novel objects from the background. We achieve this by enabling OV-DQUO to perform contrastive learning from synthetic noisy query-box pairs, as shown in Figure 2(b). Specifically, for a given unknown object box $\hat{r}_{i}$, $2N$ noise proposals $\tilde{\mathcal{R}}=\\{\tilde{r}_{1},\tilde{r}_{2},\ldots,\tilde{r}_{2N}\\}$ are generated based on its coordinates with two noise scales $\lambda_{1}$ and $\lambda_{2}$, where $\lambda_{1}<\lambda_{2}$. Among these proposals, the first $N-1$ region proposals have a smaller noise scale $\lambda_{1}$ and are regarded as positive samples during training. In contrast, the remaining proposals from $N$ to $2N-1$ have a larger noise scale $\lambda_{2}$ and are treated as negative samples. For query embedding $q_{i}$, if a noisy region proposal $\tilde{r}_{i}$ belongs to the positive samples, we query it with the correct text embedding $\mathcal{V}_{t}(T_{wc})$. In contrast, for negative samples, we randomly select a proportion $\rho$ of samples and assign incorrect text embeddings of base categories $\mathcal{V}_{t}(T_{base})$, where $\rho$ is a noise scale parameter. The whole process is as follows: $\displaystyle\tilde{r}_{i}$ $\displaystyle=\begin{cases}\hat{r}_{i}+\lambda_{1}\cdot\epsilon(\hat{r}_{i}),&\text{if }0\leq i<N,\\\ \hat{r}_{i}+\lambda_{2}\cdot\epsilon(\hat{r}_{i}),&\text{otherwise}.\end{cases}\quad q_{i}$ $\displaystyle=\begin{cases}\mathcal{V}_{t}(T_{base}),&\text{if }N\leq i<2N\text{ and }R(i)<\rho,\\\ \mathcal{V}_{t}(T_{wc}),&\text{otherwise}.\end{cases}$ (6) where $R(i)\sim\text{Uniform}(0,1)$ is a random function, and $\epsilon$ is a function randomly calculates the offset based on input boxes. Denoising text query training utilizes contrastive learning by treating accurate bounding boxes with correct queries as positive samples, and bounding boxes that partially cover objects as negative samples, regardless of the query. The denoising part is performed simultaneously with the vanilla training part while using the attention mask for isolation. The denoise training loss and overall training objective for OV-DQUO can be expressed as follows: $\displaystyle\mathcal{L}_{denoise}=\sum_{i=0}^{2N}w_{i}\mathcal{L}_{\text{focal}}\left(\tilde{m}_{i},\mathbb{I}_{(0<i<N)}\right),\ \text{where}\ \tilde{m}_{i}=\operatorname{Decoder}(q_{i},\tilde{r}_{i})$ (7) $\displaystyle\mathcal{L}_{total}=\mathcal{L}_{pseudo}+\mathcal{L}_{base}+\mathcal{L}_{denoise}$ (8) where $\mathbb{I}_{(0<i<N)}$ is the indicator function, which equals 1 if $0<i<N$ and 0 otherwise. $\tilde{m}_{i}$ denotes the probability that query embedding $q_{i}$ matches the content within bounding box $\tilde{r}_{i}$. $\mathcal{L}_{pseudo}$ and $\mathcal{L}_{base}$ are vanilla pseudo-label learning loss and base category loss mentioned above. ### 3.3 Region of Query Interests Selection Existing two-stage OVD methods select region proposals based on either class- agnostic objectness[42, 33] or region-text similarity[20]. However, as we mentioned, objectness tends to favor the known categories. Region-text similarities exhibit less bias when leveraging a frozen VLM image encoder as the backbone, but they are insensitive to localization quality. As shown in Figure 2(c), we propose Region of Query Interests (RoQIs) selection, a novel method that considers both objectness and region-text similarity for selecting region proposals, achieving a more balanced recall of base and novel category objects. Specifically, given the region proposals set $\mathcal{R}$ and corresponding objectness score vector $O$, VLM feature map $\bm{\phi}$, and category name text embedding $\mathbf{L}$, the region of query interests set $\mathcal{R}^{*}$ for the next stage is generated as follows: $\mathcal{R}^{*}=\operatorname{gather}(\mathcal{R},t,N),\ \text{where }t=(\operatorname{max}(\operatorname{RoIAlign}(\mathcal{R},\bm{\phi})\cdot\mathbf{L}^{\top}))^{\bm{\alpha}}\cdot O^{(1-\bm{\alpha})}$ (9) where $\operatorname{gather}$ denotes the operation of selecting top-$N$ regions from $R$ accordi ng to $t$. RoIAlign[9] is a method used to obtain region features within a bounding boxes from the feature map $\bm{\phi}$. $\operatorname{max}$ means the maximum similarity of each region visual embeddings to all text embeddings. $\bm{\alpha}$ is the weighted geometric mean parameter. ## 4 Experiments ### 4.1 Dataset & Training & Evaluation OV-COCO benchmark. Following [37], we divide the 80 classes in the COCO dataset [18] into 48 base classes and 17 novel classes. In this benchmark, models are trained on the 48 base classes, which contain 107,761 images and 665,387 instances. Subsequently, the models are evaluated on the validation set, which includes both the base and novel classes, containing 4,836 images and 33,152 instances. For the OV-COCO benchmark, we use $\text{AP}_{50}^{\text{Novel}}$ as our evaluation metric, which calculates the mean average precision at an IoU of 50% for novel classes. OV-LVIS benchmark. Following standard practice [42, 6], we remove categories with rare tags in the LVIS [7] training set. Models are trained on 461 common classes and 405 frequent classes, which contain 100,170 images and 1,264,883 instances. After training, the models are evaluated on the validation set, which includes the common, frequent, and rare classes, containing 19,809 images and 244,707 instances. For the OV-LVIS benchmark, we use $\text{mAP}_{r}$ as our evaluation metric, which calculates the box AP averaged on IoUs from 0.5 to 0.95 for rare classes. ### 4.2 Implementation Details Model Specifications. OV-DQUO is built on the closed-set detector DINO [39]. To adapt it for the open-vocabulary setting, we follow the previous practice[36, 35] of modifying the decoder and letting it output matching probabilities conditioned on the input query. OV-DQUO is configured to have 1,000 object queries, 6 encoder layers, and 6 decoder layers. In the OV-COCO benchmark, we use CLIP of R50 and R50x4 [35] as the backbone networks. In the OV-LVIS benchmark, we use the self-distilled CLIP of ViT-B/16 and ViT-L/14 [33] as the backbone network. For the text embedding of each category, follow the previous works[35, 36, 33], we calculate the average representation of each category under 80 prompt templates using the text encoder of VLM, including the wildcard. We employ a MLP layer to transform the text embedding dimension of VLMs into 256. Training & Hyperparameters. We train OV-DQUO using 8 GPUs with a batch size of 4 on each GPU, using the AdamW optimizer with a learning rate of $1\mathrm{e}{-4}$ and a weight decay of $1\mathrm{e}{-4}$. To stabilize training, we evaluate on the exponential moving average (EMA) of the model after training. The cost hyperparameters for class, bbox, and GIoU in the Hungarian matching algorithm are set to 2.0, 5.0, and 2.0, respectively. More details about the model settings and training parameters of OV-DQUO and open- world pseudo labeling process can be found in Appendix A.5. ### 4.3 Benchmark Results OV-COCO. Table 1 summarizes the main results of OV-DQUO on the OV-COCO benchmark. To ensure a fair comparison, we detail the use of external training resources, backbone size, and access to novel class names during training for each method, as these factors vary from methods and significantly impact performance. It can be seen that OV-DQUO consistently outperforms all state- of-the-art methods in novel object detection, achieving the best results of 39.2/45.6 $\text{AP}_{50}^{\text{Novel}}$ with backbone networks of RN50/R50x4, respectively. Note that CLIPSelf[33] is based on the EVA version of CLIP[28], which is larger than our backbone and has stronger zero-shot classification capabilities. However, OV-DQUO still outperforms CLIPSelf by 1.3 AP50 on novel categories. Table 1: Comparison with state-of-the-art open-vocabulary object detection methods on OV-COCO. Caption supervision indicates that the method learns from extra image-text pairs, while CLIP supervision refers to transferring knowledge from CLIP. The column ’Novel’ specifies whether a method requires access to novel class names during training. †: implemented with the EVA version of CLIP[28]. P-L, R-AT, and KD-based are classifications of methods, denoting pseudo-labeling, region-aware training, and knowledge distillation- based approaches, respectively, as defined in [44]. Method Taxonomy Supervision Backbone Novel AP${}_{50}^{\text{Novel}}$ AP${}_{50}^{\text{Base}}$ AP${}_{50}^{\text{All}}$ ViLD[6] KD-based CLIP RN50 ✓ 27.6 59.9 51.3 Detic[43] P-L Caption[23] RN50 ✗ 27.8 47.1 42.0 OV-DETR[36] KD-based CLIP RN50 ✓ 29.4 61.0 52.7 RegionCLIP[42] P-L Caption[26] RN50 ✗ 31.4 57.1 50.4 VLDet[17] R-AT Caption[3] RN50 ✗ 32.0 50.6 45.8 MEDet[2] R-AT Caption[3] RN50 ✗ 32.6 54.0 49.4 BARON-KD[32] KD-based CLIP RN50 ✗ 34.0 60.4 53.5 VL-PLM[40] P-L CLIP RN50 ✓ 34.4 60.2 53.5 CLIM[34] KD-based CLIP RN50 ✗ 36.9 - - SAS-Det[41] P-L CLIP RN50x4 ✓ 37.4 58.5 53.0 RegionCLIP[42] P-L Captions[26] RN50x4 ✗ 39.3 61.6 55.7 CORA[35] R-AT CLIP RN50x4 ✗ 41.7 44.5 43.8 PromptDet[27] P-L Caption[24] ViT-B/16 ✗ 30.6 63.5 54.9 RO-ViT[13] R-AT CLIP ViT-L/16 ✗ 33.0 - 47.7 CFM-ViT[12] R-AT CLIP ViT-L/16 ✗ 34.1 - 46.0 CLIPSelf[33] KD-based CLIP ViT-B/16†(87M) ✗ 37.6 54.9 50.4 CLIPSelf[33] KD- based CLIP ViT-L/14†(304M) ✗ 44.3 64.1 59.0 OV-DQUO(Ours) P-L CLIP RN50(38M) ✗ 39.2 41.8 41.1 OV-DQUO(Ours) P-L CLIP RN50x4(87M) ✗ 45.6 49.0 48.1 Table 2: Comparison with state-of-the-art open-vocabulary object detection methods on OV-LVIS. Method Supervision Backbone $\text{mAP}_{r}$ ViLD[6] CLIP RN50 16.3 OV- DETR[36] CLIP RN50 17.4 BARON-KD[32] CLIP RN50 22.6 RegionCLIP[42] Caption[26] RN50x4 22.0 CORA+[35] Caption[23] RN50x4 28.1 F-VLM[14] CLIP RN50x64 32.8 CFM- ViT[12] CLIP ViT-L/14 33.9 RO-ViT[13] CLIP ViT-H/16 34.1 CLIPSelf[33] CLIP ViT-L/14 34.9 CoDet[21] Caption[26] ViT-L/14 37.0 OV-DQUO(Ours) CLIP ViT-B/16 29.7 OV-DQUO(Ours) CLIP ViT-L/14 39.3 Table 3: Cross-datasets transfer detection from OV-LVIS to COCO and Objects365. $\dagger$: Detection specialized pretraining with SoCo[31]. COCO Objects365 Method AP AP50 AP75 AP AP50 AP75 Supervised[6] 46.5 67.6 50.9 25.6 38.6 28.0 ViLD[6] 36.6 55.6 39.6 11.8 18.0 12.6 DetPro$\dagger$[4] 34.9 53.8 37.4 12.1 18.8 12.9 BARON[32] 36.2 55.7 39.1 13.6 21.0 14.5 RO-ViT[13] - - - 17.1 26.9 19.5 F-VLM[14] 37.9 59.6 41.2 16.2 25.3 17.5 CoDet[21] 39.1 57.0 42.3 14.2 20.5 15.3 OV-DQUO (Ours) 39.2 55.8 42.5 18.4 26.8 19.6 OV-LVIS. Table 3 summarizes the main results of OV-DQUO on the OV-LVIS benchmark. Since LVIS dataset encompasses considerably more categories than COCO (1203 vs. 80), we replaced the backbone network with stronger classification capabilities ViT-B/16 and ViT-L/14 [33] in the OV-LVIS experiments. It is worth noting that this does not lead to an unfair comparison, as OV-DQUO still consistently outperforms all state-of-the-art methods, including those using the same [33] (+4.4 $\text{mAP}_{r}$) or larger backbones [14] (+5.8 $\text{mAP}_{r}$), or using external image-caption data [21] (+2.3 $\text{mAP}_{r}$), achieving the best result of 39.3 $\text{mAP}_{r}$. Transfer to Other Datasets. Since the open-vocabulary detector may encounter data from different domains in open-world applications, we further evaluate OV-DQUO under a cross-dataset setting. Table 3 summarizes the main results of transferring OV-DQUO trained on OV-LVIS to the validation sets of COCO[18] and Object365[25]. We do not finetune OV-DQUO but only replace the text query embedding with the 80 categories in COCO and the 365 categories in Object365 during testing. Experiments show that OV-DQUO achieves competitive results on COCO and outperform the previous leading method[14] by 1.3 AP on Object365, demonstrating robust cross-dataset generalization. ### 4.4 Ablation Study Ablation Study on Main Components. As presented in Table 5, with the RN50x4 backbone, the vanilla OV-DQUO achieves 41.7 $\text{AP}_{50}$ on novel categories (#1). Additional supervision from open-world unknown objects boosts this to 43.3 $\text{AP}_{50}$ (#2). Furthermore, adding denoising text query training brings an additional 1.7 $\text{AP}_{50}$ performance gain (#3), demonstrating its effectiveness in improving discriminability between novel categories and backgrounds. Finally, RoQIs selection contributes another 0.6 $\text{AP}_{50}$ to the novel categories (#5). Effects of Matching Different Wildcards. As presented in Table 5, we explore matching different wildcard text embeddings with open-world unknown objects. In addition to "Object", we select several words that can represent general foreground regions, such as "Salient Object", "Foreground Region", "Target", and "Thing", and investigate their impact on performance. Experimental results demonstrate that compared to intricate wildcards ("Foreground Region","Salient Object"), simpler and more general wildcards ("Thing","Object") can achieve better results. Table 4: Ablation study on the main effective components of OV-DQUO. # Open-World Supervision Denoising Text Query Training RoQIs Selection $\text{AP}_{50}^{\text{Novel}}$ $\text{AP}_{50}^{\text{Base}}$ $\text{AP}_{50}^{\text{All}}$ 1 - - - 41.7 48.1 46.4 2 ✓ ✗ ✗ 43.3 46.8 45.8 3 ✓ ✓ ✗ 45.0 49.0 47.9 4 ✗ ✗ ✓ 42.7 48.0 46.6 5 ✓ ✓ ✓ 45.6 49.0 48.1 Table 5: Ablation study on matching different wildcards with unknown objects. Wildcard $\text{AP}_{50}^{\text{Novel}}$ $\text{AP}_{50}^{\text{Base}}$ $\text{AP}_{50}^{\text{All}}$ "Salient Object" 44.4 47.9 47.0 "Foreground Region" 44.1 47.7 46.7 "Target" 44.5 48.6 47.5 "Thing" 44.9 48.0 47.2 "Object" 45.0 48.9 47.9 (a) Baseline Detector[35] (b) OV-DQUO Figure 3: Confidence score distributions (c) Baseline Detector[35] (d) OV-DQUO Figure 4: Embedding distributions Visualization Analysis of OV-DQUO. We visualize the prediction results of OV- DQUO and the baseline detector[35] in Figures 4 and 4, including their output confidence distributions and output embedding T-SNE results. As shown in Figure 4, compared to the baseline detector, OV-DQUO outputs a more balanced confidence distribution between novel and base classes. Additionally, the confidence distribution predicted by OV-DQUO for both base and novel classes has less overlap with the background confidence distribution. As shown in Figure 4, compared to the baseline detector, the embeddings of novel category object output by OV-DQUO exhibit better discriminability from background embeddings. The comparison of the confidence distributions between OV-DQUO and baseline detector for each novel category can be found in the Appendix A.1. Wildcard Matching .vs Relabeling. We further compare wildcard matching with existing relabeling methods [35, 42] to evaluate its superiority. Specifically, we compare it with two methods: (1) relabeling each unknown object with the most similar novel category; and (2) forcibly relabeling each unknown object with the most similar base category. As presented in Table 7, experiments show that pairing each open-world unknown object with a specific category leads to suboptimal results. We believe that this outcome arises because open-world unknown objects include many foreground objects that do not belong to base or novel categories. Forcing these objects into specific pairings introduces considerable noise during training. Conversely, matching such foreground objects with wildcard text embeddings prevents model misguidance. Effects of Different Region Proposal Selection Strategies. We explore the impact of different region proposal selection strategies on performance, including objectness, region-text similarity, and RoQIs selection. As shown in Table 7, selecting proposals based on objectness score result in the recall of regions biased towards base categories. Besides, selecting proposals based on region-text similarity tends to recall regions with low localization quality, leading to performance degradation. Consequently, fusing objectness with region-text similarity achieves best results. Table 6: Ablation study on wildcard matching and relabeling methods Match Method $\text{AP}_{50}^{\text{Novel}}$ $\text{AP}_{50}^{\text{Base}}$ $\text{AP}_{50}^{\text{All}}$ Base classes Relabeling 42.4 48.7 47.0 Novel classes Relabeling 42.9 47.4 46.2 Wildcard Matching 45.0 48.9 47.9 Table 7: Ablation study on different proposal selection strategies Selection Strategy $\text{AP}_{50}^{\text{Novel}}$ $\text{AR}_{50}^{\text{Base}}$ $\text{AR}_{50}^{\text{Novel}}$ Objectness Selection 41.7 72.4 69.9 Region-Text Similarity 29.7 58.6 69.3 RoQIs Selection 42.7 72.1 70.5 Ablation Study on Hyperparameters. We explored the impact of different hyperparameter settings in OV-DQUO on performance, including the number of open-world pseudo-labeling iterations $t$, the weight $\bm{\gamma}$ for scaling the foreground likelihood score, and the weight of the denoising loss $\bm{\beta}$. Table 10 shows the ablation study on pseudo-labeling iteration $t$. We calculated the recall for objects in the COCO training set after each pseudo-labeling iteration as a reference. Experimental results indicate that OV-DQUO achieves optimal results when $t$ equals 2. Although recall increases with more iterations, the introduced noise starts to reduce the model performance on novel categories. Table 10 presents the ablation study on scaling the foreground score. We use the power function $(w_{i})^{\bm{\gamma}}$ to scale the foreground likelihood score for each unknown object, where $\bm{\gamma}$ controls the degree of scaling. When $\bm{\gamma}$ is set to 0, it serves as an ablation for the FE module. Results show that setting $\bm{\gamma}$ to 0 significantly degrades performance due to the release of pseudo-label noise. The best performance is achieved when $\bm{\gamma}$ is set to 0.5. Table 10 presents the ablation study on the weight of the denoising loss $\bm{\beta}$. Experimental results show that changing the weight of the denoising loss does not significantly affect performance. Moreover, the best results on novel categories are achieved when the denoising loss weight equals the classification loss weight, i.e., $\bm{\beta}=2$. Table 8: Ablation study on pseudo-labeling iterations $t$ $\text{AR}_{50}^{\text{All}}$ $\text{AP}_{50}^{\text{Novel}}$ $\text{AP}_{50}^{\text{All}}$ - 80.2 41.7 46.4 1 85.7 44.0 47.9 2 86.5 45.0 47.9 3 87.1 44.8 48.5 Table 9: Ablation study on scaling foreground score $\bm{\gamma}$ $\text{AP}_{50}^{\text{Novel}}$ $\text{AP}_{50}^{\text{Base}}$ $\text{AP}_{50}^{\text{All}}$ 0.0 43.0 47.4 46.2 0.5 45.0 48.9 47.9 1.0 44.4 48.3 47.3 2.0 44.1 47.7 46.7 Table 10: Ablation study on denoising loss weight $\bm{\beta}$ $\text{AP}_{50}^{\text{Novel}}$ $\text{AP}_{50}^{\text{Base}}$ $\text{AP}_{50}^{\text{All}}$ 1.0 44.8 48.3 47.4 2.0 45.0 48.9 47.9 3.0 44.4 48.9 47.7 4.0 44.4 48.6 47.5 ## 5 Limitations and Conclusions In this paper, we reveal that confidence bias constrains the novel category detection of existing OVD methods. Inspired by open-world detection tasks that identify unknown objects, we introduce an OV-DQUO framework to address this bias, which achieves new state-of-the-art results on various OVD benchmarks. While integrating OVD with OWD into a unified end-to-end framework is promising, it remains under-explored here and reserved for future research. ## References * Bangalath et al. [2022] Hanoona Bangalath, Muhammad Maaz, Muhammad Uzair Khattak, Salman H Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. _Advances in Neural Information Processing Systems_ , 35:33781–33794, 2022. * Chen et al. [2022] Peixian Chen, Kekai Sheng, Mengdan Zhang, Mingbao Lin, Yunhang Shen, Shaohui Lin, Bo Ren, and Ke Li. Open vocabulary object detection with proposal mining and prediction equalization. _arXiv preprint arXiv:2206.11134_ , 2022. * Chen et al. [2015] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. _arXiv preprint arXiv:1504.00325_ , 2015. * Du et al. [2022] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 14084–14093, 2022. * Fang et al. [2023] Ruohuan Fang, Guansong Pang, Lei Zhou, Xiao Bai, and Jin Zheng. Unsupervised recognition of unknown objects for open-world object detection. _arXiv preprint arXiv:2308.16527_ , 2023. * Gu et al. [2021] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. _arXiv preprint arXiv:2104.13921_ , 2021. * Gupta et al. [2019] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 5356–5364, 2019. * Gupta et al. [2022] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 9235–9244, 2022. * He et al. [2017] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In _Proceedings of the IEEE international conference on computer vision_ , pages 2961–2969, 2017. * Joseph et al. [2021] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 5830–5840, 2021. * Kim et al. [2022] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In So Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. _IEEE Robotics and Automation Letters_ , 7(2):5453–5460, 2022. * Kim et al. [2023a] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In _2023 IEEE/CVF International Conference on Computer Vision (ICCV)_ , pages 15556–15566, 2023a. doi: 10.1109/ICCV51070.2023.01430. * Kim et al. [2023b] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 11144–11154, 2023b. * Kuo et al. [2022] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. _arXiv preprint arXiv:2209.15639_ , 2022. * Li et al. [2023a] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In _Proceedings of the IEEE/CVF International Conference on Computer Vision_ , pages 6501–6510, 2023a. * Li et al. [2023b] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 23390–23400, 2023b. * Lin et al. [2022] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. _arXiv preprint arXiv:2211.14843_ , 2022. * Lin et al. [2014] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In _Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13_ , pages 740–755. Springer, 2014. * Lin et al. [2017] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In _Proceedings of the IEEE international conference on computer vision_ , pages 2980–2988, 2017. * Liu et al. [2023] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. _arXiv preprint arXiv:2303.05499_ , 2023. * Ma et al. [2023] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In A. Oh, T. Neumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, _Advances in Neural Information Processing Systems_ , volume 36, pages 71078–71094. Curran Associates, Inc., 2023. * Radford et al. [2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, _Proceedings of the 38th International Conference on Machine Learning_ , volume 139 of _Proceedings of Machine Learning Research_ , pages 8748–8763. PMLR, 18–24 Jul 2021. * Ridnik et al. [2021] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. _arXiv preprint arXiv:2104.10972_ , 2021. * Schuhmann et al. [2021] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. _arXiv preprint arXiv:2111.02114_ , 2021. * Shao et al. [2019] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In _Proceedings of the IEEE/CVF international conference on computer vision_ , pages 8430–8439, 2019. * Sharma et al. [2018] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In _Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , pages 2556–2565, 2018. * Song and Bang [2023] Hwanjun Song and Jihwan Bang. Prompt-guided transformers for end-to-end open-vocabulary object detection. _arXiv preprint arXiv:2303.14386_ , 2023. * Sun et al. [2023] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. _arXiv preprint arXiv:2303.15389_ , 2023. * Wang et al. [2023] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 11186–11196, 2023. * Wang et al. [2022] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 14176–14186, 2022. * Wei et al. [2021] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning pretraining for detection via object-level contrastive learning. _Advances in Neural Information Processing Systems_ , 34:22682–22694, 2021. * Wu et al. [2023a] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 15254–15264, 2023a. * Wu et al. [2023b] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. _arXiv preprint arXiv:2310.01403_ , 2023b. * Wu et al. [2024] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Wentao Liu, and Chen Change Loy. Clim: Contrastive language-image mosaic for region representation. In _Proceedings of the AAAI Conference on Artificial Intelligence_ , volume 38, pages 6117–6125, 2024. * Wu et al. [2023c] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 7031–7040, 2023c. * Zang et al. [2022] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In _European Conference on Computer Vision_ , pages 106–122. Springer, 2022. * Zareian et al. [2021] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 14393–14402, 2021. * Zhai et al. [2020] Hongyu Zhai, Jian Cheng, and Mengyong Wang. Rethink the iou-based loss functions for bounding box regression. In _2020 IEEE 9th joint international information technology and artificial intelligence conference (ITAIC)_ , volume 9, pages 1522–1528. IEEE, 2020. * Zhang et al. [2022] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. _arXiv preprint arXiv:2203.03605_ , 2022. * Zhao et al. [2022] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In _European Conference on Computer Vision_ , pages 159–175. Springer, 2022. * Zhao et al. [2024] Shiyu Zhao, Samuel Schulter, Long Zhao, Zhixing Zhang, Vijay Kumar B. G, Yumin Suh, Manmohan Chandraker, and Dimitris N. Metaxas. Taming self-training for open-vocabulary object detection, 2024. * Zhong et al. [2022] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 16793–16803, 2022. * Zhou et al. [2022] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krähenbühl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In _European Conference on Computer Vision_ , pages 350–368. Springer, 2022. * Zhu and Chen [2023] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. _arXiv preprint arXiv:2307.09220_ , 2023. * Zohar et al. [2023] Orr Zohar, Kuan-Chieh Wang, and Serena Yeung. Prob: Probabilistic objectness for open world object detection. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 11444–11453, 2023. ## Appendix A Appendix / supplemental material ### A.1 Visualization Result of Confidence Distribution for OV-DQUO and Baseline Detector (a) Airplane (b) Bus (c) Cake (d) Cat (e) Couch (f) Cow (g) Cup (h) Dog (i) Elephant (j) Keyboard (k) Knife (l) Sink Figure 5: Visualization result of confidence distribution for OV-DQUO and baseline detector As shown in Figure 5, we present details on the differences in confidence distribution between OV-DQUO and the baseline detector [35] when detecting novel categories. The data is derived from their predictions on the OV-COCO validation set. The experimental results indicate that, for novel categories such as airplanes, buses, cats, and dogs, the high-density region of the confidence distribution for OV-DQUO lies between 0.6 and 0.8, in sharp contrast to the baseline detector. This indicates that OV-DQUO benefits from the additional supervision signals provided by the open-world detector. Additionally, we observed that for novel categories such as keyboards, knives, and sinks, the high-density region of the confidence distribution for OV-DQUO is around 0.4. These category objects share the characteristic of being small and typically not the salient objects within an image, which makes them difficult for the open-world detector to recognize. However, through denoising text query training, the confidence for these category objects still exhibits superiority compared to the baseline detector. ### A.2 Model Performance Analysis We provide more details in Table 11 regarding using VLM to classify GT boxes, classify detector predictions, and classify detector predictions with IoU confidence. It is evident that compared with existing methods, our method significantly improves the detection performance of novel categories and narrows the gap with the experimental group that uses IoU as the confidence. Simultaneously, we observed an improvement in the detection performance of known categories. We attribute this to the model learning from open-world pseudo-labels and denoising training, which enhances its ability to distinguish foreground objects from the background. However, there is still a gap between our method and the group that uses IoU as confidence. We believe that false positive detections caused by the similarity between category text embeddings are the primary reason for this phenomenon. We will explore this issue in future work. Table 11: Performance analysis on the OV-COCO validation set with backbone networks RN50 and RN50x4. Method Backbone $\text{AP}_{50}^{\text{Novel}}$ $\text{AP}_{50}^{\text{Base}}$ Ground Truth RN50 65.1 70.0 IoU Confidence 52.5 58.6 CORA[35] 35.1 35.5 OV- DQUO 39.2 41.8 Method Backbone $\text{AP}_{50}^{\text{Novel}}$ $\text{AP}_{50}^{\text{Base}}$ Ground Truth RN50x4 74.1 76.0 IoU Confidence 59.1 63.7 CORA[35] 41.7 44.5 OV- DQUO 45.6 49.0 ### A.3 Visualization of Open-World Object Proposals Figure 6: Visualization of open-world pseudo-labels. The first row shows the base category annotations from the OV-COCO training set, with missing novel category objects marked by dashed boxes for each image. The second row displays the open-world object proposals generated by OLN. The third row presents the foreground likelihood estimation results from FE for each unknown object proposal. OV-DQUO mitigates the confidence bias issue between base and novel categories by learning from open-world unknown objects. Additionally, to avoid pseudo- label noise misleading the OV-DQUO training process, we follow the OWD method and use a foreground estimator to assign weights to each open-world unknown object. In Figure 6, we visualize these open-world unknown objects along with their corresponding foreground likelihood scores. The visualization results show that the open-world detector can identify most of the novel category objects. Additionally, we observe that the output of the detector also includes some non-object areas, such as distant trees and buildings. Furthermore, it can be seen that the foreground estimator is able to assign discriminative weights to foreground objects and non-object regions, which is key to avoiding model degradation. ### A.4 Visualization of Detection Results We show the detection results of OV-DQUO on OV-COCO and OV-LVIS validation set in Figure 7 and Figure 8, respectively. On OV-COCO dataset, OV-DQUO correctly detects novel categories including couch, dog, bus, cow, scissors, and so on. On LVIS dataset, OV-DQUO detects rare categories like salad plate, fedora hat, gas mask and so on. In Figure 9, we also present the results of applying the LVIS-trained OV-DQUO to the Objects365 dataset. We observe that OV-DQUO trained on OV-LVIS is capable of accurately identifying a broad spectrum of object concepts specified in the Objects365 dataset, showcasing remarkable generalization ability. Figure 7: Visualization of detection results on OV-COCO. Red boxes are for novel categories, while blue boxes are for base categories. Figure 8: Visualization of detection results on OV-LVIS. Red boxes are for rare categories, while blue boxes are for common and frequent categories. Figure 9: Visualization of transfer detection results on Objects365[25] dataset. ### A.5 Details of OV-DQUO Hyper-Parameter Configuration Detail setting for OV-DQUO. Following previous work [35], we set the exponential moving average factor to 0.99996. The hyperparameters for the matching cost are identical to the corresponding loss coefficients. During inference, the temperature $\tau$ of the classification logits is set to 0.01. Additionally, we multiply the logits of novel classes by a factor of 3.0. There are slight differences in specific parameter settings between our experiments on the OV-COCO and OV-LVIS datasets. These differences include the number of training epochs, image processing resolution, and the application of repeat factor sampling, among other parameters. Detailed configurations are provided in Table 12. Detail setting for open-world pseudo labeling. Following previous work [5], we train OLN using 8 GPUs with a batch size of 2 per GPU. The models are initialized with SoCo weights [31] and trained for 70,000 iterations using the SGD optimizer with a learning rate of $2\times 10^{-2}$. FE are trained for 3,000 iterations with a learning rate of $2\times 10^{-7}$ and a total batch size of 16. The training of OLN and FE adheres to the settings of OV-COCO and OV-LVIS, where annotations for novel classes and rare categories are removed. Following [5], we use the region proposals generated by FreeSoLo [30] as the initial unknown object annotations. Table 12: Experimental configurations of OV-DQUO for OV-COCO and OV-LVIS experiments. Configuration OV-COCO OV-LVIS Training epochs 30 35 Repeat factor sampling No Yes Image resolution 1333 $\times$ 800 1024 $\times$ 1024 / 896 $\times$ 896 Text embedding dimensions 1024 / 640 512 / 768 Multi-scale features ResNet (C3, C4) ViT (5, 7, 11) / (10, 14, 23) Sample categories No 100 Pseudo-label iterations 2 3 ### A.6 Criterion Details for Filtering Open-World Object Proposals In this section, we detail the process of filtering open-world object proposals generated by the open-world detector. The specific steps are as follows: * • Perform non-maximum suppression based on localization quality with a threshold of 0.3. * • Ensure that the box size exceeds 2000 pixels. * • Maintain an aspect ratio between 0.25 and 4.0. * • Ensure that the Intersection over Union with base category objects is less than 0.3.
#### Continuous Uniform By definition $\displaystyle\psi_{P}^{*}(y)=\sup\left\\{y\theta-\log\left(M_{P}[\theta]\right):\theta\in\real\right\\},$ where for $a<b$ we have that $\displaystyle M_{P}[\theta]=\begin{cases}\frac{\exp(b\theta)-\exp(a\theta)}{(b-a)\theta},&\theta\neq 0,\\\ 1,&\theta=0.\end{cases}$ Since $M_{P}[\theta]$ is continuous at zero, then, without loss of generality, we obtain $\displaystyle\begin{array}[]{rl}\psi_{P}^{*}(y)&=\sup\left\\{y\theta-\log\left(\frac{\exp(b\theta)-\exp(a\theta)}{(b-a)\theta}\right):\theta\in\real\right\\}\\\ &=\sup\left\\{(y-b)\theta-\log\left(\frac{1-\exp(-(b-a)\theta)}{(b-a)\theta}\right):\theta\in\real\right\\}\\\ &=\sup\left\\{(y-a)\theta-\log\left(\frac{\exp((b-a)\theta)-1}{(b-a)\theta}\right):\theta\in\real\right\\}\\\ &=\sup\left\\{(y-\mu)\theta-\log\left(\frac{\exp((b-\mu)\theta)-\exp((a-\mu)\theta)}{(b-a)\theta}\right):\theta\in\real\right\\}.\end{array}$ (A.13) where $\mu=(a+b)/2$. If $y\geq b$ then from the second formulation above we can conclude that $\psi_{P}^{*}(y)=\infty$ by taking $\theta\rightarrow\infty$. Similarly, if, $y\leq a$, then from the third formulation above we can conclude that $\psi_{P}^{*}(y)=\infty$ by taking $\theta\rightarrow-\infty$. If $y=\mu$ then the last formulation of (A.13) can be written as $\displaystyle\sup\left\\{-\log\left(\frac{\exp(\gamma\theta)-\exp(-\gamma\theta)}{2\gamma\theta}\right):\theta\in\real\right\\}=-\log\left(\inf\left\\{\phi(\theta):\theta\in\real\right\\}\right),$ where $\gamma:=(b-a)/2>0$ and $\displaystyle\phi(\theta):=\begin{cases}\frac{\exp(\gamma\theta)-\exp(-\gamma\theta)}{2\gamma\theta},&\theta\neq 0,\\\ 1,&\theta=0.\end{cases}$ By using L’Hôpital’s rule and some straightforward arguments, it is easy to verify that $\displaystyle\lim\limits_{|\theta|\rightarrow+\infty}\phi(\theta)=+\infty,\quad\lim\limits_{|\theta|\rightarrow 0}\phi(\theta)=1\quad\text{and}\quad\phi(\theta)=\phi(-\theta).$ Thus, $\phi$ is continuous at zero (which justifies its definition), coercive and symmetric. Since the log-normalizer function $\psi_{P}(\theta)=\log\left(M_{P}[\theta]\right)$ is strictly convex we can conclude that if a solution exists it must be unique. The coercivity of $\phi$ implies that a solution exists, and due to the symmetry of $\phi$ we can conclude that it must be zero. To summarize, in this case, $\psi_{P}^{*}(\mu)=0$ (with $\theta=0$). If $y\neq\mu$ such that $a<y<b$ then the optimal solution to (A.13) is nonzero and by the first-order optimality conditions it must satisfy $\displaystyle y-\frac{b\exp(b\theta)-a\exp(a\theta)}{\exp(b\theta)-\exp(a\theta)}+\frac{1}{\theta}=0.$ (A.14) Therefore, using (A.13) we can summarize that for $y\in(a,b)=\mathrm{dom}\,{\psi}_{P}^{*}$: $\displaystyle\psi_{P}^{*}(y)=\begin{cases}0,&y=\mu,\\\ (y-\mu)\theta-\log\left(\frac{\exp((b-\mu)\theta)-\exp((a-\mu)\theta)}{(b-a)\theta}\right),&y\neq\mu,\end{cases}$ where $\theta$ is the root of (A.14). #### Logistic The moment generating function for Logistic distribution with location and scaling parameters $\mu$ and $s>0$, respectively, is given by $\displaystyle M_{P}[\theta]=\exp(\mu y)B(1-s\theta,1+s\theta),\qquad s\theta\in(-1,1),$ where $B(\cdot,\cdot)$ stands for the _Beta function_ $\displaystyle B(\alpha,\beta)=\int_{0}^{1}t^{\alpha-1}(1-t)^{\beta-1}dt.$ The beta function and the closely related _gamma function_ $\displaystyle\Gamma(\alpha)=\int_{0}^{\infty}t^{\alpha-1}\exp(-t)dt,\qquad\alpha>0,$ share the following well-known relation $\displaystyle B(\alpha,\beta)=\frac{\Gamma(\alpha)\Gamma(\beta)}{\Gamma(\alpha+\beta)}.$ (A.15) The gamma function is an extension of the factorial as for a positive integer $\alpha$ it holds that $\Gamma(\alpha)=(\alpha-1)!$. In the following, we will use the well-known function equations $\displaystyle B(\alpha+1,\beta)=B(\alpha,\beta)\frac{\alpha}{\alpha+\beta},$ (A.16) and $\displaystyle B(\alpha,1-\alpha)=\Gamma(1-\alpha)\Gamma(\alpha)=\frac{\pi}{\sin(\pi\alpha)},\qquad\alpha\notin\mathbb{Z}.$ (A.17) The latter is known as Euler’s reflection formula or Euler’s function equation. Further details and proofs for both (A.16) and (A.17) can be found, for example, in [2]. Since $s\theta\in(-1,1)$, the above relations imply that for any $\theta\neq 0$ $\displaystyle\phi_{s}(\theta):=B(1-s\theta,1+s\theta)\overset{\eqref{apndx:eq:beta_func_1}}{=}B(-s\theta,1+s\theta)\frac{-s\theta}{-s\theta+1+s\theta}\overset{\eqref{apndx:eq:beta_eulers_reflection}}{=}\frac{-\pi s\theta}{\sin(-\pi s\theta)}.$ For $\theta=0$ we can verify by (A.15) that $\displaystyle\phi_{s}(\theta)=B_{s}(1-s\theta,1+s\theta)=1.$ Thus, we can summarize $\displaystyle\phi_{s}(\theta)=B(1-s\theta,1+s\theta)=\begin{cases}1,&s\theta=0,\\\ \frac{-\pi s\theta}{\sin(-\pi s\theta)},&s\theta\in(-1,1)\setminus\\{0\\}.\end{cases}$ (A.18) Using L’Hôpital’s rule we can verify that $\phi_{s}$ is continuous at $\theta=0$. Since $-\pi s\theta\geq\sin(-\pi s\theta)$ for all $s\theta\in(-1,1)$ we can conclude that $\phi_{s}(\theta)\geq 1$ for all $s\theta\in(-1,1)$ and equality ($\phi_{s}(\theta)=1$) holds if and only if $s\theta=1$. Taking $|s\theta|\rightarrow 1$ it is evident that $\phi_{s}(\theta)\rightarrow\infty$. In addition, for any $\theta\neq 0$ the derivative of $\phi$ is given by $\displaystyle\phi_{s}^{\prime}(\theta)=-\pi s\left[\frac{\sin(-\pi s\theta)+\pi s\theta\cos(-\pi s\theta)}{\sin^{2}(-\pi s\theta)}\right],$ and consequently $\displaystyle\frac{\phi_{s}^{\prime}(\theta)}{\phi_{s}(\theta)}=\frac{\sin(-\pi s\theta)+\pi s\theta\cos(-\pi s\theta)}{\theta\sin(-\pi s\theta)}.$ (A.19) We are now ready to evaluate Cramér’s rate function that corresponds to the logistic distribution. $\displaystyle\begin{array}[]{rl}\psi_{P}^{*}(y)&=\sup\left\\{y\theta-\log\left(M_{P}[\theta]\right):\theta\in\real\right\\}\\\ &=\sup\left\\{(y-\mu)\theta-\log\left(\phi_{s}(\theta)\right):\theta\in\real\right\\}.\end{array}$ (A.22) If $y=\mu$ then the discussion that follows equation (A.18) implies that $\sup\\{-\log(\phi_{s}(\theta)):\theta\in\real\\}\leq 0$ where the upper bound is attained for $\theta=0$ (since $\phi_{s}(\theta)\geq 1$ and $\phi_{s}(0)=1$). Thus, we can conclude that $\psi_{P}^{*}(\mu)=0$. If $y\neq\mu$ then the optimal solution to (A.22) satisfies $\theta\neq 0$. Since, in addition, for $|s\theta|\rightarrow 1$ we have that $\phi_{s}(\theta)\rightarrow\infty$, and consequently, $-\log(\phi_{s}(\theta))\rightarrow-\infty$, an optimal solution to (A.22) for the case $y\neq\mu$ must satisfy the first-order optimality conditions $\displaystyle 0=y-\mu-\frac{\phi_{s}^{\prime}(\theta)}{\phi_{s}(\theta)}=y-\mu-\frac{1}{\theta}-\frac{\pi s}{\tan{(-\pi s\theta)}},$ (A.23) where the above follows from (A.19). To summarize, $\displaystyle\psi_{P}^{*}(y)=\begin{cases}0,&y=\mu,\\\ (y-\mu)\theta-\log\left(B(1-s\theta,1+s\theta)\right),&y\neq\mu,\end{cases}$ where $\theta\in\real$ is the nonzero root of (A.23).
# The impact of the uncertainties in the 12C($\alpha$, $\gamma$)16O reaction rate on the evolution of low– to intermediate–mass stars Ben T. Pepper,1 A. G. Istrate2, A. D. Romero1 and S. O. Kepler1 1Physics Institute, Universidade Federal do Rio Grande do Sul, 91501-900 Porto-Alegre, RS, Brazil 2Department of Astrophysics/IMAPP, Radboud University, P O Box 9010, NL-6500 GL Nijmegen, The Netherlands E-mail<EMAIL_ADDRESS> (Accepted XXX. Received YYY; in original form ZZZ) ###### Abstract One of the largest uncertainties in stellar evolutionary computations is the accuracy of the considered reaction rates. The 12C($\alpha$, $\gamma$)16O reaction is particularly important for the study of low- and intermediate-mass stars as it determines the final C/O ratio in the core which influences the white dwarf cooling evolution. Thus, there is a need for a study of how the computations of white dwarfs and their progenitors that are made to date may be affected by the uncertainties of the 12C($\alpha$, $\gamma$)16O reaction rates. In this work we compute fully evolutionary sequences using the MESA code with initial masses in the range of $0.90\leq M_{i}/M_{\odot}\leq 3.05$. We consider different adopted reaction rates, obtained from the literature, as well as the extreme limits within their uncertainties. As expected, we find that previous to the core helium burning stage, there are no changes to the evolution of the stars. However, the subsequent stages are all affected by the uncertainties of the considered reaction rate. In particular, we find differences to the convective core mass during the core helium burning stage which may affect pulsation properties of subdwarfs, the number of thermal pulses during the asymptotic giant branch and trends between final oxygen abundance in the core and the progenitor masses of the remnant white dwarfs. ###### keywords: nuclear reactions – stars: abundances – stars: evolution ††pubyear: 2021††pagerange: The impact of the uncertainties in the 12C($\alpha$, $\gamma$)16O reaction rate on the evolution of low– to intermediate–mass stars–B ## 1 Introduction Single stellar evolution is fuelled by nuclear reactions that occur within the stellar interior (Bethe, 1939; Hoyle, 1946, 1954; Burbidge et al., 1957). These reactions not only release energy which allows the star to support itself against gravitational collapse and remain in hydrostatic equilibrium, but also change the composition of the star: this is known as nucleosynthesis (Eddington, 1920; Hoyle, 1954; Burbidge et al., 1957). The study of these nuclear reactions is where nuclear physics and astronomy come hand-in-hand; an understanding of what happens at the fundamental level provides a better knowledge of how stars evolve and influence their environment. Particularly, improved estimations of the often uncertain reaction rate data, including formula fitted to such data, will improve the accuracy of stellar evolution codes and the understanding of stellar evolution (Caughlan & Fowler, 1988; Angulo et al., 1999; Katsuma, 2012; Xu et al., 2013; An et al., 2016). Such estimations are hereafter referred to as ’reaction rates’. The 12C($\alpha$, $\gamma$)16O reaction during the central helium burning stage is considered to be the most important mechanism for defining the white dwarf (WD) core composition (Salaris & Cassisi, 2005; D’Antona & Mazzitelli, 1990; De Gerónimo et al., 2017; Deboer et al., 2019). However, the reaction rate for this reaction has an extremely large uncertainty (Fowler et al., 1967; Caughlan & Fowler, 1988; Kunz et al., 2002; An et al., 2016; Deboer et al., 2017, 2019). The main entrance channel for the ${}^{12}\text{C}+\alpha$ mechanism ($E_{\alpha_{0}}=7.16\,\text{MeV}$) does not have a resonance channel close to this threshold, the closest occurring at $E_{x}=9.59\,\text{MeV}$. Instead, the low energy cross-section is largely influenced by the $1^{-1}$ ($E_{x}=7.12\,\text{MeV}$) and $2^{+}$ ($E_{x}=6.92\,\text{MeV}$) subthreshold states (see Figure 2 of Deboer et al., 2017, for details). The primary influence of these two nearby subthreshold states and the addition of possible resonant transitions in the wings of the broad channel at $E_{x}=9.59\,\text{MeV}$ makes the nuclear cross-section extremely difficult to estimate (see Fowler et al., 1967; Kunz et al., 2002; An et al., 2016; Deboer et al., 2017, 2019; Aliotta et al., 2021). During the core helium burning (CHB) stage, carbon is produced from the fusion of three helium nuclei via the triple-$\alpha$ process (Salpeter, 1952; Kippenhahn & Weigert, 1990; Salaris & Cassisi, 2005; Prialnik, 2009). As the abundance of helium in the core depletes, the probability of carbon interacting with helium to produce oxygen [via 12C($\alpha$, $\gamma$)16O] is larger than that of the triple-$\alpha$ process at late times during the core helium burning stage (Salaris & Cassisi, 2005). Thus, the 12C($\alpha$, $\gamma$)16O reaction is of great importance and is vital to model the carbon- oxygen (C/O) abundance in the inner chemical profiles for all stellar masses, but particularly low– and intermediate–mass stars (Woosley & Weaver, 1995; Weaver & Woosley, 1993; Wallerstein et al., 1997). The C/O abundance, therefore the 12C($\alpha$, $\gamma$)16O reaction, is important in many areas of stellar evolution. Such as, influencing the pulsation properties of ZZ Ceti stars (De Gerónimo et al., 2015, 2017). Differences between the considered 12C($\alpha$, $\gamma$)16O reaction rate will also affect the duration of the core helium burning stage (Deboer et al., 2017). In addition, the 12C($\alpha$, $\gamma$)16O reaction impacts supernova explosions as the outcome is related to the composition of the final WD (e.g. Iben & Tutukov, 1984; Wu et al., 2020) and third dredge-up episodes (TDUs) during the asymptotic giant branch (AGB) stage (Frost & Lattanzio, 1996; Karakas et al., 2002; Marigo, 2002; Karakas & Lattanzio, 2003; Marigo, 2007; Cristallo et al., 2009; Weiss & Ferguson, 2009; Ventura & Marigo, 2009; Kalirai et al., 2014; Matteucci, 2021). Furthermore, thermonuclear explosions of C/O WDs impacts the ignition of Type 1a supernovae, an important event in constraining cosmological parameters (Perlmutter et al., 1999; Riess et al., 1998). The enrichment of the outer layer of the AGB stars from dredge-up and the mass-loss affects the chemical evolution of galaxies (Matteucci, 2012; Boothroyd & Sackmann, 1988; Kobayashi et al., 2020; Ventura et al., 2020; Cristallo et al., 2015; Matteucci, 2021). Additionally, the 12C($\alpha$, $\gamma$)16O reaction governs whether a star will form a neutron star or black hole (Brown et al., 2001; Heger et al., 2002; Tur et al., 2007; West et al., 2013; Sukhbold & Adams, 2020). Gravitational wave detections from black hole mergers can also be used to constrain the 12C($\alpha$, $\gamma$)16O reaction rate by determining the mass of the black hole and the fraction of carbon and oxygen that remains (see Farmer et al., 2020, for details). De Gerónimo et al. (2015) and De Gerónimo et al. (2017) consider 3 different reaction rates: an adopted rate from Angulo et al. (1999) and the high and low rates from Kunz et al. (2002). They consider these alternate rates for the CHB until the thermally pulsing asymptotic giant branch (TP-AGB) phase with a sole focus on how the pulsational properties are affected in ZZ Ceti stars, rather than all stages as we attempt in this work. In this work, we use stellar evolutionary models as tools to study the impact of the 12C($\alpha$, $\gamma$)16O reaction rate uncertainties on the stellar structure and evolution of low– and intermediate–mass stars. The paper is organised as follows. Section 2 describes the input physics and numerical tool used to compute the evolutionary sequences, as well as a deeper discussion of the considered 12C($\alpha$, $\gamma$)16O reaction rates used in this work. In section 3 we present and discuss our results. We summarise our work in section 4, concluding our findings and indicating future areas where the impact of this work may affect. ## 2 Numerical Tools ### 2.1 MESA Input Physics In this work we employ the Modules for Experiments in Stellar Astrophysics (MESA) code version-r15140 (see Paxton et al. (2011); Paxton et al. (2013, 2015, 2018), for details). We compute the full evolutionary sequence from the zero age main sequence (ZAMS) through both core hydrogen and helium burning stages, leading to the AGB and the white dwarf stage (WD). The computation stops when the stellar model reaches a luminosity of $\log(L/L_{\odot})=-3$ on the WD cooling track. This stopping condition is applied such that the sequences have experienced their evolution through the DAV instability strip (Fontaine & Brassard, 2008; Winget & Kepler, 2008; Althaus et al., 2010). This allows for asteroseismology of ZZ Ceti stars to be performed in the future. The final WD masses obtained in this work range from $0.513M_{\odot}\leq M_{f}/M_{\odot}\leq 0.691\,M_{\odot}$. The initial mass range considered in this work is selected such that all sequences evolve into a carbon–oxygen WD (examples of works which consider/include a similar mass range are Renedo et al., 2010; Romero et al., 2015; De Gerónimo et al., 2017; Marigo et al., 2020). We compute a total of 246 sequences, with an initial metallicity of $Z_{i}=0.01$ and 41 initial masses in the range of $0.90\leq M_{i}/M_{\odot}\leq 3.05$. For each initial mass, we compute the full evolution considering 6 different formulae for the 12C($\alpha$, $\gamma$)16O reaction rate. The 6 reaction rates are adapted from Angulo et al. (1999) and An et al. (2016). Each source comprises 3 reaction rates: the adopted rate, the low and high limiting values, given by the reported uncertainties of the respective rate (see Section 2.2). The rates taken from Angulo et al. (1999) are part of the NACRE compilation and have been used extensively in other computations (Renedo et al., 2010; Romero et al., 2015; De Gerónimo et al., 2017). The reaction rates from An et al. (2016) are less recognised, but boast a lower uncertainty on their reported adopted reaction rate. More detail on these rates and their significance can be found in Section 2.2. We use the reaction network ’basic.net’, which comprises 33 individual reactions including the full p-p chain, CNO cycle, 3$\alpha$ up until 24Mg, which contains the 12C($\alpha$, $\gamma$)16O reaction. This network also includes 8 individual isotopes: 1H, 3He, 4He, 12C, 14N, 16O, 20Ne, 24Mg in addition to elementary and $\alpha$ particles. In our computations, we consider the default radiative opacity tables within MESA. These are from Ferguson et al. (2005) (for $2.7\leq\textrm{log}T\leq 4.5$) and from the OPAL project (for $3.75\leq\textrm{log}T\leq 8.7$) (Iglesias & Rogers, 1993, 1996). Furthermore, we consider OPAL Type 2 tables as they allow for varying amounts of C and O, which are needed for helium burning and beyond (Iglesias & Rogers, 1996; Paxton et al., 2011). We adopt the standard mixing length free parameter as $\alpha=2.0118$. This value is adopted from the work of Guzik et al. (2016) who found this value to be a good approximation for sequences that consider the solar metallicity when using the opacity tables from the OPAL project. To derive this value, Guzik et al. (2016) compared calculated non-adiabatic solar oscillation frequencies and solar interior sound speeds to observed frequencies and helioseismic inferences. However, it should be noted that Guzik et al. (2016) consider an initial metallicity of $Z_{i}=0.015$, rather than the value we consider in this work ($Z_{i}=0.01$). Such a difference would alter the value of the $\alpha$ parameter if a similar analysis was performed with this initial metallicity consideration. Convective mixing is treated as a time-dependent diffusion process, with the diffusion coefficient given as, $D_{\mathrm{EM}}=D_{0}\exp(-2z/fH_{P})$ (1) where $H_{P}$ is the pressure scale-height at the convective boundary, $D_{0}$ is the diffusion coefficient of the unstable regions that are near the convective boundary, and $z$ is the geometric distance from the convective boundary. $f$ is an adjustable free parameter that controls the efficiency of mixing by setting the size of the overshooting region (Herwig et al., 1997; Herwig, 2000). We take the value of $f=0.016$ for all regions of the model for this work, following the same consideration of overshooting as Herwig (2000); Weiss & Ferguson (2009); De Gerónimo et al. (2017). This treatment of the convective boundaries was also adopted by other authors for single stellar evolution computations (Weiss & Ferguson, 2009; Romero et al., 2015; De Gerónimo et al., 2017). The presence of dredge-up episodes during the core helium burning stage is relevant for the final composition of WDs (Prada Moroni & Straniero, 2002; Straniero et al., 2003; Renedo et al., 2010). During the thermally pulsing AGB phase, although overshooting was considered at the boundary of the convective H-rich envelope during the TP-AGB, the third dredge-up episodes did not occur. Therefore, the evolution of the hydrogen–exhausted core (which is hereafter simply referred to as "the helium core mass") and the final mass of the sequences for those which should experience some third dredge-up episodes will be affected (see Section 3.2). We define the "helium core mass" as the region from the centre until the local abundance of hydrogen is greater than $10^{-6}$. Additional models were computed to assess the impact of the third dredge-up on the core mass growth during the thermal pulses (see Section 3.2 and Appendix B for details). For regions stable against convection according to the Ledoux criterion, but there is an inversion of mean molecular weight, we employ thermohaline mixing. In MESA this is treated as a diffusion process, as above, with a diffusion coefficient produced by the stability analysis of Ulrich (1972) and Kippenhahn et al. (1980). For the efficiency parameter of thermohaline mixing, we consider $\alpha_{th}=1.0$ (see Equation 14 of Paxton et al., 2013, for details)). Thermohaline mixing was considered in order to smooth a discontinuity in the carbon and oxygen chemical profiles at the edge of the C/O core, during the early-AGB. Towards the end of the core helium burning stage, when the central He abundance is lower than $\sim$10%, breathing pulse–like instabilities may appear. However, these events are attributed to adopted algorithms rather than to the physics of convection (see Straniero et al., 2003; Romero et al., 2015; Constantino et al., 2016, 2017, for details). To suppress the breathing pulses, when the central abundance of He drops below 0.13, we neglect convection until the central abundance of helium decreases below $10^{-6}$, similar to the prescription used by Renedo et al. (2010) and Romero et al. (2015). Without this prescription, the final carbon-to-oxygen (C/O) ratios can vary rapidly (up to $\pm 0.1$) with small increments of initial mass ($0.05\,M_{\odot}$). During the main sequence (MS), red–giant branch (RGB) and core helium burning stages, the mass-loss due to stellar winds follows the rate based on the Reimers formula (see Reimers, 1975). The asymptotic giant branch and subsequent evolution follow a rate based on the Bloecker formula instead (see Bloecker, 1995). We set our scale factors to be $\eta_{R}=0.5$ and $\eta_{B}=0.2$ for the Reimers and Bloecker formulae, respectively. These values are chosen as they reproduce a WD with a similar final mass to that found by Renedo et al. (2010) for $M_{i}=1.00M_{\odot}$ with $Z_{i}=0.01$. A grey atmosphere is employed for the entire evolution of all sequences, which utilises the grey Eddington $\tau$ relation. We consider the equations of state ELM EOS and DT2 EOS, which are derived from the HELM EOS (Timmes & Swesty, 2000) and the SCVH tables (Saumon et al., 1995), respectively. Once the star leaves the AGB, we employ an element diffusion process from the work of Burgers (1969). We refer to element diffusion as the physical mechanism for mixing chemicals that includes gravitational settling, thermal diffusion and chemical diffusion. Gravitational settling leads to denser element diffusing towards the core, while lighter elements float towards the surface. Thermal diffusion acts in the same direction as gravitational settling, although to a lesser extent, bringing highly charged and more massive species to the central regions of the star. Chemical diffusion, however, works against this general direction (see Iben & MacDonald, 1985; Thoul et al., 1994, for details). In addition to the aforementioned processes, MESA includes radiative accelerations (Hu et al., 2011) into their element diffusion prescription. These radiative forces are negligible in hot regions, as well as being computationally demanding. Hence, we do not consider the effects of radiative levitation. Our element diffusion process is applied to the following isotopes: 1H, 3He, 4He, 12C, 14N, 16O, 20Ne, 24Mg. ### 2.2 The 12C($\alpha$, $\gamma$)16O Reaction Here we discuss a brief, yet relevant, history of 12C($\alpha$, $\gamma$)16O reaction rate evaluations. We lead this into further detail for the 12C($\alpha$, $\gamma$)16O reaction rate prescriptions from Angulo et al. (1999) and An et al. (2016), discussing their differences to the previous determinations from the literature. Fowler et al. (1967) organised the first symposium of reaction rate cross- sections that included the 12C($\alpha$, $\gamma$)16O reaction. At the time, many resonant factors were neglected and were updated by Caughlan & Fowler (1988). However, it is believed that some resonances were still neglected and the treatment of the S-factor in this work produced values that are too small and require a scale factor of $\sim$2 to produce a realistic S-Factor (Angulo et al., 1999; Kunz et al., 2002; Heil et al., 2008; An et al., 2016; Deboer et al., 2017, 2019). Built upon the works of Fowler et al. (1967); Caughlan & Fowler (1988) and those associated works in between, Angulo et al. (1999) provided a strong basis for the 12C($\alpha$, $\gamma$)16O reaction within the NACRE compilation. Angulo et al. (1999) provided the reaction rates for 86 different reactions, including 12C($\alpha$, $\gamma$)16O. For the S-factor calculations, Angulo et al. (1999) considered the values for non-resonant energies. For narrow resonances, however, they fit the resulting cross-section using a Briet-Wigner model. When the effects of different resonant energies overlap, they use a multi-resonance fit, shown in equation 29 of Angulo et al. (1999). Angulo et al. (1999) state that their analysis is numerical for the majority, although they do provide an analytical approach for each reaction, for completeness. They find that their numerical approach yields a higher accuracy for their calculated reaction rates. The quoted S-Factor value from Angulo et al. (1999) for a stellar energy of 300 keV is $S(\text{300\, \text{keV}})=199\pm 64\,\text{keV\,b}$, resulting in a reaction rate ($RR$) of $RR(\text{300\, \text{keV}})=(9.11^{+3.69}_{-3.67})\cdot 10^{-15}\text{cm\textsuperscript{3}\, mole\textsuperscript{-1} s\textsuperscript{-1}}$. A stellar energy of $E=\text{300\, \text{keV}}$ is often chosen as the energy at which to compare the S-factors across different works, as it is associated with the ignition of core helium burning. In this work, we consider the adopted rate of Angulo et al. (1999) (NACRE_A) and the highest and lowest reaction rate within the uncertainties (NACRE_H and NACRE_L, respectively). Hereafter, we refer to the collective 12C($\alpha$, $\gamma$)16O reaction rates from Angulo et al. (1999) as ’NACRE’. An et al. (2016) point out that the resonance parameters used by Kunz et al. (2002), which were taken from Tilley et al. (1993), neglect the ground state transitions from the works of Brochard, F. et al. (1975); Ophel et al. (1976). This results in a larger value for the expected reaction rate at helium burning temperatures. Instead, An et al. (2016) use the reduced R-matrix and S-factor derived by An et al. (2015) to estimate the reaction rate, which accounted for all transitions. In their computations, An et al. (2015) and An et al. (2016) found a significant reduction to the uncertainty of their S-factors when compared to that of Angulo et al. (1999), $S(\text{300\, keV})=162.7\pm 7.3\,\text{keV\, b}$. The reaction rate for the same energy resulted $RR(\text{300\,keV})=(7.83\pm 0.35)\times 10^{-15}\text{cm\textsuperscript{3}\, mole\textsuperscript{-1} s\textsuperscript{-1}}$. We consider the adopted rate from An et al. (2016) (An_A) and the highest and lowest reaction rate within the uncertainties (An_H and An_L, respectively). However, the S-factor calculation of An et al. (2015), seems to neglect external contributions for ground state energy levels, making this approximation not valid for high precision analysis (Deboer et al., 2017). Therefore, we treat the uncertainties of the 12C($\alpha$, $\gamma$)16O reaction rate from An et al. (2016) as arbitrary differences to determine the effect of the urgent need for more precise 12C($\alpha$, $\gamma$)16O reaction rate uncertainties, as claimed by Kunz et al. (2002); Tur et al. (2010). Some works further claim that the uncertainty must be less than 10% to be on par with non-nuclear physical uncertainties (see Woosley et al., 2003; Deboer et al., 2017, for details). Figure 1 shows a comparison between the adopted reaction rates from An et al. (2016), NACRE and all of their associated uncertainties. In this figure, we depict for each rate, the ratio between the rate and the value for NACRE_A, as a function of temperature. For an analysis including other works, see Figure 4 of An et al. (2016). As can be seen from Figure 1, for energies characteristic of stellar energies, the An_A, An_H and An_L reaction rates are lower than for NACRE_A for most temperatures within the blue shaded region, characteristic of core helium burning temperatures. We therefore expect to have a larger C/O ratio in the core after the central helium burning stage for the sequences which consider the rate from An et al. (2016) when compared to those sequences which consider NACRE_A. It can also be seen in Figure 1 that the range between NACRE_H and NACRE_L includes all the other prescriptions within the region of helium burning temperatures, which will lead to the largest differences in the C/O ratio after the core helium burning stage. At higher temperatures (greater than those considered to be helium burning temperatures) the reaction rate from An et al. (2016) is larger than that from NACRE. These temperatures are not reached in the sequences computed within this work. Figure 1: Ratios of each reaction rate considered when compared to the adopted NACRE rate for the 12C($\alpha$, $\gamma$)16O reaction, as a function of temperature, where $T_{9}=T/10^{9}$. The beige shaded region defines the temperatures where helium burning occurs. During the core helium burning stage is also where the 12C($\alpha$, $\gamma$)16O reaction is most prominent. The light-orange dotted and red dashed lines represent the NACRE_L and NACRE_H considerations, respectively. The solid blue line defines the adopted rate from An et al. (2016) with the An_L and An_H rates being depicted as light- blue dotted and dark-blue dashed lines, respectively. ## 3 Results and Discussions In this section we describe in detail the effects that the uncertainties of the 12C($\alpha$, $\gamma$)16O reaction rate have on the inner structure and evolution for low- and intermediate-mass single stars. As expected, during the pre-main sequence, main sequence (MS) and red-giant branch (RGB), we find no differences to the evolution since the 12C($\alpha$, $\gamma$)16O reaction only becomes important, and increasingly more dominant, during the CHB as the central helium abundance decreases (Salaris & Cassisi, 2005; Spruit, 2015; Deboer et al., 2019). Thus, we report no difference between the different 12C($\alpha$, $\gamma$)16O reaction rates at the time of, or shortly after, the helium-flash or a non-degenerate helium ignition. We only show the results from the CHB, AGB and WD stages where we expect some differences to occur due to the uncertainties and separate literature sources of the 12C($\alpha$, $\gamma$)16O reaction rate. We consider each evolutionary stage separately in chronological order. ### 3.1 The Core Helium Burning Phase Figure 2 shows the carbon–to–oxygen (C/O) ratio for each star at the end of CHB, as a function of initial mass. As expected due to the large uncertainties of the reaction rate from NACRE, the smallest and largest C/O ratios come from the NACRE_H and NACRE_L rates, respectively. Note that when all reaction rates from An et al. (2016) are considered, the values for the C/O ratios are between the values corresponding to NACRE_A and NACRE_L. We find that the C/O ratio at the end of the CHB decreases for all considered reaction rates around an initial mass of $M_{i}=1.90\,M_{\odot}$. This mass corresponds to the minimum mass for which helium burning starts in non- degenerate conditions, and will be referred to as the transition mass. The C/O ratio increases again for higher initial masses (between $2.20\leq M_{i}/M_{\odot}\leq 2.45$). We find that the initial mass where the increase of the C/O ratio occurs is dependent on the considered 12C($\alpha$, $\gamma$)16O reaction rate, such that higher reaction rates have a wider initial mass range for the decreased C/O ratio and lower reaction rates have a narrower initial mass range. For example, the NACRE_H has the widest range ($1.90\leq M_{i}/M_{\odot}\leq 2.45$) whereas the NACRE_L has the narrowest range ($1.90\leq M_{i}/M_{\odot}\leq 2.20$). Furthermore, we find no difference to the initial mass range between the adopted rate from An et al. (2016) and the An_H and An_L rates. We also add that the decrease in the C/O ratio is more pronounced for less efficient reaction rates, see Figure 2, for details. Figure 2: Central C/O ratio at the end of the CHB as a function of initial mass. The red points represent the reaction rates considered by NACRE and the blue points are those considered by An et al. (2016). Additionally, squares represent the respective adopted rates while darker-coloured triangles and lighter-coloured upside-down triangles represent the high and low limit uncertainties, respectively. Figure 3 shows the time spent in the CHB as a function of initial mass for the High and Low reaction rate formulas for NACRE (left panel) and An et al. (2016) (right panel). We consider the difference in the CHB age from the values obtained using the respective adopted reaction rate for each panel. Considering the NACRE rates (left panel of Figure 3), we find that the CHB lifetime can be up to 12 Myr shorter (longer) from the adopted rate if we consider NACRE_L (NACRE_H) reaction rate, which is roughly a 7% difference. On the other hand the differences between the An et al. (2016) rates are much lower (right panel of Figure 3), up to 4 Myr translating to a difference of 4%. Such changes to the CHB lifetimes due to limits of the uncertainties on the 12C($\alpha$, $\gamma$)16O reaction rate are not negligible, particularly for the rate taken from NACRE. Constantino et al. (2016) found that the difference in the the ratio of HB–to–AGB stars in a sample of 48 globular clusters could be explained by the differences in the CHB duration due to the uncertainties in the 12C($\alpha$, $\gamma$)16O reaction rate. Figure 3: Differences to the duration of the CHB stage due to associated reaction rate uncertainties as a function of initial mass. The differences are calculated between each limit of the reaction rate due to their uncertainties and the adopted rate of each case. The left panel shows the differences of the uncertainties of the rate calculated by NACRE and the right panel shows the same from the rate of An et al. (2016). Darker-coloured triangles and lighter- coloured upside-down triangles represent the high and low limit uncertainties, respectively. The top panel of Figure 4 shows the CHB history of the convective mass. The convective mass is defined as the mass-coordinate of the core convective boundary, such that convection occurs between this mass-coordinate and the centre. Additionally, the bottom panel of Figure 4 shows the luminosities of the 3$\alpha$ process and the 12C($\alpha$, $\gamma$)16O reaction (the latter will be referred to as C$\alpha$ luminosity), for the NACRE reaction rates. As expected, the C$\alpha$ luminosity increases when the more efficient reaction rates are considered. Furthermore, the contribution from the 3$\alpha$ process decreases for higher reaction rates due to the helium reservoir being depleted faster by the more efficient 12C($\alpha$, $\gamma$)16O reaction. Mixing episodes due to the convective core during the CHB extends from the C/O core to the He-rich layers above, so we define the convective mass as the mass of the convective core. Figure 4 also shows that higher reaction rates produce more mixing episodes which are characterised by sudden increases of the convective mass. These enhanced convective episodes bring fresh helium from the helium region above the C/O core which not only increases the duration of the CHB but also increases the abundance of oxygen in the core (Ghasemi et al., 2017; Guo & Li, 2018). Convective mixing episodes induce a chemical discontinuity between the fully mixed core and the radiative layer, increasing the opacity beyond the convective boundary. In a class of CHB pulsating stars, sdB stars (see Heber, 2009, for an in depth discussion), g-modes propagate from the surface all the way until the boundary of the convective core (Ghasemi et al., 2017). Since we find significant differences to the size of the convective core and number of mixing episodes between the NACRE adopted reaction rate and its uncertainties for the 12C($\alpha$, $\gamma$)16O reaction rate, the precision of astereoseismology for these objects is limited and must be considered in the calculations of the pulsation period spectrum. However for the adopted rate taken from An et al. (2016), the high and low limits (An_H and An_L, respectively) do not produce a significant change to the convective core mass and the total number of mixing episodes and would therefore produce a more precise study of the g-mode pulsations (see Figure 13 in Appendix A, for an example of the same case that considers the reaction rates from An et al. (2016)). The implications for asteroseismology from the treatment to mixing during the CHB has been studied by Constantino et al. (2015) who found that changes to the composition and He-burning reaction rates do not significantly change the period spacing of pulsations for pulsators during the CHB stage. However, the period values could be more sensitive to the changes in the chemical profile. Figure 4: History of the convective mass (top panel), 3$\alpha$ luminosity and the luminosity of the 12C($\alpha$, $\gamma$)16O reaction during the CHB (bottom panel). The history is given in terms of the CHB duration. This plot in particular considers all NACRE prescriptions for the 12C($\alpha$, $\gamma$)16O reaction rate for an initial mass of $M_{i}=2.45\,M_{\odot}$. Blue lines represent NACRE_H, orange-brown depicts NACRE_A and dark-brown shows NACRE_L. Furthermore, the solid line represents the convective mass, dotted lines show the luminosity of the 12C($\alpha$, $\gamma$)16O reaction and dot-dash lines portray the 3$\alpha$ luminosity. The total energy produced by the 12C($\alpha$, $\gamma$)16O reaction during the CHB is presented in Figure 5. The values shown in Figure 5 are moving averages. We compute the total energy by integrating the C$\alpha$ luminosity with respect to time for the CHB duration. Figure 5 shows the ratio between the different reaction rates and the NACRE_A (top panel) and An_A (bottom panel) reaction rates, as a function of initial mass. If we consider the reaction rates from An et al. (2016), the differences are generally smaller than 10%, the largest difference occurs for the sequence with an initial mass of $2.85\,M_{\odot}$ that considers An_H. In most cases, the differences are no larger than 5% (70.7% of the sequences for An_H and 82.9% of the sequences for An_L). We find larger differences between the limits of 12C($\alpha$, $\gamma$)16O NACRE rates when compared to the NACRE_A formula, as shown in the top panel of Figure 5. In this case we also compare the adopted reaction rate from An et al. (2016). If we consider how NACRE_H differs from NACRE_A, we find that the energy production for the majority of the sequences are greater than 10% than that of the NACRE_A case, with a few exceeding a difference of 20%. For NACRE_L, the carbon energy produced differs more than 30% from the NACRE_A rate. The extra energy produced from the high rates when compared to the adopted rates increases the temperature gradient further allowing convection to continue, causing the extra mixing episodes shown in Figure 4 (Kippenhahn & Weigert, 1990; Prialnik, 2009). Considering the adopted rate from An et al. (2016), the absolute value of the differences in carbon energy produced due to An_H and An_L appears to be independent of either selection. This is not the case for the NACRE rates. A limiting factor for the amount of energy produced is the abundance of available helium. This is more of a limit for the NACRE_H case due to lack of available helium inhibiting further reactions to occur. The NACRE_L will always produce less carbon energy and so is not limited by the helium abundance or lack thereof. The smaller uncertainties of the rates taken from An et al. (2016) are not large enough to produce such an effect. Figure 5: Ratios of the total energy produced by the 12C($\alpha$, $\gamma$)16O reaction as a function of initial mass. Values are presented in the form of moving averages. The energy produced is calculated by integrating the C$\alpha$ luminosity shown in Figure 4 and is integrated with respect to time. The ratios in the top panel are in terms of the NACRE_A rate and the ratios in the bottom panel are made in terms of An_A. The red points represent the reaction rates considered by NACRE and the blue points are those considered by An et al. (2016). Additionally, squares represent the respective adopted rates while darker-coloured triangles and lighter-coloured upside-down triangles represent the high and low limit uncertainties, respectively. The CHB stage is where the 12C($\alpha$, $\gamma$)16O reaction is the most active. In particular, we find that the largest differences due to the considered 12C($\alpha$, $\gamma$)16O reaction rate appear in the final C/O ratio, CHB duration, energy generation rate and the number of experienced mixing episodes. The primary reason that we find such changes to these properties is due to the changes in energy generation that affects the convection efficiency in this phase. Furthermore, we find that the differences between the An_H and An_L rates from the An_A rate are generally insignificant, unlike those of the NACRE uncertainties which are intrinsically larger. A final point to add is that, in future works, the use of overshooting parameters specifically designed for the CHB would be interesting. Works such as Spruit (2015) claim to keep the convective boundaries stable inhibiting the need for manual breathing pulse suppression, as performed in this work, whilst keeping "stable" convection active throughout the evolution (Spruit, 2015; Constantino et al., 2017). ### 3.2 The Asymptotic Giant Branch Phase During the AGB the energy production is given by two shell sources, the hydrogen-shell at the base of the hydrogen-rich envelope and the He-shell on top of the C/O core. Hydrogen burning occurs through the CNO cycle, while He- burning is through the 3$\alpha$ process. Towards the end of the AGB, the He- burning shell will become thin enough to trigger unstable burning, and the thermal pulses (TPs) begin (e.g. Kippenhahn & Weigert, 1990; Iben, 1991). During the interpulse period between the TPs, the outer convection zone may be deep enough to bring the products of He-shell burning to the surface, this is known as the third dredge-up (TDU) (Wallerstein et al., 1997; Busso et al., 1999; Herwig, 2005; Karakas & Lattanzio, 2014). Well known consequences of TDUs are a reduction of the helium core mass and changes to the surface composition, leading to the formation of C-stars (Frost & Lattanzio, 1996; Busso et al., 1999; Karakas et al., 2002; Weiss & Ferguson, 2009; Romero et al., 2015; Marigo et al., 2020). The extent of the reduction of the helium core mass from TDU episodes is parameterised by the dredge-up efficiency parameter, $\lambda_{d}$111The dredge-up efficiency parameter is defined as the fraction of helium core mass lost during the TDU episode over the helium core mass growth since the last TDU (see Karakas et al., 2002; Marigo et al., 2013, for details). The 12C($\alpha$, $\gamma$)16O reaction during this stage is essentially inactive. There may be some fusion reactions between 12C and alpha particles at the edge of the C/O core but they are, however, insignificant. Thus, any difference between the sequences during the AGB is due to the effect that the 12C($\alpha$, $\gamma$)16O reaction rate has during the CHB. Figure 6 shows the helium core mass at the first TP of each sequence as a function of initial mass. A minimum value occurs for an initial mass $M_{i}=1.90\,M_{\odot}$, which is the transition point as described in Section 3.1. The same result was found in the work of Kalirai et al. (2014), whose initial models come from those produced in Bressan et al. (2012). However, their transition point occurs for $M_{i}=2.00\,M_{\odot}$ due the larger initial metallicity affecting the mass for which core helium burning ignites in degenerate conditions (Bertelli et al., 1986; Romero et al., 2015). We find that there is no significant difference to the helium core mass at the first TP as a result of different 12C($\alpha$, $\gamma$)16O reaction rates for masses lower than the transition point. Above this mass, the maximum difference between the NACRE rates is $\sim 0.01M_{\odot}$, with NACRE_L producing lower helium core masses and NACRE_H producing larger helium core masses. This is due to the difference in energy outputs between the adopted rate, NACRE_A, and the NACRE_H/NACRE_L rates. Higher reaction rates during the CHB increase the temperature throughout the star which favours the CNO-cycle (Boeltzig et al., 2016), allowing the helium core mass to develop further than sequences which consider lower reaction rates. There are no significant differences in the helium core mass at the first TP between the adopted rate from An et al. (2016) and An_H/An_L for any of the considered initial masses. Figure 6: Helium core mass at the start of the first TP as a function of initial mass. All of the considered reaction rates and their uncertainties are shown within this figure. We find a minimum to the helium core mass for the same initial mass which corresponds to the transition mass where core helium burning begins on a non-degenerate core rather an electron degenerate core ($M_{i}=1.90\,M_{\odot}$). Within the uncertainties, we find differences up to 0.01 M⊙ for masses larger than $M_{i}=1.90\,M_{\odot}$. The red points represent the reaction rates considered by NACRE and the blue points are those from An et al. (2016). Additionally, squares represent the respective adopted rates while darker-coloured triangles and lighter-coloured upside-down triangles represent the high and low limit uncertainties, respectively. Figure 7 shows the growth of the helium core mass during the TP-AGB as a function of initial mass for each considered reaction rate. We find that the dramatic increase of core growth (for helium core mass growth $\geq 10\%$ (Kalirai et al., 2014)) occurs in the range $1.70\leq M_{i}/M_{\odot}\leq 2.60$, with a maximum increase of 19% occurring at $M_{i}\approx 2.00\,M_{\odot}$. This result is in agreement with that of Bird & Pinsonneault (2011) and is similar to that of Kalirai et al. (2014), who find a helium core growth up to 30%. This discrepancy between their work and ours is due to not only a different initial metallicity, but also their consideration of a less efficient mass-loss scheme for stages previous to the AGB (Reimers law with $\eta_{R}=0.2$ (Bressan et al., 2012)). Thus, the models used by Kalirai et al. (2014) have a larger mass of hydrogen fuel to produce a larger final mass (see Table 1 for our values of this variable and Bird & Pinsonneault (2011) for an in-depth discussion of the hydrogen fuel variable). Furthermore, possible differences to the energy produced in the H-rich envelope during the TP-AGB may affect the rate of the helium core growth (see Forestini & Charbonnel, 1997; Marigo et al., 2013; Kalirai et al., 2014, for details). Considering only the difference in helium core mass growth for NACRE_A rate and it’s NACRE_H/NACRE_L limits, we find that NACRE_L has a larger core growth and NACRE_H has smaller core growth. The increased core growth during the AGB for the NACRE_L sequences is due to the smaller helium core mass at the first TP (see Figure 6) and as such more fuel to keep He-shell burning sustained, particularly for initial masses above the transition point where the core growth differences are greater (see Table 1). Additionally, during the TP-AGB, we find differences in the energy generation from the CNO cycle between the NACRE_H/NACRE_L limits in comparison with the NACRE_A. The energy generation can be up to 25% lower (higher) when the NACRE_H (NACRE_L) reaction rate is considered. Figure 7: Percentage growth of the helium core mass during the AGB as a function of initial mass. Growth is calculated as the difference between the final mass of the core and the helium core mass described in Figure 6. We find that the largest growth occurs for initial masses $\approx 2.00\,M_{\odot}$, peaking at 19%. Above initial masses of $M_{i}=2.90\,M_{\odot}$, it appears that the growth begins to plateau around 8-9%. The red points represent the reaction rates considered by NACRE and the blue points are those from An et al. (2016). Additionally, squares represent the respective adopted rates while darker-coloured triangles and lighter-coloured upside-down triangles represent the high and low limit uncertainties, respectively. Figure 8 shows the number of thermal pulses as a function of initial mass for each considered reaction rate. Moreover, it shows that lower reaction rates experience more TPs than higher reaction rates. This is related to the larger amount of available hydrogen to aid the outward growth of the helium core through a greater number of unstable He-shell burning episodes - TPs. We do not find any M-star to C-star transitions (see Marigo et al., 2020, for example) as convective overshooting about the boundary between the helium core and the He–exhausted core was disregarded during the TP-AGB, inhibiting the TDU (Herwig, 2000; Romero et al., 2015). However, overshooting still occurred at the boundary of the H–rich core. We define the "He–exhausted core" as the region from the centre until the local abundance of helium is greater than $10^{-6}$. Thermal pulses are strongly dependent on the mass–loss rate, helium core mass and initial metallicity (Karakas et al., 2002; Cristallo et al., 2009; Weiss & Ferguson, 2009; Renedo et al., 2010; Romero et al., 2015; De Gerónimo et al., 2017). We find that the number of thermal pulses in our computations is lower than that from the works of Weiss & Ferguson (2009); Renedo et al. (2010) and Romero et al. (2015) for a given initial mass, a similar treatment of convection and a similar helium core mass at the beginning of the TP-AGB phase. Difference in the number of TPs could be related to the different mass–loss schemes during the RGB stage. In this work we consider the mass–loss prescription from Bloecker (1995) while the works of Weiss & Ferguson (2009); Renedo et al. (2010) and Romero et al. (2015) consider a mass–loss scheme that produces a "super wind" stage towards the last TPs, making it more efficient in these last TPs but less so in the early TP-AGB (see Vassiliadis & Wood, 1993; van Loon et al., 2005, for details). However, the trend in the number of experienced TPs as a function of initial mass obtained in our work agrees with other works (see Weiss & Ferguson, 2009; Renedo et al., 2010; Romero et al., 2015). To assess the effect of the TDU during the TP-AGB, we computed additional sequences, allowing convective overshooting to occur at all fully– or semi–convective boundaries, with $f=0.016$ (see Appendix B, for details on it’s effect). For sequences that consider the NACRE_A prescription, TDU episodes occur for initial masses larger than $M_{i}\geq 2.40\;M_{\odot}$, with the dredge-up efficiency parameter ($\lambda_{d}$) showing values of $\lambda_{d}=0.033-0.124$ that increases with increasing initial mass. The abundance of carbon and oxygen at the surface does increase during each TDU in these additional models, but the C/O is still lower than 1 meaning that our models show an oxygen dominated surface. A higher value of the overshooting parameter may be necessary to produce C–stars (see Herwig et al., 1997; Karakas et al., 2002; Weiss & Ferguson, 2009; Romero et al., 2015; Marigo et al., 2020, for examples of C-star transitions). For sequences where convective overshooting was considered across all boundaries during the AGB we find a decrease in the final helium core mass up to 0.63%. This value is much lower than the 15% decrease found by Karakas et al. (2002); Romero et al. (2015). The sequences that have initial masses $M_{i}<2.40\;M_{\odot}$ do not show any third dredge–up episodes, as such we do not expect any difference to the growth of the helium core or the final mass. For those sequences with initial masses $M_{i}\geq 2.40\;M_{\odot}$, a more detailed study of the convective boundaries during the TP-AGB is required for more thorough analysis of why we find such weak dredge–up efficiency parameters. In the case of NACRE_H and NACRE_L, we find that TDU episodes occur for the same initial mass range as that of the NACRE_A sequences ($2.40\leq M_{i}/M_{\odot}\leq 3.05$). Additionally, the dredge-up efficiency parameters are also similar to those of the NACRE_A sequences, with $\lambda_{d}=0.040-0.123$. From the results gathered in this work, we find that the uncertainties of current 12C($\alpha$, $\gamma$)16O reaction rates are not significant in modelling the TDU. The 12C($\alpha$, $\gamma$)16O reaction during the AGB is negligible during the TP-AGB. Instead, the main energy source occurs through the 3$\alpha$ reaction series and the CNO-cycle within the H-rich envelope (Herwig, 2005; Karakas & Lattanzio, 2014). Thus, we do not find any significant change to the peak TP luminosity nor the depth of each TDU, since the changes in core mass at the beginning of the TP-AGB are negligible as a result of the uncertainties of the 12C($\alpha$, $\gamma$)16O reaction rate, as shown in Figure 6 (see Wallerstein et al., 1997; Wagenhuber & Groenewegen, 1998; Busso et al., 1999; Herwig, 2005; Karakas & Lattanzio, 2014, for details). However, the uncertainties of the overshooting efficiency raises a greater uncertainty in the surface composition during the AGB, as such we leave a detailed discussion for a future work that considers the overshooting efficiency in more detail (Abia et al., 2002; Herwig, 2005; Cristallo et al., 2009; Ventura & Marigo, 2009; Karakas & Lattanzio, 2014). Figure 8: Number of TPs experienced as a function of initial mass. Each reaction rate consideration and their uncertainties are shown. We find that the number of TPs peaks at initial masses $\approx 2.00\,M_{\odot}$, in-line with the largest core growth, as in Figure 7. We also show that lower reaction rates for the 12C($\alpha$, $\gamma$)16O reaction produce more TPs. The red points represent the reaction rates considered by NACRE and the blue points are those from An et al. (2016). Additionally, squares represent the respective adopted rates while darker-coloured triangles and lighter-coloured upside-down triangles represent the high and low limit uncertainties, respectively. $M_{i}$/$M_{\odot}$ | $\Delta M_{\text{growth}}/M_{\odot}$ | $M_{\text{fuel}}/M_{\odot}$ ---|---|--- NACRE_H | NACRE_A | NACRE_L | An_A | NACRE_H | NACRE_A | NACRE_L | An_A 1.00 | 0.009 | 0.010 | 0.009 | 0.009 | 0.007 | 0.008 | 0.007 | 0.008 1.50 | 0.030 | 0.027 | 0.031 | 0.031 | 0.024 | 0.022 | 0.026 | 0.025 1.60 | 0.037 | 0.038 | 0.039 | 0.038 | 0.030 | 0.031 | 0.031 | 0.031 2.00 | 0.085 | 0.091 | 0.089 | 0.091 | 0.069 | 0.073 | 0.072 | 0.073 2.90 | 0.043 | 0.044 | 0.050 | 0.048 | 0.035 | 0.035 | 0.040 | 0.039 Table 1: Values showing the TP-AGB helium core mass growth and fuel mass. We report the values from the following reaction rate considerations: NACRE_H, NACRE_A, NACRE_L and An_A. We do not report the values from the uncertainties of the rate taken from An et al. (2016) since they are negligible when compared to their adopted rate. ### 3.3 The White Dwarf Final Cooling Track Figure 9 shows the initial-to-final mass relation (IFMR) for all sequences produced in this work. We find that there is no significant difference in the final mass of any given initial mass due to the 12C($\alpha$, $\gamma$)16O reaction rate. Considering the largest difference in the reaction rates, between NACRE_H and NACRE_L, the largest difference in the final mass for a given initial mass is less than $0.01\,M_{\odot}$ ($<2\%$). In the interest of the pursuit for a global IFMR, we compare our IFMR to those of other works of a similar metallicity. We consider the IFMRs from the works of Weidemann (2000); Salaris et al. (2009) and Renedo et al. (2010). We find a similar trend with the work of Weidemann (2000), both of which consider the same mass-loss scheme from Bloecker (1995) for the AGB phase. The IFMRs from the works of Salaris et al. (2009) and Renedo et al. (2010) consider the mass–loss scheme from Vassiliadis & Wood (1993) for the AGB and show a much steeper gradient in their IFMRs. However, the core masses between this work and the works of Weidemann (2000); Salaris et al. (2009) and Renedo et al. (2010) are similar at the first TP. Thus, it is reasonable to assume that the difference is due to their considered mass-loss scheme for the IFMR determination. By considering the third-order polynomial nature of the IFMR computed in this work, we fit a function to the NACRE_A final masses to produce a general relation from the results of this work. This allows for a comparison to other IFMRs as well as other masses to be easily estimated, if desired. The following IFMR reproduces the IFMR of NACRE_A well, such that the R-square value is $R^{2}=0.9995$: $M_{f}=0.02047M_{i}^{3}-0.1051M_{i}^{2}+0.2323M_{i}+0.3783M_{\odot}$ (2) where $M_{f}$ is the final mass and $M_{i}$ is the initial mass. The non- linear relationship described by Equation 2 is caused by the mass-loss rate adopted on the AGB. The Bloecker (1995) scheme in particular has a large dependency on luminosity. It would be interesting to see how our IFMR holds for observational data as well as it’s dependency on metallicity - an important dependence as discussed in Romero et al. (2015). Figure 9: Initial-to-final mass relation of all sequences calculated as part of this work. Also shown are other IFMRs from the works of Weidemann (2000); Salaris et al. (2009); Renedo et al. (2010) (yellow stars, purple dashed line and black squares, respectively) for a comparison of their trends. The red points represent the reaction rates considered by NACRE, and the blue points are those from An et al. (2016). Additionally, squares represent the respective adopted rates while darker-coloured triangles and lighter-coloured upside-down triangles represent the high and low limit uncertainties, respectively. We find that the slope of the IFMR has a strong dependency on the considered mass-loss scheme considered during the AGB, with the scheme from Vassiliadis & Wood (1993) producing a steeper gradient and that from Bloecker (1995) showing a shallower gradient. In Figure 10 we show, in panel a), the final ages of a WD that has cooled to an effective temperature of $T_{\textrm{eff}}=10\,000$K (log scale) as a function of initial mass for all the sequences computed in this work. The differences in the final ages due to the High/Low limits of each considered 12C($\alpha$, $\gamma$)16O reaction rate are in general negligible, with variations of the order $\sim 0.01$ Gyr for both the NACRE and An et al. (2016) 12C($\alpha$, $\gamma$)16O reaction rate. The variations in the reported final ages due to the uncertainties of the 12C($\alpha$, $\gamma$)16O reaction rate are a magnitude lower than the populations studied in the works of Hansen et al. (2013); Forbes et al. (2015); Campos et al. (2016). As such, the impact that the 12C($\alpha$, $\gamma$)16O reaction rate has on final ages of WD models is currently negligible as compared to the greater uncertainty of ageing stellar populations. Panels c) and d) of Figure 10 show the moving average for the time spent on the cooling track for the NACRE and An et al. (2016) 12C($\alpha$, $\gamma$)16O reaction rates, respectively. We define this quantity as the time taken for a star on the final cooling track to cool from it’s maximum effective temperature until an effective temperature of $T_{\text{eff}}=10\,000$K. During the final cooling track, the differences in the duration due to the reaction rates between the Adopted and High/Low limits generally differ up to $0.030$ Gyr for those of NACRE and up to $0.015$ Gyr for An et al. (2016). The general trend is in agreement with past discussions of the effect of the 12C($\alpha$, $\gamma$)16O reaction rate and cooling time during this stage of evolution, such that more oxygen-rich cores will produce a lower cooling time. This is due to the gravitational energy release during stratification occurring at earlier times for more oxygen-rich cores. As a consequence, the WD is left with a lower thermal content to feed the surface luminosity at later times. The larger the luminosity at which the stratification occurs, the shorter the resulting cooling times will be (D’Antona & Mazzitelli, 1990; Prada Moroni & Straniero, 2002; Salaris et al., 2010). Furthermore, for the High/Low limits of the NACRE rate, we find that NACRE_L produces a greater absolute difference than that of NACRE_H. This is due to the availability of helium during the CHB as discussed in Section 3.1. Figure 10: Panel a) shows the final age (log scale) of the star on the final cooling track with an effective temperature $T_{\text{eff}}=10\,000$K. Panel b) shows the time spent of the cooling track, defined as the time taken for a WD on the final cooling track to cool from its maximum effective temperature to an effective temperature of $T_{\text{eff}}=10\,000$K. Panel c) and d) show the moving average for the difference of cooling times between the High/Low limits and the Adopted rate for the NACRE and An et al. (2016) 12C($\alpha$, $\gamma$)16O reaction rate, respectively. All panels are represented as functions of initial mass. The NACRE reaction rates are shown as different shades of red and those from An et al. (2016) are depicted by shades of blue. Furthermore, squares represent the respective adopted rates while darker- coloured triangles and lighter-coloured upside-down triangles represent the high and low limit uncertainties, respectively. In general, we find that the uncertainties of the 12C($\alpha$, $\gamma$)16O reaction rate have an negligible effect on the final ages of the stars at this point, whereas the cooling time can differ up to 8%. After the settling and diffusion processes described in Section 2, the final oxygen abundances within the core of the sequences are presented in Figure 11, as a function of initial mass. We find similar trends to the oxygen mass fraction in this stage to those found at the end of the CHB. Although there are slight increases to the oxygen mass fraction due to the aforementioned diffusion processes (Unglaub & Bues, 2000). Additionally, diffusion affects the C/O ratio throughout the star up to the surface and not just in the core (see Herwig, 2000; Straniero et al., 2003, for details). The onset of crystallisation begins when the core cools to a certain temperature, $T_{c}$ (Segretain et al., 1994; Horowitz et al., 2010). This temperature is dependent on the internal composition of the star. Through observations of the globular cluster NGC 6397, Winget et al. (2009) report that the crystallisation of the WD core is similar to that of a pure carbon core. According to the phase diagram produced in Horowitz et al. (2010) and their limits for the maximum crystallisation temperature, this would require a limit to the oxygen mass fraction of $X_{\text{O}}\leq 0.64$. This requires that the maximum S-factor at 300 keV has an upper limit of $S(300\,\text{keV})\leq 170\,\text{keV b}$. Considering the relationship between oxygen mass fraction and initial mass presented in Figure 11, we find that NACRE_H and NACRE_A produce central oxygen abundances that are too large for a crystallisation process similar to that found by Horowitz et al. (2010). Meanwhile, the rates An et al. (2016) agree not only with the oxygen mass fraction limit presented by Horowitz et al. (2010), but also their derived S-factor for an energy of 300 keV. Thus, we find that sequences dedicated to studying crystallisation using the method presented by Horowitz et al. (2010) should consider a lower reaction rate than that from NACRE for the 12C($\alpha$, $\gamma$)16O reaction to keep their analysis consistent with the input physics that they use. Figure 11: Central oxygen mass fraction for the final WD as a function of initial mass. We show each calculated sequence. The trends for each considered reaction rate are similar to those found in Figure 2. There has been a slight increase in the central oxygen abundance since the CHB due to diffusion processes in the star. Additionally, squares represent the respective adopted rates while darker-coloured triangles and lighter-coloured upside-down triangles represent the high and low limit uncertainties, respectively. Figure 12 shows the abundance profiles of white dwarf models with a stellar mass of $M_{*}=0.548M_{\odot}$, $T_{\textrm{eff}}=20\,000$K and an initial mass of $M_{i}=1.30\,M_{\odot}$. Sequences that consider a reaction rate from NACRE are shown in the top panel and those from An et al. (2016) are represented in the bottom panel. All sequences finish with similar structure to those shown in Figure 12. The profiles depict a DA white dwarf configuration, with a hydrogen-rich envelope, a helium buffer and a C/O core. Where the abundance of carbon reaches it’s maximum, we hereafter refer to this as the carbon peak. We show that the interior of the star has a consistent trend where the carbon peak is higher for lower reaction rates - an outcome of a less efficient reaction rate which leaves behind a larger abundance of carbon. Furthermore, the position of the carbon peak changes with the reaction rates, moving away from the centre as the reaction rate increases. We find in general that differences between An_A and the An_H/An_L reaction rates do not affect this region drastically (bottom panel), unlike that of the NACRE 12C($\alpha$, $\gamma$)16O reaction rate considerations (top panel). The abundance profile and composition gradients in these central regions that lie within the range of $1<-\textrm{log}_{10}(1-M_{r}/M_{*})<2$ affect the peaks in the Brunt-Väisälä frequency, which disturbs the period spectrum structure (see Córsico & Althaus, 2006; Romero et al., 2012b, for more details). This is an outcome of the pulsation modes that are trapped in this region through the mode-trapping mechanism. We confirm that uncertainties of the 12C($\alpha$, $\gamma$)16O reaction rate may affect the pulsation period spectrum. Another region where the Brunt-Väisälä frequency is affected is in the He/H transition region. In particular, the position of the He/H transition will impact the period spectrum (Romero et al., 2012a, 2013). Figure 12: In both panels we show the abundance profiles of sequences considering an initial mass of $M_{i}=1.30\,M_{\odot}$. The top panel represents the adopted rate and it’s uncertainties for the NACRE rate, and the same for the An et al. (2016) rates in the bottom panel. The line-styles for each rate are shown in the legend in the bottom panel and the colours for each element is shown in the legend in the top panel. Colour version is available online. ## 4 Conclusions In this work we analyse the impact that the limits of the 12C($\alpha$, $\gamma$)16O reaction rate has on the inner structure and evolutionary properties of low- and intermediate-mass stars. We consider the 12C($\alpha$, $\gamma$)16O reaction rates from NACRE (Angulo et al., 1999) and An et al. (2016). We have computed stellar sequences from the ZAMS until the remnant white dwarf reaches a luminosity of $\text{log}(L/L_{\odot})=-3$. We applied similar starting parameters for different ensembles of reaction rates where we consider the adopted rate along with the upper and lower limits within the uncertainties of each source. We summarise our main results below. 1. 1. The C/O ratio of the core in the final model of each sequence is affected by the 12C($\alpha$, $\gamma$)16O reaction rate as expected, with lower C/O ratios for larger reaction rates. We find that the decreased C/O ratio for initial masses greater than the transition mass increase again at higher masses. The mass at which this increase occurs is dependent on the considered 12C($\alpha$, $\gamma$)16O reaction rate, such that it occurs for higher masses if higher reaction rates are considered. This is due to an increased number of mixing episodes, a cause of larger energy outputs increasing convective efficiency which brings fresh helium to the core during the CHB. Note that significant differences between the adopted rate and high/low limits occur only for those rates taken from NACRE which has a much larger uncertainty than those from An et al. (2016). 2. 2. CHB lifetime is dependent on the considered reaction rate, a higher reaction rate produces a greater lifetime. We deem this to be a consequence in the number of mixing episodes extending the core helium burning lifetime, although further research would be beneficial to confirm this. Between the adopted rate and high/low limits, we find a difference up to 12 Myr for the NACRE rates and up to 4 Myr for those from An et al. (2016). 3. 3. The helium core mass at the beginning of the first TP is independent of the considered 12C($\alpha$, $\gamma$)16O reaction rate up to and including the transition mass. Above this mass, we find a maximum difference of $\approx 0.01M_{\odot}$ between NACRE_H and NACRE_L, with lower reaction rates producing a lower helium core mass. Additionally, our minimum helium core mass at this point occurs at our transition mass. 4. 4. Growth of the helium core mass between the first TP and the final mass reaches a maximum of 19%, with growths greater than 10% occurring in the mass range $1.70\leq M_{i}/M_{\odot}\leq 2.60$ which is in agreement with Bird & Pinsonneault (2011) and Kalirai et al. (2014). The largest growths occur for the lower reaction rates due to more available hydrogen which remained after the CHB. There are no significant differences between the rates taken from An et al. (2016) due to the limits being smaller in relation to their adopted rate than those from NACRE. 5. 5. The number of TPs during the TP-AGB is dependent on the considered 12C($\alpha$, $\gamma$)16O reaction rate. We find that lower reaction rates increase the number of TPs due to a larger hydrogen fuel aiding the outward growth of the helium core mass by fuelling the unstable He-shell with a greater supply of fresh helium. 6. 6. TDU episodes occur for sequences in the initial mass range of $2.40\leq M_{i}/M_{\odot}\leq 3.05$ with dredge-up efficiency parameters $\lambda_{d}=0.033-0.124$. This mass range is independent of the considered 12C($\alpha$, $\gamma$)16O reaction rate. Additionally, the values of $\lambda_{d}$ between the considered 12C($\alpha$, $\gamma$)16O reaction rate uncertainties are not significant. Furthermore, the depth of each TDU is independent of the 12C($\alpha$, $\gamma$)16O reaction rate. 7. 7. The IFMR produced in this work has a similar trend to that of Weidemann (2000), who also consider a similar mass-loss prescription during the AGB. The IFMRs of Renedo et al. (2010) and Salaris et al. (2009) show a much steeper gradient and they consider the Vassiliadis & Wood (1993) mass-loss prescription during the AGB. 8. 8. We find that the final ages of the sequences are in general independent of the considered reaction rate. However, during the final cooling track, we find differences up to 10% between the adopted rates and high/low limits. This is true for both those rates taken from NACRE and An et al. (2016). This difference in the cooling time agrees with the works of Prada Moroni & Straniero (2002); Salaris et al. (2010); Isern et al. (2013). 9. 9. The final C/O ratio in the core shows a similar trend to that at the end of the CHB. The oxygen abundance increases slightly due to the diffusion processes. The final oxygen mass fraction for NACRE_A and NACRE_H sequences are greater than the values derived by Horowitz et al. (2010) for crystallisation of a C/O core. The reaction rates from An et al. (2016) agree closely with the derived values of Horowitz et al. (2010). As such, future works should consider a lower reaction rate than that of NACRE when considering the crystallisation process of Horowitz et al. (2010). 10. 10. The inner structure of the star is affected by the uncertainties within the considered reaction rates, particularly those from NACRE. The position and height of the carbon peak is significantly affected by the difference between the adopted rate and high/low limits of the reaction rate for the NACRE considerations. This may affect the modes in which pulsations can occur during the ZZ Ceti instability strip (Córsico & Althaus, 2006; Romero et al., 2012a). Although we analyse the possible evolutionary stages where more accurate 12C($\alpha$, $\gamma$)16O reaction rates are needed, a deeper analysis of some effects are still required. For instance, a quantification of how the pulsation modes of sdB’s and ZZ Ceti stars are affected, for example. Furthermore, we conclude that a lower reaction than that of NACRE_A is favourable for the Horowitz et al. (2010) considerations of crystallisation, however, this must be further analysed as well. By limiting the uncertainties of 12C($\alpha$, $\gamma$)16O reaction rates to 10% of the adopted rate, as in An et al. (2016), reports a much better consistency of stellar parameters. ## Acknowledgements BTP, ADR and SOK acknowledge support by CNPq and PRONEX-FAPERGS/CNPq. This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 001. This research has made use of NASA’s Astrophysics Data System. AGI acknowledges support from the Netherlands Organisation for Scientific Research (NWO). We also thank developers of the MESA software, which was used extensively in this work. Finlly, we thank the anonymous referee for their input to make it a more complete work. ## Data Availability The data is available upon request to the corresponding author. ## References * Abia et al. (2002) Abia C., et al., 2002, ApJ, 579, 817 * Aliotta et al. (2021) Aliotta M., et al., 2021, arXiv e-prints, p. arXiv:2109.14418 * Althaus et al. (2010) Althaus L. G., Córsico A. H., Isern J., García-Berro E., 2010, A&ARv, 18, 471 * An et al. (2015) An Z.-D., et al., 2015, Phys. Rev. C, 92, 045802 * An et al. (2016) An Z.-D., Ma Y.-G., Fan G.-T., Li Y.-J., Chen Z.-P., Sun Y.-Y., 2016, ApJ, 817, L5 * Angulo et al. (1999) Angulo C., et al., 1999, Nuclear Phys. A, 656, 3 * Bertelli et al. (1986) Bertelli G., Bressan A., Chiosi C., Angerer K., 1986, A&AS, 66, 191 * Bethe (1939) Bethe H. A., 1939, Physical Review, 55, 434 * Bird & Pinsonneault (2011) Bird J. C., Pinsonneault M. H., 2011, ApJ, 733, 81 * Bloecker (1995) Bloecker T., 1995, A&A, 297, 727 * Boeltzig et al. (2016) Boeltzig A., et al., 2016, European Physical Journal A, 52, 75 * Boothroyd & Sackmann (1988) Boothroyd A. I., Sackmann I. J., 1988, ApJ, 328, 653 * Bressan et al. (2012) Bressan A., Marigo P., Girardi L., Salasnich B., Dal Cero C., Rubele S., Nanni A., 2012, MNRAS, 427, 127 * Brochard, F. et al. (1975) Brochard, F. Chevallier, P. Disdier, D. Rauch, V. Scheibling, F. 1975, J. Phys. France, 36, 113 * Brown et al. (2001) Brown G. E., Heger A., Langer N., Lee C.-H., Wellstein S., Bethe H. A., 2001, New Astron., 6, 457 * Burbidge et al. (1957) Burbidge E. M., Burbidge G. R., Fowler W. A., Hoyle F., 1957, Reviews of Modern Physics, 29, 547 * Burgers (1969) Burgers J. M., 1969, Flow Equations for Composite Gases * Busso et al. (1999) Busso M., Gallino R., Wasserburg G. J., 1999, ARA&A, 37, 239 * Campos et al. (2016) Campos F., et al., 2016, MNRAS, 456, 3729 * Caughlan & Fowler (1988) Caughlan G. R., Fowler W. A., 1988, Atomic Data and Nuclear Data Tables, 40, 283 * Constantino et al. (2015) Constantino T., Campbell S. W., Christensen-Dalsgaard J., Lattanzio J. C., Stello D., 2015, MNRAS, 452, 123 * Constantino et al. (2016) Constantino T., Campbell S. W., Lattanzio J. C., van Duijneveldt A., 2016, MNRAS, 456, 3866 * Constantino et al. (2017) Constantino T., Campbell S. W., Lattanzio J. C., 2017, MNRAS, 472, 4900 * Córsico & Althaus (2006) Córsico A. H., Althaus L. G., 2006, A&A, 454, 863 * Cristallo et al. (2009) Cristallo S., Straniero O., Gallino R., Piersanti L., Domínguez I., Lederer M. T., 2009, ApJ, 696, 797 * Cristallo et al. (2015) Cristallo S., Straniero O., Piersanti L., Gobrecht D., 2015, ApJS, 219, 40 * D’Antona & Mazzitelli (1990) D’Antona F., Mazzitelli I., 1990, ARA&A, 28, 139 * De Gerónimo et al. (2015) De Gerónimo F. C., Córsico A. H., Althaus L. G., Romero A. D., 2015, The Impact of the Uncertainties in the 12C($\alpha$,$\gamma$)16O Reaction Rate on the Asteroseismology of ZZ Ceti Stars: First Results. p. 225 * De Gerónimo et al. (2017) De Gerónimo F. C., Althaus L. G., Córsico A. H., Romero A. D., Kepler S. O., 2017, A&A, 599, A21 * Deboer et al. (2017) Deboer R. J., et al., 2017, in APS Division of Nuclear Physics Meeting Abstracts. p. CC.006 * Deboer et al. (2019) Deboer R., Brune C., Wiescher M., 2019, in APS Division of Nuclear Physics Meeting Abstracts. p. FG.003 * Eddington (1920) Eddington A. S., 1920, The Observatory, 43, 341 * Farmer et al. (2020) Farmer R., Renzo M., de Mink S. E., Fishbach M., Justham S., 2020, ApJ, 902, L36 * Ferguson et al. (2005) Ferguson J. W., Alexander D. R., Allard F., Barman T., Bodnarik J. G., Hauschildt P. H., Heffner-Wong A., Tamanai A., 2005, ApJ, 623, 585 * Fontaine & Brassard (2008) Fontaine G., Brassard P., 2008, PASP, 120, 1043 * Forbes et al. (2015) Forbes D. A., Pastorello N., Romanowsky A. J., Usher C., Brodie J. P., Strader J., 2015, MNRAS, 452, 1045 * Forestini & Charbonnel (1997) Forestini M., Charbonnel C., 1997, A&AS, 123, 241 * Fowler et al. (1967) Fowler W. A., Caughlan G. R., Zimmerman B. A., 1967, ARA&A, 5, 525 * Frost & Lattanzio (1996) Frost C. A., Lattanzio J. C., 1996, ApJ, 473, 383 * Ghasemi et al. (2017) Ghasemi H., Moravveji E., Aerts C., Safari H., Vučković M., 2017, MNRAS, 465, 1518 * Guo & Li (2018) Guo J.-J., Li Y., 2018, MNRAS, 478, 3290 * Guzik et al. (2016) Guzik J. A., Fontes C. J., Walczak P., Wood S. R., Mussack K., Farag E., 2016, IAU Focus Meeting, 29B, 532 * Hansen et al. (2013) Hansen B. M. S., et al., 2013, Nature, 500, 51 * Heber (2009) Heber U., 2009, ARA&A, 47, 211 * Heger et al. (2002) Heger A., Woosley S. E., Rauscher T., Hoffman R. D., Boyes M. M., 2002, New Astron. Rev., 46, 463 * Heil et al. (2008) Heil M., et al., 2008, Phys. Rev. C, 78, 025803 * Herwig (2000) Herwig F., 2000, A&A, 360, 952 * Herwig (2005) Herwig F., 2005, ARA&A, 43, 435 * Herwig et al. (1997) Herwig F., Bloecker T., Schoenberner D., El Eid M., 1997, A&A, 324, L81 * Herwig et al. (1999) Herwig F., Blöcker T., Langer N., Driebe T., 1999, A&A, 349, L5 * Horowitz et al. (2010) Horowitz C. J., Schneider A. S., Berry D. K., 2010, Phys. Rev. Lett., 104, 231101 * Hoyle (1946) Hoyle F., 1946, MNRAS, 106, 343 * Hoyle (1954) Hoyle F., 1954, ApJS, 1, 121 * Hu et al. (2011) Hu H., Tout C. A., Glebbeek E., Dupret M. A., 2011, MNRAS, 418, 195 * Iben (1991) Iben I. J., 1991, in Michaud G., Tutukov A. V., eds, Vol. 145, Evolution of Stars: the Photospheric Abundance Connection. p. 257 * Iben & MacDonald (1985) Iben I. J., MacDonald J., 1985, ApJ, 296, 540 * Iben & Tutukov (1984) Iben I. J., Tutukov A. V., 1984, ApJS, 54, 335 * Iglesias & Rogers (1993) Iglesias C. A., Rogers F. J., 1993, ApJ, 412, 752 * Iglesias & Rogers (1996) Iglesias C. A., Rogers F. J., 1996, ApJ, 464, 943 * Isern et al. (2013) Isern J., Artigas A., García-Berro E., 2013, in European Physical Journal Web of Conferences. p. 05002 (arXiv:1212.0806), doi:10.1051/epjconf/20134305002 * Kalirai et al. (2014) Kalirai J. S., Marigo P., Tremblay P.-E., 2014, ApJ, 782, 17 * Karakas & Lattanzio (2003) Karakas A. I., Lattanzio J. C., 2003, Publ. Astron. Soc. Australia, 20, 279 * Karakas & Lattanzio (2014) Karakas A. I., Lattanzio J. C., 2014, Publ. Astron. Soc. Australia, 31, e030 * Karakas et al. (2002) Karakas A. I., Lattanzio J. C., Pols O. R., 2002, Publ. Astron. Soc. Australia, 19, 515 * Katsuma (2012) Katsuma M., 2012, ApJ, 745, 192 * Kippenhahn & Weigert (1990) Kippenhahn R., Weigert A., 1990, Stellar Structure and Evolution * Kippenhahn et al. (1980) Kippenhahn R., Ruschenplatt G., Thomas H. C., 1980, A&A, 91, 175 * Kobayashi et al. (2020) Kobayashi C., Karakas A. I., Lugaro M., 2020, ApJ, 900, 179 * Kunz et al. (2002) Kunz R., Fey M., Jaeger M., Mayer A., Hammer J. W., Staudt G., Harissopulos S., Paradellis T., 2002, ApJ, 567, 643 * Marigo (2002) Marigo P., 2002, A&A, 387, 507 * Marigo (2007) Marigo P., 2007, A&A, 467, 1139 * Marigo et al. (2013) Marigo P., Bressan A., Nanni A., Girardi L., Pumo M. L., 2013, MNRAS, 434, 488 * Marigo et al. (2020) Marigo P., et al., 2020, Nature Astronomy, * Matteucci (2012) Matteucci F., 2012, Chemical Evolution of Galaxies, doi:10.1007/978-3-642-22491-1. * Matteucci (2021) Matteucci F., 2021, A&ARv, 29, 5 * Ophel et al. (1976) Ophel T. R., Frawley A. D., Treacy P. B., Bray K. H., 1976, Nuclear Physics A, 273, 397 * Paxton et al. (2011) Paxton B., Bildsten L., Dotter A., Herwig F., Lesaffre P., Timmes F., 2011, ApJS, 192, 3 * Paxton et al. (2013) Paxton B., et al., 2013, ApJS, 208, 4 * Paxton et al. (2015) Paxton B., et al., 2015, ApJS, 220, 15 * Paxton et al. (2018) Paxton B., et al., 2018, ApJS, 234, 34 * Perlmutter et al. (1999) Perlmutter S., Turner M. S., White M., 1999, Phys. Rev. Lett., 83, 670 * Prada Moroni & Straniero (2002) Prada Moroni P. G., Straniero O., 2002, ApJ, 581, 585 * Prialnik (2009) Prialnik D., 2009, An Introduction to the Theory of Stellar Structure and Evolution * Reimers (1975) Reimers D., 1975, Circumstellar envelopes and mass loss of red giant stars. pp 229–256 * Renedo et al. (2010) Renedo I., Althaus L. G., Miller Bertolami M. M., Romero A. D., Córsico A. H., Rohrmann R. D., García-Berro E., 2010, ApJ, 717, 183 * Riess et al. (1998) Riess A. G., et al., 1998, AJ, 116, 1009 * Romero et al. (2012a) Romero A. D., Córsico A. H., Althaus L. r. G., Miller Bertolami M. M., 2012a, arXiv e-prints, p. arXiv:1204.6101 * Romero et al. (2012b) Romero A. D., Córsico A. H., Althaus L. G., Kepler S. O., Castanheira B. G., Miller Bertolami M. M., 2012b, MNRAS, 420, 1462 * Romero et al. (2013) Romero A. D., Kepler S. O., Córsico A. H., Althaus L. G., Fraga L., 2013, ApJ, 779, 58 * Romero et al. (2015) Romero A. D., Campos F., Kepler S. O., 2015, MNRAS, 450, 3708 * Salaris & Cassisi (2005) Salaris M., Cassisi S., 2005, Evolution of Stars and Stellar Populations * Salaris et al. (2009) Salaris M., Serenelli A., Weiss A., Miller Bertolami M., 2009, ApJ, 692, 1013 * Salaris et al. (2010) Salaris M., Cassisi S., Pietrinferni A., Kowalski P. M., Isern J., 2010, ApJ, 716, 1241 * Salpeter (1952) Salpeter E. E., 1952, ApJ, 115, 326 * Saumon et al. (1995) Saumon D., Chabrier G., van Horn H. M., 1995, ApJS, 99, 713 * Segretain et al. (1994) Segretain L., Chabrier G., Hernanz M., Garcia-Berro E., Isern J., Mochkovitch R., 1994, ApJ, 434, 641 * Spruit (2015) Spruit H. C., 2015, A&A, 582, L2 * Straniero et al. (2003) Straniero O., Domínguez I., Imbriani G., Piersanti L., 2003, ApJ, 583, 878 * Sukhbold & Adams (2020) Sukhbold T., Adams S., 2020, MNRAS, 492, 2578 * Thoul et al. (1994) Thoul A. A., Bahcall J. N., Loeb A., 1994, ApJ, 421, 828 * Tilley et al. (1993) Tilley D. R., Weller H. R., Cheves C. M., 1993, Nuclear Physics A, 564, 1 * Timmes & Swesty (2000) Timmes F. X., Swesty F. D., 2000, ApJS, 126, 501 * Tur et al. (2007) Tur C., Heger A., Austin S. M., 2007, ApJ, 671, 821 * Tur et al. (2010) Tur C., Heger A., Austin S. M., 2010, ApJ, 718, 357 * Ulrich (1972) Ulrich R. K., 1972, ApJ, 172, 165 * Unglaub & Bues (2000) Unglaub K., Bues I., 2000, A&A, 359, 1042 * Vassiliadis & Wood (1993) Vassiliadis E., Wood P. R., 1993, ApJ, 413, 641 * Ventura & Marigo (2009) Ventura P., Marigo P., 2009, MNRAS, 399, L54 * Ventura et al. (2020) Ventura P., Dell’Agli F., Lugaro M., Romano D., Tailo M., Yagüe A., 2020, A&A, 641, A103 * Wagenhuber & Groenewegen (1998) Wagenhuber J., Groenewegen M. A. T., 1998, A&A, 340, 183 * Wallerstein et al. (1997) Wallerstein G., et al., 1997, Reviews of Modern Physics, 69, 995 * Weaver & Woosley (1993) Weaver T., Woosley S., 1993, Physics Reports, 227, 65 * Weidemann (2000) Weidemann V., 2000, A&A, 363, 647 * Weiss & Ferguson (2009) Weiss A., Ferguson J. W., 2009, A&A, 508, 1343 * West et al. (2013) West C., Heger A., Austin S. M., 2013, ApJ, 769, 2 * Winget & Kepler (2008) Winget D. E., Kepler S. O., 2008, ARA&A, 46, 157 * Winget et al. (2009) Winget D. E., Kepler S. O., Campos F., Montgomery M. H., Girardi L., Bergeron P., Williams K., 2009, ApJ, 693, L6 * Woosley & Weaver (1995) Woosley S. E., Weaver T. A., 1995, ApJS, 101, 181 * Woosley et al. (2003) Woosley S. E., Heger A., Rauscher T., Hoffman R. D., 2003, Nuclear Phys. A, 718, 3 * Wu et al. (2020) Wu C., Wang B., Wang X., Maeda K., Mazzali P., 2020, MNRAS, 495, 1445 * Xu et al. (2013) Xu Y., Takahashi K., Goriely S., Arnould M., Ohta M., Utsunomiya H., 2013, Nuclear Physics A, 918, 61 * van Loon et al. (2005) van Loon J. T., Cioni M. R. L., Zijlstra A. A., Loup C., 2005, A&A, 438, 273 ## Appendix A Convection during CHB for An rates Figure 13 shows the CHB history of the convective mass and the luminosities of the 3$\alpha$ process and the 12C($\alpha$, $\gamma$)16O reaction, for the reaction rates taken from An et al. (2016). This figure is analogous to that of Figure 4 which shows the same for the NACRE rates. We provide this figure to prove that we do not find any significant difference between the number of mixing episodes, luminosity from the 3$\alpha$ process and the 12C($\alpha$, $\gamma$)16O reaction. Thus, the high/low limits for the 12C($\alpha$, $\gamma$)16O reaction rate from An et al. (2016) does not affect the CHB in terms of energy production, mixing episodes or CHB duration. This was not found for the NACRE case, which is discussed in Section 3.1. Figure 13: History of the convective mass (top panel), 3$\alpha$ luminosity and the luminosity of the 12C($\alpha$, $\gamma$)16O reaction during the CHB (bottom panel). The history is given in terms of the CHB duration. This plot in particular considers all An et al. (2016) prescriptions for the 12C($\alpha$, $\gamma$)16O reaction rate for an initial mass of $M_{i}=2.45\,M_{\odot}$. Blue lines represent An_H, orange-brown depicts An_A and dark-brown shows An_L. Furthermore, the solid line represents the convective mass, dotted lines show the luminosity of the 12C($\alpha$, $\gamma$)16O reaction and dot-dash lines portray the 3$\alpha$ luminosity. ## Appendix B Additional AGB Models Figure 14 shows the Kippenhahn diagram for the case of $M_{i}=3.05\,M_{\odot}$ during the TP-AGB in the original NACRE_A models. We represent the mass co- ordinate on the first y–axis and the surface C/O ratio on the second y–axis. Both values are plotted against the age of the sequence. These models did not consider convective overshooting around the border of the He–exhausted core. Green slashed areas show convective regions, red back slashed areas represent semi-convective regions and the purple regions are where overshooting occurs. The purple dotted line shows the history of the He–exhausted core mass and the blue dotted line represents the history of the helium core mass. The colour bar measures the energy generation rate from nuclear reactions. The solid orange line represents the C/O ratio at the surface. It can be seen that the overshooting occurs close to the envelope boundary and there is no overshooting about the semi-convective region of the He–exhausted core. As a result of this, we do not observe TDU episodes in the original models. We can be sure that there are no TDU episodes because of the lack of change in helium core mass and that the surface C/O ratio remains constant, which would change if TDUs were experienced (Frost & Lattanzio, 1996; Herwig et al., 1999; Karakas et al., 2002; Weiss & Ferguson, 2009; Romero et al., 2015; De Gerónimo et al., 2017; Marigo et al., 2020). Figure 15 shows the same as Figure 14 but allows for convective overshooting at each boundary. We find that, with the new prescription, convection and overshooting extends throughout the helium buffer. For this reason material can be "dredged-up" from the core to the surface. This results in the helium core mass and He–exhausted core masses changing with each convective episode - an outcome of TDU episodes (Frost & Lattanzio, 1996; Herwig et al., 1999; Karakas et al., 2002; Weiss & Ferguson, 2009; Romero et al., 2015; De Gerónimo et al., 2017; Marigo et al., 2020). Furthermore, we find an increase in the surface C/O ratio with each TDU as material travels from the stellar interior to the surface. The surface C/O ratio, however, remains less than 1. This indicates a larger overshooting parameter is required for M–star to C–star transitions. Figure 14: Kippenhahn diagram for during the TP-AGB for the case of $M_{i}=3.05\,M_{\odot}$ of the original models. We represent the mass co- ordinate on the first y–axis and the surface C/O ratio on the second y–axis. Both values are plotted against the age of the sequence. This model did not consider convective overshooting at boundary of the He–exhausted core which inhibited the TDU. The colour bar measures the energy generation rate from nuclear reactions. The blue dotted line represents the helium core mass while the purple dotted line represents the He–exhausted core. Green slashed regions show convection and the red back slashed regions represent where regions of the star are semi-convective. Finally, purple areas are where overshooting occurs. Figure 15: Kippenhahn diagram for during the TP-AGB for the case of $M_{i}=3.05\,M_{\odot}$ of the new models. We represent the mass co-ordinate on the first y–axis and the surface C/O ratio on the second y–axis. Both values are plotted against the age of the sequence. The colour bar measures the energy generation rate from nuclear reactions. This model considered convective overshooting at all convective boundaries, allowing fro TDUs to occur. The blue dotted line represents the helium core mass while the purple dotted line represents the He–exhausted core. Green slashed regions show convection and the red back slashed regions represent where regions of the star are semi-convective. Finally, purple areas are where overshooting occurs.
# Teaching Design by Contract using Snap! ††thanks: Identify applicable funding agency here. If none, delete this. 1st Marieke Huisman Formal Methods and Tools University of Twente Enschede, The Netherlands <EMAIL_ADDRESS>2nd Raúl E. Monti Formal Methods and Tools University of Twente Enschede, The Netherlands <EMAIL_ADDRESS> ###### Abstract With the progress in deductive program verification research, new tools and techniques have become available to support design-by-contract reasoning about non-trivial programs written in widely-used programming languages. However, deductive program verification remains an activity for experts, with ample experience in programming, specification and verification. We would like to change this situation, by developing program verification techniques that are available to a larger audience. In this paper, we present how we developed prototypal program verification support for Snap!. Snap! is a visual programming language, aiming in particular at high school students. We added specification language constructs in a similar visual style, designed to make the intended semantics clear from the look and feel of the specification constructs. We provide support both for static and dynamic verification of Snap! programs. Special attention is given to the error messaging, to make this as intuitive as possible. ###### Index Terms: verification, software, education ## I Introduction Research in deductive program verification has made substantial progress over the last years: tools and technique have been developed to reason about non- trivial programs written in widely-used programming languages, the level of automation has substantially increased, and bugs in widely-used libraries have been found [1, 2, 3]. However, the use of deductive verification techniques remains the field of expert users, and substantial programming knowledge is necessary to appreciate the benefits of these techniques. We believe that it is important to make deductive program verification techniques accessible also to novice programmers. Therefore, we have to teach the Design-by-Contract [4] (DbC) approach, which requires the programmer to explicitly specify the assumptions and responsibilities of code in a modular way, in parallel with actually teaching programming, i.e. DbC should be taught as an integral part of the process leading from design to implementation. In this paper, we make the Design-by-Contract idea accessible to high school students, in combination with appropriate tool support, which is currently unavailable. Concretely, this paper presents a Design-by-Contract approach for Snap! [5]. Snap! is a visual programming language targeting high school students. The design of Snap! is inspired by Scratch, another widely-used visual programming language. Compared to Scratch, Snap! has some more advanced programming features. In particular, Snap! provides the possibility to create parametrised reusable blocks, basically modelling user-defined functions. Also the look and feel of Snap! targets high school students, whereas Scratch aims at an even younger age group. Snap! has been successfully integrated in high school curricula, by its integration in the _Beauty and Joy of Computing_ course [6]. This course combines programming skills with a training in abstract computational thinking. The first step to support Design-by-Contract for Snap! is to define a suitable specification language. The visual specification language that we propose in this paper is built as a seamless extension of Snap!, i.e. we propose a number of new specification blocks and natural modifications of existing ones. These variations capture the main ingredients for the Design-by-Contract approach, such as pre- and postconditions. Moreover, we also provide blocks to add assertions and loop invariants in a program and we extend the standard expression pallets of Snap! with some common expressions to ease specifications. The choice of specification constructs is inspired by existing specification languages for Design-by-Contract, such as JML [7], choosing the most frequently used constructs with a clear and intuitive meaning. Moreover, all verification blocks are carefully designed to reflect the intended semantics of the specifications in a visual way. A main concern for a programmer, after writing the specification of the intended behaviour of their programs, should be to validate that these programs behave according to their specification. Therefore, we provide two kinds of tool support: (i) runtime assertion checking [8], which checks whether specifications are not violated during a particular program execution, and (ii) static checking (or deductive verification) [9], which verifies that all possible program executions respect its specifications. The runtime assertion checker is built as an extension of the standard Snap! execution mechanism. The deductive verification support is built by providing a translation from a Snap! program into Boogie [10]. Another important aspect to take into account for a good learning experience are the error messages that indicate that a specification is violated. We have integrated these messages in Snap!’s standard error reporting system, again sticking to the look and feel of standard Snap!. Moreover, we have put in effort to make the error messages as clear as possible, so that also a relatively novice programmer can understand why the implementation deviates from the specification. ## II Background ### II-A Snap! Snap! is a visual programming language. It has been designed to introduce children (but also adults) to programming in an intuitive way. At the same time, it is also a platform for serious study of computer science[11]. Snap! actually re-implements and extends Scratch [12]. Programming in Snap! is done by dragging and dropping blocks into the coding area. Blocks represent common program constructs such as variable declarations, control flow statements (branching and loops), function calls and assignments. Snapping blocks together, the user builds a script and visualises its behaviour by means of turtle graphics visualisations, called sprites. Sprites can change shape, move, show bubbled text, play music, etc. For all these effects, dedicated blocks are available. Figure 1: The Snap! working area. The Snap! interface divides the working area into three parts: the pallet area, the scripting area, and the stage area, indicated by labels 1, 2 and 3, respectively, in Fig. 1. On the left, the various programming blocks are organised into pallets that describe their natural use. For instance, the _Variables_ pallet contains blocks for declaring and manipulating variables. In Snap!, variables are dynamically typed. Blocks are dragged and dropped from the pallets into the scripting area, located at the centre of the working area where the Snap! program is constructed. Blocks can be arranged by snapping them together, or by inserting them as arguments of other blocks. Blocks can only be used as arguments if their shapes match with the shape of the argument slots in the target block. These shapes actually provide a hint on the expected evaluation type of a block, for instance, rounded slots for numbers and diamond slots for booleans . The behaviour of the script is shown with turtle graphics drawings in the stage area located in the rightmost part of the screen. In addition, at the bottom of the pallet area, there is a “Make a block” button. This allows the user to define his or her _Build Your Own Block_ (BYOB) blocks. When pressed, a new floating “Block Editor” window pops out with a new coding area, in which the behaviour of the personalised block can be defined (similar to how a script is made in the scripting area). Label $4$ in Fig. 1 shows a BYOB block being edited. Once defined, the BYOB block becomes available to be used just as any other predefined block. ### II-B Program Verification The basis of the Design-by-Contract approach [13] is that the behaviour of all program components is defined as a contract. For example, a function contract specifies the conditions under which a function may be called (the function’s _precondition_), and it specifies the guarantees that the function provides to its caller (the function’s _postcondition_). There exist several specification languages that have their roots in this Design-by-Contract approach. For example the Eiffel programming language has built-in support for pre- and postconditions [14], and for Java, the behavioural interface language JML [15] is widely used. As is common for such languages, we use the keyword _requires_ to indicate a precondition, and the keyword _ensures_ to indicate a postcondition. If a program behaviour is specified using contracts, various techniques can be used to validate whether an implementation respects the contract. Dynamic verification validates an implementation w.r.t. a specification at runtime. This means that, whenever during program execution a specification is reached, it will be checked for this particular execution that the property specified indeed holds. In particular, this means that whenever a function will be called, its precondition will be checked, and whenever the function returns, its postcondition will be checked. An advantage of this approach is that it is easy and fast to use it: one just runs a program and checks if the execution does not violate the specifications. A disadvantage is that it only provides guarantees about a concrete execution. In contrast, static verification aims at verifying that all possible behaviours of a function respect its contract. This is done by applying Hoare logic proof rules [16] or using Dijkstra’s predicate transformer semantics [17]. Applying these rules results in a set of first-order proof obligations; if these proof obligations can be proven it means that the code satisfies its specification. Advantage of this approach is that it guarantees correctness of all possible behaviours. Disadvantage is that it is often labour-intensive, and often many additional annotations, such as for example loop invariants, are needed to guide the prover. ## III Visual Program Specifications This section discusses how to add visual specification constructs to Snap!. Our goal was to do this in such a way that (1) the intended semantics of the specification construct is clear from the way it is visualised, and (2) that it smoothly integrates with the existing programming constructs in Snap! Often, Design-by-Contract specifications are added as special comments in the code. For example, in JML a function contract is written in a special comment, tagged with an @-symbol, immediately preceding the function declaration. The tag ensures that the comment can be recognised as part of the specification. There also exist languages where for example pre- and postconditions are part of the language (e.g., Eiffel [18], Spec# [19]). We felt that for our goal, specifications should be integrated in a natural way in the language, rather than using comments. Therefore, we introduce variations of the existing block structures, in which we added suitable slots for the specifications. This section discusses how we added pre- and postconditions, and in-code specifications such as asserts and loop invariants to Snap!. In addition, to have a sufficiently expressive property specification language, we also propose an extension of the expression constructs. ### III-A Visual Pre- and Postconditions To specify pre- and postconditions for a BYOB script, we provide a variation of the initial hat block with a slot for a precondition at the start of the block, and a slot for a postcondition at the end of the block (Fig. 2). This shape is inspired by the c-shaped style of other Snap! blocks, such as blocks for loops. The main advantage is that it visualises at which points in the execution, the pre- and the postconditions are expected to hold. In addition, it also graphically identifies which code is actually verified. Moreover, the shapes are already familiar to the Snap! programmer. If the slots are not filled, default pre- and postcondition true can be used. Notice that the pre- and postcondition slots consist of multiple boolean-argument slots, and we define the property to be the conjunction of the evaluation of each of these slots. This is similar to how Snap! extends a list or adds arguments to the header of a BYOB. Figure 2: Hat block extended with contracts ### III-B Visual Assertions and Loop Invariants For static verification, pre- and postconditions are often not sufficient, and we need additional in-code specifications to guide the prover, such as assertions, which specify properties that should hold at a particular point in the program, and loop invariants. Moreover, assertions can also be convenient for run-time assertion checking to make it explicit that a property holds at a particular point in the program. #### Visual Assertions To specify assertions, both the property specified and the location within the code are relevant. To allow the specification of assertions at arbitrary places in a script, we define a special assertion block similar to all other control blocks. #### Visual Loop Invariants Loop invariants are necessary for static verification [20]. A loop invariant should hold at the beginning and end of every loop iteration. To account for this, we provide a (multi-argument boolean) slot to specify the loop invariant in the traditional Snap! c-shaped loop block. This slot is located just after the header where the loop conditions are defined. In addition, the c-shaped loop block repeats the word invariant at the bottom of the block (see Figure 3) to visually indicate that the invariant is checked after each iteration. Figure 3: Visual loop invariants. ### III-C Visual Expressions In addition, we have introduced some specification-only keywords, as commonly found in Design-by-Contract languages. * • An _old_ expression is used in postconditions to indicate that a variable/expression should be evaluated in the pre-state of the function. To support this, we introduced an operator block with a slot for a variable name. * • A _result_ expression refers to the return value of a function inside its postcondition. We support this by introducing a constant operator, that allows to specify a property about the result value of a reporter BYOB. We also introduce syntax to ease the definition of complex Boolean expressions, by means of the operator blocks , , and , as well as syntax to write more advanced Boolean expressions, introducing support for quantified expressions (See Fig.4). Figure 4: A global quantification expression block ## IV Graphical approach to verification result reporting Another important point to consider is how to report on the outcome of the verification: (1) presenting the verdict of a passed verification, and (2) in case of failure, giving a concrete and understandable explanation for the failure. The latter is especially important in our case, as we are using the technique with inexperienced users. In order to signal a contract violation, or any assertion invalidated during dynamic verification, we use Snap!’s pop-up notification windows. These windows have the advantage that a failing block can be printed inside them even when the failing script is not currently visible to the user. This allows to be very precise about the error, even when the BYOB body is not currently visible. In order to signal errors while compiling to Boogie, such as making use of dynamic typing or nested lists in your Snap! BYOB code, we use Snap!’s speech bubbles that can emerge at specific points in the script while describing the cause of failure. This has the advantage that the failing block can easily be singled out by the location of the bubble, while the cause of failure is described by the text inside the bubble. We find this option less invasive than a pop-up window but still as precise, and we can be sure that the blocks involved will be visible since static verification is triggered from the BYOB editor window (See Fig.5). Notice that currently we do not report the results of static verification within Snap!, since our extension only returns a compiled Boogie code which has to be verified with Boogie separately. Figure 5: Static verification compilation notification. ## V Tool support We have developed our ideas into a prototypal extension to Snap! which can be found at https://git.snt.utwente.nl/montire/verifiedsnap/. This repository also contains a set of running examples to showcase the new support for verification. These are available in the _lessons_ folder under the root directory along with an exercise sheet named _exercises.pdf_. The extension uses the same technology as the original Snap! and can be run by just opening the _snap.html_ file in most common web-browsers that support java-script. Our extension supports both dynamic and static verification of BYOB blocks. Dynamic verification is automatically triggered when executing BYOB blocks in the usual way. For static verification, a dedicated button located at the top right corner of the BYOB editor window allows to trigger the compilation of the BYOB code into an intended equivalent Boogie code. The compiled code can be then downloaded and verified with Boogie. Boogie can be run locally or on the cloud at https://rise4fun.com/Boogie/. _Dynamic verification_ has been fully integrated into the normal execution flow of a Snap! program, and thus there is no real restrictions on the characteristics of the BYOB that can be dynamically verified. For _Static verification_ , we have restricted data types to be Integers, Booleans and List of Integers. Moreover, we do not support dynamic typing of variables. Finally, we only focus on compiling an interesting subset of Snap! blocks for the sake of teaching Design-by- Contract. ## VI Conclusions This paper presented a prototypal program verification extension to Snap!. The extension is intended to support the teaching of Design-by-Contract in the later years of high schools. We paid considerable attention to the didactic aspects of our tool: the looks and feel of the extension should remain familiar to Snap! users, the syntax and structure of the new blocks should give a clear intuition about their semantics, and the error reporting should be precise and expressive. Our extension allows to analyse BYOB blocks both by runtime assertion checking and static verification. Runtime assertion checking is fully integrated into Snap! and there is no limitation on the kind of blocks that can be analysed. Static verification compiles the Snap! code into a Boogie equivalent code and the verification needs to be run outside of Snap!. Moreover, we make some restrictions on the kind of BYOB blocks we can compile, in order to keep the complexity of the prototype low. As future work we would like to lift these restrictions as much as possible by integrating the remaining Snap! blocks into the compilation and by allowing other data types to be used. Also, we would like to integrate the verification into Snap!, translating Boogie messages back to the Snap! world, to help student to interpret them. We would like to carry out an empirical study on our proposed approach. This will require the development of a concrete study plan and its evaluation in a Dutch classroom. Computer science curricula that uses blocks programming is widely and freely available [21, 22, 23, 24, 25]. Nevertheless, it is hardly spotted that they include topics around design and verification of code. The words ‘test’ or ‘testing’ are also rare around the curricula and, where mentioned, they are not sufficiently motivated. The drawbacks of teaching coding with blocks without paying attention to design nor correctness has already been analysed [26, 27]. We have not found any work on teaching these concepts in schools, nor implementations on block programming that support teaching them. | Marieke Huisman is a professor in Software Reliability at the University of Twente. She obtained her PhD in 2001 from the Radboud University Nijmegen. Afterwards she worked at INRIA Sophia Antipolis, and since 2008 at University of Twente. Her research interests are in the verification of concurrent software, as implemented in the VerCors program verifier. She is in particular interested in making verification usable in a practical setting, and she works for example on annotation generation, and support for different programming languages. ---|--- | Raúl E. Monti received his PhD in 2018 from Universidad Nacional de Córdoba. He is currently a PostDoc at the Universiteit Twente. His research interests involve the development and practical application of formal foundations and tools for analysis and verification of software and hardware systems by means of model checking and deductive verification. His work involves interacting with industry to apply his research in the verification of industrial (embedded) systems and software. ---|--- ## References * [1] S. De Gouw, J. Rot, F. De Boer, R. Bubel, and R. Hähnle, “OpenJDK’s java.utils.collection.sort() is broken: The good, the bad and the worst case,” in _Proc. 27th Intl. Conf. on Computer Aided Verification (CAV), San Francisco_ , ser. LNCS, D. Kroening and C. Pasareanu, Eds., vol. 9206\. Springer, Jul. 2015, pp. 273–289. * [2] W. Oortwijn, M. Huisman, S. J. Joosten, and J. van de Pol, “Automated verification of parallel nested dfs,” in _International Conference on Tools and Algorithms for the Construction and Analysis of Systems_. Springer, 2020, pp. 247–265. * [3] M. Safari, W. Oortwijn, S. Joosten, and M. Huisman, “Formal verification of parallel prefix sum,” in _NASA Formal Methods_ , R. Lee, S. Jha, and A. Mavridou, Eds. Cham: Springer International Publishing, 2020, pp. 170–186. [Online]. Available: https://doi.org/10.1007/978-3-030-55754-6_10 * [4] B. Meyer, “Applying "design by contract",” _Computer_ , vol. 25, no. 10, pp. 40–51, 1992. [Online]. Available: https://doi.org/10.1109/2.161279 * [5] B. Harvey, D. D. Garcia, T. Barnes, N. Titterton, D. Armendariz, L. Segars, E. Lemon, S. Morris, and J. Paley, “Snap!(build your own blocks),” in _Proceeding of the 44th ACM technical symposium on Computer science education_ , 2013, pp. 759–759. * [6] D. Garcia, B. Harvey, and T. Barnes, “The beauty and joy of computing,” _Inroads_ , vol. 6, no. 4, pp. 71–79, 2015. [Online]. Available: https://doi.org/10.1145/2835184 * [7] G. T. Leavens, A. L. Baker, and C. Ruby, “JML: A notation for detailed design,” in _Behavioral Specifications of Businesses and Systems_ , H. Kilov, B. Rumpe, and I. Simmonds, Eds. Boston, MA: Springer US, 1999, pp. 175–188. * [8] Y. Cheon, “A runtime assertion checker for the Java Modeling Language,” Ph.D. dissertation, Department of Computer Science, Iowa State University, Ames, 2003, technical Report 03-09. * [9] K. R. M. Leino, “Towards reliable modular programs,” 1995. * [10] M. Barnett, B. E. Chang, R. DeLine, B. Jacobs, and K. R. M. Leino, “Boogie: A modular reusable verifier for object-oriented programs,” in _Formal Methods for Components and Objects, 4th International Symposium, FMCO 2005, Amsterdam, The Netherlands, November 1-4, 2005, Revised Lectures_ , ser. Lecture Notes in Computer Science, F. S. de Boer, M. M. Bonsangue, S. Graf, and W. P. de Roever, Eds., vol. 4111. Springer, 2005, pp. 364–387. [Online]. Available: https://doi.org/10.1007/11804192_17 * [11] B. Harvey and J. Mönig, “Snap! reference manual,” _URL http://snap. berkeley. edu/SnapManual. pdf_ , 2017. * [12] M. Resnick, J. Maloney, A. Monroy-Hernández, N. Rusk, E. Eastmond, K. Brennan, A. Millner, E. Rosenbaum, J. S. Silver, B. Silverman, and Y. B. Kafai, “Scratch: programming for all,” _Commun. ACM_ , vol. 52, no. 11, pp. 60–67, 2009. [Online]. Available: https://doi.org/10.1145/1592761.1592779 * [13] B. Meyer, J.-M. Nerson, and M. Matsuo, “Eiffel: object-oriented design for software engineering,” in _European Software Engineering Conference_. Springer, 1987, pp. 221–229. * [14] B. Meyer, “Eiffel: A language and environment for software engineering,” _Journal of Systems and Software_ , vol. 8, no. 3, pp. 199–246, 1988. * [15] G. T. Leavens, Y. Cheon, C. Clifton, C. Ruby, and D. R. Cok, “How the design of jml accommodates both runtime assertion checking and formal verification,” _Science of Computer Programming_ , vol. 55, no. 1-3, pp. 185–208, 2005. * [16] C. Hoare, “An axiomatic basis for computer programming,” _Communications of the ACM_ , vol. 12, no. 10, pp. 576–580, 1969. * [17] E. Dijkstra, _A Discipline of Programming_. Prentice-Hall, 1976. * [18] B. Meyer, _Eiffel: The Language_. Prentice-Hall, 1991. [Online]. Available: http://www.eiffel.com/doc/#etl * [19] M. Barnett, K. R. M. Leino, and W. Schulte, “The Spec# programming system: An overview,” in _Construction and Analysis of Safe, Secure and Interoperable Smart Devices: Proceedings of the International Workshop CASSIS 2004_ , ser. LNCS, G. Barthe, L. Burdy, M. Huisman, J.-L. Lanet, and T. Muntean, Eds., vol. 3362. Springer, 2005, pp. 151–171. * [20] T. Türk, “Local reasoning about while-loops,” in _VSTTE 2010. Workshop Proceedings_ , R. Joshi, T. Margaria, P. Müller, D. Naumann, and H. Yang, Eds. ETH Zürich, 2010, pp. 29–39. * [21] “The Beauty and Joy of Computing. An AP CS Principles Course,” https://bjc.edc.org/. Accessed October 2020. * [22] “CS First,” https://csfirst.withgoogle.com/s/en/home. Accessed October 2020. * [23] “The Creative Computing Curriculum,” http://creativecomputing.gse.harvard.edu/guide/. Accessed October 2020. * [24] “An Introduction to Programming. A Pencil Code Teacher’s Manual,” https://manual.pencilcode.net/. Accessed October 2020. * [25] P. Factorovich and F. Sawady, _Actividades para aprender a Program. AR: Segundo ciclo de la educación primaria y primero de la secundaria_. Miller Ed Buenos Aires, 2015. * [26] O. Meerbaum-Salant, M. Armoni, and M. Ben-Ari, “Habits of programming in scratch,” in _Proceedings of the 16th Annual SIGCSE Conference on Innovation and Technology in Computer Science Education, ITiCSE 2011, Darmstadt, Germany, June 27-29, 2011_ , G. Rößling, T. L. Naps, and C. Spannagel, Eds. ACM, 2011, pp. 168–172. [Online]. Available: https://doi.org/10.1145/1999747.1999796 * [27] E. Aivaloglou and F. Hermans, “How kids code and how we know: An exploratory study on the scratch repository,” in _Proceedings of the 2016 ACM Conference on International Computing Education Research_ , 2016, pp. 53–61.
# A low-loss ferrite circulator as a tunable chiral quantum system Ying-Ying Wang Department of Physics, University of Massachusetts-Amherst, Amherst, MA, USA Sean van Geldern Department of Physics, University of Massachusetts-Amherst, Amherst, MA, USA Thomas Connolly Present address: Department of Applied Physics, Yale University, New Haven, CT, USA Department of Physics, University of Massachusetts-Amherst, Amherst, MA, USA Yu-Xin Wang Pritzker School of Molecular Engineering, University of Chicago, Chicago, IL, USA Alexander Shilcusky Department of Physics, University of Massachusetts- Amherst, Amherst, MA, USA Alexander McDonald Pritzker School of Molecular Engineering, University of Chicago, Chicago, IL, USA Department of Physics, University of Chicago, Chicago, IL, USA Aashish A. Clerk Pritzker School of Molecular Engineering, University of Chicago, Chicago, IL, USA Chen Wang <EMAIL_ADDRESS>Department of Physics, University of Massachusetts-Amherst, Amherst, MA, USA ###### Abstract Ferrite microwave circulators allow one to control the directional flow of microwave signals and noise, and thus play a crucial role in present-day superconducting quantum technology. They are typically viewed as a black-box, with their internal structure neither specified nor used as a quantum resource. In this work, we show a low-loss waveguide circulator constructed with single-crystalline yttrium iron garnet (YIG) in a 3D cavity, and analyze it as a multi-mode hybrid quantum system with coupled photonic and magnonic excitations. We show the coherent coupling of its chiral internal modes with integrated superconducting niobium cavities, and how this enables tunable non- reciprocal interactions between the intra-cavity photons. We also probe experimentally the effective non-Hermitian dynamics of this system and its effective non-reciprocal eigenmodes. The device platform provides a test bed for implementing non-reciprocal interactions in open-system circuit QED. ## I Introduction Microwave circulators, typically composed of a transmission line Y-junction with ferrite materials Kord _et al._ (2020), are ubiquitous in superconducting circuit QED experiments Devoret and Schoelkopf (2013). They provide a crucial link in the readout chain of superconducting quantum processors, by directing the signal traffic while protecting the qubits and resonators from thermal noise Krantz _et al._ (2019). They also enable the interactions between distinct quantum circuit modules to be non-reciprocal Kurpiers _et al._ (2018); Axline _et al._ (2018), a feature which is important for eliminating long-distance cross-talks in modular quantum computation architectures. Despite their importance, microwave circulators are generally treated as broadband black-box devices in experiments. Formulating a more microscopic quantum description is often challenging, as their internal modes involving the magnetic spin excitations (magnons) are generally too lossy and complex to be analyzed using canonical circuit quantization Vool and Devoret (2017). On the other hand, there has been growing interest in studying and manipulating magnon excitations of ferromagnetic/ferrimagnetic materials in the quantum regime Lachance-Quirion _et al._ (2019); Awschalom _et al._ (2021). In particular, the ferromagnetic resonance (FMR) mode of yttrium iron garnet (YIG), a ferrimagnetic insulator with usage in commercial circulators, has shown sufficiently high quality factor and coupling cooperativity with microwave cavities to function as a quantum oscillator mode in strong-coupling circuit QED Huebl _et al._ (2013); Tabuchi _et al._ (2014); Zhang _et al._ (2014). Notably, coherent coupling of magnons with a superconducting qubit Tabuchi _et al._ (2015) and single-shot detection of a single magnon Lachance-Quirion _et al._ (2020) have been demonstrated using a millimeter- sized single-crystalline YIG sphere in a 3D cavity. Furthermore, there is a plausible pathway towards planar superconducting-magnonic devices Hou and Liu (2019); Golovchanskiy _et al._ (2021) to connect circuit QED with spintronics technologies by advancing fabrication techniques of low-damping YIG films Heyroth _et al._ (2019). It would be interesting to harness these recent advances in the study of quantum magnonics to revisit the design of microwave circulators, potentially leading to new kinds of non-reciprocal devices in circuit QED. Our work here describes a first step in this direction. Here we demonstrate a tunable non- reciprocal device based on the waveguide circulator loaded with single- crystalline YIG, which explicitly makes use of well-characterized hybrid polariton modes. Such modes are the normal modes of coupled magnon-photon systems Huebl _et al._ (2013); Tabuchi _et al._ (2014); Zhang _et al._ (2014); Boventer _et al._ (2018); Zhang _et al._ (2017), and have an intrinsic chirality that is set by the magnetic field Anderson _et al._ (2016); Owens _et al._ (2018); Zhang _et al._ (2020). While our device follows the same basic working principles underpinning textbook circulators Kord _et al._ (2020); Fay and Comstock (1965), detailed understanding of the internal modes allows us to incorporate the physical source of non-reciprocity in the full description of a larger system including two external superconducting cavities, using a non-Hermitian effective Hamiltonian. While our device can be configured to operate as a traditional circulator for its non-reciprocal transmission of travelling waves, the main focus of our study is to use the device for mediating tunable non-reciprocal interaction between localized long-lived quantum modes. Such non-reciprocal mode-mode couplings result in distinct signatures in the eigenvalues and eigenvectors of the non-Hermitian system Hamiltonian, which is relevant to the more general study of non-Hermitian dynamics in contexts ranging from classical optics to quantum condensed matter. Anomalous properties of the eigenvalues and eigenvectors of a non-Hermitian Hamiltonian have given rise to a number of striking phenomena such as the existence of exceptional points Heiss (2012); Özdemir _et al._ (2019) and the non-Hermitian skin effect Hatano and Nelson (1997); Yao and Wang (2018); McDonald _et al._ (2018), but direct experimental access to the underlying eigenmodes is often difficult. In this study, we provide comprehensive characterization of the eigenmode structure, which is a step towards effective Hamiltonian engineering of non-reciprocal non-Hermitian systems. The most tantalizing usage of non-reciprocity in quantum systems (such as entanglement stabilization using directional interactions in chiral quantum optics setups Stannigel _et al._ (2012); Lodahl _et al._ (2017)) require extremely high quality devices. In particular, they must approach the pristine limit where undesirable internal loss rates are negligible compared to the non-reciprocal coupling rates. While many experiments have focused on new avenues of achieving non-reciprocity Chapman _et al._ (2017); Lecocq _et al._ (2017); Sliwa _et al._ (2015); Ruesink _et al._ (2016); Fang _et al._ (2017); Wang _et al._ (2019); Xu _et al._ (2019), this loss-to-coupling ratio, which can be understood as the quantum efficiency of the non-reciprocal interactions, has been typically limited to approximately 10% ($\sim$ 0.5 dB) or more, which is comparable to the linear insertion loss of typical commercial circulators as measured in modular circuit QED experiments Kurpiers _et al._ (2018); Axline _et al._ (2018). This performance lags far behind the quality of unitary operations between reciprocally coupled quantum components (i.e. two-qubit gate infidelity $<1$%). Our approach provides a route for transcending this limitation on the quantum efficiency of non-reciprocal interactions. The results of our study have implications in several areas: (1) In the context of quantum magnonics, we present the first study of polariton modes with a partially magnetized ferrite material, which features a high quality factor and low operating field, both of which are crucial for constructing superconducting-magnonic devices. (2) In the context of modular superconducting quantum computing, we demonstrate the first circulator with internal loss well below 1% of the coupling bandwidth, which would enable high-fidelity directional quantum state transfer. (3) For the general non- Hermitian physics, we demonstrate an experimental probe of the non-reciprocal eigenvector composition of a non-Hermitian system. Combining these advances, we have established an experimental platform that meets the conditions for future study of nonlinear non-reciprocal interactions with superconducting qubits. ## II Experimental Setup Figure 1: Device and measurement setup. (a) A YIG cylinder (black) is placed at the center of the intersection of three rounded-rectangular waveguides placed 120 degrees away from each other. The light grey region is vacuum inside an oxygen-free copper enclosure. The device can be assembled in two different configurations: First, a drum-head shaped transition pin can be attached at the end of each waveguide section to form an impedance matched waveguide-to-SMA transition (IMT). Alternatively, a short weakly-coupled probe (WCP) can be attached to each waveguide section to explore the internal modes of the device. (b) The device is mounted to a mezzanine plate that is thermalized to the mixing chamber of a dilution refrigerator, and is positioned at the center of a superconducting solenoid magnet which operates at 4K. The device is connected to three input cables (with attenuators as marked) and two output amplifier lines (with directional couplers splitting signals) for $S$-parameter measurements using a vector network analyzer (VNA). Our experimental setup is shown in Fig. 1(a). Three rounded-rectangular waveguides, each with a cross section of 21.0 mm $\times$ 4.0 mm, placed 120 degrees away from each other, intersect to form the body of the circulator. A $\phi$-5.58 mm $\times$ 5.0 mm single-crystalline YIG cylinder is placed at the center of the Y-junction, with external magnetic fields applied along its height (the $z$ axis and the [111] orientation of the YIG crystal). At the end of the three waveguide sections, we can either attach impedance-matched waveguide-to-SMA transitions (IMT) to perform standard characterization of the circulator (as in Section IV), or attach weakly-coupled probe pins (WCP) to explore the internal modes of this YIG-loaded Y-shaped cavity (as in Section III). The use of reconfigurable probes in the same waveguide package allows us to infer the operation condition and the performance of the circulator from the properties of the internal modes. Furthermore, the copper waveguide sections can be replaced by superconducting niobium cavities, with details to be described in Section V and Fig. 5. This modular substitution introduces additional external high Q modes to the system, and understanding the resulting Hamiltonian and the hybridized mode structure of the full system will be a first step towards the study of pristine non-reciprocal interactions in circuit QED. The device package is thermalized to the mixing chamber plate ($\sim$20 mK) of a Bluefors LD-250 dilution refrigerator inside the $\phi$-100 mm bore of a 1 T superconducting magnet that applies magnetic field along the $z$ axis [Fig. 1(b)]. A vector network analyzer is used to measure the complex microwave transmission coefficients $S_{ij}$ (from Port $j$ to Port $i$, where $i,j=1,2,3$) of the device in series with a chain of attenuators, filters and amplifiers as in typical circuit QED experiments. A magnetic shield made of a steel sheet is placed outside the bottom half of the refrigerator, and all data is acquired under the persistent mode of the superconducting magnet to minimize magnetic-field fluctuations. ## III Internal mode structure Figure 2: Internal mode spectrum of the device. (a) VNA transmission measurement $S_{21}$ of multi-mode photon-magnon hybrid system formed in the waveguide circulator package with WCP. The blue (red) dashed line plots the frequency of the clockwise(counterclockwise) mode from a simplified two-mode model of photon-magnon avoided crossing with $g/2\pi$ = 1.3 GHz (2.1 GHz) to compare with an observed spectral line. (b) The right panel shows a finer sweep of $S_{21}$ in the low-field regime. The mode frequencies differ slightly from (a) since the data was acquired after some modifications to the device packaging (a piece of Teflon spacer at the top of the YIG cylinder was removed). The left panel shows the electromagnetic mode structures of the eigenmode solutions from our HFSS simulation for the WCP with good frequency agreement to the experimental data (see Fig. 8 in Appendix). The color scale from red to blue represents electric field strength from high to low in log scale. The pair of modes around 11 GHz are connected to the circulating modes of the loaded circulator and (c) shows their linewidths. We begin by discussing the internal mode structure of the device, as probed by $S_{21}$ as a function of applied magnetic field $B$ when the device is installed with WCP [Fig. 2(a)]. A series of electromagnetic modes (relatively field-independent) are observed to undergo large avoided crossings with a cluster of magnon modes of the YIG crystal, forming photon-magnon polariton modes. The magnon mode most strongly coupled to photons is known to correspond to near-uniform precession of YIG spins, or the Kittel mode of FMR, whose frequency increases linearly with magnetic field: $\omega_{m}=\gamma[B+\mu_{0}(N_{x,y}-N_{z})M_{s}]\approx\gamma B$, as marked by the dashed line in Fig. 2(a). Here $\gamma$ is the gyromagnetic ratio, and the (volume-averaged) demagnetizing factors $N_{x,y,z}$ in magnetic saturated state are very close to $1/3$ for the aspect ratio of our YIG cylinder Chen _et al._ (1991). These avoided crossings are similar to previous experiments showing strong photon-magnon coupling Tabuchi _et al._ (2014); Zhang _et al._ (2014), but due to the much larger size of the YIG in our experiment, a large cluster of higher-order magnetostatic modes, most of which have slightly higher frequency than the Kittel mode Fletcher _et al._ (1960); Klingler _et al._ (2017) also coherently interact with the microwave photons, contributing to the complex transmission spectra in the vicinity of the crossings. Nevertheless, to have a coarse estimate of the photon-magnon coupling strength, it is convenient to model each observed spectral line far away from the crossing region as a bare electromagnetic mode with frequency $\omega_{c}/2\pi$ hybridized with a single combined magnon mode. The implied coupling strengths $g/2\pi$ (in the cavity QED convention) are about 1.2 GHz and 2.1 GHz for the two modes of particular interest to this study [blue and red in Fig. 2(a)], placing the mode hybridization in the ultrastrong coupling regime (see e.g. Marković _et al._ (2018)) with $g/(\omega_{c}+\omega_{m})\sim 10\%$. Even at $B=0$, with a photon-magnon detuning of $\Delta=\omega_{c}-\omega_{m}\approx 2\pi\cdot 10$ GHz, the participation of magnon excitations in the photon-branch of the polariton modes remains quite substantial. Using finite-element simulations (Ansys HFSS, Appendix A), we identify that the five polariton modes in the frequency range of 8-12 GHz at $B=0$ include two nearly-degenerate mode pairs with two-fold symmetry and another mode with three-fold symmetry. Electric field distributions of each of the modes are illustrated in Fig. 2(b). Each degenerate mode pair can be understood using a basis of standing-wave modes polarized along the $x$ or $y$ direction. The application of a magnetic field lifts this $x$-$y$ degeneracy, as the mode pair forms clockwise and counterclockwise rotating eigenmodes with a frequency splitting Owens _et al._ (2018); Zhang _et al._ (2020); Anderson _et al._ (2016). Prior use of these chiral polariton mode pairs have been in the magnetically saturated regime Anderson _et al._ (2016); Owens _et al._ (2018); Zhang _et al._ (2020). Here we focus on the low-field regime ($|B|<0.05$ mT) where the approximately linear increase of frequency splitting between the mode pair reflects increasing magnetization of YIG under increasing applied magnetic field. After implementing demagnetization training cycles to suppress a relatively small hysteretic effect throughout our experiments, we expect an approximately linear magnetization curve ($M$-$H$) for YIG until it approaches magnetic saturation. In the limit of high permeability $\mu\gg\mu_{0}$ (with $\mu_{0}$ being the vacuum permeability), we have $M=B/{\mu_{0}N_{z}}$ (note that $B$ is the applied magnetic field strength) and $N_{z}\approx 0.285$ is the $z$-direction demagnetizing factor when the YIG is significantly below magnetic saturation Chen _et al._ (1991). Saturation magnetization $M_{s}=2440$ Oe Solt (1962) of YIG is approached on the scale of $B\sim\mu_{0}N_{z}M_{s}\approx 70$ mT, which agrees with the changing curvature of the mode-splitting spectra. On the other hand, in the completely demagnetized state ($M=0$) at zero field, the system is expected to satisfy macroscopic time-reversal symmetry. As supported by numerical simulations, the $x$-$y$ mode pairs should be in principle exactly degenerate since both the Y-junction geometry and the [111] YIG crystal has 3-fold rotational symmetry around the $z$ axis. However, appreciable zero-field splitting is observed experimentally. We attribute this splitting to some anisotropy in the x-y plane breaking this symmetry and allowing a preferred magnetization axis of the YIG at 0 field. Some possible explanations for this anisotropy are a small visible damage to our YIG crystal on one edge or possible imperfections in eccentricity and alignment. If the magnetic domains of unsaturated YIG preferentially align with one in-plane axis compared to its orthogonal axis within the $x$-$y$ plane, this anisotropy would result in a relative frequency shift between the standing-wave modes along the in-plane easy and hard axes. This anisotropy-induced frequency shift $\pm\beta$ for the $x$ and $y$ modes can be modeled in numerical simulations employing a permeability tensor of unsaturated ferromagnets Schlömann (1970); Green and Sandy (1974) with certain anisotropic assumption, which can plausibly explain the data (Appendix A). As $B$ increases, we expect $\beta$ to decay towards 0 when the magnetic domains are increasingly aligned towards the $z$ direction, thus making any $x$-$y$ plane energetic preference of negligible effect. We model this decay with a thermodynamic toy model (Appendix B) whose details do not affect the conclusions of this study. For the rest of this article, we will focus on the pair of polariton modes near 11 GHz in Fig. 2(b), and refer to them as “the circulator modes” for reasons that will become apparent. We can model their frequencies in the partially magnetized regime ($|B|<50$ mT) using a phenomenological model accounting for the degeneracy-lifting anisotropy and the field-dependent magnetization of YIG. Let the zero-field frequencies of the $x$ and $y$ modes be $\omega_{x}=\omega_{y}$ if the device had perfect 3-fold symmetry, $\beta$ and $\theta/2$ be the magnitude of anisotropy caused degeneracy-lifting and the direction of the in-plane anisotropy axis (relative to the $x$ axis), and off-diagonal imaginary coupling term $\pm ikB$ be the magnetic field induced degeneracy-lifting, linearly increasing with a real coefficient $k$. We use the following Hamiltonian to characterize the pair of circulator modes in the basis of $x$ and $y$ mode amplitudes: $H/\hbar=\begin{pmatrix}\omega_{x}+\beta\cos{\theta}+mB^{2}&\beta\sin{\theta}+ikB\\\ \beta\sin{\theta}–ikB&\omega_{y}-\beta\cos{\theta}+mB^{2}\end{pmatrix}$ (1) This effective model of the polariton modes has absorbed the magnon contributions in the regime where they have been adiabatically eliminated. The formation of clockwise and counterclockwise eigenmodes is due to magnon- mediated interactions modeled by $\pm ikB$. The level repulsion from the far- detuned magnon modes is approximated by a small quadratic shift in frequency $mB^{2}$. The quadratic dependence was empirically chosen because the sum of the mode frequencies over field displayed a roughly quadratic relationship with $B$ over the plotted field range. By fitting the mode spectrum in Fig. 2(b), we obtain $\omega_{x}/2\pi=\omega_{y}/2\pi=11.054$ GHz, $k/2\pi=9.82$ GHz/T, $m/2\pi=50$ GHz/T2, $\beta/2\pi=139$ MHz. Figure 3: Illustration of the circulator working principle and low-temperature characterization of the non-reciprocity. In the circulator package with IMT, the frequency splitting of clockwise and counterclockwise rotating modes as shown in (a) can be tuned such that the phase of the modes are $\pi/6$ and $-\pi/6$ as shown in (b). This then produces a node at the upper port, thereby preventing any signal from leaving there at all times where $\omega t=0$ and $\omega t=\pi/4$ are shown pictorially in (c). (d, e) Measured microwave transmission (d) $|S_{12}|$ and (e) $|S_{21}|$ spectra as a function of magnetic field B. (f-i) The isolation performance, (f) $\mathcal{I}_{12}=|S_{12}/S_{21}|$,(h) $\mathcal{I}_{21}=|S_{21}/S_{12}|$, (g) $\mathcal{I}_{23}=|S_{23}/S_{32}|$, (i) $\mathcal{I}_{13}=|S_{13}/S_{31}|$. $S_{21}$ is obtained by measuring the $S_{12}$ at –$B$, which provides a self- calibrated way to determine the isolation of the circulator. It is well-known that the FMR modes of partially magnetized ferrimagnetic insulators, where the magnetic domains are not aligned in equilibrium, have large damping. Therefore, one may expect broad linewidths for photon-magnon polariton modes below magnetic saturation. Indeed, we have observed linewidths exceeding 100 MHz for another polariton mode at 5 GHz at $B<$ 50 mT (not shown). Surprisingly, the polariton modes at higher frequency display narrow linewidths, $\kappa_{i}\approx 2$ MHz for the pair of circulator modes [Fig. 2(c)], which corresponds to quality factors on par with some circuit QED elements such as the readout resonators. The narrow internal linewidth of the circulator modes is crucial for constructing a low-loss circulator and eventually achieving high quantum efficiency of non-reciprocal interactions in circuit QED. It is primarily aided by the use of single crystalline YIG and the relatively low magnon participation in the circulator modes compared to commercial circulators. The observed $\kappa_{i}$ may be limited by either the spin relaxation in YIG or the Ohmic loss in copper. The former remains to be investigated in this partially magnetized regime, and the latter may be further reduced through better surface treatment or the use of superconducting materials in low-field regions of the waveguide package. ## IV Circulator characterization The device acts as a circulator when the end of each waveguide section is in IMT rather than WCP with an applied magnetic field in the $\hat{z}$ direction. In this configuration, the linewidths of all internal modes are substantially broadened forming a transmission continuum in the measurement, as shown in Fig. 3(d,e) for $|S_{12}|$ and $|S_{21}|$. Nevertheless, the operating condition of the circulator can be conceptually understood as having a pair of counter-propagating internal modes with their magnetic-field-induced splitting ($\delta$) satisfying the relationship $\delta=2\kappa_{c}/{\sqrt{3}}$ versus their half linewidths ($\kappa_{c}$) Fay and Comstock (1965). As illustrated in Fig. 3(a-c), when driven at a frequency in the middle of the two resonances, the two circulator modes are excited with equal amplitude and a phase shift of $\pm\frac{\pi}{6}$ relative to the drive. The resultant standing wave pattern forms a node at the isolation port of the circulator. This condition can be satisfied by choosing the correct combination of frequency and magnetic field. We characterize the non-reciprocity of the circulator by the isolation ratio $\mathcal{I}_{12}=|S_{12}/S_{21}|$, which may be computed from Fig. 3(d,e). However, since $S_{12}$ and $S_{21}$ are measured through different cables and amplifier chains [Fig. 1(b)], it is challenging to calibrate their absolute values precisely. A much better self-calibrated technique to extract the isolation ratio in our system is to use the Onsager-Casimir relation Casimir (1945), $S_{21}(B)=S_{12}(-B)$, resulting from the microscopic time reversal symmetry. Therefore, we use $\mathcal{I}_{12}=|S_{12}(B)/S_{12}(-B)|$ to determine the isolation ratio of the circulator as shown in Fig. 3(f), with the (field-independent) contribution from same transmission chain cancelled out. The result indicates the circulator working condition is met for the pair of counter-propagating modes at $\sim$11.2 GHz with external field $\sim$0.022 T. We see $\geq$20 dB of isolation over a bandwidth of about 250 MHz, with maximum isolation of at least 35 dB. The same analysis on $S_{21}$ data yields the same isolation property [Fig. 3(h)] as expected. Similarly, $\mathcal{I}_{23}$ and $\mathcal{I}_{13}$ are measured as in Fig. 3(g) and (i), each showing a slightly different working field and frequency (possibly due to imperfections of the device geometry) but similar isolation magnitude and bandwidth. These data are measured at an estimated circulating photon number on the order of 10’s, but when we lower the power to below single photon level, the isolation property does not show notable changes. An important motivation of our work is to ultimately implement pristine non- reciprocal interactions between superconducting qubits or cavities. It is crucial to minimize the ratio between the undesirable internal dissipation ($\kappa_{i}$) and the external bath coupling ($\kappa_{c}$) that enables non- reciprocity. In the case of a circulator, this ratio sets the limit for the circulator’s microwave insertion loss $\mathcal{L}_{21}$ Kord _et al._ (2020); Fay and Comstock (1965): $\mathcal{L}_{21}=1-|S_{21}|^{2}\geq 1-|S_{21}|^{2}-|S_{11}|^{2}-|S_{31}|^{2}\approx\frac{\kappa_{i}}{\kappa_{c}}$ (2) Figure 4: Characterization of the internal loss of the circulator at room temperature. (a) Linewidths of the pair of circulator modes measured at room temperature. (b) Transmission $S_{21}$ and reflection $S_{11}$ near the maximum isolation regime of the circulator, measured at $B=24.8$ mT (top panel) and the internal loss of the circulator calculated from it (bottom panel) at room temperature. This lower limit is obtained in principle when the circulator has perfect impedance matching ($S_{11}=0$) and isolation ratio ($S_{31}=0$). Typical commercial ferrite circulators used in circuit QED experiments have shown insertion loss around 10% Kurpiers _et al._ (2018); Axline _et al._ (2018), which is dominated by internal loss. Experimental Josephson circulators so far have also reported insertion loss of -0.5 dB (11%) or higher Chapman _et al._ (2017); Lecocq _et al._ (2017). The lowest quoted insertion loss for a commercially-listed waveguide circulator is -0.1 dB (or 2.2%) but that is untested in the quantum regime. In order for the quantum efficiency of a non- reciprocal two-qubit interaction channel to match the fidelity of state-of- the-art two-qubit operations, the insertion loss would need to be improved to the sub-percent level. To the best of our knowledge, it is an open challenge to calibrate the insertion loss of a microwave component in a dilution refrigerator with a precision better than 1%. Even using specialized Thru-Reflect-Line calibration components and well-characterized cryogenic switches, the resultant precision would still be limited to about 0.1dB (or 2.3%) Ranzani _et al._ (2013). In order to infer the loss of our circulator at 20 mK, we measure its $S$ parameters at room temperature after a careful calibration procedure that uses attenuators in series to suppress standing waves. We find a conservative upper bound for room-temperature internal loss of $\leq 1–|S_{21}|^{2}–|S_{11}|^{2}\approx 2\%$, as shown in Fig. 4(b). Assuming $\kappa_{c}$ does not change as a function of temperature, comparing the intrinsic linewidth of the circulator mode pair at room temperature versus 20 mK would inform the internal loss at 20 mK. The intrinsic linewidths, measured in WCP, are 4.1 and 6.3 MHz at room temperature [Fig. 4(a)] and 1.8 and 2.2 MHz at low temperature [Fig. 2(c)], indicating that $\kappa_{c}\gtrsim 260$ MHz and $\kappa_{i}/\kappa_{c}\lesssim 0.8\%$. If we instead use the relation of $\kappa_{c}=\sqrt{3}\delta/2$, which yields $\kappa_{c}$ in the range of 430 MHz to 550 MHz (and data in Section V would further suggest $\kappa_{c}$ at the high end of this range), or $\kappa_{i}/\kappa_{c}\approx 0.4\%$. Further improvement of the circulator bandwidth and the coupling ratio can be achieved by applying impedance transformation techniques to increase $\kappa_{c}$ Helszajn (2008). Translating this small internal loss ratio to a sub-percent insertion loss for a circulator as a peripheral transmission-line device would further require excellent impedance matching. However, we emphasize that this requirement is not fundamental if the circulator is modeled as part of the quantum system itself mediating interactions between other quantum resonance modes. Unlike most ferrite circulators, our device operates in the regime of partial magnetization for YIG. It only requires a moderate external magnetic field that is significantly below the critical field of a variety of superconducting materials. This allows for 3D integration of superconducting niobium cavities and shielded transmon qubits for studying circuit QED with non-reciprocal interactions. In the following section, we demonstrate direct coupling of two external superconducting cavity modes with the circulator modes and analyze the resultant non-reciprocal hybrid system as a whole. ## V Tuning non-reciprocity of eigenmode structure Figure 5: Waveguide circulator-cavity integration. (a) The photo image, (b) a schematic top-down view, and (c) a diagrammatic illustration of the effective Hamiltonian (see Eq. (4), for clarity the $\beta$ and $mB^{2}$ terms have been neglected in the illustration) of our integrated non-reciprocal device. It is composed of a Cu waveguide Y-junction loaded with a YIG cylinder, two Nb cavity segments with weakly-coupled drive ports (Port 1 and 2), and an output port with IMT (Port 3). For each cavity, the sidewall closest to the copper Y-junction is formed by a standalone niobium plate in the assembly [enclosed in blue in (a)], which contains a 5 mm-diameter aperture to create an evanescent coupling between the superconducting cavity mode and the circulator modes. One of the cavities is loaded with a transmon qubit [marked as $\times$ in (b)] which stays unused in its ground state in this study. We integrate superconducting cavities with the ferrite device by replacing the rectangular waveguide extensions with superconducting 3D cavities made of niobium [Fig. 5(a)]. Two cavities, attached at Port 1 and 2, are tuned to have resonance frequencies close to each other, $\omega_{1}\approx\omega_{2}\sim 10.8$ GHz, both of which are within the bandwidth of the circulator. Each cavity is coupled to the central Y-junction via a coupling aperture. As a result, the circulator modes will mediate an interaction between these two external cavities. Crucially, this circulator-mediated interaction can have both coherent and dissipative aspects, and can be non-reciprocal. The degree and the direction of non-reciprocity of the coupling can be tuned via the external magnetic field. Note that Port 3 remains impedance-matched to a transmission line. This is also essential: it serves as the dominant dissipative bath that is necessary for achieving non-reciprocal inter-mode interactions Metelmann and Clerk (2015). To probe the hybridized mode structure of the composite system, we measure $S_{31}$ from a weakly-coupled drive port on Cavity 1 to Port 3. The measured amplitude of $S_{31}$ as a function of magnetic field and frequency is shown in Fig. 6(a). There are a total of four bare oscillator modes in the vicinity (within 0.5 GHz) of the frequency range of interest: two superconducting cavity modes and two internal circulator modes. Since the loaded circulator modes with very broad linewidths ($>100$ MHz) are difficult to observe in the presence of the standing-wave background of the coaxial cables, this spectroscopy measurement primarily reveals the eigenmodes that are localized in the external superconducting cavities. Indeed, at $|B|>0.03$ T, the spectrum shows two sharp resonances which we identify as the bare cavity modes to a good approximation. At lower fields, the cavities appear to more strongly hybridize with each other and with the lossy circulator modes, but the spectrum can nontheless be captured relatively well by the sum of two Lorentzian modes $a$ and $b$: $S_{31}=\frac{A_{a}e^{i\phi_{a}}}{{-i(\omega-\omega_{a})-\kappa_{a}/2}}+\frac{A_{b}e^{i\phi_{b}}}{{-i(\omega-\omega_{b})-\kappa_{b}/2}}$ (3) By fitting the spectrum to Eq. (3), we can extract the linewidth ($\kappa_{i}$), frequency ($\omega_{i}$) and amplitude ($A_{i}$) of the two Lorentzians at each magnetic field, as plotted in Fig. 6(c-e). The magnetic field dependence of the two prominent Lorentzians can be connected to the eigenmode solutions of an effective Hamiltonian model of system. We describe the system using the following 4$\times$4 non-Hermitian matrix, written in the basis of the amplitudes of the two cavity modes and the two circulator modes: $H_{\mathrm{eff}}/\hbar=\\\ \begin{pmatrix}\;\omega_{1}-i\frac{\kappa_{1}}{2}\;&0&g_{y}&g_{x}\\\ 0&\;\omega_{2}-i\frac{\kappa_{2}}{2}\;&g_{y}&-g_{x}\\\ g_{y}&g_{y}&\omega_{y}-\beta\cos\theta+mB^{2}-i\frac{\kappa_{3}}{2}&\beta\sin\theta- ikB\\\ g_{x}\ &-g_{x}\ &\beta\sin\theta+ikB&\omega_{x}+\beta\cos\theta+mB^{2}\\\ \end{pmatrix}$ (4) The two niobium cavities have bare frequencies $\omega_{1}$, $\omega_{2}$, and input coupling rates of $\kappa_{1}$ and $\kappa_{2}$. The bottom right block of Eq. (4) describes the two circulator modes, with their anisotropy dependence and imaginary coupling due to magnon hybridization following the same description as in Eq. (1). The zero-field frequencies of the two circulator modes $\omega_{x}$, $\omega_{y}$ are no longer equal since the device is no longer 3-fold symmetric. The $y$-mode with frequency $\omega_{y}$ is symmetric with respect to the $y$ axis, and therefore has an equal and in- phase coupling rate $g_{y}$ with the two cavities. It rapidly leaks to the waveguide output Port 3, with a decay rate $\kappa_{3}\gg\kappa_{1},\kappa_{2},g_{x},g_{y}$. $\kappa_{3}$ is related to $\kappa_{c}$ of the loaded circulator as in Section IV by $\kappa_{3}=4\kappa_{c}/3$. The $x$-mode is anti-symmetric with respect to the $y$ axis, preventing it from coupling to the output port. This also leads to a 180∘ phase difference in cavity coupling as accounted for by the negative sign on two of the $g_{x}$ parameters Figure 6: Spectroscopy of the hybridized non-reciprocal modes of a circulator- cavity system. (a) VNA transmission measurement and (b) model prediction of $|S_{31}|$ frequency spectrum over external magnetic field $B$. Remaining panels show magnetic field dependence of the system’s eigenmodes and wavefunctions: (c) eigenmode linewidths $\kappa_{n}/2\pi$, (d) eigenmode frequencies $\omega_{n}/2\pi$, (e) amplitude parameter $A_{n}$ (c.f. Eq.(3)), and (f) amplitude ratio (c.f. Eq.(8)) of experimental data (dots) from two- mode Lorentzian fit (Eq.(3)) and theory predictions (dashed lines). The symmetry of $\kappa_{n}$ and $\omega_{n}$ ($i.e.~{}$ complex eigen-energy of the hybrid system) with respect to $B$ exemplifies the microscopic time- reversal symmetry of the non-Hermitian system. The non-reciprocity is reflected in the difference in $A_{n}$ at $\pm B$, which reveals the asymmetry in the left/right eigen-vector structure (c.f. Eq.(8))). The effective Hamiltonian parameters from the fit are: $\omega_{1}/2\pi=10.8104$ GHz, $\omega_{2}/2\pi=10.8040$ GHz, $\omega_{x}/2\pi=10.707$ GHz, $\omega_{y}/2\pi=10.813$ GHz, $\theta=37.7^{\circ}$, $\kappa_{3}/2\pi=730$ MHz, $g_{x}/2\pi=(9.0+0.011\beta)$ MHz, $g_{y}/2\pi=(5.0+0.006\beta)$ MHz with $\beta/2\pi=139$ MHz at $B=0$ and decays with $|B|$. This effective non-Hermitian Hamiltonian can be diagonalized as: $H_{\mathrm{eff}}=\sum_{n}\hbar\omega_{n}\ket{n_{R}}\bra{n_{L}}$ (5) where $n=a,b,c,d$ are the eigenmode indices of the system, $\omega_{n}$ the complex eigen-frequencies, and $\ket{n_{R}}$ and $\ket{n_{L}}$ the right and left eigenvectors of the non-Hermitian Hamiltonian, defined as: $H_{\mathrm{eff}}\ket{n_{R}}=\hbar\omega_{n}\ket{n_{R}}$ and $H_{\mathrm{eff}}^{\dagger}\ket{n_{L}}=\hbar\omega_{n}^{*}\ket{n_{L}}$. The scattering matrix element $S_{ij}$ from Port $j$ to Port $i$ can be generally drived from the input-output theory relation: $S_{ij}=\delta_{ij}-i\sqrt{\kappa_{i}\kappa_{j}}G_{ij}^{R}(\omega)$, where the $4\times 4$ retarded matrix Green’s function is defined as: $G^{R}(\omega)=(\omega-H_{\mathrm{eff}})^{-1}$, and $\kappa_{i}$ and $\kappa_{j}$ are the output and input coupling rates, respectively. Applying this formalism to the $S_{31}$ measurement of our device, we arrive at the following Lorentzian spectral decomposition to describe the spectrum: $S_{31}(\omega)=\sum_{n}\frac{-i\sqrt{\kappa_{1}\kappa_{3}}\bra{y}\ket{n_{R}}\bra{n_{L}}\ket{1}}{\omega-\omega_{n}}$ (6) where the real and imaginary parts of the eigen-frequency $\omega_{n}$ correspond to the observed Lorentzian frequencies and half linewidths, respectively. The amplitudes of the Lorentzians are proportional to the product of the left eigenvector overlap with the bare cavity mode $\ket{1}$ and the right eigenvector overlap with the output circulator mode $\ket{y}$. By fitting the extracted Lorentzian parameters of the two prominent eigenmodes in Fig. 6(c-e) to the predictions of the $4\times 4$ Hamiltonian model across all fields (Eq. 4), we can determine all the free Hamiltonian parameters in this model. This includes $\kappa_{3}=730$ MHz, implying $\kappa_{c}=550$ MHz for the loaded circulator, consistent with (and at the high end of) the estimates in Section IV. Somewhat surprisingly, the experimental data strongly suggests that the cavity-circulator coupling rates $g_{x}$ and $g_{y}$ must be magnetic field dependent. (For example, it heavily constrains that $g_{x}/2\pi>16$ MHz near $B$ = 0 and $g_{x}/2\pi<12$ MHz at $|B|>30$ mT.) We attribute this varying coupling to the change in electromagnetic field distribution of the $x$\- and $y$-modes around the coupling aperture due to the $x$-$y$ anisotropy of YIG. Assuming $g_{x}$ and $g_{y}$ contains a contribution proportional to $\beta(B)$ with the same decay shape over applied field, the effective Hamiltonian model fits the Lorentzian parameters quite well and also reproduces the overall transmission spectrum [Fig. 6(b)]. The eigenmode features of the system can be understood intuitively by considering first the inter-mixing of the $x$, $y$ circulator modes (i.e. diagonalization of the lower right block of $H_{\mathrm{eff}}$) and then their mixing with the two cavity modes. At $B=0$, the circulator modes are relatively close in frequency to the bare cavities, resulting in strong four- mode hybridization and substantial linewidth-broadening and frequency shift to Mode $a$. As $B$ increases, the block-diagonalized circulator modes split further in frequency in response to increasing magnetization of YIG (analogous to the unloaded internal mode spectrum in Fig. 2b), and become more detuned from the bare cavities, so the cavity-circulator hybridization is continuously reduced. This is reflected in the eventual flattening of the frequency and linewidth of the observed Lorentzians at high fields. In our device, opposite magnetic fields produce opposite directions of non- reciprocity, hence the transmission spectra observed at $\pm B$ in Fig. 6(a) are markedly different. Interestingly, the extracted data in Fig. 6(c,d) shows that the underlying eigenmode frequency and linewidths at $\pm B$ are equal, unchanged under the mapping $\mathcal{P}$ of $B:\mapsto-B$. This is no coincidence, but is rather the direct consequence of microscopic symmetry requirements. Recall again that the Onsager-Casimir relation Casimir (1945) requires that the full scattering matrix $S$ satisfy $S(-B)=S^{\mathrm{T}}(B)$. As $S$ is however directly determined by our non- Hermitian Hamiltonian, this necessarily implies that $H^{\phantom{T}}_{\mathrm{eff}}(-B)=H_{\mathrm{eff}}^{\mathrm{T}}(B)$. This in turn implies that the complex eigenvalues of $H_{\mathrm{eff}}$ are unchanged under $\mathcal{P}$. Note that the operation $\mathcal{P}$ is not just a simple time-reversal operation, as it does not involve transforming loss to gain (and vice versa). This property of $H_{\mathrm{eff}}$ can be easily seen to hold for our specific model in Eq. (4). Nonetheless, we emphasize that our experimental observation here of eigenvalue invariance under the mapping $\mathcal{P}$ is a demonstration of a general physical property; it is by no means contingent on the specifics of our model. While the eigenvalues of $H_{\mathrm{eff}}$ do not directly reflect the non- reciprocal physics of our system, the same is not true of its eigenvectors. As it involves matrix transposition, the operation $\mathcal{P}$ exchanges the left and right eigenvectors of the effective Hamiltonian: $\ket{n_{R}(B)}=\ket{n_{L}(-B)}^{*}$. A defining feature of a non-reciprocal Hamiltonian is that the left and right eigenvectors generally differ in their spatial structures (i.e. they look very different when expressed in a basis of bare modes): $R_{i,n}=\frac{|\bra{n_{L}}\ket{i}|}{|\bra{i}\ket{n_{R}}|}\neq 1$ (7) As has been discussed elsewhere McDonald and Clerk (2020); Schomerus (2020), the $R_{i,n}$ characterizes a fundamental asymmetry in the response of our system. The numerator characterizes the susceptibility of the eigenmode $n$ to a perturbation or excitation entering from bare mode $i$. In contrast, the denominator tells us the amplitude on bare mode $i$ that would result given that the system eigenmode $n$ is excited. In a Hermitian system these quantities are necessarily identical, expressing a fundamental kind of reciprocity between susceptibility and response. In our non-Hermitian system, the non-unity ratio here reflects the effective non-reciprocity of the inter- mode interactions. This non-reciprocal eigenvector structure is experimentally verified by the asymmetry of the Lorentzian amplitudes with respect to $B$ in Fig. 6(e), $\frac{A_{n}(B)}{A_{n}(-B)}=\bigg{|}\frac{\bra{n_{L}}\ket{1}}{\bra{1}\ket{n_{R}}}\frac{\bra{y}\ket{n_{R}}}{\bra{n_{L}}\ket{y}}\bigg{|}=\frac{R_{1,n}}{R_{y,n}}$ (8) We plot this ratio in Fig. 6(f). For our device, a calculation based on Eq. (4) shows that $R_{y,n}\approx 1$ for most of the field range (near zero field and $|B|\gtrsim 15$ mT), allowing Fig. 6(f) to be understood as a measurement of the non-reciprocity ratio $R_{1,n}$, in this field range, showing the role of Cavity 1 in the two prominent eigenmodes of the system. In particular, the most pronounced asymmetry is observed near the optimal working point of the circulator ($B=\pm 28$ mT) for Mode $b$, which can leak through cavity 1 but cannot be excited from Cavity 1 or vice versa, as expected for a mode dominated by photons in Cavity 2. Figure 7: Mediated non-reciprocal coupling rates between the external superconducting cavities. Red curves show the off-diagonal coupling terms in the effective two-mode Hamiltonian (c.f. Eq.(9)), $\lvert H_{12}\rvert$ and $\lvert H_{21}\rvert$, as a function of magnetic field, and blue shows $r=\sqrt{\lvert H_{21}\rvert/\lvert H_{12}\rvert}$. It is also interesting to discuss our system in the context of general systems exhibiting non-reciprocal interactions between constituent parts. Such systems are commonly described by phenomenological non-Hermitian Hamiltonian matrices $H$, whose matrix elements in a local basis encode interactions with directionality: $|H_{ij}|\neq|H_{ji}|$. A prominent example is the Hatano- Nelson model Hatano and Nelson (1997) of asymmetric tunneling on a lattice. In our case, we have a microscopically-motivated model that is fully consistent with the requirements of microscopic reversability, but which encodes non- reciprocity. As shown in appendix C, one can adiabatically eliminate the internal circulator modes from our system to obtain an effective two-mode non- Hermitian Hamiltonian that describes the external cavity modes and their circulator-mediated interaction: $\displaystyle H^{\prime}_{\mathrm{eff}}/\hbar=\begin{pmatrix}\omega_{1,\mathrm{eff}}-i\frac{\kappa_{1,\mathrm{eff}}}{2}&H_{12}\\\ H_{21}&\omega_{2,\mathrm{eff}}-i\frac{\kappa_{2,\mathrm{eff}}}{2}\\\ \end{pmatrix}$ (9) The field-tunable non-reciprocity can be seen in the asymmetry of the off- diagonal coupling values $H_{12}$ and $H_{21}$ based on the model, as plotted in Fig. 7. We note that the scale of $H_{12}$ and $H_{21}$ of a few MHz (which can be increased by using a larger coupling hole) is much larger than the achievable internal loss of the superconducting cavities , making the non- reciprocal coupling the dominant interaction. Furthermore, this reduced Hamiltonian and its eigenvectors can be mapped to a reciprocal Hamiltonian and associated eigenvectors using a similarity transformation $S(r)$, where $\sqrt{\lvert H_{21}\rvert/\lvert H_{12}\rvert}\equiv r$, as outlined in Appendix C. The similarity transform effectively localizes the mode participation on the lattice site in the direction of stronger coupling more than would be expected in the reciprocal case, causing the amplitude ratios ($R_{i,n}$) to deviate from 1, which can be viewed as a consequence of the non-Hermitian skin effect on a two site Hatano- Nelson lattice Yao and Wang (2018). Furthermore,the similarity transformation can be used to explain the qualitative behavior of the disparate amplitude ratios seen in Fig. 6(f). In particular, for $B\gtrsim 20$ mT, we have $R_{1,a}\approx 1$ and $R_{1,b}\approx r^{2}$ (see Appendix C for details). ## VI Outlook In this work, we have revisited the working principles of a Y-junction ferrite circulator Fay and Comstock (1965), a microwave engineering classic from the 1960’s, in an entirely new context of hybrid quantum systems and non-Hermitian Hamiltonian. The use of reconfigurable probes and single-crystalline YIG in a low-loss waveguide package allows us to connect the properties of the photon- magnon polaritons to the circulator performance. We have further leveraged our direct access to the internal modes and our ability to tune their coupling in situ to construct a multi-mode chiral system and unambiguously reveal its non- reciprocal eigenvector structure. An understanding of the circulator modes and the non-reciprocal eigen-vector structure of the multi-mode chiral system provides a foundation for future engineering of any target non-Hermitian Hamiltonian. This is achieved by our creation of a template model that one can use to couple any circuit QED element to in order to understand how it would integrate into the non-reciprocal dynamics, as was done here with the superconducting cavities. Looking forward, our device architecture provides a versatile testbed for studying non-reciprocal interactions in circuit QED by integration of superconducting qubits. This is enabled by two of its highlighted properties: the low internal loss of the circulator modes ($<$1% of the demonstrated coupling rates, compatible with potential high-fidelity operations), and the relatively low-field operation of the circulator ($\sim$25 mT, below ferrimagnetic saturation). The latter allows niobium waveguides or cavities to conveniently act as magnetic shields for superconducting qubits. We have preliminarily tested that the coherence times of a transmon qubit housed in one of the niobium cavities are unaffected by in-situ application of a global magnetic field up to at least 0.1 T. We expect a transmon housed in a niobium waveguide should receive a similar level of protection from magnetic field. Direct non-reciprocal coupling of superconducting qubits would open a new frontier in the study of non-reciprocal dynamics currently dominated by linear systems Sounas and Alù (2017); Fang _et al._ (2017); Xu _et al._ (2019); Ruesink _et al._ (2016). The physics of a $N$-mode linear non-reciprocal system can always be described efficiently by a $N\times N$ non-Hermitian Hamiltonian matrix (exemplified by our application of such a model) and its dynamics are always in the classical correspondence limit. Direct participation of multiple nonlinear modes (such as superconducting qubits) in non-reciprocal coupling, as envisioned in chiral quantum optics Lodahl _et al._ (2017), would lead to novel forms of entanglement stabilization and many- body phases Stannigel _et al._ (2012); Ramos _et al._ (2014). Our system presents another potential platform to implement this regime in circuit QED in additional to those proposed using dynamic control Guimond _et al._ (2020); Gheeraert _et al._ (2020). Strong coupling of Josephson circuits with low- loss non-reciprocal elements can even produce degenerate and protected ground states for robust encoding of qubits Rymarz _et al._ (2021). ###### Acknowledgements. We thank Juliang Li and Dario Rosenstock for experimental assistance. This research was supported by U.S. Army Research Office under grants W911-NF-17-1-0469 and W911-NF-19-1-0380. ## Appendix A Numerical simulation of the ferrite device Finite element analysis software that supports magnetodynamic simulations, such as Ansys HFSS, can be used to simulate our ciculator system with both driven mode and eigenmode solutions. Eigenmode analysis can solve for the frequency and field distributions of our device’s eigenmodes, while driven mode analysis reports the S-parameters over frequency. Here we discuss eigenmode simulations, but driven mode analysis can be carried out similarly. It is well known that when the applied field is large and magnetization is saturated along $z$ axis, one can generalize to the whole ferrite the equations of motion derived from the torque experienced by an electron dipole moment under the presence of an applied field. This approach, augmented by the small signal approximation of the Landau-Lifshitz equation of motion, yields the textbook Polder (relative) permeability tensor: $[\mu]_{z}=\begin{pmatrix}\mu_{r}&i\kappa&0\\\ -i\kappa&\mu_{r}&0\\\ 0&0&1\\\ \end{pmatrix}$ (10) where $\mu_{r}=1+\frac{\omega_{0}\omega_{m}}{\omega_{0}^{2}-\omega^{2}}$, $\kappa=\frac{\omega\omega_{m}}{\omega_{0}^{2}-\omega^{2}}$, with $\omega_{0}=\gamma\mu_{0}H_{0}$ and $\omega_{m}=\gamma\mu_{0}M_{s}$ being the internal field strength and saturation magnetization converted to frequencies, respectively. The Polder permeability tensor is implimented in HFSS by default to solve for the interaction of a saturated ferrite with an AC microwave field. However, in our experiment we operate the circulator at a low bias field, where the ferrite is not fully saturated. We adopt a permeability tensor model proposed by Sandy and Green Green and Sandy (1974) for a partially-magnetized ferrite: $\displaystyle[\mu]_{z}=\begin{pmatrix}\mu_{p}&i\kappa_{p}&0\\\ -i\kappa_{p}&\mu_{p}&0\\\ 0&0&\mu_{z}\\\ \end{pmatrix}$ (11) where $\displaystyle\mu_{p}$ $\displaystyle=\mu_{d}+(1-\mu_{d})\bigg{(}\frac{|M_{p}|}{M_{s}}\bigg{)}^{3/2}$ (12) $\displaystyle\kappa_{p}$ $\displaystyle=\kappa\bigg{(}\frac{M_{p}}{M_{s}}\bigg{)}$ (13) $\displaystyle\mu_{z}$ $\displaystyle=\mu_{d}^{\big{(}1-\frac{|M_{p}|}{M_{s}}\big{)}^{5/2}}$ (14) $\displaystyle\mu_{d}$ $\displaystyle=\frac{1}{3}+\frac{2}{3}\sqrt{1-\bigg{(}\frac{\omega_{m}}{\omega}\bigg{)}^{2}}$ (15) with $M_{p}$ being the net magnetization of the partially magnetized ferrite. This model contains functional forms for $\mu_{p}$ and $\mu_{z}$ that are purely empirical. However, the expressions for $\kappa_{p}$, which dictates the chiral splitting of the circulator modes, and $\mu_{d}$, which represents the permeability in fully demagnetized state, are well motivated Schlömann (1970). Figure 8: Eigenfrequencies of the device from finite-element simulations. For the circulator device with WCP, HFSS eigenmode simulation gives the mode frequency over different magnetic fields(dots) and agree semi-quantitatively with experimental data. This model is implemented in simulations by defining materials with the customized permeability tensor as given above. Whereas HFSS does not by default support eigenmode simulations for a ferrite under a DC bias field, manually defining the permeability tensor components allows us to simulate the circulator’s eigenmode structure at any magnetization (bias field) as shown in Fig. 8. The simulation results agree semi-quantitatively with the experimental data in Fig. 2(a) with a linear relationship between applied magnetic field and magnetization $M=\mu_{0}B/N_{z}$ mentioned earlier, including a dielectric resonance mode with steep magnetic field dependence that is visible in the experimental data. Relating to the anisotropy mentioned earlier in section III, the simulation is treating the YIG cylinder as completely isotropic, leading to degenerate Mode x and y at 0 field. To account for the anisotropy, we introduce a general energetic preference along the $x$ axis, thus making the domains of the unsaturated YIG preferentially align along the $x$ axis and breaking the rotational symmetry. When all domains are oriented along the $z$ axis with net magnetization of zero, the permeability is calculated to be $[\mu]_{z}=\begin{pmatrix}\mu_{\mathrm{eff}}&0&0\\\ 0&\mu_{\mathrm{eff}}&0\\\ 0&0&1\\\ \end{pmatrix}$ (16) where $\mu_{\mathrm{eff}}=\sqrt{\frac{\omega^{2}-\omega_{m}^{2}}{\omega^{2}}}$. To get the permeability matrix for domains align along the $x$ and $y$ axes ($[\mu]_{x}$, $[\mu]_{y}$), one can apply a change of coordinates to Eq. (A7). The matrix for completely random domain orientations would be an equal average of the three permeability matrices, $[\mu]_{x},[\mu]_{y},[\mu]_{z}$ Schlömann (1970). Applying a weighted average to the matrices will then allow for representation of an energetic preference, as shown for a preference along the $x$ axis: $[\mu]=(\frac{1}{3}+\delta)[\mu]_{x}+(\frac{1}{3}-\delta)[\mu]_{y}+\frac{1}{3}[\mu]_{z}$ (17) Using $\delta=0.1$ in Eq. (A8) gives 260 MHz of splitting between Mode x and Mode y, which is in good agreement with experimental results from Fig. 2(b). ## Appendix B Modeling YIG anisotropy in system Hamiltonian As mentioned earlier, there is a clear broken rotational symmetry in the $x$-$y$ plane apparent from the splitting in Fig. 2(b). Since the exact origin of the anisotropy is unknown, we will treat it as a general energetic favoring in the x-y plane. As the $\hat{z}$ bias field is increased, the magnetization will align more along $\hat{z}$, making $x$-$y$ plane preferences less impactful. Based off this understanding, we wanted a simple functional form to describe how the effect of this anisotropy decays with an increase in bias field strength that we could use to describe the decay of $\beta$. Since we just want the general form of how the effect of an energetic preference decays over field, the actual form of the energetic preference in the $x$-$y$ plane is not important. We chose to use a toy model of a magnetic domain with a simple Hamiltonian ($H_{\mathrm{an}}$) with a simple energetic preference given by $K$ along the $x$ axis and a total net magnetic moment $M$: $H_{\mathrm{an}}=-BM\cos(\theta)-K\sin^{2}(\theta)\cos^{2}(\phi)$ (18) To see how the effect of this anisotropy changes as we vary the magnetic field $B$, we utilized classical Boltzmann statistics. We define a partition function: $Z=\int_{0}^{\pi}\int_{0}^{2\pi}e^{-H_{\mathrm{an}}(\theta,\phi)/(k_{b}T)}\,\sin(\theta)d\theta d\phi\ $ (19) so we can calculate the expectation of the magnetic moment direction using Eq. (B3) for $A=M_{x},M_{y},M_{z}$; $M_{x}=M\sin(\theta)\cos(\phi),M_{y}=M\sin(\theta)\sin(\phi),M_{z}=M\cos(\theta)$. $\langle A^{2}\rangle=\int_{0}^{\pi}\int_{0}^{2\pi}A^{2}e^{-H_{\mathrm{an}}(\theta,\phi)/(k_{b}T)}\,\sin(\theta)d\theta d\phi\ /Z$ (20) We calculate these expectation values numerically, and find that the difference of $\langle M_{x}^{2}\rangle-\langle M_{y}^{2}\rangle$ follows approximately a $\sech(B)$ function. This motivates us to use this simple functional form to model the anisotropy-induced term $\beta$: $\displaystyle\beta(B)$ $\displaystyle=\beta_{0}\sech(B/B_{0})$ (21) The scaling factor $B_{0}$ was fit to the $S_{31}$ spectrum giving a value of 18.5 mT. While this is a rather crude phenomenological treatment of the anisotropy, since the detuning of the circulator modes becomes large enough that there is little hybridization with the cavities at relatively small magnetic fields ($\sim$20 mT), the exact dependence on magnetic field becomes less important to understand the non reciprocal dynamics of the cavities. ## Appendix C Two-mode Hamiltonian and gauge symmetry We aim to elucidate the non-reciprocity from the Hamiltonian given in Eq. (4) by reducing it to the form written in Eq. (9). In order to do this, we adiabatically integrate out the two circulator modes to reduce the Hamiltonian to a simple $2\times 2$ matrix ($H^{\prime}_{\mathrm{eff}})$ involving only the two cavity modes. The adiabatic elimination is justified due to the large loss rate on the hybridized circulator modes, making their relevant time scales much faster than the time scale set by the coupling parameters to the cavities. The form of the of the effective Hamiltonian is written out in Eq. (9). Due to the complicated dependence on the four mode model parameters, we have written simple frequency and loss terms on the diagonal entries and simple non-reciprocal couplings on the off diagonal entries where their explicit values change as a function of magnetic field. The coupling terms ($H_{12},H_{21}$) along with $r=\sqrt{\lvert H_{21}\rvert/\lvert H_{12}\rvert}$ are plotted in Fig. 7. The non-reciprocal nature of the system then becomes immediately apparent as the $H_{21}$ and $H_{12}$ Hamiltonian terms are different outside of 0 field, showing a clear directionality in the interaction. We can map this Hamiltonian to a reciprocal one using the similarity transformation outlined in Eq. (C1) with the transformation matrix written in Eq. (C2). This new reciprocal Hamiltonain is now symmetric under flipping the sign of the magnetic field $H_{\mathrm{rec}}(B)=H_{\mathrm{rec}}(-B)$. $H^{\prime}_{\mathrm{eff}}\rightarrow SH^{\prime}_{\mathrm{eff}}S^{-1}\equiv H_{\mathrm{rec}}$ (22) $S=\begin{pmatrix}r^{1/2}&0\\\ 0&r^{-1/2}\\\ \end{pmatrix}$ (23) This means that plotting the ratio of $R_{i,n,\mathrm{rec}}$ from $H_{\mathrm{rec}}$ will always yield 1 for all B values. One can also map the eigenvectors of the original system ($\ket{\psi_{i}}$) to the reciprocal system ($\ket{\psi_{i,\mathrm{rec}}}$) by $\ket{\psi_{Ri,\mathrm{rec}}}$ = $S\ket{\psi_{Ri}},\ket{\psi_{Li,\mathrm{rec}}}$ = $S^{-1}\ket{\psi_{Li}}$. Starting from the ratio $R_{i,n,\mathrm{rec}}=1$ using the eigenvectors of $H_{\mathrm{rec}}$, it is then apparent that transforming the eigenvectors back to those of $H^{\prime}_{\mathrm{eff}}$ will allow one to simply caluclate $R_{i,n}$. To illustrate this, we start with the explicit change in components from the transformation of the right and left eigenvectors as written in Eqns. (C3, C4), we can then substitute these in to the earlier expression for the ratio $R_{i,n}$ and see how the ratio deviates from the reciprocal case of 1, as done in Eq. (C5) with $i=1$ as an example. $\displaystyle\ket{\psi_{R,\mathrm{rec}}}=\begin{pmatrix}x\\\ y\\\ \end{pmatrix}\xrightarrow[\text{transform}]{\text{similarity}}\ket{\psi_{R}}=\frac{1}{\sqrt{\frac{|x|^{2}}{r}+|y|^{2}r}}\begin{pmatrix}\frac{1}{\sqrt{r}}x\\\ \sqrt{r}y\\\ \end{pmatrix}$ (24) $\ket{\psi_{L,\mathrm{rec}}}=\begin{pmatrix}x^{*}\\\ y^{*}\\\ \end{pmatrix}\xrightarrow[\text{transform}]{\text{similarity}}\ket{\psi_{L}}=\frac{1}{\sqrt{|x|^{2}r+\frac{|y|^{2}}{r}}}\begin{pmatrix}\sqrt{r}x^{*}\\\ \frac{1}{\sqrt{r}}y^{*}\\\ \end{pmatrix}$ (25) $R_{1,n}=\frac{|\bra{n_{L}}\ket{1}|}{|\bra{1}\ket{n_{R}}|}=\frac{|r^{1/2}x\sqrt{|x|^{2}r^{-1}+|y|^{2}r|}}{{|r^{-1/2}x\sqrt{|x|^{2}r+|y|^{2}r^{-1}}|}}$ (26) It is important to note two simplifying limits for Eq. (C5) that the reader may verify themselves, for $x/y\gg r$, $R_{1,n}\approx 1$ and for $y/x\gg r$, $R_{1,n}\approx r^{2}$. One can use this similarity transformation to understand the qualitative behavior of the disparate amplitude ratios seen in Fig. 6(f). As mentioned earlier, the amplitude ratio in this case can be roughly approximated as $R_{1,n}$ at $|B|\gtrsim$ 15 mT so we can focus primarily on this ratio to understand the behavior in this field range. At larger fields ($B\gtrsim 20$ mT) Modes a and b are dominated by participation in the bare cavity modes so we can approximate these modes by using the eigenmode values from $H^{\prime}_{\mathrm{eff}}$ for the cavity mode components and zeros for the circulator mode components. Under this approximation we can look at the inner products in $R_{1,n}$ just from the components in the eigenmodes of $H^{\prime}_{\mathrm{eff}}$. Mode a is largely dominated by the cavity 1 component with little circulator participation with $|\bra{a_{\mathrm{rec}}}\ket{1}|/|\bra{a_{\mathrm{rec}}}\ket{2}|\gg r$ for all values of $r$, thus we can invoke the limit of Eq.(C5) previously mentioned to find the ratio $R_{1,b}\approx 1$ which is what is seen in Fig. 6(f). The same argument can be made for mode b, but in the opposite limit of $|\bra{b_{\mathrm{rec}}}\ket{2}|/|\bra{b_{\mathrm{rec}}}\ket{1}|\gg r$, leading to the other limit of Eq.(C5), making the ratio $R_{1,b}\approx r^{2}$ which can be seen by comparing Fig. 6(f) with Fig. 7. ## References * Kord _et al._ (2020) A. Kord, D. L. Sounas, and A. Alù, Proceedings of the IEEE 108, 1728 (2020). * Devoret and Schoelkopf (2013) M. H. Devoret and R. J. Schoelkopf, Science 339, 1169 (2013). * Krantz _et al._ (2019) P. Krantz, M. Kjaergaard, F. Yan, T. P. Orlando, S. Gustavsson, and W. D. Oliver, Applied Physics Reviews 6, 021318 (2019). * Kurpiers _et al._ (2018) P. Kurpiers, P. Magnard, T. Walter, B. Royer, M. Pechal, J. Heinsoo, Y. Salathé, A. Akin, S. Storz, J.-C. Besse, S. Gasparinetti, A. Blais, and A. Wallraff, Nature 558, 264 (2018). * Axline _et al._ (2018) C. J. Axline, L. D. Burkhart, W. Pfaff, M. Zhang, K. Chou, P. Campagne-Ibarcq, P. Reinhold, L. Frunzio, S. M. Girvin, L. Jiang, M. H. Devoret, and R. J. Schoelkopf, Nature Physics 14, 705 (2018). * Vool and Devoret (2017) U. Vool and M. Devoret, International Journal of Circuit Theory and Applications 45, 897 (2017). * Lachance-Quirion _et al._ (2019) D. Lachance-Quirion, Y. Tabuchi, A. Gloppe, K. Usami, and Y. Nakamura, Applied Physics Express 12, 070101 (2019). * Awschalom _et al._ (2021) D. D. Awschalom, C. H. R. Du, R. He, F. J. Heremans, A. Hoffmann, J. T. Hou, H. Kurebayashi, Y. Li, L. Liu, V. Novosad, J. Sklenar, S. E. Sullivan, D. Sun, H. Tang, V. Tiberkevich, C. Trevillian, A. W. Tsen, L. R. Weiss, W. Zhang, X. Zhang, L. Zhao, and C. W. Zollitsch, arXiv:2102.03222 (2021). * Huebl _et al._ (2013) H. Huebl, C. W. Zollitsch, J. Lotze, F. Hocke, M. Greifenstein, A. Marx, R. Gross, and S. T. B. Goennenwein, Physical Review Letters 111, 127003 (2013). * Tabuchi _et al._ (2014) Y. Tabuchi, S. Ishino, T. Ishikawa, R. Yamazaki, K. Usami, and Y. Nakamura, Physical Review Letters 113, 083603 (2014). * Zhang _et al._ (2014) X. Zhang, C.-L. Zou, L. Jiang, and H. X. Tang, Physical Review Letters 113, 156401 (2014). * Tabuchi _et al._ (2015) Y. Tabuchi, S. Ishino, A. Noguchi, T. Ishikawa, R. Yamazaki, K. Usami, and Y. Nakamura, Science 349, 405 (2015). * Lachance-Quirion _et al._ (2020) D. Lachance-Quirion, S. P. Wolski, Y. Tabuchi, S. Kono, K. Usami, and Y. Nakamura, Science 367, 425 (2020). * Hou and Liu (2019) J. T. Hou and L. Liu, Physical Review Letters 123, 107702 (2019). * Golovchanskiy _et al._ (2021) I. A. Golovchanskiy, N. N. Abramov, V. S. Stolyarov, M. Weides, V. V. Ryazanov, A. A. Golubov, A. V. Ustinov, and M. Y. Kupriyanov, Science Advances 7, eabe8638 (2021). * Heyroth _et al._ (2019) F. Heyroth, C. Hauser, P. Trempler, P. Geyer, F. Syrowatka, R. Dreyer, S. G. Ebbinghaus, G. Woltersdorf, and G. Schmidt, Physical Review Applied 12, 054031 (2019). * Boventer _et al._ (2018) I. Boventer, M. Pfirrmann, J. Krause, Y. Schön, M. Klaui, and M. Weides, Physical Review B 97, 184420 (2018). * Zhang _et al._ (2017) D. Zhang, X.-Q. Luo, Y.-P. Wang, T.-F. Li, and J. Q. You, Nature Communications 8, 1368 (2017). * Anderson _et al._ (2016) B. M. Anderson, R. Ma, C. Owens, D. I. Schuster, and J. Simon, Physical Review X 6, 041043 (2016). * Owens _et al._ (2018) C. Owens, A. LaChapelle, B. Saxberg, B. M. Anderson, R. Ma, J. Simon, and D. I. Schuster, Physical Review A 97, 013818 (2018). * Zhang _et al._ (2020) X. Zhang, A. Galda, X. Han, D. Jin, and V. M. Vinokur, Physical Review Applied 13, 044039 (2020). * Fay and Comstock (1965) C. E. Fay and R. L. Comstock, IEEE Transactions on Microwave Theory and Techniques 13, 15 (1965). * Heiss (2012) W. D. Heiss, 45, 444016 (2012). * Özdemir _et al._ (2019) S. K. Özdemir, S. Rotter, F. Nori, and L. Yang, Nature Materials 18, 783 (2019). * Hatano and Nelson (1997) N. Hatano and D. R. Nelson, Physical Review B 56, 8651 (1997). * Yao and Wang (2018) S. Yao and Z. Wang, Physical Review Letters 121, 086803 (2018). * McDonald _et al._ (2018) A. McDonald, T. Pereg-Barnea, and A. A. Clerk, Physical Review X 8, 041031 (2018). * Stannigel _et al._ (2012) K. Stannigel, P. Rabl, and P. Zoller, New Journal of Physics 14, 063014 (2012). * Lodahl _et al._ (2017) P. Lodahl, S. Mahmoodian, S. Stobbe, A. Rauschenbeutel, P. Schneeweiss, J. Volz, H. Pichler, and P. Zoller, Nature 541, 473 (2017). * Chapman _et al._ (2017) B. J. Chapman, E. I. Rosenthal, J. Kerckhoff, B. A. Moores, L. R. Vale, J. A. B. Mates, G. C. Hilton, K. Lalumière, A. Blais, and K. W. Lehnert, Physical Review X 7, 041043 (2017). * Lecocq _et al._ (2017) F. Lecocq, L. Ranzani, G. A. Peterson, K. Cicak, R. W. Simmonds, J. D. Teufel, and J. Aumentado, Physical Review Applied 7, 024028 (2017). * Sliwa _et al._ (2015) K. M. Sliwa, M. Hatridge, A. Narla, S. Shankar, L. Frunzio, R. J. Schoelkopf, and M. H. Devoret, Physical Review X 5, 041020 (2015). * Ruesink _et al._ (2016) F. Ruesink, M.-A. Miri, A. Alù, and E. Verhagen, Nature Communications 7, 13662 (2016). * Fang _et al._ (2017) K. Fang, J. Luo, A. Metelmann, M. H. Matheny, F. Marquardt, A. A. Clerk, and O. Painter, Nature Physics 13, 465 (2017). * Wang _et al._ (2019) Y.-P. Wang, J. W. Rao, Y. Yang, P.-C. Xu, Y. S. Gui, B. M. Yao, J. Q. You, and C.-M. Hu, Physical Review Letters 123, 127202 (2019). * Xu _et al._ (2019) H. Xu, L. Jiang, A. A. Clerk, and J. G. E. Harris, Nature 568, 65 (2019). * Chen _et al._ (1991) D.-X. Chen, J. Brug, and R. Goldfarb, IEEE Transactions on Magnetics 27, 3601 (1991). * Fletcher _et al._ (1960) R. C. Fletcher, R. C. LeCraw, and E. G. Spencer, Physical Review 117, 955 (1960). * Klingler _et al._ (2017) S. Klingler, H. Maier-Flaig, C. Dubs, O. Surzhenko, R. Gross, H. Huebl, S. T. B. Goennenwein, and M. Weiler, Applied Physics Letters 110, 092409 (2017). * Marković _et al._ (2018) D. Marković, S. Jezouin, Q. Ficheux, S. Fedortchenko, S. Felicetti, T. Coudreau, P. Milman, Z. Leghtas, and B. Huard, Physical Review Letters 121, 040505 (2018). * Solt (1962) I. H. Solt, Journal of Applied Physics 33, 1189 (1962). * Schlömann (1970) E. Schlömann, Journal of Applied Physics 41, 204 (1970). * Green and Sandy (1974) J. Green and F. Sandy, IEEE Transactions on Microwave Theory and Techniques 22, 641 (1974). * Casimir (1945) H. B. G. Casimir, Reviews of Modern Physics 17, 343 (1945). * Ranzani _et al._ (2013) L. Ranzani, L. Spietz, Z. Popovic, and J. Aumentado, Review of Scientific Instruments 84, 034704 (2013). * Helszajn (2008) J. Helszajn, _The Stripline Circulator: Theory and Practice_ , 1st ed. (Wiley-IEEE Press, Hoboken, NJ, 2008). * Metelmann and Clerk (2015) A. Metelmann and A. A. Clerk, Physical Review X 5, 021025 (2015). * McDonald and Clerk (2020) A. McDonald and A. A. Clerk, Nature Communications 11, 5382 (2020). * Schomerus (2020) H. Schomerus, Physical Review Research 2, 013058 (2020). * Sounas and Alù (2017) D. L. Sounas and A. Alù, Nature Photonics 11, 774 (2017). * Ramos _et al._ (2014) T. Ramos, H. Pichler, A. J. Daley, and P. Zoller, Physical Review Letters 113, 237203 (2014). * Guimond _et al._ (2020) P.-O. Guimond, B. Vermersch, M. L. Juan, A. Sharafiev, G. Kirchmair, and P. Zoller, npj Quantum Information 6, 1 (2020). * Gheeraert _et al._ (2020) N. Gheeraert, S. Kono, and Y. Nakamura, Physical Review A 102, 053720 (2020). * Rymarz _et al._ (2021) M. Rymarz, S. Bosco, A. Ciani, and D. P. DiVincenzo, Physical Review X 11, 011032 (2021).
# Direct Nash Optimization: ​​Teaching Language Models to Self-Improve with General Preferences Corby Rosset Ching-An Cheng Arindam Mitra Michael Santacroce Ahmed Awadallah11footnotemark: 1 Tengyang Xie11footnotemark: 1 Microsoft Research Correspondence to <EMAIL_ADDRESS> ###### Abstract This paper studies post-training large language models (LLMs) using preference feedback from a powerful oracle to help a model iteratively improve over itself. The typical approach for post-training LLMs involves Reinforcement Learning from Human Feedback (RLHF), which traditionally separates reward learning and subsequent policy optimization. However, such a reward maximization approach is limited by the nature of “point-wise” rewards (such as that of the Bradley-Terry model), which fails to express complex intransitive or cyclic preference relations. While advances on RLHF show reward learning and policy optimization can be merged into a single contrastive objective for stability, they yet still remain tethered to the reward maximization framework. Recently, a new wave of research sidesteps the reward maximization presumptions in favor of directly optimizing over “pair- wise” or general preferences. In this paper, we introduce _Direct Nash Optimization_ (DNO), a _provable_ and _scalable_ algorithm that marries the _simplicity_ and _stability_ of contrastive learning with _theoretical generality_ from optimizing general preferences. Because DNO is a _batched on- policy_ algorithm using a regression-based objective, its implementation is straightforward and efficient. Moreover, DNO enjoys _monotonic improvement_ across iterations which helps it improve even over a strong teacher (such as GPT-4). In our experiments, a resulting 7B parameter Orca-2.5 model aligned by DNO achieves the state-of-the-art win-rate against GPT-4-Turbo of 33% on AlpacaEval 2.0 (even after controlling for response length), an absolute gain of 26% ($7\%\\!\to\\!33\%$) over the initializing model. It outperforms models with far more parameters, including Mistral Large, Self-Rewarding LM (70B parameters), and older versions of GPT-4. Our ablation studies analyze critical design decisions surrounding the choice of preference pairs, and the use of LLMs-as-preference-annotators. These results underscore the promise of DNO in the LLMs post-training, as well as offer actionable insights for the AI research community. ## 1 Introduction Figure 1: Direct Nash Optimization achieves state-of-the-art results for a 7B parameter large language model, being the first to surpass 30% in both raw win-rate and length-controlled (LC) win-rate against GPT-4-Turbo. Win Rate and LC Win Rate have $0.93$ to $0.98$ correlation with ChatBot Arena scores. The field of artificial intelligence is evolving towards advanced models that can understand, reason, follow complex instructions, and create nuanced content, while aligning with human values and preferences. Large Language Models (LLMs) (e.g., Brown et al., 2020; Ouyang et al., 2022; Touvron et al., 2023; OpenAI et al., 2023) have demonstrated remarkable capabilities in generating human-like text, answering questions, and coding, yet they still face challenges in tasks that require a high degree of reliability, safety, and ethical alignment. To address these challenges, fine-tuning LLMs using Reinforcement Learning from Human Feedback (RLHF) (Christiano et al., 2017; Bai et al., 2022a; Ouyang et al., 2022) has demonstrates strong potential for making LLMs more helpful by aligning them with human values. The RLHF framework has been long studied in the context of preference-based reinforcement learning (RL) or RL from human preferences (e.g., Knox and Stone, 2008; Akrour et al., 2012; Griffith et al., 2013; Wirth et al., 2017; Christiano et al., 2017). The conventional methods for RLHF typically assume that the preference is determined by a scalar reward function through some model, such as the frequently used Bradley-Terry (BT) model (Bradley and Terry, 1952).111We use “reward model” to denote a framework that translates preferences into rewards, e.g., Bradley-Terry, while “reward function” is a (possibly learned) function that outputs reward scalars. RLHF then optimizes toward the preference in a two-step procedure: reward learning, and policy optimization (through RL) to maximize the learned reward. Under certain conditions, the two-step procedure can be streamlined into a single-step contrastive learning approach (Rafailov et al., 2023), eliminating the need for explicit reward learning. Algorithms of this kind (e.g., Rafailov et al., 2023, DPO) leverage the insight that a policy can be expressed equivalently by an “internal reward function” that the policy is optimal to, so they reduce the RLHF problem to regressing the policy’s internal reward function to that of the preference model. These algorithms are originally offline, and boast enhanced stability and ease of optimization. Nonetheless, two-step RLHF algorithms and their single-step contrastive variants still fundamentally rely on the reward maximization framework, wherein reward-based preferences are governed by, e.g., the BT model. The reward maximization framing poses a major limitation. Reward functions, defined to output a scalar score $r(x,y)$ for a _single_ response $y$ to input $x$, cannot express general preferences $y\succ y^{\prime}\mid x$ between a _pair_ of outputs in all cases, e.g., intransitive or cyclic preferences (Elo, 1978). Hence, LLMs trained under reward maximization cannot always align with human preference. Furthermore, recent works show that even in settings where preferences can be perfectly expressed under the reward-based BT models, optimizing towards rewards yields problematic behaviors; we refer the reader to Bertrand et al. (2023); Azar et al. (2023); Munos et al. (2023) for more details. Lastly, reward functions in practice can quickly become “stale” as the distribution of the policy shifts under training (Ross et al., 2011; Cheng et al., 2023; Azar et al., 2023; Munos et al., 2023) – leaving them vulnerable to “reward hacking” (Amodei et al., 2016) In response to these weaknesses, an appealing line of work on RLHF proposes to directly optimize the _general preference function_ itself, instantiated as some oracle. These studies re-frame RLHF as finding a Nash equilibrium of a two-player game with “payoffs” from a regularized (Munos et al., 2023) or un- regularized (Swamy et al., 2024) general preference function. To solve it, they further approximate such Nash equilibrium using single-player algorithms by leveraging the symmetry of the preference function. Then, they instead define the reward of a response as the expected win-rate against the policy’s own behavior, as judged by the preference function, e.g., $r(x,y)={\mathbb{E}}_{y^{\prime}\sim\pi(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right]$. Hence, rewards are maximized by responses that are preferred over the policy’s expected response, and a Nash equilibrium is achieved when both players deploy a $\pi^{\star}$ that is preferred over any competing policy. However, these proposed single-player algorithms are primarily (purely) on-policy, and they sometimes require a separately estimated preference function or a time-varying reward function. How to scale these algorithms up faithfully is still under-investigated. We are motivated to overcome two separate challenges: the limited expressivity of reward-based RLHF, and the lack of clarity on how to scale up optimizing with respect to general preferences. Recent advances in reward-based optimization e.g., DPO, already have efficient and scalable implementations – we seek a similarly efficient solution under the framework of general preferences. We propose a provable and scalable RLHF algorithm – Direct Nash Optimization (DNO) (Algorithm 1) that achieves the best of both worlds, combining the scalability of contrastive objectives with the theoretical soundness of general preference optimization. DNO is designed as a _batched on-policy_ algorithm with a regression-based learning objective; this design choice makes DNO stable and scalable, striking a balance between deployment efficiency and adaptability. We summarize at a high level the key ingredients and insights of DNO below. 1. 1. To address the issue that reward functions cannot express general preferences, we leverage recent insights that the notion of reward of ought to be expressed as _expected_ win-rates with regard to a general preference function.222E.g., for a fixed $y\mid x$, the expected win-rate of $y$ against the policy itself is: ${\mathbb{E}}_{y^{\prime}\sim\pi(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right]$. 2. 2. To address the issue found in previous work that optimizing this more general objective with online algorithms is sample-inefficient or unstable, we decompose the learning procedure into a sequence of “batched on-policy” iterations, wherein each step instead optimizes a simple regression objective. 3. 3. The regression objective (we choose binary cross-entropy) aligns the “internal reward function” of the policy to the expected win-rate compared with itself (as defined in 3 of Algorithm 1). By sampling outputs from the current policy to use for training (i.e., “self-play”), this procedure incentivizes self- improving behavior. 4. 4. Our framework is general enough to admit off-policy samples into training, importantly, those from a more powerful teacher (See choice of $\mu_{1}$ and $\mu_{2}$ in Algorithm 1). 5. 5. Furthermore, to ensure stability and computational efficiency, we propose a filtering scheme such that the reward regression is only performed on preference pairs with a sufficiently large margin (for theoretical explanation, see Section 4; in practice, see Section 5.2). 6. 6. DNO repeats this procedure for multiple iterations to let the policy optimize toward the general preference. Since each step involves a regression problem it can be easily implemented at scale. Theoretically, we prove DNO converges to the intended Nash equilibrium on average, and that it can improve monotonically across iterations (see Section 3.1). Furthermore, our finite-sample analysis shows that approximation error at any iteration between the learned policy and the target is tightly bounded (Theorem 1). On the practical side, we provide a scalable implementation of DNO (Algorithm 2): an iterative self-improving algorithm with contrastive updates, which approximates Algorithm 1 under several critical design choices. Those choices include: sampling multiple online outputs from the policy being trained, using GPT-4 as the preference oracle, comparing on-policy samples to GPT-4’s own (teacher) outputs, and training only on pairs with “large margin” (for theoretical explanation, see Section 4; in practice, see Section 5.2). The primary distinction of our work over related works of Nash-MD (Munos et al., 2023) and SPO (Swamy et al., 2024) is that they both exhibit sample efficiency issues (two timescale updates or sample-inefficient RL steps), and both use purely on-policy samples. We resolve the efficiency issue with a sample-efficient objective that works in practice, and DNO is more flexible to incorporate off-policy samples from e.g., a powerful teacher. Most importantly, DNO works in practice – we provide comprehensive empirical evaluations, resulting in state-of-the-art performance: * • The resulting 7B parameter Orca-2.5 model, aligned using the practical implementation of DNO (Algorithm 2), achieves the state-of-the-art win-rate of any 7B model, exceeding $33\%$ against GPT-4-Turbo beyond on the AlpacaEval 2.0, even after controlling for length. This is an over $26\%$ absolute gain ($7\%\\!\to\\!33\%$) compared to the initialized model. It outperforms several recent advanced closed-source models, including Mistral Large and GPT-4-0613, as well as open-source models with far more ($10\times$) parameters, such as Self-Rewarding LM (Yuan et al., 2024) which has 70B parameters. * • Our thorough ablation studies in Section 5.2 examine critical design touchpoints surrounding choice of loss function (supervised finetuning or contrastive), training paradigm (with or without on-policy samples), preference annotator quality (large margin or not), and training pair construction (self-play, teacher-vs-student, etc). Our findings highlight that carefully-crafted methods encoded in Algorithm 2 lead to substantial gains. * • We show some examples of outputs across iterations which demonstrate qualitative improvements such as better addressing nuanced issues and presumptious questions (LABEL:ex:example-1), better organization and clarity while refraining from making misleading statements (LABEL:ex:example-2), and higher information density in answers (LABEL:ex:example-3). We hope that the results presented herein will provide clarity to the community regarding the use of AI feedback for post-training LLMs. ## 2 Preliminaries This section provides an overview of the RL from human feedback (RLHF) pipeline. We do not differentiate between RLHF and RLAIF (e.g., Bai et al., 2022b; Lee et al., 2023), as the distinction is outside our scope of discussion. Thus, we will uniformly refer to both concepts as RLHF. However, we want to make a clear delineation between two subtle differences: RLHF maximizing point-wise reward functions, and RLHF optimizing general preferences. It should be noted that this discussion is more broadly applicable in scope to general contextual bandits setup as well. Throughout this paper, we use $x\in\mathcal{X}$ to denote the input (i.e. the prompt) received by the LLM from a space $\mathcal{X}$. In this paper, we do not consider the distribution shift over the prompts, following the standard contextual bandits setup of RLHF (e.g., Ouyang et al., 2022; Rafailov et al., 2023), and we use $\rho$ to denote the distribution of the prompts. We use $y\in\mathcal{Y}$ to denote the response from the LLM given the prompt $x$ (this corresponds to action in the contextual bandits setup). We also use $\pi:\mathcal{X}\to\Delta(\mathcal{Y})$ to denote the policy, which is a LLM here, and $\Pi$ is the policy class. Our discussion throughout this paper will also regularly involve the following three learning paradigms, which are originally introduced and commonly used in the RL literature: 1. (1) _Offline_ : The learning algorithm operates without any active data collection, e.g., sampling from the current policy. The algorithm relies solely on an offline dataset for training. 2. (2) _Purely on-policy_ : technically, online on-policy. In this setup, learning takes place by sampling outputs from the latest policy and immediately updating it based on the newly collected data. No data reuse or additional offline data is considered. 3. (3) _Batched on-policy_ 333 We acknowledge abuse of terminology. Our algorithm is not entirely online, as it only contains batched data collection. It is also not strictly on-policy because it uses examples from other policies, like a teacher. While “offline” or “off-policy” may be technically more relevant, they might lead to misunderstanding among readers and detract from the emphasis we want to place on the collection samples from the current policy, which constitute the majority of our training data.: This setup is the middle of the offline and purely on-policy setups, striking a balance between deployment efficiency and adaptability. It involves iterative online data collection and can use other offline data. Its distinctive feature is that here the data collection in each iteration occurs in a batched fashion (e.g., akin to a dataset scale, much larger than the size of a typical mini-batch), and the amount of policy change can be more significant (e.g., running gradient steps over multiple epochs of a dataset, as opposed to tens of updates). ### 2.1 RLHF Based on Reward Models One typical approach to conducting RLHF is a two-step procedure through a reward function (Christiano et al., 2017). Suppose a preference dataset $\mathcal{D}_{\mathsf{pref}}\coloneqq\\{(x,y^{+},y^{-})\\}$ is given, where $(y^{+},y^{-})\sim\pi_{\mathsf{ref}}(\cdot\mid x)$, $\pi_{\mathsf{ref}}$ is some reference policy such as the policy obtained after supervised fine-tuning (SFT), and a preference $y^{+}\succ y^{-}\mid x$ is labeled by some human or AI annotator. In RLHF with reward functions, the preference is assumed to be generated based on some latent reward $r^{\star}$. The first step is to learn a reward function $r\in\mathcal{R}$ under some reward model assumption, where $\mathcal{R}$ is the reward class. A number of reward model assumptions have been studied, and the Bradley-Terry (BT) model (Bradley and Terry, 1952) is the most commonly used one. The BT model assumes the probability of $y^{+}\succ y^{-}\mid x$ satisfies $\displaystyle\mathcal{P}(y^{+}\succ y^{-}\mid x)\coloneqq\frac{\exp(r^{\star}(x,y^{+}))}{\exp(r^{\star}(x,y^{+}))+\exp(r^{\star}(x,y^{-}))}.$ This leads to the maximum-likelihood reward learning objective: $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:ppo_rm}}{e}q:ppo_{r}m}\widehat{r}\leftarrow\mathop{\mathrm{argmax}}_{r\in\mathcal{R}}{\mathbb{E}}_{(x,y^{+},y^{-})\sim\mathcal{D}}\left[\log\sigma(r(x,y^{+})-r(x,y^{-}))\right],$ (1) where $\sigma(\cdot)\coloneqq\frac{\exp(\cdot)}{1+\exp(\cdot)}$ is the sigmoid function. After that, the LLM is finetuned using the learned ${\widehat{r}}$ with RL, $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:ppo_pg}}{e}q:ppo_{p}g}{\widehat{\pi}}\leftarrow\mathop{\mathrm{argmax}}_{\pi\in\Pi}\mathbb{E}_{x\sim\mathcal{D}_{\mathsf{pref}},y\sim\pi(\cdot\mid x)}\left[{\widehat{r}}(x,y)-\beta\log\frac{\pi(x,y)}{\pi_{\mathsf{ref}}(x,y)}\right],$ (2) where the KL penalty term, $\mathbb{E}_{x\sim\mathcal{D}_{\mathsf{pref}},y\sim\pi(\cdot\mid x)}\left[\beta\log\frac{\pi(x,y)}{\pi_{\mathsf{ref}}(x,y)}\right]$, is used to mitigate overoptimization of the reward model (Ouyang et al., 2022), controlled by the coeffcient $\beta$. For the purposes of our discussion, we will call the objective above as “PPO objective”, and this two-step learning procedure as “PPO”. #### DPO Direct Preference Optimization (DPO) is proposed by Rafailov et al. (2023) as an alternative RLHF approach for combining the two-step procedure of PPO into a single objective. It utilizes the closed form solution ${\widehat{\pi}}$ in Eq. 2, so that solving ${\widehat{\pi}}$ directly from Eq. 1 becomes possible via $\displaystyle{\widehat{\pi}}\leftarrow\mathop{\mathrm{argmax}}_{\pi\in\Pi}{\mathbb{E}}_{(x,y^{+},y^{-})\sim\mathcal{D}_{\mathsf{pref}}}\log\left[\sigma\left(\beta\log\frac{\pi(y^{+}\mid x)}{\pi_{\mathsf{ref}}(y^{+}\mid x)}-\beta\log\frac{\pi(y^{-}\mid x)}{\pi_{\mathsf{ref}}(y^{-}\mid x)}\right)\right].$ ### 2.2 RLHF with General Preferences We now introduce the setup for directly optimizing a general preference function, as well as provide an overview of existing solutions to achieve this goal (mostly by leveraging the symmetry of the preferences), especially those proposed by Munos et al. (2023); Swamy et al. (2024). Here we assume that the learner is given query access to a general preference function $\mathcal{P}(y\succ y^{\prime}\mid x)\in[0,1]$, for any $(x,y,y^{\prime})\in\mathcal{X}\times\mathcal{Y}\times\mathcal{Y}$. This function indicates the probability that action $y$ is preferred over $y^{\prime}$ given the context $x$. In practice, this setup can be viewed as the theoretical mode of RLAIF (e.g., Bai et al., 2022b; Yuan et al., 2024), human-in-the-loop RLHF (e.g., Ouyang et al., 2022), or distillation fine- tuning (e.g., Tunstall et al., 2023). One common difficulty in optimizing a general preference function is its _intransitivity_ , e.g., it is possible that $\mathcal{P}(a\succ b)=\mathcal{P}(b\succ c)=\mathcal{P}(c\succ a)=1$, for some options $(a,b,c)$ (details see, e.g., Bertrand et al., 2023; Munos et al., 2023; Swamy et al., 2024). Therefore, the learning goal of optimizing general preferences can be the Nash equilibrium of the two-player zero-sum game with the payoffs as the general preference function $\mathcal{P}$. The formal definition of such Nash equilibrium is defined by the _Minimax Winner_ , $\mathsf{MW}$ (see, e.g., Kreweras, 1965; Simpson, 1969; Kramer, 1973; Fishburn, 1984), or the _von Neumann Winner_ (see, e.g., Dudík et al., 2015), $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:defMW}}{e}q:defMW}\mathsf{MW}(\mathcal{P})\coloneqq\mathop{\mathrm{argmax}}_{\pi\in\Pi}\mathop{\mathrm{argmin}}_{\pi^{\prime}\in\Pi}\mathcal{P}(\pi\succ\pi^{\prime})=\left(\mathop{\mathrm{argmax}}_{\pi\in\Pi}\min_{\pi^{\prime}\in\Pi}\mathcal{P}(\pi\succ\pi^{\prime}),\mathop{\mathrm{argmin}}_{\pi^{\prime}\in\Pi}\max_{\pi\in\Pi}\mathcal{P}(\pi\succ\pi^{\prime})\right),$ (3) where $\displaystyle\mathcal{P}(\pi\succ\pi^{\prime})\coloneqq{\mathbb{E}}_{x\sim\rho,y\sim\pi(\cdot\mid x),y^{\prime}\sim\pi^{\prime}(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right].$ #### SPO To approximate the Nash equilibrium as defined in Eq. 3, Swamy et al. (2024) proposed a single-player algorithm, SPO. This algorithm applies results from no-regret algorithms (e.g., Freund and Schapire, 1997). The SPO algorithm is executed essentially using the following two-step iterative process: for each $t=1,2,\dotsc,T$, (i) $\displaystyle~{}r_{t}(x,y)\leftarrow{\mathbb{E}}_{y^{\prime}\sim\pi_{t}(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right],~{}~{}\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$ (ii) $\displaystyle~{}\pi_{t+1}(\cdot\mid x)\leftarrow\frac{1}{Z_{t}(x)}\pi_{t}(\cdot\mid x)\exp\left(\frac{r_{t}(x,\cdot)}{\eta}\right),~{}~{}\forall x\in\mathcal{X},$ (4) where $\eta$ is the learning rate, $\pi_{1}$ is the uniform policy, i.e., $\pi_{1}(\cdot\mid x)\leftarrow{\sf unif}(\mathcal{Y}),~{}\forall x\in\mathcal{X}$, and $Z_{t}(x)\coloneqq\sum_{y\in\mathcal{Y}}\pi_{t}(y\mid x)\exp\left(\frac{r_{t}(x,y)}{\eta}\right)$ is the partition function for iteration $t$. Using the no-regret update of soft policy iteration, as shown in Eq. 4, Swamy et al. (2024) proved that the uniform mixture of $\pi_{1:T}$ from SPO is an approximation of the Nash equilibrium of $\mathsf{MW}(\mathcal{P})$, as defined in Eq. 3. #### Nash-MD Munos et al. (2023) proposed Nash-MD to approximate the Nash equilibrium of a KL-regularized preference function, $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:def_reg_pref}}{e}q:def_{r}eg_{p}ref}\mathcal{P}^{\pi,\pi^{\prime}}_{\tau}(y\succ y^{\prime}\mid x)\coloneqq$ $\displaystyle~{}\mathcal{P}(y\succ y^{\prime}\mid x)-\tau\log\frac{\pi(y\mid x)}{\pi_{\mathsf{ref}}(y\mid x)}+\tau\log\frac{\pi^{\prime}(y\mid x)}{\pi_{\mathsf{ref}}(y\mid x)},$ (5) $\displaystyle\mathcal{P}_{\tau}(\pi\succ\pi^{\prime})\coloneqq$ $\displaystyle~{}{\mathbb{E}}_{x\sim\rho,y\sim\pi(\cdot\mid x),y^{\prime}\sim\pi^{\prime}(\cdot\mid x)}\left[\mathcal{P}^{\pi,\pi^{\prime}}_{\tau}(y\succ y^{\prime}\mid x)\right]$ (6) $\displaystyle=$ $\displaystyle~{}\mathcal{P}(\pi\succ\pi^{\prime})-\tau{\mathbb{E}}_{x\sim\rho}\left[D_{\mathrm{KL}}(\pi(\cdot\mid x)~{}\|~{}\pi_{\mathsf{ref}}(\cdot\mid x))\right]+\tau{\mathbb{E}}_{x\sim\rho}\left[D_{\mathrm{KL}}(\pi^{\prime}(\cdot\mid x)~{}\|~{}\pi_{\mathsf{ref}}(\cdot\mid x))\right].$ Following this, Munos et al. (2023) demonstrate that the Nash Equilibrium of $\mathsf{MW}(\mathcal{P}_{\tau})$ can be approximated using a mirror descent (Nemirovskij and Yudin, 1983; Bubeck, 2015; Lattimore and Szepesvári, 2020) inspired algorithm, Nash-MD, which has a last-iteration guarantee. The Nash-MD algorithm can be viewed as a two-step iterative process: for each $t=1,2,\dotsc,T$, (i) $\displaystyle~{}r_{t}(x,y)\leftarrow{\mathbb{E}}_{y^{\prime}\sim\pi^{\tau}_{t}(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right],~{}~{}\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$ (ii) $\displaystyle~{}\pi_{t+1}(\cdot\mid x)\leftarrow\frac{1}{Z_{t}(x)}\pi_{t}^{\tau}(\cdot\mid x)\exp\left(\frac{r_{t}(x,\cdot)}{\eta}\right),~{}~{}\forall x\in\mathcal{X},$ where $\eta$ is the learning rate, $\pi_{t}^{\tau}$ is the geometric mixture between $\pi_{t}$ and $\pi_{\mathsf{ref}}$, $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:def_smooth_pit}}{e}q:def_{s}mooth_{p}it}\pi_{t}^{\tau}(y\mid x)\coloneqq\frac{\pi_{t}(y\mid x)^{1-\nicefrac{{\tau}}{{\eta}}}\pi_{\mathsf{ref}}(y\mid x)^{\nicefrac{{\tau}}{{\eta}}}}{\sum_{y^{\prime}\in\mathcal{Y}}\pi_{t}(y\mid x)^{1-\nicefrac{{\tau}}{{\eta}}}\pi_{\mathsf{ref}}(y\mid x)^{\nicefrac{{\tau}}{{\eta}}}},~{}\forall(x,y)\in\mathcal{X}\times\mathcal{Y},$ (7) and $Z_{t}(x)\coloneqq\sum_{y\in\mathcal{Y}}\pi_{t}^{\tau}(y\mid x)\exp\left(\frac{r_{t}(x,y)}{\eta}\right)$ is the partition function for iteration $t$. ## 3 Direct Nash Optimization While the no-regret update of soft policy iteration used in SPO and Nash-MD has inspired many standard (deep) reinforcement learning algorithms (e.g., Kakade, 2001, NPG,; Schulman et al., 2015, TRPO,; Schulman et al., 2017, PPO,; Haarnoja et al., 2018, SAC,), its faithful implementation still usually involves the two-timescale update. This could potentially lead to complex hyperparameter tuning and unstable performance. In this section, we propose a direct and iterative algorithm, Direct Nash Optimization (Algorithm 1), to approximate the Nash equilibrium of $\mathsf{MW}(\mathcal{P})$. This algorithm is primarily inspired by SPO. It can be readily adapted to Nash-MD for approximating the Nash equilibrium of $\mathsf{MW}(\mathcal{P}_{\tau})$ with the last-iteration guarantee, and we will discuss this in Appendix A. Algorithm 1 Direct Nash Optimization (DNO) input: General preference function $\mathcal{P}$, learning rate $\eta$, number of iterations $T$, prompt distribution $\rho$. 1:Initialize $\pi_{1}\leftarrow{\sf unif}(\mathcal{A})$. 2:for iteration $t=1,2,\dotsc,T$ do 3: Compute $r_{t}(x,y)\leftarrow{\mathbb{E}}_{y^{\prime}\sim\pi_{t}(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right]$, $\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$. 4: Obtain $\pi_{t+1}$ by, $\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:dnoloss}}{e}q:dnoloss}\begin{gathered}\pi_{t+1}\leftarrow\mathop{\mathrm{argmax}}_{\pi\in\Pi}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\mathcal{D}_{t}}\bigg{\\{}\sigma\left(r_{t}(x,y_{1})-r_{t}(x,y_{2})\right)\log\left[\sigma\left(\eta\log\frac{\pi(y_{1}\mid x)}{\pi_{t}(y_{1}\mid x)}-\eta\log\frac{\pi(y_{2}\mid x)}{\pi_{t}(y_{2}\mid x)}\right)\right]\\\ \hskip 120.0pt+\sigma\left(r_{t}(x,y_{2})-r_{t}(x,y_{1})\right)\log\left[\sigma\left(\eta\log\frac{\pi(y_{2}\mid x)}{\pi_{t}(y_{2}\mid x)}-\eta\log\frac{\pi(y_{1}\mid x)}{\pi_{t}(y_{1}\mid x)}\right)\right]\bigg{\\}},\end{gathered}$ (8) where $\mathcal{D}_{t}$ is generated by $x\sim\rho,y_{1}\sim\mu_{1,t}(\cdot\mid x),y_{2}\sim\mu_{2,t}(\cdot\mid x)$; $\mu_{1,t}$ and $\mu_{2,t}$ can be either off-policy (e.g., pre-defined) or on-policy (based on $\pi_{t}$). 5:end for 6:return $\bar{\pi}={\sf unif}(\pi_{1:T})$. ### 3.1 Derivation of Algorithm 1 In most practical algorithms which are inspired by soft policy iteration, including the original practical version of SPO, they typically adopt the following approach: “pushing” $\pi$ towards this subsequent learning goal in each iteration (we will refer to this as the soft policy iteration target throughout the paper): $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:def_pistar}}{e}q:def_{p}istar}\pi^{\star}_{t+1}(\cdot\mid x)\coloneqq\frac{1}{Z_{t}(x)}\pi_{t}(\cdot\mid x)\exp\left(\frac{r_{t}(x,\cdot)}{\eta}\right),$ (9) where $Z_{t}(x)=\sum_{y\in\mathcal{Y}}\pi_{t}(y\mid x)\exp\left(\frac{r_{t}(x,y)}{\eta}\right)$ is the partition function. It can be realized by minimizing a distance metric between $\pi_{t+1}$ and $\pi$. For example, the PPO algorithm for RLHF (e.g., Christiano et al., 2017; Ouyang et al., 2022) essentially minimizes the reverse KL divergence as follows, $\displaystyle(\pi_{t+1}^{\text{PPO}}\leftarrow)$ $\displaystyle~{}\mathop{\mathrm{argmin}}_{\pi\in\Pi}{\mathbb{E}}_{x\sim\rho}\left[D_{\mathrm{KL}}(\pi(\cdot\mid x)~{}\|~{}\pi^{\star}_{t+1}(\cdot\mid x))\right]$ $\displaystyle=$ $\displaystyle~{}\mathop{\mathrm{argmax}}_{\pi\in\Pi}\mathbb{E}_{x\sim\rho,y\sim\pi(\cdot\mid x)}\left[\eta\log\frac{\pi^{\star}_{t+1}(x,y)}{\pi_{t}(x,y)}-\eta\log\frac{\pi(x,y)}{\pi_{t}(x,y)}\right]$ $\displaystyle=$ $\displaystyle~{}\mathop{\mathrm{argmax}}_{\pi\in\Pi}\mathbb{E}_{x\sim\rho,y\sim\pi(\cdot\mid x)}\left[r_{t}(x,y)-\eta Z_{t}(x)-\eta\log\frac{\pi(x,y)}{\pi_{t}(x,y)}\right]$ $\displaystyle=$ $\displaystyle~{}\mathop{\mathrm{argmax}}_{\pi\in\Pi}\mathbb{E}_{x\sim\rho,y\sim\pi(\cdot\mid x)}\left[r_{t}(x,y)-\eta\log\frac{\pi(x,y)}{\pi_{t}(x,y)}\right].$ ($\Leftrightarrow$ PPO objective, as $Z_{t}$ is independent of $\pi$) However, implementing the above approach typically necessitates _on-policy_ sampling from the current policy $\pi$. Ignoring the $Z_{t}(x)$ term could also lead to high variance in the empirical gradient estimation. This is a persistent issue in actor-critic style algorithms that usually suggests the need for an additional baseline (details see, e.g., Mnih et al., 2016), which also requires on-policy estimation. When $r_{t}$ also varies over iterations, as in SPO or Nash-MD, we then need to update all of the policy, baseline, and reward online simultaneously. These challenges have hindered the scalability of existing algorithms which are based on learning the Nash equilibrium of general preference functions. #### Regressing “internal rewards” towards preference-based rewards Different from the mentioned approaches above which are mostly focusing on the concept of “pushing” $\pi\to\pi^{\star}_{t+1}$. We now consider the following mechanism: regressing $r_{\pi,t}\to r_{t}$, where $r_{\pi,t}$ is the internal reward function of a given $\pi$ at iteration $t$: $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:def_rpit}}{e}q:def_{r}pit}r_{\pi,t}(x,y)\coloneqq\eta\log\frac{\pi(y\mid x)}{\pi_{t}(y\mid x)}+\eta Z_{t}(x).$ (10) This can be interpreted as a reparameterization trick, where $\pi$ is exactly the soft policy iteration target (refer to Eq. 9) induced by $\pi_{t}$ and the defined $r_{\pi,t}$. Therefore, regressing that specifically parameterized $r_{\pi,t}$ to $r_{t}$ allows us to directly optimize the soft policy iteration target with respect to $r_{\pi,t}$ and $\pi_{t}$. This idea is inspired by techniques from inverse RL (e.g., Finn et al., 2016b, a, Guided Cost Learning) as well as recent advances in RLHF (Rafailov et al., 2023, DPO). To avoid the issues arising from the partition function $Z_{t}(x)$, we consider learning from the $(x,y_{1},y_{2})$ tuple, where $y_{1}$ and $y_{2}$ are both responses to textual input $x$. Note that, due to the offline learning nature of the regressive objective, the sampling distribution of $y_{1}$ and $y_{2}$ does not impact the learning objective (i.e., $r_{\pi,t}\to r_{t}$, but it may affect the sample complexity from the coverage reason as we will discuss later), whereas pushing $\pi\to\pi_{t+1}^{\star}$ requires sampling $y$ on-policy, as previously discussed. Therefore, given an arbitrary $(x,y_{1},y_{2})$ tuple, we regress the “prediction” $\hat{z}$ to the “goal” $z$ (both defined below), using binary logarithmic/cross-entropy loss to measure the prediction error (see, e.g., Foster and Krishnamurthy, 2021), $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:def_z}}{e}q:def_{z}}{\widehat{z}}\coloneqq\sigma\left(r_{\pi,t}(x,y_{1})-r_{\pi,t}(x,y_{2})\right)=\sigma\bigg{(}\underbrace{\eta\log\frac{\pi(y_{1}\mid x)}{\pi_{t}(y_{1}\mid x)}-\eta\log\frac{\pi(y_{2}\mid x)}{\pi_{t}(y_{2}\mid x)}}_{\eqqcolon\Delta_{\pi,t}(x,y_{1},y_{2})}\bigg{)},\quad z\coloneqq\sigma\big{(}\underbrace{r_{t}(x,y_{1})-r_{t}(x,y_{2})}_{\eqqcolon\Delta_{t}^{\star}(x,y_{1},y_{2})}\big{)};$ (11) $\displaystyle\ell_{\pi,t}(x,y_{1},y_{2})\coloneqq z\log(1/{\widehat{z}})+(1-z)\log(1/(1-{\widehat{z}}))$ $\displaystyle=-\sigma\big{(}\Delta_{t}^{\star}(x,y_{1},y_{2})\big{)}\log\left[\sigma\big{(}\Delta_{\pi,t}(x,y_{1},y_{2})\big{)}\right]-\sigma\big{(}\Delta_{t}^{\star}(x,y_{2},y_{1})\big{)}\log\left[\sigma\big{(}\Delta_{\pi,t}(x,y_{2},y_{1})\big{)}\right].$ Therefore, we obtain the following objective to learn $\pi_{t+1}$, $\displaystyle~{}\mathop{\mathrm{argmin}}_{\pi\in\Pi}\mathcal{L}_{\mathcal{D}_{t}}(\pi;\pi_{t})$ $\displaystyle\coloneqq$ $\displaystyle~{}\mathop{\mathrm{argmin}}_{\pi\in\Pi}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\mathcal{D}_{t}}\left[\ell_{\pi,t}(x,y_{1},y_{2})\right]$ $\displaystyle=$ $\displaystyle~{}\mathop{\mathrm{argmax}}_{\pi\in\Pi}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\mathcal{D}_{t}}\Big{[}\sigma\big{(}\Delta_{t}^{\star}(x,y_{1},y_{2})\big{)}\log\left[\sigma\big{(}\Delta_{\pi,t}(x,y_{1},y_{2})\big{)}\right]+\sigma\big{(}\Delta_{t}^{\star}(x,y_{2},y_{1})\big{)}\log\left[\sigma\big{(}\Delta_{\pi,t}(x,y_{2},y_{1})\big{)}\right]\Big{]}$ $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:reward_regression}}{e}q:reward_{r}egression}=$ $\displaystyle~{}\mathop{\mathrm{argmax}}_{\pi\in\Pi}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\mathcal{D}_{t}}\Bigg{[}\sigma\left(r_{t}(x,y_{1})-r_{t}(x,y_{2})\right)\log\left[\sigma\left(\eta\log\frac{\pi(y_{1}\mid x)}{\pi_{t}(y_{1}\mid x)}-\eta\log\frac{\pi(y_{2}\mid x)}{\pi_{t}(y_{2}\mid x)}\right)\right]$ (12) $\displaystyle~{}\hskip 100.0pt+\sigma\left(r_{t}(x,y_{2})-r_{t}(x,y_{1})\right)\log\left[\sigma\left(\eta\log\frac{\pi(y_{2}\mid x)}{\pi_{t}(y_{2}\mid x)}-\eta\log\frac{\pi(y_{1}\mid x)}{\pi_{t}(y_{1}\mid x)}\right)\right]\Bigg{]}.$ Here, $\mathcal{D}_{t}$ is generated by $x\sim\rho,y_{1}\sim\mu_{1,t}(\cdot\mid x),y_{2}\sim\mu_{2,t}(\cdot\mid x)$ with some policies $\mu_{1,t}$ and $\mu_{2,t}$. It should be noted that $\mu_{1,t}$ and $\mu_{2,t}$ for each $t\in[T]$ are parts of our algorithm’s design decisions. We will provide choices for them in Section 3.2 to promote sample efficiency, which are informed by our finite-sample analysis. #### Monotonic improvement from the _batched on-policy_ updates One key distinction between DNO and existing algorithms for learning Nash equilibrium (such as SPO and Nash-MD) is that those algorithms aim to approach the Nash equilibrium in a purely on-policy manner, which can be potentially unstable and may need to incorporate two-timescale updates (that change the reward function used in the inner problem more frequently). On the other hand, DNO is a batched on-policy algorithm with single-timescale updates. From a purely theoretical perspective, it seems that DNO may require many iterations to ensure the convergence of $\bar{\pi}$ to the Nash equilibrium, which could potentially be costly. Additionally, DNO only converges on- average, and it is unrealistic to deploy in practice that uniform mixture policy $\bar{\pi}$ (note that, as inspired by Munos et al. (2023), DNO could be extended to regularized preferences with last-iteration convergence, which is discussed in Appendix A). However, from a practical perspective, we can leverage the following two desirable properties from LLMs scenario to eliminate these concerns and ensure _monotonic improvement_ over the DNO iterations: Firstly, the soft policy iteration target Eq. 9 is actually the analytical solution for maximizing the following loss, $\ell_{t}(\pi)\coloneqq\mathcal{P}(\pi\succ\pi_{t})-\eta{\mathbb{E}}_{x\sim\rho}\left[D_{\mathrm{KL}}(\pi(\cdot\mid x)~{}\|~{}\pi_{t}(\cdot\mid x))\right]$, and $\pi_{t+1}^{\star}=\mathop{\mathrm{argmax}}_{\pi}\ell_{t}(\pi)$. We can notice that $\ell_{t}(\pi_{t})=0.5$ and ${\mathbb{E}}_{x\sim\rho}\left[D_{\mathrm{KL}}(\pi(\cdot\mid x)~{}\|~{}\pi_{t}(\cdot\mid x))\right]\geq 0$. This means $0.5\leq\ell_{t}(\pi_{t+1}^{\star})=\mathcal{P}(\pi_{t+1}^{\star}\succ\pi_{t})-\eta{\mathbb{E}}_{x\sim\rho}\left[D_{\mathrm{KL}}(\pi_{t+1}^{\star}(\cdot\mid x)~{}\|~{}\pi_{t}(\cdot\mid x))\right]\Longrightarrow\mathcal{P}(\pi_{t+1}^{\star}\succ\pi_{t})\geq 0.5+\eta{\mathbb{E}}_{x\sim\rho}\left[D_{\mathrm{KL}}(\pi_{t+1}^{\star}(\cdot\mid x)~{}\|~{}\pi_{t}(\cdot\mid x))\right]$. This means $\pi_{t+1}^{\star}$ is guaranteed to be more preferred than $\pi_{t}$ with respect to the preference $\mathcal{P}$, and there is even a computable lower bound of the amount of improvement—$\eta{\mathbb{E}}_{x\sim\rho}\left[D_{\mathrm{KL}}(\pi_{t+1}^{\star}(\cdot\mid x)~{}\|~{}\pi_{t}(\cdot\mid x))\right]$. Therefore, if $\pi_{t+1}$ learned from 4 of Algorithm 1 is a accurate enough approximation of $\pi_{t+1}^{\star}$ (which is proved in Section 3.2), we could expect that the policy is monotonically improved over DNO iterations. Note that the monotonic improvement guarantee is _exclusive_ to our design choice of _batched on- policy_ updates in DNO, because the alternatives are either unclear or unstable: it is undefined how to perform iterative updates offline, and one gradient update from a purely online algorithm may not be able able to accurately approximate the the soft policy iteration target $\pi_{t+1}^{\star}$. Secondly, in practice, we usually have validation data available, which allows us to deploy the best policy over $\pi_{1:(T+1)}$. ### 3.2 Theoretical Analysis One of our major proposals is to use a regression-based objective to approximate the explicit soft policy iteration; in this section we show the approximation error from this regression is tightly bounded with finite-sample analysis. The following proposition discusses how well the solution of the regression-based objective (defined in Eq. 12 or 4 of Algorithm 1) can approximate the soft policy iteration (Eq. 9) in terms of the total variation metric at each iteration. ###### Theorem 1 (informal). Fix an arbitrary iteration $t\in[T]$. Suppose $\pi_{t+1}$ is from 4 of Algorithm 1, and $\pi_{t+1}^{\star}$ is defined in Eq. 9. Then, under mild assumptions (realizability and boundedness, formally introduced in Appendix B), we have $\displaystyle{\mathbb{E}}_{x\sim\rho}\left[\left(D_{\mathrm{TV}}(\pi_{t+1}(\cdot\mid x),\pi_{t+1}^{\star}(\cdot\mid x))\right)^{2}\right]\leq\mathcal{O}\left(\frac{\mathfrak{C}_{t}R_{\max}^{2}\log(\nicefrac{{|\Pi|}}{{\delta}})}{N}\right),$ where the concentrability coefficient $\mathfrak{C}_{t}$ is defined as below, $\displaystyle\mathfrak{C}_{t}\coloneqq\frac{{\mathbb{E}}_{x\sim\rho,y_{1}\sim\pi_{t+1}^{\star}(\cdot\mid x),y_{2}\sim\pi_{t+1}(\cdot\mid x)}\left[\left(\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}-\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)^{2}\right]}{{\mathbb{E}}_{x\sim\rho,y_{1}\sim\mu_{1,t}(\cdot\mid x),y_{2}\sim\mu_{2,t}(\cdot\mid x)}\left[\left(\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}-\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)^{2}\right]}.$ If $\pi_{t}=\pi_{t}^{\star}$ for all $t\in[T]$, the reader can refer to (Swamy et al., 2024, Section 3) for the convergence of $\bar{\pi}$ (returned by Algorithm 1) to the Nash equilibrium. We expect the total variation difference between $\pi_{t}$ and $\pi_{t}^{\star}$ provided by Theorem 1 will be additive errors on top of the guarantees from Swamy et al. (2024). Note that, we present the concentrability coefficient $\mathfrak{C}_{t}$ as data-dependent, with $\pi_{t+1}$ (learned from data) as part of its definition. We aim to make this guiding the design choices of $\mu_{1,t}$ and $\mu_{2,t}$ from such $\mathfrak{C}_{t}$ for the purpose of sample efficiency. The formal statement and detailed proof of Theorem 1, without involving $\pi_{t+1}$, are deferred to Appendix B. Although it shares a similar expression to the concentrability coefficient in offline reinforcement learning (e.g., Chen and Jiang, 2019; Xie et al., 2021), the policies $\mu_{1,t}$ and $\mu_{2,t}$ are flexible here due to the generative nature of large language models. This flexibility allows for additional intervention, enhancing sample efficiency. We can notice that the value of $\mathfrak{C}_{t}$ can be always bounded by $\mathfrak{C}_{t}\leq\max_{(x,y)\in\mathcal{X}\times\mathcal{Y}}\frac{\pi_{t+1}^{\star}(y\mid x)\pi_{t+1}(y\mid x)}{\mu_{1,t}(y\mid x)\mu_{2,t}(y\mid x)}$ in the worst case. However, as $\pi_{t+1}$ is likely to be restricted within a certain region, for instance, because fine-tuning will not significantly alter the behavior of the language model, we anticipate that such a coefficient will not depend on the per-$(x,y)$ worst case. On the other hand, as a direct observation, we notice that the ideal selection of $\mu_{1,t}$ and $\mu_{2,t}$ should be close to the target of soft policy iteration $\pi_{t+1}^{\star}$ (assuming $\pi_{t+1}^{\star}$ and $\pi_{t+1}$ are close). Interestingly, this theoretical observation coincides with recent empirical results. Here, Liu et al. (2024b) suggests that using statistical rejection sampling to sample from the soft policy iteration target (which is almost equivalent to sampling $y_{1}$ and $y_{2}$ from $\pi_{t+1}^{\star}$) could benefit preference tuning. However, in our case, if we use similar statistical rejection sampling techniques on $\pi_{t}$ to sample $\pi_{t+1}^{\star}$ (and $\pi_{t+1}$), the cost of rejection sampling is likely to be comparable to the concentrability coefficient $\mathfrak{C}_{t}$ when choosing $\mu_{1,t}$ and $\mu_{2,t}$ to be $\pi_{t}$ (see, e.g., Owen, 2013). This suggests that both $\pi_{t}$ and $\pi_{t+1}^{\star}$ (via rejection sampling) as the choices of $\mu_{1,t}$ and $\mu_{2,t}$ will be comparable options in terms of sample efficiency. On the other hand, as we will demonstrate in the next section, since $r_{t}$ is defined based on $\pi_{t}$ (as shown in 3 of Algorithm 1), choosing $\mu_{1,t}$ and $\mu_{2,t}$ to be $\pi_{t}$ can easily adapt to such a reward of $r_{t}$. Another interesting observation is that despite Eq. 12 sharing a similar form with Bradley-Terry style reward modeling with using MLE, the target distributions used to measure distribution shift appear to be quite different. This disparity is due to the different objectives: fitting soft policy iteration versus reward estimation. For the Bradley-Terry style reward modeling using MLE, the desired distribution of $y_{1}$ and $y_{2}$ should be two distinct distributions (see, e.g., Zhan et al., 2024; Xiong et al., 2023). However, in our case where the learning goal is to fit the soft policy iteration, we may prefer $y_{1}$ and $y_{2}$ from two (near) on-policy distributions as discussed above, as long as we expect the learned $\pi_{t+1}$ will be accurate enough. To the best of our knowledge, this is the first theoretical result that illustrates the importance of on-policy sampling beyond policy optimization style algorithms for RLHF. ## 4 Practical Algorithm – Iterative Contrastive Self-Improvement In this section, we shift our focus to the algorithmic design of the practically scalable version of DNO, following the principles discussed in the last section. A primary challenge encountered in the implementation of the conceptual algorithm DNO (Algorithm 1) stems from the necessity to compute the expectation with respect to the preference function $\mathcal{P}$ under the current policy $\pi_{t}$. Perhaps surprisingly, as we will show, all we need is a properly implemented iterative DPO-like contrastive learning algorithm. Algorithm 2 DNO-Prct: Practical Implementation of DNO via Iterative Contrastive Self-Improvement input: General preference function $\mathcal{P}$, learning rate ${\widetilde{\eta}}$, iterations $T$, reference policy $\pi_{\mathsf{ref}}$, prompt distribution $\rho$. 1:Initialize $\pi_{1}\leftarrow\pi_{\mathsf{ref}}$. 2:for iteration $t=1,2,\dotsc,T$ do 3: Construct $\mathcal{D}_{t}=\\{(x,y^{\mathsf{gold}})\\}$ where $x\sim\rho$ and $y\sim\pi_{\mathsf{gold}}(\cdot\mid x)$. 4: Sample _batched on-policy_ responses: Sample $K$ outputs per per prompt using the current $\pi_{t}$: $\\{y_{t}^{1},y_{t}^{2},\dotsc,y_{t}^{K}\\}\sim\pi_{t}(\cdot\mid x)$, $\forall x\in\mathcal{D}_{t}$. 5: Rank responses: For each $x\in\mathcal{D}_{t}$, rank the corresponding $\\{y_{t}^{1},y_{t}^{2},\dotsc,y_{t}^{K},y^{\mathsf{gold}}\\}$ using the pair- wise win-rate by sampling from the general preference function $\mathcal{P}$. 6: Filter preference pairs: Construct $\mathcal{D}_{t+1}=\\{(x,y_{t}^{+},y_{t}^{-})\\}$, for all $x\in\mathcal{D}_{t+1}$, and $(y_{t}^{+},y_{t}^{-})$ are large-margin pairs (based on the win-rate rank) within the responses for $x$ from the previous step. 7: Contrastive learning: Obtain $\pi_{t+1}$ by, $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:nashdpo}}{e}q:nashdpo}\pi_{t+1}\leftarrow\mathop{\mathrm{argmax}}_{\pi\in\Pi}{\mathbb{E}}_{(x,y_{t}^{+},y_{t}^{-})\sim\mathcal{D}_{t+1}}\log\left[\sigma\left({\widetilde{\eta}}\log\frac{\pi(y_{t}^{+}\mid x)}{\pi_{t}(y_{t}^{+}\mid x)}-{\widetilde{\eta}}\log\frac{\pi(y_{t}^{-}\mid x)}{\pi_{t}(y_{t}^{-}\mid x)}\right)\right].$ (13) 8:end for 9:return best of $\pi_{1:(T+1)}$ on the validation data. We present our the practical implementation of DNO in Algorithm 2 (DNO-Prct), which is a batched on-policy algorithm that conducts self-improvement iteratively via contrastive learning. One key consideration in our algorithmic design is that we only need to implicitly use the reward function $r_{t}$. This comes from the specifically designed on-policy sampling, data filtering, and pair construction. While these specific design choices make DNO-Prct seem similar to simply performing DPO iteratively, there are significant reasons for these design decisions, as we will discuss below. #### Batched on-policy sampling The use of batched on-policy sampling in 4 of Algorithm 2 is crucial to avoid explicit use of $r_{t}$ (defined as ${\mathbb{E}}_{y^{\prime}\sim\pi_{t}(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right]$ in 3 of Algorithm 1). This means we essentially choose $\mu_{1}$ and $\mu_{2}$ in DNO to be $\pi_{t}$ in DNO-Prct, but we are free to let them vary slightly as a mixture of other policies, e.g., from a stronger teacher. Specifically, it is unrealistic to assume in practice that we can access the exact value of $\mathcal{P}(y\succ y^{\prime}\mid x)$ given an $(x,y,y^{\prime})$ tuple. Based on the definition of $r_{t}$ and the fact of $\\{y_{t}^{1},y_{t}^{2},\dotsc,y_{t}^{K}\\}$ are sampled from $\pi_{t}$, DNO-Prct essentially uses the following sampled based approach to estimate $r_{t}$: $r_{t}(x,y)\approx\frac{1}{K}\sum_{y^{\prime}\in\\{y_{t}^{1},y_{t}^{2},\dotsc,y_{t}^{K},y^{\mathsf{gold}}\\}\setminus y}\mathds{1}_{\mathcal{P}}(\text{Is $y$ better than $y^{\prime}$ on $x$}?)$, for any $x$ and $y\in\\{y_{t}^{1},y_{t}^{2},\dotsc,y_{t}^{K},y^{\mathsf{gold}}\\}$, where $\mathds{1}_{\mathcal{P}}$ denotes one sample from $\mathcal{P}$ and output $\\{0,1\\}$. This is implemented in 5 of Algorithm 2, and its precise implementation on this is discussed in the Section 5. On the other hand, as we discussed in the last section, the batched on-policy sampling from $\pi_{t}$ is an appropriate option due to the consideration of sample efficiency when we use Eq. 13 to approximate the soft policy iteration (see Theorem 1 and its discussion). #### Preference pair construction Another key design choice in Algorithm 2 is that Eq. 13 of Algorithm 2 only uses the purely contrastive loss, whereas Eq. 8 of Algorithm 1 also contains the regression target $\sigma\left(r_{t}(x,y)-r_{t}(x,y^{\prime})\right)$ (for a given $(x,y,y^{\prime})$ tuple), which is not necessarily $\\{0,1\\}$. As we discussed above, it is unrealistic to expect access to the exact value of $\mathcal{P}(y\succ y^{\prime}\mid x)$, so it is also unlikely to get an accurate value of the regression target $\sigma(r_{t}(x,y)-r_{t}(x,y^{\prime}))$. Thus, we add an additional data filtering step to address this issue as in 6 of Algorithm 2. Ideally, we want the selected $(x,y^{+},y^{-})$ tuple to satisfy $\sigma(r_{t}(x,y_{t}^{+})-r_{t}(x,y_{t}^{-}))\approx 1$, so that Eq. 8 can be approximated by Eq. 13. However, one can notice that it requires $r_{t}(x,y_{t}^{+})-r_{t}(x,y_{t}^{-})\to\infty$, but we know $r_{t}(x,y)\in[0,1]$, $\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$. From the derivation of DNO in Section 3, it is clear that scaling up $r_{t}$ and $\eta$ with the same absolute constant $c$ does not affect the soft policy iteration target of Eq. 9, but it will slightly change the DNO objective (Eq. 8 in Algorithm 1) by $r_{t}\to c\cdot r_{t}$ and $\eta\to c\cdot\eta\eqqcolon{\widetilde{\eta}}$. This scaling strategy helps us sidestep the problem of bounded $r_{t}$, and in this sense, we may expect the proper ${\widetilde{\eta}}$ in DNO-Prct to be relatively larger (than, e.g., $\eta$ in Algorithm 1). However, an enlarged ${\widetilde{\eta}}$ in Eq. 8 will worsen the sample complexity suggested in Theorem 1 (for details, refer to its proof in Appendix B, especially for the derivation of Eq. 18). So, to avoid the proper ${\widetilde{\eta}}$ being too large, we only use pairs with large margin as in 6 of Algorithm 2 to make sure $r_{t}(x,y_{t}^{+})-r_{t}(x,y_{t}^{-})$ is not too small. This decision is also supported empirically in techniques like RLCD (Yang et al., 2023) and Axiomatic Preference Models (Rosset et al., 2023) which highlight the importance of having large margin or clear directional differences between positive and negative LLM responses when training preference models. #### Relationship between DNO-Prct and DPO The reader may discern that DNO-Prct (Algorithm 2)—the practical implementation of DNO—can be described as an iterative version of the DPO algorithm. Such similarity is by design, intended to harness the simplicity and effectiveness of DPO (Rafailov et al., 2023) and build on empirical advancements from recent work that applies DPO iteratively (e.g., Yuan et al., 2024; Tran et al., 2024). Our experiments point to the importance of several design choices which help accommodate the general preferences, such as rankings derived from pair-wise win rates. More interestingly, our findings point to a surprising connection—that _“a meticulously designed iterative DPO algorithm” could approach the Nash equilibrium of any given general preferences._ Our general algorithmic framework—DNO (Algorithm 1)—is broader and fundamentally different from iterative DPO. For example, the DNO framework could also be directly extended to the regularized preference case (as discussed in Appendix A) or equipped with other advanced sample techniques (e.g., Liu et al., 2024b, RSO) as suggested by Theorem 1 for sample efficiency. On the other hand, although the soft policy iteration (or the KL- regularized reward optimization) is used in both DNO and DPO, they arise from fundamentally different reasons. For DNO, KL-regularization originates from online learning, no-regret learning through mirror descent (Nemirovskij and Yudin, 1983) or follow-the-regularized-leader (FTRL) (Kalai and Vempala, 2005; Cesa-Bianchi and Lugosi, 2006; Shalev-Shwartz et al., 2012; Hazan et al., 2016). For DPO and PPO, the KL-regularization is an approximation for the total variation penalty to ensure monotonic improvement of the policy (Kakade and Langford, 2002; Schulman et al., 2015). Later, this approach was simplified by Schulman et al. (2017, PPO), and recently used for post-training LLMs (Ouyang et al., 2022). ## 5 Experiments Figure 2: Comparison of various post-training techniques showing that Direct Nash Optimization (DNO) is the most effective. All methods with colorful error bands are 1) implemented by ourselves, 2) initialized with a 7B parameter Orca-2.5 LLM, and 3) are “batched on-policy” (except SFT and Offline DPO which are epochs), all else being equal. Algorithm 2 is chosen for its efficiency and simplicity from an implementation standpoint (in this section, we will use DNO to denote Algorithm 2 or DNO-Prct for simplicity). Once the input dataset $\\{x_{i}\in\mathcal{X}\\}$ is chosen, each iteration of DNO proceeds in three phrases: sampling outputs from the current policy, annotating outputs for preference pair generation, and then training the next policy with the new training pairs. Iteration 0 is defined to start by sampling from the initial SFT model to produce training data for iteration 1. ### 5.1 Experimental Setup Data: We mainly use Ultrafeedback (Cui et al., 2023), which consists of 60k prompts, several models’ outputs to those prompts, and preference annotations from GPT-4-Turbo. This dataset thus provides a source of offline preferences. For our iterative experiments, we split this dataset into three non- overlapping partitions of the inputs to be used for separate iterations of batched on-policy learning. For each input, we also collect the GPT-4-Turbo output if it was not already present in the original dataset to be reserved for $y^{\text{gold}}$. Every experiment except one in this study solely uses UltraFeedback. The exception is one “scaled up” experiment with about 10x more data sourced from a mixture of datasets aggregated including Anthropic HH (Bai et al., 2022a), UltraChat (Ding et al., 2023), MetaMathQA (Yu et al., 2023), EvolInstruct (Xu et al., 2023a), UltraFeedback (Cui et al., 2023) and Orca-2 (Mitra et al., 2023). Note that we only use the input prompts for these datasets and collect a GPT-4-Turbo responses for all 600k of these input prompts. Sampling from the Policy: At the end of training, we sample 5 outputs from the resulting student policy using top p sampling with $p=0.95$ and temperature 0.7. Several works have shown the benefit of sampling and comparing multiple diverse outputs from the policy (Yuan et al., 2023a; Mitra et al., 2024; Liu et al., 2024b; Dong et al., 2023; Wang et al., 2022). We implement a simple defect detection system which flags any sample that has a high amount of repeated n-grams as automatic negative. Preference Annotation: We use GPT-4-Turbo “as a judge” to label preferences among the 5 policy samples and 1 gold sample (which is also GPT-4-Turbo) as shown in Fig. 3. This prompt contains a few minor modifications from the that used in (Yuan et al., 2024). It implements an additive scoring framework on a 6-point scale where a score of 6 represents the highest quality answer according to certain dimensions like “correctness”, “expert knowledge”, “conciseness” etc. By following this rubric, GPT-4 acting as an annotator represents a best-effort general preference model because it compares multiple candidate responses side-by-side in the context window, and stratifies them along meaningful dimensions of quality. | | Alpaca Eval 2 | MT Bench ---|---|---|--- Technique | | Epoch --- or Iter | Len-control. --- Win Rate | Win Rate --- vs. GPT-4 | Avg. len --- (chars) | 1st --- Turn | 2nd --- Turn Avg Orca-2.5 SFT | Epoch 1 | 10.76 | 6.99 | 1174 | 7.72 | 6.02 | 6.88 Orca-2.5 SFT on Positives | Epoch 4 | 11.62 | 7.96 | 1420 | 7.62 | 6.23 | 6.92 Offline DPO (ours) | Epoch 4 | 19.49 | 18.22 | 1884 | 7.69 | 7.08 | 7.38 Self-Rewarding 70B | Iter 3 | - | 20.44 | 2552 | - | - | 7.25 SPIN (ours) | Iter 3 | 16.18 | 16.13 | 1922 | 7.58 | 7.53 | 7.55 DNO-Restrictive | Iter 3 | 21.61 | 19.21 | 1795 | 7.59 | 7.35 | 7.46 DNO-Lookahead | Epoch 1 | 18.58 | 18.28 | 1907 | 8.09 | 7.32 | 7.70 DNO | Iter 3 | 22.59 | 24.97 | 2228 | 7.62 | 7.35 | 7.48 Table 1: AlpacaEval 2.0 and MT-Bench results in our controlled setting after training on UltraFeedback. Training Pair Construction: Adhering to 6 in Algorithm 2 implies that not all pairs are suitable for training. Firstly, we must enforce the positives to be high quality in an absolute sense, and secondly, the negatives are directionally worse by a large margin. On the 6 point annotation scale, only samples that score a 5 or 6 are allowed to be positives. From the positives that meet this criteria, if any, we then construct all pairs such that the negative is at least 2 points lower. If the positive happens to be from the student, we relax this constraint to 1 point margin since the GPT-4-Turbo teacher outputs rarely receive a score less than 5 (as shown by the average teacher score in Table 2). Additionally, we are motivated to preserve the preference behavior from previous iterations so that new policies do not inadvertently regress to past bad behavior. To enforce this, we incorporate an exponentially decaying proportion of prior iterations’ training pairs into the current iteration, i.e. we sample at most 30% of training pairs from iteration $t-1$, 15% from $t-2$, and so on. We do not re-inference outputs for those inputs from the most recent policy. Recall that previous iterations’ inputs are non- overlapping with the splits for other iterations. Training: To prevent overfitting, we train our batched on-policy methods for at most one epoch on newly constructed pairs. Our effective batch size is fixed to 64 for all experiments. Our learning rate, beta, and alpha are found with brief hyperparameter searches. For most experiments, the learning rate is 5E-5, beta is either 0.1 or 0.05, and alpha is 0.005. We found that at higher iterations, the learning rate needs to be lowered. In SFT (supervised fine- tuning) experiments, our learning rate is 5E-6 and we mask out loss for the inputs. We use the open-source TRL library’s implementation to run our experiments. Evaluation: Our primary goal is to train a policy that is comparable to the most powerful state-of-the-art langauge models. Hence, AlpacaEval 2.0 (Dubois et al., 2023) is an appropriate benchmark because it computes win-rate against GPT-4-Turbo in a head-to-head fashion on a dataset of 805 input prompts that is shown to correlate with human preferences (0.93 spearman correlation with Chatbot Arena). While it is known that auto-eval methods also correlate with spurious features such as length, a new version of AlpacaEval 2.0 corrects for this with a length-controlled win-rate that has an even higher spearman correlation (0.98) with Chatbot Arena 444https://github.com/tatsu- lab/alpaca_eval. We also evaluate on MT-Bench (Zheng et al., 2023) which allows the llm-as-a- judge to first explain its reasoning before providing a scalar score on 1-10 for the candidate response to a bank of 80 questions. One crucial difference between AlpacaEval 2.0 and MT Bench is that the former asks GPT-4-Turbo to predict which of two side-by-side responses humans would prefer, weighted by the logits to represent its uncertainty, whereas MT-Bench asks the model to first generate a justification and then output a score on 1-10, but it neither defines the ratings (e.g. how a 7 is different than a 5) nor accounts for uncertainty in the logits of the score. We also evaluate on the OpenLLM leaderboard (Beeching et al., 2023), which measures reasoning ability on downstream NLP tasks like coding and question answering by evaluating the accuracy of the multiple choice answer option with the highest logit. Since our training data is primarily instruction-following and not trained to output just the sole answer option, this benchmark is not the primary target of this study; nonetheless, DNO on instruction tuning tasks ought to show no regression on reasoning tasks. | | | Annotations of Training Data | New Training Pairs ---|---|---|---|--- | inputs | | student --- length (words) | best-of-n --- student win-rate | Avg. # --- student wins | Avg. --- student score | Avg. --- teacher score $T\succ S$ | $S\succ T$ | $S\succ S$ DNO-Restrictive Iter 0 | 19.6k | 162 +/- 190 | 15.9% | 0.486 | 3.46 | 4.99 | 42.4k | 0 | 0 DNO-Restrictive Iter 1 | 19.9k | 359 +/- 350 | 34.2% | 1.11 | 4.86 | 4.77 | 17.5k | 0 | 0 DNO-Restrictive Iter 2 | 19.8k | 256 +/- 207 | 35.0% | 1.31 | 5.21 | 4.87 | 9.9k | 0 | 0 DNO Iter 0 | 19.6k | 162 +/- 190 | 15.9% | 0.486 | 3.46 | 4.99 | 30.7k | 4.2k | 25.9k DNO Iter 1 | 19.9k | 671 +/- 546 | 34.6% | 1.22 | 4.61 | 4.62 | 20.3k | 19.4k | 62.9k DNO Iter 2 | 19.8k | 361 +/- 251 | 43.6% | 1.90 | 5.25 | 4.59 | 7.1k | 32.4k | 10.9k Table 2: The dynamics of how sampled outputs from a previous iteration’s policy compare to their teacher, and how many new training pairs they give rise to in the next iteration. The crucial point is that DNO constructs new pairs where the student is compared to the teacher, whereas DNO-Restrictive, SPIN, and IPO-MD do not. ### 5.2 Results and Analysis We run several head-to-head experiments that control for hyperparameters and input data. We often refer to the policy being trained as the “student” and GPT-4 as a “teacher”; GPT-4 is also used as an annotator when prompted. SFT Baselines The first baseline is Orca-2.5 itself, which is a mistralai/Mistral-7B-v0.1 raw pretrained model fine-tuned on a new collection of Orca-2 data (Mitra et al., 2023). This model was finetuned for three epochs and achieves scores shown in the top of Table 4. All other experiments in this study are initialized with Epoch 1 of Orca-2.5. This is the solid horizontal line in Fig. 2. The second baseline is continue-SFT of Orca-2.5 training towards the positives in UltraFeedback (and masking out loss over the input prompts). If the original positive in that dataset was not from GPT-4-Turbo, we replace it with one that is. This is the red line in Fig. 2. It is clear that even offline contrastive training methods are more beneficial than additional SFT, showing that the difference between the positive and negative output provides more valuable training signal than the positive in isolation. Large Margin Filtering of Training Pairs: We ran a simple experiment of Offline DPO for one epoch on UltraFeedback data. In the control, we trained on all 63k preference pairs in the original dataset, whereas in the treatment we filtered the 42k pairs that met a large margin requirement enforcing that the positive’s scores exceeded that of the negative by at least 1.0 (out of 10) according to their GPT-4-Turbo annotator. All else was equal. Even though the treatment was trained for fewer steps on less data, it achieved an AlpacaEval 2.0 win rate of 11.60 vs 9.60 for the control, showing that fewer higher quality preference pairs is better than a higher quantity of noisy pairs (not shown in the tables). On-Policy is Better than Off-Policy One of the critical questions in this study whether to sample “on-policy” outputs from the current student to use in training pairs, or whether “off-policy” outputs collected from other models different than the student will suffice. We ran 4 epochs of Offline DPO on UltraFeedback (filtered for large margin), and as shown in Table 1, on-policy methods especially DNO surpass the off-policy DPO, even when trained for 4 epochs while the on-policy models were granted only three iterations. Recall that each iteration of batched on-policy training sees only a third of the UltraFeedback input data, whereas an epoch of Offline DPO sees the entire dataset. | | ARC-C --- (25-shot) | GSM8K --- (5-shot) | HellaSwag --- (10-shot) | MMLU --- (5-shot) | TruthfulQA --- (0-shot) | WinoGrande --- (5-shot) Avg Orca-2.5 Epoch 1 | 0.609 | 0.635 | 0.818 | 0.614 | 0.489 | 0.738 | 0.652 Orca-2.5 Epoch 3 | 0.624 | 0.641 | 0.826 | 0.624 | 0.506 | 0.746 | 0.661 SPIN (ours) Iter 3 | 0.668 | 0.448 | 0.862 | 0.623 | 0.601 | 0.759 | 0.660 DNO Iter 1 | 0.657 | 0.572 | 0.834 | 0.623 | 0.568 | 0.755 | 0.668 DNO Iter 2 | 0.663 | 0.562 | 0.845 | 0.624 | 0.580 | 0.753 | 0.671 DNO Iter 3 | 0.672 | 0.542 | 0.852 | 0.622 | 0.606 | 0.753 | 0.675 Table 3: Results on Open-LLM Leaderboard reasoning tasks, which we do not expect to decrease. Higher Quality Annotators In our study, we use GPT-4-Turbo to provide the annotations for preference pairs. However, the Self-Rewarding Language Model uses the Llama-2-70B (Touvron et al., 2023) model trained to also give feedback as the annotator, which in their study starts off with a 65% agreement rate with human-labeled preferences improving to 80% in the last iteration (Yuan et al., 2024). While it was not reported how well GPT-4-Turbo’s annotations agree with their held-out human labels, we believe that having a higher-quality annotator to start with will lead to higher quality policies. Since both our studies use UltraFeedback data, and our annotation prompt is based on their annotation prompt, we believe there is a valid comparison. We observe DNO initialized with a 7B base model outperforms the 70B parameter Self-Rewarding model over the same number of training iterations (24.97 win- rate vs 20.44 on AlpacaEval 2.0, and 7.46 MT-Bench vs 7.25), at least in part due to the higher quality preference annotations. See the dark blue band versus the gray line in Fig. 2 and the corresponding row in Table 1. However, unlike Self-Rewarding LM, we saw a slight gain rather than a drop reasoning benchmarks like ARC-Challenge (Clark et al., 2018) and HellaSwag (Zellers et al., 2019). Granted, the evaluation of OpenLLM predicts the answer with the max logit corresponding to one of the multiple-choice options, which is not congruous with how these techniques are trained. Training Pair Construction One of the most critical implementation questions in this study is how to construct training pairs that help the student policy exceed a strong teacher like GPT-4-Turbo. One approach, Self-Play Finetuning (SPIN), removes the preference annotation step and automatically assigns the teacher output to be the positive, and all student samples to be negative (Chen et al., 2024). We find in our re-implementation of SPIN that this is detrimental, presumably because this automatic assignment could lead to noisy training pairs in cases where the student might actually be preferred. The resulting win-rate of SPIN is only 16.13 after three epochs of iterative training compared to 24.97 for DNO as shown in Table 1, all else being equal. Similar results hold in the OpenLLM results in Table 3. In a second experiment, which we denote DNO-Restrictive, we annotate all preference pairs with GPT-4-Turbo as usual, but only admit training pairs where the teacher’s output is the preferred one. The difference between DNO and DNO-Restrictive is illustrated in Table 2 where 0 student-vs-teacher and student-vs-student pairs are created. The same is also true for SPIN, but SPIN would admit a greater quantity of noisy teacher-vs-student examples even when they are dis-preferred: Table 2 shows that after Iteration 2 of DNO- Restrictive, only 9.9k instances exist of the teacher being preferred over the student, whereas SPIN would have automatically created about 100k (5 samples $\times$ 20k inputs). While DNO-Restrictive is slightly better (19.21 win-rate) than SPIN, it still does not give the student a chance to compare its behavior to a powerful teacher. Absence of this signal is a major oversight, since the last row of Table 2 shows that by Iter 3, over 64% of the DNO training data (32k pairs) are cases where the student is in fact preferred over the teacher, a number which increases with iteration. We conclude it is imperative to “allow the student to become the teacher” i.e. learn from comparisons where its own outputs are preferred over a more powerful teacher. One curious phenomenon in Table 2 is that while the teacher outputs are fixed ahead of time, the annotator gives slightly lower scores to the teacher as the student improves; we are not sure if this is an innocuous artifact of preference annotations, or symptomatic of a deeper problem. Also, the total quantity of new “large margin” training pairs (not counting those sampled from previous iterations) in DNO tends to decrease as the policy improves across iterations, but we do not have enough data to quantify how this relates to a change in quality. Lookahead to Future Iterations As a curiosity, we experimented with whether a model could benefit from the knowledge of which training pairs it would generate if it could look into the future. We tested this by running three- iterations of DNO, accumulating all the preference pairs across iterations, combining and shuffling them, and then re-starting training from the initial model. In essence, this turns the batch-online DNO into an offline learning algorithm we denote as DNO-Lookahead. We trained for one epoch on the three iterations’ worth of preference data. It deteriorated more than we expected on AlpacaEval 2.0 win-rate (24.97 to 18.18), however, even more surprisingly, the MT-Bench numbers improved significantly (7.48 to 7.70). While the reasons for the relatively low correlation between MT-Bench and AlpacaEval 2.0 are not entirely clear, it is important to consider the disparity in the size of the datasets. Given that MT-Bench consists of merely 80 examples, whereas AlpacaEval 2.0 contains 10x more, we conjecture that the statistical significance and reliability of the findings from AlpacaEval 2.0 are regarded with greater confidence. DNO Scales with More Data: One of the reasons we split UltraFeedback into three non-overlapping partitions is to avoid overfitting. Another strategy to avoid overfitting is to collect more data, so we increased by a factor of 10 the instruction data based on publicly available datasets. We split a large mixture of datasets into six non-overlapping partitions of roughly 100k inputs each (and inference GPT-4-Turbo outputs for all inputs), and show that DNO- More-Data scales well in this expanded regime (see the purple line in Fig. 2 and the last row of Table 4. We make some notes on the behavior of this experiment: because each iteration builds on outputs of the previous iteration, if there are any anomalies or errors in critical components such as preference annotation, those errors will propagate and the only way to combat them is “roll back” to the iteration that introduced them. This can result in wasted time and cost, which are both already very high as shown in Appendix C. We suspect that the “depth” of iterations matters more than the “width” or number of samples within each iteration, and furthermore, that having equal number of inputs per iteration may not be optimal, but we did not test this thoroughly. From an efficiency standpoint, although this algorithm is “batched”, some optimizations can be made, such as starting to annotate sampled policy outputs are soon as they are ready instead of waiting for all inference jobs to finish. “Exploding” Lengths It is known that contrastive LLM training techniques, especially DPO, lead to longer outputs from the model which is widely suspected to be a form of “reward hacking”. Curiously, Table 2 shows that the largest jump comes after the first round of contrastive training (Iteration 1), where lengths explode by at least a factor of 2 over the initializing SFT model, before inching down again in the next iteration. We interpret this “length spike” as wasted computation optimizing towards a spurious signal; we wish we were better equipped to control this phenomenon. | | Alpaca Eval 2 | MT Bench ---|---|---|--- Technique | | Epoch --- or Iter | Len-control. --- Win Rate | Win Rate --- vs. GPT-4 | Avg. len --- (chars) | 1st --- Turn | 2nd --- Turn Avg Orca-2.5 SFT | Epoch 1 | 10.76 | 6.99 | 1174 | 7.72 | 6.02 | 6.88 Orca-2.5 SFT | Epoch 2 | 15.29 | 7.88 | 1060 | 7.56 | 6.38 | 6.98 Orca-2.5 SFT | Epoch 3 | 15.90 | 8.17 | 1058 | 7.53 | 6.73 | 7.13 DNO-More-Data | Iter 1 | 8.96 | 10.67 | 2795 | 7.00 | 6.06 | 6.53 DNO-More-Data | Iter 2 | 14.61 | 16.94 | 2782 | 7.62 | 7.23 | 7.42 DNO-More-Data | Iter 3 | 21.81 | 26.74 | 2539 | 7.74 | 6.66 | 7.21 DNO-More-Data | Iter 4 | 22.93 | 29.08 | 3033 | 7.54 | 6.92 | 7.24 DNO-More-Data | Iter 5 | 32.06 | 34.98 | 2856 | 7.10 | 6.39 | 6.75 DNO-More-Data | Iter 6 | 33.05 | 34.38 | 2683 | 7.28 | 6.65 | 6.97 Table 4: DNO-More-Data is trained on 10x more instruction data than DNO. It is still initialized with Epoch 1 of Orca-2.5 SFT, so the delta it provides in AlpacaEval 2.0 win rate is 27.39 absolute (22.29 length-controlled) ## 6 Related Work We divide the space of related work into whehter or not the techniques use SFT or contrastive losses, in offline or online update settings. Online RLHF algorithms: RLHF innovated how to align language models with human preferences (Christiano et al., 2017; Stiennon et al., 2020), but it is unstable to train and memory-intensive, requiring all three of the parameterized policy model, reward model, and advantage model to be on device for training. Reward-model Augmented SFT: Since the introduction of RLHF, several emergent techniques apply reward models in various ways, such as to filter training data or rank responses. Reward rAnked Finetuning (RAFT) (Dong et al., 2023) and RRHF (Yuan et al., 2023b) offer the conceptually simplest solution for offline preference learning, which is to sample multiple outputs from a policy, rank them with a reward model, and then finetune on the best sampled output using SFT. This resembles the iterative behavior-cloning technique DAgger (Ross et al., 2011). Offline Contrastive Preference Learning: There exist several loss functions for contrastive preference learning, first introduced in the offline setting, namely Direct Preference Optimization (Rafailov et al., 2023, DPO) and Calibrated Sequence Likelihood Estimation a.k.a. SLiC (Zhao et al., 2023). Azar et al. (2023) make it clear that point-wise reward estimates are no substitute for pair-wise preferences, and that a policy can easily overfit to deterministic preferences without proper regularization. They derive a more general objective for RLHF, IPO, to directly optimize offline preference probabilities. Statistical Rejection Sampling Optimization (RSO) generates multiple samples from an initial model, ranks them to create training pairs, and optimizes them under a unified framework encompassing DPO and SLiC (Liu et al., 2024b). Inspired by the learning-to-rank literature, Listwise preference optimization (LIPO) extends pair-wise preference learning to list-wise (Liu et al., 2024a). Preference Ranking Optimization (PRO) also learns towards list-wise preferences (Song et al., 2024). The KTO algorithm takes a different approach from DPO and does not assume that a pair of good-vs-bad outputs for the same input exist, but rather a pool of good outputs and a pool of bad outputs for any inputs exist and optimizes an “unpaired” loss (Ethayarajh et al., 2024). Iterative Reward-based Finetuning: Reinforced Self-Training (ReST) is one of the first methods to explore iterative self-improving training strategies framed as a two-stage “Grow” step that samples from the current policy, and a “Improve” step that uses a reward model to filter ever-higher quality samples that are then used to improve the policy with offline RL (Gulcehre et al., 2023). A follow-up work explores the use of AI feedback rather than reward ranking (Singh et al., 2023). On-policy Contrastive Learning: Self-Rewarding Language Models (Yuan et al., 2024) is in practice very similar to DNO. They study the benefits of batched iteratively training on preferences derived from a recent policy’s sampled outputs, but in their work, they use the policy itself as the annotator, which starts off being able to provide only weak preference signals. Self-Play Fine- Tuning (Chen et al., 2024) a.k.a SPIN and Adversarial Preference Optimization a.k.a APO (Cheng et al., 2023) are both iterative LLM training techniques that are compatible with contrastive losses, but they make a very limiting assumption that the teacher is better than the student (without regard to any annotator feedback). The Cringe Loss (Adolphs et al., 2022) is a token-level loss function that contrasts the correct next token with a hard-negative token from the vocabulary that has high logit weight but still incorrect. The Pairwise Cringe Loss (Xu et al., 2023b) applies the cringe loss to an iterative self-improving style of training. On-Policy General Preference Optimization: Wang et al. (2023) consider finding the von Neumann winner of general preferences via multi-agent RL from the theoretical perspective. Nash-MD optimizes a policy towards the Nash equilibrium of a generalized preference model using policy gradients, showing that by sampling from a mixture of policies, one can converge to the Nash equilibrium in the last iteration (Munos et al., 2023). Self-play Preference Optimization (SPO) is another online two-player mini-max game that converges to a Nash equilibrium with no-regret guarantees (Swamy et al., 2024). However, these techniques are not as data efficient as contrastive losses and are difficult to implement faithfully without cumbersome two-timescale updates (Munos et al., 2023). A concurrent improvement, IPO-MD, mitigates these difficulties by using purely on-policy IPO updates and is empirically evaluated on an article summarization task (Calandriello et al., 2024). Guo et al. (2024) also propose to eliminate rewards in online AI-feedback (OAIF) by using another LLM to annotate which of two online-sampled outputs from the current policy is preferred. However, all the above studies only consider training pairs constructed between self-play “student vs student” samples, and between student and initial $\pi_{\text{ref}}$. That is, there is no concept of a more powerful “teacher” to compare against in their training pairs. We showed in Table 2 that omitting these “student vs teacher” preferences may hinder performance. ## 7 Conclusion In this paper we achieve dual goals of post-training LLMs against a more general class of preference models while providing a practical and scalable implementation with finite-sample analysis. Our strong empirical results are based on the insight that optimizing general preference functions can be reduced to finding the Nash equilibrium of a two-player game with the payoff as the preference, and further solved by a single-play algorithm. Most techniques to optimize for this objective use soft policy iteration, which is difficult to implement faithfully and may require unstable on-policy and two- timescale updates. Our contribution, Direct Nash Optimization, addresses these challenges by approximating soft policy iteration updates with a regression- based contrastive objective in a batched manner, which is a much more stable and forgiving learning objective, and we establish a concentration bound of $\widetilde{O}(\nicefrac{{1}}{{N}})$ on the squared total variation error between the learned policy and its target of the soft policy iteration update at any given iteration $t$. Theoretically, DNO converges to the Nash equilibrium on-average, but in practice enjoys monotonic improvement across iterations. Training a 7B parameter LLM with DNO achieves state-of-the-art performance on AlpacaEval 2.0, exceeding both Mistral Large and older versions of GPT-4. We illuminate many of the practical design choices that will aid future development of iterative self-improving algorithms. ## References * Adolphs et al. [2022] Leonard Adolphs, Tianyu Gao, Jing Xu, Kurt Shuster, Sainbayar Sukhbaatar, and Jason Weston. The cringe loss: Learning what language not to model. _arXiv preprint arXiv:2211.05826_ , 2022. * Akrour et al. [2012] Riad Akrour, Marc Schoenauer, and Michèle Sebag. April: Active preference learning-based reinforcement learning. In _Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2012, Bristol, UK, September 24-28, 2012. Proceedings, Part II 23_ , pages 116–131. Springer, 2012. * Amodei et al. [2016] Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. Concrete problems in ai safety. _arXiv preprint arXiv:1606.06565_ , 2016. * Azar et al. [2023] Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. A general theoretical paradigm to understand learning from human preferences. _arXiv preprint arXiv:2310.12036_ , 2023. * Bai et al. [2022a] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. Training a helpful and harmless assistant with reinforcement learning from human feedback. _arXiv preprint arXiv:2204.05862_ , 2022a. * Bai et al. [2022b] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukosuite, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan. Constitutional ai: Harmlessness from ai feedback. _arXiv preprint arXiv:2212.08073_ , 2022b. * Beeching et al. [2023] Edward Beeching, Clémentine Fourrier, Nathan Habib, Sheon Han, Nathan Lambert, Nazneen Rajani, Omar Sanseviero, Lewis Tunstall, and Thomas Wolf. Open llm leaderboard. https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard, 2023. * Bertrand et al. [2023] Quentin Bertrand, Wojciech Marian Czarnecki, and Gauthier Gidel. On the limitations of the elo, real-world games are transitive, not additive. In _International Conference on Artificial Intelligence and Statistics_ , pages 2905–2921. PMLR, 2023. * Bradley and Terry [1952] Ralph Allan Bradley and Milton E. Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. _Biometrika_ , 39(3/4):324–345, 1952. * Brown et al. [2020] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. _Advances in neural information processing systems_ , 33:1877–1901, 2020. * Bubeck [2015] Sébastien Bubeck. Convex optimization: Algorithms and complexity. _Foundations and Trends® in Machine Learning_ , 8(3-4):231–357, 2015. * Calandriello et al. [2024] Daniele Calandriello, Daniel Guo, Remi Munos, Mark Rowland, Yunhao Tang, Bernardo Avila Pires, Pierre Harvey Richemond, Charline Le Lan, Michal Valko, Tianqi Liu, et al. Human alignment of large language models through online preference optimisation. _arXiv preprint arXiv:2403.08635_ , 2024. * Cesa-Bianchi and Lugosi [2006] Nicolo Cesa-Bianchi and Gábor Lugosi. _Prediction, learning, and games_. Cambridge university press, 2006. * Chen and Jiang [2019] Jinglin Chen and Nan Jiang. Information-theoretic considerations in batch reinforcement learning. In _International Conference on Machine Learning_ , pages 1042–1051. PMLR, 2019. * Chen et al. [2024] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. _arXiv preprint arXiv:2401.01335_ , 2024. * Cheng et al. [2023] Pengyu Cheng, Yifan Yang, Jian Li, Yong Dai, and Nan Du. Adversarial preference optimization. _arXiv preprint arXiv:2311.08045_ , 2023. * Christiano et al. [2017] Paul F. Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. _Advances in neural information processing systems_ , 30, 2017. * Clark et al. [2018] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. _arXiv:1803.05457v1_ , 2018. * Cui et al. [2023] Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Wei Zhu, Yuan Ni, Guotong Xie, Zhiyuan Liu, and Maosong Sun. Ultrafeedback: Boosting language models with high-quality feedback. _arXiv preprint arXiv:2310.01377_ , 2023. * Ding et al. [2023] Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. Enhancing chat language models by scaling high-quality instructional conversations. _arXiv preprint arXiv:2305.14233_ , 2023. * Dong et al. [2023] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, SHUM KaShun, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. _Transactions on Machine Learning Research_ , 2023. * Dubois et al. [2023] Yann Dubois, Chen Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy S. Liang, and Tatsunori B. Hashimoto. Alpacafarm: A simulation framework for methods that learn from human feedback. _Advances in Neural Information Processing Systems_ , 36, 2023. * Dudík et al. [2015] Miroslav Dudík, Katja Hofmann, Robert E. Schapire, Aleksandrs Slivkins, and Masrour Zoghi. Contextual dueling bandits. In _Conference on Learning Theory_ , pages 563–587. PMLR, 2015. * Elo [1978] Arpad E. Elo. _The rating of chessplayers, past and present_. Arco Pub., New York, 1978. ISBN 0668047216 9780668047210. URL http://www.amazon.com/Rating-Chess-Players-Past-Present/dp/0668047216. * Ethayarajh et al. [2024] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. _arXiv preprint arXiv:2402.01306_ , 2024. * Finn et al. [2016a] Chelsea Finn, Paul Christiano, Pieter Abbeel, and Sergey Levine. A connection between generative adversarial networks, inverse reinforcement learning, and energy-based models. _arXiv preprint arXiv:1611.03852_ , 2016a. * Finn et al. [2016b] Chelsea Finn, Sergey Levine, and Pieter Abbeel. Guided cost learning: Deep inverse optimal control via policy optimization. In _International conference on machine learning_ , pages 49–58. PMLR, 2016b. * Fishburn [1984] Peter C. Fishburn. Probabilistic social choice based on simple voting comparisons. _The Review of Economic Studies_ , 51(4):683–692, 1984. * Foster and Krishnamurthy [2021] Dylan J. Foster and Akshay Krishnamurthy. Efficient first-order contextual bandits: Prediction, allocation, and triangular discrimination. _Advances in Neural Information Processing Systems_ , 34:18907–18919, 2021. * Freund and Schapire [1997] Yoav Freund and Robert E. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. _Journal of computer and system sciences_ , 55(1):119–139, 1997. * Griffith et al. [2013] Shane Griffith, Kaushik Subramanian, Jonathan Scholz, Charles L. Isbell, and Andrea L. Thomaz. Policy shaping: Integrating human feedback with reinforcement learning. _Advances in neural information processing systems_ , 26, 2013. * Gulcehre et al. [2023] Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, Arnaud Doucet, Orhan Firat, and Nando de Freitas. Reinforced self-training (rest) for language modeling. _arXiv preprint arXiv:2308.08998_ , 2023. * Guo et al. [2024] Shangmin Guo, Biao Zhang, Tianlin Liu, Tianqi Liu, Misha Khalman, Felipe Llinares, Alexandre Rame, Thomas Mesnard, Yao Zhao, Bilal Piot, Johan Ferret, and Mathieu Blondel. Direct language model alignment from online ai feedback. _arXiv preprint arXiv:2402.04792_ , 2024. * Haarnoja et al. [2018] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In _International conference on machine learning_ , pages 1861–1870. PMLR, 2018. * Hazan et al. [2016] Elad Hazan et al. Introduction to online convex optimization. _Foundations and Trends® in Optimization_ , 2(3-4):157–325, 2016. * Kakade and Langford [2002] Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. In _Proceedings of the Nineteenth International Conference on Machine Learning_ , pages 267–274, 2002. * Kakade [2001] Sham M. Kakade. A natural policy gradient. _Advances in neural information processing systems_ , 14, 2001. * Kalai and Vempala [2005] Adam Kalai and Santosh Vempala. Efficient algorithms for online decision problems. _Journal of Computer and System Sciences_ , 71(3):291–307, 2005. * Knox and Stone [2008] W Bradley Knox and Peter Stone. Tamer: Training an agent manually via evaluative reinforcement. In _2008 7th IEEE international conference on development and learning_ , pages 292–297. IEEE, 2008. * Kramer [1973] Gerald H. Kramer. On a class of equilibrium conditions for majority rule. _Econometrica: Journal of the Econometric Society_ , pages 285–297, 1973. * Kreweras [1965] Germain Kreweras. Aggregation of preference orderings. In _Mathematics and Social Sciences I: Proceedings of the seminars of Menthon-Saint-Bernard, France (1–27 July 1960) and of Gösing, Austria (3–27 July 1962)_ , pages 73–79, 1965. * Lattimore and Szepesvári [2020] Tor Lattimore and Csaba Szepesvári. _Bandit algorithms_. Cambridge University Press, 2020. * Lee et al. [2023] Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. _arXiv preprint arXiv:2309.00267_ , 2023. * Liu et al. [2024a] Tianqi Liu, Zhen Qin, Junru Wu, Jiaming Shen, Misha Khalman, Rishabh Joshi, Yao Zhao, Mohammad Saleh, Simon Baumgartner, Jialu Liu, Peter J. Liu, and Xuanhui Wang. Lipo: Listwise preference optimization through learning-to-rank. _arXiv preprint arXiv:2402.01878_ , 2024a. * Liu et al. [2024b] Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J. Liu, and Jialu Liu. Statistical rejection sampling improves preference optimization. In _The Twelfth International Conference on Learning Representations_ , 2024b. * Mitra et al. [2023] Arindam Mitra, Luciano Del Corro, Shweti Mahajan, Andres Codas, Clarisse Simoes, Sahaj Agarwal, Xuxi Chen, Anastasia Razdaibiedina, Erik Jones, Kriti Aggarwal, Hamid Palangi, Guoqing Zheng, Corby Rosset, Hamed Khanpour, and Ahmed Awadallah. Orca 2: Teaching small language models how to reason. _arXiv preprint arXiv:2311.11045_ , 2023. * Mitra et al. [2024] Arindam Mitra, Hamed Khanpour, Corby Rosset, and Ahmed Awadallah. Orca-math: Unlocking the potential of slms in grade school math. _arXiv preprint arXiv:2402.14830_ , 2024. * Mnih et al. [2016] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In _International conference on machine learning_ , pages 1928–1937. PMLR, 2016. * Munos et al. [2023] Rémi Munos, Michal Valko, Daniele Calandriello, Mohammad Gheshlaghi Azar, Mark Rowland, Zhaohan Daniel Guo, Yunhao Tang, Matthieu Geist, Thomas Mesnard, Andrea Michi, Marco Selvi, Sertan Girgin, Nikola Momchev, Olivier Bachem, Daniel J. Mankowitz, Doina Precup, and Bilal Piot. Nash learning from human feedback. _arXiv preprint arXiv:2312.00886_ , 2023. * Nemirovskij and Yudin [1983] Arkadij Semenovič Nemirovskij and David Borisovich Yudin. _Problem complexity and method efficiency in optimization_. Wiley-Interscience, 1983. * OpenAI et al. [2023] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Łukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Łukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mély, Ashvin Nair, Reiichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O’Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report. _arXiv preprint arXiv:2303.08774_ , 2023. * Ouyang et al. [2022] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. _Advances in neural information processing systems_ , 35:27730–27744, 2022. * Owen [2013] Art B. Owen. _Monte Carlo theory, methods and examples_. https://artowen.su.domains/mc/, 2013. * Rafailov et al. [2023] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. _Advances in Neural Information Processing Systems_ , 36, 2023. * Ross et al. [2011] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In _Proceedings of the fourteenth international conference on artificial intelligence and statistics_ , pages 627–635. JMLR Workshop and Conference Proceedings, 2011. * Rosset et al. [2023] Corby Rosset, Guoqing Zheng, Victor Dibia, Ahmed Awadallah, and Paul Bennett. Axiomatic preference modeling for longform question answering. In _Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing_ , pages 11445–11475, 2023. * Schulman et al. [2015] John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In _International conference on machine learning_ , pages 1889–1897. PMLR, 2015. * Schulman et al. [2017] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. _arXiv preprint arXiv:1707.06347_ , 2017. * Shalev-Shwartz et al. [2012] Shai Shalev-Shwartz et al. Online learning and online convex optimization. _Foundations and Trends® in Machine Learning_ , 4(2):107–194, 2012. * Simpson [1969] Paul B. Simpson. On defining areas of voter choice: Professor tullock on stable voting. _The Quarterly Journal of Economics_ , 83(3):478–490, 1969. * Singh et al. [2023] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Gamaleldin Elsayed, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshiteej Mahajan, Laura Culp, Lechao Xiao, Maxwell L. Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Warkentin, Yundi Qian, Yamini Bansal, Ethan Dyer, Behnam Neyshabur, Jascha Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. _arXiv preprint arXiv:2312.06585_ , 2023. * Song et al. [2024] Feifan Song, Bowen Yu, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. In _Proceedings of the AAAI Conference on Artificial Intelligence_ , volume 38, pages 18990–18998, 2024. * Stiennon et al. [2020] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F. Christiano. Learning to summarize with human feedback. _Advances in Neural Information Processing Systems_ , 33:3008–3021, 2020. * Swamy et al. [2024] Gokul Swamy, Christoph Dann, Rahul Kidambi, Zhiwei Steven Wu, and Alekh Agarwal. A minimaximalist approach to reinforcement learning from human feedback. _arXiv preprint arXiv:2401.04056_ , 2024. * Touvron et al. [2023] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and fine-tuned chat models. _arXiv preprint arXiv:2307.09288_ , 2023. * Tran et al. [2024] Hoang Tran, Chris Glaze, and Braden Hancock. snorkelai/snorkel-mistral-pairrm-dpo, 2024. https://huggingface.co/snorkelai/Snorkel-Mistral-PairRM-DPO. * Tunstall et al. [2023] Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes Belkada, Shengyi Huang, Leandro von Werra, Clémentine Fourrier, Nathan Habib, Nathan Sarrazin, Omar Sanseviero, Alexander M. Rush, and Thomas Wolf. Zephyr: Direct distillation of lm alignment. _arXiv preprint arXiv:2310.16944_ , 2023. * Wang et al. [2022] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. _arXiv preprint arXiv:2203.11171_ , 2022. * Wang et al. [2023] Yuanhao Wang, Qinghua Liu, and Chi Jin. Is rlhf more difficult than standard rl? _arXiv preprint arXiv:2306.14111_ , 2023. * Wirth et al. [2017] Christian Wirth, Riad Akrour, Gerhard Neumann, and Johannes Fürnkranz. A survey of preference-based reinforcement learning methods. _Journal of Machine Learning Research_ , 18(136):1–46, 2017. * Xie et al. [2021] Tengyang Xie, Ching-An Cheng, Nan Jiang, Paul Mineiro, and Alekh Agarwal. Bellman-consistent pessimism for offline reinforcement learning. _Advances in neural information processing systems_ , 34:6683–6694, 2021. * Xie et al. [2023] Tengyang Xie, Dylan J. Foster, Yu Bai, Nan Jiang, and Sham M. Kakade. The role of coverage in online reinforcement learning. In _The Eleventh International Conference on Learning Representations_ , 2023. * Xiong et al. [2023] Wei Xiong, Hanze Dong, Chenlu Ye, Han Zhong, Nan Jiang, and Tong Zhang. Iterative preference learning from human feedback: Bridging theory and practice for rlhf under kl-constraint. _arXiv preprint arXiv:2312.11456_ , 2023. * Xu et al. [2023a] Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. Wizardlm: Empowering large language models to follow complex instructions, 2023a. * Xu et al. [2023b] Jing Xu, Andrew Lee, Sainbayar Sukhbaatar, and Jason Weston. Some things are more cringe than others: Preference optimization with the pairwise cringe loss. _arXiv preprint arXiv:2312.16682_ , 2023b. * Yang et al. [2023] Kevin Yang, Dan Klein, Asli Celikyilmaz, Nanyun Peng, and Yuandong Tian. Rlcd: Reinforcement learning from contrast distillation for language model alignment. _arXiv preprint arXiv:2307.12950_ , 2023. * Yu et al. [2023] Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T. Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models, 2023. * Yuan et al. [2024] Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. _arXiv preprint arXiv:2401.10020_ , 2024. * Yuan et al. [2023a] Zheng Yuan, Hongyi Yuan, Chengpeng Li, Guanting Dong, Chuanqi Tan, and Chang Zhou. Scaling relationship on learning mathematical reasoning with large language models. _arXiv preprint arXiv:2308.01825_ , 2023a. * Yuan et al. [2023b] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrhf: Rank responses to align language models with human feedback without tears. _arXiv preprint arXiv:2304.05302_ , 2023b. * Zellers et al. [2019] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? _arXiv preprint arXiv:1905.07830_ , 2019. * Zhan et al. [2024] Wenhao Zhan, Masatoshi Uehara, Nathan Kallus, Jason D. Lee, and Wen Sun. Provable offline preference-based reinforcement learning. In _The Twelfth International Conference on Learning Representations_ , 2024. * Zhao et al. [2023] Yao Zhao, Misha Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, and Peter J. Liu. Calibrating sequence likelihood improves conditional language generation. In _The Eleventh International Conference on Learning Representations_ , 2023. * Zheng et al. [2023] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena. _Advances in Neural Information Processing Systems_ , 36, 2023. Appendix ## Appendix A Extension to Regularized Preferences In this section, we discuss how to extend the DNO framework to the case of regularized preferences (defined in Eq. 5), $\displaystyle\mathcal{P}^{\pi,\pi^{\prime}}_{\tau}(y\succ y^{\prime}\mid x)=\mathcal{P}(y\succ y^{\prime}\mid x)-\tau\log\frac{\pi(y\mid x)}{\pi_{\mathsf{ref}}(y\mid x)}+\tau\log\frac{\pi^{\prime}(y\mid x)}{\pi_{\mathsf{ref}}(y\mid x)},$ which was first introduced and solved by Munos et al. [2023] via Nash-MD introduced earlier. One can notice that the only difference between SPO and Nash-MD is that SPO uses the last iteration policy $\pi_{t}$ for both constructing reward $r_{t}$ and performing a soft policy iteration update, whereas Nash-MD uses the smoothed version $\pi_{t}^{\tau}$ (firstly defined in Eq. 7), $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:def_pitsmooth_app}}{e}q:def_{p}itsmooth_{a}pp}\pi_{t}^{\tau}(y\mid x)\coloneqq\frac{\pi_{t}(y\mid x)^{1-\nicefrac{{\tau}}{{\eta}}}\pi_{\mathsf{ref}}(y\mid x)^{\nicefrac{{\tau}}{{\eta}}}}{\sum_{y^{\prime}\in\mathcal{Y}}\pi_{t}(y\mid x)^{1-\nicefrac{{\tau}}{{\eta}}}\pi_{\mathsf{ref}}(y\mid x)^{\nicefrac{{\tau}}{{\eta}}}},~{}\forall(x,y)\in\mathcal{X}\times\mathcal{Y},$ (14) for both. This allows Nash-MD to obtain a late-iteration guarantee. On the other hand, due to the symmetry of regularized preferences, if we consider on-average convergence case, it is likely that SPO can be adapted with a simpler way as follows: for each $t=1,2,\dotsc,T$, (i) $\displaystyle~{}r_{t}(x,y)\leftarrow{\mathbb{E}}_{y^{\prime}\sim\pi_{t}(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right],~{}~{}\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$ (ii) $\displaystyle~{}\pi_{t+1}(\cdot\mid x)\leftarrow\frac{1}{Z_{t}(x)}\pi_{t}^{\tau}(\cdot\mid x)\exp\left(\frac{r_{t}(x,\cdot)}{\eta}\right),~{}~{}\forall x\in\mathcal{X},$ where $Z_{t}(x)\coloneqq\sum_{y\in\mathcal{Y}}\pi_{t}^{\tau}(y\mid x)\exp\left(\frac{r_{t}(x,y)}{\eta}\right)$ is the partition function for iteration $t$. Here, the smoothed policy $\pi_{t}^{\tau}$ is only used in the soft policy iteration step, and this coincides with the OMD algorithm from Munos et al. [2023]. Algorithm 3 DNO (Regularized Preferences Version) input: General preference function $\mathcal{P}$, learning rate $\eta$, coefficient of KL-regularization $\tau$, number of iterations $T$, prompt distribution $\rho$. 1:Initialize $\pi_{1}\leftarrow{\sf unif}(\mathcal{A})$. 2:for iteration $t=1,2,\dotsc,T$ do 3: Compute $r_{t}(x,y)$ by, 4: Option I:$\triangleright$ for on-average convergence 5: $r_{t}(x,y)\leftarrow{\mathbb{E}}_{y^{\prime}\sim\pi_{t}(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right],~{}\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$. 6: Option II:$\triangleright$ for last-iteration convergence 7: $r_{t}(x,y)\leftarrow{\mathbb{E}}_{y^{\prime}\sim\pi_{t}^{\tau}(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right],~{}\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$, where $\pi_{t}^{\tau}$ is defined in Eq. 14. 8: Obtain $\pi_{t+1}$ by, $\begin{gathered}\pi_{t+1}\leftarrow\mathop{\mathrm{argmax}}_{\pi\in\Pi}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\mathcal{D}_{t}}\bigg{\\{}\sigma\left(r_{t}(x,y_{1})-r_{t}(x,y_{2})\right)\log\left[\sigma\left(\eta\log\frac{\pi(y_{1}\mid x)}{\widetilde{\pi}_{t}^{\tau}(y_{1}\mid x)}-\eta\log\frac{\pi(y_{2}\mid x)}{\widetilde{\pi}_{t}^{\tau}(y_{2}\mid x)}\right)\right]\\\ \hskip 120.0pt+\sigma\left(r_{t}(x,y_{2})-r_{t}(x,y_{1})\right)\log\left[\sigma\left(\eta\log\frac{\pi(y_{2}\mid x)}{\widetilde{\pi}_{t}^{\tau}(y_{2}\mid x)}-\eta\log\frac{\pi(y_{1}\mid x)}{\widetilde{\pi}_{t}^{\tau}(y_{1}\mid x)}\right)\right]\bigg{\\}},\end{gathered}$ where $\mathcal{D}_{t}$ is generated by $x\sim\rho,y_{1}\sim\mu_{1,t}(\cdot\mid x),y_{2}\sim\mu_{2,t}(\cdot\mid x)$ with some policies $\mu_{1,t}$ and $\mu_{2,t}$, and $\widetilde{\pi}_{t}^{\tau}(y\mid x)\coloneqq\pi_{t}(y\mid x)^{1-\nicefrac{{\tau}}{{\eta}}}\pi_{\mathsf{ref}}(y\mid x)^{\nicefrac{{\tau}}{{\eta}}},\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$ (the unnormalized version of $\pi_{t}^{\tau}(y\mid x)$ defined in Eq. 14). 9:end for 10:return $\bar{\pi}={\sf unif}(\pi_{1:T})$. Algorithm 4 DNO-Prct (Regularized Preferences Version) input: General preference function $\mathcal{P}$, learning rate ${\widetilde{\eta}}$, coefficient of KL-regularization $\tau$, number of iterations $T$, reference policy $\pi_{\mathsf{ref}}$, seed dataset $\mathcal{D}_{0}=\\{(x,y^{\mathsf{gold}})\\}$ where $x\sim\rho$ and $y\sim\pi_{\mathsf{gold}}(\cdot\mid x)$, reference model $\pi_{\mathsf{ref}}$. 1:Initialize $\pi_{1}\leftarrow\pi_{\mathsf{ref}}$. 2:for iteration $t=1,2,\dotsc,T$ do 3: Sample _batched on-policy_ responses: 4: Option I: $\triangleright$ for on-average convergence 5: Sample $K$ outputs per per prompt using the current $\pi_{t}$: $\\{y_{t}^{1},y_{t}^{2},\dotsc,y_{t}^{K}\\}\sim\pi_{t}(\cdot\mid x)$, $\forall x\in\mathcal{D}_{0}$. 6: Option II: $\triangleright$ for last-iteration convergence 7: Sample $K$ outputs per per prompt using the smoothed current policy $\pi_{t}^{\tau}$: $\\{y_{t}^{1},y_{t}^{2},\dotsc,y_{t}^{K}\\}\sim\pi_{t}^{\tau}(\cdot\mid x)$, $\forall x\in\mathcal{D}_{0}$, where $\pi_{t}^{\tau}$ is defined in Eq. 14 with accommodating $\eta\to{\widetilde{\eta}}$. 8: Rank responses: For each $x\in\mathcal{D}_{0}$, rank the corresponding $\\{y_{t}^{1},y_{t}^{2},\dotsc,y_{t}^{K},y^{\mathsf{gold}}\\}$ using the pair- wise win-rate by sampling from the general preference function $\mathcal{P}$. 9: Filter and construct preference pairs: Construct $\mathcal{D}_{t}=\\{(x,y_{t}^{+},y_{t}^{-})\\}$, for all $x\in\mathcal{D}_{0}$, and $(y_{t}^{+},y_{t}^{-})$ are large-margin pairs (based on the win-rate rank) within the responses for $x$ from the previous step. 10: Contrastive learning: Obtain $\pi_{t+1}$ by, $\displaystyle\pi_{t+1}\leftarrow\mathop{\mathrm{argmax}}_{\pi\in\Pi}{\mathbb{E}}_{(x,y_{t}^{+},y_{t}^{-})\sim\mathcal{D}_{t}}\log\left[\sigma\left({\widetilde{\eta}}\log\frac{\pi(y_{t}^{+}\mid x)}{\widetilde{\pi}_{t}^{\tau}(y_{t}^{+}\mid x)}-{\widetilde{\eta}}\log\frac{\pi(y_{t}^{-}\mid x)}{\widetilde{\pi}_{t}^{\tau}(y_{t}^{-}\mid x)}\right)\right],$ where $\widetilde{\pi}_{t}^{\tau}(y\mid x)\coloneqq\pi_{t}(y\mid x)^{1-\nicefrac{{\tau}}{{{\widetilde{\eta}}}}}\pi_{\mathsf{ref}}(y\mid x)^{\nicefrac{{\tau}}{{{\widetilde{\eta}}}}},\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$ (the unnormalized version of $\pi_{t}^{\tau}(y\mid x)$ defined in Eq. 14, after accommodating $\eta\to{\widetilde{\eta}}$). 11:end for 12:return best of $\pi_{1:(T+1)}$ on the validation data. Based on discuss above, we can then obtain the extension of DNO to the regularized preferences in Algorithm 3, and its practical implementation in Algorithm 4. Note that, similar to Nash-MD, the late-iteration option for both Algorithm 3 and Algorithm 4 requires sampling from the smoothed policy $\pi_{t}^{\tau}$ (the mixture between $\pi_{t}$ and $\pi_{\mathsf{ref}}$, defined in Eq. 14). One solution to address this can be sampling from the token-level between $\pi_{t}$ and $\pi_{\mathsf{ref}}$ instead as suggested by Munos et al. [2023]. ## Appendix B Detailed Proofs In this section, we provide detailed proofs for our theoretical results. Note that, the definitions and assumptions presented heavily adopts the ideas related to version space and concentrability from reinforcement learning theory literature [esp., Xie et al., 2021, 2023]. Nevertheless, the descriptions provided herein are intentionally simplified to elucidate the core insights into the algorithmic design. A full and exhaustive theoretical analysis falls outside the primary scope of this paper. We now make the following definitions and assumptions. ###### Definition 1 (Feasible solution space). For each iteration $t\in[T]$, we define $\Pi_{t}\subseteq\Pi$ as the feasible solution space for iteration $t$. The $\pi_{t}$ obtained by Algorithm 1 is always belong to $\Pi_{t}$, regardless of the randomness of the data sampling procedure in Algorithm 1. Here, Definition 1 follows a similar spirit as the version space in RL theory literature, where $\Pi_{t}$ only contains policies that have a small empirical loss, which can be further converted to a small population loss under standard concentration procedures. ###### Definition 2 (Concentrability coefficient over the feasible solution space). For all $t\in[T]$, suppose $\Pi_{t}$ is defined in Definition 1, and $\mu_{1,t}$ and $\mu_{2,t}$ are some given data generate policy. Now, for any $t\in[T]$, we define $\mathfrak{C}_{t}$ to be the concentrability coefficient at iteration $t$ over its feasible solution space, where $\displaystyle\frac{{\mathbb{E}}_{x\sim\rho,y_{1}\sim\pi_{t+1}^{\star}(\cdot\mid x),y_{2}\sim\pi_{t+1}(\cdot\mid x)}\left[\left(\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}-\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)^{2}\right]}{{\mathbb{E}}_{x\sim\rho,y_{1}\sim\mu_{1,t}(\cdot\mid x),y_{2}\sim\mu_{2,t}(\cdot\mid x)}\left[\left(\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}-\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)^{2}\right]}\leq\mathfrak{C}_{t},$ for any $\pi_{t+1}\in\Pi_{t+1}$ and any $\pi_{t+1}^{\star}\in\left\\{\frac{1}{Z_{\pi}(x)}\pi(\cdot\mid x)\exp\left(\frac{r_{\pi}(x,\cdot)}{\eta}\right):\pi\in\Pi_{t}\right\\}$; and here we use the definition of $r_{\pi}(x,y)\coloneqq{\mathbb{E}}_{y^{\prime}\sim\pi(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right],\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$, and $Z_{\pi}(x)=\sum_{y\in\mathcal{Y}}\pi(y\mid x)\exp\left(\frac{r_{\pi}(x,y)}{\eta}\right),\forall x\in\mathcal{X}$. Definition 2 can be viewed as a natural extension of concentrability from the (offline) reinforcement learning literature to our setup. ###### Assumption 1 (Realizability over the feasible solution space). For any $\pi\in\Pi_{t}$ where $\Pi_{t}$ is defined in Definition 1 for all $t\in[T]$, we assume the following soft-policy iteration update $\displaystyle\pi^{\sf SPI}(\cdot\mid x)\coloneqq\frac{1}{Z_{\pi}(x)}\pi(\cdot\mid x)\exp\left(\frac{r_{\pi}(x,\cdot)}{\eta}\right),$ where $r_{\pi}(x,y)\coloneqq{\mathbb{E}}_{y^{\prime}\sim\pi(\cdot\mid x)}\left[\mathcal{P}(y\succ y^{\prime}\mid x)\right],\forall(x,y)\in\mathcal{X}\times\mathcal{Y}$, and $Z_{\pi}(x)=\sum_{y\in\mathcal{Y}}\pi(y\mid x)\exp\left(\frac{r_{\pi}(x,y)}{\eta}\right),\forall x\in\mathcal{X}$ is the partition function. ###### Assumption 2 (Boundedness over the feasible solution space). Suppose $\Pi_{t}$ is defined in Definition 1 for all $t\in[T]$, then we assume $\log\frac{\pi(y\mid x)}{\pi_{t}(y\mid x)}\in[-R_{\max},R_{\max}]$ for all $\pi\in\Pi$, $\pi_{t}\in\Pi_{t}$, and $(x,y)\in\mathcal{X}\times\mathcal{Y}$. Assumption 2 may appear somewhat unconventional, as it explicitly assumes boundedness on the log probabilities. Nonetheless, it is important to note that the value of $\log\frac{\pi(y\mid x)}{\pi_{t}(y\mid x)}$ is directly measurable and controllable in practice, which is different from the common use case, such as maximum likelihood problems. ###### Theorem 2 (Formal Version of Theorem 1). Under Assumptions 1 and 2, and fix an arbitrary iteration $t\in[T]$. Suppose $\pi_{t+1}$ is from 4 of Algorithm 1, and $\pi_{t+1}^{\star}$ is defined in Eq. 9. Then, we have $\displaystyle{\mathbb{E}}_{x\sim\rho}\left[\left(D_{\mathrm{TV}}(\pi_{t+1}(\cdot\mid x),\pi_{t+1}^{\star}(\cdot\mid x))\right)^{2}\right]\leq\mathcal{O}\left(\frac{\mathfrak{C}_{t}R_{\max}^{2}\log(\nicefrac{{|\Pi|}}{{\delta}})}{N}\right),$ where $\mathfrak{C}_{t}$ is defined in Definition 2. ###### Proof of Theorem 2. We will now present the proof using the following two-step procedure. Step 1: From regression with log loss to squared error bound. By standard results on the regression with the logarithmic loss, we know, $\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:logloss_reg_bound}}{e}q:logloss_{r}eg_{b}ound}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\mathcal{D}_{t}}\Big{[}\sigma\big{(}\Delta_{t}^{\star}(x,y_{1},y_{2})\big{)}\log\left[\sigma\big{(}\Delta_{\pi_{t+1},t}(x,y_{1},y_{2})\big{)}\right]+\sigma\big{(}\Delta_{t}^{\star}(x,y_{2},y_{1})\big{)}\log\left[\sigma\big{(}\Delta_{\pi_{t+1},t}(x,y_{2},y_{1})\big{)}\right]\Big{]}\lesssim\frac{\log(\nicefrac{{|\Pi|}}{{\delta}})}{N}.$ (15) Note that similar results could also apply beyond finite $\Pi$. For simplicity, we omit the detailed discussion in our paper. For more in-depth discussions about regression with the logarithmic loss, the reader can refer to, e.g., Foster and Krishnamurthy [2021]. Next, by the Pinsker’s inequality, we have for any $z,{\widehat{z}}\in[0,1]$, $\displaystyle\frac{(z-{\widehat{z}})^{2}}{2}\leq z\log\left(\frac{z}{{\widehat{z}}}\right)+(1-z)\log\left(\frac{1-z}{1-{\widehat{z}}}\right).$ Substituting the $z$ and ${\widehat{z}}$ with Eq. 11 and combining with Eq. 15, we obtain that $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:reward_sq_Dt}}{e}q:reward_{s}q_{D}t}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\mathcal{D}_{t}}\left[\big{(}\sigma\left(r_{t}(x,y_{1})-r_{t}(x,y_{2})\right)-\sigma\left(r_{\pi_{t+1},t}(x,y_{1})-r_{\pi_{t+1},t}(x,y_{2})\right)\big{)}^{2}\right]\lesssim\frac{\log(\nicefrac{{|\Pi|}}{{\delta}})}{N},$ (16) where $a\lesssim b$ means $a\leq c\cdot b$ for some absolute constant $c$. Then, by the standard concentration for squared loss, e.g., Lemma A.4 of Xie et al. [2021] with $\gamma=0$, Eq. 16 implies $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:reward_sq_pop}}{e}q:reward_{s}q_{p}op}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\rho\times\mu_{1:2,t}}\left[\big{(}\sigma\left(r_{t}(x,y_{1})-r_{t}(x,y_{2})\right)-\sigma\left(r_{\pi_{t+1},t}(x,y_{1})-r_{\pi_{t+1},t}(x,y_{2})\right)\big{)}^{2}\right]\lesssim\frac{\log(\nicefrac{{|\Pi|}}{{\delta}})}{N},$ (17) where we use “$\times$” as the shorthand of joint distribution for the sake of simplicity, for example, $(x,y_{1},y_{2})\sim\rho\times\mu_{1:2,t}$ is shorthand for $x\sim\rho,y_{1}\sim\mu_{1,t}(\cdot\mid x),y_{2}\sim\mu_{2,t}(\cdot\mid x)$. By the definition of $r_{t}$ in 3 of Algorithm 1, we know $r_{t}(x,y)\in[0,1]$ for all $(x,y)\in\mathcal{X}\times\mathcal{Y}$. Thus, by a variant of mean value theorem, we know $\displaystyle~{}\big{|}r_{t}(x,y_{1})-r_{t}(x,y_{2})-r_{\pi_{t+1},t}(x,y_{1})+r_{\pi_{t+1},t}(x,y_{2})\big{|}$ (18) $\displaystyle\leq$ $\displaystyle~{}\frac{\eta R_{\max}}{1-\sigma(1)}\big{|}\sigma\left(r_{t}(x,y_{1})-r_{t}(x,y_{2})\right)-\sigma\left(r_{\pi_{t+1},t}(x,y_{1})-r_{\pi_{t+1},t}(x,y_{2})\right)\big{|},$ for any $(x,y_{1},y_{2})\in\mathcal{X}\times\mathcal{Y}\times\mathcal{Y}$, where $R_{\max}$ is introduced from Assumption 2. This is because: let $a\coloneqq r_{t}(x,y_{1})-r_{t}(x,y_{2})\in[-1,1]$, and $b\coloneqq r_{\pi_{t+1},t}(x,y_{1})-r_{\pi_{t+1},t}(x,y_{2})\in[-\eta R_{\max},\eta R_{\max}]$, and, then, we can directly verify that the slope we need to bound $\nicefrac{{\big{|}a-b\big{|}}}{{\big{|}\sigma\left(a\right)-\sigma\left(b\right)\big{|}}}$ reaches its maximum at $a=1$ and $b=\eta R_{\max}$. Combining Eqs. 17 and 18, we obtain $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:dno_reward_bound}}{e}q:dno_{r}eward_{b}ound}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\rho\times\mu_{1:2,t}}\left[\big{(}r_{t}(x,y_{1})-r_{t}(x,y_{2})-r_{\pi_{t+1},t}(x,y_{1})+r_{\pi_{t+1},t}(x,y_{2})\big{)}^{2}\right]\lesssim\frac{\eta^{2}R_{\max}^{2}\log(\nicefrac{{|\Pi|}}{{\delta}})}{N}.$ (19) Step 2: Concentration in the policy space. We now reason about the concentration of $\pi_{t+1}\to\pi_{t+1}^{\star}$ from Eq. 19, where $\pi_{t+1}^{\star}$ is defined in Eq. 9 and $\pi_{t+1}$ is the policy corresponding to the learned $r_{\pi_{t+1},t}$. By the definition of $r_{\pi,t}$ in Eq. 10, we have $\displaystyle~{}r_{t}(x,y_{1})-r_{t}(x,y_{2})-r_{\pi_{t+1},t}(x,y_{1})+r_{\pi_{t+1},t}(x,y_{2})$ $\displaystyle=$ $\displaystyle~{}r_{t}(x,y_{1})-r_{t}(x,y_{2})-\eta\log\frac{\pi_{t+1}(y_{1}\mid x)}{\pi_{t}(y_{1}\mid x)}+\eta\log\frac{\pi_{t+1}(y_{2}\mid x)}{\pi_{t}(y_{2}\mid x)}$ $\displaystyle=$ $\displaystyle~{}\eta\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}-\eta\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}.$ This implies $\displaystyle{\mathbb{E}}_{(x,y_{1},y_{2})\sim\rho\times\mu_{1:2,t}}\left[\left(\eta\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}-\eta\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)^{2}\right]\lesssim\frac{\eta^{2}R_{\max}^{2}\log(\nicefrac{{|\Pi|}}{{\delta}})}{N}$ $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:oneside_dno_bound}}{e}q:oneside_{d}no_{b}ound}\Longrightarrow{\mathbb{E}}_{(x,y_{1},y_{2})\sim\rho\times\pi_{t+1}^{\star}\times\pi_{t+1}}\left[\left(\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}-\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)^{2}\right]\lesssim\frac{\mathfrak{C}_{t}R_{\max}^{2}\log(\nicefrac{{|\Pi|}}{{\delta}})}{N},$ (20) where the last step follows from the definition of $\mathfrak{C}_{t}$ (Definition 2). On the other hand, we have $\displaystyle~{}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\rho\times\pi_{t+1}^{\star}\times\pi_{t+1}}\left[\left(\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}-\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)^{2}\right]$ $\displaystyle=$ $\displaystyle~{}{\mathbb{E}}_{(x,y_{1},y_{2})\sim\rho\times\pi_{t+1}^{\star}\times\pi_{t+1}}\left[\left(\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}\right)^{2}+\left(\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)^{2}-2\left(\log\frac{\pi_{t+1}^{\star}(y_{1}\mid x)}{\pi_{t+1}(y_{1}\mid x)}\right)\cdot\left(\log\frac{\pi_{t+1}^{\star}(y_{2}\mid x)}{\pi_{t+1}(y_{2}\mid x)}\right)\right]$ $\displaystyle=$ $\displaystyle~{}{\mathbb{E}}_{(x,y)\sim\rho\times\pi_{t+1}^{\star}}\left[\left(\log\frac{\pi_{t+1}^{\star}(y\mid x)}{\pi_{t+1}(y\mid x)}\right)^{2}\right]+{\mathbb{E}}_{(x,y)\sim\rho\times\pi_{t+1}}\left[\left(\log\frac{\pi_{t+1}^{\star}(y\mid x)}{\pi_{t+1}(y\mid x)}\right)^{2}\right]$ $\displaystyle~{}+2{\mathbb{E}}_{x\sim\rho}\Bigg{[}\underbrace{{\mathbb{E}}_{y\sim\pi_{t+1}^{\star}(\cdot\mid x)}\left[\log\frac{\pi_{t+1}^{\star}(y\mid x)}{\pi_{t+1}(y\mid x)}\right]}_{=D_{\mathrm{KL}}(\pi^{\star}_{t+1}(\cdot\mid x)~{}\|~{}\pi_{t+1}(\cdot\mid x))}\cdot\underbrace{{\mathbb{E}}_{y\sim\pi_{t+1}(\cdot\mid x)}\left[\log\frac{\pi_{t+1}(y\mid x)}{\pi_{t+1}^{\star}(y\mid x)}\right]}_{=D_{\mathrm{KL}}(\pi_{t+1}(\cdot\mid x)~{}\|~{}\pi^{\star}_{t+1}(\cdot\mid x))}\Bigg{]}$ $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:dno_mid}}{e}q:dno_{m}id}\geq$ $\displaystyle~{}{\mathbb{E}}_{(x,y)\sim\rho\times\pi_{t+1}^{\star}}\left[\left(\log\frac{\pi_{t+1}^{\star}(y\mid x)}{\pi_{t+1}(y\mid x)}\right)^{2}\right]+{\mathbb{E}}_{(x,y)\sim\rho\times\pi_{t+1}}\left[\left(\log\frac{\pi_{t+1}^{\star}(y\mid x)}{\pi_{t+1}(y\mid x)}\right)^{2}\right].$ (21) Next, we fix an arbitrary ${\widetilde{x}}\in\mathcal{X}$, and we have $\displaystyle~{}{\mathbb{E}}_{y\sim\pi_{t+1}^{\star}(\cdot\mid{\widetilde{x}})}\left[\left(\log\frac{\pi_{t+1}^{\star}(y\mid{\widetilde{x}})}{\pi_{t+1}(y\mid{\widetilde{x}})}\right)^{2}\right]+{\mathbb{E}}_{y\sim\pi_{t+1}(\cdot\mid{\widetilde{x}})}\left[\left(\log\frac{\pi_{t+1}^{\star}(y\mid{\widetilde{x}})}{\pi_{t+1}(y\mid{\widetilde{x}})}\right)^{2}\right]$ $\displaystyle\geq$ $\displaystyle~{}\left({\mathbb{E}}_{y\sim\pi_{t+1}^{\star}(\cdot\mid{\widetilde{x}})}\left[\left|\log\frac{\pi_{t+1}^{\star}(y\mid{\widetilde{x}})}{\pi_{t+1}(y\mid{\widetilde{x}})}\right|\right]\right)^{2}+\left({\mathbb{E}}_{y\sim\pi_{t+1}(\cdot\mid{\widetilde{x}})}\left[\left|\log\frac{\pi_{t+1}^{\star}(y\mid{\widetilde{x}})}{\pi_{t+1}(y\mid{\widetilde{x}})}\right|\right]\right)^{2}$ (by Jensen’s inequality) $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:dno_fixx}}{e}q:dno_{f}ixx}\gtrsim$ $\displaystyle~{}\left({\mathbb{E}}_{y\sim\pi_{t+1}^{\star}(\cdot\mid{\widetilde{x}})}\left[\left|\log\frac{\pi_{t+1}^{\star}(y\mid{\widetilde{x}})}{\pi_{t+1}(y\mid{\widetilde{x}})}\right|\right]+{\mathbb{E}}_{y\sim\pi_{t+1}(\cdot\mid{\widetilde{x}})}\left[\left|\log\frac{\pi_{t+1}^{\star}(y\mid{\widetilde{x}})}{\pi_{t+1}(y\mid{\widetilde{x}})}\right|\right]\right)^{2},$ (22) where $a\gtrsim b$ means $a\geq c\cdot b$ for some absolute constant $c$. We now recall the definition of $f$-divergence: $D_{f}(p,q)\coloneqq{\mathbb{E}}_{y\sim q}[f(\nicefrac{{p(y)}}{{q(y)}})]$ for two distributions $p$ and $q$, where $f:\mathbb{R}^{+}\to\mathbb{R}$ is convex with $f(1)=0$. Thus, we can notice that, $\displaystyle\addcontentsline{lla}{section}{\numberline{\string\crtrefnumber{eq:dno_df}}{e}q:dno_{d}f}{\mathbb{E}}_{y\sim p}\left[\left|\log\frac{p(y)}{q(y)}\right|\right]+{\mathbb{E}}_{y\sim q}\left[\left|\log\frac{p(y)}{q(y)}\right|\right]=D_{f_{1}}(p,q),\quad\text{where}~{}f_{1}(u)\coloneqq(1+u)\cdot\left|\log(u)\right|,~{}u\in\mathbb{R}^{+}.$ (23)
# Efficient Long-Range Entanglement using Dynamic Circuits Elisa Bäumer<EMAIL_ADDRESS>IBM Quantum, IBM Research – Zurich, 8803 Rüschlikon, Switzerland Vinay Tripathi IBM Quantum, IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA Department of Physics & Astronomy, University of Southern California, Los Angeles, California 90089, USA Derek S. Wang IBM Quantum, IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA Patrick Rall IBM Quantum, MIT-IBM Watson AI Lab, Cambridge, MA 02142, USA Edward H. Chen IBM Quantum, Almaden Research Center, San Jose, CA 95120, USA IBM Quantum, Research Triangle Park, NC 27709, USA Swarnadeep Majumder IBM Quantum, IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA Alireza Seif IBM Quantum, IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA Zlatko K. Minev IBM Quantum, IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA ###### Abstract Quantum simulation traditionally relies on unitary dynamics, inherently imposing efficiency constraints on the generation of intricate entangled states. In principle, these limitations can be superseded by non-unitary, dynamic circuits. These circuits exploit measurements alongside conditional feed-forward operations, providing a promising approach for long-range entangling gates, higher effective connectivity of near-term hardware, and more efficient state preparations. Here, we explore the utility of shallow dynamic circuits for creating long-range entanglement on large-scale quantum devices. Specifically, we study two tasks: CNOT gate teleportation between up to 101 qubits by feeding forward 99 mid-circuit measurement outcomes, and the preparation of Greenberger–Horne–Zeilinger (GHZ) states with genuine entanglement. In the former, we observe that dynamic circuits can outperform their unitary counterparts. In the latter, by tallying instructions of compiled quantum circuits, we provide an error budget detailing the obstacles that must be addressed to unlock the full potential of dynamic circuits. Looking forward, we expect dynamic circuits to be useful for generating long- range entanglement in the near term on large-scale quantum devices. ## I Introduction Quantum systems present two distinct modes of evolution: deterministic unitary evolution, and stochastic evolution as the consequence of quantum measurements. To date, quantum computations predominantly utilize unitary evolution to generate complex quantum states for information processing and simulation. However, due to inevitable errors in current quantum devices [1], the computational reach of this approach is constrained by the depth of the quantum circuits that can realistically be implemented on noisy devices. The introduction of non-unitary dynamic circuits, or adaptive circuits, may be able to overcome some of these limitations by employing mid-circuit measurements and feed-forward operations. Such conditional operations are a necessary ingredient for quantum error correction (see, e.g., Ref. [2]). In the near term, dynamic circuits present a promising avenue for generating long-range entanglement, a task at the heart of quantum algorithms. This includes both implementation of long-range entangling gates that, due to local connectivity among the qubits in many quantum platforms, can require deep unitary quantum circuits, and preparation of many-qubit entangled [3, 4] and topologically ordered quantum states [5, 6, 7, 8, 9, 10, 11, 12, 13]. From a physical standpoint, the entanglement needs to propagate across the entire range between the qubits. Given that the entanglement cannot spread faster than its information light cone [14, 15], entangling two qubits that are a distance $n$ apart requires a minimum two-qubit gate-depth that scales as $\mathcal{O}(n)$, and even when assuming all-to-all connectivity, the generation of entanglement over $n$ qubits necessitates a minimum two-qubit gate depth of $\mathcal{O}(\log n)$. Thus, the task becomes challenging when applying only unitary gates. Using dynamic circuits, the spread of information can be mostly conducted by classical calculations, which can be faster and with a higher fidelity than the quantum gates, and long-range entanglement can be created in a shallow quantum circuit [16, 17, 18], i.e. the depth of quantum gates is constant for any $n$. While dynamic circuits have been explored in small-scale experiments [19, 20, 21, 22, 23], only recently have there been experimental capabilities on large- scale quantum devices. However, most demonstrations (with the exception of e.g. Refs. [24, 25, 26]) have utilized post-selection [27] or post-processing [28, 29] instead of feed-forward to prepare entangled states. Such approaches enable the study of properties of the state prepared in isolation, but have limited applicability when the state preparation is part of a larger quantum information processing task. Here, we explore the utility of shallow dynamic circuits for creating long- range entanglement on large-scale superconducting quantum devices. In Section II, we demonstrate an advantage with dynamic circuits by teleporting a long- range entangling CNOT gate over up to 101 locally connected superconducting qubits. We also discuss how this approach can be generalized to more complex gates, such as the three-qubit Toffoli gate. Then, in Section III, we prepare a long-range entangled state, the GHZ state [3], with a dynamic circuit. We show that—with a composite error mitigation stack customized for the hardware implementation of dynamic circuits—we can prepare genuinely entangled GHZ states but fall short of state-of-the-art system sizes achieved with unitary gates due to hardware limitations. We predict conditions under which dynamic circuits should be advantageous over unitary circuits based on our error budget calculation. Figure 1: Teleporting a CNOT gate for long-range entanglement. (a) Left: Circuit for a long-range CNOT gate spanning a 1D chain of $n$-qubits subject to nearest-neighbor connections only. Middle: Equivalent unitary decomposition into implementable CNOT gates; circuit depth $\mathcal{O}(n)$. Right: Equivalent circuit employing measurements with feed-forward operations; circuit depth $\mathcal{O}(1)$. If the post-measurement state is unused, feed- forward operations can be handled in post-processing, eliminating the need for their experimental implementation. Yellow regions indicate the idle time during CNOT gates on other qubits as well as during measurement and feed- forward (which is denoted by duration $\mu$). (b) Error model inputs for unitary, measurement-based, and dynamic-circuit CNOT protocols comprise the total number of: non-zero idle-block times, CNOT gates, and additional measurements. (c) Experimental results, where dynamic circuits offer improved fidelity for CNOT gate teleportation across a qubit chain $\gtrsim$ 10 qubits. (d) Map of a 127-qubit heavy-hexagonal processor, ibm_sherbrooke, overlaid with system configurations for long-range gate teleportation across a locally connected bus. To establish an effective all-to-all connectivity, we show one possible strategy of dividing the qubits into system (purple and orange) and sacrificial ancilla (turquoise and blue for extra connections) qubits. To parallellize gate execution with increased connectivity, orange qubits can be used as ancillas. We show how a particular long-range CNOT can be implemented through an ancilla bus marked as turquoise spins. ## II Gate teleportation The limited connectivity between qubits in many quantum computational platforms can result in the compilation of non-local unitary circuits into deep and error-prone unitary circuits. A potential solution is the use of shallow dynamic circuits. The crucial ingredient for such protocols is long- range CNOT gates from the first to $n$th qubit, as shown on the left in Fig. 1(a). In the following, we demonstrate a regime under which dynamic circuits enable higher-fidelity long-range CNOT gates via gate teleportation. We first describe the dynamic circuit and compare to its equivalent unitary counterpart. We argue, using a simple error budget, that there exists a regime in which the dynamic circuit implementation has an advantage over the unitary one, see Fig. 1(b). Then, using up to $101$ qubits on a superconducting processor, we demonstrate a crossover in the fidelity of CNOT gate teleportation, where dynamic circuits perform better for entangling qubits over longer ranges; see Fig. 1(c). This gate teleportation scheme enables an effective all-to-all connectivity in devices with a more limited connectivity, such as those on a heavy-hexagonal lattice. By using some of the qubits as ancillas for measurement and classical feed-forward operations, the ancilla qubits form a bus that connects all system qubits with each other. Therefore, by sacrificing some of the qubits in a large device with limited connectivity, we gain effective access to an all-to-all connected device with fewer qubits; see Fig 1(d). Finally, we show that these ideas can be generalized to teleporting multi-qubit gates, such as the Toffoli or CCZ gate. ### II.1 CNOT We describe the dynamic circuit for CNOT gate teleportation, shown on the right in Fig. 1(a) and derived in Appendix A.1. Importantly, this dynamic circuit can be straightforwardly extended for any number of qubits $n$ (where $n$ is the number of ancillas) such that the depth remains constant for any initial states $\ket{\varphi_{1}}$ $\left(\ket{\varphi_{2}}\right)$ of the control (target) qubit. We expect the error to be dominated by the $n$ mid- circuit measurements, $n+1$ CNOT gates parallelized over 2 gate layers, and idle time mostly over the classical feed-forward time. Note that in this particular realization, each of the $n$ ancilla qubits between the two system qubits must be in the state $\ket{0}$. Therefore, during the course of the gate teleportation, the ancillas cannot also be used as memory qubits, further motivating the division of qubits into system and sacrificial ancilla qubits in Fig. 1(d). We also present an equivalent, low-error unitary counterpart in the middle of Fig. 1(a). (In Appendix B, we propose several different unitary implementations of the long-range CNOT gate. Based on experimental results, as well as the noise model described in Appendix E that gives rise to the error budget described in Appendix B.2, we select this one.) In this unitary realization, the system qubits are connected by a bus of ancilla qubits that are initialized in and returned to the $|0\rangle$ state, just as in its dynamic counterpart. In our particular compilation, throughout the execution of the circuit, qubits that are not in the $|\phi_{1}\rangle$ or $|\phi_{2}\rangle$ state are in the $|0\rangle$ state. Doing so minimizes both decoherence and cross-talk errors intrinsic to our superconducting qubit design. Therefore, relative to the dynamic version, there is no error due to idle time or mid-circuit measurements, although there are $\sim$4 times more CNOT gates. A summary of the error budgets for the dynamic and unitary circuits is in Fig. 1(b). Based on this table, we expect that dynamic circuits should be advantageous over unitary circuits if the additional $n$ mid-circuit measurements in the dynamic circuit introduce less error than the $3n$ extra CNOT gates in the unitary circuit, assuming $n$ is large enough such that the idling error $\mu$ incurred during measurement and classical feed-forward in the dynamic circuit is relatively small. Importantly, we should note that these error analyses only consider the gate error on the two respective qubits, but not the error introduced on other qubits, which we expect to be much larger in the unitary case due to the linear depth. Thus, the constant- depth dynamic circuit might be even more advantageous than what we can see from the gate fidelity. To determine the experimental gate fidelity, let our ideal unitary channel be $\mathcal{U}(\rho)\coloneqq U\rho U^{\dagger}$ and its noisy version be $\tilde{\mathcal{U}}(\rho)\coloneqq\mathcal{U}(\Lambda(\rho))$, where $\Lambda$ is the effective gate noise channel and $\rho$ is a quantum state. The average gate fidelity of the noisy gate is $\mathcal{F}_{\mathrm{avg}}\left(\mathcal{U},\tilde{\mathcal{U}}\right)=\int\mathrm{d}\psi\,\operatorname{Tr}\left[\mathcal{U}\left(\rho_{\psi}\right)\tilde{\mathcal{U}}\left(\rho_{\psi}\right)\right]$, where the Haar average is taken over the pure states $\rho_{\psi}=\left|\psi\vphantom{\psi}\right\rangle\left\langle\vphantom{\psi}\psi\right|$. This fidelity can be faithfully estimated from Pauli measurements on the system, using Monte Carlo process certification [30, 31], as detailed in Appendix C.2. The results from a superconducting quantum processor are shown in Fig. 1(c). As expected, for a small number of qubits $n\lesssim 10$ the unitary implementation yields the best fidelities. However, for increasing $n$ it converges much faster to the fidelity of a random gate (0.25) than the dynamic circuits implementation, which converges to a value slightly below 0.4. These align well with the error budget analysis in Appendix B.2 and the noise model predictions depicted in Appendix E. Note that, in the limit of large $n$, the fidelities of the measurement-based scheme are limited by the $Z$ and $X$ corrections on $\ket{\phi_{1}}$ and $\ket{\phi_{2}}$ (see Fig. 1(a)). A straightforward derivation using this noise model shows that the minimum possible process fidelity due to only incorrect $Z$ and $X$ corrections (without the fixed infidelity from the idle time and CNOT gates) is 0.25, which converts to a gate fidelity of 0.4. The measurement-based protocol with post-processing performs slightly better than the dynamic circuits as the former does not incur errors from the classical feed-forward, allowing us to isolate the impact of classical feed- forward from other errors, such as the $n+1$ intermediate CNOT gates and mid- circuit measurements. Note, however, that the post-processing approach is generally not scalable if further circuit operations follow the teleported CNOT due to the need to simulate large system sizes, further emphasizing the advantage of dynamic circuits as errors rooted in classical feedforward are reduced. Overall, we find that CNOT gates over large distances are more efficiently executed with dynamic circuits than unitary ones. ### II.2 Toffoli or CCZ Dynamic circuits can also be applied to more efficiently compile multi-qubit gates. As an example, we describe how the CCZ, or Toffoli gate up to two single-qubit Hadamard gates, can be implemented by optimizing multiple teleported CNOT gates. Compilation of the unitary circuit on a 1D chain of $n+3$ qubits using CNOT gates naïvely requires a two-qubit gate depth of $\mathcal{O}\left(n\right)$. Using dynamic circuits, we can implement this long-range entangling gate in shallow depth. Naïvely, one could successively implement each CNOT gate of the typical Toffoli decomposition (shown at the top of Fig. 2(a)) using the gate teleportation described previously. However, involving an ancillary qubit between the three system qubits to merge the teleported gates, as shown at the bottom of Fig. 2(a), allows for a more efficient implementation with the dynamic circuit; see Fig. 2(b). In total, this formulation requires $n+1$ measurements, $n+6$ CNOT gates, and 5 feed- forward operations divided across two sequential steps. Notably, as most qubits are projectively measured early in the circuit, the idling error should be low. Thus, we expect this shallow implementation with dynamic circuits to be advantageous over its unitary counterpart, especially for large $n$. Figure 2: CCZ with (a) unitary circuit and (b) a dynamic circuit over long ranges. ## III State preparation: GHZ Figure 3: Preparing long-range entangled states. (a) Illustration of a GHZ state with chosen qubit spins (spheres) in a superposition of “all up” and “all down” polarizations (arrows), overlaid on a quantum processor. (b) Circuits to prepare an $n$-qubit GHZ state using either a unitary (left) or dynamic (right) circuit. For a 1D qubit chain, the depth of the unitary (resp., dynamic) circuit scales as $\mathcal{O}(n)$ (resp., $\mathcal{O}(1)$). If the final state is not directly used, the feed-forward operations can be implemented in classical post-processing on the output bits (classically controlled X gates and resets can be omitted). Yellow regions indicate the idle time during CNOT gates on other qubits as well as during measurement and feed-forward (which is denoted by duration $\mu$). (c) Error model inputs for the GHZ preparation circuits. The model incorporates the noisy components of the circuits: non-zero idle circuit periods (yellow), number of CNOT gates (pink), and the number of mid-circuit measurements (green). These parameters are used to derive an error model that yields a lower-bound on the protocol fidelity, shown in the following panel. (d) Fidelity of preparing the GHZ state on quantum hardware using unitary, measurement-based post-processing, or dynamic circuits in the absence or presence of dynamical decoupling (DD). Data shown with dots. Theory curves based on the error model parameters of panel (c) shown in dashed lines. Dynamic circuits can also be used to prepare long-range entangled states. A prototypical example is the GHZ state [3], shown schematically in Fig. 3(a). While it can be created using only Clifford gates and thus can be simulated efficiently on a classical computer [32], it becomes non-simulatable when followed by a sufficient number of non-Clifford gates in a larger algorithm, or when inserted as a crucial ingredient in e.g. efficient compilation of multi-qubit gates [33, 34]. Here, we show that GHZ states with long-range entanglement can be prepared with dynamic circuits. Although we do not see a clear advantage of dynamic circuits over unitary ones in this case, we provide a detailed description of the challenges that must be addressed to realize such an advantage. For preparation of a GHZ state on a 1D $n$-qubit chain, in Fig. 3, we show the equivalence between the unitary circuit (left) and dynamic circuit (right). (For a detailed derivation, see Appendix A.2.) Notably, the unitary equivalent has a two-qubit gate depth that scales as $\mathcal{O}\left(n\right)$ with quadratically increasing idle time and $n-1$ total CNOT gates, while the depth of the dynamic circuits remains constant with linearly increasing idle time, $3n/2-1$ total CNOT gates, and $n/2-1$ mid-circuit measurements (see Fig. 3(c)). The dynamic circuit incurs less idle time and fewer two-qubit gate depth at the cost of increased CNOT gates and mid-circuit measurements. Therefore, we expect dynamic circuits to be advantageous for large system sizes $n$ and low errors in mid-circuit measurement. For a more detailed analysis of the error budget, see Appendix D.1. We explore whether current large-scale superconducting quantum devices enable an advantage with dynamic circuits for preparation of the entangled GHZ state. To efficiently verify the preparation of a quantum state $\sigma$, we use the Monte Carlo state certification that samples from Pauli operators with non- zero expectation values, as implemented in Ref. [27] and described in detail in Appendix C.1. As the $n$-qubit GHZ state is a stabilizer state, we can randomly sample $m$ of the $2^{n}$ stabilizers $\\{S_{i}\\}_{i=1..2^{n}}$ and approximate the fidelity by $F=\frac{1}{m}\sum_{k=1}^{m}\langle S_{k}\rangle_{\sigma}+\mathcal{O}\left(\frac{1}{\sqrt{m}}\right)$. The experimental results of GHZ state preparation with unitary and dynamic circuits are shown in Fig. 3(d). They all include measurement error mitigation on the final measurements [35]. On the left, we show the results without dynamical decoupling. In the unitary case, we observe genuine multipartite entanglement, defined as state fidelity $F>0.5$ [36], within a confidence interval of $95\%$ up to 7 qubits with a rapid decay in fidelity with increasing system size due to coherent errors in two-qubit gates and ZZ crosstalk errors during idling time [37]. In the dynamic case, we observe genuine entanglement up to 6 qubits. Here, we do not find a crossover point after which dynamic circuits have an advantage over unitary circuits. We attribute the performance of dynamic circuits to several factors, including the fact that the current implementation results in an average classical feedforward time that scales with the number of potential mid-circuit measurement bitstring outcomes, which itself grows exponentially with system size. By reducing the error induced by idle time during classical feedforward, we expect dynamic circuits to surpass unitary circuits at $\gtrsim$10 qubits—we can see this by studying the post-processing case, which is equivalent to the dynamic circuit implementation except that the classical logic is executed in post-processing, not during execution of the quantum circuit itself. We expect the exponential scaling of classical feedforward time to be reduced to linear or constant scaling in the near term. On the right of Fig. 3(d), we show the results using dynamical decoupling (DD) [38, 39]. We observe improved fidelities for both the unitary and dynamic circuit cases, but not for the post-processing case as there is little error induced by idle times to quench with dynamical decoupling in the first place. For the unitary case, we observe genuine multipartite entanglement up to 17 qubits, more than twice as many compared to the unmitigated unitary case. This result is close to the state of the art on superconducting quantum processors and is limited by the fact that we do not leverage the 2D connectivity of the device, as in Ref. [40]. While the fidelities are improved with DD for dynamic circuits, the improvement is less dramatic. We attribute this difference to two reasons: First, the unitary circuit has a quadratic idling error term in contrast to a leading linear term for dynamic circuits, resulting in comparatively smaller improvement for dynamic circuits with dynamical decoupling. Second, with the current controls, we are not able to apply DD pulses during the classical feedforward time, which is the main source of idling error in the dynamic circuit. As in the unmitigated case, we observe rapid decay of the fidelity with increasing system. This can again be partially attributed to exponential growth of the classical feedforward time. In the near future, we expect to reduce this scaling to linear, in which case we expect drastically improved performance and genuine entanglement up to $\sim$15 qubits. Still, however, we do not expect to observe an advantage with dynamic circuits for preparation of GHZ states over unitary ones. To realize an advantage with dynamic circuits, we require a scenario where the quadratically scaling idle error of the unitary circuit dominates over sufficiently small CNOT and mid-circuit measurement error; see Appendix D.2 for a more detailed analysis. We anticipate these conditions can be realized through a combination of hardware improvements and the extension of error mitigation techniques, such as probabilistic error cancellation [41, 42], toward mid-circuit measurements. ## IV Conclusion and Outlook Dynamic circuits are a promising feature toward overcoming connectivity limitations of large-scale noisy quantum hardware. Here, we demonstate their potential for efficiently generating long-range entanglement with two useful tasks: teleporting entangling gates over long ranges to enable effective all- to-all connectivity, and state preparation with the GHZ state as an example. For CNOT gate teleportation, we show a regime in which dynamic circuits result in higher fidelities on up to 101 qubits of a large-scale superconducting quantum processor. We leave incorporating this more efficient implementation of long-range entangling gates as a subroutine in another quantum algorithm to future work; potential studies can include simulating many-body systems with non-local interactions. As we demonstrate theoretically, gate teleportation schemes can be extended beyond CNOT gates to multi-qubit ones, such as the CCZ gate. Its experimental implementation is also a promising project for the future. For state preparation, based on both unmitigated and mitigated hardware experiments, we expect to see the value of dynamic circuits once the classical post-processing becomes more efficient and the mid-circuit measurement errors can be reduced. We plan on revising the experiments as soon these capabilities are available. We anticipate that further experiments with dynamic circuits and the development of noise models describing them will be vital contributions toward efficient circuit compilation, measurement-based quantum computation, and fault-tolerant quantum computation. ## V Acknowledgements We thank Diego Ristè, Daniel Egger, and Alexander Ivrii for valuable discussions and feedback. We thank Emily Pritchett, Maika Takita, Abhinav Kandala, and Sarah Sheldon for their help. We also thank Thomas Alexander, Marius Hillenbrand, and Reza Jokar for their support with implementing dynamic circuits. ## References * Preskill [2018] J. Preskill, Quantum computing in the NISQ era and beyond, Quantum 2, 79 (2018). * Terhal [2015] B. M. Terhal, Quantum error correction for quantum memories, Rev. Mod. Phys. 87, 307 (2015). * Greenberger _et al._ [1989] D. M. Greenberger, M. A. Horne, and A. Zeilinger, Going beyond bell’s theorem, in _Bell’s Theorem, Quantum Theory and Conceptions of the Universe_, edited by M. Kafatos (Springer Netherlands, Dordrecht, 1989) pp. 69–72. * Raussendorf _et al._ [2005] R. Raussendorf, S. Bravyi, and J. Harrington, Long-range quantum entanglement in noisy cluster states, Phys. Rev. A 71, 062313 (2005). * Dennis _et al._ [2002] E. Dennis, A. Kitaev, A. Landahl, and J. Preskill, Topological quantum memory, J. Math. Phys. 43, 4452 (2002). * Piroli _et al._ [2021] L. Piroli, G. Styliaris, and J. I. Cirac, Quantum circuits assisted by local operations and classical communication: Transformations and phases of matter, Phys. Rev. Lett. 127, 220503 (2021). * Lu _et al._ [2022] T.-C. Lu, L. A. Lessa, I. H. Kim, and T. H. Hsieh, Measurement as a shortcut to long-range entangled quantum matter, PRX Quantum 3, 040337 (2022). * Verresen _et al._ [2021] R. Verresen, N. Tantivasadakarn, and A. Vishwanath, Efficiently preparing schrödinger’s cat, fractons and non-abelian topological order in quantum devices,(2021), arXiv:2112.03061 (2021). * Tantivasadakarn _et al._ [2021] N. Tantivasadakarn, R. Thorngren, A. Vishwanath, and R. Verresen, Long-range entanglement from measuring symmetry-protected topological phases, arXiv:2112.01519 (2021). * Tantivasadakarn _et al._ [2022] N. Tantivasadakarn, R. Verresen, and A. Vishwanath, The shortest route to non-abelian topological order on a quantum processor, arXiv:2209.03964 (2022). * Tantivasadakarn _et al._ [2023] N. Tantivasadakarn, A. Vishwanath, and R. Verresen, Hierarchy of topological order from finite-depth unitaries, measurement, and feedforward, PRX Quantum 4, 020339 (2023). * Bravyi _et al._ [2022] S. Bravyi, I. Kim, A. Kliesch, and R. Koenig, Adaptive constant-depth circuits for manipulating non-abelian anyons, arXiv:2205.01933 (2022). * Lu _et al._ [2023] T.-C. Lu, Z. Zhang, S. Vijay, and T. H. Hsieh, Mixed-state long-range order and criticality from measurement and feedback, PRX Quantum 4, 030318 (2023). * Lieb and Robinson [1972] E. H. Lieb and D. W. Robinson, The finite group velocity of quantum spin systems, Commun. Math. Phys. 28, 251 (1972). * Bravyi _et al._ [2006] S. Bravyi, M. B. Hastings, and F. Verstraete, Lieb-robinson bounds and the generation of correlations and topological quantum order, Phys. Rev. Lett. 97, 050401 (2006). * Raussendorf _et al._ [2003] R. Raussendorf, D. E. Browne, and H. J. Briegel, Measurement-based quantum computation on cluster states, Phys. Rev. A 68, 022312 (2003). * Jozsa [2005] R. Jozsa, An introduction to measurement based quantum computation, arXiv:0508124 (2005). * Beverland _et al._ [2022] M. Beverland, V. Kliuchnikov, and E. Schoute, Surface code compilation via edge-disjoint paths, PRX Quantum 3, 10.1103/prxquantum.3.020342 (2022). * Barrett _et al._ [2004] M. D. Barrett, J. Chiaverini, T. Schaetz, J. Britton, W. M. Itano, J. D. Jost, E. Knill, C. Langer, D. Leibfried, R. Ozeri, and D. J. Wineland, Deterministic quantum teleportation of atomic qubits, Nature 429, 737 (2004). * Pfaff _et al._ [2012] W. Pfaff, T. H. Taminiau, L. Robledo, H. Bernien, M. Markham, D. J. Twitchen, and R. Hanson, Demonstration of entanglement-by-measurement of solid-state qubits, Nat. Phys. 9, 29 (2012). * Ristè _et al._ [2013] D. Ristè, M. Dukalski, C. A. Watson, G. de Lange, M. J. Tiggelman, Y. M. Blanter, K. W. Lehnert, R. N. Schouten, and L. DiCarlo, Deterministic entanglement of superconducting qubits by parity measurement and feedback, Nature 502, 350 (2013). * Wan _et al._ [2019] Y. Wan, D. Kienzler, S. D. Erickson, K. H. Mayer, T. R. Tan, J. J. Wu, H. M. Vasconcelos, S. Glancy, E. Knill, D. J. Wineland, A. C. Wilson, and D. Leibfried, Quantum gate teleportation between separated qubits in a trapped-ion processor, Science 364, 875 (2019). * Córcoles _et al._ [2021] A. D. Córcoles, M. Takita, K. Inoue, S. Lekuch, Z. K. Minev, J. M. Chow, and J. M. Gambetta, Exploiting dynamic quantum circuits in a quantum algorithm with superconducting qubits, Phys. Rev. Lett. 127, 100501 (2021). * Foss-Feig _et al._ [2023] M. Foss-Feig, A. Tikku, T.-C. Lu, K. Mayer, M. Iqbal, T. M. Gatterman, J. A. Gerber, K. Gilmore, D. Gresh, A. Hankin, N. Hewitt, C. V. Horst, M. Matheny, T. Mengle, B. Neyenhuis, H. Dreyer, D. Hayes, T. H. Hsieh, and I. H. Kim, Experimental demonstration of the advantage of adaptive quantum circuits, arXiv:2302.03029 (2023). * Iqbal _et al._ [2023] M. Iqbal, N. Tantivasadakarn, T. M. Gatterman, J. A. Gerber, K. Gilmore, D. Gresh, A. Hankin, N. Hewitt, C. V. Horst, M. Matheny, T. Mengle, B. Neyenhuis, A. Vishwanath, M. Foss-Feig, R. Verresen, and H. Dreyer, Topological order from measurements and feed-forward on a trapped ion quantum computer, arXiv:2302.01917 (2023). * Moses _et al._ [2023] S. Moses, C. Baldwin, M. Allman, R. Ancona, L. Ascarrunz, C. Barnes, J. Bartolotta, B. Bjork, P. Blanchard, M. Bohn, _et al._ , A race track trapped-ion quantum processor, arXiv:2305.03828 (2023). * Cao _et al._ [2023] S. Cao, B. Wu, F. Chen, M. Gong, Y. Wu, Y. Ye, C. Zha, H. Qian, C. Ying, S. Guo, Q. Zhu, H.-L. Huang, Y. Zhao, S. Li, S. Wang, J. Yu, D. Fan, D. Wu, H. Su, H. Deng, H. Rong, Y. Li, K. Zhang, T.-H. Chung, F. Liang, J. Lin, Y. Xu, L. Sun, C. Guo, N. Li, Y.-H. Huo, C.-Z. Peng, C.-Y. Lu, X. Yuan, X. Zhu, and J.-W. Pan, Generation of genuine entanglement up to 51 superconducting qubits, Nature 619, 738 (2023). * Smith _et al._ [2023] K. C. Smith, E. Crane, N. Wiebe, and S. Girvin, Deterministic constant-depth preparation of the aklt state on a quantum processor using fusion measurements, PRX Quantum 4, 020315 (2023). * Chen _et al._ [2023] E. H. Chen, G.-Y. Zhu, R. Verresen, A. Seif, E. Bäumer, D. Layden, N. Tantivasadakarn, G. Zhu, S. Sheldon, A. Vishwanath, S. Trebst, and A. Kandala, Realizing the nishimori transition across the error threshold for constant-depth quantum circuits, _to appear_ (2023). * Flammia and Liu [2011] S. T. Flammia and Y.-K. Liu, Direct fidelity estimation from few pauli measurements, Phys. Rev. Lett. 106, 230501 (2011). * da Silva _et al._ [2011] M. P. da Silva, O. Landon-Cardinal, and D. Poulin, Practical characterization of quantum devices without tomography, Phys. Rev. Lett. 107, 210404 (2011). * Gottesman [1998] D. Gottesman, The heisenberg representation of quantum computers, arXiv:9807006 (1998). * Hoyer and Spalek [2005] P. Hoyer and R. Spalek, Quantum fan-out is powerful, Theory Comput. 1, 81 (2005). * Yang and Rall [2023] W. Yang and P. Rall, Harnessing the power of long-range entanglement for clifford circuit synthesis, arXiv:2302.06537 (2023). * Nation and Treinish [2023] P. D. Nation and M. Treinish, Suppressing quantum circuit errors due to system variability, PRX Quantum 4, 010327 (2023). * Leibfried _et al._ [2005] D. Leibfried, E. Knill, S. Seidelin, J. Britton, R. B. Blakestad, J. Chiaverini, D. B. Hume, W. M. Itano, J. D. Jost, C. Langer, R. Ozeri, R. Reichle, and D. J. Wineland, Creation of a six-atom ‘schrödinger cat’ state, Nature 438, 639 (2005). * Tripathi _et al._ [2022] V. Tripathi, H. Chen, M. Khezri, K.-W. Yip, E. Levenson-Falk, and D. A. Lidar, Suppression of crosstalk in superconducting qubits using dynamical decoupling, Phys. Rev. Appl. 18, 024068 (2022). * Viola _et al._ [1999] L. Viola, E. Knill, and S. Lloyd, Dynamical decoupling of open quantum systems, Phys. Rev. Lett. 82, 2417 (1999). * Jurcevic _et al._ [2021] P. Jurcevic, A. Javadi-Abhari, L. S. Bishop, I. Lauer, D. F. Bogorin, M. Brink, L. Capelluto, O. Günlük, T. Itoko, N. Kanazawa, A. Kandala, G. A. Keefe, K. Krsulich, W. Landers, E. P. Lewandowski, D. T. McClure, G. Nannicini, A. Narasgond, H. M. Nayfeh, E. Pritchett, M. B. Rothwell, S. Srinivasan, N. Sundaresan, C. Wang, K. X. Wei, C. J. Wood, J.-B. Yau, E. J. Zhang, O. E. Dial, J. M. Chow, and J. M. Gambetta, Demonstration of quantum volume 64 on a superconducting quantum computing system, Quantum Sci. Technol. 6, 025020 (2021). * Mooney _et al._ [2021] G. J. Mooney, G. A. L. White, C. D. Hill, and L. C. L. Hollenberg, Generation and verification of 27-qubit greenberger-horne-zeilinger states in a superconducting quantum computer, J. Phys. Commun. 5, 095004 (2021). * Temme _et al._ [2017] K. Temme, S. Bravyi, and J. M. Gambetta, Error mitigation for short-depth quantum circuits, Phys. Rev. Lett. 119, 180509 (2017). * van den Berg _et al._ [2023] E. van den Berg, Z. K. Minev, A. Kandala, and K. Temme, Probabilistic error cancellation with sparse pauli–lindblad models on noisy quantum processors, Nat. Phys. 19, 1116 (2023). * Jamiołkowski [1972] A. Jamiołkowski, Linear transformations which preserve trace and positive semidefiniteness of operators, Rep. Math. Phys. 3, 275 (1972). * Horodecki _et al._ [1999] M. Horodecki, P. Horodecki, and R. Horodecki, General teleportation channel, singlet fraction, and quasidistillation, Phys. Rev. A 60, 1888 (1999). ## Appendix A Circuit Derivations Figure 4: Useful circuit identities that are used in the illustrative derivation of the CNOT gate teleportation and GHZ state preparation. In the following we show the circuit equivalences of the CNOT gate teleportation (Fig. 1(a)) and the GHZ state preparation (Fig. 3(b)). We are not claiming any novelty with this “proof”, but just wanted to show the reader how to derive them in an illustrative way. Before, let us start with some features that we will be using: * • The Bell state $\frac{1}{\sqrt{2}}(\ket{00}+\ket{11})$ can be illustrated as a so-called “cup”, as shown in Fig. 4(a), We can move gates along wires including along the cup, as in Fig. 4(b). * • Principle of deferred measurement: a controlled gate followed by a measurement of the controlled qubit results in the same as first performing the measurement and then applying a classically-controlled gate as in Fig. 4(c). * • While CNOT gates commute when they are conditioned on the same qubit or have the same target qubit, we get an extra gate when they act on the same qubit differently as shown in Fig. 4(d). ### A.1 Long-Range CNOT Figure 5: Graphical derivation for reducing a long-range CNOT gate into gate teleportation executed with measurements and feed-forward operations, i.e., a dynamic circuit. Roman numerals indicate sequential step numbers described in main text. In Fig. 5 we illustrate a derivation of the CNOT gate teleportation, as exemplified for $7$ qubits, which can be straightforwardly extended to an arbitrary number of qubits. In the following, we provide explanations for each step of the derivation, labeled by roman numerals in the figure: 1. (i) In the first step, we observe that entangling, measuring, and resetting the ancilla qubits does not affect the circuit. 2. (ii) We insert CNOT gates that would cancel each other. From now on we omit to write down the reset of the ancilla qubits following the measurement. 3. (iii) We move the pink CNOT gates along the Bell states to the respective qubits above. Also, we add Hadamard gates to flip the direction of the orange CNOT gates (except for the one at the bottom). Note that we can omit the hadamard gates right before the measurements, as they are not affecting the other qubits anymore. 4. (iv) By moving the bottom orange CNOT “up” along the Bell state and passing a pink CNOT, we get the extra purple CNOT gate. 5. (v) Moving the new purple CNOT “up” along the Bell state, an extra gate appears that cancels with the initial long-range CNOT gate when pushed to the left (and then it is controlled on state $\ket{0}$, so can be omitted as well). 6. (vi) Now we make use of the principle of deferred measurement. 7. (vii) In a final step we merge the classically-conditioned gates. The orange $\oplus$ correspond to XOR gates, i.e. addition mod 2. We also represented the initial Bell states again with their circuit representation. ### A.2 GHZ state preparation Figure 6: Graphical derivation for the preparation of a GHZ state by converting its canonical but deep unitary circuit into a constant-depth circuit utilizing measurement and feed-forward operations—a dynamic circuit. Roman numerals indicate sequential step numbers described in main text. In Fig. 6 we have illustrated a derivation of the GHZ state preparation, exemplary for $7$ qubits, but it can be straightforwardly extended to an arbitrary number of qubits. In the following, we provide explanations for each step of the derivation, labeled by roman numerals in the figure: 1. (i) Pushing every second CNOT gate to the very right introduces the extra pink CNOT gates. 2. (ii) We can omit CNOT gates that are conditioned on state $\ket{0}$. 3. (iii) As every second qubit is only involved at the very end, we can use those before and reset them. 4. (iv) A Bell state followed by a CNOT gate results in two uncorrelated qubits in states $\ket{+}$ and $\ket{0}$. 5. (v) We move the pink CNOT gates along the Bell states to the respective qubits above (they commute with the other CNOTs they are “passing”). 6. (vi) Pushing the pink CNOT gates to the left through the purple CNOT gates introduces the extra orange CNOT gates. 7. (vii) We make use of the principle of deferred measurement. 8. (viii) In a final step we merge the classically-conditioned gates. As the classical calculation can be done extremely fast compared to quantum gates, we draw it as a vertical line. The orange $\oplus$ correspond to XOR gates, i.e. addition mod 2. We also represented the initial Bell states again with their circuit representation. ## Appendix B CNOT circuits ### B.1 Unitary variants In order to compare the dynamic circuits implementation to a solely unitary one, let us first consider different unitary strategies that might be more or less powerful in different regimes: Strategy I: Ancilla-based implementation We can consider a similar setting as for dynamic circuits, where we place the system qubits in a way that they are connected by a bus of empty ancilla qubits. In this case, we need to swap the system qubits towards each other and back, so that the ancillas are empty in the end again. The swaps can be simplified since the ancillas are empty in the beginning. Here we can divide into different scenarios: * • Circuit Ia: To minimize the number of CNOT gates, we could swap the controlled qubit all the way to the target qubit and back, which results in the circuit depicted in Fig. 7. Here, a lot of gates cancel, so given $n$ ancilla qubits, the number of CNOT gates is $2n+1$. However, here the idle time of the qubits while they are not in state $\ket{0}$ equals $n^{2}+2n$ times the CNOT gate time. * • Circuit Ib: In order to decrease the idle time, we could essentially swap both, the controlled qubit and the target qubit half-way and back as illustrated in Fig. 7. In that case, less gates “cancel”, so for $n$ ancilla qubits we get $3n+1$ CNOT gates, but the idle time reduces to $\frac{n^{2}}{4}+n$ times the CNOT gate time. * • Circuit Ic: If we wanted to reduce the idle time even further, it might be beneficial to not cancel the CNOT gates in scenario 1b), but keep them to bring the swapped qubits back to state $\ket{0}$ as shown in Fig. 7. In that case, we have essentially no idle time (as qubits in state $\ket{0}$ are not prone to idling errors). Here, the number of CNOT gates increased to $4n+1$ though. Strategy II: SWAP-based implementation without ancillas This is the case that happens if we just feed our circuit to the transpiler. Here we do not use any ancilla qubits, but only system qubits and apply swaps to move them around. The qubits can be at a different location in the end, so we do not need to swap back. The corresponding circuit is shown in Fig. 7. In this case we require $3\tilde{n}+1$ CNOT gates and the idle time is $\frac{3}{2}\tilde{n}^{2}-2\tilde{n}$ times the CNOT gate time. However, it is important to note here that the number of qubits lying between the two qubits of interest $\tilde{n}$ is on average much shorter than the number of ancillas between two system qubits in the first scenario. Considering the connectivity illustrated in Fig. 1 (c), the relation is approximately $n=2\tilde{n}+3$. Figure 7: Comparison of the different unitary implementations of a long-range CNOT gate. While the circuits in panels (Ia), (Ib), and (Ic) realize ancilla- based implementations, the circuit of panel (II) realizes a SWAP-based implementation without ancillas. The shaded regions indicate idle periods that accumulate errors. ### B.2 Error budget Let us now compare the regimes in which we expect the different implementations to be most useful to demonstrate the benefit of dynamic circuits. In Appendix E we derive a simple noise model that allows us to compute the combined effect of different sources of decoherence as a single Pauli noise rate: $\displaystyle\lambda_{\mathrm{tot}}=t_{\mathrm{idle}}\lambda_{\mathrm{idle}}+N_{\mathrm{CNOT}}\lambda_{\mathrm{CNOT}}+N_{\mathrm{meas}}\lambda_{\mathrm{meas}}\;.$ (1) In Lemma 1 we show that the final process fidelity is loosely lower-bounded by $e^{-\lambda_{\mathrm{tot}}}$. The quantity $\lambda_{\mathrm{tot}}$ combines the following noise sources: * • The total amount of time $t_{\mathrm{idle}}$ that qubits spend idle within the circuit, and a conversion factor $\lambda_{\mathrm{idle}}$ that quantifies the strength of decoherence. $t_{\mathrm{idle}}$ is expressed in multiples of the CNOT gate time (i.e. $t_{\mathrm{idle}}=3$ for 3 CNOT gate times). The time for a mid-circuit measurement, including the additional time waiting for feedback, is defined as $\mu$ times the time for a CNOT gate. * • The total number of CNOT gates $N_{\mathrm{CNOTs}}$ and an average Pauli noise rate $\lambda_{\mathrm{CNOT}}$ per CNOT. * • The total number of mid-circuit measurements $N_{\mathrm{meas}}$ and an average Pauli noise rate $\lambda_{\mathrm{meas}}$ per measurement. In Table 1, we have summarized the error budget for each of the cases. Case | | $t_{\mathrm{idle}}$ --- | $N_{\mathrm{CNOT}}$ --- | $N_{\mathrm{meas}}$ --- | Two-qubit gate depth --- Unitary Ia) | $n^{2}+2n$ | $2n+1$ | $0$ | $2n+1$ Unitary Ib) | $\frac{n^{2}}{4}+n$ | $3n+1$ | $0$ | $2n+1$ Unitary Ic) | $0$ | $4n+1$ | $0$ | $2n+1$ Unitary II) | $\frac{3}{4}\tilde{n}^{2}-\frac{3}{2}\tilde{n}$ | $3\tilde{n}+1$ | $0$ | $\frac{3}{2}\tilde{n}+1$ Unitary II) with normed $n$ | $\approx\frac{3}{16}n^{2}-\frac{15}{8}n+\frac{45}{16}$ | $\approx\frac{3}{2}n-2$ | $0$ | $\approx\frac{3}{4}n-\frac{5}{4}$ Dynamic circuits | $2\mu+2$ | $n+1$ | $n$ | $2+\mu$, or $O(1)$ Table 1: Comparison of the error budget of the unitary and the dynamic circuits implementation in terms of idle time, number of CNOT gates and mid- circuit measurements and two-qubit gate depth. Note, that as the number of involved qubits $\tilde{n}$ needed for the unitary implementation II) is in general much smaller, we rescale it for the error budget with the relation $n\approx 2\tilde{n}+3$ Comparing the different unitary cases it becomes clear that for large $n$ the unitary implementation Ic) will be the best, as all other implementations have an error in the idling time that scales quadratically. This might be slightly counter-intuitive, as it tells us that even without measurement and feed- forward, it can be still beneficial to use ancilla qubits and thereby increase the distances. For small $n$ we need to keep in mind that for the swap-based implementation (unitary II) the number of involved qubits $\tilde{n}$ is smaller than the number of qubits $n$ needed for the same task in the ancilla- based implementation. Respecting the rescaled errors, unitary II would be the most promising implementation for small $n$. In addition to the CNOT errors and idling errors, for dynamic circuits we also need to consider the error from the additional measurements, as well as a constant term $\mu$ that comes from the idling error during measurement and feed-forward. Given this rough error analysis in Table 1, we can infer that for large $n$ dynamic circuits will be beneficial if the measurement of $n$ qubits introduces less error than $3n$ CNOT gates, that is, when $\lambda_{\mathrm{meas}}<3\lambda_{CNOT}$. A sketch of how the fidelities for the different cases decrease with $n$ is illutrated in Fig. 8. Note, that these error analyses only take into account the error on the involved qubits though. Considering also the fact that there are potentially a lot of other $m$ qubits “waiting” for this operation to be performed would add another idling error of $m\cdot(2n+1)$. So the fact that dynamic circuits can perform entangled gates between arbitrary qubits in constant depth instead of linear depth with only unitary operations speeds up the whole algorithm and therefore might be much more powerful than what we can see in the error on the respective qubits. Figure 8: Comparison of the process fidelities of the different unitary implementations as well as the dynamic circuits implementation considering the error budget indicated in Table 1. In this figure we use $\mu=(t_{\text{meas}}+t_{\text{feed-forward}})/t_{\text{cnot}}\approx 3.65$, $\lambda_{\mathrm{idle}}=0.03$, $\lambda_{\mathrm{CNOT}}=0.02$ and $\lambda_{\mathrm{meas}}=0.03$. ## Appendix C Estimation of the state and gate fidelities using Monte-Carlo sampling ### C.1 State fidelity In order to determine the fidelity of the experimentally prepared quantum state, denoted as $\sigma$, we employ the Monte Carlo state certification method, which was introduced inRefs. [31, 30]. We first briefly review the notion of fidelity between two quantum states. #### Quantum state fidelity. Let us introduce the Uhlmann-Jozsa state fidelity between two general quantum states $\rho$ and $\sigma$. These objects are elements of the space of valid density operators associated with the system Hilbert space, $\mathcal{H}$, i.e., $\rho,\sigma\in D\left(\mathcal{H}\right)$. Assuming one of them is a pure state $\sigma=\left|\phi\vphantom{\phi}\right\rangle\left\langle\vphantom{\phi}\phi\right|$, we can simplify the general expression as shown in the following: $\displaystyle F\left(\rho,\sigma\right)$ $\displaystyle\coloneqq\left[\operatorname{Tr}\left(\sqrt{\sqrt{\rho}\sigma\sqrt{\rho}}\right)\right]^{2}$ (2) $\displaystyle=\left\langle\phi\middle|\rho\middle|\phi\right\rangle$ (3) $\displaystyle=\operatorname{Tr}\left[\rho\sigma\right]\;.$ (4) If $\rho$ is also a pure state $\rho=\left|\psi\vphantom{\psi}\right\rangle\left\langle\vphantom{\psi}\psi\right|$, the expression reduces to a simple overlap $F\left(\rho,\sigma\right)=\left|\left\langle\psi\middle|\phi\right\rangle\right|^{2}$. We note that some authors define the square root of this as the fidelity. #### Pauli decomposition. To connect to experimental measurements, let us decompose the quantum sates in the standard Pauli basis. The set of all Pauli operators on $n$ qubits $\left\\{I,X,Y,Z\right\\}^{\otimes n}$ forms an orthogonal Hermitian operator basis. The inner product in operator space $L\left(\mathcal{H}\right)$ between two Pauli operators $P_{i},P_{j}\in\mathcal{L}\left(\mathcal{H}\right)$ is $\left\langle P_{i},P_{j}\right\rangle=\operatorname{Tr}\left(P_{i}P_{j}\right)=d\delta_{ij}$, where the dimension of the pure state Hilbert space $d\coloneqq\dim\mathcal{H}=2^{n}$. In terms of this basis, any quantum state $\rho\in D\left(\mathcal{H}\right)$, can be decomposed into $\rho=\sum_{i=0}^{4^{n}-1}\frac{\left\langle P_{i},\rho\right\rangle}{\left\langle P_{i},P_{i}\right\rangle}P_{i}=\frac{1}{d}\sum_{i=0}^{4^{n}-1}\rho_{i}P_{i}\;,\quad\mathrm{with}\quad\rho_{i}\coloneqq\left\langle P_{i},\rho\right\rangle=\operatorname{Tr}\left(P_{i}\rho\right)\;,$ where the Pauli expectation value of the state with respect to the $i$-th Pauli is $\rho_{i}$—an easily measurable quantity. We can similarly define the expectation values of the Pauli $P_{i}$ with respect to the prepared state $\sigma$ and the desired state $\rho$ as $\sigma_{i}:=\langle P_{i}\rangle_{\sigma}=\text{Tr}\left(\sigma P_{i}\right)$ and $\rho_{i}:=\langle P_{i}\rangle_{\rho}=\text{Tr}\left(\rho P_{i}\right)$, respectively. #### Fidelity in terms of Pauli expectation values. The state fidelity between the measured $\sigma$ and ideally expected pure $\rho$ state, see Eq. (2), in terms of the Pauli decomposition of each is $\displaystyle F(\rho,\sigma)=\text{Tr}\left[\rho\sigma\right]=\sum_{i}\frac{\rho_{i}\sigma_{i}}{d}=\sum_{i}\frac{\rho_{i}^{2}}{d}\frac{\sigma_{i}}{\rho_{i}}\;,$ (5) where $\sigma_{i}$ is an experimentally measured expectation value and $\rho_{i}$ is a theoretically calculated one. Given this, we can now define the relevance distribution $r(P_{i}):=\frac{\rho_{i}^{2}}{d}$, such that $F(\rho,\sigma)=\sum_{i:\rho_{i}\neq 0}r(P_{i})\frac{\sigma_{i}}{\rho_{i}}$. #### Random sampling of expectation values. When sampling $m$ random operators $\\{P_{k}\\}_{k=1..m}$ according to the relevance distribution $r(P_{k})$ and determining their expectation values $\sigma_{k}$, the estimated fidelity $\tilde{F}:=\sum_{k=1}^{m}\frac{\sigma_{k}}{\rho_{k}}$ approximates the actual fidelity $F$ with an uncertainty that decreases as $\frac{1}{\sqrt{m}}$. Note that there is also an uncertainty in estimating each $\sigma_{k}$, where for an additive precision $\epsilon$ roughly $(\epsilon\rho_{k})^{-2}$ shots are required. #### Random sampling of GHZ stabilizers. As the GHZ state is a stabilizer state, for each $n$ there are exactly $2^{n}$ non-zero Pauli operators $P_{i}$ that each have eigenvalue $\pm 1$. Note that some stabilizers of the GHZ state have a minus sign, e.g. $-YYX$. For the $n$-qubit GHZ state, by defining the set of stabilizers $\\{S_{i}\\}_{i=1..2^{n}}$, we can express the fidelity in terms of only expectation values on the stabilizers $\displaystyle F(\rho,\sigma)$ $\displaystyle=\frac{1}{2^{n}}\sum_{i=1}^{2^{n}}\langle S_{i}\rangle_{\sigma}\;.$ (6) This expression can be approximated by randomly sampling $m$ of the $2^{n}$ stabilizers, defining the unbiased estimator $\tilde{F}=\frac{1}{m}\sum_{k=1}^{m}\langle S_{k}\rangle_{\sigma}=F+\mathcal{O}\left(\frac{1}{\sqrt{m}}\right)$, which converges with the number of random samples chosen to the ideal fidelity. ### C.2 Average gate fidelity Similarly to the state fidelity, we use the Monte Carlo process certification following [31] to determine the average gate fidelity of our noisy CNOT gate. #### Average gate fidelity. Consider the case in which we want to implement an ideal gate $\mathcal{U}(\rho)\coloneqq U\rho U^{\dagger}$. However, instead we can implement only a noisy gate $\tilde{\mathcal{U}}(\rho)\coloneqq\mathcal{U}(\Lambda(\rho))$, where $\Lambda$ is some effective noise channel and $\rho$ is a quantum state. What is the gate fidelity of noisy $\tilde{\mathcal{U}}$ relative to the ideal $\mathcal{U}$. For a single given pure state $\rho=\left|\phi\vphantom{\phi}\right\rangle\left\langle\vphantom{\phi}\phi\right|$, the state fidelity of the output of the ideal and noisy channels is $\displaystyle F\left(\mathcal{U},\tilde{\mathcal{U}};\rho\right)$ $\displaystyle=\left[\operatorname{Tr}\left[\sqrt{\sqrt{\mathcal{U}\left(\rho\right)}\tilde{\mathcal{U}}\left(\rho\right)\sqrt{\mathcal{U}\left(\rho\right)}}\right]\right]^{2}$ (7) $\displaystyle=\operatorname{Tr}\left[\mathcal{U}\left(\rho\right)\tilde{\mathcal{U}}\left(\rho\right)\right]$ (8) $\displaystyle=\operatorname{Tr}\left[\rho\Lambda\left(\rho\right)\right]\;,$ (9) which can be used to obtain the average gate fidelity devised by a uniform Haar average over the fidelity of the ideal and noisy output states, with $\rho_{\psi}=\left|\psi\vphantom{\psi}\right\rangle\left\langle\vphantom{\psi}\psi\right|$, $\displaystyle\mathcal{F}_{\mathrm{avg}}\left(\mathcal{U},\tilde{\mathcal{U}}\right)$ $\displaystyle=\int\mathrm{d}\psi\,F\left(\mathcal{U},\tilde{\mathcal{U}};\rho_{\psi}\right)$ (10) $\displaystyle=\int\mathrm{d}\psi\,\operatorname{Tr}\left[\mathcal{U}\left(\rho\right)\tilde{\mathcal{U}}\left(\rho\right)\right]$ (11) $\displaystyle=\operatorname{Tr}\left[\int\mathrm{d}\psi\,\left|\psi\vphantom{\psi}\right\rangle\left\langle\vphantom{\psi}\psi\right|\Lambda\left(\left|\psi\vphantom{\psi}\right\rangle\left\langle\vphantom{\psi}\psi\right|\right)\right]\;.$ (12) To estimate $\mathcal{F}_{\mathrm{avg}}\left(\mathcal{U},\tilde{\mathcal{U}}\right)$, we will use the process (or entanglement) fidelity as a more experimentally- accessible quantity. #### Process fidelity. Compared to the gate fidelity, the process fidelity is more readily estimated. It can in turn serve as a direct proxy to the gate fidelity. To make the connection, recall that the Choi-Jamiolkowski isomorphism [43] maps every quantum operation $\Lambda$ on a $d$-dimensional space to a density operator $\rho_{\Lambda}=(\mathbb{I}\otimes\Lambda)\ket{\phi}\bra{\phi}$, where $\ket{\phi}=\frac{1}{\sqrt{d}}\sum_{i=1}^{d}\ket{i}\otimes\ket{i}$. For a noise-free, ideal unitary channel $\mathcal{U}$ and its experimental, noisy implementation $\tilde{\mathcal{U}}$, the process fidelity $\mathcal{F}_{\mathrm{proc}}$ is the state fidelity of the respective Choi states $\rho_{\mathcal{U}}$ and $\rho_{\tilde{\mathcal{U}}}$: $\displaystyle\mathcal{F}_{\mathrm{proc}}({{\mathcal{U}}},{\tilde{\mathcal{U}}}):=F(\rho_{{\mathcal{U}}},\rho_{\tilde{\mathcal{U}}})\;.$ (13) From this fidelity, the gate fidelity can be extracted using the following relation derived in [44]: $\displaystyle\mathcal{F}_{\mathrm{gate}}({{\mathcal{U}}},{\tilde{\mathcal{U}}})=\frac{d\mathcal{F}_{\mathrm{proc}}(\rho_{{\mathcal{U}}},\rho_{\tilde{\mathcal{U}}})+1}{d+1}\;.$ (14) #### Estimating the process fidelity. As described in Ref. [31], instead of a direct implementation of $(\mathbb{I}\otimes\tilde{\mathcal{U}})\ket{\phi}\bra{\phi}$ followed by measuring random Pauli operators on all qubits, we follow the more practical approach, where $\tilde{\mathcal{U}}$ is applied to the complex conjugate of a random product of eigenstates of local Pauli operators $P_{i}\otimes P_{j}$, followed by a measurement of random Pauli operators $P_{k}\otimes P_{l}$. This leads to the same expectation values $\displaystyle\rho_{ijkl}:=\text{Tr}\left[(P_{i}\otimes P_{j}\otimes P_{k}\otimes P_{l})(\mathbb{I}\otimes\mathcal{U})\ket{\phi}\bra{\phi}\right]=\text{Tr}\left[(P_{k}\otimes P_{l})\mathcal{U}(P_{i}\otimes P_{j})^{\ast}\right]/d\;.$ (15) The operators are then sampled according to the relevance distribution $\displaystyle r_{ijkl}:=r(P_{i}P_{j}P_{k}P_{l})=\frac{\rho_{ijkl}^{2}}{d}\;.$ (16) For $\Lambda(\rho)=\mathrm{CNOT}\rho\mathrm{CNOT}^{\dagger}$, there are only $16$ combinations of Pauli operators with a non-zero expectation value $\rho_{ijkl}$: $\rho_{ijkl}=-\frac{1}{2}$ for $P_{i}P_{j}P_{k}P_{l}\in\ \\{YYXZ,XZYY\\}$ and $\rho_{ijkl}=\frac{1}{2}$ for the remaining 14. Thus, the relevance distribution is uniform amongst those with $r=\frac{1}{16}$ and we can just take the average expectation value of those $16$ operators. ## Appendix D Error analysis for GHZ states ### D.1 Error budget As in Appendix B.2, we leverage (1) to estimate the total noise $\lambda_{\mathrm{tot}}$ of a quantum circuit as motivated by the model discussed in Appendix E. There, it is derived that $e^{-\lambda_{\mathrm{tot}}}$ gives a lower bound on the _process_ fidelity of the circuit. For GHZ states however, we are interested the _state_ fidelity, so the bound from Lemma 1 no longer applies in a rigorous sense. However, we find that the same model can still provide useful intuition if we accept that the model parameters $\lambda_{\mathrm{CNOT}},\lambda_{\mathrm{meas}}$ no longer have a direct interpretation in terms of worst-case Pauli-Lindblad noise or a combination of amplitude- and phase-damping noise respectively. See Appendix E for details. For the unitary approach we require $n$ CNOT gates to entangle $n$ qubits. For simplicity we assume (and implement) only a one-dimensional connectivity chain in our protocols and the following numbers correspond to an even number $n$ (only constant terms change when considering odd $n$). To minimize the idling time, we start in the middle and apply CNOT gates simultaneously towards both ends. This leads to an idle time of $\frac{n^{2}}{4}-\frac{3}{2}n+2$ times the CNOT gate time, as displayed in Table 2. In the dynamic circuits approach we require $\frac{3}{2}n-2$ CNOT gates in total, while the idling time is $\frac{\mu}{2}n+1$ times the CNOT gate time, where $\mu$ corresponds to the measurement and feed-forward time (as a multiple of the CNOT gate time). However, here we also need to consider the errors of the additional $\frac{n}{2}-1$ measurements. As the error coming from the CNOT gates and the measurements is usually substantially larger than the error from the idling time, we expect that for small $n$ the standard unitary preparation succeeds. However, as the idling time there scales as $\mathcal{O}\left(n^{2}\right)$ in contrast to all errors in the measurement-based approach scaling only as $\mathcal{O}\left(n\right)$, we expect a crossover for large $n$, where the implementation with dynamic circuits will become more beneficial. The error budget is summarized in Table 2. Case | | $t_{\mathrm{idle}}$ --- | $N_{\mathrm{CNOT}}$ --- | $N_{\mathrm{meas}}$ --- | Two-qubit gate depth --- Unitary | $n^{2}/4-3n/2+2$ | $n-1$ | $0$ | $n-1$ Dynamic circuits | $1+\mu n/2$ | $3n/2-2$ | $n/2-1$ | $3+\mu$, or $O(1)$ Table 2: Comparison of the error budget of the unitary and the dynamic circuits implementation in terms of idle time, number of CNOT gates and mid- circuit measurements and two-qubit gate depth. ### D.2 Expected cross-over for lower mid-circuit measurement errors Here we use the values of $t_{\mathrm{idle}},N_{\mathrm{CNOT}},$ and $N_{\mathrm{meas}}$ shown in Table 2 to predict how many qubits are required to see and the state fidelity at the cross-over, or where the performance of dynamic circuits becomes higher than that of its unitary counterpart, as a function of the mid-circuit measurement errors. Note that in this noise model we assume that we can eliminate all ZZ errors by applying dynamical decoupling. We keep the idling error constant at $\lambda_{\mathrm{idle}}=0.001$ and consider different CNOT errors $\lambda_{\mathrm{CNOT}}\in\\{0.001,0.01,0.02\\}$. We can reach a fidelity $>0.5$ for a CNOT error of $\lambda_{\mathrm{CNOT}}=0.01$ with mid-circuit measurement errors $\lambda_{\mathrm{meas}}\lesssim 0.003$ and for a CNOT error $\lambda_{\mathrm{CNOT}}=0.001$ with mid-circuit measurement errors $\lambda_{\mathrm{meas}}\lesssim 0.012$ Figure 9: Noise-model predictions that indicate how many qubits are required to see a cross-over and what the corresponding fidelity would be as a function of the mid-circuit measurement errors. ## Appendix E Pauli-Lindblad Noise Model In this section we present a simple framework for computing lower bounds on fidelities using the Pauli-Lindblad noise model discussed in [42]. Pauli- Lindblad noise channels have several nice properties that we can use to simplify calculations, and also allow us to reduce estimates of the noise properties of our hardware to relatively few parameters. Normally, Pauli-Lindblad noise is the workhorse of probabilistic error cancellation - an error mitigation scheme that leverages characterization of noise in order to trade systematic uncertainty for statistical uncertainty. But we are more interested in using Pauli-Lindblad noise as a tool for capturing the behavior of fidelity as a function of circuit size with an appropriate balance of rigor and simplicity. As such, our central goal in this section is to develop mathematical tools that allow us develop a Pauli-Lindblad representation of various noise sources such as decoherence and gate noise and to find a method to combine all of this noise into a fidelity for the entire process. In particular, we aim to give a justification for modeling noise via the quantity $\lambda_{\mathrm{tot}}$ as in (1). This is achieved by Lemma 1, which states that $e^{-\lambda_{\mathrm{tot}}}$ gives a lower bound on the process fidelity. We leave the majority of our mathematical exposition without proof for sake of brevity, but present the proof of Lemma 1 at the end of this section. #### Pauli-Lindblad noise. Pauli-Lindblad noise is a quantum channel defined as follows. Let $\mathcal{P}$ be the $n$-qubit Pauli group modulo phase, and consider some $P\in\mathcal{P}$. Then for some noise rate $\lambda\in\mathbb{R}^{+}$ the noise channel $\Gamma^{\lambda}_{P}$ is given by: $\displaystyle\Gamma_{P}^{\lambda}(\rho)=(1-\omega)\rho+\omega P\rho P^{\dagger}\quad\mathrm{where}\quad\omega:=\frac{1-e^{-2\lambda}}{2}\;.$ (17) This is essentially applying $P$ with probability $\omega$. Pauli noise channels also have a representation as time evolution with respect to a simple Lindbladian: for $P\in\mathcal{P}$, let $\mathcal{L}_{P}(\rho):=P\rho P-\rho$. This way $\Gamma_{P}^{\lambda}=e^{\lambda\mathcal{L}_{P}}$. The main justification for why we can restrict to Pauli noise channels is twirling. Conjugating an arbitrary noise channel by a random Pauli matrix yields a channel that is always expressible as a product of Pauli noise. Although our experiments do not feature twirling, even for untwirled circuits we expect the Pauli-Lindblad noise to capture the first-order noise behavior. Another reason why we expect our noise model to only capture the behavior to first-order is that we assume the noise rates are the same for all qubits. All CNOT gates and idle times are assumed to contribute the same amount of noise. But this is not a realistic representation of our hardware - in actuality different qubits have different coherence times and gate qualities also vary. When we consider circuits on many qubits we expect these differences to average out. Let $\Lambda$ be a quantum channel. Then let $\tilde{\Lambda}$ be its Pauli- twirled version given by: $\displaystyle\tilde{\Lambda}:=\frac{1}{|\mathcal{P}|}\sum_{P\in\mathcal{P}}P\Lambda(P\rho P)P\;.$ (18) For $Q\in\mathcal{P}$, twirled channels $\tilde{\Lambda}$ satisfy $\tilde{\Lambda}(Q)=c_{Q}Q$ for some coefficients $c_{Q}$. For every $\tilde{\Lambda}$ there exist noise rates $\lambda_{P}$ for $P\in\mathcal{P}/\\{I\\}$ such that $\tilde{\Lambda}=\prod_{P}\Gamma_{P}^{\lambda_{P}}$. These noise rates satisfy: $\displaystyle c_{Q}=e^{-2\sum_{P}(\lambda_{P}\cdot 1_{PQ=-QP})}\;.$ (19) A central convenience of Pauli noise channels is that they do not interfere with each other when propagated: Pauli noise channels commute $\Gamma^{\lambda_{P}}_{P}\Gamma^{\lambda_{Q}}_{Q}=\Gamma^{\lambda_{Q}}_{Q}\Gamma^{\lambda_{Q}}_{P}$, and the noise rates can be added together when the Pauli is the same $\Gamma^{\lambda_{1}}_{P}\Gamma^{\lambda_{2}}_{P}=\Gamma^{\lambda_{1}+\lambda_{2}}_{P}$ . #### Combining noise channels into a single fidelity. Say we are trying to compute the overall amount of noise in a particular quantum circuit that has been appropriately twirled. Gates and idle time of the qubits all contribute some amount of Pauli noise. We propagate all of the Pauli noise to the end of the circuit, thereby removing any noise that does not affect certain mid-circuit measurements. Finally, we must tally up the noise Paulis on the resulting quantum state. One metric for measuring the error on the final state is trace distance, or diamond norm if we are considering a channel. For a single Pauli noise source, we have the simple relation that for any $P$ we have $\left|\Gamma^{\lambda}_{P}-I\right|_{\diamond}=1-e^{-2\lambda}$. To generalize this to multiple Paulis, a simple approach could be to just apply the triangle inequality to all of the different Paulis. But it turns out we can do much better using the following bound on the process fidelity: ###### Lemma 1. Consider a channel $\Lambda=\prod_{P}\Gamma_{P}^{\lambda_{P}}$ for some rates $\lambda_{P}$. Then $\mathcal{F}_{\mathrm{proc}}(\Lambda,\mathcal{I})\geq\exp(-\sum_{P}\lambda_{P})$. This bound is still pretty loose, but it is very simple and does better than adding up diamond norms. This can be seen by, for example, looking at the channel $\prod_{i=1}^{N}\Gamma_{P_{i}}^{c/N}$. Lemma 1 gives $\mathcal{F}_{\mathrm{proc}}\geq\exp(-c)$ while adding up diamond norms and and converting them to a fidelity bound gives $\mathcal{F}_{\mathrm{proc}}\geq 1-\frac{1}{2}N(1-e^{-2c/N})$. The latter is looser for $N\geq 2$ and for any $c$. Lemma 1 also has the key advantage that it makes computation of the overall noise rate very simple: just add up all the noise rates. This allows us to simply tally the total idle time and count the number of CNOTs to obtain the total amount of noise, as in Appendix B.2. An issue with using Lemma 1 is that it becomes increasingly loose in the limit of large $\sum_{P}\lambda_{P}$. The quantity $\exp(-\sum_{P}\lambda_{P})$ vanishes in this limit, but in general we have $\mathcal{F}_{\mathrm{proc}}(\Lambda,\Lambda^{\prime})\geq 1/d$ for all $\Lambda,\Lambda^{\prime}$. When we only have one source of Pauli noise $\Gamma_{P}^{\lambda}$ then not even the lower limit of $1/d$ can be reached as $\lambda\to\infty$. Unfortunately, we see now way of overcoming this limitation while preserving the mathematical elegance of this tool: we would like to simply consider the quantity $\sum_{P}\lambda_{P}$. The reason for this shortcoming is that we do not account for cancellations between Pauli errors - we discuss the details of the derivation at the end of this section. Another limitation of this analysis is that it completely ignores crosstalk. Every gate is assumed to behave independently. Assuming independent errors corresponds to a worst-case analysis analogous to the union bound, so we would expect the bounds resulting from Lemma 1 to still roughly capture average error from crosstalk by accounting for it as $T_{2}$ dephasing noise, an error that we include when modeling experiments without dynamical decoupling. #### Propagating noise to the end of the circuit. Next, we discuss how to move all the noise sources to the end of the circuit. This is particularly easy since we are considering Clifford circuits. Once all the noise is in one place, we can use Lemma 1 to combine it into a single fidelity. With $\mathcal{U}\cdot\coloneqq U\cdot U^{\dagger}$ as before, elementary calculation shows that $\mathcal{U}\Gamma^{\lambda}_{P}=\Gamma^{\lambda}_{\mathcal{U}(P)}\mathcal{U}$, so Pauli-Lindblad noise propagated through a unitary Clifford circuit is still Pauli-Lindblad noise. Our circuits also feature several adaptive gates, propagation through which can be achieved as follows. Let $\Lambda_{\mathrm{disc}}$ be the channel that traces out the first of two qubits. Then $\Lambda_{\mathrm{disc}}\Gamma^{\lambda}_{P\otimes Q}=\Gamma^{\lambda}_{Q}\Lambda_{\mathrm{disc}}$. Similarly, let $\Lambda_{\mathrm{corr},P}$ be the channel that measures the first qubit and applies a correction $P$ onto the second qubit. If $P$ and $Q$ commute, then $\Lambda_{\mathrm{corr},P}\Gamma^{\lambda}_{Q\otimes R}=\Gamma^{\lambda}_{R}\Lambda_{\mathrm{corr},P}$. Otherwise, $\Lambda_{\mathrm{corr},P}\Gamma^{\lambda}_{Q\otimes R}=\Gamma^{\lambda}_{PR}\Lambda_{\mathrm{corr},P}$. Now that we have established how to move noise to the end of the circuit and to tally it into a bound on the fidelity, all that remains is to show how to bring various noise sources into Pauli-Lindblad form. #### Decoherence noise. We begin with decoherence noise that affects idling qubits. We consider depolarizing, dephasing, and amplitude damping noise. Conveniently, depolarizing and dephasing noise are already Pauli noise channels. A depolarizing channel $\Lambda_{\mathrm{dep},q}$ replaces the input $\rho$ with the maximally mixed state with probability $1-q$: $\displaystyle\Lambda_{\mathrm{dep},q}(\rho)=q\rho+(1-q)\frac{I}{2^{n}}\;.$ (20) We derive that $\Lambda_{\mathrm{dep},q}=\prod_{P\in\mathcal{P}_{n}/\\{I\\}}\Gamma^{\lambda}_{P}$ with $q=\exp(-4^{n}\lambda)$. The phase damping process is given by the Lindbladian with $L_{0}=\ket{0}\bra{0}$ and $L_{1}=\ket{1}\bra{1}$: $\displaystyle\mathcal{L}_{\mathrm{ph}}=\sum_{i\in\\{0,1\\}}L_{i}\rho L_{i}^{\dagger}-\frac{1}{2}\left\\{L^{\dagger}_{i}L_{i},\rho\right\\}\;.$ (21) Since $\mathcal{L}_{\mathrm{ph}}=\frac{1}{2}\left(Z\rho Z-\rho\right)$, it satisfies $e^{\lambda\mathcal{L}_{\mathrm{ph}}}=\Gamma^{\lambda/2}_{Z}$. We can easily compute $\lambda$ from a phase damping experiment: since $\bra{+}\Lambda^{\lambda}_{\mathrm{damp}}(\ket{+}\bra{+})\ket{+}=\frac{1}{2}(1+e^{-\lambda})$ we have $\lambda=t/T_{2}$. The amplitude damping channel is not a Pauli-Lindblad channel, and must be twirled in order to bring into Pauli-Lindblad form. The amplitude damping process $\mathcal{L}_{\mathrm{damp}}$ is given by $L=\ket{0}\bra{1}$ with: $\displaystyle\mathcal{L}_{\mathrm{damp}}(\rho)=L\rho L-\frac{1}{2}\left\\{L^{\dagger}L,\rho\right\\}\;.$ (22) If we let $\Lambda^{\lambda}_{\mathrm{damp}}:=e^{\lambda\mathcal{L}_{\mathrm{damp}}}$ then we have $\tilde{\Lambda}^{\lambda}_{\mathrm{damp}}=\Gamma^{\lambda/4}_{X}\Gamma^{\lambda/4}_{Y}$. Similarly, $\lambda$ can be obtained from an amplitude damping experiment: since $\bra{1}\Lambda^{\lambda}_{\mathrm{damp}}(\ket{1}\bra{1})\ket{1}=e^{-\lambda}$ we straightforwardly have $\lambda=t/T_{1}$. If we have both dephasing and amplitude damping noise, we can combine the two together as follows. For some $T_{1},T_{2}$, consider the combined noise channel $\Lambda^{t}_{\mathrm{noise}}=\exp\left(\frac{t}{T_{1}}\mathcal{L}_{\mathrm{damp}}+\frac{t}{T_{2}}\mathcal{L}_{\mathrm{ph}}\right)$. Then: $\displaystyle\tilde{\Lambda}^{t}_{\mathrm{noise}}=\Gamma_{X}^{\frac{t}{4T_{1}}}\Gamma_{Y}^{\frac{t}{4T_{1}}}\Gamma_{Z}^{\frac{t}{2T_{2}}}\;.$ (23) This follows from the fact that $\mathcal{L}_{\mathrm{damp}}$ and $\mathcal{L}_{\mathrm{ph}}$ commute. #### Noise from unitary gates. In principle we could perform experiments, as in [42], to determine the exact Pauli rates for each unitary, as is necessary for probabilistic error cancellation. However, two-qubit gates like the CNOT gate have fifteen noise parameters corresponding to the $4^{2}-1$ nontrivial two-qubit Pauli operators. For our purposes we would prefer to model CNOT noise using just a single number. One approach could be to just assume that the CNOT noise is simply depolarizing noise. In this case, all fifteen Pauli noise rates are equal and can be connected to the process fidelity. Say we aim to implement an ideal unitary $U$, but our hardware can only implement $\bar{\mathcal{U}}=\mathcal{U}\Lambda_{\mathrm{dep},q}$ up to a known fidelity $F(\mathcal{U},\bar{\mathcal{U}})$. Then $q=(4^{n}F(\mathcal{U},\bar{\mathcal{U}})-1)/(4^{n}-1).$ However, it turns out that spreading out the error uniformly over all the Paulis is rather cumbersome because it requires propagating every possible Pauli error. A more tractable approach is to just consider the worst case Pauli error. In that case, For any unitary $U$ and $P\in\mathcal{P}$, we have $F(\mathcal{U},\mathcal{U}\Gamma_{P}^{\lambda})=(1+e^{-2\lambda})/2$. #### Conclusions. We have derived a rigorous justification for a rather simple strategy for deriving theoretical predictions of noisy superconducting quantum hardware. Expressions for noise as a function of circuit size can be derived simply by counting the amount of idle time, CNOT gates, and number of mid-circuit measurements. The model has very few parameters, which are simply the Pauli- Lindblad noise rates corresponding to each of these operations (sometimes per unit time). These different noise rates are added up and converted to a fidelity via Lemma 1. The advantage of a rigorous derivation is that we can directly see the ways in which this model fails to tightly capture the actual error. A central issue is that Lemma 1 does not take into account cancellation between various noise sources, causing the fidelity to approach zero in the limit of high rate. This is despite the fact that the worst possible process fidelity is nonzero. Another oversimplification is that we do not capture the fact that not all possible Pauli noise rates can affect a given observable. We also cannot capture correlations between errors, as may be the case with crosstalk, and instead take a worst-case approach reminiscent of the union-bound. All of these reasons indicate that this model should produce relatively loose lower bounds. ###### Proof of Lemma 1.. Say $\Lambda(\rho)=\sum_{P,Q}c_{P,Q}P\rho Q$. Then $\mathcal{F}_{\mathrm{proc}}(I,\Lambda)=\mathcal{F}_{\mathrm{proc}}(I,\tilde{\Lambda})=c_{I,I}$. The proof proceeds with two loose lower bounds that notably fail to capture cancellations between different error sources. Given $\Lambda=\prod_{P}\Gamma_{P}^{\lambda_{P}}$, recall that $\Gamma_{P}^{\lambda_{P}}(\rho)=(1-\omega_{P})\rho+\omega_{P}P\rho P^{\dagger}$. Expanding out $\Lambda$, we see that: $\displaystyle c_{I,I}\geq\prod_{P}(1-\omega_{P})=\prod_{P}\frac{1+e^{-2\lambda_{P}}}{2}\;.$ (24) Next, observing that $(1+e^{-2x})/2\geq e^{-x}$ for $x>0$: $\displaystyle...\geq\prod_{P}e^{-\lambda_{P}}=\exp\left(-\sum_{P}\lambda_{P}\right)\;.$ (25) ∎ #### Convergence to 0.4. In the main text, we remarked that the fidelities of the measurement-based CNOT experiments converge to a value slightly below 0.4, as is observed in Figure 1 (c). As discussed, this is due to the structure of the measurement- based circuit in Figure 1 (a). While the circuit also experiences infidelity on the top and bottom qubits due to idle time and some CNOTs, the only infidelity that actually scales with $n$ is due to incorrect $Z$ and $X$ corrections on the top and bottom qubits respectively. We can model this noise as $\Gamma_{ZI}^{\lambda_{ZI}}\Gamma_{IX}^{\lambda_{IX}}$ in the limit of large $\lambda_{ZI},\lambda_{IX}$, in which case $\omega_{ZI},\omega_{IX}$ approach $1/2$. We proceed as in (24). Since these Pauli errors cannot cancel, the calculation is exact. $\displaystyle\mathcal{F}_{\mathrm{proc}}(I,\Gamma_{ZI}^{\lambda_{ZI}}\Gamma_{IX}^{\lambda_{IX}})=c_{I,I}=(1-\omega_{ZI})(1-\omega_{IX})=1/4\;.$ (26) This converts to $\mathcal{F}_{\mathrm{gate}}(I,\Gamma_{ZI}^{\lambda_{ZI}}\Gamma_{IX}^{\lambda_{IX}})=(4\mathcal{F}_{\mathrm{proc}}(I,\Gamma_{ZI}^{\lambda_{ZI}}\Gamma_{IX}^{\lambda_{IX}})+1)/(4+1)=0.4$.
11institutetext: Language Technologies Research Center, KCIS, IIIT Hyderabad 11email<EMAIL_ADDRESS><EMAIL_ADDRESS> 22institutetext: TCS Research, Hyderabad, India 22email<EMAIL_ADDRESS> # LimGen: Probing the LLMs for Generating Suggestive Limitations of Research Papers Abdur Rahman Bin Md Faizullah 11 * * Ashok Urlana 22 * * Rahul Mishra 11 ###### Abstract Examining limitations is a crucial step in the scholarly research reviewing process, revealing aspects where a study might lack decisiveness or require enhancement. This aids readers in considering broader implications for further research. In this article, we present a novel and challenging task of Suggestive Limitation Generation (SLG) for research papers. We compile a dataset called LimGen, encompassing 4068 research papers and their associated limitations from the ACL anthology. We investigate several approaches to harness large language models (LLMs) for producing suggestive limitations, by thoroughly examining the related challenges, practical insights, and potential opportunities. Our LimGen dataset and code can be accessed at https://github.com/armbf/LimGen. ###### Keywords: Limitations LLM Constrained Text Generation. **footnotetext: Equal contribution. ## 1 Introduction The process of reviewing research articles lies at the core of the scientific publication process. It serves as a pivotal stage wherein experts meticulously scrutinize the content, methodology, and findings of a study before it is disseminated to the broader scientific community. This rigorous evaluation and critique of the strengths and limitations of the research work ensures the integrity, reproducibility, and credibility of the research being presented. Identifying the potential limitations of the research papers is a very tedious and time-consuming process that involves delving into the intricacies of the research methodology, understanding the theoretical framework, and grasping the nuances of the subject matter. This endeavor demands not only a broad knowledge base but also keen attention to detail and analytical thinking. Table 1: Comparison of actual vs LLM generated suggestive limitations (best viewed in color). Magenta indicates the matching content between the actual and LLM-generated limitations, whereas Blue represents the novel limitations. Actual Limitation: Contamination is a significant problem in the era of pretrained language models trained on large web corpora… Additionally, this data’s distribution is fixed to pre-college Physics, Chemistry, and Mathematics. There are more gradations and difficulty levels at which the model can be evaluated, which have not been tested as part of our analysis. --- Llama2-DPR: The limitations of the approach are: The dataset is limited to only three subjects (Physics, Chemistry, and Mathematics) and does not cover other important subjects like Biology, Geology, and Computer Science. The dataset does not include any real-world applications or scenarios, limiting the generalization of the model’s capabilities. The evaluation metric used is based on the number of correct answers, which does not fully capture the model’s reasoning abilities. The dataset does not include any adversarial examples or challenging cases to test the model’s robustness and adaptability. This paper presents a novel task of Suggestive Limitation Generation (SLG) for research papers, which aims to generate a diverse array of potential limitations specific to each paper, providing reviewers with valuable insights to facilitate their assessment process. The task of SLG poses a greater challenge compared to text summarization, question-answering (QnA), and open- ended text generation tasks. It demands not only reasoning abilities but also the capacity to discern and incorporate associations from a corpus of previously encountered papers during the fine-tuning process. For example, in Table 1, we show the actual limitation block from a research paper in row 1 and the corresponding generated limitations in row 2. Illustrated in magenta, the proposed model demonstrates the capability to generate limitations akin to those originally outlined in the paper (the dataset is limited to only three subjects, Physics, Chemistry, and Mathematics) Additionally, it showcases the ability to propose novel, valid limitations (depicted in blue) related to adversarial examples, a facet not included by authors in original limitations. The key idea behind this approach is to capitalize on cues regarding the similarities or differences among research papers and learn to recommend comparable limitations for a paper, especially when its underlying methodology closely aligns with that of a set of other papers. To this end, we create a dataset called LimGen, comprising 4068 research papers and corresponding limitations from the ACL anthology. Subsequently, we probe many state-of-the- art large language models (LLMs) in a multitude of experimental setups to generate the suggestive limitations of the research papers. We conduct a very thorough evaluation by utilizing automatic (rouge-based), human (manual expert-based), and model assistive (LLM-based) approaches. The key contributions of this work are: 1) To the best of our knowledge, we are the first to propose the task of Suggestive Limitation Generation (SLG) for research papers. 2) We release a SLG dataset LimGen, consisting of 4068 papers and corresponding limitations. 3) We propose and experiment with several schemes to utilize LLMs for SLG. 4) We perform thorough evaluations using automated, human, and LLM-driven methodologies. ## 2 Related work Scientific document understanding poses a persistent challenge, primarily attributable to its structure, diverse content modalities (such as tables and figures), and the incorporation of citations within the text. The recent emergence of large-scale scientific document summarization datasets were automatically collected from the public repositories [3, 7]. Several works on scientific documents encompass tasks such as abstract generation [12], delving into the contributions outlined in a paper [17, 10], scientific papers summarization [2] and formulate multi-perspective summaries by leveraging reviews of the research papers [4, 24]. Moreover, few works delve into other forms of supervision for scientific document understanding, including citations [18, 25], author-written highlights [5], transcripts from conference presentations of the research papers [14] and annotations [19]. In our work, we collected the research papers and corresponding limitations from the ACL anthology. In contrast to existing works, we attempt to generate suggestive limitations of the research papers using LLMs. To the best of our knowledge, this is the first work towards generating limitations of the research papers. ## 3 LimGen Dataset ### 3.1 Dataset collection We obtain the dataset from ACL Anthology111https://aclanthology.org/ website. We take advantage of the recent mandatory inclusion of the ‘limitations’ section in the research paper for the submission of Computational Linguistics- related venues. We scrap the proceedings of EMNLP, ACL, and EACL venues of 2022 and 2023 years respectively. After obtaining the papers, the initial step involves using scipdf_parser222https://github.com/titipata/scipdf_parser to parse the PDFs. The parser segregates the content of the paper into section- wise information. From the extracted sections, to create the ‘source’ text for the SLG task, we discard some of the sections Abstract, Introduction, Related Works, Acknowledgements, Conclusion, Ethics Statement, Appendix, References, Limitations and preserve the main contribution of the paper in form of methodology, experiments, results, and discussions, etc. We use the corresponding actual ‘limitations’ section as the reference limitations. In total, we utilize 4068 peer-reviewed short and long papers from three different ACL venues. As an initial exploratory analysis of the proposed LimGen dataset, we computed several key statistics. Notably, the average length of the research papers stands at approximately 5000 tokens and 187 sentences. Whereas the limitation sections, average around 230 tokens, and 9 sentences. The longer length of the papers poses a challenge for the large language models to process the longer context length. For detailed statistics corresponding to each conference, please refer to the provided Table 2. Table 2: LimGen Dataset Statistics Number of research papers 4068 --- ACL 2022 | 1750 | | #Avg words per paper | 5122 EMNLP 2022 | 1227 | | #Avg sentences per paper | 188 EACL 2022 | 456 | | #Avg words per limitation | 230 EMNLP 2023 | 635 | | #Avg sentences per limitation | 10 Table 3: Manual analysis of 60 research papers; Relevance | Deduce Limitation | Future work or Limitation ---|---|--- Yes | No | Partial | Yes | No | Partial | Yes | No | Partial 37 | 3 | 20 | 12 | 13 | 35 | 15 | 13 | 32 Limitations related to Methodology | Experimental setup | Dataset | Evaluation 41 | 17 | 22 | 7 ### 3.2 Nature of the limitations To understand the nature of the limitations in research papers, we conduct a manual analysis of 60 papers. We maintain diversity (short, and long papers from diverse venues) in paper selection to capture the stylistic variations of the limitations present in the research papers. The analysis aims to understand, 1) The relevance of the underlying limitation to the research paper, 2) whether the given limitation can be deduced by reading the paper or not?, 3) establish whether the mentioned limitation represents a real constraint or suggests potential avenues for future research, 4) classify the limitation according to its relevance to specific sections of the paper, including Methodology, dataset, evaluation or experimental setup. We present the outcome of our manual analysis in Table 3. We note that within our sample, numerous papers regard their future prospects as limitations. The extent of the limitation section ranges from mere sentences to substantially lengthy paragraphs. Additionally, our observations indicate that in over 50% of the papers, discerning the stated limitation directly from the text is not straightforward. Furthermore, a significant portion of these limitations are predominantly associated with the methodology section of the paper. There are few instances, where the limitations cover more than one section of the research paper. ## 4 Benchmark Experiments ### 4.1 Task formulation This section introduces the Suggestive Limitation Generation (SLG) task formulation. To produce the limitations of the papers, we approach the task as a Seq2Seq problem. Precisely, we craft a model designed to intake a scientific paper R as input and systematically produce a structured limitation block L = l(1:n), where l(1:n) represents the combination of n limitations of R, sequentially generated. ### 4.2 Methodology This section describes the various approaches to generate the suggestive limitations. We explore two suitable text generation paradigms to generate the limitations of the research papers. Firstly, we consider the limitation generation as a summarization task and utilize the summarization-specific pre- trained models including BART333https://huggingface.co/facebook/bart-large-cnn [15] and PEGASUS444https://huggingface.co/google/pegasus-large [26] to generate the suggestive limitations. Given that the objective of the SLG task surpasses the complexity of the summarization task, which typically involves limited or constrained generation entropy, the SLG task demands a higher degree of generation entropy. Unlike summarization, where the model’s task is to condense information, in SLG, the model must infer and recommend limitations from the source content, drawing from its understanding during fine-tuning. Hence, the generative scheme becomes pivotal. To this end, as the second paradigm, we utilize popular Large Language Models (LLMs) namely T5555https://huggingface.co/google-t5/t5-base [20], Cerebras-GPT [9] and Llama 2 [22]. To experiment with both of these paradigms, we utilize the following three schemes. Figure 1: General architecture diagram for the suggestive limitation generation. Figure 2: Architecture diagram for DPR fine-tuning. Figure 3: Architecture diagram for Chain modeling. #### 4.2.1 Non-truncated research paper In this scheme, we employ the entire research paper and its associated limitations to experiment with both summarization-specific and generative models, as illustrated in Figure 1. We fine-tune the summarization models such as BART, PEGASUS and also utilize generative models including T5, Llama 2, and Cerbras GPT to perform the zero-shot prompting and fine-tuning. To experiment with the generative models in this scheme, we use the prompt depicted in Table 4. However, this approach is constrained by the models’ max input token limit, resulting in a lack of comprehensive context for the research paper. Table 4: Prompt for fine-tuning with the non-truncated research papers. | Prompt: Generate limitations or shortcomings for the following scientific paper --- {Paper Text} Limitations: End-of-Prompt #### 4.2.2 Dense Passage Retrieval (DPR) To address the length constraint in the Non-truncated research paper scheme, we employ the DPR approach. This approach processes only the relevant passages for each sentence present in the limitation section. To obtain the relevant passages, we utilize the following three-stage approach. 1. 1. Paragraph and sentence processing: We segment the research papers into paragraphs and obtain the sentences of each paragraph by using Spacy666https://spacy.io/api/tokenizer. 2. 2. Tokenization and paragraph management: We tokenize these passages using BertTokenizer [8] to manage the length of each passage to maintain the max input token limit of 1024 for Llama 2 and 2048 for Cerebras. We merge smaller passages for optimization and split larger passages to adhere to the max input token limit. 3. 3. Limitation sentence extraction and encoding: We encode limitation sentences for each document and the passages using SentenceTransformer [21] (all- MiniLM-L6-v2777https://www.sbert.net/). We compute the cosine similarity between each sentence in the limitations and all sentences within each passage. Further, discard the passages with a similarity score of less than 0.5. Ultimately, we identify the top three passages with the highest cosine similarity for each sentence in the limitations. Table 5: Prompt for DPR-based fine-tuning. | Prompt: Generate limitations or shortcomings for the following passage from a --- scientific paper passage:{DPR paragraph} A brief technical summary of the scientific paper for context:{Summary of the paper} Limitations: End-of-Prompt We fine-tune the model by employing individual limitation sentences as target outputs, with the input comprising the top three passages retrieved through DPR (Dense Passage Retrieval). These passages are specifically obtained using the respective limitation sentence as the query. The prompt for the DPR-based fine-tuning is depicted in Table 5. The fine-tuned model is utilized to generate limitations for each passage extracted from the paper. These limitations are then compiled to form the comprehensive set of limitations associated with the research paper as shown in Figure 2. Both the summarization and generative models utilize the DPR-retrieved passages for the experiments. While generally effective, this approach may produce irrelevant limitations due to the lack of a broader context of the entire research paper. To tackle this challenge, we investigate the application of chain modeling techniques as a prospective solution. #### 4.2.3 Chain Modeling To overcome the constraints posed by the lack of contextual information in the DPR approach, we implement the chain modeling approach inspired by LangChain888https://github.com/LangChain-ai/LangChain. The chain modeling consists of two stages. In stage 1 (limitation generation), we generate the limitations for all passages of the research paper obtained by following the steps 1 and 2 in DPR model creation (see Section 4.2.2). However, going through individual passages, we observe that, the model lacks the comprehensive context provided by the entire research paper. To overcome this, in the prompt, we include the summary of the research paper along with each passage. The prompt to obtain the summary is mentioned in Table 6. In stage 2 (refinement), we refine and distill all the generated limitations, eliminate obvious duplicates, and standardize the language as depicted in Figure 3. The prompt for both stages is shown in Table 7. The chain modeling approach utilizes two distinct fine-tuned models. The former one (LLM1 in Figure 3) is fine-tuned on DPR passages along with summary of the research paper as input and limitation as output. The prompt for the same is mentioned in Table 5. Whereas the latter one (LLM2 in Figure 3) is fine-tuned on full non-truncated paper as input and corresponding paper limitations as output using the prompt depicted in Table 4. This approach is indicated as Llama2-DPR-Distilled in Table 9. Another approach using the full non-truncated fine-tuned Llama 2 model for both LLM1 and LLM2, which is indicated as Llama2-FT-Distilled (FT refers to Full-text) in Table 9. We also implement chain modelling approach where the result of the previous iteration is passed to the current iteration with the passage. But it is observed that when the model hallucinates or causes repetitions, all the subsequent steps are affected. For this setting, the corresponding results are reported in the table 9 under Llama2-Continuous. Table 6: Prompt for generating the summary of the research paper. | Initial Prompt: Write a concise technical summary of the following: {first passage} --- CONCISE TECHNICAL SUMMARY: {Summary of the paper} End-of-Prompt Iteration Prompt: Your job is to produce a final technical summary of a research paper. We have provided a summary generated by you up to a certain point: {summary from the previous step} We have the opportunity to refine the existing summary (only if needed) with some more context below. {next passage} Given the new context, refine the original technical summary. Keep the technical summary less than 350 words. If the context isn’t useful in the technical research context, return the original summary. Do not ask any questions in the response. Refined Technical Summary: End-of-Prompt ### 4.3 Experimental Setup To conduct experiments with generative models with DPR and chain modeling approaches, we use the Cerebras-GPT [9] 1.3B and Llama 2 [22] 7B versions. Due to hardware constraints, we utilize the smaller versions of the LLMs and perform the fine-tuning with the LoRA [11] configuration with 8-bit quantization. We utilize the XTuring code base999https://github.com/stochasticai/xTuring for the fine-tuning. For the DPR approach, we limit the number of top passages passed to the model to 2 for Llama 2 and 3 for Cerbras-GPT. Moreover, we utilize four Nvidia GeForce RTX 2080 Ti GPUs (11GB). Due to increased memory requirements for finetuning Llama 2 models for chain modeling with non-truncated research papers and processing the DPR dataset with summaries as context, we temporarily upgrade to 4 NVIDIA GeForce RTX 3090 GPUs. We use 80-10-10 split of the LimGen dataset for the creation of train, valid, and test datasets. ##### Specifics for Chain Modelling: The chain modeling approach requires more computation due to an increase in the input context length and the need for higher inference speed to process all the passages in a paper. To accommodate these requirements, we fine-tune the Llama 2 model using Axolotl101010https://github.com/OpenAccess-AI- Collective/axolotl with LoRA and FlashAttention [6]. Further, we use AWQ via AutoAWQ111111https://github.com/casper-hansen/AutoAWQ with zero-point, group size of 128, 4 bits, and GEMM version on vLLM [13] for efficient inference. Table 7: Prompts for the chain modeling approach. | Iteration Prompt: Your job is to take in a passage from a research paper --- and a concise summary of that research and identify one or two main limitations from the given passage using the summary as context. Paper Passage: {passage} Paper Summary: {summary} Limitations: End-of-Prompt Distill Prompt: The following is set of limitations:. {list of generated limitations} Take these and distill them into a final, consolidated list of limitations: End-of-Prompt ## 5 Experimental Results and Analysis Summarization and generative approaches. The results for experiments with summarization and generative models are listed in Table 8. Although the objective of generating a summary and suggestive limitation may appear similar, our experiments reveal that models trained specifically for summarization do not effectively generate insightful limitations, often merely extracting sentences from the texts. Whereas, in case of generative models, we observe a significant decline in output quality when excluding ‘Limitations:’ keyword in the prompt. In the generation approach, Llama 2 and Cerebras models demonstrate proficiency in limitations generation, but their effectiveness is hindered by the truncation of the input paper due to token length limitations. Moreover, they fail to generate limitations across different sections of the research paper due to the handling of limited context length. Dense Passage Retrieval. To overcome the limited context issue, we utilized the relevant passages obtained from the DPR approach to fine-tine the model. This model then iteratively takes passages and generates limitations for each passage, which are collated to get the set of limitations for a paper. This experiment’s performance is generally effective when a limitation could be derived from a passage. However, despite the DPR approach being effective in highlighting relevant limitations, it inadvertently points to less pertinent, paragraph-level limitations due to the lack of overall context of the paper. Chain modeling. In this approach, we experiment with fine-tuning the models by providing the summary of the entire paper along with the passages to avoid the overall context issue. After all the limitations are generated using the input passages and summary, in the refinement stage, all the duplicates are discarded. Table 8: Limitation generation experimental results; BS refers to BERTScore. Model | Approach | Without DPR data | With DPR data ---|---|---|--- R-1 | R-2 | R-L | BS | R-1 | R-2 | R-L | BS BART-large | Fine-tuning | 30.8 | 4.5 | 15.8 | 82.8 | 10.7 | 0.6 | 8.3 | 82.8 PEGASUS-large | 20.2 | 3.1 | 12.7 | 82.3 | 16.7 | 7.1 | 14.2 | 84.7 T5-base | 27.7 | 4.3 | 16.4 | 83.6 | 18.8 | 7.6 | 16.2 | 85.9 Llama-2-7b | Zero-shot | 21.3 | 3.3 | 12.1 | 81.9 | 16.7 | 5.2 | 8.9 | 82.4 Cerebras GPT2.7B | 17.6 | 2.1 | 12.1 | 78.9 | 19.8 | 4.8 | 10.3 | 80.5 Llama-2-7b | Fine-tuning | 21.4 | 3.1 | 12.7 | 81.1 | 34.8 | 11.0 | 17.7 | 83.5 Cerebras GPT2.7B | 16.9 | 1.9 | 12.0 | 79.1 | 32.4 | 9.6 | 15.9 | 83.4 Table 9: Limitation generation experimental results for chain modeling. The model utilized for distillation/refinement consistently involves Llama2 fine-tuned on Full paper. ‘Fine-tuning’ column indicates the type of the dataset used for fine-tuning. Chain Modeling | Fine-tuning | R-1 | R-2 | R-L | BS ---|---|---|---|---|--- | Llama2-Continuous --- Full Paper | 24.3 | 6.2 | 15.6 | 82.9 | Llama2-DPR --- DPR dataset | 33.5 | 9.9 | 16.3 | 83.2 Llama2-FT-Distilled | Full paper | 28.5 | 5.8 | 15.2 | 83.5 Llama2-DPR-Distilled | DPR dataset | 30.4 | 7.6 | 16.2 | 83.3 ### 5.1 Evaluation To assess the performance of the proposed models, we adopt the three types of evaluation strategies, 1) Automatic evaluation, 2) LLM-base evaluation, and 3) Human evaluation. Automatic evaluation. We conduct the automatic evaluation by using the standard evaluation metrics such as ROUGE [16] and BERTScore [27]. We use the RoBERTa121212https://huggingface.co/FacebookAI/roberta-large large model for the BERTScore calculation. Table 10: Prompts for GPT-4-based evaluation of the generated limitations. Evaluation Prompt: For the below sets of limitations created for the above research paper, tell me if they are actually limitations and if each limitation set is a proper limitation of the paper even though it might not be mentioned as a limitation in the original paper. Rate them with respect to the original limitations section in the above paper and with respect to the paper itself. Also give each set a score of 1 to 5 on the above qualities, with 5 indicating very good limitations for the above paper. Defend the score. --- {Limitations generated by each model} End-of-Prompt LLM-based Evaluation. We perform the LLM-based evaluation by utilizing the GPT-4 [1] to evaluate the quality of the generated limitations. We use the zero-shot prompting strategy to obtain the evaluation scores from GPT-4. We instructed the GPT-4 to assign a score between 1 (least) to 5 (best). As shown in Table 11, we observe that the ‘Llama2-FT-Distilled’ approach outperforms the remaining models and the T5-base obtains the lowest score. It indicates that summarization-specific models fail to generate limitations of research paper due to their inherent nature of training objective. The prompt used to perform the LLM-based evaluation is mentioned in Fig 10. We notice that GPT-4 does not thoroughly analyze the limitations. It tends to assign high scores to general limitations of the model or approach, even if they may not be accurate within the context of the provided research paper. When prompted again to verify the limitations once more, it might fail to identify incorrect limitations if they’re presented in a manner that implies they pertain to any aspect discussed in the paper. When a specific limitation is singled out and prompted for re-evaluation, GPT-4 shows improved performance but struggles until the actual issue with the limitation is pointed out explicitly. Human Evaluation. We perform the human evaluation of 50 research papers (selected at random from the test set) and corresponding generated limitations. The limitations are generated by four different systems including Llama2-DPR, Llama2, T5-base, and Llama2-FT-Distilled approaches. We asked the evaluators to rate Yes, No, or Partial for each of the following questions. Q1. Whether the limitation generated by the model makes sense or not?, Q2. Is there any overlap between the gold and generated limitation?, Q3. The generated limitation is an actual limitation or not (can be a summary or prospective work), Q4. The generated limitation contains any hallucinations, repetitive sentences, or grammatical errors. The results of the human evaluation are detailed in Table 12. Our findings and manual evaluations suggest a superior performance of chain modeling using Llama2 trained on the full paper dataset for both limitation generation and reduction step (Llama2-FT-Distilled). This approach yields fewer but proper, coherent limitations, effectively reducing non-limitation and speculative content. The enhancement in the quality of generated limitations can be credited to the comprehensive summary provided to the model, enabling it to grasp the entirety of the paper’s context. Also, the distillation step of the chain modeling adds coherence and structure to the results. The DPR dataset- trained model (Llama2-DPR-Distilled) closely trails behind, likely owing to its proficiency in generating highly pertinent limitations. However, it occasionally produces generic limitations rather than focused ones, such as “The model employs Micro-F1 scores as the primary evaluation metric, yet other metrics might be more suitable depending on the particular task or application”. Table 11: Automatic evaluation using GPT-4 | T5-base | | Llama2 --- | Llama2-DPR --- | Llama2-FT-Distilled --- Score | 2.71 | 3.60 | 3.12 | 4.10 Table 12: Human evaluation of LLMs generated limitations; For Q1-Q3, the higher values of ‘Yes’ are preferred, whereas for Q4 the higher values of ‘No’ are desired. | Llama2-DPR | Llama2 | T5-base | Llama2-FT-Distilled ---|---|---|---|--- | Yes | No | Partial | Yes | No | Partial | Yes | No | Partial | Yes | No | Partial Q1 | 10 | 24 | 16 | 19 | 12 | 19 | 10 | 22 | 18 | 35 | 04 | 11 Q2 | 06 | 36 | 08 | 05 | 31 | 14 | 02 | 34 | 14 | 14 | 15 | 21 Q3 | 08 | 24 | 18 | 22 | 13 | 15 | 10 | 21 | 19 | 31 | 04 | 15 Q4 | 18 | 26 | 07 | 11 | 30 | 09 | 27 | 12 | 11 | 10 | 31 | 09 Error analysis. As part of human evaluation, Q4 helps to obtain insights for the error analysis of the models. We notice that the ‘Llama2-FT-Distilled’ model generates better limitations compared to other models. The limitations generated by ‘Llama2-DPR’ and ‘T5-base’ do not make sense and contain a lot of noisy information or produce the summary of the research paper rather than limitations. Moreover, the T5-base model is highly prone to directly copying several phrases from the research paper. Llama2-DPR considers local limitations, referring to limitations mentioned in a passage that refers to another paper or approach but are not the limitations of the current paper. We have also observed that more than 50% of limitations generated from T5-base are prone to hallucination or contain repetitive phrases. Despite generating lengthy limitations, most of the ‘Llama2-FT-Distilled’ model-generated limitations make sense. ## 6 Challenges and Future work In this section, we explore the key challenges and potential opportunities associated with the SLG task. Complexity of the SLG task. As illustrated in Table 3, it is often difficult to infer the actual limitations solely from the content of the papers due to the lack of detailed context surrounding these limitations. Consequently, predicting such nuanced limitations poses a challenge for the model. Our experiments reveal that both summarization and open-text generation strategies struggle to generate these intricate limitations. Therefore, in our experiments, we place emphasis on the fine-tuning phase, with the expectation that the model will learn to discern similarities and differences across research papers, enabling it to infer nuanced limitations more effectively. Furthermore, the concept of similarity among papers can be explicitly modeled by incorporating auxiliary information, such as citations. Evaluation metrics suitability. Lexical overlapping-based metrics such as ROUGE operate on n-gram-based matching, yet many generated limitations feature valid novel sentences and phrase formulations compared to those mentioned in the research papers. This disparity makes lexical matching-based metrics imperfect for evaluating the task of SLG. As shown in Tables 8 and 9, We find no notable discernable variation in the BERTScore values across the models. However, our human evaluation reveals considerable variations in the quality of the limitations generated by different models. Developing a tailored evaluation metric for the SLG task stands out as the most promising path forward. Multi-modal content. We do not explore the utilization of non-textual elements such as images and tables present in the research paper to generate the limitations. Images such as architecture diagrams and plots, along with tables like result tables and ablation tables, can provide supplementary context, especially for grasping nuanced potential limitations. Coherence and relevance. We observe that despite the apparent superior performance of DPR-based models, they fail to generate coherent limitations. The best Llama2-FT-Distilled model also generates a few speculative limitations and has difficulty in generating highly relevant limitations for every paper. Open-ended generation of LLMs. As illustrated in Table 12, around 20% of the limitations generated with the aid of LLMs are susceptible to issues such as hallucination, repetitions, and grammatically incorrect sentence structures. Controlling these issues leads to the generation of more faithful limitations. Controllability. Recent advancements in controllable generation emphasize the ability to address the specific intentions of users [23]. However, we note that LLMs occasionally struggle to consistently generate specific structured limitations. At times, the models produce limitations as bullet points, while in other instances, they generate them as paragraphs. ## 7 Limitations In this work, our focus is solely on extracting limitations from only textual content within research papers. Nevertheless, incorporating non-textual elements like tables and images could enhance the generation of more contextually relevant limitations. Due to compute constraints, we experimented with the Llama 2 7B and Cerebras-GPT 1.3B models. The automated evaluation metrics used are insufficient for the complexity of this task, so we supplemented our evaluation process with manual evaluations. Since the LimGen dataset, sourced from the ACL anthology, it may exhibit bias towards generating limitations specific to computational linguistic aspects. As a result, its applicability to other domains like physics, chemistry, and mathematical research studies could be limited. We listed the limitations generated by our best model for this paper in Table 13. We observe that, despite the broad coverage of the limitations, the model generates few of the sub-standard limitations. Table 13: Limitations generated by our best-performing model for our paper (best viewed in color). Magenta indicates the appropriate limitations, whereas Blue represents sub-standard limitations. Hallucination: The model may generate limitations that are not present in the input text, which can lead to inaccurate or irrelevant information in the generated limitations. Repetition: The model may repeat similar phrases or sentences multiple times in the generated limitations, which can result in a lack of diversity in the generated limitations. Limited training data: The model may struggle to generate limitations that are not present in the training data, which can affect the quality of the output. Inadequate evaluation metrics: The authors do not use adequate evaluation metrics to assess the performance of their approach, which can affect the validity of their results. Lack of consideration of the research question: The authors do not consider the research question when generating limitations, which can lead to inaccurate or irrelevant limitations. Lack of consideration of the methodology: The authors do not consider the methodology used in the research paper when generating limitations, which can lead to inaccurate or irrelevant limitations. --- ## 8 Conclusions In this paper, we introduce the novel task of Suggestive Limitation Generation (SLG) for research papers, aiming to provide reviewers with potential limitations of the underlying work, thereby assisting in the review process. We compile a dataset of 4068 research papers and corresponding limitations from the ACL anthology. We propose an LLM-based baseline for SLG tasks and conduct several ablation studies. Subsequently, we perform a thorough evaluation of the proposed models with automatic, LLM-based, and human evaluation schemes. Moving forward, our plans involve incorporating images and tabular content in addition to text for the SLG task. ## 9 Ethics Statement In the creation of the LimGen dataset, we did not collect any personal data. Instead, we rely on the publicly accessible dataset from the ACL anthology. This research did not involve any user data. In the release of LimGen, all personal information pertaining to the authors of the research papers will be excluded. ## References * [1] Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) * [2] Cachola, I., Lo, K., Cohan, A., Weld, D.S.: Tldr: Extreme summarization of scientific documents. In: Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 4766–4777 (2020) * [3] Cohan, A., Dernoncourt, F., Kim, D.S., Bui, T., Kim, S., Chang, W., Goharian, N.: A discourse-aware attention model for abstractive summarization of long documents. In: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers). pp. 615–621 (2018) * [4] Cohan, A., Feigenblat, G., Ghosal, T., Shmueli-Scheuer, M.: Overview of the first shared task on multi perspective scientific document summarization (mup). In: Proceedings of the Third Workshop on Scholarly Document Processing. pp. 263–267 (2022) * [5] Collins, E., Augenstein, I., Riedel, S.: A supervised approach to extractive summarisation of scientific papers. In: Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017). pp. 195–205 (2017) * [6] Dao, T., Fu, D., Ermon, S., Rudra, A., Ré, C.: Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems 35, 16344–16359 (2022) * [7] Delort, J.Y., Alfonseca, E.: Dualsum: a topic-model based approach for update summarization. In: Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics. pp. 214–223 (2012) * [8] Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). pp. 4171–4186 (2019) * [9] Dey, N., Gosal, G., Zhiming, Chen, Khachane, H., Marshall, W., Pathria, R., Tom, M., Hestness, J.: Cerebras-gpt: Open compute-optimal language models trained on the cerebras wafer-scale cluster (2023) * [10] Hayashi, H., Kryściński, W., McCann, B., Rajani, N., Xiong, C.: What’s new? summarizing contributions in scientific literature. In: Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. pp. 1019–1031 (2023) * [11] Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) * [12] Kumarasinghe, D., de Silva, N.: Automatic generation of abstracts for research papers. In: Proceedings of the 34th Conference on Computational Linguistics and Speech Processing (ROCLING 2022). pp. 221–229 (2022) * [13] Kwon, W., Li, Z., Zhuang, S., Sheng, Y., Zheng, L., Yu, C.H., Gonzalez, J., Zhang, H., Stoica, I.: Efficient memory management for large language model serving with pagedattention. In: Proceedings of the 29th Symposium on Operating Systems Principles. pp. 611–626 (2023) * [14] Lev, G., Shmueli-Scheuer, M., Herzig, J., Jerbi, A., Konopnicki, D.: Talksumm: A dataset and scalable annotation method for scientific paper summarization based on conference talks. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 2125–2131 (2019) * [15] Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 7871–7880 (2020) * [16] Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. 74–81. Association for Computational Linguistics, Barcelona, Spain (Jul 2004) * [17] Liu, M.H., Yen, A.Z., Huang, H.H., Chen, H.H.: Contributionsum: Generating disentangled contributions for scientific papers. In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. pp. 5351–5355 (2023) * [18] Mao, Y., Zhong, M., Han, J.: CiteSum: Citation text-guided scientific extreme summarization and domain adaptation with limited supervision. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 10922–10935. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022) * [19] Meng, R., Thaker, K., Zhang, L., Dong, Y., Yuan, X., Wang, T., He, D.: Bringing structure into summaries: a faceted summarization dataset for long scientific documents. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). pp. 1080–1089 (2021) * [20] Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485–5551 (2020) * [21] Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 3982–3992 (2019) * [22] Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) * [23] Urlana, A., Mishra, P., Roy, T., Mishra, R.: Controllable text summarization: Unraveling challenges, approaches, and prospects–a survey. arXiv preprint arXiv:2311.09212 (2023) * [24] Urlana, A., Surange, N., Shrivastava, M.: Ltrc@ mup 2022: Multi-perspective scientific document summarization using pre-trained generation models. In: Proceedings of the Third Workshop on Scholarly Document Processing. pp. 279–284 (2022) * [25] Yasunaga, M., Kasai, J., Zhang, R., Fabbri, A.R., Li, I., Friedman, D., Radev, D.R.: Scisummnet: A large annotated corpus and content-impact models for scientific paper summarization with citation networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 7386–7393 (2019) * [26] Zhang, J., Zhao, Y., Saleh, M., Liu, P.: Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In: International Conference on Machine Learning. pp. 11328–11339. PMLR (2020) * [27] Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675 (2019)
# Unisolvence of random Kansa collocation by Thin-Plate Splines for the Poisson equation F. Dell’Accio University of Calabria, Rende (CS), Italy A. Sommariva M. Vianello ###### Abstract Existence of sufficient conditions for unisolvence of Kansa unsymmetric collocation for PDEs is still an open problem. In this paper we make a first step in this direction, proving that unsymmetric collocation matrices with Thin-Plate Splines for the 2D Poisson equation are almost surely nonsingular, when the discretization points are chosen randomly on domains with analytic boundary. ## 1 Introduction Kansa unsymmetric collocation, originally proposed in the mid ’80s [11], has become over the years a popular meshless method for the discretization of boundary value problems for PDEs. Despite its wide and successful adoption for the numerical solution of a variety of physical and engineering problems (cf. e.g. [4] with the references therein), a sound theoretical foundation concerning unisolvence of the corresponding linear systems is still missing. Indeed, it was shown by Hon and Schaback [10] that there exist point configurations that lead to singularity of the collocation matrices, though these are very special and “rare”cases. For this reason greedy and other approaches have been developed to overcome the theoretical problem and ensure invertibility, cf. e.g. [14, 18]. On the other hand, in the textbook [8] one can read : “Since the numerical experiments by Hon and Schaback show that Kansa’s method cannot be well-posed for arbitrary center locations, it is now an open question to find sufficient conditions on the center locations that guarantee invertibility of the Kansa matrix”, and the situation does not seem to have changed so far. In this paper we make a first step in this direction, proving that unsymmetric collocation matrices with Thin-Plate Splines (without polynomial addition) for the 2D Poisson equation are almost surely nonsingular, when the discretization points are chosen randomly on domains with analytic boundary. Though TPS are not the most adopted option for Kansa collocation, they have been often used in the meshless literature, cf. e.g. [4, 5, 20] with the references therein. One of their most relevant features is that they are scale invariant, thus avoiding the delicate matter of the scaling choice with scale dependent RBF, which is still an active research topic, cf. e.g. [2, 13]. On the other hand, the fact that TPS without polynomial addition can guarantee unisolvence in the interpolation framework has been recently recognized experimentally in [17] and theoretically in [1, 6]. As we shall see, one of the key aspects is that Thin-Plate Splines $\phi(\|P-A\|_{2})$, which correspond to the radial functions $\phi(r)=r^{2\nu}\log(r)\;,\;\;\nu\in\mathbb{N}\;,$ (1) are real analytic functions off their center $A$, due to analyticity of the univariate functions $\log(\cdot)$ and $\sqrt{\cdot}$ in $\mathbb{R}^{+}$. Analiticity together with the presence of a singularity at the center will be the key ingredients of our unisolvence result by random collocation. ## 2 Unisolvence of random Kansa collocation Consider the Poisson equation with Dirichlet boundary conditions (cf. e.g. [7]) $\left\\{\begin{array}[]{l}\Delta u(P)=f(P)\;,\;P\in\Omega\\\ u(P)=g(P)\;,\;P\in\partial\Omega=\gamma([a,b])\;,\end{array}\right.$ (2) where we assume that $\Omega\subset\mathbb{R}^{2}$ is a domain with analytic boundary (a bounded connected open set whose boundary is an analytic curve), namely a curve $\gamma:[a,b]\to\mathbb{R}^{2}\;,\;\gamma(a)=\gamma(b)$, that is analytic and regular (i.e. $\gamma^{\prime}(t)\neq(0,0)$ for every $t\in[a,b]$). In Kansa collocation (see e.g. [8, 10, 11, 14, 18, 19]) one determines a function $u_{N}(P)=\sum_{j=1}^{n}{c_{j}\,\phi_{j}(P)}+\sum_{k=1}^{m}{d_{k}\,\psi_{k}(P)}\;,\;\;N=n+m\;,$ (3) where $\phi_{j}(P)=\phi(\|P-P_{j}\|_{2})\;,\;\;\\{P_{1},\dots,P_{n}\\}\subset\Omega\;,$ (4) $\psi_{k}(P)=\phi(\|P-Q_{k}\|_{2})\;,\;\;\\{Q_{1},\dots,Q_{m}\\}\subset\partial\Omega\;,$ (5) such that $\left\\{\begin{array}[]{l}\Delta u_{N}(P_{i})=f(P_{i})\;,\;i=1,\ldots,n\\\ u_{N}(Q_{h})=g(Q_{h})\;,\;h=1,\ldots,m\;.\end{array}\right.$ (6) The following facts will be used below. Defining $\phi_{A}(P)=\phi(\|P-A\|)$, we have $\phi_{A}(B)=\phi_{B}(A)$ and $\Delta\phi_{A}(B)=\Delta\phi_{B}(A)$. In fact, the Laplacian in polar coordinates centered at $A$ (cf. e.g. [7, Ch.2]) is the radial function $\Delta\phi_{A}={\frac{\partial^{2}\phi}{\partial^{2}r}}+{\frac{1}{r}}\frac{\partial\phi}{\partial r}=4\nu r^{2(\nu-1)}(\nu\log(r)+1)\;.$ (7) Moreover, $\phi_{A}(A)=0$ and $\Delta\phi_{A}(A)=0$ for $\nu\geq 2$, since $\Delta\phi\to 0$ as $r\to 0$. Kansa collocation can be rewritten in matrix form as $\left(\begin{array}[]{cc}\Delta\Phi&\Delta\Psi\\\ \\\ \Phi&\Psi\end{array}\right)\left(\begin{array}[]{c}\mathbf{c}\\\ \\\ \mathbf{d}\end{array}\right)=\left(\begin{array}[]{c}\mathbf{f}\\\ \\\ \mathbf{g}\end{array}\right)$ (8) where the block matrix is $K_{N}=K_{N}(\\{P_{i}\\},\\{Q_{h}\\})=\left(\begin{array}[]{cc}\Delta\Phi&\Delta\Psi\\\ \\\ \Phi&\Psi\end{array}\right)$ $=\left(\begin{array}[]{ccccccc}0&\cdots&\cdots&\Delta\phi_{n}(P_{1})&\Delta\psi_{1}(P_{1})&\cdots&\Delta\psi_{m}(P_{1})\\\ \\\ \vdots&\ddots&&\vdots&\vdots&\cdots&\vdots\\\ \vdots&&\ddots&\vdots&\vdots&\cdots&\vdots\\\ \\\ \Delta\phi_{1}(P_{n})&\cdots&\cdots&0&\Delta\psi_{1}(P_{n})&\cdots&\Delta\psi_{m}(P_{n})\\\ \\\ \phi_{1}(Q_{1})&\cdots&\cdots&\phi_{n}(Q_{1})&0&\cdots&\psi_{m}(Q_{1})\\\ \\\ \vdots&\cdots&\cdots&\vdots&\vdots&\ddots&\vdots\\\ \\\ \phi_{1}(Q_{m})&\cdots&\cdots&\phi_{n}(Q_{m})&\psi_{1}(Q_{m})&\cdots&0\\\ \\\ \end{array}\right)$ and $\textbf{f}=\\{f(P_{i})\\}_{i=1,\ldots,n}$, ${\textbf{g}}=\\{g(Q_{h})\\}_{h=1,\ldots,m}$. We can now state and prove our main result. ###### Theorem 1 Let $K_{N}$ be the TPS Kansa collocation matrix defined above, with $N=n+m\geq 2$, where $\\{P_{i}\\}$ is a sequence of independent uniformly distributed random points in $\Omega$, and $\\{Q_{h}\\}$ a sequence of independent uniformly distributed points on $\partial\Omega$. Namely, $\\{Q_{h}\\}=\\{\gamma(t_{h})\\}$ with $\\{t_{h}\\}$ sequence of independent identically distributed random abscissas in $(a,b)$ with respect to the arclength density $\|\gamma^{\prime}(t)\|_{2}/L$, $L=length(\gamma([a,b]))$. Then for every $N\geq 2$ the matrix $K_{N}$ is a.s. (almost surely) nonsingular. Proof. The proof proceeds by complete induction on $N$. For the induction base, we prove that $\mbox{det}(K_{N})$ is a.s. nonzero for $N=2$, that is for $n=2$ and $m=0$, or $n=0$ and $m=2$, or $n=1$ and $m=1$. In the first case, $\mbox{det}(K_{2})=-\Delta\phi_{2}(P_{1})\Delta\phi_{1}(P_{2})=-(\Delta\phi_{1}(P_{2}))^{2}$ $=-16\nu^{2}\|P_{2}-P_{1}\|_{2}^{4\nu-4}\left(\nu\log(\|P_{2}-P_{1}\|_{2})+1\right)^{2}$ which vanishes iff $P_{2}=P_{1}$ (an event with null probability) or $P_{2}$ falls on (the intersection with $\Omega$ of) the curve $\nu\log(\|P-P_{1}\|_{2})+1=0$, that is on the circle $\|P-P_{1}\|_{2}^{2}=\exp(-2/\nu)\;.$ But this event has null probability, since any algebraic curve is a null set in $\mathbb{R}^{2}$. In the second case, $\mbox{det}(K_{2})=-\psi_{2}(Q_{1})\psi_{1}(Q_{2})=-\psi_{1}^{2}(Q_{2})=-\psi_{1}^{2}(\gamma(t_{2}))\;.$ Now, given $P_{1}=\gamma(t_{1})$, the function $\lambda(t)=\psi_{1}^{2}(\gamma(t))$ is analytic in $(a,t_{1})$ and in $(t_{1},b)$. Then $\psi_{1}^{2}(\gamma(t_{2}))$ is zero iff $t_{2}=t_{1}$ (an event that has null probability), or $t_{2}$ falls on the zero set of $\lambda$ in $(a,t_{1})$ or $(t_{1},b)$. Again this event has null probability since the zero set of an univariate analytic function in an open interval is a null set (cf. [12, 15]). As for the third case, assume that $Q_{1}$ is chosen on the boundary (randomly or not) and that $P_{1}$ is chosen randomly in the interior. Since $\mbox{det}(K_{2})=-\phi_{1}(Q_{1})\Delta\psi_{1}(P_{1})$ $=-4\nu\phi_{1}(Q_{1})\|P_{1}-Q_{1}\|_{2}^{2\nu-2}\left(\nu\log(\|P_{1}-Q_{1}\|_{2})+1\right)\;,$ and $\phi_{1}(Q_{1})\neq 0$ being $P_{1}\neq Q_{1}$, the determinant vanishes if and only if $P_{1}$ falls on (the intersection with $\Omega$ of) the curve $\nu\log(\|P-Q_{1}\|_{2})+1=0$, that is on the circle $\|P-Q_{1}\|_{2}^{2}=\exp(-2/\nu)$ and again this event has null probability. For the inductive step, we consider separately the case where a boundary point is added, for which we define the matrix $U(P)=\left(\begin{array}[]{cccccccc}0&\cdots&\cdots&\Delta\phi_{n}(P_{1})&\Delta\psi_{1}(P_{1})&\cdots&\Delta\psi_{m}(P_{1})&\Delta\phi_{1}(P)\\\ \\\ \vdots&\ddots&&\vdots&\vdots&\cdots&\vdots&\vdots\\\ \vdots&&\ddots&\vdots&\vdots&\cdots&\vdots&\vdots\\\ \\\ \Delta\phi_{1}(P_{n})&\cdots&\cdots&0&\Delta\psi_{1}(P_{n})&\cdots&\Delta\psi_{m}(P_{n})&\Delta\phi_{n}(P)\\\ \\\ \phi_{1}(Q_{1})&\cdots&\cdots&\phi_{n}(Q_{1})&0&\cdots&\psi_{m}(Q_{1})&\psi_{1}(P)\\\ \\\ \vdots&\cdots&\cdots&\vdots&\vdots&\ddots&\vdots&\vdots\\\ \\\ \phi_{1}(Q_{m})&\cdots&\cdots&\phi_{n}(Q_{m})&\psi_{1}(Q_{m})&\cdots&0&\psi_{m}(P)\\\ \\\ \phi_{1}(P)&\cdots&\cdots&\phi_{n}(P)&\psi_{1}(P)&\cdots&\psi_{m}(P)&0\\\ \\\ \end{array}\right)$ Observe that in this case $K_{N+1}=U(Q_{m+1})$. Indeed, $\psi_{k}(Q_{h})=\psi_{h}(Q_{k})$ and $\Delta\phi_{i}(Q_{m+1})=\Delta\psi_{m+1}(P_{i})$. Differently, if an interior point is added, we define the matrix $V(P)=\left(\begin{array}[]{cccccccc}0&\cdots&\cdots&\Delta\phi_{n}(P_{1})&\Delta\phi_{1}(P)&\Delta\psi_{1}(P_{1})&\cdots&\Delta\psi_{m}(P_{1})\\\ \\\ \vdots&\ddots&&\vdots&\vdots&\vdots&\cdots&\vdots\\\ \vdots&&\ddots&\vdots&\vdots&\vdots&\cdots&\vdots\\\ \\\ \Delta\phi_{1}(P_{n})&\cdots&\cdots&0&\Delta\phi_{n}(P)&\Delta\psi_{1}(P_{n})&\cdots&\Delta\psi_{m}(P_{n})\\\ \\\ \Delta\phi_{1}(P)&\cdots&\cdots&\Delta\phi_{n}(P)&0&\Delta\psi_{1}(P)&\cdots&\Delta\psi_{m}(P)\\\ \\\ \phi_{1}(Q_{1})&\cdots&\cdots&\phi_{n}(Q_{1})&\psi_{1}(P)&0&\cdots&\psi_{m}(Q_{1})\\\ \\\ \vdots&\cdots&\cdots&\vdots&\vdots&\vdots&\ddots&\vdots\\\ \\\ \phi_{1}(Q_{m})&\cdots&\cdots&\phi_{n}(Q_{m})&\psi_{m}(P)&\psi_{1}(Q_{m})&\cdots&0\\\ \\\ \end{array}\right)$ Observe that in this case $K_{N+1}=V(P_{n+1})$ since $\psi_{k}(P_{n+1})=\phi_{n+1}(Q_{k})$ and $\Delta\phi_{j}(P_{i})=\Delta\phi_{i}(P_{j})$. Concerning the determinants, applying Laplace determinantal rule on the last row of $U(P)$ we see that for every $\ell$, $1\leq\ell\leq m$, we get the representation $F(P)=\mbox{det}(U(P))=\delta_{N-1}\psi_{\ell}^{2}(P)+A(P)\psi_{\ell}(P)+B(P)$ (9) where $|\delta_{N-1}|=|\mbox{det}(K_{N-1}(\\{P_{i}\\},\\{Q_{h}\\}_{h\neq\ell}))|$ $A\in\mbox{span}\\{\phi_{j},\Delta\phi_{j},\psi_{k}\,;\,1\leq j\leq n\,,\,1\leq k\leq m\,,\,k\neq\ell\\}$ $B\in\mbox{span}\\{\phi_{i}\Delta\phi_{j},\psi_{k}\phi_{i},\psi_{k}\Delta\phi_{i},\psi_{k}\psi_{h}\,;\,1\leq i,j\leq n\,,\,1\leq k,h\leq m\,,\,k,h\neq\ell\\}\;.$ Similarly, developing $det(V(P))$ by the $(n+1)$-row we have $G(P)=\mbox{det}(V(P))=-\mbox{det}(K_{N-1})(\Delta\phi_{n}(P))^{2}+C(P)\Delta\phi_{n}(P)+D(P)$ (10) where $C\in\mbox{span}\\{\Delta\phi_{j},\psi_{k},\Delta\psi_{k}\,;\,1\leq j\leq n-1\,,\,1\leq k\leq m\\}$ $D\in\mbox{span}\\{\Delta\phi_{i}\Delta\phi_{j},\Delta\phi_{i}\Delta\psi_{h},\psi_{k}\Delta\phi_{i},\psi_{k}\Delta\psi_{h}\,;\,1\leq i,j\leq n-1\,,\,1\leq k,h\leq m\\}\;.$ First, we prove that $G$ is not identically zero in $\Omega$ if $\mbox{det}(K_{N-1})\neq 0$ (the latter a.s. holds by inductive hypothesis). Let $P(t)=P_{n}+t(1,0)$, $t\in\mathbb{R}$, and $r(t)=\|P(t)-P_{n}\|_{2}=|t|$. If $G\equiv 0$ then $G(P(t))\equiv 0$ in neighborhood of $t=0$. Then, we would locally have $u^{2}(t)=c(t)u(t)+d(t)\;,\;\;u(t)=\Delta\phi_{n}(P(t))\;,$ (11) where $c(t)=C(P(t))/\mbox{det}(K_{N-1})$ and $d(t)=D(P(t))/\mbox{det}(K_{N-1})$. Notice that both $c$ and $d$ are analytic in a neighborhood of $t=0$, since $C$ and $D$ are analytic in a neighborhood of $P_{n}$. By (11) and (7) we get $u(t)=4\nu t^{2(\nu-1)}\left(\nu\log(|t|)+1\right)\;.$ (12) Clearly $c$ cannot be identically zero there, otherwise $u^{2}$ would be analytic at $t=0$ and thus would have an algebraic order of infinitesimal as $t\to 0$, whereas by (12) we have $u^{2}(t)\sim 16\nu^{4}t^{4(\nu-1)}\log^{2}(|t|)$. Hence taking the Maclaurin expansion of $c$ we get $c(t)\sim c_{s}t^{s}$ as $t\to 0$ for some $s\geq 0$, the order of the first nonvanishing derivative at $t=0$. Now, $u^{2}(t)\sim 16\nu^{4}t^{4(\nu-1)}\log^{2}(|t|)$, whereas by $u^{2}\equiv cu+d$ we would have $u^{2}(t)\sim 4\nu^{2}c_{s}t^{s+2(\nu-1)}\log(|t|)+d_{p}t^{p}$, where either $d(0)\neq 0$ and $p=0$, or $d(0)=0$ and $p>0$ (the order of the first nonvanishing derivative at $t=0$). Then we get a contradiction, since $u^{2}$ cannot have two distinct limits or orders of infinitesimal at the same point. Moreover, $G$ is clearly continuous in $\Omega$ and analytic in $\Omega\setminus\\{P_{1},\dots,P_{n}\\}$, since all the functions involved in its definition (10) are analytic up to their own center. Consequently, if $\mbox{det}(K_{N-1})\neq 0$ by continuity $G$ is not identically zero also in $\Omega\setminus\\{P_{1},\dots,P_{n}\\}$. Then, $\mbox{det}(K_{N+1})=\mbox{det}(V(P_{n+1}))=G(P_{n+1})$ is a.s. nonzero, since the zero set of a not identically zero real analytic function on an open connected set in $\mathbb{R}^{d}$ is a null set (cf. [15] for an elementary proof). More precisely, denoting by $Z_{G}$ the zero set of $G$ in $\Omega$, we have that $Z_{G}=(Z_{G}\cap\\{P_{1},\dots,P_{n}\\})\cup(Z_{G}\cap(\Omega\setminus\\{P_{1},\dots,P_{n}\\}))\;.$ Hence $Z_{G}$ is a null set if $G\not\equiv 0$, because the first intersection is a finite set, and the second is the zero set of a not identically zero real analytic function. Considering the probability of the corresponding events and recalling that $\mbox{det}(K_{N-1})\neq 0$ (which a.s. holds) implies $G\not\equiv 0$, we can then write $\mbox{prob}\\{\mbox{det}(K_{N+1})=0\\}=\mbox{prob}\\{G(P_{n+1})=0\\}$ $=\mbox{prob}\\{G\equiv 0\\}+\mbox{prob}\\{G\not\equiv 0\;\&\;P_{n+1}\in Z_{G}\\}=0+0=0\;,$ and this branch of the inductive step is completed. We turn now to the branch of the inductive step where a boundary point is added. In this case we consider the function $F$ in (9) restricted to the boundary, that is $F(P(t))$ with $P(t)=\gamma(t)$, $t\in(a,b)$, which for every fixed $\ell\in\\{1,\dots,m\\}$ has the representation $F(\gamma(t))=\mbox{det}(U(\gamma(t)))=\delta_{N-1}v^{2}(t)+A(\gamma(t))v(t)+B(\gamma(t))$ where $v(t)=\psi_{\ell}(\gamma(t))=r_{\ell}^{2\nu}(t)\log(r_{\ell}(t))\;,\;\;r_{\ell}(t)=\|\gamma(t)-Q_{\ell}\|_{2}$ (13) with $Q_{\ell}=\gamma(t_{\ell})$, $\;t_{\ell}\in(a,b)$. We claim that if $\delta_{N-1}\neq 0$ (which a.s. holds by inductive hypothesis), $F\circ\gamma$ cannot be identically zero in any of the two connected components of $(a,b)\setminus\\{t_{1},\dots,t_{m}\\}$ (i.e., the subintervals) having $t_{\ell}$ as extremum. Otherwise, we would have in a left or in a right neighborhood of $t_{\ell}$ $v^{2}(t)=\alpha(t)v(t)+\beta(t)\;,$ (14) where $\alpha(t)=A(\gamma(t))/\delta_{N-1}$ and $\beta(t)=B(\gamma(t))/\delta_{N-1}$ are both analytic in a full neighborhood of $t_{\ell}$. Notice that, since $\gamma^{\prime}(t_{\ell})\neq(0,0)$ (the curve is regular), $r_{\ell}(t)\sim\|\gamma^{\prime}(t_{\ell})\|_{2}|t-t_{\ell}|$ which by (13) gives $v(t)\sim\|\gamma^{\prime}(t_{\ell})\|_{2}^{2\nu}(t-t_{\ell})^{2\nu}\l og(|t-t_{\ell}|)$ and $v^{2}(t)\sim\|\gamma^{\prime}(t_{\ell})\|_{2}^{4\nu}(t-t_{\ell})^{4\nu}\log^{2}(|t-t_{\ell}|)$ as $t\to t_{\ell}$. Now $\alpha$ cannot be identically zero in any left or right neighborhood, otherwise $v^{2}\equiv\beta$ there and would have an algebraic order of infinitesimal at $t_{\ell}$. Hence taking the Taylor expansion of $\alpha$ we get $\alpha(t)\sim\alpha_{s}(t-t_{\ell})^{s}$ as $t\to t_{\ell}$ for some $s\geq 0$, the order of the first nonvanishing derivative at $t=t_{\ell}$. On the other hand, by $v^{2}\equiv\alpha v+\beta$ locally, we would have $v^{2}(t)\sim\|\gamma^{\prime}(t_{\ell})\|_{2}^{2\nu}\alpha_{s}(t-t_{\ell})^{s+2\nu}\log(|t-t_{\ell}|)+\beta_{p}(t-t_{\ell})^{p}$, where either $\beta(t_{\ell})\neq 0$ and $p=0$, or $\beta(t_{\ell})=0$ and $p>0$ (the order of the first nonvanishing derivative at $t=t_{\ell}$). Again we get a contradiction, since $v^{2}$ cannot have two distinct limits or orders of infinitesimal at the same point. The result is that $F\circ\gamma$ is a.s. not identically zero in any connected component of $(a,b)\setminus\\{t_{1},\dots,t_{m}\\}$. Then, $\mbox{det}(K_{N+1})=\mbox{det}(U(Q_{m+1}))=F(\gamma(t_{m+1}))$ is a.s. nonzero. In fact, observe that $F\circ\gamma$ is analytic in $(a,b)\setminus\\{t_{1},\dots,t_{m}\\}$, since $F$ is analytic in $\mathbb{R}^{2}\setminus(\\{Q_{1},\dots,Q_{m}\\}\cup\\{P_{1},\dots,P_{n}\\})$. Moreover, denoting by $Z_{F\circ\gamma}$ the zero set of $F\circ\gamma$ in $(a,b)$, we have that $Z_{F\circ\gamma}=(Z_{F\circ\gamma}\cap\\{t_{1},\dots,t_{m}\\})\cup(Z_{F\circ\gamma}\cap((a,b)\setminus\\{t_{1},\dots,t_{m}\\}))\;.$ Hence $Z_{F\circ\gamma}$ is a null set if ${F\circ\gamma}\not\equiv 0$, because the first intersection is a finite set, and the second is the componentwise finite union of the zero sets of a not identically zero real analytic function on each connected component. Considering the probability of the corresponding events and recalling that $\mbox{det}(K_{N-1})\neq 0$ (which a.s. holds) implies $F\circ\gamma\not\equiv 0$, we can then write $\mbox{prob}\\{\mbox{det}(K_{N+1})=0\\}=\mbox{prob}\\{F(Q_{m+1})=0\\}$ $=\mbox{prob}\\{F\circ\gamma\equiv 0\\}+\mbox{prob}\\{F\circ\gamma\not\equiv 0\;\&\;t_{m+1}\in Z_{F\circ\gamma}\\}=0+0=0\;,$ and also the boundary branch of the inductive step is completed. $\square$ ### 2.1 Remarks on possible extensions The result of Theorem 1 is a first step towards a theory of Kansa collocation unisolvence, and could be extended in several directions within the random framework. The first extension comes immediately from the fact that a null set has also measure zero for any continuous measure with density (that is, absolutely continuous with respect to the Lebesgue measure). We can state indeed the following ###### Theorem 2 The assertion of Theorem 1 holds true if the points $\\{P_{i}\\}$ are independent identically distributed with respect any continuous probability measure with density on $\Omega$, say $\sigma\in L^{1}_{+}(\Omega)$, and the abscissas $\\{t_{h}\\}$ are independent identically distributed with respect any continuous probability measure with density on $(a,b)$, say $w\in L^{1}_{+}(a,b)$. This extension could be interesting whenever it is known that the solution has steep gradients or other regions where it is useful to increase the discretization density. Concerning the implementation of random sampling with respect to continuous probability densities, we recall the well-known “acceptance-rejection method”, cf. e.g. [3, 9, 16] with the references therein. More difficult but worth of further investigations are: * • extension to $\Omega\subset\mathbb{R}^{d}$, $d\geq 3$; * • extension to other analitic RBF up to the center, e.g. Radial Powers; * • extension to piecewise analytic boundaries; * • extension to other differential operators and/or boundary conditions. The latter in particular could be challenging, since the operators involved in the equation and in the boundary conditions may not be radial. Acknowledgements. Work partially supported by the DOR funds of the University of Padova, and by the INdAM-GNCS 2024 Projects “Kernel and polynomial methods for approximation and integration: theory and application software”. This research has been accomplished within the RITA “Research ITalian network on Approximation” and the SIMAI Activity Group ANA&A, and the UMI Group TAA “Approximation Theory and Applications”. ## References * [1] L.P. Bos, A. Sommariva, M. Vianello, A note on polynomial-free unisolvence of polyharmonic splines at random points. https://arxiv.org/abs/2312.13710 * [2] R. Cavoretto, A. De Rossi, Adaptive procedures for meshfree RBF unsymmetric and symmetric collocation methods, Appl. Math. Comput. 382 (2020), 125354. * [3] A. Chalkis, C. Katsamaki, J. Tonelli-Cueto, On the Error of Random Sampling: Uniformly Distributed Random Points on Parametric Curves, ISSAC ’22: Proceedings of the 2022 Intern. Symp. on Symb. and Alg. Comput., July 2022, pp. 273–282. https://arxiv.org/pdf/2203.02832.pdf * [4] W. Chen, Z.-J. Fu, C.S. Chen, Different Formulations of the Kansa Method: Domain Discretization, in W. Chen et al., Recent Advances in Radial Basis Function Collocation Methods, SpringerBriefs in Applied Sciences and Technology, 2014. * [5] P.P. Chinchapatnam, K. Djidjeli, P.B. Nair, Unsymmetric and symmetric meshless schemes for the unsteady convection–diffusion equation, Comput. Methods Appl. Mech. Engrg. 195 (2006), 2432–-2453. * [6] F. Dell’Accio, A. Sommariva, M. Vianello, Random sampling and unisolvent interpolation by almost everywhere analytic functions, Appl. Math. Lett. 145 (2023). * [7] L.C. Evans, Partial Differential Equations, Graduate Studies in Mathematics 19, AMS, 1998. * [8] G.E. Fasshauer, Meshfree Approximation Methods with Matlab, Interdisciplinary Mathematical Sciences, Vol. 6, World Scientific, 2007. * [9] B.D. Flury, Acceptance-Rejection Sampling Made Easy, SIAM Rev. 32 (1990), 474–476. * [10] Y.C. Hon, R. Schaback, On unsymmetric collocation by radial basis functions, Appl. Math. Comput. 119 (2001), 177–-186. * [11] E.J. Kansa, Application of Hardy’s multiquadric interpolation to hydrodynamics, in Proc. 1986 Simul. Conf., Vol. 4, pp. 111–-117. * [12] S.G. Krantz, H.R. Parks, A Primer of Real Analytic Functions, Second Edition, Birkhäuser, Boston, 2002. * [13] E. Larsson, R. Schaback, Scaling of radial basis functions, IMA Journal of Numerical Analysis (2023), drad035. * [14] L. Ling, R. Opfer, R. Schaback, Results on meshless collocation techniques, Eng. Anal. Bound. Elem. 30 (2006), 247–253. * [15] B.S. Mityagin, The Zero Set of a Real Analytic Function, Math. Notes 107 (2020), 529–530. * [16] N. Nguyen, G. Ökten, The acceptance-rejection method for low-discrepancy sequences, Monte Carlo Methods Appl. 22 (2016), 133–148. * [17] A. Pasioti, On the Constrained Solution of RBF Surface Approximation, MDPI Mathematics 10 (2022), 2582. * [18] R. Schaback, H. Wendland, Kernel techniques: From machine learning to meshless methods, Acta Numer. 15 (2006) 543–639. * [19] H. Wendland, Scattered Data Approximation, Cambridge Monogr. Appl. Comput. Math., vol. 17, Cambridge Univ. Press, Cambridge, 2005. * [20] M. Zerroukat, K. Djidjeli, A. Charafi, Explicit and implicit meshless methods for linear advection–diffusion-type partial differential equations, Int. J. Numer. Methods Eng. 48 (2000), 19–35.
# Conversion efficiency in Kerr microresonator optical parametric oscillators: From three modes to many modes Jordan R. Stone<EMAIL_ADDRESS>Joint Quantum Institute, NIST/University of Maryland, College Park, MD 20742 National Institute for Standards and Technology, Gaithersburg, MD 20899 Gregory Moille Joint Quantum Institute, NIST/University of Maryland, College Park, MD 20742 National Institute for Standards and Technology, Gaithersburg, MD 20899 Xiyuan Lu National Institute for Standards and Technology, Gaithersburg, MD 20899 Institute for Research in Electronics and Applied Physics and Maryland NanoCenter, University of Maryland, College Park, MD 20742, USA Kartik Srinivasan Joint Quantum Institute, NIST/University of Maryland, College Park, MD 20742 National Institute for Standards and Technology, Gaithersburg, MD 20899 ###### Abstract We study optical parametric oscillations in Kerr-nonlinear microresonators, revealing an intricate solution space – parameterized by the pump-to-signal conversion efficiency – that arises from an interplay of nonlinear processes. Using a three-mode approximation, we derive an efficiency-maximizing relation between pump power and frequency mismatch. To move beyond a three-mode approximation, a necessity for geometries such as integrated microring resonators, we numerically simulate the Lugiato-Lefever Equation that accounts for the full spectrum of nonlinearly-coupled resonator modes. We observe and characterize two nonlinear phenomena linked to parametric oscillations in multi-mode resonators: Mode competition and cross phase modulation-induced modulation instability. Both processes may impact conversion efficiency. Finally, we show how to increase the conversion efficiency by tuning the microresonator loss rates. Our analysis will guide microresonator designs that aim for high conversion efficiency and output power. ## I Introduction Integrated photonics offers scalable options for generating, processing, and routing optical signals within classical and quantum networks Liang and Bowers (2010); Agrell _et al._ (2016); Sipahigil _et al._ (2016); Elshaari _et al._ (2020); Jin _et al._ (2021). In general, optical processors apply linear and/or nonlinear operations to light. A notable case is the optical microresonator, whose small size and large quality factor ($Q$) work to intensify circulating light and promote efficient nonlinear interactions Vahala (2003); Yang _et al._ (2018). Indeed, microresonators host a veritable zoo of nonlinear eigenstates, including soliton frequency combs Kippenberg _et al._ (2018), Raman frequency combs Liu _et al._ (2018), Hz-linewidth lasers based on stimulated Brilluoin scattering Loh _et al._ (2015), $\chi^{(2)}$ and $\chi^{(3)}$-type parametric oscillators Bruch _et al._ (2019); Lu _et al._ (2019), and more for applications in communications, timekeeping, and sensing Spencer _et al._ (2018); Marin-Palomo _et al._ (2017); Lai _et al._ (2020); Newman _et al._ (2019). Many of the experiments cited above were motivated by a high demand for coherent light sources on a chip. One important type of coherent source is the optical parametric oscillator (OPO), which is often employed to reach wavelengths not directly accessible by conventional laser gain Vodopyanov _et al._ (2000); Mieth _et al._ (2014). Optical parametric oscillations occur in $\chi^{(3)}$-nonlinear media when vacuum fluctuations are amplified by stimulated four wave mixing (FWM), if the FWM gain exceeds the resonator losses Boyd (2020). Degenerately-pumped OPOs are a special case in which two frequency-degenerate pump photons are converted into one higher-frequency signal photon and one lower-frequency idler photon. In principle, a degenerately-pumped OPO can generate coherent light within the frequency range $\textrm{DC}-2\omega_{\rm{p}}$, where $\omega_{\rm{p}}$ is the pump laser frequency Lu and Srinivasan (2021). Hence, a chip-scale, degenerately-pumped OPO could offer superior scalability, higher efficiencies, and broader spectral coverage than alternatives. It would be readily implemented in miniaturized technologies, from optical clocks in which the OPO could be tuned to clock-type or cooling-type atomic transitions, to quantum processors in which the OPO could be tuned to qubit frequencies. Figure 1: Introduction to the microresonator-based, degenerately-pumped optical parametric oscillator ($\mu$OPO). (a) Schematic of a microring resonator coupled to an access waveguide that carries the input and output fields. (b) Energy diagram for the degenerate four wave mixing (FWM) process that drives parametric oscillation. (c) Depictions of mode spectra, nonlinear couplings, and frequency shifts in both a three-mode approximation (TMA) and multi-moded model. Dashed lines correspond to zero frequency mismatch. Red arrows indicate mode frequency shifts induced by Kerr nonlinearity. The TMA considers FWM between only the pump, signal, and idler modes, while multi- moded models account for the nonlinear couplings (orange, hollow arrows) between all mode sets conducive to FWM. Recently, several experiments have been reported that advance the microresonator-based, degenerately-pumped OPO ($\mu$OPO) and make real headway towards a chip-scale, wavelength-by-design light source. Achievements include sub-milliwatt oscillation thresholds Lu _et al._ (2019), octave-spanning and tunable spectra Sayson _et al._ (2019); Tang _et al._ (2020), visible-light generation spanning red to green Lu _et al._ (2020); Domeneguetti _et al._ (2021), and a $\mu$OPO that uses a 2D photonic crystal cavity Marty _et al._ (2021). Nonetheless, the reported pump-to-signal conversion efficiencies are typically (except in instances involving narrow spectral bandwidths) $<0.1$% – a nonstarter for applications Lu _et al._ (2020); Sayson _et al._ (2019); Domeneguetti _et al._ (2021); Tang _et al._ (2020). Indeed, while the demand for efficiency calls for a deeper understanding of the underlying nonlinear physics, experiments have so far relied on a simplified theoretical framework for $\mu$OPOs. For instance, frequency matching is considered in either a cold-cavity limit Lu _et al._ (2019), or otherwise only accounts for a populated pump mode Sayson _et al._ (2019). A more accurate description of frequency matching should account for the exact distribution of intraresonator photons. Moreover, analyses have relied on a three-mode approximation (TMA), in which only the pump, signal, and idler modes interact through Kerr nonlinearity Sayson _et al._ (2019); Hansson _et al._ (2013). Of course, real microresonators comprise a more complex spectrum of modes that are nonlinearly coupled together. As a result, there is presently a gap between theoretical and experimental progress revolving around $\mu$OPOs; ultimately, there is little theoretical basis on which to design microresonators to meet end-user demands. Here, we construct a generalized $\mu$OPO solution space; thereby, we reveal connections between the $\mu$OPO state and experimental parameters, and we identify processes that limit conversion efficiency. We adopt a model based on the Lugiato-Lefever Equation (LLE) Coen _et al._ (2013); Chembo and Menyuk (2013) and support our main numerical results with theoretical analyses. In the next section, we explain our modeling and present simulation results using a TMA. Then, we expand the model to include a spectrum of nonlinearly-coupled resonator modes. We demonstrate two nonlinear phenomena that cannot be explained within a TMA. In the first, a mode competition takes place between multiple signal and idler mode pairs. In the second, modulation instability induced by cross-phase modulation constrains the $\mu$OPO conversion efficiency. Finally, we propose two strategies for increasing the $\mu$OPO conversion efficiency and output power. Surprisingly, when comparing microresonators with different loss rates but identical geometries, we find that the resonator with greater losses will, in some cases, promote higher efficiency. ## II Modeling the $\mu$OPO: The Lugiato-Lefever Equation, three-mode approximation, and dispersion To study $\mu$OPOs, we consider a microring resonator coupled to an optical access waveguide and pumped by a continuous-wave (CW) laser, as depicted in Fig. 1a. This structure supports whispering gallery modes, in which azimuthal modes are grouped into families sharing a transverse spatial mode profile. Modes within a family are spaced (in the frequency domain) by a free-spectral range (FSR) that is inversely proportional to the ring circumference, $L$. In our model, we consider a single mode family and denote its resonant frequencies as $\omega_{\rm{\mu}}$, where $\mu$ is the azimuthal mode number shifted to make $\omega_{\rm{0}}$ the frequency of the pumped mode. The pump laser has frequency $\omega_{\rm{p}}$ and waveguide power $P_{\rm{in}}$, and the intraresonator field, $a$, obeys the Lugiato-Lefever Equation (LLE) Chembo and Menyuk (2013): ${}\frac{da}{dt}=\sqrt{\frac{\kappa_{\rm{c}}(0)}{\hbar\omega_{\rm{p}}}P_{\rm{in}}}-\left(\frac{\kappa_{i}}{2}+i\frac{\kappa(0)}{2}\alpha- ig_{0}|a|^{2}\right)a-i\mathcal{D}(\mu)\tilde{a},$ (1) where $|a|^{2}$ gives the intraresonator energy in units of photon number, $\kappa_{\rm{c}}(\mu)$ is the mode-dependent coupling rate to the access waveguide, $\kappa_{\rm{i}}$ is the mode-independent intrinsic loss rate, $\kappa(\mu)=\kappa_{\rm{i}}+\kappa_{\rm{c}}(\mu)$ is the mode-dependent total loss rate, $\alpha=\frac{\omega_{\rm{0}}-\omega_{\rm{p}}}{\kappa(0)/2}$ is the normalized pump-resonator frequency detuning, $g_{\rm{0}}=\frac{n_{\rm{2}}c\hbar\omega_{\rm{0}}^{2}}{n^{2}V}$ is the nonlinear gain per photon, $n_{\rm{2}}$ is the Kerr index, $c$ is the speed of light in vacuum, $n$ is the refractive index, $V$ is the mode volume, and $\mathcal{D}(\mu)=\omega_{\rm{\mu}}-(\omega_{\rm{0}}+\mu D_{\rm{1}})+i\kappa_{\rm{c}}(\mu)$, where $D_{\rm{1}}=2\pi\times\textrm{FSR}$; $\tilde{a}$ indicates that operations to $a$ are performed in the frequency domain. Notably, the integrated dispersion, $D_{\rm{int}}$, is contained in Eq. 1 as $D_{\rm{int}}=Re(\mathcal{D})/\kappa(0)$. For concreteness, we use $n_{\rm{2}}=2.4\times 10^{-19}$ m2/W and $n=1.9$, which are typical values for silicon nitride (SiN) microrings, and $\omega_{\rm{p}}\approx 2\pi\times 384$ THz ($\approx$ 780 nm wavelength). Unless otherwise stated, we use $\kappa_{\rm{i}}=2\pi\times 200$ MHz and $\kappa_{\rm{c}}=\kappa_{\rm{i}}$ (i.e., critical coupling). Figure 2: Study of the $\mu$OPO conversion efficiency using a three-mode approximation. (a) Conversion efficiency, $CE$ (gray), and effective mismatch, $\Delta_{s}$ (green), versus the pump-resonator frequency detuning, $\alpha$, for a pump laser power $P_{\rm{in}}=20~{}\rm{mW}$. From top to bottom, the mismatch values are $\delta=1.25,2,4$, and $7$, respectively. Dashed sections indicate unstable or oscillatory solutions. (b) A generalized $CE$ map showing the relationship between maximum $CE$, normalized pump power, $X$, and $\delta$. The white dashed line is expressed as $X=8\delta$ and closely follows the contour of highest maximum $CE$. (c) $CE$ (gray) and effective pump-resonator detuning, $\Delta_{\rm{p}}$ (green), versus $\alpha$. (d) $\Delta_{\rm{p}}$ versus $\delta$ at values of $\alpha$ that maximize $CE$ for $P_{\rm{in}}=15$ mW (red, $X\approx 5$) and $P_{\rm{in}}=25$ mW (orange, $X\approx 8$). Figure 1c introduces some key concepts and illustrates differences between the TMA and the multi-mode LLE (here, ’multi-mode’ refers to the inclusion of many longitudinal modes from the same spatial mode family). The main difference is that a multi-mode model accounts for many more modes that can be coupled together by the Kerr nonlinearity. In contrast, a TMA only considers nonlinear interactions (e.g., FWM) between the pump, signal, and idler modes. Importantly, both models account for imperfect frequency matching. In general, phase-matched mode pairs (i.e., with azimuthal numbers $\pm\mu$) are not frequency matched – the associated FWM process does not conserve energy. Imperfect frequency matching is often quantified by the frequency mismatch parameter ${}\delta_{\rm{\mu}}=\frac{\omega_{\mu}+\omega_{-\mu}-2\omega_{\rm{0}}}{\kappa(0)}.$ (2) Throughout, we use $\delta_{\rm{\mu}}$ to refer to the dispersive mismatch spectrum and $\delta$ to refer to the value of $\delta_{\rm{\mu(-\mu)}}$ at the targeted signal (idler) mode. A degenerate FWM process only conserves energy and momentum if $\delta$ is compensated by nonlinear frequency shifts of $\omega_{\rm{p(s,i)}}$, which we denote as $N_{\rm{p(s,i)}}$ for the pump (signal, idler) mode. Nonlinear shifts arise from self- and cross-phase modulation (SPM and XPM, respectively) and are related to the intraresonator intensity spectrum. Indeed, an occupied resonator necessarily implies $N_{p(s,i)}<0$; therefore, $\delta=0$ is not conducive to parametric oscillation. Using a TMA, we investigate how changes in $\delta$ impact the $\mu$OPO, and then we introduce other modes into our simulations. To ensure the consistency of our numerical methods, we use the split-step Fourier method to simulate Eq. 1 for both the TMA and multi-mode LLE. Our primary goal is to understand the efficiency with which pump photons are converted into signal or idler photons. Accordingly, we define the conversion efficiency in units of photon flux as ${}CE=\frac{\omega_{\rm{p}}}{\omega_{\rm{s(i)}}}\frac{P_{\rm{s(i)}}}{P_{\rm{in}}},$ (3) where $\omega_{\rm{s(i)}}$ is the signal (idler) frequency and $P_{\rm{s(i)}}$ is the signal (idler) output power in the access waveguide. Importantly, the underlying symmetry of the FWM process implies that the signal and idler fields have equal $CE$ values and that $N_{\rm{s}}=N_{\rm{i}}$ for a critically-coupled resonator. Throughout, we present $CE$ for the signal; the same results apply to the idler. In Fig. 2, we present simulation results obtained using a TMA. In a single simulation, $P_{\rm{in}}$ and $\delta$ are held fixed; we vary $\alpha$ to simulate a $\omega_{\rm{p}}$ scan across resonance from blue to red detuning, as shown in Fig. 2a. During the simulation, we record various data, including the $CE$, $N_{\rm{p(s,i)}}$, and optical spectra (see Supplemental Material for details). In Fig. 2a, each panel corresponds to $P_{\rm{in}}=20$ mW, but $\delta$ is increased from top to bottom. Our results indicate a highest obtainable $CE$ of $12.5$ % for critically-coupled resonators, in agreement with Ref. Sayson _et al._ (2019). Additionally, each panel depicts the effective signal detuning, defined as $\Delta_{\rm{s}}=\delta+\frac{\alpha}{2}-N_{\rm{s}}$. Notably, when $CE>0$, $\Delta_{\rm{s}}\approx 0$, which indicates that dispersion is perfectly compensated by detuning and nonlinearity in the $\mu$OPO state. To fully characterize the relationship between $CE$, $P_{\rm{in}}$, and $\delta$, we construct the universal $CE$ map shown in Fig. 2b. The parameter space is defined by $P_{\rm{in}}$ and $\delta$; in Fig. 2b, we normalize $P_{\rm{in}}$ as ${}X=\frac{P_{\rm{in}}}{P_{\rm{th}}},$ (4) where $P_{\rm{th}}=\frac{\hbar\omega_{\rm{0}}\kappa^{3}(0)}{8g_{\rm{0}}\kappa_{\rm{c}}(0)}$ is the oscillation threshold power Kippenberg _et al._ (2004). Each data pixel in the $CE$ map represents the maximum $CE$ value taken from the corresponding LLE simulation, as indicated by the dashed line connecting Figs. 2a and 2b. The $CE$ map has a few notable features. First, we do not observe parametric oscillation for any values of $P_{\rm{in}}$ when $\delta\leq 0$. This indicates that nonlinear frequency shifts of $\omega_{\rm{\mu}}$ inhibit frequency matching even when $X\approx 1$. Second, $CE$ contours follow clear trends through the parameter space. In particular, to maintain $CE$ at larger values of $X$, $\delta$ must be increased, apparently to compensate for larger $N_{s(p,i)}$. Remarkably, we can derive an analytical expression for the contour of highest $CE$. We provide the derivation in the Supplemental Material; here, we present the final result, $X=8\delta$, and indicate it with the white dashed line in Fig. 2b. Clearly, to maximize $CE$ for a given $P_{\rm{in}}$, the microresonator dispersion (i.e., $\delta$) should be designed appropriately. Figure 3: Depictions of microresonator dispersion. (a) $\delta_{\rm{\mu}}$ (green circles), and integrated dispersion, $D_{\rm{int}}$ (purple stripe), for the TE1 mode family calculated using finite-element method eigenfrequency simulations. The microresonator has ring radius $RR=15$ $\mu$m, ring width $RW=800$ nm, height $H=600$ nm, and pump frequency $\omega_{\rm{p}}=2\pi\times 388$ THz. Inset: Illustration of the test device cross section. (b) Group velocity refractive index, $n_{\rm{g}}$, versus $\omega/2\pi$ for the same device as in (a). The dashed lines illustrate how a turning point in $\delta_{\rm{\mu}}$ occurs when $n_{\rm{g}}(-\mu)=n_{\rm{g}}(\mu)$. In Fig. 2a, it is perhaps remarkable that $CE$ values tend to increase monotonically until an abrupt cutoff. To elucidate this observation, we monitor the effective pump detuning, $\Delta_{\rm{p}}=\frac{\alpha}{2}-N_{\rm{p}}$, during our simulations and present a sample result in Fig. 2c. When $CE$ is greatest, $\Delta_{\rm{p}}$ reaches a minimum value near zero. Physically, this condition implies that $\omega_{\rm{p}}$ is nearly resonant; intuitively, this is necessary to maximize the dropped power and, in turn, the nonlinear gain. The cutoff results from Kerr bistability; i.e., when further increases in $\alpha$ are no longer compensated by $N_{\rm{p}}$, the intraresonator field abruptly transitions to the CW state. To characterize the universality of this feature, we record $\Delta_{\rm{p}}$ (evaluated for the highest $CE$) versus $\delta$ for two different powers, as shown in Fig. 2d. Apparently, realizing high $CE$ requires $\Delta_{\rm{p}}\approx 0$. As a first step towards transitioning from a TMA to a multi-mode LLE, we briefly discuss the microresonator dispersion and its different representations. In Fig. 3a, we present $\delta_{\rm{\mu}}$ and $D_{\rm{int}}$ spectra for the fundamental transverse electric (TE) mode family, labeled TE1, of a SiN microring resonator (hereafter referred to as our test device) with ring radius $RR=15$ $\mu$m, ring width $RW=800$ nm, and height $H=600$ nm. We extract $\delta_{\rm{\mu}}$ and $D_{\rm{int}}$ from the mode spectrum, $\omega_{\rm{\mu}}$, that we calculate from finite-element method eigenfrequency simulations 111Finite element simulations performed using Comsol Multiphysics, which is identified here to foster understanding, without implying recommendation or endorsement by NIST.. It is easy to show that $\delta_{\rm{\mu}}=2\times\mathcal{S}(D_{\rm{int}}(\mu))$, where $\mathcal{S}(D_{\rm{int}}(\mu))$ indicates the symmetric part of $D_{\rm{int}}$; i.e., the even orders of $D_{\rm{int}}$ when expressed as a Taylor expansion around $\mu=0$. We have assessed that analyzing $\delta_{\rm{\mu}}$ is sufficient to understand the $\mu$OPO dynamics, which are not sensitive to odd orders of $D_{\rm{int}}$ Sayson _et al._ (2019). This is a notable departure from other nonlinear microresonator eigenstates, e.g., dissipative Kerr solitons. There are two defining features of the $\delta_{\rm{\mu}}$ spectrum shown in Fig. 3a. First, its negative curvature around $\mu=0$ indicates normal dispersion, which is required to suppress comb formation. Second, to overcome the normal dispersion and achieve frequency matching, the $D_{\rm{int}}$ expansion must contain higher order (even) terms such that $\delta_{\rm{\mu}}$ turns and becomes positive. Physically, this means that an anomalous-to-normal dispersion transition is necessary. Indeed, the dashed lines connecting Figs. 3a and 3b illustrate how $n_{\rm{g}}(\mu)=n_{\rm{g}}(-\mu)$ corresponds to a $\delta_{\rm{\mu}}$ turning point, where $n_{\rm{g}}(\mu)$ is the dispersive group velocity refractive index. Figure 4: Transition from a TMA to multi-mode LLE simulations. (a) Depiction of the four wave mixing (FWM) processes studied in this work. Mode competition (orange arrows) suppresses FWM to the target signal and idler mode pair (with mode numbers $\pm\mu$) in favor of spectrally-adjacent mode pairs (with mode numbers $\pm(\mu\pm 1)$), and modulation instability induced by cross-phase modulation (XPM-MI, purple arrows) reroutes energy in the pump mode to its spectrally-nearest neighbors. (b) $CE$ map for mode $\mu=46$; the resonator parameters are $RR=15\mu$m, $RW=800$ nm, $H=600$ nm, and $\omega_{\rm{p}}=2\pi\times 382$ THz. (c) Difference in $CE$ between simulations based on a TMA and the $CE$ map in part (b). Next, we preview the differences between $CE$ maps calculated from multi-mode and TMA models. We observe two FWM processes, depicted in Fig. 4a, that require the multi-mode model. In one case, we observe mode competition between mode pairs with consecutive $\mu$ values. In the second case, we observe modulation instability (MI) in the normally-dispersive spectral region around $\omega_{\rm{p}}$, and we explain it as arising from XPM between the pump, signal, and idler fields. We mostly constrain our multi-mode LLE simulations to $P_{\rm{in}}\leq 25$ mW ($X\leq 8$); at higher values of $P_{\rm{in}}$, the $\mu$OPO dynamics become more complex. Figure 5: Mode competition and switching. (a) Illustration of mode competition in a $\mu$OPO. Oscillations on the target signal and idler mode pair (with mode numbers $\pm\mu$) are suppressed in favor of spectrally-adjacent mode pairs (with mode numbers $\pm(\mu\pm 1)$). (b) $\delta_{\rm{\mu}}$ for the test device with $\omega_{\rm{p}}=2\pi\times 382$ THz. The pale orange stripe overlaps modes that are predicted to oscillate when $P_{\rm{in}}=20$ mW. (c) $CE$ maps for modes $\mu=46$, $\mu=45$, and $\mu=44$. The maps indicate a mode competition in which the mode pair with smallest $\delta$ is favored for oscillation. (d) $CE$ versus $\alpha$ for modes $\mu=46$ (blue) and $\mu=45$ (green), where $P_{\rm{in}}=20$ mW and $\beta=5\times 10^{-4}$. The pale stripes come from a TMA. (e) Optical spectra that correspond to different values of $\alpha$ in (d). The blue, orange, green, and red spectra correspond to $\alpha=1,1.5,2.5$, and $3.5$, respectively. Figure 4b shows a $CE$ map for our test device pumped near $\omega_{\rm{p}}=2\pi\times 382$ THz. To make a straightforward comparison between these data and Fig. 2b, we calculate the difference in $CE$ between the two $CE$ maps, as shown in Fig. 4c. In general, XPM-induced MI (XPM-MI) explains $CE$ differences for small $\delta$, while mode competition is responsible for the abrupt cutoff in $CE$ (marked by the sharp transition to zero $CE$ in Fig. 4b or the bold yellow stripe in Fig. 4c) that indicates a different mode pair is oscillating. In the following sections, we explore mode competition and XPM-MI in detail. ## III Mode competition and switching Mode competition is ubiquitous in laser systems with multi-mode resonators Narducci _et al._ (1986); Gong _et al._ (2007). In general, mode competition occurs when several resonator modes experience amplification simultaneously; hence, modes compete for gain and become coupled. In this section, we present the results of multi-mode LLE simulations in which several mode pairs are simultaneously nearly frequency matched. We use our findings to establish general principles for mode competition that will inform future microresonator designs. In Fig. 5, we present simulation results for our test device pumped near $\omega_{\rm{p}}=2\pi\times 382$ THz. The FWM process related to mode competition is illustrated in Fig. 5a. Modes that are spectrally adjacent to the targeted signal and idler pair may be nearly frequency matched; hence, light in these modes may become amplified through FWM. To explore this phenomenon, we consider the $\delta_{\rm{\mu}}$ spectrum shown in Fig. 5b. According to the TMA, any mode pairs with $\delta_{\rm{\mu}}>0$ may oscillate, provided $P_{\rm{in}}$ is large enough. In Fig. 5b, the data points covered by the pale gold stripe indicate mode pairs that would oscillate when $P_{\rm{in}}=20$ mW, if they (along with the pump mode) were the only modes in the system (i.e., in a TMA). Hence, we endeavour to understand how a mode pair (or pairs) is chosen for oscillation over its spectral neighbors. To study mode competition in our test device, we perform multi-mode LLE simulations, from which we construct $CE$ maps for the modes corresponding to $\mu=46$, $\mu=45$, and $\mu=44$, as shown in Fig. 5c. To vary $\delta_{\rm{\mu}}$ (which is fixed for a given geometry), we apply a quadratic dispersion to $\omega_{\rm{\mu}}$, such that $\omega_{\rm{\mu}}\rightarrow\omega_{\rm{\mu}}+\beta\kappa(0)\mu^{2}$, where $\beta$ quantifies the added dispersion. We choose this approach in order to maintain the overall shape of $\delta_{\rm{\mu}}$. Note that $\delta_{\rm{45}}$ and $\delta_{\rm{44}}$ are both negative in Fig. 5b, but they will become positive as $\beta$ is increased. Figure 6: Characterization of XPM-MI in microresonators with normal dispersion. (a) Illustration of the XPM-MI FWM process (purple arrows), in which XPM between the pump, signal, and idler fields induces MI that distributes pump energy into spectrally-nearby sidebands. (b) $\delta_{\rm{\mu}}$ for $\beta=0.0625$ (green circles), $\beta=1.25$ (cyan upward-pointing triangles), and $\beta=5$ (blue downward-pointing triangles). One mode pair ($\mu=\pm 50$) is frequency shifted to be frequency matched to the pump mode. (c) $CE$ versus $\alpha$ for the $\delta_{\rm{\mu}}$ spectra in (b), with $P_{\rm{in}}=20$ mW and $\delta=3$. (d) Optical spectrum associated with part (c), with $\alpha=4$ and $\beta=0.0625$. (e) $CE$ maps for microresonators with the $\delta_{\rm{\mu}}$ spectra in (b), ordered from least to greatest $\beta$. The $CE$ maps in Fig. 5c present clear evidence for mode competition, and we identify three notable features in them. First, the $CE$ map for mode $\mu=46$ resembles that of the TMA for small $\delta_{\rm{46}}$. Second, as $\beta$ is increased, the maximum $CE$ for mode $\mu=46$ declines sharply, and this coincides with $\delta_{\rm{45}}>0$ and oscillation on mode $\mu=45$. This pattern repeats as $\beta$ is increased further – the oscillating mode changes from $\mu=45$ to $\mu=44$, and so on. Hence, we assess that the mode pair with smallest positive $\delta$ is favored for oscillation. Finally, there are small regions of parameter space where multiple mode pairs seem to oscillate simultaneously – we explore this phenomenon in Figs. 5d and 5e. Figure 5d shows $CE$ for modes $\mu=46$ and $\mu=45$ during a $\omega_{\rm{p}}$ scan (i.e., varying $\alpha$) with $\beta=5\times 10^{-4}$ ($\delta_{\rm{46}}\approx 3$) and $P_{\rm{in}}=20$ mW. During the scan, mode $\mu=46$ begins to oscillate first; its $CE$ closely follows predictions made using a TMA (pale blue stripe) and reaches a maximum when $\alpha\approx 1.3$. In this regime, the $\mu$OPO spectrum is tri-modal, as depicted by the blue spectrum in Fig. 5e (the spectral bandwidth in Fig. 5e only spans $\omega_{\rm{p}}$ and $\omega_{\rm{s}}$; we have confirmed the spectrum is symmetric around $\omega_{\rm{p}}$). Beyond $\alpha\approx 1.3$, mode $\mu=45$ begins to oscillate, and $CE$ for mode $\mu=46$ deviates from its TMA counterpart. Between $\alpha\approx 1.3$ and $\alpha\approx 2.2$, both modes oscillate simultaneously, but their respective $CE$ values are not predicted by a TMA. Moreover, in this region the $\mu$OPO spectrum is not tri-modal; rather, it is strongly multi-moded with intensity peaks occurring near $\omega_{\rm{p}}$, $\omega_{\rm{s}}$, and $\omega_{\rm{i}}$, as shown by the orange spectrum in Fig. 5e. Beyond $\alpha\approx 2.2$, $CE$ for mode $\mu=46$ is zero, and $CE$ for mode $\mu=45$ follows its TMA counterpart. We term this phenomenon, wherein $CE$ values for different mode pairs conform to their TMA counterparts for different portions of a $\omega_{\rm{p}}$ scan, mode switching. Finally, at $\alpha\geq 3$, the $\mu$OPO decays in favor of MI. Remarkably, the MI state is supported in the normally-dispersive region around $\mu=0$ and without strong signal or idler fields. The MI spectrum is shown by the shaded red curve in Fig. 5e. In the next section, we explore MI and its impact on $CE$. ## IV XPM-MI: Characterization and Theory It is clear from Fig. 5 that mode competition is not the only process that differentiates the multi-mode LLE and TMA. Specifically, Figs. 5d and 5e establish that MI can suppress or extinguish the signal and idler fields, and this occurs in regions of parameter space where a TMA predicts efficient parametric oscillation. In this section, we isolate the MI process in our simulations by considering a special dispersion profile that eliminates mode competition, and we develop a theory for MI as arising from XPM between the pump, signal, and idler fields. We term this process XPM-MI to link it to previous investigations Agrawal (1987). The XPM-MI FWM process is illustrated by the purple arrows in Fig. 6a. Energy from the pump laser is distributed to sidebands in a narrow spectral bandwidth around $\omega_{\rm{p}}$. If we use a multi-mode LLE to simulate our test device, the effects of mode competition can obfuscate the impact of XPM-MI. Therefore, we contrive the heuristically useful $\delta_{\rm{\mu}}$ spectra depicted in Fig. 6b. Here, we define $\delta_{\rm{\mu}}=-2\beta\mu^{2}$ for all $\mu\neq\pm 50$. The modes $\mu=\pm 50$ are designated for parametric oscillation and assigned a frequency mismatch value $\delta$ that can be manipulated apart from $\beta$. Figure 7: Theory of XPM-MI. (a) MI gain parameter, $\lambda$, versus $\alpha$ for $\mu=1$. Orange and gold curves correspond to $P_{\rm{in}}=15$ mW, $\delta=1.5$, $\beta=0.25$, and $P_{\rm{in}}=25$ mW, $\delta=2.5$, $\beta=0.25$, respectively. The red curves all use $P_{\rm{in}}=20$ mW, with bold, dotted, and dashed curves corresponding to $\delta=2.5$ and $\beta=1$, $\delta=5$ and $\beta=0.25$, and $\delta=2$ and $\beta=0.25$, respectively. (b) The value of $\alpha$ at which MI first appears, $\alpha_{\rm{ON}}$, versus $\beta$ for $P_{\rm{in}}=20$ mW and $\delta=2.5$. Light blue circles indicate theoretical values using Eq. 5, while dark blue diamonds mark values predicted by LLE simulations. (c) $\lambda$ map for $\mu=1$ and $\beta=0.25$. In Fig. 6c, we present LLE simulations of $CE$ versus $\alpha$ for microresonators with the $\delta_{\rm{\mu}}$ spectra in Fig. 6b; these data overlay the corresponding simulation using a TMA (pale gray stripe). They are representative of the entire ($P_{\rm{in}},\delta)$ parameter space, and they are noteworthy for two reasons. First, in all cases $CE$ follows the gray stripe until $\alpha\geq\alpha_{\rm{ON}}$, where $\alpha_{\rm{ON}}$ corresponds to the onset of XPM-MI. An example XPM-MI spectrum is shown in Fig. 6d. Unlike the XPM-MI spectrum shown in Fig. 5e, here the XPM-MI and $\mu$OPO states co-exist, albeit with suppressed $CE$. Second, $\alpha_{\rm{ON}}$ increases with increasing $\beta$. As a result, $CE$ fully converges to its TMA counterpart in the high-$\beta$ limit. To investigate this convergence, we construct the $CE$ maps for different values of $\beta$, and we present the results in Fig. 6e. We observe steady convergence to the TMA $CE$ map as $\beta$ is increased. Clearly, it is crucial to realize strong normal dispersion near $\omega_{\rm{p}}$ to observe efficient parametric oscillation when $X>>1$. Next, we present an expression for the XPM-MI gain that we derive from a set of coupled mode equations (see Supplemental Material for details), and we analyze this expression to support our conclusion that XPM underlies the observed MI states. As noted in Ref. Hansson _et al._ (2013), MI may occur in normally-dispersive microresonators; however, it usually requires a hard excitation, i.e., the MI sidebands are not amplified from vacuum fluctuations. This fact explains how the MI state can sometimes persist after parametric oscillations decay. Still, in our simulations we have never observed MI emerge before parametric oscillations. Indeed, it was predicted in Ref. Agrawal (1987) that XPM between two waves can make them modulationally unstable, even when one or both waves propagates in normally-dispersive media. Motivated by this study, we analyze a set of coupled mode equations that assume the pump, signal, and idler modes are occupied, and we derive the MI gain, $\lambda_{\rm{\mu}}$, as ${}\lambda_{\rm{\mu}}=-1+\sqrt{I_{\rm{0}}^{2}+I_{\rm{s}}^{2}+2I_{\rm{0}}I_{\rm{s}}\rm{cos}(\Delta\phi)-\textit{k}_{\rm{\mu}}^{2}},$ (5) where $I_{\rm{0}}$ and $I_{\rm{s}}$ are the photon numbers for the pump and signal modes, respectively, $\Delta\phi=2\phi_{\rm{p}}-\phi_{\rm{s}}-\phi_{\rm{i}}$ is the relative phase mismatch between the pump, signal, and idler fields, which are expressed as $\tilde{a}_{\rm{p(s,i)}}=\sqrt{I_{\rm{p(s,i)}}}e^{-i\phi_{\rm{p(s,i)}}}$, and $k_{\rm{\mu}}=2I_{\rm{0}}+4I_{\rm{s}}-\beta\mu^{2}-\alpha$. MI occurs when $\lambda_{\rm{\mu}}>0$, for any $\mu$. Note from these definitions that $\beta\mu^{2}$ is the frequency mismatch parameter for the MI sidebands, with mode numbers $\pm\mu$, and not the frequency mismatch for the $\mu$OPO mode pair, which we denote as $\delta$. In the limit $I_{\rm{s}}\rightarrow 0$, Eq. 5 is equivalent to the MI gain expression derived in Ref. Hansson _et al._ (2013). Moreover, it is clear that $\lambda_{\rm{\mu}}$ is maximized when the pump, signal, and idler fields are in phase (i.e. when $\Delta\phi=0$), which corresponds to the in-phase addition of two FWM processes: the degenerate FWM process $2\omega_{\rm{p}}\rightarrow\omega_{\rm{\mu}}+\omega_{\rm{-\mu}}$ and the non-degenerate process $\omega_{\rm{s}}+\omega_{\rm{i}}\rightarrow\omega_{\rm{\mu}}+\omega_{\rm{-\mu}}$. To analyze XPM-MI using Eq. 5, we perform LLE simulations, using a TMA, for various values of $P_{\rm{in}}$ and $\delta$ and extract values for $I_{\rm{0}},I_{\rm{s}}$, and $\Delta\phi$ that we use to calculate $\lambda_{\rm{\mu}}$. When $\lambda_{\rm{\mu}}>0$, the mode pair with mode numbers $\pm\mu$ will be amplified and steal energy from the $\mu$OPO. In general, a large normal dispersion leads to the sideband pair with $\mu=\pm 1$ having the largest gain; therefore, in what follows we drop the $\mu$ subscript and define $\lambda$ as the MI gain for this pair. We have confirmed that $\lambda_{\rm{\mu}}<0$ whenever $I_{\rm{s}}=0$ and $\beta>0$; physically, this means that XPM between the pump, sigal, and idler fields is required to initiate MI for microresonators with normal dispersion. Hence, we term this process XPM-MI. Figure 7a presents $\lambda$ calculations for various values of $P_{\rm{in}}$, $\delta$, and $\beta$. They indicate that $\lambda$ grows with increasing $\alpha$, which is explained by the stronger XPM that results from more powerful $\mu$OPO sidebands at large $\alpha$. To compare our XPM-MI theory (i.e., Eq. 5) to multi-mode LLE simulations, we calculate $\alpha_{\rm{ON}}$ in both cases, for different values of $\beta$. Overall, we observe good agreement, but Eq. 5 generally predicts higher values of $\alpha_{\rm{ON}}$ than the multi-mode LLE. A sample comparison of this type is presented in Fig. 7b. Figure 8: Strategies to achieve high $CE$ and increase power output from a $\mu$OPO. (a) Depiction of a microring resonator and its loss rates. (b) Simulated mode transmission vs $\omega_{\rm{p}}$ for $\kappa_{\rm{i}}=2\pi\times 125$ MHz (red), $\kappa_{\rm{i}}=2\pi\times 200$ MHz (orange), and $\kappa_{\rm{i}}=2\pi\times 275$ MHz (gold). In each case, $\kappa_{\rm{c}}=\kappa_{\rm{i}}$. (c) $CE$ maps corresponding to (b), for a resonator with $RR=15~{}\mu$m, $RW=800$ nm, $H=600$ nm, and $\omega_{\rm{p}}=2\pi\times 385$ THz. (d) Simulated mode transmissions for $\kappa_{\rm{c}}=\kappa_{\rm{i}}$ (green) and $\kappa_{\rm{c}}=10\times\kappa_{\rm{i}}$ (blue), where $\kappa_{\rm{i}}=2\pi\times 200$ MHz. (e) Maximum $CE$ versus $\kappa_{\rm{c}}/\kappa_{\rm{i}}$ and $\delta$, where $\kappa_{\rm{c}}$ is the signal mode coupling rate. The pump and idler modes are critically coupled. To more comprehensively compare our theory with multi-mode LLE simulations, we calculate the maximum gain, $\lambda_{\rm{max}}$, for $\beta=0.25$ and different values of $P_{\rm{in}}$ and $\delta$. We use these data to construct the $\lambda_{\rm{max}}$ map presented in Fig. 7c. It accurately predicts the threshold $P_{\rm{in}}$ values for XPM-MI, and it has the overall trend that $\lambda_{\rm{max}}$ grows with increasing $P_{\rm{in}}$ and $\delta$. This trend is consistent with the $CE$ maps presented in Fig. 6e, which deviate from the TMA as $P_{\rm{in}}$ and $\delta$ are increased. Importantly, our observations may explain prior experimental results, which consistently report that increasing $P_{\rm{in}}$ leads to the formation of undesired sidebands Lu _et al._ (2019); Tang _et al._ (2020); Fujii _et al._ (2019). Overall, our theory supports the hypothesis that XPM drives MI in a $\mu$OPO. Moreover, we can now form a concise description of parasitic FWM in $\mu$OPOs: Mode competition dictates which mode pair oscillates, while XPM-MI constrains $CE$. ## V Towards efficient and high-power $\mu$OPO Our results suggest several design rules for avoiding mode competition and mitigating XPM-MI through dispersion engineering. For example, Fig. 6 demonstrates that larger normal dispersion around $\omega_{\rm{p}}$ will suppress XPM-MI. Moreover, increasing $\frac{d\delta_{\rm{\mu}}}{d\mu}$ will help prevent mode competitions. This can be accomplished by using smaller resonators or in resonators with greater dispersion. Still, in practice it is nontrivial to fabricate devices with ideal dispersion characteristics. Therefore, in this section we describe two strategies to increase $CE$ by tuning the microresonator loss rates, $\kappa_{\rm{c}}(\mu)$ and $\kappa_{\rm{i}}$. We assume that users of real-world $\mu$OPO devices will value high output power; therefore, we focus on ways to increase $CE$ for fixed $P_{\rm{in}}$. Figure 8a depicts the coupling and intrinsic loss rates that are quantified by $\kappa_{\rm{c}}(\mu)$ and $\kappa_{\rm{i}}$, respectively. In practice, $\kappa_{\rm{c}}(\mu)$ is controllable by various design parameters, including the waveguide-resonator separation and waveguide-resonator coupling length, for example, in a ‘pulley’ configuration Moille _et al._ (2019). Moreover, $\kappa_{\rm{i}}$ can be reduced within some spectral bands by annealing Spencer _et al._ (2014); Graziosi _et al._ (2018); hence, one gains some control over $\kappa_{\rm{i}}$ by making a suitable choice for the annealing time or temperature (also, increasing $\kappa_{\rm{i}}$ can be realized through, for example, intentionally-introduced surface roughness). We consider two ways to increase $CE$. First, by increasing $P_{\rm{th}}$, one increases the $P_{\rm{in}}$ values for which XPM-MI occurs. To demonstrate this approach, we perform multi-mode LLE simulations of our test device for three different values of $\kappa$. In every simulation, $\kappa_{\rm{c}}(\mu)=\kappa_{\rm{i}}$ and $\omega_{\rm{p}}=2\pi\times 385$ THz. The simulated modal lineshapes for these resonators are shown in Fig. 8b, and the corresponding $CE$ maps are shown in Fig. 8c. In Fig. 8c, $\kappa$ is increased from the left-most panel to the right-most panel. As $\kappa$ is increased, the region of highest $CE$ is shifted towards higher values of $P_{\rm{in}}$. Moreover, this region is broadened along both axes, indicating that devices with larger $\kappa$ will have a greater tolerance for design errors in $\delta$. While increases in $\kappa$ prevent high $CE$ at low $P_{\rm{in}}$, the obtainable output power generated with high $P_{\rm{in}}$ has grown; for instance, with $P_{\rm{in}}=30$ mW, the maximum output power increases from $P_{\rm{sig}}\approx 2.25$ mW at $\kappa/2\pi=250$ MHz (Fig. 8c, left-most panel) to $P_{\rm{sig}}\approx 3.75$ mW at $\kappa/2\pi=550$ MHz (Fig. 8c, right-most panel). Finally, we consider the relationship between the signal mode coupling rate, $\kappa_{\rm{c}}$, and $CE$. This relationship was explored in Ref. Sayson _et al._ (2019), with the result that $CE$ may be as large as $25$% (overcoupling the pump mode may further increase $CE$). Here, we reiterate this result and explore it within the $\mu$OPO parameter space. Figure 8d shows simulated signal mode lineshapes for two values of $\kappa_{\rm{c}}$, and Fig. 8e shows the $CE$ map for $P_{\rm{in}}=20$ mW in a parameter space defined by the coupling ratio, $\kappa_{\rm{c}}/\kappa_{\rm{i}}$, and $\delta$. In our simulations, we use a TMA and keep the pump and idler modes critically coupled. We not only observe an increase in the highest obtainable $CE$ to nearly $25$%, but we also observe that the region of highest $CE$ is broadened for increasing $\kappa_{\rm{c}}/\kappa_{\rm{i}}$. Broadening occurs until $\kappa_{\rm{c}}/\kappa_{\rm{i}}\approx 30$, at which point the advantages of overcoupling are overcome by the corresponding increases in $P_{\rm{th}}$. ## VI Discussion In conclusion, we have established a foundation of simulation results for $\mu$OPOs that moves beyond a TMA and will help guide experimental efforts to realize the high conversion efficiencies predicted by the simplified theory. We introduced a $CE$ map that encapsulates the $\mu$OPO solution space, and through multi-mode LLE simulations we reveal nonlinear dynamics not predictable from a TMA. In particular, we identified mode competition and demonstrated how it determines the oscillating mode pair. Meanwhile, the range of parameter space over which high $CE$ can be obtained is constrained by XPM- MI. Mode competition and XPM-MI both depend on the microresonator dispersion, $\delta_{\rm{\mu}}$, and suitable dispersion engineering may circumvent these processes. Still, optimizing $\delta_{\rm{\mu}}$ may be nontrivial in practice; therefore, we have proposed two strategies to increase $CE$ apart from dispersion engineering. Ultimately, suitable control of both resonator loss (including waveguide coupling) and dispersion makes it possible to tailor microresonator geometries for high $CE$ at a targeted pump power, that is, to produce a useful amount of output power. Such engineering will be crucial in the development of compact, coherent light sources that take advantage of the enormous wavelength flexibility inherent to $\chi^{(3)}$ OPOs. ###### Acknowledgements. We thank Travis Briles and Edgar Perez for a careful reading of the paper. This project is funded by the DARPA LUMOS program. Supplemental Material ## I Calculations of $\mu$OPO variables In this section, we describe mathematical formulae that relate LLE variables to the data (e.g. $CE$ and optical spectra) presented in the main text. Simulations of the Lugiato-Lefever Equation (LLE) yield solutions for the complex intraresonator field, $a$; we denote the intraresonator field spectrum $\tilde{a}=\mathcal{F}(a)$, where $\mathcal{F}$ denotes the Fourier transform, and $a$ is normalized such that $|\tilde{a}|^{2}_{\omega=\omega_{\mu}}$ gives the number of intraresonator photons in the mode $\mu$. Table 1 lists the important variables, along with their physical descriptions and how they are calculated from $a$, $\tilde{a}$, and simulation parameters. Notably, the expressions for $\phi_{\rm{p}}$ and $N_{\rm{p}}$ also apply to the signal and idler modes when the expressions are evaluated at the appropriate value of $\omega$. We follow Ref. Yu _et al._ (2021) in our calculation of the nonlinear mode frequency shifts. Variable | Description (units) | LLE Calculation ---|---|--- $P_{\rm{sig}}$ | Output signal power (W) | $\kappa_{\rm{c}}\hbar\omega_{\rm{s}}|\tilde{a}|^{2}_{\omega=\omega_{\rm{s}}}$ $\phi_{\rm{p}}$ | Pump phase (rad) | $\angle\tilde{a}_{\omega=\omega_{\rm{0}}}$ $N_{\rm{p}}$ | Pump mode nonlinear frequency shift ($\kappa(0)/2\pi$) | $\frac{1}{\kappa(0)}\left(\frac{\mathcal{F}(g_{\rm{0}}|a|^{2}a)}{\tilde{a}}\right)_{\omega=\omega_{\rm{0}}}$ $P_{\rm{out}}$ | Pump laser output power ($W$) | $\hbar\omega_{\rm{p}}|\sqrt{\kappa_{\rm{c}}(0)}\tilde{a}_{\omega=\omega_{\rm{0}}}-\sqrt{\frac{P_{\rm{in}}}{\hbar\omega_{\rm{p}}}}|^{2}$ Table 1: List and descriptions of important variables derived from the LLE. ## II Derivation of highest maximum $CE$ contour Next, we seek to derive the result - stated in the main text - that the contour of highest maximum $CE$ in the TMA is given by $X=8\delta$. Our starting point is the following set of coupled equations Sayson _et al._ (2019): $\displaystyle\frac{dA_{\rm{p}}}{dt}$ $\displaystyle=-(1+i\alpha)A_{\rm{p}}+i(|A_{\rm{p}}|^{2}+2|A_{\rm{s}}|^{2}+2|A_{\rm{i}}|^{2})A_{\rm{p}}+2A_{\rm{p}}^{*}A_{\rm{s}}A_{\rm{i}}+\sqrt{X}$ (6) $\displaystyle\frac{dA_{\rm{s}}}{dt}$ $\displaystyle=-(1+i\alpha+i\delta)A_{\rm{s}}+i(|A_{\rm{s}}|^{2}+2|A_{\rm{p}}|^{2}+2|A_{\rm{i}}|^{2})A_{\rm{s}}+2A_{\rm{s}}^{*}A_{\rm{p}}A_{\rm{i}}$ (7) $\displaystyle\frac{dA_{\rm{i}}}{dt}$ $\displaystyle=-(1+i\alpha+i\delta)A_{\rm{i}}+i(|A_{\rm{i}}|^{2}+2|A_{\rm{p}}|^{2}+2|A_{\rm{s}}|^{2})A_{\rm{i}}+2A_{\rm{i}}^{*}A_{\rm{p}}A_{\rm{s}},$ (8) where $A_{\rm{p(s,i)}}=\sqrt{I_{\rm{p(s,i)}}}e^{-i\phi_{\rm{p(s,i)}}}$ denotes the pump (signal, idler) field, and $I_{\rm{p(s,i)}}=|A_{\rm{p(s,i)}}|^{2}$ is proportional to the pump (signal, idler) intraresonator photon number. If we substitute the field expressions into the above equations, we obtain six coupled equations for the real variables $I_{\rm{p(s,i)}}$ and $\phi_{\rm{p(s,i)}}$, $\displaystyle\frac{dI_{\rm{p}}}{dt}$ $\displaystyle=-2I_{\rm{p}}-4I_{\rm{p}}\sqrt{I_{\rm{s}}I_{\rm{i}}}\textrm{sin}(\Delta\phi)+2\sqrt{I_{\rm{p}}X}\textrm{cos}(\phi_{\rm{p}})$ (9) $\displaystyle\frac{dI_{\rm{s}}}{dt}$ $\displaystyle=-2I_{\rm{s}}+2I_{\rm{p}}\sqrt{I_{\rm{s}}I_{\rm{i}}}\textrm{sin}(\Delta\phi)$ (10) $\displaystyle\frac{dI_{\rm{i}}}{dt}$ $\displaystyle=-2I_{\rm{i}}+2I_{\rm{p}}\sqrt{I_{\rm{s}}I_{\rm{i}}}\textrm{sin}(\Delta\phi)$ (11) $\displaystyle\frac{d\phi_{\rm{p}}}{dt}$ $\displaystyle=I_{\rm{p}}+2I_{\rm{s}}+2I_{\rm{i}}+2\sqrt{I_{\rm{s}}I_{\rm{i}}}\textrm{cos}(\Delta\phi)-\alpha-\sqrt{\frac{X}{I_{\rm{p}}}}\textrm{sin}(\phi_{\rm{p}})$ (12) $\displaystyle\frac{d\phi_{\rm{s}}}{dt}$ $\displaystyle=I_{\rm{s}}+2I_{\rm{p}}+2I_{\rm{i}}+I_{\rm{p}}\sqrt{\frac{I_{\rm{i}}}{I_{\rm{s}}}}\textrm{cos}(\Delta\phi)-\alpha-\delta$ (13) $\displaystyle\frac{d\phi_{\rm{i}}}{dt}$ $\displaystyle=I_{\rm{i}}+2I_{\rm{p}}+2I_{\rm{s}}+I_{\rm{p}}\sqrt{\frac{I_{\rm{s}}}{I_{\rm{i}}}}\textrm{cos}(\Delta\phi)-\alpha-\delta,$ (14) where $\Delta\phi=2\phi_{\rm{p}}-\phi_{\rm{s}}-\phi_{\rm{i}}$. We next reduce this system to four equations by first noting that, from the symmetry of Eqs. 7 and 8, $I_{\rm{s}}=I_{\rm{i}}$. Therefore, we define $M=I_{\rm{s}}=I_{\rm{i}}$. Moreover, momentum conservation implies that $\frac{d}{dt}(\phi_{\rm{s}}+\phi_{\rm{i}})=0$. Hence, for steady state conditions we obtain $\displaystyle I_{\rm{p}}$ $\displaystyle=-2I_{\rm{p}}M\textrm{sin}(\Delta\phi)+\sqrt{I_{\rm{p}}X}\textrm{cos}(\phi_{\rm{p}})$ (15) $\displaystyle 1$ $\displaystyle=I_{\rm{p}}\textrm{sin}(\Delta\phi)$ (16) $\displaystyle\sqrt{\frac{X}{I_{\rm{p}}}}\textrm{sin}(\phi_{\rm{p}})$ $\displaystyle=I_{\rm{p}}+4M+2M\textrm{cos}(\Delta\phi)-\alpha$ (17) $\displaystyle I_{\rm{p}}\textrm{cos}(\Delta\phi)$ $\displaystyle=\alpha-3M-2I_{\rm{p}}+\delta.$ (18) Equations 15-18 are still sufficiently general to study the stationary solutions. However, we proceed to simplify them via two ansatzes associated with a maximally efficient $\mu$OPO. Specifically, we set $\phi_{\rm{p}}=0$, which corresponds to the pump laser being on resonance, and $M=\frac{I_{\rm{p}}}{2}$. The validity of these assumptions is confirmed by our numerical results, but they are also physically intuitive. For instance, because the FWM process that drives $\mu$OPO ($2\omega_{\rm{p}}\rightarrow\omega_{\rm{s}}+\omega_{\rm{i}}$) is reversible, one expects the intraresonator photons to be evenly distributed between the pump mode and sideband pair. After inserting our two ansatzes into Eqs. 15-18, we combine Eqs. 15 and 16 to obtain $I_{\rm{p}}=X/4$. Then, Eqs. 17 and 18 are combined to obtain $\frac{I_{\rm{p}}}{2}=\delta$. Insertion of the former into the latter yields the desired result, $X=8\delta$. ## III Derivation of XPM-MI gain Finally, in this section we derive Eq. 4 from the main text. We follow the classical procedure of linearizing a set of coupled equations around a steady- state solution and then allowing plane-wave perturbations to grow exponentially. We start from the equations of motion for the fields $A_{\rm{\mu}}$ and $A_{\rm{-\mu}}$ when the pump, signal, and idler modes are occupied, $\displaystyle\frac{dA_{\rm{\mu}}}{dt}$ $\displaystyle=i\beta\mu^{2}+i(2I_{\rm{p}}+2I_{\rm{s}}+2I_{\rm{i}}+2I_{\rm{-\mu}}+I_{\rm{\mu}})A_{\rm{\mu}}-(1+i\alpha)A_{\rm{\mu}}+iA_{\rm{\mu}}^{*}A_{\rm{p}}^{2}+iA_{\rm{\mu}}^{*}A_{\rm{s}}A_{\rm{i}}$ (19) $\displaystyle\frac{dA_{\rm{-\mu}}}{dt}$ $\displaystyle=i\beta\mu^{2}+i(2I_{\rm{p}}+2I_{\rm{s}}+2I_{\rm{i}}+I_{\rm{-\mu}}+2I_{\rm{\mu}})A_{\rm{-\mu}}-(1+i\alpha)A_{\rm{-\mu}}+iA_{\rm{-\mu}}^{*}A_{\rm{p}}^{2}+iA_{\rm{-\mu}}^{*}A_{\rm{s}}A_{\rm{i}},$ (20) where $\beta$ quantifies the dispersion. Next, we introduce the perturbation $\delta f_{\rm{\mu}}(t)$, so that the equations of motion for the perturbations, after simplifying, read $\displaystyle\frac{d\delta f_{\rm{\mu}}}{dt}$ $\displaystyle=(ik_{\rm{\mu}}-1)\delta f_{\rm{\mu}}+i(A_{\rm{p}}^{2}+A_{\rm{s}}A_{\rm{i}})\delta f_{\rm{\mu}}^{*}$ (21) $\displaystyle\frac{d\delta f_{\rm{-\mu}}^{*}}{dt}$ $\displaystyle=-(ik_{\rm{\mu}}+1)\delta f_{\rm{-\mu}}^{*}-i(A_{\rm{p}}^{*2}+A_{\rm{s}}^{*}A_{\rm{i}}^{*})\delta f_{\rm{\mu}},$ (22) where $k_{\rm{\mu}}=\beta\mu^{2}+2I_{\rm{p}}+4I_{\rm{s}}-\alpha$, and we have again assumed $I_{\rm{s}}=I_{\rm{i}}$. If we set $\delta f_{\rm{\mu}}(t)=ae^{\lambda_{\rm{\mu}}t}$, then we obtain a set of linear homogenous equations with eigenvalue $\lambda_{\rm{\mu}}$. Solving the eigenvalue problem yields $\lambda_{\rm{\mu}}=-1\pm\sqrt{I_{\rm{p}}^{2}+I_{\rm{s}}^{2}+2I_{\rm{p}}I_{\rm{s}}\textrm{cos}(\Delta\phi)-k_{\rm{\mu}}^{2}},$ (23) which is the desired result. Notably, the original set of coupled equations are approximations in several ways. For instance, we only consider one sideband pair, but the full XPM-MI dynamics involves the collective interactions of multiple sideband pairs simultaneously. In addition, we neglect FWM processes that do not involve both MI sidebands (e.g. the interaction $A_{\rm{\mu}}^{*}A_{\rm{s}}^{*}A_{\rm{p}}A_{\rm{s+\mu}}$). These approximations may explain the small discrepancies between our theory and LLE simulations. ## References * Liang and Bowers (2010) D. Liang and J. E. Bowers, Nature Photonics 4, 511 (2010). * Agrell _et al._ (2016) E. Agrell, M. Karlsson, A. Chraplyvy, D. J. Richardson, P. M. Krummrich, P. Winzer, K. Roberts, J. K. Fischer, S. J. Savory, B. J. Eggleton, _et al._ , Journal of Optics 18, 063002 (2016). * Sipahigil _et al._ (2016) A. Sipahigil, R. E. Evans, D. D. Sukachev, M. J. Burek, J. Borregaard, M. K. Bhaskar, C. T. Nguyen, J. L. Pacheco, H. A. Atikian, C. Meuwly, _et al._ , Science 354, 847 (2016). * Elshaari _et al._ (2020) A. W. Elshaari, W. Pernice, K. Srinivasan, O. Benson, and V. Zwiller, Nature Photonics 14, 285 (2020). * Jin _et al._ (2021) W. Jin, Q.-F. Yang, L. Chang, B. Shen, H. Wang, M. A. Leal, L. Wu, M. Gao, A. Feshali, M. Paniccia, _et al._ , Nature Photonics 15, 346 (2021). * Vahala (2003) K. J. Vahala, Nature 424, 839 (2003). * Yang _et al._ (2018) K. Y. Yang, D. Y. Oh, S. H. Lee, Q.-F. Yang, X. Yi, B. Shen, H. Wang, and K. Vahala, Nature Photonics 12, 297 (2018). * Kippenberg _et al._ (2018) T. J. Kippenberg, A. L. Gaeta, M. Lipson, and M. L. Gorodetsky, Science 361 (2018). * Liu _et al._ (2018) X. Liu, C. Sun, B. Xiong, L. Wang, J. Wang, Y. Han, Z. Hao, H. Li, Y. Luo, J. Yan, _et al._ , ACS Photonics 5, 1943 (2018). * Loh _et al._ (2015) W. Loh, A. A. Green, F. N. Baynes, D. C. Cole, F. J. Quinlan, H. Lee, K. J. Vahala, S. B. Papp, and S. A. Diddams, Optica 2, 225 (2015). * Bruch _et al._ (2019) A. W. Bruch, X. Liu, J. B. Surya, C.-L. Zou, and H. X. Tang, Optica 6, 1361 (2019). * Lu _et al._ (2019) X. Lu, G. Moille, A. Singh, Q. Li, D. A. Westly, A. Rao, S.-P. Yu, T. C. Briles, S. B. Papp, and K. Srinivasan, Optica 6, 1535 (2019). * Spencer _et al._ (2018) D. T. Spencer, T. Drake, T. C. Briles, J. Stone, L. C. Sinclair, C. Fredrick, Q. Li, D. Westly, B. R. Ilic, A. Bluestone, _et al._ , Nature 557, 81 (2018). * Marin-Palomo _et al._ (2017) P. Marin-Palomo, J. N. Kemal, M. Karpov, A. Kordts, J. Pfeifle, M. H. Pfeiffer, P. Trocha, S. Wolf, V. Brasch, M. H. Anderson, _et al._ , Nature 546, 274 (2017). * Lai _et al._ (2020) Y.-H. Lai, M.-G. Suh, Y.-K. Lu, B. Shen, Q.-F. Yang, H. Wang, J. Li, S. H. Lee, K. Y. Yang, and K. Vahala, Nature Photonics 14, 345 (2020). * Newman _et al._ (2019) Z. L. Newman, V. Maurice, T. Drake, J. R. Stone, T. C. Briles, D. T. Spencer, C. Fredrick, Q. Li, D. Westly, B. R. Ilic, _et al._ , Optica 6, 680 (2019). * Vodopyanov _et al._ (2000) K. Vodopyanov, F. Ganikhanov, J. Maffetone, I. Zwieback, and W. Ruderman, Optics Letters 25, 841 (2000). * Mieth _et al._ (2014) S. Mieth, A. Henderson, and T. Halfmann, Optics express 22, 11182 (2014). * Boyd (2020) R. W. Boyd, _Nonlinear optics_ (Academic press, 2020). * Lu and Srinivasan (2021) X. Lu and K. Srinivasan, Physical Review Applied 16, 014027 (2021). * Sayson _et al._ (2019) N. L. B. Sayson, T. Bi, V. Ng, H. Pham, L. S. Trainor, H. G. Schwefel, S. Coen, M. Erkintalo, and S. G. Murdoch, Nature Photonics 13, 701 (2019). * Tang _et al._ (2020) Y. Tang, Z. Gong, X. Liu, and H. X. Tang, Optics Letters 45, 1124 (2020). * Lu _et al._ (2020) X. Lu, G. Moille, A. Rao, D. A. Westly, and K. Srinivasan, Optica 7, 1417 (2020). * Domeneguetti _et al._ (2021) R. R. Domeneguetti, Y. Zhao, X. Ji, M. Martinelli, M. Lipson, A. L. Gaeta, and P. Nussenzveig, Optica 8, 316 (2021). * Marty _et al._ (2021) G. Marty, S. Combrié, F. Raineri, and A. De Rossi, Nature Photonics 15, 53 (2021). * Hansson _et al._ (2013) T. Hansson, D. Modotto, and S. Wabnitz, Physical Review A 88, 023819 (2013). * Coen _et al._ (2013) S. Coen, H. G. Randle, T. Sylvestre, and M. Erkintalo, Optics Letters 38, 37 (2013). * Chembo and Menyuk (2013) Y. K. Chembo and C. R. Menyuk, Physical Review A 87, 053852 (2013). * Kippenberg _et al._ (2004) T. Kippenberg, S. Spillane, and K. Vahala, Physical review letters 93, 083904 (2004). * Note (1) Finite element simulations performed using Comsol Multiphysics, which is identified here to foster understanding, without implying recommendation or endorsement by NIST. * Narducci _et al._ (1986) L. Narducci, J. Tredicce, L. Lugiato, N. Abraham, and D. Bandy, Physical Review A 33, 1842 (1986). * Gong _et al._ (2007) M. Gong, Y. Yuan, C. Li, P. Yan, H. Zhang, and S. Liao, Optics Express 15, 3236 (2007). * Agrawal (1987) G. P. Agrawal, Physical Review Letters 59, 880 (1987). * Fujii _et al._ (2019) S. Fujii, S. Tanaka, M. Fuchida, H. Amano, Y. Hayama, R. Suzuki, Y. Kakinuma, and T. Tanabe, Optics letters 44, 3146 (2019). * Moille _et al._ (2019) G. Moille, Q. Li, T. C. Briles, S.-P. Yu, T. Drake, X. Lu, A. Rao, D. Westly, S. B. Papp, and K. Srinivasan, Optics Letters 44, 4737 (2019). * Spencer _et al._ (2014) D. T. Spencer, J. F. Bauters, M. J. Heck, and J. E. Bowers, Optica 1, 153 (2014). * Graziosi _et al._ (2018) T. Graziosi, S. Mi, M. Kiss, and N. Quack, in _2018 International Conference on Optical MEMS and Nanophotonics (OMN)_ (IEEE, 2018) pp. 1–5. * Yu _et al._ (2021) S.-P. Yu, D. C. Cole, H. Jung, G. T. Moille, K. Srinivasan, and S. B. Papp, Nature Photonics 15, 461 (2021).
# AutoProtoNet: Interpretability for Prototypical Networks Pedro Sandoval-Segura Department of Computer Science University of Maryland <EMAIL_ADDRESS> Wallace Lawson Navy Center for Applied Research in Artificial Intelligence Naval Research Laboratory <EMAIL_ADDRESS> ###### Abstract In meta-learning approaches, it is difficult for a practitioner to make sense of what kind of representations the model employs. Without this ability, it can be difficult to both understand what the model knows as well as to make meaningful corrections. To address these challenges, we introduce AutoProtoNet, which builds interpretability into Prototypical Networks by training an embedding space suitable for reconstructing inputs, while remaining convenient for few-shot learning. We demonstrate how points in this embedding space can be visualized and used to understand class representations. We also devise a prototype refinement method, which allows a human to debug inadequate classification parameters. We use this debugging technique on a custom classification task and find that it leads to accuracy improvements on a validation set consisting of in-the-wild images. We advocate for interpretability in meta-learning approaches and show that there are interactive ways for a human to enhance meta-learning algorithms. ## 1 Introduction ††DISTRIBUTION STATEMENT A. Approved for public release: distribution unlimited. It is expensive and time-consuming to collect data to train current state-of- the-art image classification systems [13]. When a classification algorithm is deployed, new classes or labels cannot be easily added without incurring new costs related to re-training the model [1][2]. Meta-learning approaches for few-shot learning solve both these problems by training networks that learn quickly from little data with computationally inexpensive fine-tuning [23][20][15]. Despite these methods performing well on benchmark few-shot image classification tasks, these methods are not interpretable; a human may have no way of knowing why a certain classification decision was made. Additionally, the lack of interpretability limits any kind of debugging of network representations. In this work, we take a step toward the development of a meta-learning algorithm which can learn in a few-shot setting, can handle new classes at test time, is interpretable enough for a human to understand how the model makes decisions, and which can be debugged in a simple way. We revisit Prototypical Networks (ProtoNets) [20] as the focus of our study. ProtoNets are based on a simple idea: there exists an embedding space where images cluster around a single “prototype” for each class. Given the simplicity of this few-shot learning approach, it makes sense to ask: what does a prototype look like? And, have we learned an adequate prototype representation? The outcomes of our study can be summarized as follows: * • We introduce AutoProtoNet, which merges ideas from autoencoders and Prototypical Networks, to perform few-shot image classification and prototype reconstruction. * • We use AutoProtoNet to visualize prototypes and find that they are comparable in quality to those of an autoencoder. AutoProtoNet also remains accurate on few-shot image classification benchmarks. * • We devise a prototype refinement method, which can be used to debug inadequate prototypes, and we validate the performance of the resulting model using a novel validation set of in-the-wild images. Our goal in this work is to elucidate the benefits of learning embeddings that can be visualized and interpreted by humans. To the best of our knowledge, there is no meta-learning approach that allows for a human to play a role in the fine-tuning of the base model. ## 2 Related Work ### 2.1 Meta-learning and Prototypical Networks Before meta-learning, transfer learning was used to handle few-shot problems. In transfer learning, a feature extractor is trained on a large dataset, then fine-tuned for new tasks [2]. However, transfer learning has some drawbacks. For example, adding a new class may require re-training the model and, in the few-shot setting, overfitting few example images is possible. Meta-learning algorithms aim to learn a “base” model that can be quickly fine- tuned for a new task. The base model is trained using a set of training tasks $\\{\mathcal{T}_{i}\\}$, sampled from some task distribution. Each task consists of _support_ data, $\mathcal{T}_{i}^{s}$, and _query_ data, $\mathcal{T}_{i}^{q}$. Support data is used to fine-tune the model, while query data is used to evaluate the resulting model. Practically speaking, each task is an image classification problem involving only a small number of classes. The number of examples per class in the support set is called the _shot_ , and the number of classes is called the _way_. For example, in 5-way 1-shot learning, we are given 1 example for each of the 5 classes to use for fine-tuning. Following the meta-learning framework presented in [8], Algorithm 1 can be used as a general way to understand both metric-learning methods [23] [20] and gradient-based methods like MAML [6]. Algorithm 1 The meta-learning framework Input: Base model, $F_{\theta}$ Input: Fine-tuning algorithm, $A$ Input: Learning rate, $\gamma$ Input: Distribution over tasks, $p(\mathcal{T})$ 1:Initialize $\theta$, the weights of $F$ 2:while not done do 3: Sample batch of tasks $\\{\mathcal{T}_{i}\\}_{i=1}^{n}$, where $\mathcal{T}_{i}\sim p(\mathcal{T})$ and $\mathcal{T}_{i}=(\mathcal{T}_{i}^{s},\mathcal{T}_{i}^{q})$ 4: for i=1,…,n do 5: $\theta_{i}\leftarrow A(\theta,\mathcal{T}_{i}^{s})$ $\triangleright$ Fine- tune model on $\mathcal{T}_{i}^{s}$ (inner loop) 6: $g_{i}\leftarrow\nabla_{\theta}\mathcal{L}(F_{\theta_{i}},\mathcal{T}_{i}^{q})$ 7: end for 8: $\theta\leftarrow\theta-\frac{\gamma}{n}\sum_{i}g_{i}$ $\triangleright$ Update base model parameters (outer loop) 9:end while For ProtoNets [20], the base model $F_{\theta}:\mathbb{R}^{D}\rightarrow\mathbb{R}^{M}$ is an embedding network which takes an image $x\in\mathbb{R}^{D}$ as input and outputs an embedding vector of dimension $M$. Suppose, for example, we have a $K$-way task $\mathcal{T}_{i}=(\mathcal{T}_{i}^{s},\mathcal{T}_{i}^{q})$ where $\mathcal{T}_{i}^{s}=\\{(x_{i,1},y_{i,1}),(x_{i,2},y_{i,2}),...,(x_{i,N},y_{i,N})\\}$, and where $y_{i,j}\in\\{1,...,K\\}$. Additionally, let $S_{k}\subset\mathcal{T}_{i}^{s}$ denote the set of support examples of class $k$. Then, a prototypical network computes a prototype $p_{k}$ for each class $k$ by computing a class-wise mean of embedded support examples: $p_{k}=\frac{1}{|S_{k}|}\sum_{(x,y)\in S_{k}}F_{\theta}(x)$ (1) Thus, in the case of ProtoNets, the fine-tuning algorithm $A$ does not update model parameters $\theta$, but instead it computes a set of prototypes which the base model will use to classify query data. We can think of $A$ as a function taking both embedding network parameters $\theta$ and support data $\mathcal{T}_{i}^{s}$ and returning a tuple $\theta_{i}$ consisting of a set of prototypes and an unchanged set of model parameters; i.e., $A(\theta,\mathcal{T}_{i}^{s})=(\\{p_{k}\\}_{i=0}^{k},\theta)=\theta_{i}$. In this way, $F_{\theta_{i}}$ in Algorithm 1 refers to using the base model parameters $\theta$ and the set of prototypes $\\{p_{k}\\}_{i=0}^{k}$ during inference. Given a distance function $d:\mathbb{R}^{M}\times\mathbb{R}^{M}\rightarrow[0,\infty)$ and a query point $x$, a ProtoNet produces a distribution over classes based on a softmax over distances to the prototypes in embedding space: $p_{\theta}(y=k|x)=\frac{\exp(-d(F_{\theta}(x),p_{k}))}{\sum_{k^{\prime}}\exp(-d(F_{\theta}(x),p_{k^{\prime}}))}$ (2) Training proceeds by minimizing the negative log-likelihood $\mathcal{L}(\theta)=-\log p_{\theta}(y=k|x)$ of the true class $k$ using SGD. Unfortunately, ProtoNet does not provide a way to understand the embedding space or visualize $p_{k}$ – a problem we directly address in this work. ### 2.2 Understanding Meta-learning Approaches Investigating the ability of meta-learning methods to adapt to new tasks has been the subject of numerous studies. The success of meta-learning approaches certainly seems to suggest that the representations learned by meta-learning must be different than those learned through standard training [9]. Goldblum et al. [9] find that meta-learned feature extractors outperform classically trained models of the same architecture and suggest that meta-learned features are qualitatively different from conventional features. While work has been done to understand how the meta-learning networks train [10][7], there has been little to no focus on developing tools to interpret the meta-learned models. ### 2.3 Interpretability in Convolutional Models In safety or security-critical applications, understanding why a classification system made a certain prediction is important. Just because a classification system is highly accurate, does not mean the network has learned the right kinds of features [11]. We believe that a system that can demonstrate its logic semantically or visually is more likely to be trusted and used. Being that a ProtoNet is primarily a convolutional neural network, it is appropriate to understand progress on interpretability of convolutional neural networks (CNN). There are many research branches within the umbrella of CNN interpretability including visualizations of intermediate network layers [25][16][19][21], diagnosis of CNN representations [27][26], and building explainable models [28]. In contrast to works which focus their attention on CNN layers and activations, we take a more specific approach in visualizing embedding space for ProtoNets. Zhang et al. [28] propose a compelling method of modifying convolutional layers so that each filter learns to represent a particular object part, thus allowing for each filter to correspond to a semantically meaningful image feature. We believe there could be interesting work incorporating this technique into meta-learning approaches, but is not appropriate for a shallow embedding network like the one we employ for ProtoNets. ### 2.4 Generative Models Work on Variational Prototyping Encoder (VPE) [12] is most similar to ours in that a meta-task is used to learn an embedding space suitable for both few- shot learning and unseen data representation. In contrast, we do not focus on the image translation task from real images to prototypes and instead focus our attention on visualizing prototypes for interpretability and refinement. There are also a number of works which investigate connections between autoencoder architectures and meta-learning, but which are not directly applicable for interpretability of few-shot image classification. For example, Wu et al. [24] propose the Meta-Learning Autoencoder (MeLA) framework which learns a recognition and generative model to transform a single-task model into one that can quickly adapt to new tasks using few examples. However, their framework is meant for the more general understanding of _tasks_ like physical state estimation and video prediction, as opposed to the image classification tasks which we focus on. Similarly, Epstein et al. [5] develop a meta-learning framework consisting of joint autoencoders for the purpose of learning multiple tasks simultaneously, but this approach is tailored more for the field of multi-task learning. ## 3 Algorithm Our interpretability algorithm takes advantage of the simplicity of the ProtoNet classification method. In particular, a ProtoNet classifies query data according to the class of the prototype which the query data’s embedding is nearest to, typically in Euclidean space. This classification method raises an obvious question: what does a prototype look like? To answer this question, we extend ProtoNets with a decoder to reconstruct images from embeddings. ### 3.1 Data The CIFAR-FS dataset [3] is a recent few-shot image classification benchmark consisting of all $100$ classes from CIFAR-100 [14]. Classes are randomly split into 64, 16, and 20 for meta-training, meta-validation, and meta-testing respectively. Every class contains $600$ images of size $32\times 32$. The _mini_ ImageNet dataset [23] is another standard benchmark for few-shot image classification. It consists of 100 randomly chosen classes from ILSVRC 2012 [4], which are split into 64, 16, and 20 classes for meta-training, meta- validation, and meta-testing respectively. For every class, there are $600$ images of size $84\times 84$. We adopt the commonly-used Ravi and Larochelle split proposed in [18]. ### 3.2 Architecture AutoProtoNet consists of an encoder-decoder architecture which compresses the input to produce an embedding which must be reconstructed by the decoder. There 4 sequential convolution blocks for the encoder and 4 sequential transpose convolution blocks for the decoder. The details of these blocks can be found in Table 2 of Appendix B. A forward pass through the model is shown in Figure 1. Output padding is used in the second transpose convolution block of the decoder to ensure that the output size of the final transpose convolution block matches the input $84\times 84$ dimensions of _mini_ ImageNet images, but no output padding modifications are necessary for CIFAR-FS images. Our architectural design choices imply that a $84\times 84$ _mini_ ImageNet image is embedded as $1600$-dimensional vector, while a $32\times 32$ CIFAR-FS image is embedded as $256$-dimensional vector. Figure 1: Visualization of the forward pass through AutoProtoNet. ### 3.3 Training Algorithm 2 AutoProtoNet Meta-Learning Input: Encoder and decoder networks, $F_{\theta}$ and $G_{\phi}$, where $\psi=[\theta;\phi]$ Input: Fine-tuning algorithm, $A$ Input: Reconstruction loss weight, $\lambda$ Input: Learning rate, $\gamma$ Input: Distribution over tasks, $p(\mathcal{T})$ 1:Initialize $\theta,\phi$, the weights of encoder and decoder 2:while not done do 3: Sample batch of tasks $\\{\mathcal{T}_{i}\\}_{i=1}^{n}$, where $\mathcal{T}_{i}\sim p(\mathcal{T})$ and $\mathcal{T}_{i}=(\mathcal{T}_{i}^{s},\mathcal{T}_{i}^{q})$ 4: for i=1,…,n do 5: $\hat{\mathcal{T}_{i}}\leftarrow G_{\phi}(F_{\theta}(\mathcal{T}_{i}))$ $\triangleright$ Reconstruct task data 6: $\mathcal{L}_{R}\leftarrow\mathrm{MSE}(\mathcal{T}_{i},\hat{\mathcal{T}_{i}})$ $\triangleright$ Compute reconstruction loss 7: $\theta_{i}\leftarrow A(\theta,\mathcal{T}_{i}^{s})$ $\triangleright$ Compute prototypes (inner loop) 8: $\mathcal{L}_{C}\leftarrow\mathrm{NLL}(F_{\theta_{i}},\mathcal{T}_{i}^{q})$ $\triangleright$ Compute classification loss 9: $\mathcal{L}\leftarrow\mathcal{L}_{C}+\lambda\mathcal{L}_{R}$ 10: $g_{i}\leftarrow\nabla_{\psi}\mathcal{L}$ 11: end for 12: $\psi\leftarrow\psi-\frac{\gamma}{n}\sum_{i}g_{i}$ $\triangleright$ Update base model parameters (outer loop) 13:end while Training AutoProtoNet is not much different from training a ProtoNet. The main difference is that we augment the meta-training loop with a reconstruction loss to regularize the embedding space and make it suitable for image reconstruction. We display the forward pass through AutoProtoNet in Figure 1 and adapt the meta-learning framwork from Section 2.1 to describe the meta- training of AutoProtoNet in Algorithm 2. Our “base” model now consists of parameters $\psi$ which is a concatenation of encoder network parameters $\theta$ and decoder network parameters $\phi$. In Line 5 of Algorithm 2, we pass both support and query data from the current task $\mathcal{T}_{i}$ through the encoder and decoder to produce a reconstruction $\hat{\mathcal{T}_{i}}$. This reconstruction is then compared to the original data using mean squared error (MSE) loss. The finetuning algorithm in Line 7 of Algorithm 2 is identical to the description in Section 2.1, where $\theta_{i}=(\\{p_{k}\\}_{i=0}^{k},\theta)$ is a tuple consisting of a set of prototypes for every class and the encoder network’s model parameters. Both of these are used to compute the likelihood of the true labels of our query data as in Equation 2, which is maximized by minimizing the negative log-likelihood (NLL). Finally, the classification loss $\mathcal{L}_{C}$ and the reconstruction loss $\mathcal{L}_{R}$ are summed so they can be jointly optimized. We meta-train ProtoNet and AutoProtoNet on both _mini_ ImageNet and CIFAR-FS. To create a prototype reconstruction baseline, we also train two models which make use of ILSVRC 2012 [4], which we refer to as ImageNet Autoencoder and ImageNet AutoProtoNet. Note that because _mini_ ImageNet is a subset of ILSVRC 2012, the ImageNet models also provide insight into whether more data during pretraining offers any benefit for meta-learning or prototype reconstructions. All training was performed on a single NVIDIA Quadro P6000 from our internal cluster. Training details for each model used in this work are described below. #### ProtoNet Using Algorithm 1, we meta-train a standard ProtoNet for 30 epochs using SGD. Our SGD optimizer uses Nesterov momentum of $0.9$, weight decay of $5\times 10^{-4}$, and a learning rate of $0.1$, which we decrease to $0.06$ after $20$ epochs. #### AutoProtoNet Using Algorithm 2, we meta-train an AutoProtoNet for 30 epochs using SGD. We use the same SGD settings as in ProtoNet training. We use a reconstruction loss weight $\lambda=1$. Following [20], both ProtoNet and AutoProtoNet models were trained using 20-way 5-shot episodes, where each class contains $15$ query points per episode, for $30$ epochs. #### ImageNet Autoencoder We train an autoencoder of the same architecture as AutoProtoNet using only mean squared error (MSE) loss on ILSVRC 2012 [4] for 20 epochs. We use the SGD optimizer with Nesterov momentum of $0.9$, weight decay of $5\times 10^{-4}$, and a learning rate of $0.1$, which we decrease by a factor of $10$ every $5$ epochs. To evaluate this model’s performance on benchmark few-shot image classification datasets, we make use of the only the encoder to produce embeddings and produce classification labels using the standard ProtoNet classification rule. #### ImageNet AutoProtoNet We use the encoder and decoder weights from the ImageNet Autoencoder as a starting point for the weights of an AutoProtoNet. All other training details are identical to that of AutoProtoNet, which we meta-train using Algorithm 2. The 5-way 5-shot test set accuracies of all models used in this work are shown in Table 1. AutoProtoNet is able to maintain the same level of few-shot image classification accuracy on benchmark datasets as a standard ProtoNet. While we expected AutoProtoNet to have an advantage due to having to incorporate features useful for reconstruction into embeddings, our results suggest that these reconstruction features are not always useful. Given the additional ILSVRC 2012 [4] data during pretraining, we also expected that ImageNet AutoProtoNet would outperform all other models, but our test results demonstrate that representations learned for image reconstruction are not too helpful for few-shot image classification. Test set accuracies for ImageNet Autoencoder underscore the point that an embedding space trained for only reconstruction is by no means competitive for few-shot classification, though it does achieve better than chance accuracy. ## 4 Experiments ### 4.1 Prototype Visualization While a standard ProtoNet employs an intuitive nearest-neighbor classification rule for query points, there is no intuitive way for a user to understand what a prototype embedding represents. Prototypical embeddings are crucial to understanding the decision boundaries of ProtoNets. The idea is that a ProtoNet embeds similar images nearby in embedding space, but without a way to visualize these embeddings, we argue that a human practitioner would be unable to debug or improve their deployed model. AutoProtoNet addresses this issue by learning an embedding space that is suitable for image reconstruction. Figure 2 displays prototype visualizations given a validation support set from _mini_ ImageNet and CIFAR-FS. The ImageNet Autoencoder (IA) and ImageNet AutoProtoNet (IAP) were both pretrained on all of ILSVRC 2012 [4], and so classes present in this validation support set are not novel classes because _mini_ ImageNet is a subset of ILSVRC 2012. However, in the case of the AutoProtoNet (AP), the classes in this validation support set are novel and the synthesized prototype images remain qualitatively on-par with the models trained with more data (such as ImageNet Autoencoder), suggesting that meta- tasks during training were sufficient to regularize an embedding space suitable for image synthesis. Analyzing the prototype reconstructions from CIFAR-FS in Figure 2(b), we see that prototype visualizations are generally too blurry to help a human determine whether the model has learned a sufficient representation of a class. We believe part of the problem is the low resolution and size of CIFAR-FS images. (a) _mini_ ImageNet (b) CIFAR-FS Figure 2: Support sets for a 5-way 5-shot validation task of _mini_ ImageNet (a) and CIFAR-FS (b). The embeddings of every image within a class are averaged to form a prototype embedding which is then synthesized as an image by using the decoder of an ImageNet Autoencoder (IA), an ImageNet AutoProtoNet (IAP), and an AutoProtoNet (AP). Table 1: 5-way 5-shot test set accuracies with 95% confidence intervals. Model | _mini_ ImageNet | CIFAR-FS ---|---|--- ImageNet Autoencoder | $36.83\pm 0.48$% | $46.08\pm 0.58$% ImageNet AutoProtoNet | $70.76\pm 0.51$% | $79.65\pm 0.52$% ProtoNet | $70.20\pm 0.52$% | $80.31\pm 0.51$% AutoProtoNet | $70.61\pm 0.52$% | $80.16\pm 0.52$% ### 4.2 Human-guided Prototype Refinement (a) Support set and prototype visualizations (b) New image and corresponding embedding (c) Interpolating 10 steps from initial prototype to new image embedding (d) New set of prototypes Figure 3: Steps for human-guided prototype selection in a 5-way 1-shot task. Step (a): a human chooses an initial prototype to refine. Step (b): a human captures one additional image to guide prototype refinement. Step (c): Interpolations between the initial prototype and the new image embedding (index 9) are shown to the human and a new prototype selection is made. Step (d): A new set of prototypes is set, with class 2 having been refined. To highlight the benefits of an embedding space suitable for image reconstruction, we designed an experiement to demonstrate how a human can guide prototype selection at test-time using AutoProtoNet. Assuming the user knows the kinds of images the model will encounter at inference time and given the ability to capture one more image, could we refine an initial prototype to achieve higher accuracy on the validation set? #### Data Collection Based on objects we had around the house, we chose to formulate a 5-way 1-shot classification problem between “door knob”, “frying pan”, “light switch”, “orange”, and “water bottle”. Note that “orange” and “frying pan” are classes in the _mini_ ImageNet training split, but all other classes are novel. Because we sought to demonstrate how one might use an AutoProtoNet in a real- world setting, all $55$ images in this task are novel, in-the-wild images, captured using an iPhone 12. Our support set consists of $5$ images ($1$ image per class). Our validation set consists of $50$ images ($10$ images per class) and can be found in Figure 4 of Appendix A. #### Prototype Refinement Prototype refinement is a debugging technique meant for cases in which a human believes prototype visualization may not be representative of the class. To exaggerate the idea of prototype refinement, we purposefully choose the back- side of a frying pan as a support image for class 1 (“frying pan”) so that the prototype visualization has undesirable image features. Generally, a prototype for an arbitrary object of a novel class is likely to be visually ambiguous if the embedding network did not train on a suitable dataset, so this setup is conceivable in the real-world. For our classification model, we make use of the AutoProtoNet described in Section 3.3. To apply AutoProtoNet to this new classification task, we “fine- tine” AutoProtoNet by providing a support set shown in Figure 3(a). After meta-learning, an AutoProtoNet’s only changeable parameters are its prototypes which, by design, can be reconstructed into images using the decoder. By visually understanding an AutoProtoNet’s embedding space, a user can choose to change image features of a prototype reconstruction, thus changing the prototype itself. In contrast, a standard ProtoNet performs inference using its support data, which is visually inaccessible and uninterpretable. Using a newly captured image $x\in\mathbb{R}^{d}$, we use the encoder $F_{\theta}$ to generate an embedding $p=F_{\theta}(x)$. Given an initial prototype $p_{k}$ for class $k$, we use the decoder $G_{\phi}$ to synthesize images $\hat{x}_{i}\in\mathbb{R}^{d}$ for interpolations between $p_{k}$ and $p$ as follows: $\hat{x}_{i}=G_{\phi}((1-\alpha)p_{k}+\alpha p)\qquad\alpha\in[0,1]$ (3) #### Results Using the initial prototypes from Figure 3(a), AutoProtoNet achieves $80\%$ accuracy on the validation set consisting of $50$ images from all $5$ classes. The $10$ misclassified images are all of the “frying pan” class. After debugging the “frying pan” prototype by capturing an additional image of a correctly-oriented frying pan and choosing an interpolation, the resulting embedding is used as the new support as shown in Figure 3(d). Under the new human-guided prototypes, AutoProtoNet achieves an accuracy of $98\%$ on the validation set, where the single misclassified image is of the “door knob” class. The novelty of our method lies in the ability for a human to fine-tune the model in an interactive way, leading to a performance increase in validation set accuracy. In this example, AutoProtoNet’s decoder allowed for the visualization of the prototype embedding, which we found to be visually incorrect. Thus, we captured an additional, more representative image to designate the direction in which to move the initial prototype to fit a human- designated criteria. ## 5 Conclusion With AutoProtoNet, we present a step toward meta-learning approaches capable of giving some insight into their learned parameters. We argue that if meta- learning approaches are to be useful in practice, there should be ways for a human to glean some insight into why a classification might have been made. Through prototype visualizations and a prototype refinement method, we highlight the benefits of AutoProtoNet and take steps to improve a simple few- shot classification algorithm by making it more interpretable while maintaining the same degree of accuracy as a standard ProtoNet. Our proposed method could likely be extended to Relation Networks [22], MetaOptNet [15], or R2D2 [3], with a decoder network to visualize embeddings. It may also be possible to meta-train a variational autoencoder to learn a latent space more suitable for detailed image synthesis. We believe generative models can play a larger role in interpretability of meta-learning algorithms. To confirm the effectiveness of our interpretability results, we intend to perform a human subjects study where a human determines whether prototype visualizations help in understanding classification results. We also recognize the limits of using a small dataset to evaluate the performance of our prototype refinement method. We leave the creation of a larger, more diverse validation set to future work. ## Acknowledgments and Disclosure of Funding This work was supported by the Office of Naval Research (WL). The views and conclusions contained in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the U.S. Navy. ## References * Altae-Tran et al. [2016] H. Altae-Tran, B. Ramsundar, A. S. Pappu, and V. S. Pande. Low data drug discovery with one-shot learning. _CoRR_ , abs/1611.03199, 2016. URL http://arxiv.org/abs/1611.03199. * Bengio [2012] Y. Bengio. Deep learning of representations for unsupervised and transfer learning. In _Proceedings of ICML workshop on unsupervised and transfer learning_ , pages 17–36. JMLR Workshop and Conference Proceedings, 2012. * Bertinetto et al. [2018] L. Bertinetto, J. F. Henriques, P. H. S. Torr, and A. Vedaldi. Meta-learning with differentiable closed-form solvers. _CoRR_ , abs/1805.08136, 2018. URL http://arxiv.org/abs/1805.08136. * Deng et al. [2009] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In _2009 IEEE Conference on Computer Vision and Pattern Recognition_ , pages 248–255, 2009. doi: 10.1109/CVPR.2009.5206848. * Epstein et al. [2019] B. Epstein, R. Meir, and T. Michaeli. Joint autoencoders: A flexible meta-learning framework. In M. Berlingerio, F. Bonchi, T. Gärtner, N. Hurley, and G. Ifrim, editors, _Machine Learning and Knowledge Discovery in Databases_ , pages 494–509, Cham, 2019. Springer International Publishing. ISBN 978-3-030-10925-7. * Finn et al. [2017] C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. _CoRR_ , abs/1703.03400, 2017. URL http://arxiv.org/abs/1703.03400. * Frosst et al. [2019] N. Frosst, N. Papernot, and G. Hinton. Analyzing and improving representations with the soft nearest neighbor loss, 2019. * Goldblum et al. [2019] M. Goldblum, L. Fowl, and T. Goldstein. Robust few-shot learning with adversarially queried meta-learners. _CoRR_ , abs/1910.00982, 2019. URL http://arxiv.org/abs/1910.00982. * Goldblum et al. [2020] M. Goldblum, S. Reich, L. Fowl, R. Ni, V. Cherepanova, and T. Goldstein. Unraveling meta-learning: Understanding feature representations for few-shot tasks. _CoRR_ , abs/2002.06753, 2020. URL https://arxiv.org/abs/2002.06753. * Huang et al. [2019] W. R. Huang, Z. Emam, M. Goldblum, L. Fowl, J. K. Terry, F. Huang, and T. Goldstein. Understanding generalization through visualizations. _CoRR_ , abs/1906.03291, 2019. URL http://arxiv.org/abs/1906.03291. * Ilyas et al. [2019] A. Ilyas, S. Santurkar, D. Tsipras, L. Engstrom, B. Tran, and A. Madry. Adversarial examples are not bugs, they are features, 2019. * Kim et al. [2019] J. Kim, T. Oh, S. Lee, F. Pan, and I. S. Kweon. Variational prototyping-encoder: One-shot learning with prototypical images. _CoRR_ , abs/1904.08482, 2019. URL http://arxiv.org/abs/1904.08482. * Kolesnikov et al. [2019] A. Kolesnikov, L. Beyer, X. Zhai, J. Puigcerver, J. Yung, S. Gelly, and N. Houlsby. Large scale learning of general visual representations for transfer. _CoRR_ , abs/1912.11370, 2019. URL http://arxiv.org/abs/1912.11370. * Krizhevsky [2009] A. Krizhevsky. Learning multiple layers of features from tiny images. 2009\. * Lee et al. [2019] K. Lee, S. Maji, A. Ravichandran, and S. Soatto. Meta-learning with differentiable convex optimization. _CoRR_ , abs/1904.03758, 2019. URL http://arxiv.org/abs/1904.03758. * Mahendran and Vedaldi [2014] A. Mahendran and A. Vedaldi. Understanding deep image representations by inverting them. _CoRR_ , abs/1412.0035, 2014. URL http://arxiv.org/abs/1412.0035. * Paszke et al. [2019] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, _Advances in Neural Information Processing Systems 32_ , pages 8024–8035. Curran Associates, Inc., 2019. URL http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf. * Ravi and Larochelle [2017] S. Ravi and H. Larochelle. Optimization as a model for few-shot learning. In _ICLR_ , 2017. * Simonyan et al. [2014] K. Simonyan, A. Vedaldi, and A. Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps, 2014. * Snell et al. [2017] J. Snell, K. Swersky, and R. S. Zemel. Prototypical networks for few-shot learning. _CoRR_ , abs/1703.05175, 2017. URL http://arxiv.org/abs/1703.05175. * Springenberg et al. [2015] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller. Striving for simplicity: The all convolutional net, 2015. * Sung et al. [2017] F. Sung, Y. Yang, L. Zhang, T. Xiang, P. H. S. Torr, and T. M. Hospedales. Learning to compare: Relation network for few-shot learning. _CoRR_ , abs/1711.06025, 2017. URL http://arxiv.org/abs/1711.06025. * Vinyals et al. [2016] O. Vinyals, C. Blundell, T. P. Lillicrap, K. Kavukcuoglu, and D. Wierstra. Matching networks for one shot learning. _CoRR_ , abs/1606.04080, 2016. URL http://arxiv.org/abs/1606.04080. * Wu et al. [2018] T. Wu, J. Peurifoy, I. L. Chuang, and M. Tegmark. Meta-learning autoencoders for few-shot prediction, 2018. * Zeiler and Fergus [2013] M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. _CoRR_ , abs/1311.2901, 2013. URL http://arxiv.org/abs/1311.2901. * Zhang et al. [2017a] Q. Zhang, R. Cao, F. Shi, Y. N. Wu, and S.-C. Zhu. Interpreting cnn knowledge via an explanatory graph, 2017a. * Zhang et al. [2017b] Q. Zhang, R. Cao, Y. N. Wu, and S.-C. Zhu. Growing interpretable part graphs on convnets via multi-shot learning, 2017b. * Zhang et al. [2018] Q. Zhang, Y. N. Wu, and S.-C. Zhu. Interpretable convolutional neural networks. In _Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition_ , pages 8827–8836, 2018. ## Appendix A Validation Set for Custom Classification Task In Figure 4, we display the $50$ images of our custom $5$-way validation set. The images from the “light switch” and “door knob” classes are diverse in terms of shape, pose, and lighting condition. Figure 4: Validation set for experiment described in Section 4.2 ## Appendix B Architecture Details In our description of the AutoProtoNet architecture in Table 2, we display output sizes for the first Conv Block of the encoder and the first Conv Transpose Block of the decoder, assuming an $84\times 84$ _mini_ ImageNet image is used as input. Table 2: AutoProtoNet Architecture Components Conv Block | | Conv Transpose Block ---|---|--- Layer | Parameters | Output Size | | Layer | Parameters | Output Size Conv | $3\times 3$, $64$ | $64\times 84\times 84$ | | Conv Transpose | $2\times 2$, $*2$ | $64\times 10\times 10$ Batch Norm | | | | Batch Norm | | Max Pool | $3\times 3$, /2 | $64\times 42\times 42$ | | Conv | $3\times 3$, $64$ | $64\times 10\times 10$ ## Appendix C Implementation Details We use PyTorch [17] and work on a fork of code used for [8], which uses the MIT License. Our fork can be used to reproduce experiments and is available here: https://github.com/psandovalsegura/AdversarialQuerying.
# ${\bf{\rm NK}_{1}}$ of Bak’s unitary group over Graded Rings Rabeya Basu Indian Institute of Science Education and Research (IISER) Pune, India<EMAIL_ADDRESS><EMAIL_ADDRESS>and Kuntal Chakraborty Indian Institute of Science Education and Research (IISER) Pune, India <EMAIL_ADDRESS> Research by the first author was supported by SERB-MATRICS grant for the financial year 2020–2021. And, research by the second author was supported by IISER (Pune) post-doctoral research grant. Corresponding Author<EMAIL_ADDRESS><EMAIL_ADDRESS> Abstract: For an associative ring $R$ with identity, we study the absence of $k$-torsion in ${\rm NK_{1}GQ}(R)$; Bass nil-groups for the general quadratic or Bak’s unitary groups. By using a graded version of Quillen–Suslin theory we deduce an analog for the graded rings. 2020 Mathematics Subject Classification: 19-XX, 15-XX, 16-XX, 20-XX Key words: General linear groups, Elementary subgroups, Quadratic forms, Higman linearisation, $k$-torsion, Whitehead group - $\rm K_{1}$. ## 1\. Introduction Let $R$ be an associative ring with identity element $1$. When $R$ is commutative, we define ${\rm SK_{1}}(R)$ as the kernel of the determinant map from the Whitehead group ${\rm K_{1}}(R)$ to the group of units of $R$. The Bass nil-group ${\rm NK_{1}}(R)=\textnormal{ker}({\rm K_{1}}(R[X])\rightarrow{\rm K_{1}}(R))$; $X=0$. i.e., the subgroup consisting of elements $[\alpha(X)]\in{\rm K_{1}}(R[X])$ such that $[\alpha(0)]=[{\rm I}]$. Hence ${\rm K_{1}}(R[X])\cong{\rm NK}_{1}(R)\oplus{\rm K_{1}}(R)$. The aim of this paper is to study some properties of Bass nil-groups ${\rm NK_{1}}$ for the general quadratic groups or Bak’s unitary groups. It is well-known that for many rings, e.g. if $R$ is regular Noetherian, Dedekind domain, or any ring with finite global dimension, the group ${\rm NK}_{1}(R)$ is trivial. On the other hand, if it is non-trivial, then it is not finitely generated as a group. e.g. if $G$ is a non-trivial finite group, the group ring $\mathbb{Z}G$ is not regular. In many such cases, it is difficult to compute ${\rm NK}_{1}(\mathbb{Z}[G])$. In [13], D.R. Harmon proved the triviality of this group when $G$ is finite group of square free order. C. Weibel, in [20], has shown the non-triviality of this group for $G$ = $\mathbb{Z}/2\oplus\mathbb{Z}/2$, $\mathbb{Z}/4$, and $D_{4}$. Some more results are known for finite abelian groups from the work of R.D. Martin; cf.[16]. It is also known (cf.[12]) that for a general finite group $G$, ${\rm NK}_{1}(R[G])$ is a torsion group for the group ring $R[G]$. In fact, for trivial ${\rm NK}_{1}(R)$, every element of finite order of ${\rm NK}_{1}(R[G])$ is some power of the cardinality of $G$. For $R=\mathbb{Z}$, this is a result of Weibel. In particular, if $G$ is a finite $p$-group ($p$ a prime), then every element of ${\rm NK}_{1}(\mathbb{Z}[G])$ has $p$-primary order. In [17], J. Stienstra showed that ${\rm NK_{1}}(R)$ is a ${\rm W}(R)$-module, where ${\rm W}(R)$ is the ring of big Witt vectors (cf.[11] and [19]). Consequently, in ([18], §3), C. Weibel observed that if $k$ is a unit in $R$, then ${\rm SK_{1}}(R[X])$ has no $k$-torsion, when $R$ is a commutative local ring. Note that if $R$ is a commutative local ring then ${\rm SK_{1}}(R[X])$ coincides with ${\rm NK_{1}}(R)$; indeed, if $R$ is a local ring then ${\rm SL}_{n}(R)={\rm E}_{n}(R)$ for all $n>0$. Therefore, we may replace $\alpha(X)$ by $\alpha(X)\alpha(0)^{-1}$ and assume that $[\alpha(0)]=[{\rm I}]$. In [7], the first author extended Weibel’s result for arbitrary associative rings. In this paper we prove the analog result for $\lambda$-unitary Bass nil-groups, viz. ${\rm NK_{1}GQ}^{\lambda}(R,\Lambda)$, where $(R,\Lambda)$ is the form ring as introduced by A. Bak in [1]. The main ingredient for our proof is an analog of Higman linearisation (for a subclass of Bak’s unitary group) due to V. Kopeiko; cf.[15]. For the general linear groups, Higman linearisation (cf.[6]) allows us to show that ${\rm NK_{1}}(R)$ has a unipotent representative. The same result is not true in general for the unitary nil-groups. Kopeiko’s results in [15] explain a complete description of the elements of ${\rm NK_{1}GQ}^{\lambda}(R,\Lambda)$ that have (unitary) unipotent representatives. Followings are the main results in this article. ###### Theorem 1.1. Let $[\alpha(X)]=\big{[}\begin{pmatrix}A(X)&B(X)\\\ C(X)&D(X)\end{pmatrix}\big{]}\in{\rm NK_{1}GQ}^{\lambda}(R,\Lambda)$ with $A(X)\in{\rm GL}_{r}(R[X])$ for some $r\in\mathbb{N}$. Then $[\alpha(X)]$ has no $k$-torsion if $kR=R$. And, an analog for the graded rings: ###### Theorem 1.2. Let $R=R_{0}\oplus R_{1}\oplus\dots$ be a graded ring. Let $k$ be a unit in $R_{0}$. Let $N=N_{0}+N_{1}+\dots+N_{r}\in{\rm M}_{r}(R)$ be a nilpotent matrix, and ${\rm I}$ denote the identity matrix. If $[({\rm I}+N)]^{k}=[{\rm I}]$ in ${\rm K_{1}GQ}^{\lambda}(R,\Lambda)$, then $[{\rm I}+N]=[{\rm I}+N_{0}]$. In the proof of 1.2, we have used a graded version of Quillen–Suslin’s local- global principle for Bak’s unitary group over graded rings. This unify and generalize the results proved in [5], [7], [9], and [10]. ###### Theorem 1.3. (Graded local-global principle) Let $R=R_{0}\oplus R_{1}\oplus R_{2}\oplus\cdots$ be a graded ring with identity $1$. Let $\alpha\in{\rm GQ}(2n,R,\Lambda)$ be such that $\alpha\equiv{\rm I}_{2n}\pmod{R_{+}}$. If $\alpha_{\mathfrak{m}}\in{\rm EQ}(2n,R_{\mathfrak{m}},\Lambda_{\mathfrak{m}})$, for every maximal ideal $\mathfrak{m}\in\rm Max(C(R_{0}))$, then $\alpha\in{\rm EQ}(2n,R,\Lambda).$ ## 2\. Preliminaries Let $R$ be an associative ring with identity element $1$. Let ${\rm M}(n,R)$ denote the additive group of $n\times n$ matrices, and ${\rm GL}(n,R)$ denote the multiplicative group of $n\times n$ invertible matrices. Let $e_{ij}$ be the matrix with $1$ in the $ij$-th position and $0$’s elsewhere. The elementary subgroup of ${\rm GL}(n,R)$ plays a key role in classical algebraic K-theory. We recall, ###### Definition 2.1. Elementary Group ${\rm E}(n,R)$: The subgroup of all matrices in ${\rm GL}(n,R)$ generated by $\\{{\rm E}_{ij}(\lambda):\lambda\in R,i\neq j\\}$, where ${\rm E}_{ij}(\lambda)={\rm I}_{n}+\lambda e_{ij}$, and $e_{ij}$ is the matrix with $1$ in the $ij$-position and $0$’s elsewhere. ###### Definition 2.2. For $\alpha\in{\rm M}(r,R)$ and $\beta\in{\rm M}(s,R)$, the matrix $\alpha\perp\beta$ denotes its embedding in ${\rm M}(r+s,R)$ (here $r$ and $s$ are even integers in the non-linear cases), given by $\alpha\perp\beta=\left(\begin{array}[]{cc}\alpha&0\\\ 0&\beta\end{array}\right).$ There is an infinite counterpart: Identifying each matrix $\alpha\in{\rm GL}(n,R)$ with the large matrix $(\alpha\perp\\{1\\})$ gives an embedding of ${\rm GL}(n,R)$ into ${\rm GL}(n+1,R)$. Let ${\rm GL}(R)=\underset{n=1}{\overset{\infty}{\cup}}{\rm GL}(n,R)$, and ${\rm E}(R)=\underset{n=1}{\overset{\infty}{\cup}}{\rm E}(n,R)$ be the corresponding infinite linear groups. As a consequence of classical Whitehead Lemma (cf.[3]) due to A. Suslin, one gets $[{\rm GL}(R),{\rm GL}(R)]={\rm E}(R).$ ###### Definition 2.3. The quotient group ${\rm K_{1}}(R)=\frac{{\rm GL}(R)}{[{\rm GL}(R),{\rm GL}(R)]}=\frac{{\rm GL}(R)}{{\rm E}(R)}$ is called the Whitehead group of the ring $R$. For $\alpha\in{\rm GL}(n,R)$, let $[\alpha]$ denote its equivalence class in ${\rm K_{1}}(R)$. In the similar manner we define ${\rm K_{1}}$ group for the other types of classical groups; viz., the symplectic Whitehead group ${\rm K_{1}}{\rm Sp}(R)$ and the orthogonal Whitehead group ${\rm K_{1}}{\rm O}(R)$. This paper explores a uniform framework for classical type groups over graded structures. Let us begin by recalling the concept of form rings and form parameter as introduced by A. Bak in [1]. This allows us to give a uniform definition for classical type groups. ###### Definition 2.4. (Form rings): Let $R$ be an associative ring with identity, and with an involution $-:R\rightarrow R$, $a\mapsto\overline{a}$. Let $\lambda\in C(R)$ = the center of $R$, with the property $\lambda\overline{\lambda}=1$. We define two additive subgroups of $R$, viz. $\Lambda_{\rm max}=\\{a\in R\mid a=-\lambda\overline{a}\\}~{}\textit{and}~{}\Lambda_{\rm min}=\\{a-\lambda\overline{a}\mid a\in R\\}.$ One checks that for any $x\in R$, $\Lambda_{\rm max}$ and $\Lambda_{\rm min}$ are closed under the conjugation operation $a\mapsto\overline{x}ax$. A $\lambda$-form parameter on $R$ is an additive subgroup $\Lambda$ of $R$ such that $\Lambda_{\rm min}\subseteq\Lambda\subseteq\Lambda_{\rm max}$, and $\overline{x}\Lambda x\subseteq\Lambda$ for all $x\in R$. i.e., a subgroup between two additive groups which is also closed under the conjugation operation. A pair $(R,\Lambda)$ is called a form ring. To define Bak’s unitary group or the general quadratic group, we fix a central element $\lambda\in R$ with $\lambda\overline{\lambda}=1$, and then consider the form $\psi_{n}=\begin{pmatrix}0&{\rm I}_{n}\\\ \lambda{{\rm I}}_{n}&0\end{pmatrix}.$ For more details, see [7], and [8]. Bak’s Unitary or General Quadratic Groups ${\rm GQ}$: ${\rm GQ}(2n,R,\Lambda)~{}=~{}\\{\sigma\in{\rm GL}(2n,R,\Lambda)\,|\,\overline{\sigma}\psi_{n}\sigma=\psi_{n}\\}.$ ### Elementary Quadratic Matrices : Let $\rho$ be the permutation, defined by $\rho(i)=n+i$ for $i=1,\dots,n$. For $a\in R$, and $1\leq i,j\leq n$, we define $q\varepsilon_{ij}(a)={\rm I}_{2n}+ae_{ij}-\overline{a}e_{\rho(j)\rho(i)}$ for $i\neq j$, $qr_{ij}(a)=\left\\{\begin{array}[]{ll}{\rm I}_{2n}+ae_{i\rho(j)}-\lambda\overline{a}e_{j\rho(i)}&\text{for}~{}i\neq j\\\ {\rm I}_{2n}+ae_{\rho(i)j}&\text{for}~{}i=j\end{array}\right.$ $ql_{ij}(a)=\left\\{\begin{array}[]{ll}{\rm I}_{2n}+ae_{\rho(i)j}-\overline{\lambda}\overline{a}e_{\rho(j)i}&\text{for}~{}i\neq j\\\ {\rm I}_{2n}+ae_{\rho(i)j}&\text{for}~{}i=j\end{array}\right.$ (Note that for the second and third type of elementary matrices, if $i=j$, then we get $a=-\lambda\overline{a}$, and hence it forces that $a\in\Lambda_{\rm max}(R)$. One checks that these above matrices belong to $\rm GQ(2n,R,\Lambda)$; cf.[1]. $n$-th Elementary Quadratic Group ${\rm EQ}(2n,R,\Lambda)$: The subgroup generated by $q\varepsilon_{ij}(a),qr_{ij}(a)\text{and }ql_{ij}(a)$, for $a\in R$ and $1\leq i,j\leq n$. For uniformity we denote the elementary generators of ${\rm EQ}(2n,R,\Lambda)$ by $\eta_{ij}(*)$. Stabilization map: There are standard embeddings: ${\rm GQ}(2n,R,\Lambda)\longrightarrow{\rm GQ}(2n+2,R,\Lambda)$ given by $\begin{pmatrix}a&b\\\ c&d\end{pmatrix}\mapsto\begin{pmatrix}a&0&b&0\\\ 0&1&0&0\\\ c&0&d&0\\\ 0&0&0&1\end{pmatrix}.$ Hence we have ${\rm GQ}(R,\Lambda)=\underset{\longrightarrow}{\lim}\,\,{\rm GQ}(2n,R,\Lambda)$. It is clear that the stabilization map takes generators of ${\rm EQ}(2n,R,\Lambda)$ to the generators of ${\rm EQ}(2(n+1),R,\Lambda)$. Hence we have ${\rm EQ}(R,\Lambda)=\underset{\longrightarrow}{\lim}\,\,{\rm EQ}(2n,R,\Lambda)$ There are standard formulas for the commutators between quadratic elementary matrices. For details, we refer [1] (Lemma 3.16). In later sections there are repeated use of those relations. The analogue of the Whitehead Lemma for the general quadratic groups (cf.[1]) due to Bak allows us to write: $[{\rm GQ}(R,\Lambda),{\rm GQ}(R,\Lambda)]=[{\rm EQ}(R,\Lambda),{\rm EQ}(R,\Lambda)]={\rm EQ}(R,\Lambda).$ Hence we define the Whitehead group of the general quadratic group ${\rm K_{1}}{\rm GQ}=\frac{{\rm GQ}(R,\Lambda)}{{\rm EQ}(R,\Lambda)}.$ And, the Whitehead group at the level $m$ ${\rm K}_{1,m}{\rm GQ}=\frac{{\rm GQ}_{m}(R,\Lambda)}{{\rm EQ}_{m}(R,\Lambda)},$ where $m=2n$ in the non-linear cases. Let $(R,\Lambda)$ be a form ring. We extend the involution of $R$ to the ring $R[X]$ of polynomials by setting $\overline{X}=X$. As a result we obtain a form ring $(R[X],\Lambda[X])$. ###### Definition 2.5. The kernel of the group homomorphism ${\rm K_{1}GQ}(R[X],\Lambda[X])\rightarrow{\rm K_{1}GQ}(R,\Lambda)$ induced from the form ring homomorphism $(R[X],\Lambda[X])\rightarrow(R,\Lambda):X\mapsto 0$ is denoted by ${\rm NK_{1}GQ}(R,\Lambda)$. We often say it as Bass nilpotent unitary ${\rm K_{1}}$-group of $R$, or just unitary nil-group. From the definition it follows that ${\rm K_{1}GQ}(R[X],\Lambda[X])={\rm K_{1}GQ}(R,\Lambda)\oplus{\rm NK_{1}GQ}(R,\Lambda).$ In this context, we will use following two types of localizations, mainly over graded ring $R=R_{0}\oplus R_{1}\oplus R_{2}\oplus\cdots$. 1. (1) Principal localization: for a non-nilpotent, non-zero divisor $s$ in $R_{0}$ with $\overline{s}=s$, we consider the multiplicative subgroup $S=\\{1,s,s^{2},\dots\\}$, and denote localized form ring by $(R_{s},\Lambda_{s})$. 2. (2) Maximal localization: for a maximal ideal $\mathfrak{m}\in\rm Max(R_{0})$, we take the multiplicative subgroup $S=R_{0}-\mathfrak{m}$, and denote the localized form ring by $(R_{\mathfrak{m}},\Lambda_{\mathfrak{m}})$. Blanket assumption: We always assume that $2n\geq 6$. Next, we recall the well-known “Swan–Weibel’s homotopy trick”, which is the main ingredient to handle the graded case. Let $R=R_{0}\oplus R_{1}\oplus R_{2}\oplus\cdots$ be a graded ring. An element $a\in R$ will be denoted by $a=a_{0}+a_{1}+a_{2}+\cdots$, where $a_{i}\in R_{i}$ for each $i$, and all but finitely many $a_{i}$ are zero. Let $R_{+}=R_{1}\oplus R_{2}\oplus\cdots$. Graded structure of $R$ induces a graded structure on ${\rm M}_{n}(R)$ (ring of $n\times n$ matrices). ###### Definition 2.6. Let $a\in R_{0}$ be a fixed element. We fix an element $b=b_{0}+b_{1}+\cdots$ in $R$ and define a ring homomorphism $\epsilon:R\rightarrow R[X]$ given by $\epsilon(b)=\epsilon(b_{0}+b_{1}+\cdots)\;=\;b_{0}+b_{1}X+b_{2}X^{2}+\cdots+b_{i}X^{i}+\cdots.$ Then we evaluate the polynomial $\epsilon(b)(X)$ at $X=a$ and denote the image by $b^{+}(a)$ i.e., $b^{+}(a)=\epsilon(b)(a)$. Note that $\big{(}b^{+}(x)\big{)}^{+}(y)=b^{+}(xy)$. Observe, $b_{0}=b^{+}(0)$. We shall use this fact frequently. The above ring homomorphism $\epsilon$ induces a group homomorphism at the ${\rm GL}(2n,R)$ level for every $n\geq 1$, i.e., for $\alpha\in{\rm GL}(2n,R)$ we get a map $\epsilon:{\rm GL}(2n,R,\Lambda)\rightarrow{\rm GL}(2n,R[X],\Lambda[X])\text{ defined by}$ $\alpha=\alpha_{0}\oplus\alpha_{1}\oplus\alpha_{2}\oplus\cdots\mapsto\alpha_{0}\oplus\alpha_{1}X\oplus\alpha_{2}X^{2}\cdots,$ where $\alpha_{i}\in{\rm M}(2n,R_{i})$. As above for $a\in R_{0}$, we define $\alpha^{+}(a)$ as $\alpha^{+}(a)=\epsilon(\alpha)(a).$ ###### Notation 2.7. By ${\rm GQ}(2n,R[X],\Lambda[X],(X))$ we shall mean the group of all quadratic matrices over $R[X]$, which are ${\rm I}_{2n}$ modulo $(X)$. Also if $R$ is a graded ring, then by ${\rm GQ}(2n,R,\Lambda,(R_{+}))$ we shall mean the group of all quadratic matrices over $R$ which are ${\rm I}_{2n}$ modulo $R_{+}$. The following lemma highlights very crucial fact which we use (repeatedly) in the proof of “Dilation Lemma”. ###### Lemma 2.8. Let $R$ be a Noetherian ring and $s\in R$. Then there exists a natural number $k$ such that the homomorphism ${\rm GQ}(2n,R,\Lambda,s^{k}R)\rightarrow{\rm GQ}(2n,R_{s},\Lambda_{s})$ $($induced by localization homomorphism $R\rightarrow R_{s})$ is injective. Moreover, it follows that the induced map ${\rm EQ}(2n,R,\Lambda,s^{k}R)\rightarrow{\rm EQ}(2n,R_{s},\Lambda_{s})$ is injective. For the proof of the above lemma we refer [14], Lemma 5.1. Recall that any module finite ring $R$ is direct limit of its finitely generated subrings. Also, ${\rm G}(R,\Lambda)=\underset{\longrightarrow}{\lim}\,{\rm G}(R_{i},\Lambda_{i})$, where the limit is taken over all finitely generated subring of $R$. Thus, one may assume that $C(R)$ is Noetherian. Hence one may consider module finite (form) rings $(R,\Lambda)$ with identity. Now we recall few technical definitions and useful lemmas. ###### Definition 2.9. A row $(a_{1},a_{2},\dots,a_{n})\in R^{n}$ is said to be unimodular if there exists $(b_{1},b_{2},\dots,b_{n})\in R^{n}$ such that $\sum_{i=1}^{n}a_{i}b_{i}=1$. The set of all unimodular rows of length $n$ is denoted by ${\rm Um}_{n}(R)$. For any column vector $v\in(R^{2n})^{t}$ we define the row vector $\widetilde{v}=\overline{v}^{t}\psi_{n}$. ###### Definition 2.10. We define the map $M:(R^{2n})^{t}\times(R^{2n})^{t}\rightarrow M(2n,R)$ and the inner product $\langle,\rangle$ as follows: $\displaystyle M(v,w)$ $\displaystyle=v.\widetilde{w}-\overline{\lambda}\,\overline{w}.\widetilde{v}$ $\displaystyle\langle v,w\rangle$ $\displaystyle=\widetilde{v}.w$ Note that the elementary generators of the groups ${\rm EQ}(2n,R,\Lambda)$ are of the form ${\rm I}_{2n}+M(*_{1},*_{2})$ for suitably chosen standard basis vectors. ###### Lemma 2.11. $($cf.[1]$)$ The group ${\rm E}(2n,R,\Lambda)$ is perfect for $n\geq 3$, i.e., $[{\rm EQ}(2n,R,\Lambda),{\rm EQ}(2n,R,\Lambda)]={\rm EQ}(2n,R,\Lambda).$ ###### Lemma 2.12. For all elementary generators of ${\rm GQ}(2n,R,\Lambda)$ we have the following splitting property: for all $x,y\in R$, $\eta_{ij}(x+y)=\eta_{ij}(x)\eta_{ij}(y).$ ${{\bf Proof:}}$ See pg. 43-44, Lemma 3.16, [1]. ###### Lemma 2.13. Let $G$ be a group, and $a_{i},b_{i}\in G$, for $i=1,2,\ldots,n$. Then for $r_{i}=\Pi_{j=1}^{i}a_{j}$, we have $\Pi_{i=1}^{n}r_{i}b_{i}r_{i}^{-1}\Pi_{i=1}^{n}a_{i}$. ###### Lemma 2.14. The group ${\rm GQ}(2n,R,\Lambda,R_{+})\cap{\rm EQ}(2n,R,\Lambda)$ generated by the elements of the type $\varepsilon\eta_{ij}(*)\varepsilon^{-1}$, where $\varepsilon\in{\rm EQ}(2n,R,\Lambda)$ and $*\in R_{+}$. ${{\bf Proof:}}$ Let $\alpha\in{\rm GQ}(2n,R,\Lambda,R_{+})\cap{\rm EQ}(2n,R,\Lambda)$. Then we can write $\alpha=\Pi_{k=1}^{r}\eta_{i_{k}j_{k}}(a_{k})$ for some element $a_{k}\in R$, $k=1,\dots,r$. We can write $a_{k}$ as $a_{k}=(a_{0})_{k}+(a_{+})_{k}$ for some $(a_{0})_{k}\in R_{0}$ and $(a_{+})_{k}\in R_{+}$. Using Lemma 2.12, we can write $\alpha$ as, $\alpha=\Pi_{k=1}^{r}(\eta_{i_{k}j_{k}}(a_{0})_{k})(\eta_{i_{k}j_{k}}(a_{+})_{k}).$ Let $\epsilon_{t}=\Pi_{k=1}^{t}\eta_{i_{k}j_{k}}((a_{0})_{k})$ for $1\leq t\leq r$. By the Lemma 2.13, we have $\alpha=\left(\Pi_{k=1}^{r}\epsilon_{k}\eta_{i_{k}j_{k}}((a_{+})_{k})\epsilon_{k}^{-1}\right)\left(\Pi_{k=1}^{r}\eta_{i_{k}j_{k}}((a_{0})_{k})\right).$ Let us write $A=\Pi_{k=1}^{r}\epsilon_{k}\eta_{i_{k}j_{k}}((a_{+})_{k})\epsilon_{k}^{-1}$ and $B=\Pi_{k=1}^{r}\eta_{i_{k}j_{k}}((a_{0})_{k})$. Hence $\alpha=AB$. Let ‘over-line’ denotes the quotient ring modulo $R_{+}$. Now going modulo $R_{+}$, we have $\overline{\alpha}=\overline{AB}=\bar{A}\bar{B}=\overline{\rm I}_{2n}\bar{B}=\overline{\rm I}_{2n}$, the last equality holds as $\alpha\in{\rm GQ}(2n,R,\Lambda,R_{+})$. Hence, $\overline{B}=\overline{{\rm I}}_{2n}$. Since the entries of $B$ are in $R_{0}$, it follows that $B={\rm I}_{2n}$. Therefore it follows that $\alpha=\Pi_{k=1}^{r}\epsilon_{k}\eta_{i_{k}j_{k}}((a_{+})_{k})\epsilon_{k}^{-1}.$ $\Box$ ## 3\. Quillen–Suslin Theory for Bak’s Group over Graded Rings ### 3.1. Local–Global Principle ###### Lemma 3.1. Let $(R,\Lambda)$ be a form ring and $v\in{\rm EQ}(2n,R,\Lambda)e_{1}$. Let $w\in R^{2n}$ be a column vector such that $\langle v,w\rangle=0$. Then ${\rm I}_{2n}+M(v,w)\in{\rm EQ}(2n,R,\Lambda)$. ${{\bf Proof:}}$ Let $v=\varepsilon e_{1}$. Then we have ${\rm I}_{2n}+M(v,w)=\varepsilon({\rm I}_{2n}+M(e_{1},w_{1}))\varepsilon^{-1}$, where $w_{1}=\varepsilon^{-1}w$. Since $\langle e_{1},w_{1}\rangle=\langle v,w\rangle=0$, we have $w_{1}^{T}=(w_{11},\dots,w_{1n-1},0,\dots,w_{12n})$. Therefore, since $\lambda\bar{\lambda}=1$, we have ${\rm I}_{2n}+M(v,w)=\prod_{\begin{subarray}{c}1\leq j\leq n\\\ 1\leq i\leq n-1\end{subarray}}\varepsilon ql_{in}(-\bar{\lambda}\overline{w}_{1n+i})q\varepsilon_{jn}(-\bar{\lambda}\overline{w}_{1j})ql_{nn}^{-1}(*)\varepsilon^{-1}$ ###### Lemma 3.2. Let $R$ be a graded ring. Let $\alpha\in{\rm EQ}(2n,R,\Lambda)$. Then for every $a\in R_{0}$ one gets $\alpha^{+}(a)\in{\rm EQ}(2n,R,\Lambda)$. ${{\bf Proof:}}$ Let $\alpha=\Pi_{k=1}^{t}({\rm I}_{2n}+aM(e_{i_{k}},e_{j_{k}})),$ where $a\in R$ and $t\geq 1$. Hence for $b\in R_{0}$, we have $\alpha^{+}(b)=\Pi_{k=1}^{t}({\rm I}_{2n}+a^{+}(b)M(e_{i_{k}},e_{j_{k}}))$. Now taking $v=e_{i}$ and $v=a^{+}(b)e_{j}$ we have $\langle v,w\rangle=0$ and ${\rm I}_{2n}+M(v,w)={\rm I}_{2n}+a^{+}(b)M(e_{i},e_{j}))$ which belongs to ${\rm EQ}(2n,R,\Lambda)$ by Corollary 3.1. Hence we have $\alpha^{+}(b)\in{\rm EQ}(2n,R,\Lambda)$ for $b\in R_{0}$. $\Box$ ###### Lemma 3.3. (Graded Dilation Lemma) Let $\alpha\in{\rm GQ}(2n,R,\Lambda)$ with $\alpha^{+}(0)={\rm I}_{2n}$ and $\alpha_{s}\in{\rm EQ}(2n,R_{s},\Lambda_{s})$ for some non-zero-divisor $s\in R_{0}$. Then there exists $\beta\in{\rm EQ}(2n,R,\Lambda)$ such that $\beta_{s}^{+}(b)=\alpha_{s}^{+}(b)$ for some $b=s^{l}$ and $l\gg 0$. ${{\bf Proof:}}$ Since $\alpha_{s}\in{\rm EQ}(2n,R_{s},\Lambda_{s})$ with $(\alpha_{0})_{s}={\rm I}_{2n}$, then $\alpha_{s}=\gamma$, where $\gamma_{ii}=1+g_{ii}$ where $g_{ii}\in(R^{+})_{s}$ and $\gamma_{ij}=g_{ij}$ for $i\neq j$, where $g_{ij}\in(R^{+})_{s}$. Choose $l$ large enough such that every denominator of $g_{ij}$ for all $i,j$ divides $s^{l}$. Then by Lemma 3.2, we have $\alpha_{s}^{+}(s^{l})\in{\rm EQ}(2n,R_{s},\Lambda_{s})$. As all denominator is cleared then $\alpha_{s}^{+}(s^{l})$ permits a natural pullback. Hence we have $\alpha^{+}(s^{l})\in{\rm EQ}(2n,R,\Lambda).$ Call this pullback as $\beta$. $\Box$ ###### Lemma 3.4. Let $\alpha_{s}\in{\rm EQ}(2n,R_{s},\Lambda_{s})$ with $\alpha_{s}^{+}(0)={\rm I}_{2n}$. Then one gets $\alpha_{s}^{+}(b+d)\alpha_{s}^{+}(d)^{-1}\in{\rm EQ}(2n,R,\Lambda)$ for some $s,d\in R_{0}$ and $b=s^{l},l\gg 0$. ${{\bf Proof:}}$ Since $\alpha_{s}\in{\rm EQ}(2n,R_{s},\Lambda_{s})$, we have $\alpha_{s}^{+}(X)\in{\rm EQ}(2n,R_{s}[X],\Lambda_{s}[X])$. Let $\beta^{+}(X)=\alpha^{+}(X+d)\alpha^{+}(d)^{-1},$ where $d\in R_{0}$. Then we have $\beta^{+}_{s}(X)\in{\rm EQ}(2n,R_{s}[X],\Lambda_{s}[X])$ and $\beta^{+}(0)={\rm I}_{2n}$. Hence by Lemma 3.3, we have, there exists $b=s^{l}$, $l\gg 0$, such that $\beta^{+}(bX)\in{\rm EQ}(2n,R[X],\Lambda[X])$. Putting $X=1$, we get our desired result. $\Box$ Proof of Theorem 1.3 – Graded Local-Global Principle: Since $\alpha_{\mathfrak{m}}\in{\rm EQ}(2n,R_{\mathfrak{m}},\Lambda_{\mathfrak{m}})$ for all $\mathfrak{m}\in{\rm Max}(C(R_{0}))$, for each $\mathfrak{m}$ there exists $s\in C(R_{0})\setminus\mathfrak{m}$ such that $\alpha_{s}\in{\rm EQ}(2n,R_{s},\Lambda_{s})$. Using Noetherian property we can consider a finite cover of $C(R_{0})$, say $s_{1}+\dots+s_{r}=1$. From Lemma 3.3, we have $\alpha^{+}(b_{i})\in{\rm EQ}(2n,R,\Lambda)$ for some $b_{i}=s_{i}^{l_{i}}$ with $b_{1}+\dots+b_{r}=1$. Now consider $\alpha_{s_{1}s_{2}\dots s_{r}}$, which is the image of $\alpha$ in $R_{s_{1}s_{2}\dots s_{r}}$. By Lemma 2.8, $\alpha\mapsto\alpha_{s_{1}s_{2}\dots s_{r}}$ is injective. Hence we can perform our calculation in $R_{s_{1}s_{2}\dots s_{r}}$ and then pull it back to $R$. $\begin{split}\alpha_{s_{1}s_{2}\dots s_{r}}=&\alpha_{s_{1}s_{2}\dots s_{r}}^{+}(b_{1}+b_{2}+\dots+b_{r})\\\ =&((\alpha_{s_{1}})_{s_{2}s_{3}\dots})^{+}(b_{1}+\dots+b_{r})((\alpha_{s_{1}})_{s_{2}s_{3}\dots})^{+}(b_{2}+\dots+b_{r})^{-1}\dots\\\ &((\alpha_{s_{i}})_{s_{1}\dots\hat{s_{i}}\dots s_{r}})^{+}(b_{i}+\dots+b_{r})((\alpha_{s_{i}})_{s_{1}\dots\hat{s_{i}}\dots s_{r}})^{+}(b_{i+1}+\dots+b_{r})^{-1}\\\ &((\alpha_{s_{r}})_{s_{1}s_{2}\dots s_{r-1}})^{+}(b_{r})((\alpha_{s_{r}})_{s_{1}s_{2}\dots s_{r-1}})^{+}(0)^{-1}\end{split}$ Observe that $((\alpha_{s_{i}})_{s_{1}\dots\hat{s_{i}}\dots s_{r}})^{+}(b_{i}+\dots+b_{r})((\alpha_{s_{i}})_{s_{1}\dots\hat{s_{i}}\dots s_{r}})^{+}(b_{i+1}+\dots+b_{r})^{-1}\in{\rm EQ}(2n,R,\Lambda)$ due to Lemma 3.4 (here $\hat{s_{i}}$ means we omit $s_{i}$ in the product $s_{1}\dots\hat{s_{i}}\dots s_{r}$), and hence $\alpha_{s_{1}s_{2}\dots s_{r}}\in{\rm EQ}(2n,R_{s_{1}\dots s_{r}},\Lambda_{s_{1}\dots s_{r}})$. This proves $\alpha\in{\rm EQ}(2n,R,\Lambda)$. $\Box$ ### 3.2. Normality and Local–Global Next we are going to show that if $K$ is a commutative ring with identity and $R$ is an associative $K$-algebra such that $R$ is finite as a left $K$-module, then the normality criterion of elementary subgroup is equivalent to the Local-Global principle for quadratic group. (One can also consider $R$ as a right $K$-algebra.) ###### Lemma 3.5. $($Bass; cf.[4]$)$ Let $A$ be an associative $B$-algebra such that $A$ is finite as a left $B$-module and $B$ be a commutative local ring with identity. Then $A$ is semilocal. ###### Theorem 3.6. $($cf.[7]$)$ Let $A$ be a semilocal ring $($not necessarily commutative$)$ with involution. Let $v\in{\rm Um}_{2n}(A)$.Then $v\in e_{1}{\rm EQ}(2n,A)$. In other words the group ${\rm EQ}(2n,A)$ acts transitively on ${\rm Um}_{2n}(A)$. Before proving the next theorem we need to recall a theorem from [7]: ###### Theorem 3.7. $($Local-Global Principle$)$ Let $A$ be an associative $B$-algebra such that $A$ is finite as a left $B$-module and $B$ be a commutative ring with identity.. If $\alpha(X)\in{\rm GQ}(2n,A[X],\Lambda[X])$, $\alpha(0)=\rm{\rm I}_{2n}$ and $\alpha_{\mathfrak{m}}(X)\in{\rm EQ}(2n,A_{\mathfrak{m}}[X],\Lambda_{\mathfrak{m}}[X])$ for every maximal ideal $\mathfrak{m}\in{\rm Max}(B)$, then $\alpha\in{\rm EQ}(2n,A[X],\Lambda[X])$. ###### Theorem 3.8. Let $K$ be a commutative ring with unity and $R=\oplus_{i=0}^{\infty}R_{i}$ be a graded $K$-algebra such that $R_{0}$ is finite as a left $K$-module. Then for $n\geq 3$ the following are equivalent: $(1)$ ${\rm EQ}(2n,R,\Lambda)$ is a normal subgroup of ${\rm GQ}(2n,R,\Lambda)$. $(2)$ If $\alpha\in{\rm GQ}(2n,R,\Lambda)$ with $\alpha^{+}(0)={\rm I}_{2n}$ and $\alpha_{\mathfrak{m}}\in{\rm EQ}(2n,R_{\mathfrak{m}},\Lambda_{\mathfrak{m}})$ for every maximal ideal $\mathfrak{m}\in{\rm Max}(K)$, then $\alpha\in{\rm EQ}(2n,R,\Lambda)$. ${{\bf Proof:}}$ $(1)\Rightarrow(2)$ We have proved the Lemma 3.1 for any form ring with identity and shown that the local-global principle is a consequence of Lemma 3.1. So, the result is true in particular if we have ${\rm EQ}(2n,R,\Lambda)$ is a normal subgroup of ${\rm GQ}(2n,R,\Lambda)$. $(2)\Rightarrow(1)$ Since polynomial rings are special case of graded rings, the result follows by using the Theorem 3.7. Let $\alpha\in{\rm EQ}(2n,R,\Lambda)$ and $\beta\in{\rm GQ}(2n,R,\Lambda)$. Then we have $\alpha$ can be written as product of the matrices of the form $({\rm I}_{2n}+\beta M(*_{1},*_{2})\beta^{-1})$, with $\langle*_{1},*_{2}\rangle=0$ where $*_{1}$ and $*_{2}$ are suitably chosen basis vectors. Let $v=\beta*_{1}$. Then we can write $\beta\alpha\beta^{-1}$ as a product of the matrices of the form ${\rm I}_{2n}+M(v,w)$ for some $w\in R^{2n}$. We must show that each ${\rm I}_{2n}+M(v,w)\in{\rm EQ}(2n,R,\Lambda)$. Consider $\gamma={\rm I}_{2n}+M(v,w)$. Then $\gamma^{+}(0)={\rm I}_{2n}$. By Lemma 3.5 we have the ring $S^{-1}R$ is semilocal where $S=K\setminus\mathfrak{m}$, and $\mathfrak{m}\in{\rm Max}(K)$. Since $v\in{\rm Um}_{2n}(R)$, then by Theorem 3.6, we have $v\in{\rm EQ}(2n,S^{-1}R,S^{-1}\Lambda)e_{1}$. Therefore by applying Lemma 3.1 to the ring $(S^{-1}R,S^{-1}\Lambda)$, we have $\gamma_{\mathfrak{m}}\in{\rm EQ}(2n,R_{\mathfrak{m}},\Lambda_{\mathfrak{m}})$ for every maximal ideal $\mathfrak{m}\in{\rm Max}(K)$. Hence by hypothesis we have $\gamma\in{\rm EQ}(2n,R,\Lambda)$. This completes the proof. $\Box$ ###### Remark 3.9. We conclude that the local-global principle for the elementary subgroups and their normality properties are equivalent. ## 4\. Bass Nil Group ${\rm{\rm NK_{1}}{\rm GQ}(R)}$ In this section recall some basic definitions and properties of the representatives of ${\rm{\rm NK_{1}}{\rm GQ}(R)}$. We represent any element of ${\rm M}_{2n}(R)$ as $\begin{pmatrix}a&b\\\ c&d\end{pmatrix},$ where $a,b,c,d\in{\rm M}_{n}(R)$. For $\alpha=\begin{pmatrix}a&b\\\ c&d\end{pmatrix}$ we call $\begin{pmatrix}a&b\end{pmatrix}$ the upper half of $\alpha$. Let $(R,\lambda,\Lambda)$ be a form ring. By setting $\bar{\Lambda}=\\{\bar{a}:a\in\Lambda\\}$ we get another form ring $(R,\bar{\lambda},\bar{\Lambda})$. We can extend the involution of $R$ to ${\rm M}_{n}(R)$ by setting $(a_{ij})^{*}=(\overline{a}_{ji})$. ###### Definition 4.1. Let $(R,\lambda,\Lambda)$ be a form ring. A matrix $\alpha=(a_{ij})\in{\rm M}_{n}(R)$ is said to be $\Lambda$-Hermitian if $\alpha=-\lambda\alpha^{*}$ and all the diagonal entries of $\alpha$ are contained in $\Lambda$. A matrix $\beta\in{\rm M}_{n}(R)$ is said to be $\bar{\Lambda}$-Hermitian if $\beta=-\bar{\lambda}\beta^{*}$ and all the diagonal entries of $\beta$ are contained in $\bar{\Lambda}$. ###### Remark 4.2. A matrix $\alpha\in{\rm M}_{n}(R)$ is $\Lambda$-Hermitian if and only if $\alpha^{*}$ is $\bar{\Lambda}$-Hermitian. The set of all $\Lambda$-Hermitian matrices forms a group under matrix multiplication. ###### Lemma 4.3. [15, Example 2] Let $\beta\in{\rm GL}_{n}(R)$ be a $\Lambda$-Hermitian matrix. Then the matrix $\alpha^{*}\beta\alpha$ is $\Lambda$-Hermitian for every $\alpha\in{\rm GL}_{n}(R)$. ###### Definition 4.4. Let $\alpha=\begin{pmatrix}a&b\\\ c&d\end{pmatrix}\in{\rm M}_{2n}(R)$ be a matrix. Then $\alpha$ is said to be a $\Lambda$-quadratic matrix if one of the following equivalent conditions holds: 1. (1) $\alpha\in{\rm GQ}(2n,R,\Lambda)$ and the diagonal entries of the matrices $a^{*}c,b^{*}d$ are in $\Lambda$, 2. (2) $a^{*}d+\lambda c^{*}d={\rm I}_{n}$ and the matrices $a^{*}c,b^{*}d$ are $\Lambda$-Hermitian, 3. (3) $\alpha\in{\rm GQ}(2n,R,\Lambda)$ and the diagonal entries of the matrices $ab^{*},cd^{*}$ are in $\Lambda$, 4. (4) $ad^{*}+\lambda bc^{*}={\rm I}_{n}$ and the matrices $ab^{*},cd^{*}$ are $\Lambda$-Hermitian. ###### Remark 4.5. The set of all $\Lambda$-quadratic matrices of order $2n$ forms a group called $\Lambda$-quadratic group. We denote this group by ${\rm GQ}^{\lambda}(2n,R,\Lambda)$. If $2\in R^{*}$, then we have $\Lambda_{\rm min}=\Lambda_{\rm max}$. In this case notions of quadratic groups and notions of $\Lambda$-quadratic groups coincides. Also this happens when $\Lambda=\Lambda_{\rm max}$. Hence quadratic groups are special cases of $\Lambda$-quadratic groups. Other classical groups appear as $\Lambda$-quadratic groups in the following way. Let $R$ be a commutative ring with trivial involution. Then ${\rm GQ}^{\lambda}(2n,R,\Lambda)=\begin{cases}{\rm Sp}_{2n}(R),&\text{if }\lambda=-1\text{ and }\Lambda=\Lambda_{\rm max}=R\\\ {\rm O}_{2n}(R),&\text{if }\lambda=1\text{ and }\Lambda=\Lambda_{\rm min}=0\end{cases}$ And for general linear group ${\rm GL}_{n}(R)$, we have, ${\rm GL}_{n}(R)={\rm GQ}^{1}(2n,H(R),\Lambda=\Lambda_{\rm max})$, where $\mathbb{H}(R)$ denotes the ring $R\oplus R^{op}$ with $R^{op}$ is the opposite ring of $R$ and the involution on $\mathbb{H}(R)$ is defined by $\overline{(x,y)}=(y,x)$. Thus the study of $\Lambda$-quadratic matrices unifies the study of quadratic matrices. We recall following results from [15]. ###### Lemma 4.6. Let $\alpha=\begin{pmatrix}a&0\\\ 0&d\end{pmatrix}\in{\rm M}_{2n}(R)$. Then $\alpha\in{\rm GQ}^{\lambda}(2n,R,\Lambda)$ if and only if $a\in{\rm GL}_{n}(R)$ and $d=(a^{*})^{-1}$. ${{\bf Proof:}}$ Let $\alpha\in{\rm GQ}^{\lambda}(2n,R,\Lambda)$. In view of $(2)$ of Definition 4.4, we have, $a^{*}d={\rm I}_{n}$. Hence $a$ is invertible and $d=(a^{*})^{-1}$. Converse holds by $(2)$ of Definition 4.4. $\Box$ ###### Definition 4.7. Let $\alpha\in{\rm GL}_{n}(R)$ be a matrix. A matrix of the form $\begin{pmatrix}\alpha&0\\\ 0&(\alpha^{*})^{-1}\end{pmatrix}$ is denoted by $\mathbb{H}(\alpha)$ and is said to be hyperbolic. ###### Remark 4.8. In a similar way we can show that matrices of the form $T_{12}(\beta):=\begin{pmatrix}{\rm I}_{n}&\beta\\\ 0&{\rm I}_{n}\end{pmatrix}$ is $\Lambda$-quadratic matrix if and only if $\beta$ is $\bar{\Lambda}$-Hermitian. And the matrix of the form $T_{21}(\gamma):=\begin{pmatrix}{\rm I}_{n}&0\\\ \gamma&{\rm I}_{n}\end{pmatrix}$ is $\Lambda$-quadratic matrix if and only if $\gamma$ is $\Lambda$-Hermitian. Likewise in the quadratic case we can define the notion of $\Lambda$-elementary quadratic groups in the following way: ###### Definition 4.9. The $\Lambda$-elementary quadratic group is denoted by ${\rm EQ}^{\lambda}(2n,R,\Lambda)$ and defined by the group generated by $2n\times 2n$ matrices of the form $\mathbb{H}(\alpha)$ where $\alpha\in{\rm E}_{n}(R)$, $T_{12}(\beta)$ and $\beta$ is $\bar{\Lambda}$-Hermitian and $T_{21}(\gamma)$ is $\gamma$ $\Lambda$-Hermitian. ###### Lemma 4.10. Let $A=\begin{pmatrix}\alpha&\beta\\\ 0&\delta\end{pmatrix}\in{\rm M}_{2n}(R)$. Then $A\in{\rm GQ}^{\lambda}(2n,R,\Lambda)$ if and only if $\alpha\in{\rm GL}_{n}(R)$, $\delta=(\alpha^{*})^{-1}$ and $\alpha^{-1}\beta$ is $\bar{\Lambda}$-Hermitian. In this case $A\equiv\mathbb{H}(\alpha)\pmod{{\rm EQ}^{\lambda}(2n,R,\Lambda)}$. ${{\bf Proof:}}$ Let $A\in{\rm GQ}^{\lambda}(2n,R,\Lambda)$. Then by $(4)$ of Definition 4.4, we have $\alpha\delta^{*}={\rm I}_{n}$ and $\alpha\beta^{*}$ is $\Lambda$-Hermitian. Hence $\alpha$ is invertible and $\delta=(\alpha^{*})^{-1}$. For $\alpha^{-1}\beta$, we get $(\alpha^{-1}\beta)^{*}=\beta^{*}(\alpha^{-1})^{*}=\alpha^{-1}(\alpha\beta^{*})(\alpha^{-1})^{*},$ which is $\Lambda$-Hermitian by Lemma 4.3. Hence $\alpha^{-1}\beta$ is $\bar{\Lambda}$-Hermitian. Conversely, the condition on $A$ will fulfill the condition $(4)$ of Definition 4.4. Hence $A$ is $\Lambda$-quadratic. Since $\alpha^{-1}\beta$ is $\bar{\Lambda}$-Hermitian, $T_{12}(-\alpha^{-1}\beta)\in{\rm EQ}^{\lambda}(2n,R,\Lambda)$ and $AT_{12}(\alpha^{-1}\beta)=\mathbb{H}(\alpha)$. Thus $A\equiv\mathbb{H}(\alpha)\pmod{{\rm EQ}^{\lambda}(2n,R,\Lambda)}$. $\Box$ A similar proof will prove the following: ###### Lemma 4.11. Let $B=\begin{pmatrix}\alpha&0\\\ \gamma&\delta\end{pmatrix}\in{\rm M}_{2n}(R)$. Then $B\in{\rm GQ}^{\lambda}(2n,R,\Lambda)$ if and only if $\alpha\in{\rm GL}_{n}(R)$, $\delta=(\alpha^{*})^{-1}$ and $\gamma$ is $\Lambda$-Hermitian. In this case $B\equiv\mathbb{H}(\alpha)\pmod{{\rm EQ}^{\lambda}(2n,R,\Lambda)}.$ ###### Lemma 4.12. Let $\alpha=\begin{pmatrix}a&b\\\ c&d\end{pmatrix}\in{\rm GQ}^{\lambda}(2n,R,\Lambda)$. Then $\alpha\equiv\mathbb{H}(a)\pmod{{\rm EQ}^{\lambda}(4n,R,\Lambda)}$ if $a\in{\rm GL}_{n}(R).$ Moreover, if $a\in{\rm E}_{n}(R)$, then $\alpha\equiv\mathbb{H}(a)\pmod{{\rm EQ}^{\lambda}(2n,R,\Lambda)}$. ${{\bf Proof:}}$ By same argument as given in Lemma 4.10, we have $a^{-1}b$ is $\Lambda$-Hermitian. Hence $T_{12}(-a^{-1}b)\in{\rm EQ}^{\lambda}(2n,R,\Lambda)$, and consequently $\alpha T_{12}(-a^{-1}b)=\begin{pmatrix}a&0\\\ c&d^{\prime}\end{pmatrix}\in{\rm GQ}^{\lambda}(2n,R,\Lambda)$ for some $d^{\prime}\in{\rm GL}_{n}(R)$. Hence by Lemma 4.11, we get $\alpha T_{12}(-a^{-1}b)\equiv H(a)\pmod{{\rm EQ}^{\lambda}(2n,R,\Lambda)}.$ Hence $\alpha\equiv H(a)\pmod{{\rm EQ}^{\lambda}(2n,R,\Lambda)}$. $\Box$ ###### Definition 4.13. Let $\alpha=\begin{pmatrix}a_{1}&b_{1}\\\ c_{1}&d_{1}\end{pmatrix}\in{\rm M}_{2r}(R)$, $\beta=\begin{pmatrix}a_{2}&b_{2}\\\ c_{2}&d_{2}\end{pmatrix}\in{\rm M}_{2s}(R)$. As before, we define $\alpha\perp\beta$, and consider an embedding $\rm GQ^{\lambda}(2n,R,\Lambda)\rightarrow\rm GQ^{\lambda}(2n+2,R,\Lambda),\,\,\,\alpha\mapsto\alpha\perp{\rm I}_{2}.$ We denote ${\rm GQ}^{\lambda}(R,\Lambda)=\underset{n=1}{\overset{\infty}{\cup}}{\rm GQ}^{\lambda}(2n,R,\Lambda)$ and ${\rm EQ}^{\lambda}(R,\Lambda)=\underset{n=1}{\overset{\infty}{\cup}}{\rm EQ}^{\lambda}(2n,R,\Lambda)$. In view of quadratic analog of Whitehead Lemma, we have the group ${\rm EQ}^{\lambda}(R,\Lambda)$ coincides with the commutator of ${\rm GQ}^{\lambda}(R,\Lambda)$. Therefore the group ${\rm K_{1}}{\rm GQ}^{\lambda}(R,\Lambda):=\frac{{\rm GQ}^{\lambda}(R,\Lambda)}{{\rm EQ}^{\lambda}(R,\Lambda)}$ is well-defined. The class of a matrix $\alpha\in{\rm GQ}^{\lambda}(R,\Lambda)$ in the group ${\rm K_{1}}{\rm GQ}^{\lambda}(R,\Lambda)$ is denoted by $[\alpha]$. In this way we obtain a ${\rm K_{1}}$-functor ${\rm K_{1}}{\rm GQ}^{\lambda}$ acting form the category of form rings to the category of abelian groups. ###### Remark 4.14. Likewise in the quadratic case, the kernel of the group homomorphism ${\rm K_{1}GQ}^{\lambda}(R[X],\Lambda[X])\rightarrow{\rm K_{1}GQ}^{\lambda}(R,\Lambda)$ induced from the form ring homomorphism $(R[X],\Lambda[X])\rightarrow(R,\Lambda);X\mapsto 0$ is denoted by ${\rm NK_{1}GQ}^{\lambda}(R,\Lambda)$. Since the $\Lambda$-quadratic groups are subclass of the quadratic groups, the Local-global principle holds for $\Lambda$-quadratic groups. We use this throughout for the next section. ## 5\. Absence of torsion in ${\rm NK_{1}}{\rm GQ}^{\lambda}(R,\Lambda)$ In this section we give the proof of Theorem 1.1 and Theorem 1.2. In [6], the proof of the theorem for the linear case is based on two key results, viz. the Higman linearisation, and a lemma on polynomial identity in the truncated polynomial rings. Here we recall the lemma with its proof to highlight its connection with the big Witt vectors. Recently, in [15], V. Kopeiko deduced an analog of Higman linearisation process for a subclass of the general quadratic groups. ###### Definition 5.1. For a associative ring $R$ with unity we consider the truncated polynomial ring $R_{t}=\frac{R[X]}{(X^{t+1})}.$ ###### Lemma 5.2. $($cf.[6], Lemma 4.1$)$ Let $P(X)\in R[X]$ be any polynomial. Then the following identity holds in the ring $R_{t}:$ $(1+X^{r}P(X))=(1+X^{r}P(0))(1+X^{r+1}Q(X)),$ where $r>0$ and $Q(X)\in R[X]$, with $\deg(Q(X))<t-r$. ${{\bf Proof:}}$ Let us write $P(X)=a_{0}+a_{1}X+\cdots+a_{t}X^{t}$. Then we can write $P(X)=P(0)+XP^{\prime}(X)$ for some $P^{\prime}(X)\in R[X]$. Now, in $R_{t}$ $\displaystyle(1+X^{r}P(X))(1+X^{r}P(0))^{-1}$ $\displaystyle=(1+X^{r}P(0)+X^{r+1}P^{\prime}(X))(1+X^{r}P(0))^{-1}$ $\displaystyle=1+X^{r+1}P^{\prime}(X)(1-X^{r}P(0)+X^{2r}(P(0))^{2}-\cdots)$ $\displaystyle=1+X^{r+1}Q(X)$ ​​where $Q(X)\in R[X]$ with $\deg(Q(X))<t-r$. Hence the lemma follows. $\Box$ Remark. Iterating the above process we can write for any polynomial $P(X)\in R[X]$, $(1+XP(X))=\Pi_{i=1}^{t}(1+a_{i}X^{i})$ in $R_{t}$, for some $a_{i}\in R$. By ascending induction it will follow that the $a_{i}$’s are uniquely determined. In fact, if $R$ is commutative then $a_{i}$’s are the $i$-th component of the ghost vector corresponding to the big Witt vector of $(1+XP(X))\in{\rm W}(R)=(1+XR[[X]])^{\times}$. For details see ([11], $\mathcal{x}$I). ###### Lemma 5.3. Let $R$ be a ring with $1/k\in R$ and $P(X)\in R[X]$. Assume $P(0)$ lies in the center of $R$. Then $(1+X^{r}P(X))^{k^{r}}=1\Rightarrow(1+X^{r}P(X))=(1+X^{r+1}Q(X))$ in the ring $R_{t}$ for some $r>0$ and $Q(X)\in R[X]$ with $\deg(Q(X))<t-r$. Following result is due to V. Kopeiko, cf. [15]. ###### Proposition 5.4. (Higman linearisation) Let $(R,\Lambda)$ be a form ring. Then, every element of the group ${\rm NK_{1}}{\rm GQ}^{\lambda}(R,\Lambda)$ has a representative of the form $[a;b,c]_{n}=\begin{pmatrix}{\rm I}_{r}-aX&bX\\\ -cX^{n}&{\rm I}_{r}+a^{*}X+\cdots+(a^{*})^{n}X^{n}\end{pmatrix}\in{\rm GQ}^{\lambda}(2r,R[X],\Lambda[X])$ for some positive integers $r$ and $n$, where $a,b,c\in{\rm M}_{r}(R)$ satisfy the following conditions: 1. (1) the matrices $b$ and $ab$ are Hermitian and also $ab=ba^{*}$, 2. (2) the matrices $c$ and $ca$ are Hermitian and also $ca=a^{*}c$, 3. (3) $bc=a^{n+1}$and $cb=(a^{*})^{n+1}$. ###### Corollary 5.5. Let $[\alpha]\in{\rm NK_{1}}{\rm GQ}^{\lambda}(R,\Lambda)$ has the representation $[a;b,c]_{n}$ for some $a,b,c\in{\rm M}_{n}(R)$ according to Proposition 5.4. Then $[\alpha]=[\mathbb{H}({\rm I}_{r}-aX)]$ in ${\rm NK_{1}}{\rm GQ}^{\lambda}(R,\Lambda)$ if $({\rm I}_{r}-aX)\in{\rm GL}_{r}(R)$. ${{\bf Proof:}}$ By Lemma 4.12 we have $[a;b,c]_{n}\equiv\mathbb{H}({\rm I}_{r}-aX)\pmod{{\rm EQ}^{\lambda}(2r,R[X],\Lambda[X])}$. Hence we have $[\alpha]=[\mathbb{H}({\rm I}_{r}-aX)]$ in ${\rm NK_{1}}{\rm GQ}^{\lambda}(R,\Lambda)$. $\Box$ Proof of Theorem 1.1: By the Theorem 5.4, we have $[\alpha]=[[a;b,c]_{n}]$ for some $a,b,c\in{\rm M}_{s}(R)$ and for some natural numbers $n$ and $s$. Note that in the Step $1$ of the Proposition 5.4, the invertibility of the first corner of the matrix $\alpha$ will not be changed during the linearisation process. Also the invertibility of the first corner is preserved in the remaining steps of the Proposition 5.4. Therefore since the first corner matrix $A(X)\in{\rm GL}_{r}(R[X])$, then we have $({\rm I}_{s}-aX)\in{\rm GL}_{s}(R[X])$. By Corollary 5.5, we have $[\alpha]=[\mathbb{H}({\rm I}_{s}-aX)]$. Now let $[\alpha]$ be a $k$-torsion. Then we have $[\mathbb{H}({\rm I}_{r}-aX)]$ is a $k$-torsion. Since $({\rm I}_{r}-aX)$ is invertible, it follows that $a$ is nilpotent. Let $a^{t+1}=0$. Since $[({\rm I}_{r}-aX)]^{k}=[{\rm I}]$ in ${\rm K_{1}}{\rm GQ}^{\lambda}(R[X],\Lambda[X])$, then by arguing as given in [7], we have $[{\rm I}_{r}-aX]=[I]$ in ${\rm K_{1}}{\rm GQ}^{\lambda}(R[X],\Lambda[X])$. This completes the proof. $\Box$ Proof of Theorem 1.2 – (Graded Version): Consider the ring homomorphism $f:R\rightarrow R[X]$ defined by $f(a_{0}+a_{1}+\dots)=a_{0}+a_{1}X+\dots.$ Then $\displaystyle[({\rm I}+N)^{k}]=[{\rm I}]$ $\displaystyle\Rightarrow f([{\rm I}+N]^{k})=[f({\rm I}+N)]^{k}=[{\rm I}]$ $\displaystyle\Rightarrow[({\rm I}+N_{0}+N_{1}X+\dots+N_{r}X^{r})]^{k}=[{\rm I}].$ Let $\mathfrak{m}$ be a maximal ideal in $R_{0}$. By Theorem 1.1, we have $[({\rm I}+N_{0}+N_{1}X+\dots+N_{r}X^{r})]=[{\rm I}]$ in ${\rm NK_{1}}{\rm GQ}^{\lambda}((R)_{\mathfrak{m}},\Lambda_{\mathfrak{m}})$. Hence by using the local-global principle we conclude $[({\rm I}+N)]=[{\rm I}+N_{0}]$ in ${\rm NK_{1}}{\rm GQ}^{\lambda}(R,\Lambda)$, as required. $\Box$ Acknowledgment: We thank Sergey Sinchuk and V. Kopeiko for many useful discussions. ## References * [1] A. Bak; ${\rm K}$-Theory of forms. Annals of Mathematics Studies, 98\. Princeton University Press, Princeton, N.J. University of Tokyo Press, Tokyo (1981). * [2] A. Bak, R. Basu, R.A. Rao; Local-global principle for transvection groups. Proceedings of The American Mathematical Society 138 (2010), no. 4, 1191–1204. * [3] H. Bass; Algebraic K-Theory, Benjamin, New York-Amsterdam (1968). * [4] H. Bass; K-Theory and stable algebra, Publ. Math. I.H.E.S. No. 22 (1964), 5–60. * [5] R. Basu, R.A. Rao, Reema Khanna; On Quillen’s local global principle. Commutative algebra and algebraic geometry, Contemp. Math., 390, Amer. Math. Soc., Providence, RI (2005), 17–30. * [6] R. Basu; Absence of torsion for $NK_{1}(R)$ over associative rings, J. Algebra Appl. 10(4) (2011), 793–799. * [7] R. Basu; Local-global principle for general quadratic and general Hermitian groups and the nilpotence of ${\rm KH}_{1}$. Zap. Nauchn. Sem. S.-Petersburg. Otdel. Mat. Inst. Steklov. (POMI) 452 (2016), Voprosy Teorii Predstavleniĭ Algebr i Grupp. 30, 5–31. translation in J. Math. Sci. (N.Y.) 232 (2018), no. 5, 591–609. * [8] R. Basu; On Transvection Subgroups of General Quadratic Modules. Journal of Algebra and Its Application. Vol. 17, No. 11, 1850217 (2018). * [9] R. Basu, Ravi A. Rao, Reema Khanna; Pillars of relative Quillen–Suslin Theory. “Leavitt Path Algebras and K-theory”, ISI Series, Springer (2020) 211–223. * [10] R. Basu, Manish Kumar Singh; On Quillen–Suslin Theory for Classical Groups; Revisited over Graded Rings. Contemp. Math. Amer. Math. Soc., Vol. 751, (2020), 5–18. * [11] S. Bloch; Algebraic ${\rm K}$-Theory and crystalline cohomology, Publ. Math. I.H.E.S. 47 (1977), 187–268. itary $K_{1}$-group of unitary ring, Journal of Mathematical Sciences, Vol 240, No 4 (2019), 459–473. * [12] I. Hambleton, W. Lück; Induction and computation of Bass Nil groups for finite groups. Pure Appl. Math. Q. 8 (2012), no. 1, 199–219. * [13] D.R. Harmon; ${\rm NK}_{1}$ of finite groups. Proc. Amer. Math. Soc. 100(2) (1987), 229–232. * [14] R. Hazrat, N. Vavilov; ${\rm K_{1}}$ of Chevalley groups are nilpotent. Journal of Pure and Applied Algebra 179 (2003), no. 1-2, 99–116. * [15] V.I. Kopeiko; Bass nilpotent unitary ${\rm K_{1}}$ group of unitary ring. Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 460 (2017), Voprosy Teorii Predstavleniĭ Algebr i Grupp. 33, 111–119. translation in J. Math. Sci. (N.Y.) 243 (2019), no. 4, 577–582. * [16] R.D. Martin; Nilgroups of finite abelian groups. Ph.D.Thesis, Columbia Iniversity, ProQuest LLC, Ann Arbor, MI, 1976. * [17] J. Stienstra; Operation in the linear ${\rm K}$-theory of endomorphisms, Current Trends in Algebraic Topology, Conf. Proc. Can. Math. Soc. 2 (1982). * [18] C. Weibel; Mayer-Vietoris Sequence and module structure on ${\rm K_{1}0}$, Springer Lecture Notes in Mathematics 854 (1981), 466–498. * [19] C. Weibel; Module Structures on the ${\rm K}$-theory of Graded Rings, Journal of Algebra 105 (1987), 465–483. * [20] C. Weibel; ${\rm NK}_{0}$ and ${\rm NK}_{1}$ of the groups $C_{4}$ and $D_{4}$. Addendum to ”Lower algebraic K-theory of hyperbolic 3-simplex reflection groups” by J.-F. Lafont and I. J. Ortiz [MR2495796]. Comment. Math. Helv. 84 (2009), no. 2, 339–349.
# Resilience through Scene Context in Visual Referring Expression Generation Simeon Junker Sina Zarrieß Bielefeld University <EMAIL_ADDRESS> ###### Abstract Scene context is well known to facilitate humans’ perception of visible objects. In this paper, we investigate the role of context in Referring Expression Generation (REG) for objects in images, where existing research has often focused on distractor contexts that exert pressure on the generator. We take a new perspective on scene context in REG and hypothesize that contextual information can be conceived of as a resource that makes REG models more resilient and facilitates the generation of object descriptions, and object types in particular. We train and test Transformer-based REG models with target representations that have been artificially obscured with noise to varying degrees. We evaluate how properties of the models’ visual context affect their processing and performance. Our results show that even simple scene contexts make models surprisingly resilient to perturbations, to the extent that they can identify referent types even when visual information about the target is completely missing. ## 1 Introduction Objects do not appear randomly in the world that surrounds us, but they occur in predictable spatial, semantic, or functional configurations and relations to their environment. Research on human perception shows that we “see the world in scenes” (Bar, 2004), and that prior experience and knowledge of the world helps us to efficiently process visual stimuli. Even with an extremely short glimpse at an image, humans remember essential semantic aspects of the scene and object arrangement (Oliva and Torralba, 2006). This rapid scene understanding allows us to handle the complexity of the visual world and to recognize objects in context, e.g., when they are not fully visible (Võ, 2021). In this paper, we take a new perspective on how visual context facilitates the automatic generation of descriptions for visual objects, a task well-known as Referring Expression Generation (REG). TRFv-0.0 | couch on right (A) ---|--- TRFv-0.5 | right brown chair (F) TRFv-1.0 | right couch (A) TRFt-0.0 | right couch (A) TRFt-0.5 | right couch (A) TRFt-1.0 | right elephant (F) TRFs-0.0 | couch on right (A) TRFs-0.5 | right couch (A) TRFs-1.0 | right couch (A) Figure 1: Example from RefCOCO testB (displayed with noise level $0.5$) with generated expressions and human judgements. Scene context enables target identification even with full occlusion (TRFv-1.0, TRFs-1.0) In past years, datasets have become available that provide referring expressions for objects in images, with objects appearing in relatively complex real-world contexts (Kazemzadeh et al., 2014; Mao et al., 2016). Yet, recent work in this area has largely followed the traditional REG paradigm by Dale and Reiter 1995, where (visual) context is mainly considered in terms of so-called distractor objects, that are similar to the target and must therefore be excluded by naming differences (Krahmer and van Deemter, 2012). These distractors do not facilitate the description task, but even exert “contextual pressure”, as the speaker needs to reason about which attributes and words make the expression unambiguous (Cohn-Gordon et al., 2018; Schüz and Zarrieß, 2021). The main goal of this paper is to widen this commonly accepted view on the role of visual context in visual REG and investigate how contextual information can be conceived as a resource that makes the generation of descriptions easier rather than harder. In visual REG from images, scene and object information is not available a priori: Whereas classical REG algorithms mostly rely on symbolic scene representations, neural generation models in visual REG have to extract object properties from low-level visual representations of the target and its context (Schüz et al., 2023). This even applies to properties as fundamental as the type of an object, i.e. how it is named in the expression. Under ideal conditions, determining a referent’s type and properties can be regarded as a relatively simple task, but it becomes non-trivial in the presence of imperfect visual information, occlusion or noise. At the same time, global visual scene context can be expected to be of great support in this task, in light of previous findings on human scene understanding (cf. Section 2). However, to date, little is known as to how processes of scene understanding and object type identification interact in REG. In this work, we hypothesize that visual scene context makes REG models more resilient, i.e., it allows them to recalibrate predictions that were based on imperfect target representations. To test this, we adopt a novel experimental setup for REG: we train and test different model architectures with target representations that have been artificially obscured with varying degrees of noise (cf. Figure 1). We provide the models with different context representations and compare their performance concerning common quality metrics and a focused human evaluation of their ability to determine referent types. We test how certain properties of the visual context affect the processing and performance of REG models, and verify our results with experiments using further datasets that are substantially different from the ones commonly used in existing REG research. Our results show that context makes models surprisingly resilient to perturbations in target representations, to the extent that they can identify referent types even when information about the objects themselves is completely missing. We believe that these results offer new perspectives on the role of scene context in visual REG. ## 2 Background #### Human scene understanding Research on human vision and perception emphasizes the fact that scenes are not mere collections of objects (Võ, 2021). When humans view a scene, they do not simply recognize the objects in it, but they understand it is a coherent whole. Oliva and Torralba (2006) observe that humans perceive the so-called gist of a scene rapidly and even when local information is missing (e.g. blurred). Experiments in this field indicate that contextual information can facilitate the recognition of visible objects across different tasks (Oliva and Torralba, 2007; Divvala et al., 2009; Galleguillos and Belongie, 2010; Parikh et al., 2012), and that, on the other hand, incongruent context can also be misleading (Zhang et al., 2020; Gupta et al., 2022). This means that the human vision exploits learned knowledge about these regularities of the visual word for visual processing (Biederman, 1972; Bar, 2004; Greene, 2013; Pereira and Castelhano, 2014; Sadeghi et al., 2015). To model these regularities, Võ (2021) proposed the notion of a “scene grammar” that can account for the interaction of global and local visual perception and understanding in humans. #### Scenes, objects, and image captioning Much research on V&L is currently concerned with modeling the generation and understanding of image descriptions, e.g. in tasks like image captioning or retrieval. Yet, many captioning tasks focus on rather object-centric descriptions that mention objects and their spatial relationships (Cafagna et al., 2021). A common representation of scene context in image captioning is scene graphs, cf. (Yang et al., 2023), which are usually modeled via spatial relations between bounding boxes of objects. Cafagna et al. 2023 propose a new task and dataset that foregrounds scene-level instead of object-centric descriptions. Another perspective on scene knowledge in captioning models is coming from work that focuses on probing them with perturbed or systematically varied images: Yin and Ordonez (2017) found that captioning with extremely reduced inputs of labeled object layouts performs surprisingly well. Related to this, Nikolaus et al. (2019) find that image captioning models often rely on regularities in object occurrences, to the extent that they fail to generalize to new combinations of objects. Their solution is to generate unseen combinations and challenge models on these. Our goal in this work is complementary: we aim to understand how exactly generation models may be able to leverage regular scene knowledge and patterns of object co-occurrence, and how this may facilitate the handling of imperfect visual information. #### REG and scene context REG is concerned with the generation of descriptions that distinguish a particular object in a given visual context, cf. Krahmer and van Deemter 2012. Recent visual REG models usually build on image captioning models but are adapted to generate more pragmatically informative expressions, using e.g. training objectives (Mao et al., 2016), comprehension modules (Luo and Shakhnarovich, 2017), reinforcement agents (Yu et al., 2017) or decoding strategies (Schüz and Zarrieß, 2021). REG models usually process different forms of context information. Whereas some models encode differences in appearance between targets and surrounding objects (Yu et al., 2016, 2017; Tanaka et al., 2019; Kim et al., 2020; Liu et al., 2020), others use representations of the global image (Mao et al., 2016; Luo and Shakhnarovich, 2017; Zarrieß and Schlangen, 2018; Panagiaris et al., 2020, 2021). Visual context is often supplemented with the relative position and size of the target in the image (Mao et al., 2016; Yu et al., 2017; Luo and Shakhnarovich, 2017; Li and Jiang, 2018; Tanaka et al., 2019; Kim et al., 2020; Panagiaris et al., 2020; Liu et al., 2020). #### Research gap Little is known about how visual REG models internally exploit their context representations and in what way context exactly enhances the generation of expressions. Here, the implicit assumption often is that models exploit context in a similar way as symbolic REG models, e.g. the Incremental Algorithms by Reiter and Dale (2000). However, a key difference to symbolic REG is that in visual REG failures in the scene and object understanding can arise from model hallucination or imperfect visual input, cf. (Schüz et al., 2023). This is especially evident for the type of objects: this attribute had a privileged role in early works (Dale and Reiter, 1995) as they are essential as the heads of referential noun phrases. In visual REG, referents must first be correctly identified to name them appropriately (Zarrieß and Schlangen, 2017; Silberer et al., 2020a, b), which is challenging in cases of deficient input, e.g. small or partially occluded objects (Yao and Fei-Fei, 2010). In this paper, we aim to close this gap and investigate how visual context information helps REG models to be more resilient to deficits in their target inputs. ## 3 Experimental Set-Up ### 3.1 Outline and Research Hypotheses The main idea of this work is to train and test standard REG models on visual target representations occluded with varying amounts of noise, to investigate how different combinations of target and context can compensate for this perturbation. For this, we draw on existing model architectures, and evaluate the trained models using both out-of-the-box quality metrics and more fine- grained human evaluation capturing the validity of assigned referent type labels. The evaluation results are also supported by supplementary analyses as well as further experiments with an additional data set. Generally, we expect that automatic metrics and human evaluation scores will drop for increasing amounts of target noise. However, we also hypothesize that visual context makes models more resilient, i.e., for the same amount of noise, models supplied with context outperform variants with only target information. While we expect this general effect across all conditions, it should be more pronounced as the amount of occlusion increases. ### 3.2 Models We set up two transformer-based REG models: TRF and CC. TRF is a transformer trained from scratch on REG data, CC builds upon a pre-trained language model. We define variants of both models using a) different combinations of target and context representations, as the respective model inputs, and b) the amount of target noise during training and inference. Implementation and training details for our models can be found in appendix B. Target representations include the visual contents of the target bounding box ($V_{t}$), its location, and size relative to the global image ($Loc_{t}$). As context representations, we use the embedding of the global image with the target masked out ($V_{c}$). For TRF, which is our better performing model, we also experiment with scene-level information (or scene summaries) about what kinds of objects are present in the surrounding scene ($S_{c}$), which are derived from panoptic segmentations (Kirillov et al. 2018, see Section 3.3). Models processing only target information are indicated with the subscript $t$, whereas models processing $V_{c}$ and $S_{c}$ context information are indexed with $v$ and $s$, respectively. To test our systems for perturbed target representations, we randomly replaced a fixed proportion of the pixels in the bounding box contents with random noise during both training and inference. All systems are trained and tested with three noise settings: $0.0$ as our baseline setting, where no pixels are perturbed; $0.5$, where $50\%$ of the pixels are replaced with noise; and $1.0$, where the entire content of the target bounding box is occluded, i.e. no visual information about the target is available. Noise levels for training and evaluation are shown in the index of the model identifiers. #### Vanilla Transformer (TRF) We use the model from Schüz and Zarrieß (2023), which is based on an existing implementation for image captioning.111https://github.com/saahiluppal/catr The model builds on ResNet (He et al., 2015) encodings for targets and context, which are passed on to an encoder/decoder transformer in the style of Vaswani et al. (2017). The model is largely comparable to the system in Panagiaris et al. (2021), but without self-critical sequence training and layer-wise connections between encoder and decoder. Unlike e.g. Mao et al. (2016), we train the model using Cross Entropy Loss. We compare three variants of this model, which take as input concatenated feature vectors comprised of the representations described above. TRFt receives only target information, i.e. the input vector is composed as $[V_{t};Loc_{t}]$. In addition to this, TRFv receives representations of the global image, with the input vector structure $[V_{t};Loc_{t};V_{c}]$. Finally, TRFs takes scene-level representations about the relative area occupied by different object classes in the visual context, i.e. $[V_{t};Loc_{t};S_{c}]$. For both $V_{t}$ and $V_{c}$, the respective parts of the image are scaled to $224\times 224$ resolution (keeping the original ratio and masking out the padding) and encoded with ResNet-152 (He et al., 2015), resulting in representations with 196 features ($14\times 14$) and hidden size $512$ for both target and context. $Loc_{t}$ is a vector of length 5 with the corner coordinates of the target bounding box and its area relative to the whole image, projected to the model’s hidden size. The scene summary input for TRFs consists of 134 features, which represent the relative area all of the 134 object or stuff types in COCO occupy in the visual context. To use this information in our model, we embed each of the object and stuff types in an additional embedding layer with $512$, which is jointly trained with the model. In the model’s forward pass, we weight each of the individual category embeddings with the relative area it occupies in the respective input image, and form $S_{c}$ by concatenating the embeddings. #### Fine-tuned GPT-2 (CC) We adapt the ClipCap model in Mokady et al. (2021) to the REG task. The authors use a simple MLP-based mapping network to construct fixed-size prefixes for GPT-2 (Radford et al., 2019) from CLIP encodings (Radford et al., 2021), and fine-tune both the mapping network and the language model for the image captioning task. To the best of our knowledge, this is the first model tested for REG which utilizes a pre-trained language model. As for the TRF model, we compare different variants of this base architecture. First, in CCt, the GPT-2 prefix is constructed as $[V_{t};Loc_{t}]$, where $V_{t}$ is computed like the CLIP prefix in the original paper (but for the contents of the target bounding box) and $Loc_{t}$ is the location features described above, projected into a single prefix token. For CCv, we again add the global image (minus the target) as context: $V_{c}$ is computed in the same way as $V_{t}$, but with a separate mapping network and with the global image as the visual input. Here, the final prefix is $[V_{t};V_{c};Loc_{t}]$. ### 3.3 Data We use RefCOCO (Kazemzadeh et al., 2014) for training and evaluation, which contains bounding boxes and expressions for MSCOCO images (Lin et al., 2014). The dataset contains separate testA and testB splits (1.9k and 1.8k items), where testA only contains humans as referential targets and testB all other object classes (but not humans). We also conduct experiments on the detection dataset PACO-EGO4D (Ramanathan et al., 2023), which contains annotations for objects and object parts in first- person video frames (Grauman et al., 2022). In comparison to RefCOCO, PACO is larger (75k items in test split), data is less standardized and objects are often harder to recognize. To construct the scene summaries for our TRFs model and analyze attention allocation patterns, we rely on annotations for panoptic segmentation (Kirillov et al., 2018), i.e. dense pixel-level segmentation masks for both thing and stuff classes in MSCOCO images (Caesar et al., 2016). ### 3.4 Evaluation #### Generation Quality / N-Gram Metrics To estimate the general generation capabilities of our models we rely on BLEU (Papineni et al., 2002) and CIDEr (Vedantam et al., 2014) as established metrics for automatic evaluation. #### Referent Type Assignment / Human Evaluation To test whether our models succeed in assigning valid types to referents, we collect human judgements for generated expressions for a subset of 200 items from the RefCOCO testB split. This split contains only non-human referents, and was chosen due to preliminary tests indicating that judgments about the validity of type labels for human referents are often difficult, e.g. due to ambiguity regarding the gender of depicted individuals. The annotators were instructed to rate only those parts of the expressions that refer to the type of the referential target. For example, “the black dog” should be rated as correct if the target is of the type dog, but is actually white. All items should be assigned exactly one of the following categories: * • Adequate / A: The generated expression contains a valid type description for the referent. * • Misaligned / M: Type designators do not apply to the intended target, but to other objects (partially) captured by the bounding box. * • Omission / O: Omission of the target type, e.g. description via non-type attributes, pronominalization or general nouns such as “thing”. * • False / F: Type designations that do not apply to the intended target or other objects captured by the bounding box. Previous research has shown considerable variation in object naming (Silberer et al. 2020a, b, among others). Therefore, for the A category, type descriptions do not have to match the ground truth annotations, but different labels can be considered adequate if they represent valid descriptions of the target type. For example, dog, pet and animal would be considered equally correct for depicted dogs. Subsequent to the human evaluation, we investigate correlations between the evaluation results and further properties of the visual context. #### Referent Type Assignment / Classification Accuracy We complement the human evaluation with RefCOCO with further experiments using PACO-EGO4D. While PACO does not provide referring expressions, we treat the object and object-part annotations similarly, i.e. our models generate the category strings (instead of predicting the respective category in a multiclass classification scheme). We evaluate the identification of object and part types by measuring the accuracy of the models in exactly reproducing the respective category strings (for entire objects) or the strings in the object-part tuples (for object parts). #### Attention Allocation We also examine how our TRFv model allocates attention over different parts of the input as a result of different noise levels during training. First, we follow Schüz and Zarrieß (2023) in measuring the ratio between target and context partitions, i.e. the summed attention weights directed to the target and its context in both the encoder and decoder multi-head attention. For this, we compute $\alpha_{t}$, $\alpha_{l}$ and $\alpha_{c}$ as the cumulative attention weights directed to $V_{t}$, $Loc_{t}$ and $V_{c}$, respectively, by calculating the sum of the attention weights assigned to each input partition, normalized such that $\alpha_{t}+\alpha_{l}+\alpha_{c}=1$. We also quantify the attention difference between $\alpha_{t}$ and $\alpha_{c}$ as $\Delta_{t,c}$, by excluding $\alpha_{l}$ and normalizing the target and context scores such that $\alpha_{t}+\alpha_{c}=1$. Then, we calculate $\Delta_{t,c}=\alpha_{t}-\alpha_{c}$, i.e. $0<\Delta_{t,c}\leq 1$ if there is relative focus on the target, $-1\leq\Delta_{t,c}<0$ if there is relative focus on the context, and $\Delta_{t,c}=0$ when both parts are weighted equally. Second, we measure the model attention allocated to different classes of objects in the visual context, using the panoptic segmentation data described in Section 3.3. Here, we first interpolate the model attention map to fit the original dimensions of the image, and retrieve the segmentation masks for the respective image. For each category $x\in X$, we then compute the cumulative attention weight $\alpha_{x}$ by computing the sum of pixels attributed to this category, weighted by the model attention scores over the image and normalized such that $\sum_{x\in X}\alpha_{x}=1$. We report $\alpha_{x=tgt}$, i.e. attention allocated to areas of the visual context assigned the same category as the referential target. As the covered area varies between object categories, we get different scores even if the model attention is perfectly balanced over the image. To address this, we also report scores that are normalized by the area covered by the category. Scores $>1$ indicate that the category is attended more than to be expected based on the coverage area. ## 4 Results on RefCOCO ### 4.1 Automatic Quality Metrics | testA | testB ---|---|--- | Bl1 | Bl2 | CDr | Bl1 | Bl2 | CDr system | | | | | | TRFt-0.0 | 0.55 | 0.35 | 0.86 | 0.57 | 0.35 | 1.28 TRFt-0.5 | 0.49 | 0.32 | 0.73 | 0.51 | 0.32 | 1.04 TRFt-1.0 | 0.35 | 0.17 | 0.34 | 0.30 | 0.14 | 0.20 TRFv-0.0 | 0.58 | 0.39 | 0.93 | 0.61 | 0.39 | 1.36 TRFv-0.5 | 0.54 | 0.35 | 0.81 | 0.56 | 0.36 | 1.24 TRFv-1.0 | 0.46 | 0.29 | 0.60 | 0.55 | 0.36 | 1.14 TRFs-0.0 | 0.54 | 0.34 | 0.84 | 0.57 | 0.35 | 1.27 TRFs-0.5 | 0.52 | 0.35 | 0.81 | 0.56 | 0.35 | 1.28 TRFs-1.0 | 0.42 | 0.24 | 0.51 | 0.53 | 0.33 | 1.12 CCt-0.0 | 0.48 | 0.30 | 0.70 | 0.47 | 0.28 | 0.88 CCt-0.5 | 0.38 | 0.22 | 0.48 | 0.36 | 0.20 | 0.52 CCt-1.0 | 0.35 | 0.16 | 0.37 | 0.29 | 0.12 | 0.16 CCv-0.0 | 0.57 | 0.38 | 0.92 | 0.58 | 0.37 | 1.25 CCv-0.5 | 0.51 | 0.32 | 0.77 | 0.49 | 0.31 | 0.97 CCv-1.0 | 0.40 | 0.23 | 0.46 | 0.38 | 0.21 | 0.46 Table 1: BLEU1, BLEU2 and CIDEr scores on RefCOCO testA and testB for all TRF and CC variants. Systems indicated with t can only access target information, v and s models are supplied with visual context and scene summaries, respectively. Target noise proportions ($0.0$, $0.5$, $1.0$) are denoted in the indices. Generally, context information leads to improved results, especially for high noise settings. Table 1 shows the results of the automatic evaluation of our systems on RefCOCO testA and testB. Interestingly, throughout all conditions, the simpler TRF model outperforms CC, although the latter builds on pre-trained CLIP and GPT-2 which are known to be effective for image captioning (Mokady et al., 2021). It is possible that CC cannot fully benefit from CLIP pre-training due to the structural differences between bounding box contents and full images, or that performance drops result from higher compression when constructing the GPT prefixes. Also, TRF achieves a considerably larger performance gain than CC when adding scene context, indicating that this model is more effective at exploiting contextual information. For both TRF and CC, scores consistently drop with increasing target noise. However, this is mitigated if context is available: For both model types, variants incorporating visual context are substantially more robust against target noise, even in cases where target representations are entirely occluded by noise ($1.0$ in the subscripts). A striking example is RefCOCO testB, where CIDEr drops to $0.20$ for TRFt-1.0 and $0.16$ for CCt-1.0, but TRFv-1.0 achieves scores as high as $1.14$. Here, CCv-1.0 drastically underperforms with CIDEr $0.46$, but still outperforms its no-context counterpart. Interestingly, we see considerable differences between testA and testB. Both TRFt-1.0 and CCt-1.0 achieve better results on testA, but the scores are generally higher on testB. Importantly, testA is restricted to human referents, while testB encompasses all non-human object classes. Therefore, models without any access to meaningful visual input could often guess right on the frequent human classes, but struggle with the higher variation in testB. This is supported by the inverse pattern that visual context particularly improves the testB results. Here, differences between $t$ and $v$ variants are much higher, suggesting that context is more informative for non- human objects, i.e. there are stronger associations between certain types of objects and the contexts in which they occur. Another striking result is that the same patterns emerge if we exchange visual context representations with more abstract scene summaries: TRFs-1.0 achieves CIDEr $1.12$ for entirely occluded targets in testB, comparable to TRFv-1.0. Interestingly, between TRFs-0.0 and TRFs-0.5 the scene model slightly improves in CIDEr scores, i.e. it can fully compensate for partial target occlusion. ### 4.2 Target Identification Human judgements were collected from 5 expert annotators, including the first author. Every system was evaluated independently by three annotators, with a Fleiss’ Kappa of 0.85, indicating almost perfect agreement (Landis and Koch, 1977). The final judgments are determined by majority vote. The human evaluation results for the 200-item subset of RefCOCO testB are shown in Table 2 and illustrated in Figure 2. Generally, the results mirror the pattern in the BLEU and CIDEr scores discussed previously: Across all conditions, A scores drop if noise ratios increase, while F scores increase at the same time. For M and O the results are less clear, but higher noise settings generally lead to higher rates than the baseline setting for both categories. This holds for TRF and CC, but TRF again performs better throughout all conditions. Again, visual context clearly allows the models to compensate for deficient target representations: While CCv-1.0 assigns adequate types in almost $20\%$ of all cases (as compared to $0.5\%$ without context information), TRFv-1.0 scores an impressive $66\%$, only $15.5\%$ less than without any target noise. Interestingly, scene summaries allow the model to compensate for deficient target representations even more effectively. Across all noise settings, TRFs achieves the highest A scores and the lowest F and O rates, even without any target noise, unlike for BLEU and CIDEr (cf. Section 4.1). | % A | % F | % O | % M ---|---|---|---|--- TRFt-0.0 | 84.0 | 10.5 | 5.0 | 0.5 TRFt-0.5 | 66.0 | 27.5 | 4.0 | 2.5 TRFt-1.0 | 1.5 | 75.5 | 19.5 | 3.5 TRFv-0.0 | 81.5 | 12.0 | 5.5 | 1.0 TRFv-0.5 | 70.5 | 18.5 | 7.0 | 4.0 TRFv-1.0 | 66.0 | 26.5 | 4.0 | 3.5 TRFs-0.0 | 89.0 | 7.0 | 3.5 | 0.5 TRFs-0.5 | 81.0 | 14.5 | 2.5 | 2.0 TRFs-1.0 | 68.0 | 22.0 | 1.5 | 8.5 CCn-0.0 | 45.5 | 46.5 | 7.0 | 1.0 CCn-0.5 | 23.0 | 61.0 | 13.0 | 3.0 CCn-1.0 | 0.5 | 84.5 | 11.0 | 4.0 CCv-0.0 | 76.0 | 21.0 | 3.0 | 0.0 CCv-0.5 | 55.0 | 36.0 | 6.5 | 2.5 CCv-1.0 | 19.5 | 68.5 | 9.0 | 3.0 $human$ | 91.0 | 2.0 | 6.5 | 0.5 Table 2: Results of the human evaluation on 200 items from RefCOCO-testB. Generally, contextual information leads to more adequate type descriptions, even if target representations are entirely occluded. ### 4.3 How do models exploit scene context? So far, our results indicate that the scene context of referential targets greatly improves the resilience of REG, to the extent that correct predictions are possible to a surprising rate even if target information is missing. Here, we aim to analyze how exactly contextual information is exploited by the models. As discussed in Section 2, previous research indicates that regularities of object co-occurrence and scene properties facilitate e.g. object recognition in context. However, qualitative inspection of our data indicates that for high noise, our systems often copy from context, i.e. correctly predict referent types that are also present in the surrounding scene, given that many classes of objects tend to appear in groups. We investigate this in more detail and (a) perform statistical tests to check whether similar objects in context support identification performance and (b) analyze the attention distribution for TRFv to see whether the model learns to attend to the respective objects in context. #### Statistical analysis: Target categories in context We hypothesize that recalibration through context should be more effective when the target class is also present in the scene. To test this, we conduct a correlation analysis between identification accuracy and the relative coverage of the target class in the context. For this, we again rely on panoptic segmentation annotations (cf. Section 3.3) to compute the proportion of pixels of the same class as the referential target, normalized by the total size of the context. We binarize the human evaluation scores (True if rated as A, else False) and compute the Point-biserial correlation coefficient between the relative coverage of the target class in context and the identification accuracy. For both TRFv-1.0 (corr: 0.321, p < 0.001) and TRFs-1.0 (corr: 0.277, p < 0.001) we found that a higher prevalence of the target class in the visual context leads to significantly higher scores in human evaluation, i.e. systems can easier compensate a lack of visual target information if the context contains similar objects. For CCv-1.0, we see the same trend, but without statistical significance (corr: 0.136, p = 0.055). For cases, where neither system can identify the target class, we see a strong inverse correlation (corr: -0.267, p < 0.001). #### Model attention to target category in context | Encoder | Decoder ---|---|--- | $\alpha_{x=tgt}$ | norm. | $\alpha_{t}$ | $\alpha_{l}$ | $\alpha_{c}$ | $\Delta_{t,c}$ | $\alpha_{x=tgt}$ | norm. | $\alpha_{t}$ | $\alpha_{l}$ | $\alpha_{c}$ | $\Delta_{t,c}$ TRFv-0.0 | 36.70 | 1.77 | 44.49 | 9.20 | 46.31 | -0.02 | 26.94 | 1.20 | 52.65 | 9.69 | 37.67 | 0.16 TRFv-0.5 | 35.27 | 1.64 | 18.90 | 16.06 | 65.04 | -0.55 | 40.56 | 2.05 | 32.65 | 14.41 | 52.94 | -0.24 TRFv-1.0 | 35.63 | 1.70 | 41.05 | 0.67 | 58.28 | -0.17 | 43.66 | 2.26 | 43.75 | 0.48 | 55.78 | -0.12 Table 3: Attention allocation scores for TRFv, averaged over all RefCOCO testB items. $\alpha$ scores are reported in %. To see whether the TRFv model has indeed learned the hypothesized copying strategy, we compute the distribution of attention mass directed to target, location and context partitions as well as to objects sharing the target category in the visual context, as described in Section 3.4. In Table 3, we report the analysis results, averaged for all items in the RefCOCO testB split. The $\Delta_{t,c}$ scores show that the context partition receives more attention if the target is occluded with noise during the training, in line with our previous results. However, surprisingly, more attention is allocated to the context in the $0.5$ noise setting than if no target information is accessible. The $\alpha$ scores also indicate that location features are especially focused in this case, suggesting that this source of information is especially helpful if visual target information is reduced, but not entirely missing. As shown by the $\alpha_{x=tgt}$ scores in Table 3, target noise during model training does not seem to have a consistent effect on encoder attention to context objects sharing the target category. For the decoder, however, we see a notable increase: Whereas the baseline model assigns an average of 26.94 % of its attention mass on context objects with the target class, this is significantly increased for higher noise settings (40.56 % and 43.66 %). The normalized results exhibit the same patterns, i.e. as a result of target noise, context objects sharing the target class receive more than double of the attention mass as to be expected based on their size in the image. These results suggest that the TRF model learns to exploit the occurrence of similar objects in target and context as a common property of scenes in RefCOCO. However, due to the prevalence of frequent object classes and the reliance on published photos, it is unclear how representative these results are. In the next section, we examine whether these patterns can be replicated for the PACO dataset. ## 5 Results on PACO In our experiments on the EGO4D portion of the PACO dataset (Section 3.3), we treat the category strings in the detection dataset as expressions and train TRFt and TRFc to generate those strings given the contents of the target bounding box and (for the latter variant) the visual context (see Section 3.4). We report accuracy scores for the test split in Table 4. Here, notably, the TRFt variant achieves higher accuracy scores than TRFv, unless the entire visual target representation is covered with random noise. This suggests that visual context is less informative or more difficult to process in PACO than RefCOCO. However, the (comparably small) gain of TRFv-1.0 over TRFt-1.0 indicates that the model can leverage the visual context to a certain degree. While some of the differences to RefCOCO may result from different experimental settings (e.g. class strings instead of expressions), the PACO results also hint towards general problems with datasets relying on scraped images such as RefCOCO, in that they may not be not sufficiently representative of the visual complexity in everyday scenes. | $acc_{obj}$ | $acc_{obj-part}$ ---|---|--- TRFt-0.0 | 60.93 | 34.03 TRFt-0.5 | 45.30 | 25.57 TRFt-1.0 | 16.47 | 7.55 TRFv-0.0 | 56.97 | 33.22 TRFv-0.5 | 34.92 | 20.31 TRFv-1.0 | 22.34 | 11.11 Table 4: Results for TRFt and TRFv on PACO-EGO4D. $acc_{obj}$ describes the accuracy for reproducing object category strings, $acc_{obj-part}$ for reproducing (object, part) tuples for annotated object parts. ## 6 Discussion and Conclusion Our findings show that contextual information about the surroundings of referential targets makes REG models more resilient against pertubations in visual target representations. Even for conditions where no target information is present at all, REG models maintain good results in automatic quality metrics and identify referent types with high accuracy, as shown in the human evaluation results. This holds for different kinds of context: While especially the TRFv model is able to leverage ResNet encodings of image contents outside the target bounding box, the same applies to scene-level representations of depiced objects, as included in the TRFs model. Interestingly, our subsequent analysis suggest that our context models implicitly learned to copy from the visual context, i.e. assign labels to referents which also apply to visible context objects. While the weaker context effects in our PACO results suggest that this strategy is not universally applicable, it appears to be highly effective the more regular RefCOCO data. This is in stark contrast to basic assumptions of the REG paradigm, where context information is considered important mainly to ensure that references can be resolved without ambiguity. Here, we show, that is also a valuable source for further communicative goals, i.e. the truthfulness of generated expressions. Overall, our results indicate that the influence of visual context in REG is more multifaceted than reflected in previous studies. Importantly, however, this study only provides an initial spotlight, as research in related fields suggests that there are other and more complex ways in which visual scene context may facilitate reference production. With this in mind, we strongly advocate further research into scene context at the interface of perceptual psychology, computer vision and language generation. ## 7 Limitations We identify the following limitations in our study: First, in both training and evaluation, we do not consider pragmatic informativeness as a core criterion for the REG task. We train our models using Cross Entropy Loss and do not test whether the generated expressions unambiguously describe the referential target, instead focusing on semantic adequacy as an important prerequisite for the generation of successful referential expressions. However, we acknowledge that a comprehensive view would require the consideration of both semantic and pragmatic aspects. Also, we do not consider recent developments such as multimodal LLMs, although the high diversity of their training data would contribute an interesting aspect to this study. Here, we selected our models with a focus on both modifiability and transparent processing. Finally, additional vision and language datasets such as VisualGenome (Krishna et al., 2016) would have made the results more representative. However, due to time and space constraints, we leave this for future research. ## References * Bar (2004) Moshe Bar. 2004. Visual objects in context. _Nature Reviews Neuroscience_ , 5(8):617–629. * Biederman (1972) Irving Biederman. 1972. Perceiving real-world scenes. _Science_ , 177(4043):77–80. * Caesar et al. (2016) Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. 2016. Coco-stuff: Thing and stuff classes in context. * Cafagna et al. (2021) Michele Cafagna, Kees van Deemter, and Albert Gatt. 2021. What vision-language models ‘see’ when they see scenes. * Cafagna et al. (2023) Michele Cafagna, Kees van Deemter, and Albert Gatt. 2023. HL dataset: Visually-grounded description of scenes, actions and rationales. In _Proceedings of the 16th International Natural Language Generation Conference_ , pages 293–312, Prague, Czechia. Association for Computational Linguistics. * Cohn-Gordon et al. (2018) Reuben Cohn-Gordon, Noah Goodman, and Christopher Potts. 2018. Pragmatically informative image captioning with character-level inference. * Dale and Reiter (1995) Robert Dale and Ehud Reiter. 1995. Computational interpretations of the gricean maxims in the generation of referring expressions. _Cognitive Science_ , 19(2):233–263. * Divvala et al. (2009) Santosh K. Divvala, Derek Hoiem, James H. Hays, Alexei A. Efros, and Martial Hebert. 2009. An empirical study of context in object detection. In _2009 IEEE Conference on Computer Vision and Pattern Recognition_. IEEE. * Galleguillos and Belongie (2010) Carolina Galleguillos and Serge Belongie. 2010. Context based object categorization: A critical survey. _Computer Vision and Image Understanding_ , 114(6):712–722. * Grauman et al. (2022) Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. 2022. Ego4d: Around the world in 3,000 hours of egocentric video. In _2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)_. IEEE. * Greene (2013) Michelle R. Greene. 2013. Statistics of high-level scene context. _Frontiers in Psychology_ , 4. * Gupta et al. (2022) Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. 2022. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In _2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)_. IEEE. * He et al. (2015) Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep residual learning for image recognition. * Kazemzadeh et al. (2014) Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. 2014. ReferItGame: Referring to objects in photographs of natural scenes. In _Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)_ , pages 787–798, Doha, Qatar. Association for Computational Linguistics. * Kim et al. (2020) Jungjun Kim, Hanbin Ko, and Jialin Wu. 2020. CoNAN: A complementary neighboring-based attention network for referring expression generation. In _Proceedings of the 28th International Conference on Computational Linguistics_ , pages 1952–1962, Barcelona, Spain (Online). International Committee on Computational Linguistics. * Kirillov et al. (2018) Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollár. 2018. Panoptic segmentation. * Krahmer and van Deemter (2012) Emiel Krahmer and Kees van Deemter. 2012. Computational generation of referring expressions: A survey. _Computational Linguistics_ , 38(1):173–218. * Krishna et al. (2016) Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. 2016. Visual genome: Connecting language and vision using crowdsourced dense image annotations. * Landis and Koch (1977) J. Richard Landis and Gary G. Koch. 1977. The measurement of observer agreement for categorical data. _Biometrics_ , 33(1):159. * Li and Jiang (2018) Xiangyang Li and Shuqiang Jiang. 2018. Bundled object context for referring expressions. _IEEE Transactions on Multimedia_ , 20(10):2749–2760. * Lin et al. (2014) Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In _Computer Vision – ECCV 2014_ , pages 740–755, Cham. Springer International Publishing. * Liu et al. (2020) Jingyu Liu, Wei Wang, Liang Wang, and Ming-Hsuan Yang. 2020. Attribute-guided attention for referring expression generation and comprehension. _IEEE Transactions on Image Processing_ , 29:5244–5258. * Luo and Shakhnarovich (2017) R. Luo and Gregory Shakhnarovich. 2017. Comprehension-guided referring expressions. _2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)_ , pages 3125–3134. * Mao et al. (2016) Junhua Mao, J. Huang, A. Toshev, Oana-Maria Camburu, A. Yuille, and Kevin Murphy. 2016. Generation and comprehension of unambiguous object descriptions. _2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)_ , pages 11–20. * Mokady et al. (2021) Ron Mokady, Amir Hertz, and Amit H. Bermano. 2021. Clipcap: Clip prefix for image captioning. * Nikolaus et al. (2019) Mitja Nikolaus, Mostafa Abdou, Matthew Lamm, Rahul Aralikatte, and Desmond Elliott. 2019. Compositional generalization in image captioning. In _Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL)_ , pages 87–98, Hong Kong, China. Association for Computational Linguistics. * Oliva and Torralba (2006) Aude Oliva and Antonio Torralba. 2006. Chapter 2 building the gist of a scene: the role of global image features in recognition. In _Progress in Brain Research_ , pages 23–36. Elsevier. * Oliva and Torralba (2007) Aude Oliva and Antonio Torralba. 2007. The role of context in object recognition. _Trends in Cognitive Sciences_ , 11(12):520–527. * Panagiaris et al. (2020) Nikolaos Panagiaris, Emma Hart, and Dimitra Gkatzia. 2020. Improving the naturalness and diversity of referring expression generation models using minimum risk training. In _Proceedings of the 13th International Conference on Natural Language Generation_ , pages 41–51, Dublin, Ireland. Association for Computational Linguistics. * Panagiaris et al. (2021) Nikolaos Panagiaris, Emma Hart, and Dimitra Gkatzia. 2021. Generating unambiguous and diverse referring expressions. _Computer Speech & Language_, 68:101184. * Papineni et al. (2002) Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In _Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics_ , pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. * Parikh et al. (2012) Devi Parikh, C. Lawrence Zitnick, and Tsuhan Chen. 2012. Exploring tiny images: The roles of appearance and contextual information for machine and human object recognition. _IEEE Transactions on Pattern Analysis and Machine Intelligence_ , 34(10):1978–1991. * Pedregosa et al. (2011) Fabian Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. _Journal of Machine Learning Research_ , 12:2825–2830. * Pereira and Castelhano (2014) Effie J. Pereira and Monica S. Castelhano. 2014. Peripheral guidance in scenes: The interaction of scene context and object content. _Journal of Experimental Psychology: Human Perception and Performance_ , 40(5):2056–2072. * Radford et al. (2021) Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In _Proceedings of the 38th International Conference on Machine Learning_ , volume 139 of _Proceedings of Machine Learning Research_ , pages 8748–8763. PMLR. * Radford et al. (2019) Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. _OpenAI blog_ , 1(8):9. * Ramanathan et al. (2023) Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, Amir Mousavi, Yiwen Song, Abhimanyu Dubey, and Dhruv Mahajan. 2023. Paco: Parts and attributes of common objects. In _2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)_. IEEE. * Reiter and Dale (2000) Ehud Reiter and Robert Dale. 2000. _Building natural language generation systems_. Cambridge University Press, Cambridge, U.K. New York. * Sadeghi et al. (2015) Zahra Sadeghi, James L. McClelland, and Paul Hoffman. 2015. You shall know an object by the company it keeps: An investigation of semantic representations derived from object co-occurrence in visual scenes. _Neuropsychologia_ , 76:52–61. * Schüz and Zarrieß (2021) Simeon Schüz and Sina Zarrieß. 2021. Decoupling pragmatics: Discriminative decoding for referring expression generation. In _Proceedings of the Reasoning and Interaction Conference (ReInAct 2021)_ , pages 47–52, Gothenburg, Sweden. Association for Computational Linguistics. * Schüz and Zarrieß (2023) Simeon Schüz and Sina Zarrieß. 2023. Keeping an eye on context: Attention allocation over input partitions in referring expression generation. In _Proceedings of the Workshop on Multimodal, Multilingual Natural Language Generation and Multilingual WebNLG Challenge (MM-NLG 2023)_ , pages 20–27, Prague, Czech Republic. Association for Computational Linguistics. * Schüz et al. (2023) Simeon Schüz, Albert Gatt, and Sina Zarrieß. 2023. Rethinking symbolic and visual context in referring expression generation. _Frontiers in Artificial Intelligence_ , 6. * Silberer et al. (2020a) Carina Silberer, Sina Zarrieß, and Gemma Boleda. 2020a. Object naming in language and vision: A survey and a new dataset. In _Proceedings of the Twelfth Language Resources and Evaluation Conference_ , pages 5792–5801, Marseille, France. European Language Resources Association. * Silberer et al. (2020b) Carina Silberer, Sina Zarrieß, Matthijs Westera, and Gemma Boleda. 2020b. Humans meet models on object naming: A new dataset and analysis. In _Proceedings of the 28th International Conference on Computational Linguistics_ , pages 1893–1905, Barcelona, Spain (Online). International Committee on Computational Linguistics. * Tanaka et al. (2019) M. Tanaka, Takayuki Itamochi, K. Narioka, Ikuro Sato, Y. Ushiku, and T. Harada. 2019\. Generating easy-to-understand referring expressions for target identifications. _2019 IEEE/CVF International Conference on Computer Vision (ICCV)_ , pages 5793–5802. * Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. * Vedantam et al. (2014) Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. 2014. Cider: Consensus-based image description evaluation. * Võ (2021) Melissa Le-Hoa Võ. 2021. The meaning and structure of scenes. _Vision Research_ , 181:10–20. * Yang et al. (2023) Xu Yang, Jiawei Peng, Zihua Wang, Haiyang Xu, Qinghao Ye, Chenliang Li, Songfang Huang, Fei Huang, Zhangzikang Li, and Yu Zhang. 2023. Transforming visual scene graphs to image captions. In _Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , pages 12427–12440, Toronto, Canada. Association for Computational Linguistics. * Yao and Fei-Fei (2010) Bangpeng Yao and Li Fei-Fei. 2010. Modeling mutual context of object and human pose in human-object interaction activities. In _2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition_. IEEE. * Yin and Ordonez (2017) Xuwang Yin and Vicente Ordonez. 2017. Obj2Text: Generating visually descriptive language from object layouts. In _Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing_ , pages 177–187, Copenhagen, Denmark. Association for Computational Linguistics. * Yu et al. (2016) Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. 2016\. Modeling context in referring expressions. In _Computer Vision – ECCV 2016_ , pages 69–85, Cham. Springer International Publishing. * Yu et al. (2017) Licheng Yu, Hao Tan, Mohit Bansal, and Tamara L Berg. 2017. A joint speaker-listener-reinforcer model for referring expressions. In _Computer Vision and Pattern Recognition (CVPR)_ , volume 2. * Zarrieß and Schlangen (2017) Sina Zarrieß and David Schlangen. 2017. Obtaining referential word meanings from visual and distributional information: Experiments on object naming. In _Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , pages 243–254, Vancouver, Canada. Association for Computational Linguistics. * Zarrieß and Schlangen (2018) Sina Zarrieß and David Schlangen. 2018. Decoding strategies for neural referring expression generation. In _Proceedings of the 11th International Conference on Natural Language Generation_ , pages 503–512, Tilburg University, The Netherlands. Association for Computational Linguistics. * Zhang et al. (2020) Mengmi Zhang, Claire Tseng, and Gabriel Kreiman. 2020. Putting visual object recognition in context. In _2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)_ , pages 12982–12991. ## Appendix A Risks and Ethical Considerations We do not believe that there are significant risks associated with this work, as we consider the generation of general expressions for generic objects in freely available datasets with limited scale. When selecting samples for human evaluation, we refrain from descriptions of people (that could potentially be perceived as hurtful). No ethics review was required. Our data (published upon acceptance) does not contain any protected information and will be fully anonymized prior to publication. ## Appendix B Model implementation and training For the hyperparameters of our models, we largely followed Panagiaris et al. (2021) (TRF) and Mokady et al. (2021) (CC). During inference, we relied on (deterministic) greedy decoding. The TRF model has 3 encoder and 3 decoder layers with 8 attention heads, hidden dimension and feedforward dimension of 512, and was trained with a initial learning rate of 0.0001 for the transformer encoder and decoder, and 0.00001 for the pre-trained ResNet-152 backbone. Our TRF models have approximately 103,000,000 parameters. For our CC model, we kept the settings defined by Mokady et al. (2021). From the two models proposed in this work, we used the variant where a simple MLP is used as a mapping network and the GPT-2 language model is fine-tuned during training. However, we have different prefix sizes than in the original paper: For CCt, we have a prefix size of 11, i.e. 10 for the visual target representation and 1 for the target location information. For CCv, our prefix size is 21, with additional 10 tokens for the visual context representation. The model was trained using a learning rate of 0.00001. CCv has approximately 338,000,000 and CCt has 307,000,000 parameters. We trained our models on an Nvidia RTX A40. RefCOCO contains 42k and PACO- EGO4D contains 116k items for training. The number of training epochs per system and the final CIDEr scores over the validation sets are displayed in Table 5. We trained all our models for a maximum of 15 epochs, with early stopping if no new maximum for CIDEr over the validation set has been achieved for three consecutive epochs. Per epoch, the compute time was approximately 2.30 h for TRF and CC on RefCOCO and 4.30 h for TRF on PACO. | dataset | epochs | CIDEr (val) ---|---|---|--- TRF${}_{t}-0.0$ | RefCOCO | 8 | 1.074 TRF${}_{t}-0.5$ | RefCOCO | 11 | 0.936 TRF${}_{t}-1.0$ | RefCOCO | 5 | 0.302 TRF${}_{v}-0.0$ | RefCOCO | 6 | 1.156 TRF${}_{v}-0.5$ | RefCOCO | 9 | 1.035 TRF${}_{v}-1.0$ | RefCOCO | 6 | 0.869 TRF${}_{s}-0.0$ | RefCOCO | 8 | 1.075 TRF${}_{s}-0.5$ | RefCOCO | 14 | 1.032 TRF${}_{s}-1.0$ | RefCOCO | 12 | 0.818 CG${}_{t}-0.0$ | RefCOCO | 7 | 0.824 CG${}_{t}-0.5$ | RefCOCO | 8 | 0.554 CG${}_{t}-1.0$ | RefCOCO | 2 | 0.294 CG${}_{v}-0.0$ | RefCOCO | 4 | 1.103 CG${}_{v}-0.5$ | RefCOCO | 10 | 0.894 CG${}_{v}-1.0$ | RefCOCO | 7 | 0.526 TRF${}_{t}-0.0$ | PACO | 3 | 3.236 TRF${}_{t}-0.5$ | PACO | 8 | 2.662 TRF${}_{t}-1.0$ | PACO | 7 | 0.814 TRF${}_{v}-0.0$ | PACO | 3 | 3.554 TRF${}_{v}-0.5$ | PACO | 14 | 3.047 TRF${}_{v}-1.0$ | PACO | 5 | 2.15 Table 5: Training information for all TRF and CC variants. ## Appendix C Scientific Artifacts In our work, we mainly used scientific artifacts in the form of existing model implementations, all of which are cited or referenced in Section 3. The model implementations were published under permissive licences, i.e. MIT (TRF) and Apache 2.0 (CC). Upon acceptance, we will publish our modifications to the model implementations using the same licences, and our other code and data using permissive licences. Apart from this, we relied on scikit-learn (version 1.2.0, Pedregosa et al. 2011) for our statistic analysis and the RefCOCO API (Kazemzadeh et al., 2014; Yu et al., 2016)222https://github.com/lichengunc/refer for computing BLEU and CIDEr scores. ## Appendix D Human Evaluation We conducted a human evaluation in which the adequacy of assigned referent types in English referring expressions was assessed. The annotation guidelines are published as supplementary material. Our annotators were undergrad student assistants from linguistics and computational linguistics, which were paid by the hour according to the applicable pay scale. The annotators were informed about the intended use of their produced data. Along with our code, upon acceptance, we will publish the fully anonymized raw and aggregated results of the human evaluation. Figure 2 illustrates the results of the human evaluation described in Section 4.2. Figure 2: Visualization of human evaluation results for all tested systems and a sample of human annotations in RefCOCO testB.
# _Ab initio_ design of charge-mismatched ferroelectric superlattices Claudio Cazorla Institut de Ci$\grave{e}$ncia de Materials de Barcelona (ICMAB-CSIC), 08193 Bellaterra, Spain Massimiliano Stengel Institut de Ci$\grave{e}$ncia de Materials de Barcelona (ICMAB-CSIC), 08193 Bellaterra, Spain ICREA - Institució Catalana de Recerca i Estudis Avançats, 08010 Barcelona, Spain<EMAIL_ADDRESS> ###### Abstract We present a systematic approach to modeling the electrical and structural properties of charge-mismatched superlattices from first principles. Our strategy is based on bulk calculations of the parent compounds, which we perform as a function of in-plane strain and out-of-plane electric displacement field. The resulting two-dimensional phase diagrams allow us to accurately predict, without performing further calculations, the behavior of a layered heterostructure where the aforementioned building blocks are electrostatically and elastically coupled, with an arbitrary choice of the interface charge (originated from the polar discontinuity) and volume ratio. By using the [PbTiO3]m/[BiFeO3]n system as test case, we demonstrate that interface polarity has a dramatic impact on the ferroelectric behavior of the superlattice, leading to the stabilization of otherwise inaccessible bulk phases. ###### pacs: 71.15.-m, 77.22.Ej, 77.55.+f, 77.84.Dy ## I Introduction When layers of perovskite oxides are epitaxially stacked to form a periodically repeated heterostructure, new intriguing functionalities can emerge in the resulting superlattice [ghosez08, ; junquera11, ]. These are further tunable via applied electric fields and thermodynamic conditions, and thus attractive for nanoelectronics and energy applications. An excellent example is the [PbTiO3]m/[SrTiO3]n system, where the polarization, tetragonality, piezoelectric response, and ferroelectric transition temperature strongly change with the volume ratio of the parent compounds [dawber05, ; dawber07, ; dawber12, ]. Such a remarkable tunability is usually rationalized in terms of epitaxial strains [dawber05b, ], electrostatic coupling (see Fig. 1a) [zubko12, ; wu12, ], and local interface effects [junquera12, ; bousquet08, ]. While perovskite titanates with ATiO3 formula (A=Sr, Pb, Ba or Ca) have traditionally been the most popular choice as the basic building blocks, a much wider range of materials (e.g., BiFeO3) is currently receiving increasing attention by the community. The motivation for such an interest is clear: a superlattice configuration provides the unique opportunity of enhancing materials properties via “strain engineering”, and a multifunctional compound such as BiFeO3 appears to be a natural candidate in this context. (For example, strain has been predicted to enhance the magnetoelectric response of BiFeO3 by several orders of magnitude compared to bulk samples [wojdel09, ; wojdel10, ].) Also, a superlattice geometry can alleviate the leakage issues of pure BiFeO3 films [ranjith07, ; ranjith08, ]. Combining a III–III perovskite like BiFeO3 (or I–V, like KNbO3) with a II–IV titanate appears, however, problematic from the conceptual point of view. In fact, the charge-family mismatch inevitably leads to polar (and hence electrostatically unstable) interfaces between layers murray09 . This is not necessarily a drawback, though: recent research has demonstrated that polar interfaces can be, rather than a nuisance to be avoided, a rich playground to be exploited for exploring exciting new phenomena. The prototypical example is the LaAlO3/SrTiO3 system, where a metallic two-dimensional electron gas appears at the heterojunction to avoid a “polar catastrophe” nakagawa06 ; ohtomo04 . Remarkably, first-principles calculations have shown that interfaces in oxide superlattices can remain insulating provided that the layers are thin enough, and produce rather dramatic effects on the respective polarization of the individual components bristowe09 ; murray09 . This means that, in a superlattice, polar discontinuities need not be compensated by electronic or ionic reconstructions; they can, instead, be used as an additional, powerful materials-design tool to control the behavior of the polar degrees of freedom therein. Such a control may be realized, for instance, by altering the stoichiometry at the interfaces (see Fig. 1b). To fully explore the potential that this additional degree of freedom (the interface built-in polarity) provides, and guide the experimental search for the most promising materials combinations, one clearly needs to establish a general theoretical framework where the “compositional charge” murray09 is adequately taken into account. Figure 1: (Color online) (a) Description of the electrostatic coupling in a ferroelectric (orange)/paraelectric (blue) bilayer; $P$, $\mathcal{E}$, and $D$ represent the component of the polarization, electric field and electric displacement vectors along the stacking direction, and $\sigma_{\rm int}$ is the interface charge density. (b) Intermixed AO-type interfaces in a [BiFeO3]m/[PbTiO3]n superlattice and the resulting interface charge densities. (c) Illustration of the $20$-atom simulation cell used in our calculations; red, blue and black spheres represent O, B, and A atoms in the ABO3 perovskite. In this Letter, we present a general first-principles approach to predict the behavior of charge-mismatched perovskite oxide superlattices based exclusively on the properties of their individual bulk constituents. Our formalism combines the constrained-$D$ strategies of Wu et al. wu08 , which are key to decomposing the total energy of the system into the contributions of the individual layers, with the rigorous description of the interface polarity proposed in Ref. stengel11 . As a result, we are able to exactly describe the electrostatic coupling and mechanical boundary conditions, enabling a clear separation between genuine interface and bulk effects. Crucially, the present method allows one to quantify, in a straightforward way, the impact that interface polarity has on the equilibrium (and metastable) phases of the superlattice. As a proof of concept we apply our formalism to the study of [PbTiO3]m/[BiFeO3]n (PTO/BFO) heterostructures. We find that (i) our _bulk_ model accurately matches earlier first-principles predictions obtained for ultrashort-period superlattices (i.e., $m=n=3$) by using explicit _supercell_ simulations stengel12 , and (ii) by assuming interface terminations with different nominal charge, we obtain a radical change in the overall ferroelectric properties of the superlattice, which demonstrates the crucial role played by the polar mismatch. Figure 2: (Color online) Energy of PTO/BFO superlattices with $a=3.81$ Å expressed as a function of $D$, for selected values of $\lambda$ and $\sigma_{\rm int}$. Equilibrium and metastable superlattice states are represented with solid and empty dots. Red (green) vertical lines indicate phase transitions occurring in bulk BFO (PTO) under different $D$ conditions. (a) and (b) represent the cases of neutral and polar interfaces, respectively. We start by expressing the total energy of a monodomain two-color superlattice (i.e., composed of species A and B) as, $U_{\rm tot}(D,\lambda,a)=\lambda\cdot U_{\rm A}(D,a)+\left(1-\lambda\right)\cdot U_{\rm B}(D,a)~{}.$ (1) Here $U_{\rm A}$ and $U_{\rm B}$ are the internal energies of the individual constituents, $D$ is the electric displacement along the out-of-plane stacking direction (i.e., $D\equiv{\cal E}+4\pi P$ where ${\cal E}$ is the electric field and $P$ is the _effective_ polarization, relative to the centrosymmetric reference configuration), $\lambda$ is the relative volume ratio of material A (i.e, $\lambda\equiv m/(n+m)$ where $m$ and $n$ are the thicknessess of layers A and B, respectively), and $a$ is the in-plane lattice parameter (we assume heterostructures that are coherently strained to the substrate). Note that short-range interface effects have been neglected. (While it is certainly possible to incorporate the latter in the model, e.g. along the guidelines described in Ref. wu08 , we believe these would have been an unnecessary complication in the context of the present study.) By construction, Eq. (1) implicitly enforces the continuity of $D$ along the out-of-plane stacking direction (which we label as $z$ henceforth), which is appropriate for superlattices where the interfaces are nominally uncharged [ghosez08, ; junquera11, ]. In presence of a polar mismatch, one has a net “external” interface charge, of compositional origin murray09 , $\sigma_{\rm int}$ (see Fig. 1a), which is localized at the interlayer junctions. In such a case, Eq. (1) needs to be revised as follows, $U_{\rm tot}(D,\sigma_{\rm int},\lambda,a)=\lambda\cdot U_{\rm A}(D,a)+\left(1-\lambda\right)\cdot U_{\rm B}(D-\sigma_{\rm int},a)~{},$ (2) i.e. the $U_{\rm B}$ curve is shifted in $D$-space to account for the jump in $D$ produced by $\sigma_{\rm int}$. (Recall the macroscopic Maxwell equation, $\nabla\cdot{\bf D}=\rho_{\rm ext}$, where $\rho_{\rm ext}$, the “external” charge, encompasses all contributions of neither dielectric nor ferroelectric origin.) Once the functions $U_{\rm A}$ and $U_{\rm B}$ are computed and stored (e.g. by using the methodology of Ref. stengel09b ), one can predict the ground-state of a hypothetical A/B superlattice by simply finding the global minimum of $U_{\rm tot}$ with respect to $D$ at fixed values of $\sigma_{\rm int}$, $\lambda$ and $a$. The advantage of this procedure is that, for a given choice of A and B, the aforementioned four-dimensional parameter space can be explored very efficiently, as no further _ab initio_ calculations are needed. It is useful, before going any further, to specify the physical origin of $\sigma_{\rm int}$ in the context of this work. Consider, for example, a periodic BiFeO3/PbTiO3 superlattice, which we assume (i) to be stoichiometric (and therefore charge-neutral) as a whole, (ii) to have an ideal AO-BO2-AO-BO2 stacking along the (001) direction, and (iii) to form (say) AO-type interfaces (see Fig. 1b). (The same arguments can be equally well applied to the case of BO2-type interfaces.) Depending on the growth conditions, one can have a certain degree of intermixing in the boundary AO layers, which will adopt an intermediate composition Bix Pb(1-x)O. As a pure BiO layer is formally charged $+1$ and PbO is neutral, we can readily write $\sigma_{\rm int}=\pm\left(x-\frac{1}{2}\right)$ (expressed in units of $e/S$ with $S$ being the surface of the corresponding 5-atom perovskite cell), where the choice of plus or minus depends on the arbitrary assignment of BiFeO3 and PbTiO3 as the A or B component in Eq. (2) [see Fig. 1b]. In the following we shall illustrate the crucial role played by $\sigma_{\rm int}$ (and hence, by the interface stoichiometry) on the ferroelectric properties of a BFO/PTO superlattice, by combining Eq. (2) with the bulk $U_{\rm BFO}(D,a)$ and $U_{\rm PTO}(D,a)$ curves that we calculate from first principles. Our calculations are performed with the “in-house” LAUTREC code within the local spin density approximation to density-functional theory. (We additionally apply a Hubbard $U=3.8$ eV to Fe ions kornev07 ; yang12 .) We use the $20$-atom simulation cell depicted in Fig. 1c for both BFO and PTO, which allows us to describe the ferroelectric and anti-ferrodistortive (AFD) modes of interest (i.e. in-phase AFDzi and out-of-phase AFDzo and AFDxy, see Ref. [bousquet08, ]). Atomic and cell relaxations are performed by constraining the out-of-plane component of $D$ stengel09b and the in-plane lattice constant $a$ to a given value. [Calculations are repeated many times in order to span the physically relevant two-dimensional $(D,a)$ parameter space.] We start by illustrating the results obtained at fixed strain, $a=3.81$ Å (see Fig. 2), by assuming $\sigma_{\rm int}=0$, which corresponds to fully intermixed junctions ($x=0.5$), and we vary the BFO volume ratio, $\lambda$. At the extreme values of $\lambda$, the results are consistent with the expectations: the equilibrium configuration of BFO (i.e., the minimum of $U_{tot}$ with $\lambda=1$) at this value of $a$ is the well-known R-type $Cc$-I phase alison10 , derived from the bulk ground state via the application of epitaxial compression; PTO ($\lambda=0$), on the other hand, is in a tetragonal $P4mm$ phase with the polarization vector oriented out of plane. Intermediate values of $\lambda$ yield a linear combination of the two single- component $U(D)$ curves, where the spontaneous $P_{z}$ at equilibrium gradually moves from the pure PTO to the pure BFO value. Unfortunately, the possible equilibrium states that can be attained by solely varying $\lambda$ (at this value of $a$ and $\sigma_{\rm int}$) lie far from any physically “interesting” region of the phase diagram. For example, note the kink at $|D|\sim$0.3 C/m2 in the pure BFO case, which corresponds to a first-order transition to an orthorhombic $Pna2_{1}$ phase (a close relative of the higher-symmetry $Pnma$ phase, occurring at $D=0$). A huge piezoelectric and dielectric response is expected in BFO in a vicinity of the transition cazorla14 , raising the question of whether one could approach this region by playing with $\sigma_{\rm int}$, in addition to $\lambda$. The answer is yes: when oxide superlattices with $\sigma_{\rm int}=0.5$ are considered [corresponding to “ideal” (BiO)+/TiO2 and (FeO2)-/PbO interfaces], the stable minimum of the system favors a smaller spontaneous polarization in the BFO layers, approaching the aforementioned ($Cc{\rm-I}\to Pna2_{1}$) phase boundary in the limit of small $\lambda$. Interestingly, the $U_{\rm tot}(D)$ curve becomes asymmetric (the interfacial charge breaks inversion symmetry), and a secondary, metastable minimum appears. Overall, the resulting phase diagram turns out to be much richer, with new combinations of phases emerging (e.g. in region II’, where BFO exists in the orthorhombic $Pna2_{1}$ phase and PTO in the tetragonal $P4mm$ phase), and highly non-trivial changes in the electrical properties occurring as a function of $\lambda$. Figure 3: Total energy (a) and out-of-plane electric displacement $D$ (b) of the equilibrium (solid symbols) and metastable (empty symbols) states of PTO/BFO superlattices with $\lambda=\frac{1}{2}$ and $\sigma_{\rm int}=0.5$, expressed as a function of the in-plane lattice parameter. Regions in which PTO and BFO exist in different phases are delimited with vertical dashed lines; the corresponding space groups and AFD distortion patterns in Glazer’s notation are shown in (a), and the components of the ferroelectric polizarization in (b). In order to further illustrate the power of our approach, we shall now fix the volume ratio to $\lambda=0.5$ (corresponding to alternating BFO and PTO layers of equal thickness) and vary the in-plane lattice parameter in the range $3.6\leq a\leq 4.2$ Å . We shall first consider the case of charged interfaces with $\sigma_{\rm int}=0.5$, as this choice allows for a direct comparison with the results of Yang _et al._ (obtained via standard supercell simulations) [stengel12, ]. In Fig. 3 we show the energy and spontaneous electric displacement of the equilibrium and metastable states as a function of $a$. Four regions can be identified in the diagrams depending on the phases adopted by BFO and PTO at each value of the in-plane strain. (Their crystal space groups, AFD pattern and in-plane / out-of-plane ferroelectric polarization, respectively $P_{xy}$ and $P_{z}$, are specified in compact form in the figure.) In region I’ both PTO and BFO adopt a monoclinic $Pc$ phase characterized by large in-phase AFDz distortions and non-zero $P_{xy}$ and $P_{z}$. Such a monoclinic $Pc$ phase is closely related to the orthorhombic $Pmc2_{1}$ structure which has been recently predicted in PTO and BFO at large tensile strains [yang12, ]. In region II’ PTO adopts an orthorhombic $Ima2$ phase, characterized by vanishing AFD distortions and a large in-plane ${\bf P}$ (we neglect the small out-of-plane $P_{z}$), while BFO is in its well- known $Cc$-I state. In region III, BFO remains $Cc$-I, while PTO adopts a $P4mm$ phase, both with _opposite_ out-of-plane polarization with respect to region II’. These structures switch back to a positively oriented $P_{z}$ in region IV’, respectively transforming into a monoclinic $Cc$-II and a tetragonal $I4cm$ phase. The $I4cm$ phase is characterized by anti-phase AFDz distorsions and an out-of-plane ${\bf P}$, while the $Cc$-II corresponds to the “supertetragonal” T-type phase of BFO zeches09 . Note that, as observed already while discussing Fig. 2, the net interface charge leads to an asymmetric double-well potential, and consequently to an energy difference (typically of $\sim 20$ meV/f.u. or less, see Fig. 3a) between the two oppositely polarized states. (Only one minimum survives at large tensile strains, where the superlattice is no longer ferroelectric.) At the phase boundaries such energy difference vanishes; the obvious kinks in the $U_{\rm tot}$ curve shown in Fig. 3(a) indicate that the transitions (at $a=3.71$, $3.87$, and $4.05$ Å) are all of first-order type. The above results are in remarkable agreement with those of Yang _et al._ stengel12 . The only apparent discrepancy concerns the ordering of the stable/metastable states in region III, which anyway involves a very subtle energy difference (and is therefore sensitive to short-range interface effects, not considered here). Obtaining such an accurate description of superlattices where the individual layers are as thin as three perovskite units stengel12 provides a stringent benchmark for our method, and validates it as a reliable modeling tool. From the physical point of view this comparison suggest that, even in the ultrathin limit, PTO/BFO superlattices can be well understood in terms of macroscopic bulk effects, i.e., short-range interface-specific phenomena appear to play a relatively minor role. Figure 4: Same as in Fig. 3, but considering neutral interfaces. The out-of- plane polarization is the same in PTO and BFO layers. Having gained confidence in our method, we can use it to predict the behavior of a hypothetical superlattice with $\sigma_{\rm int}=0$, corresponding to a centrosymmetric reference structure with fully intermixed Pb0.5Bi0.5O interface layers (see Fig. 4). Note the symmetry of the two opposite polarization states, and the common value of the spontaneous electric displacement adopted by BFO and PTO. The resulting phase diagram consists, again, in four regions, with a first-order and two second-order phase transitions occurring at $a=4.07$, $3.88$ and $3.73$ Å , respectively (see Fig. 4a). In three of these regions, the individual layers display structures which are different from those obtained in the $\sigma_{\rm int}=0.5$ case: in region I both PTO and BFO stabilize in an orthorhombic $Pmc2_{1}$ phase [yang12, ], characterized by a vanishing $P_{z}$; in region II PTO adopts a monoclinic $Cm$ phase with the polarization roughly oriented along (111) ($P_{z}\neq P_{xy}\neq 0$) and no AFD, while BFO stabilizes in the already discussed $Cc$-I phase; finally, in region IV, PTO is tetragonal $P4mm$ and BFO is monoclinic $Cc$-II. These findings quantitatively demonstrates that the interface charge mismatch can have a tremendous impact on the physical properties of oxide superlattices. Our simple and general method allows one to model and quantify accurately these effects, and most importantly to rationalize them in terms of intuitive physical concepts. In summary, we have discussed a general theoretical framework to predict the behavior of charge-mismatched superlattices. We have showed that the effect of the interface stoichiometry, which we describe via the “compositional” interface charge $\sigma_{\rm int}$, is quite dramatic, and needs to be properly accounted for in the models. More generally, we argue that $\sigma_{\rm int}$ can be regarded, in addition to $\lambda$ and $a$, as a further degree of freedom in designing oxide heterostructures with tailored functionalities, opening exciting new avenues for future research. ###### Acknowledgements. This work was supported by MICINN-Spain [Grants No. MAT2010-18113 and No. CSD2007-00041], and the CSIC JAE-DOC program (C.C.). We thankfully acknowledge the computer resources, technical expertise and assistance provided by RES and CESGA. ## References * (1) P. Ghosez and J. Junquera, J. Comp. Theor. Nanosci. 5, 2071 (2008). * (2) C. Lichtensteiger et al., in Oxides Ultrathin Films: Science and Technology, edited by G. Pacchioni and S. Valeri, Ch. 12, 265 (Wiley-VCH, Germany, 2011). * (3) M. Dawber et al., Phys. Rev. Lett. 95, 177601 (2005). * (4) M. Dawber et al., Adv. Mater. 19, 4153 (2007). * (5) J. Sinsheimer et al., Phys. Rev. Lett. 109, 167601 (2012). * (6) M. Dawber, K. M. Rabe, and J. F. Scott, Rev. Mod. Phys. 77, 1083 (2005). * (7) P. Zubko et al., Nano Letters 12, 2846 (2012). * (8) C. W. Swartz and X. Wu, Phys. Rev. B 85, 054102 (2012). * (9) P. Aguado-Puente, P. García-Fernández, and J. Junquera, Phys. Rev. Lett. 107, 217601 (2011). * (10) E. Bousquet, M. Dawber, N. Stucki, C. Lichtensteiger, P. Hermet, S. Gariglio, J.-M. Triscone, and P. Ghosez, Nature (London) 452, 732 (2008). * (11) J. C. Wojdeł and J. ${\rm\acute{I}}$${\rm\tilde{n}}$iguez, Phys. Rev. Lett. 103, 267205 (2009). * (12) J. C. Wojdeł and J. ${\rm\acute{I}}$${\rm\tilde{n}}$iguez, Phys. Rev. Lett. 105, 037208 (2010). * (13) R. Ranjith, B. Kundys, and W. Prellier, Appl. Phys. Lett. 91, 222904 (2007). * (14) R. Ranjith et al., Appl. Phys. Lett. 92, 232905 (2008). * (15) E. D. Murray and D. Vanderbilt, Phys. Rev. B 79, 100102 (2009). * (16) N. Nakagawa, H. Y. Hwang, and D. A. Muller, Nature Mater. 5, 204 (2006). * (17) A. Ohtomo and H. Y. Hwang, Nature (London) 427, 423 (2004). * (18) N. C. Bristowe, E. Artacho, and P. B. Littlewood, Phys. Rev. B 80, 045425 (2009). * (19) X. Wu, M. Stengel, K. M. Rabe, and D. Vanderbilt, Phys. Rev. Lett. 101, 087601 (2008). * (20) M. Stengel and D. Vanderbilt, Phys. Rev. B 80, 241103(R) (2009). * (21) Y. Yang, M. Stengel, W. Ren, X. H. Yan, and L. Bellaiche, Phys. Rev. B 86, 144114 (2012). * (22) M. Stengel, N. A. Spaldin, and D. Vanderbilt, Nature Physics 5, 304 (2009). * (23) I. A. Kornev, S. Lisenkov, R. Haumont, B. Dkhil, and L. Bellaiche, Phys. Rev. Lett. 99, 227602 (2007). * (24) Y. Yang, W. Ren, M. Stengel, X. H. Yan, and L. Bellaiche, Phys. Rev. Lett. 109, 057602 (2012). * (25) A. J. Hatt, N. A. Spaldin, and C. Ederer, Phys. Rev. B 81, 054109 (2010). * (26) C. Cazorla and M. Stengel, to be published. * (27) R. J. Zeches et al., Science 326, 977 (2009).
# Fuzzy Logic-based Robust Failure Handling Mechanism for Fog Computing Ranesh Kumar Naha, Saurabh Garg, Muhammad Bilal Amin School of Technology, Environments and Design University of Tasmania Hobart, Australia <EMAIL_ADDRESS>Rajiv Ranjan School of Computing Newcastle University Newcastle, United Kingdom <EMAIL_ADDRESS> ###### Abstract Fog computing is an emerging computing paradigm which is mainly suitable for time-sensitive and real-time Internet of Things (IoT) applications. Academia and industries are focusing on the exploration of various aspects of Fog computing for market adoption. The key idea of the Fog computing paradigm is to use idle computation resources of various handheld, mobile, stationery and network devices around us, to serve the application requests in the Fog-IoT environment. The devices in the Fog environment are autonomous and not exclusively dedicated to Fog application processing. Due to that, the probability of device failure in the Fog environment is high compared with other distributed computing paradigms. Solving failure issues in Fog is crucial because successful application execution can only be ensured if failure can be handled carefully. To handle failure, there are several techniques available in the literature, such as checkpointing and task migration, each of which works well in cloud based enterprise applications that mostly deals with static or transactional data. These failure handling methods are not applicable to highly dynamic Fog environment. In contrast, this work focuses on solving the problem of managing application failure in the Fog environment by proposing a composite solution (combining fuzzy logic- based task checkpointing and task migration techniques with task replication) for failure handling and generating a robust schedule. We evaluated the proposed methods using real failure traces in terms of application execution time, delay and cost. Average delay and total processing time improved by 56% and 48% respectively, on an average for the proposed solution, compared with the existing failure handling approaches. ###### Index Terms: Fog Computing, Application Failure, Dynamic Resource, Fault Tolerance, Robustness ## I Introduction Fog computing is a distributed computing paradigm in which any device that has computation capability can contribute to application processing. These devices include mobile, network, handheld and mobile devices [1]. Cloud computing has latency issues in which time-sensitive real-time application execution cannot be performed. Hence, Fog computing has emerged; it can process user application requests with minimum latency, since Fog devices are located close to the users. Applications in a smart environment must be able to respond instantly, without delay. Some example of these kinds of applications are smart cars, augmented reality applications, online multiplayer games and emergency response applications. Fog is a highly distributed environment in which numerous autonomous devices contribute to processing application requests; these are known as Fog devices [1]. The contributing devices can produce a financial benefit by allowing the Fog platform to use their resources for Fog application processing [2]. Unlike cloud resources, fog computing resources are not managed service, hence, they have high probability of failure. Furthermore, the devices in Fog environment are not completely dedicated to Fog application processing [3]. Hence, there is no guarantee that devices are always available. The device can even fail after starting the processing of the Fog application. In such a scenario, it is important to make the scheduling of the application robust for successful execution in the Fog environment, despite any failure of the Fog devices at this stage. This will reduce the impact of any application failure in the Fog environment. Without ensuring proper failure handling mechanisms, it is not possible to run time-sensitive real-time applications in the Fog environment. This is because the failure of a Fog application or device might contribute to the high rates of delay which lead to the execution of the application being unsuccessful. Thus, this is an important and challenging area for research. The need for the successful execution of Fog applications in the dynamic Fog environment motivated us to explore the reasons for failure and to investigate possible solutions. In order to meet the time sensitivity of the applications, handling failure in the Fog environment is an important and challenging task [4]. For application processing, the cloud computing environment mostly depends on the cloud data centre [5] in which the rate of failure is not that high compared with the Fog environment. Fog devices are controlled by decentralised entities while cloud data centres are managed by some central entities. Hence, predicting application failure in the cloud environment is less complicated compared with such predictions in Fog computing. In the Fog environment, it is difficult to predict the failure of the computation resources due to the unstable characteristics of the available resources in the Fog devices. Thus, a robust scheduling algorithm needs to be developed. On the other hand, to prove the correctness of the robust scheduling, an evaluation with real failure traces is required. Robust scheduling is a mechanism by which application will be executed in a way in which there will no opportunity for application failure. Even if there is a failure robust adaptation mechanism can bring back the system to stable state. To ensure this, applications might need to execute in different places at the same time, in order to avoid the risk of failure. Two methods are available for handling failures in service-oriented computing; these are proactive and reactive failure handling mechanisms [6] [7]. Proactive failure needs to be considered in a highly distributed environment in such a way that failure handling has taken place before occurrence of the failure. However, proactive failure handling will not always be suitable because some failures might occur beyond our prediction. This is because of the possibilities of device malfunction, user interruption and uncertain changes in resource availability. Therefore, we need to consider reactive failure handling which usually takes place after the failure has occurred. Due to the unstable nature of Fog devices, applying either of the methods is not always useful for ensuring the successful execution of the application in a time-sensitive manner. Hence, we used task replication methods, along with proactive and reactive failure handling methods. In this paper, we propose a composite solution by using proactive and reactive failure handling methods with replication. The key contributions of this paper are as follows: 1. 1. We propose a fuzzy logic-based method to deal with unpredicted and predicted failures. 2. 2. Our failure handling method considers time sensitivity of the application, as well as dynamic changes in the available resources in the devices. 3. 3. We evaluate the proposed failure handling method using real failure traces. The rest of the paper is organised as follows: Section II presents related literature of failure handling mechanisms in the P2P system, a cluster, a grid, the cloud and the Fog computing paradigm. Section III discusses the application scenario for the proposed solution. Section IV discusses the definition of the problem. A brief description of various resources failures and their solutions is presented in Section V. Section VI gives a detailed description of the proposed Fuzzy logic-based solution. Section VII discusses the experimental setup and evaluation technique. Section VIII presents experimental results and discussion. Finally, Section IX concludes the paper. ## II Related Work This section describes some related work on failure handling mechanisms in different distributed computing paradigms. We chose to survey failure handling methods in order to verify the uniqueness of our research. We reviewed some methods that are being used for the P2P system, as well as cluster, grid, cloud and Fog computing paradigms, in order to make the system fault-tolerant. ### II-A Failure handling in P2P System Samant and Bhattacharyya [8] examined the impact of node failure on the accessibility of resources on P2P networks. Their work examined how search efforts, the topology of the network and redundant resources can influence accessibility when various level node failures take place. Vanthournout and Deconinck [9] proposed three strategies to realise the use of self- organisation mechanisms for failure handling and failure detection. Lin et al. [10] presented an efficient fault-tolerant approach for the super- peers of peer-to-peer (P2P) file-sharing systems. Mawlood-Yunis et al. [11] identified the disconnection failure problem, due to temporary semantic mapping faults, and proposed a game theory based Fault Tolerant Adaptive Query Routing (FTAQR) algorithm to resolve it. ### II-B Failure handling in Cluster Li and Lan [12] proposed FT-Pro, an adaptive fault management mechanism that optimally chooses migration, checkpointing or no action to reduce the application execution time in the presence of failures based on the failure prediction. Various methods have been used in cluster computing to predict failure event. These methods include genetic algorithms [13], rule-based data mining [14, 15, 16] and Markov chain [17]. Many other works focus on fault management techniques which are based on prediction. Leangsuksun et al. [18] proposed a predictive failure handling mechanism which scheduled deliberate job checkpointing and migration. In another work, Castelli et al. [19] employed a different approach to failure prediction. In their approach, they first predicted the resource exhaustion failure proactively and then conducted software rejuvenation. To maximise system throughput, Oliner et al. [20] used the coordinative checkpointing strategy that optimistically skips a checkpoint when no failure prediction exists in the near future. Chakravorty et al. [21] proposed software-based prediction of failure which basically migrates a task before the failure actually occurs. ### II-C Failure handling in Grid Hwang and Kesselman [22] proposed a flexible failure handling framework for the grid which is comprised of two phases: failure detection and recovery phases. In the failure detection phase, an event notification mechanism reports failures. A failure handler deals with the failures at two levels: the task level and the workflow level. Task level failures are handled by retrying, checkpointing and replication. At the workflow level, they are managed by alternative task and redundancy. Jin et al. [23] proposed the Fault Tolerant Grid Platform (FTGP) approach from the perspective of grid users, taking the nature of grid faults into account. Lee et al. [24] proposed a resource manager for optimal resource selection. The proposed resource manager automatically selects a set of optimal resources from candidate resources which achieve optimal performance using a genetic algorithm with a fault tolerance service that satisfies QoS requirements. Lee et al. [24] implemented a fault detector and a fault manager which will handle failure by job migration, using a checkpoint. Kandaswamy et al. [25] proposed a fault-tolerance technique using over-provisioning and migration for computational grids. Khoo and Veeravalli [26] proposed a failure handling mechanism based on pro-active failure handling strategies for grid environments. ### II-D Failure handling in cloud computing Much research on handling failures in the cloud environment has been undertaken to provide a failure-prone environment. Two review works [6, 7], in which all kinds of failures were categorised into reactive and proactive failure methods, extensively evaluated failure handling mechanisms in the cloud. Reactive failure mechanisms were further divided into three sub- categories: checkpointing, replication and logging. While Virtual Machine (VM) migration was considered as proactive failure management, Gill and Buyya [6] suggested continuous monitoring of resource allocation to manage failures in the cloud environment during operation. Sharma et al. [7] point out that predicting resource behaviour is critical in the cloud environment. Sharma et al. [27] proposed a failure-aware VM consolidation technique based on exponential smoothing. They employed checkpointing and migration of VM in their proposed method. Luo et al. [28] proposed a Markov-based model to examine failures in large-scale cloud computing systems. They employed reliability-aware resource scheduling to improve fault tolerance. Although cloud computing is a mature technology, it still lacks service reliability. Hence, Buyya et al. [29] suggested investigating failure-aware provisioning and reliability-aware service management mechanisms for the cloud computing environment. ### II-E Failure handling in Fog computing Existing failure handling methods in P2P, distributed and cloud computing mainly scale infrastructure to utilise extra resources to cover failure but in Fog computing, fault tolerance is challenging due to some unfavourable factors, such as resource constraints and multiple procedures [30]. Most of those methods considered only one failure handling method (for example, checkpointing or replication or resubmission) for fault tolerance [31]. Also, they did not consider any time sensitivity of the user request [31]. Hence, some researchers proposed new methods for failure handling in the Fog computing environment [31, 32]. A Fault-Tolerant Scheduling Method (FTSM) was proposed by Alarifi et al. [31] for the Fog-Cloud environment. In their approach, the system submits time- tolerant requests to the cloud and time-sensitive requests to the edge devices. FTSM finds the checkpoint interval based on the operation time between failures for the devices. However, Alarifi et al. [31] did not consider any prediction of the failure for devices based on the fluctuating availability of computation resources in the devices. Tajiki et al. [32] proposed the Heuristic Fast Failure Recovery (HFFR) algorithm for software- defined service function chaining for Fog computing with failure consideration. The main idea of their proposed method is to find failure probability based on the predefined threshold. Similar to FTSM, HFFR did not consider the dynamic changes in the available resources. In addition, neither of the works considered real failure traces for evaluating their proposed methods. Battula et al. [33] proposed an efficient resource monitoring service for Fog computing which suggested failure handling is essential for efficient resource management in the Fog environment. In summary, existing failure handling methods in Fog computing did not take into account fully the dynamic availability of Fog resources. In this paper, we propose a combined approach of proactive and reactive failure handling with task replication to tackle highly dynamic behaviour of Fog resources. Thus, this research was carried out to propose a composite solution of utilising proactive, reactive and replication failure handling methods with dynamic changes of the resources in the Fog devices. The bivalent proposition of the Fuzzy logic technique motived us to employ this approach for failure handling. ## III Failure Issues and Scenario In this section we describe the application scenario and the research problem. We also discuss the reasons for resource failure and some possible solutions for handling failures in the Fog environment. ### III-A Application Scenario To demonstrate the problem solved in this paper, an application scenario is presented in this section. Let us assume that an emergency vehicle is using a smart transportation application and moving from point A to point B. The vehicle has to choose the shortest route to the destination. To fulfil this requirement the system needs to process data generated or stored in a dash cam, surveillance camera and sensors. Based on the traffic conditions, the following actions need to be taken: (i) inform other vehicles ahead that an emergency vehicle is approaching; (ii) override signals if there are multiple road junctions along the way; (iii) do the relevant processing in the Fog devices, and (iv) take action following the processing. The overall scenario and system architecture is presented in Figure 1. Figure 1: Application scenario and system architecture. Other incidents might also occur while the emergency vehicle is enroute. The system should act promptly to minimise the delay in reaching the destination. Here, the system needs to process data from sensors, as well as video data from dashcams and surveillance cameras. All of the processing for the above application scenarios is done in Fog devices to comply with the need for time sensitivity. Therefore, the utilisation of processing power and on-time processing are important. It is possible to ensure time sensitivity of the application by distributing the application workload among Fog devices . But the issue is what will happen if the Fog node has failed? We need to ensure that the outcome of the application should meet time-sensitive requirements in which the robustness of the scheduling approach will be assured. Robustness is a feature of the scheduling process in which application execution will be successful by ensuring time sensitivity, even if the resource has failed, any errors have occurred in the system components or any erroneous input has taken place. In our application scenario, the application always requests the completion of the processing by defining a deadline. However, our concern is how to deal with the failure of the resources during operation. We are specifically focused on minimising the impact of the failure on the applications, due to resource failure, by handling the situations in which Fog device resources have failed. ### III-B Problem Definition This research was carried out to solve the following problem: How to meet user requirements for applications in the Fog environment, with consideration of device failure, in order to satisfy any time-sensitive requirements of the application, while available resources in the devices are changing dynamically? Scheduling all related tasks to Fog devices is not such a complicated task if we can assume that all devices are up and running, and there are no chances of their failure. But, in reality, the chance of failure in the Fog environment is very high since the devices are not dedicated to running Fog applications. On the other hand, most of the devices in the Fog depend on wireless connectivity. Also, the devices are mobile and are moving frequently from one cluster to another. Next, most of the Fog devices are not stationary, meaning that the devices have limited battery power. Furthermore, the application might be interrupted by the owner of the devices (for example, the owner turns off the device for some reason; the owner does not want to participate at that moment or the owner wants to run another application which requires some resources to be freed up). Due to all of the above reasons, the chances of failure of the computation resources are very much higher than in any other distributed system. To ensure the robustness of the scheduling algorithm, we need to deal with resource failures in a way that the application user would not affected. ### III-C Resource Failure and Counter-measures The resources could fail in the Fog environment for many reasons. The reasons for failure can be categorised, such as the termination of the application to run the native application, network failure, the device being moved to another cluster, power outage, human interruption, software and hardware failure, and network attacks. Due to the mobility and dynamicity of the available resources in the devices, we can categorise all types of failure into two basic types: (i) unpredicted/immediate failure and (ii) predicted failure. We can handle failures in two different ways. Firstly, we can manage the resource failure after it took place; this is referred to as reactive failure. Secondly, it is possible to have countermeasures before the occurrence of the resource failure; this is known as proactive failure handling. Both types of failure handling mechanism have different approaches to manage resource failures. In a reactive failure handling mechanism, we can employ checkpointing and replication. In application checkpointing the state of the application is stored in reliable storage and, if the application has failed, it does not need to rerun the application from the beginning. It will start the application from the point where the latest state has been saved. There are two types of checkpointing: i) coordinated or periodic checkpointing and ii) uncoordinated or random checkpointing. In coordinated checkpointing, the checkpoint should be consistent for the processes. In uncoordinated checkpointing, each process checkpoints its state. The other type of reactive failure handling mechanism is replication which always run replicas of the running processes in different devices. The basic way to solve immediate failure is re-running the whole application but this is not the optimum way to solve the problem. For example, if a certain percentage of processing is completed, there is no point in processing the same portion of the application. Hence, the only solution for immediate failure is checkpointing. Some researchers have argued that checkpointing is not a good solution for the Fog environment because the Fog is a highly dynamic environment [3, 34]. Yi et al. [3] suggested that replications are more suitable for the Fog but multiple Fog nodes would need to be working together. Madsen et al. [35] suggested using checkpointing for the Fog which would save computation time for the faulty tasks. Some researchers used checkpointing in the Fog as a fault-tolerant technique [36, 37]. We needed to ascertain if there were any way to accommodate checkpointing in the Fog environment. To do that we needed to evaluate our solution in simulation and also in a real environment. We evaluated our proposed method in a simulated environment with real failure traces. In a proactive failure handling process, we can employ the migration process before the resource failed. Since the Fog is for time-sensitive applications, we were required to migrate the application without disconnecting devices. Hence, we needed to employ live migration for this process. Two basic types of migration are i) pre-copy migration and ii) post-copy migration [7]. In post- copy migration, application migration needs to be initiated by suspending the application at the source which will increase down-time. To minimise downtime, pre-copy migration needed to have been employed. To resolve the predicted failure, we could have employed pre-copy live migration. Once we could predict that an application was going to fail then we could migrate the application to another Fog device. But again, the question is raised: how to decide when and where to migrate? However, this research only dealt with when to migrate, not where to migrate to. To ensure the robustness of the scheduling algorithm, we needed to handle both predicted and unpredicted failures which would minimise their impact. ## IV Fuzzy logic-based failure handling mechanism To handle predicted and unpredicted failure we employed the fuzzy logic-based solution. Classical logic usually has a bivalent proposition, which may be true or false in principle. On the other hand, fuzzy logic can represent actual membership of both true and false for a function. Some propositions might be true and false to a certain degree, rather than being true or false only. For example, for a Fog device, mobility, response time and power availability might cause the failure of a device. However, the chances of failure completely depend on the membership of each parameter (for example, mobility, response time and power availability). To represent the exact degree of membership of each parameter, a fuzzy logic-based approach was undertaken. If the unpredicted failure for a Fog device were high then the Fog device would be flagged as unreliable. To handle failure for unreliable Fog devices, replication was used to ensure the robustness of the scheduling algorithm. A predicted failure handling mechanism basically acts before the resource failure takes place. However, due to decentralised management of the Fog devices, an application might have failed but this was beyond our prediction. Thus, an unpredicted failures handling mechanism allows seamless application processing. Frequent unpredicted failures caused by a Fog device will trigger replication to ensure successful application execution. Therefore, to ensure a reliable application processing environment, all three approaches (predicted failure, unpredicted failure and replication) need to be considered. Figure 2 shows what action will be taken after calculating the failure score. Figure 2: Proposed failure handling mechanism. Over utilisation of resources always causes failure. Suleiman and Venugopal [38] modelled elasticity rules for the cloud when the resource has been scaled, when utilisation is either 75% or 80%. This indicated that the chances of failure was high when the utilisation was more than 80%. Hence, we assumed that 80% to 100% utilisation was unsafe and service migration needed to have been triggered. In another work Liu et al. [39] mentioned that the chances of server crash were high when utilisation was more than 60%. Therefore, they have chosen a workload threshold of 50% to 70%. Al-Haidari et al. [40] revealed that the upper threshold for cloud resources utilisation should be 70% to 90%. Based on these studies [38, 39, 40] we assumed that less than 50% utilisation was safe and it was necessary to checkpoint in case of failure if the utilisation were 50% to 80%. A user could change these thresholds while they were being implemented in a real environment through the proposed algorithm. ### IV-A MRP score calculation for unpredicted failure To find an unpredicted failure, the system always calculates the degree of failure by calculating membership of the following parameters: (i) device mobility, (ii) device response time and (iii) device power availability. Based on the degree of failure, the system will decide how frequently checkpointing needs to be undertaken. Based on the percentage of the device movement, we defined how readily the device could be completely moved to another network. Device mobility can be represented as $D_{m}$ which could be 0% to 100%. Device response time always maps with required response time to meet the application time sensitivity. For example, to complete an application request, the device response time should be 2ms but the device is responding in 1m; therefore, the degree of failure is within the group of 0%. On the other hand, if the device response time suddenly changed to 2.5ms then the degree of failure is within the group of 100% since it is not meeting the application time-sensitive requirements. Device response time can be represented as $D_{r}$ which could be 0% to 100%. Similarly, the power available score can be calculated based on the power that is required to complete the submitted application. All the parameters of device characteristics are transformed into a normalised range [0 to 1] during fuzzification. Fuzzy logic usually includes three phases: fuzzification, fuzzy inference and defuzzification. The fuzzy sets for the above parameters are as follows: * • Device mobility: $D_{m}\in\\{Low,Normal,High\\}$ * • Device response time: $D_{r}\in\\{Fast,Normal,Slow\\}$ * • Device power: $D_{p}\in\\{Rich,Standard,Poor\\}$ Using Equation 1, the value can be normalised to fall in the interval [0 to 1]. $\overline{D_{x}}=\frac{D_{x}-\alpha_{x}}{\beta_{x}-\alpha_{x}}$ (1) In the Equation 1, $D_{x}$ is the numerical value of $x$ where $x$ is either mobility, response time or power. The value of $x$ is within the range of $\alpha_{x}$ to $\beta_{x}$. The normalised value of the parameters’ mobility, response time and power were calculated for further operation. The degree of membership of each parameter is shown in Figure 3. (a) Mobility (b) Response time (c) Power availability Figure 3: Membership of different parameters for unpredicted failure. Figure 4: Membership for mrp score. The mobility parameter of 0% to 50% is considered as low mobility; 30% to 90% is normal mobility and 70% to 100% is considered to be high mobility. Until 30% mobility membership, we considered that the system was in safe zone. However, at the point of 30%, the mobility membership was low and was decreasing, and normal mobility membership was increasing. At the point of 50%, the low mobility membership was 0 and normal mobility membership was 1. However, at the point of 70% normal mobility, membership decreased and became 0 at 90%. On the other hand, a 70% high mobility score, which started to increase and reach 1 at 90%, meant that the device was about to fail. A similar approach was employed for response time and membership of the power availability parameters. Based on the membership of each parameter, fuzzification was completed in the Fuzzy Interference System (FIS). To develop FIS we used the jfuzzylogic toolbox [41]. The membership function for low, normal and high mobility is shown in Equations 2, 3 and 4. A similar equation is used for response time and power parameters. $\mu_{mL}(x)=\begin{cases}0,&x>d\\\ \frac{d-x}{d-c},&c\leq x\leq d\\\ 1,&x<c\end{cases}$ (2) $\mu_{mN}(x)=\begin{cases}0,&(x<a)\ or\ (x>d)\\\ \frac{x-a}{b-a},&a\leq x\leq b\\\ 1,&b\leq x\leq c\\\ \frac{d-x}{d-c},&c\leq x\leq d\end{cases}$ (3) $\mu_{mH}(x)=\begin{cases}0,&x<a\\\ \frac{x-a}{b-a},&a\leq x\leq b\\\ 1,&x>b\end{cases}$ (4) We used max function as an accumulation method by which the fuzzy outcome of a particular application is represented as $X_{i}$ Fuzzy rules: Based on the behaviour of the system fuzzy, rules were generated. If any of the parameters were high, the system would not have been capable of running any application. More clearly, if a system were highly mobile, there was a high chance of resource failure in that device. In the same way, if response time or power membership were high, then resource failure in that particular device was also high. For this particular instance the rule should be as follows: * • If $D_{m}$ is high or $D_{r}$ is slow or $D_{p}$ is poor then $UF_{{mrp}_{m}}$ is high In the above rule, $UF_{{mrp}_{m}}$ represents an unpredicted failure score for application m. In order to consider some devices as being in a safe zone, all scores of all parameters should have safe zone scores with a 0% to 50% variation. For this particular instance, the rule is as follows: * • If $D_{m}$ is low and $D_{r}$ is fast and $D_{p}$ is rich then $UF_{{mrp}_{m}}$ is low To define device membership in the checkpoint zone, mobility should be low or normal, response time should be fast or normal, and power availability should be rich or standard. The mobility membership should not be high; response time should not be slow and power should not be poor, to be in the checkpointing zone. In addition, mobility should not be low, response time should not be fast and power should not be rich at the same time. To represent the situations described above, we defined seven different rules. An example of such a rule is given as follows: * • If $D_{m}$ is low and $D_{r}$ is fast and $D_{p}$ is standard then $UF_{{mrp}_{m}}$ is normal Fuzzy inference and defuzzification: To generate an mrp score we used 0% 50% as a low score, 50% to 80% as a normal score and 80% to 100% as a high score (See Figure 3). A Center of Gravity (CoG) defuzzification method was used for calculating the mrp score. The equation for CoG is shown in equation 5. $UF_{{mrp}_{x}}=\frac{\sum_{i=1}^{n}X_{i}\times\mu_{i}}{\sum_{i=1}^{n}X_{i}}$ (5) In the above equation $n$ is the number of rules needing to be triggered. $\mu_{i}$ is the singleton value which refers to the maximum score for a particular parameter. The defuzzification value for an application was used to make decisions about application failure handling ($mrp$ score). ### IV-B CPMNR score calculation for predicted failure Some failures can be predicted based on the following criteria: * • Effect on processing based on the current CPU utilisation * • Effect on processing based on available power in the device * • Effect on processing based on device mobility * • Effect on processing based on network communication (Device is capable of completing the request but network communication might be the cause of not meeting time-sensitive requirements) * • Effect on processing based on device response time All device behaviour parameters were transformed into a normalised range [0 to 1] during fuzzification. The fuzzy sets for the above parameters are as follows: * • CPU utilisation: $D_{mc}\in\\{Low,Normal,High\\}$ * • Device power: $D_{mp}\in\\{Rich,Standard,Poor\\}$ * • Device mobility: $D_{mm}\in\\{Low,Normal,High\\}$ * • Network communication: $D_{mn}\in\\{Fast,Medium,Slow\\}$ * • Response time $D_{mr}\in\\{Fast,Normal,Slow\\}$ Using Equation 6, the value can be normalised to fall into the interval [0 to 1]. $\overline{D_{mx}}=\frac{D_{mx}-\alpha_{mx}}{\beta_{mx}-\alpha_{mx}}$ (6) In the Equation 6, $D_{mx}$ is the numerical value of $mx$ where $mx$ is either CPU utilisation, power, mobility, network communication or response time. The value of $mx$ was within the range of $\alpha_{mx}$ to $\beta_{mx}$. The normalised value of the parameters’ CPU utilisation, power, mobility, network communication and response time was calculated for further operation. The degree of membership of each parameter is shown in Figures 5. (a) CPU utilisation (b) Power availability (c) Mobility (d) Network communication (e) Response time (f) cpmnr score Figure 5: Membership of different parameters for predicted failure and cpmnr score. The CPU utilisation parameter 0% to 50% was considered to be low CPU utilisation; 30% to 90% was normal CPU utilisation and 70% to 100% was considered to be high CPU utilisation. Until 30% CPU utilisation membership, we considered that the system was in safe zone. However, at the point of 30%, low CPU utilisation membership decreased and normal CPU utilisation membership increased. At the point of 50%, low CPU utilisation membership was 0 and normal CPU utilisation membership was 1. However, at the point of 70%, normal CPU utilisation membership started to decrease and became 0 at 90%. On the other hand, a 70% high CPU utilisation score starting to increase and reaching 1 at 90% meant that the device was about to fail due to the over utilisation of the CPU. A similar approach was employed for power availability, mobility, network communication and response time parameters. Based on the membership of each parameter, fuzzification was undertaken in the FIS system. Similar to the calculation of the MRP score, we used Equations 2, 3 and 4 for the membership function of low, normal and high for CPU utilisation, power availability, mobility, network communication and response time parameters. Similar to the MRP score calculation we used the max function as an accumulation method by which the fuzzy outcome of a particular application is represented as $X_{i}$. Fuzzy rules: Based on the behaviour of the system, fuzzy rules have been generated. If any of the parameters are high the system will not be capable of running any application. More clearly, if a system has high CPU utilisation, there is a high chance of application failure in that device. In the same way, if power, mobility, network communication and response time membership are high, then application failure in that particular device will be high as well. For this particular instance the rule should be as follows: * • If $D_{mc}$ is high or $D_{mp}$ is poor or $D_{mm}$ is high or $D_{mn}$ is slow or $D_{mr}$ is slow then $PF_{{cpmnr}_{m}}$ is high In the above rule, $PF_{{cpmnr}_{m}}$ represents the unpredicted failure score for application $m$. In order to consider some devices as being in a safe zone, all scores of all parameters should have safe zone scores which are within 0% to 50% variation. For this particular instance the rule is as follows: * • If $D_{mc}$ is low and $D_{mp}$ is rich and $D_{mm}$ is low and $D_{mn}$ is fast and $D_{mr}$ is fast then $PF_{{cpmnr}_{m}}$ is low To define device membership in a checkpoint zone, CPU utilisation should be low or normal, power availability should be rich or standard, mobility should be low or normal, network communication should be fast or medium and response time should be fast or normal. The CPU utilisation membership should not be high, power should not be poor, mobility membership should not be high, network communication membership should not be slow and response time should not be slow to arrive in the checkpoint zone. Also, CPU utilisation should not be low, power should not be rich, mobility should not be low, network communication should not be fast, and response time should not be fast at the same time. To represent the situations described above, we need to defined 31 different rules. An example of such a rule is given as follows: * • If $D_{mc}$ is low and $D_{mp}$ is rich and $D_{mm}$ is low and $D_{mn}$ is fast and $D_{mr}$ is normal then $PF_{{cpmnr}_{m}}$ is normal Fuzzy inference and defuzzification: To generate the cpmnr score we used 0% 50% as a low score, 50% to 80% as a normal score and 80% to 100% as a high score (See Figure 5(f)). The Centre of Gravity (CoG) defuzzification method was used for calculating the cpmnr score. The equation for CoG is shown in Equation 7. $PF_{{cpmnr}_{x}}=\frac{\sum_{i=1}^{n}X_{i}\times\mu_{i}}{\sum_{i=1}^{n}X_{i}}$ (7) In the above equation $n$ is the number of rules needing to be triggered. $\mu_{i}$ is the singleton value which refers to the maximum score for a particular parameter. The defuzzification value for an application was used for making decisions about the predicted application failure handling (cpmnr score). ### IV-C Replication Replication of the application only applies if the rate of unpredicted (immediate) failure is high. The failure rate cannot be calculated within the execution of a few application attempts. Due to this we considered at least 10 application executions before deciding whether replication was required or not. The overall process of the failure handling process is presented in Figure 6. Figure 6: Failure handling process in the proposed method. Algorithm 1 Fuzzy- logic-based failure handling (FLBFH). Input: $App[id,D_{m},D_{r},D_{p},D_{mc},D_{mp},D_{mm},D_{mn},D_{mr},SD_{ft},A_{c}]$ Output:$Ac_{tr}[App_{id},Actions]$ for all $App[id]$ do Calculate degree of changes in mobility Calculate degree of response time changes Calculate degree of power profile changes Calculate degree of CPU utilization changes Calculate degree of changes in network comm Calculate degree of failure ($d_{f}$) if $d_{f}\geq 50$ and $d_{f}\leq 80$ then $Ac_{tr}.INSERT[App_{id},Checkpoint]$ else if $d_{f}\geq 80$ and $d_{f}<100$ then $Ac_{tr}.INSERT[App_{id},Migrate]$ else if $d_{f}\geq 100$ then $Ac_{tr}.INSERT[App_{id},CheckpointRecover]$ end if $ASD_{f}$ = ($SD_{ft}$ \+ $d_{f}$)/$A_{c}$ if $ASD_{f}\geq 50$ and $A_{c}>10$ then $Ac_{tr}.INSERT[App_{id},Replicate]$ end if end for return $Ac_{tr}[]$ ### IV-D Mapping If the mrp score for unpredicted failure is in the unsafe zone, then the system will migrate the application to other available Fog devices. It is obvious that the cpmnr score will also be in the unsafe zone, if the mrp score is in the unsafe zone. However, the cpmnr score also considers CPU utilisation and network communication for making more accurate decisions about checkpointing and migration. If either of the two scores is in the checkpointing zone, then the application checkpointing will be triggered. However, if any of these two scores are in the failed zone, then the system will see if any replicated application is running or not. If so, the system will interact with one of the replicated applications. In the case of there being no running replicated application, the system will check if there are any checkpoints there or not. If there are any checkpoints, then the application will recover from that checkpoint. In the case of no checkpoint and no replicated application, $n$, then the system will rerun the whole application which is a worse case scenario. A proposed Fuzzy-Logic-based Failure Handling (FLBFH) algorithm is presented in Algorithm 1. In Algorithm 1, $SD_{f}$ is the score for degree of failure, $A_{c}$ is the app count (The total number of times a task for an application is running), $SD_{ft}$ is the total score for degree of failure and $ASD_{f}$ is the average score for degree of failure. ## V Experimental Setup and Evaluation Technique ### V-A Failure Modelling Since no failure traces are available for the Fog, we used failure traces from the Failure Trace Archive (FTA) [42]. There are 27 real failure traces available in FTA. Most of those traces have two events: failed or not failed (available). Among them, only Los Alamos National Laboratory (LANL) [43] has failure traces with reasons such as CPU failure, power failure or network failure. Therefore, we selected LANL failure traces to model failure in the Fog environment. LANL has failure traces for nine years (1996 to 2005) which consist of 4750 nodes that form 22 High-Performance Computing (HPC) systems [43]. This trace has the records for every failure that takes place within the system and which needs administrator attention. We selected those devices from LANL failure traces which had comparatively high failure rates compared with other devices. Those selected devices did not have failure traces for the year 1996 and 2005. Due to that, we used failure traces from 1997 to 2004. The Fog environment consists of numerous nodes, each HPC nodes being considered as a single Fog node. The LANL failure traces only have the information as to whether the resource failed or not. By keeping the Fog device characteristics intact, we utilised the failure information of the Fog node during simulation- based experiments. Hence, it is logical to use LANL failure traces in our experimental scenario. ### V-B Experimental Setup In order to control over the experimental environment, we chose simulation to evaluate the proposed method. We adopted a simulation environment and performance parameters from our previous works [44] [45]. In addition, we modelled a realistic Fog environment by extending the CloudSim [46] toolkit, similar to our previous work. All submitted tasks followed deadlines which varied dynamically from 10% to 80%, similar to our previous work [45]. Successful execution of application by maintaining deadlines indicated successful processing. ### V-C Performance Metrics All the performance metrics were adopted from our previous works [44, 45]. Delay: We considered the delay between the user to the Fog resources. Delay is the time between task submission and the starting of task execution. It can be calculated as follows: $\begin{split}\color[rgb]{0,0,0}d_{t}^{x}=E_{st}^{x}-U_{S}\end{split}$ (8) In Equation 8, $d_{t}^{x}$ denotes the delay for the $x$ Fog device which is involved in task execution. $E_{st}$ is the task start time and $x-U_{S}$ is the time when the user requested the execution of the task. Processing time: Processing time is the required time to process a task. It is the time between the task processing start time $p_{st}$ and the task processing end time $p_{en}$ which can be calculated by using the Equation 10. $\begin{split}\color[rgb]{0,0,0}Pt_{t}^{x}=p_{en}^{x}-p_{st}^{x}\end{split}$ (9) In the above equation, $x$ is the Fog device which is involved in task execution, and $Pt_{t}$ is the processing time for task $t$. Processing Cost: We considered connectivity and messaging costs for processing costs. These costs are based on the AWS IoT pricing model. Cost is from $1 to $1.65 per million messages for messaging and from $0.08 to $0.132 for connectivity cost for per million minutes for various regions. We considered the price that has been allocated for the Sydney region. Processing cost can be calculated as follows: $\begin{split}\color[rgb]{0,0,0}Pc_{t}=\sum_{k=a}^{n}(M_{c}+C_{c})\end{split}$ (10) In the above Equation, $M_{c}$ is the messaging cost and $C_{c}$ is the connectivity cost and $Pc_{t}$ is the total processing cost. We calculated cost for Fog device $a$ to Fog device $n$. ### V-D Evaluation Technique We compared the proposed FLBFH with two recent works HFFR [32] and FTSM [31]. Since those two works were implemented in a different simulation environment and did not consider real failure traces, we adopted the key idea of both proposed methods to fit with our simulation environment and failure traces. We compared both methods with our proposed method in the results and discussion section to show the improvement of the FLBFH failure handling method over previously proposed methods. ## VI Results and Discussion We took eight years of failure traces to perform simulations and simulate each year’s failure traces separately for HFFR, FTSM and the proposed FLBFH methods. Performance comparison of each metric is presented below in different sub-sections. ### VI-A Delay We measured average, maximum and minimum delays for each task, as shown in Figure 7, 8 and 9. The average delay for the proposed FLBFH method was improved by around 52% and 58% for HFFR and FTSM respectively, on an average for all failure traces (Figure 7). Figure 7: Average delay for different failure traces. For the failure traces of the year 2000, 2001 and 2003, the improvement was around 51% for HFFR compared with the proposed method. The maximum improvement was found for the 2004 failure traces which was 54%. Delay improvement for the rest of the failure traces was between 52% to 53%. On the other hand, compared with FTSM, the maximum improvement was 60% for the 2003 failure traces in the proposed method. The minimum delay improvement found for the 2001 failure traces compared with the FTSM, was 55%. However, the improvement was 59% for 1998, 2000 and 2002 failure traces, and 56%, 57% and 58% respectively for 1997, 1999 and 2004 failure traces. The delay improvement was different because of the difference failure handling technique. However, our proposed FLBFH method performed better over both HFFR and FTSM methods. The maximum delay for the proposed FLBFH method was improved by around 50% and 56% for HFFR and FTSM respectively (Figure 8). Figure 8: Maximum delay for different failure traces. For all failure traces, the maximum delay was improved 49% to 50% in the proposed method, compared with the HFFR method. On the other hand, the improvement was in between of 55% and 56% in the proposed method, compared with the FTSM method. Maximum improvement was found over FTSM method for average delay and the same trend was found for maximum delay. The minimum delay was more in FLBFH for most of the cases, as compared with HFFR and FTSM (Figure 9). Compared with HFFR, it was 19% more, while it was 13% more for FRSM on an average, for all failure traces. However, for 1997, 1998 and 2001, the failure traces minimum delay improved compared with FTSM, the improvement being 17% to 20%. Since the average delay was improved for the proposed algorithm, the minimum delay will not have much effect on application processing. Figure 9: Minimum delay for different failure traces. ### VI-B Processing time There is no significant difference in the average processing time for HFFR and FTSM, compared with the proposed FLBFH method (Figure 10). However, the number of failed tasks was less in the proposed FLBFH method. Since the proposed method used a Fuzzy-logic based approach for failure handling and prediction, it was able to handle failure more efficiently, with a resulting improvement in the total processing time. Figure 10: Average processing time for different failure traces. On average, the total processing time improved by 51% and 45% for the FLBFH method, compared with HFFR and FTSM respectively as shown in Figure 11. Compared with HFFR, the improvement was around 50% to 51% and compared with FTSM, it was around 44% to 46%. Total processing time improved in the proposed method because the number of failed tasks in the proposed method was fewer which provided better failure handling and robust scheduling. Figure 11: Improvement in processing time for different failure traces with FLBFH. Figure 12: Processing cost for different failure traces. ### VI-C Cost The total processing cost was less for the proposed FLBFH method, compared with HFFR and FTSM, as shown in Figure 12. HFFR had around 77% higher cost on average for all failure traces. On the other hand, FTSM had around 44% higher cost compared with FLBFH. This indicates that the number of failed tasks was higher in HFFR and FTSM, compared with the proposed FLBFH method. ## VII Conclusion The Fog computing environment is highly dynamic in terms of available resources in the devices and the chances of failure are very high. This research contributes to minimising the total number of application failures due to the failure of the resources; it helps to improve delay and processing time by proposing a Fuzzy-logic-based failure handling method. The proposed failure handling method was evaluated using real failure traces from LANL. Compared with the existing failure handling approaches, we found an improvement in average delay and total processing time which were 56% and 48% respectively on average. For future work, we will consider the implementation and evaluation of the proposed method in the real Fog computing environment, as well as power-aware resource allocation. The proposed method can be improved further by selecting more appropriate Fog devices which have less chance of failure. ## References * [1] R. K. Naha, S. Garg, D. Georgakopoulos, P. P. Jayaraman, L. Gao, Y. Xiang, and R. Ranjan, “Fog computing: survey of trends, architectures, requirements, and research directions,” _IEEE access_ , vol. 6, pp. 47 980–48 009, 2018\. * [2] S. K. Battula, S. Garg, R. K. Naha, P. Thulasiraman, and R. Thulasiram, “A micro-level compensation-based cost model for resource allocation in a fog environment,” _Sensors_ , vol. 19, no. 13, p. 2954, 2019. * [3] S. Yi, C. Li, and Q. Li, “A survey of fog computing: concepts, applications and issues,” in _Proceedings of the 2015 workshop on mobile big data_. ACM, 2015, pp. 37–42. * [4] D. Bermbach, F. Pallas, D. G. Pérez, P. Plebani, M. Anderson, R. Kat, and S. Tai, “A research perspective on fog computing,” in _International Conference on Service-Oriented Computing_. Springer, 2017, pp. 198–210. * [5] R. Buyya, C. S. Yeo, S. Venugopal, J. Broberg, and I. Brandic, “Cloud computing and emerging it platforms: Vision, hype, and reality for delivering computing as the 5th utility,” _Future Generation computer systems_ , vol. 25, no. 6, pp. 599–616, 2009. * [6] S. S. Gill and R. Buyya, “Failure management for reliable cloud computing: A taxonomy, model and future directions,” _Computing in Science & Engineering_, pp. 1–10, 2018. * [7] Y. Sharma, B. Javadi, W. Si, and D. Sun, “Reliability and energy efficiency in cloud computing systems: Survey and taxonomy,” _Journal of Network and Computer Applications_ , vol. 74, pp. 66–85, 2016. * [8] K. Samant and S. Bhattacharyya, “Topology, search, and fault tolerance in unstructured p2p networks,” in _37th Annual Hawaii International Conference on System Sciences, 2004. Proceedings of the_. IEEE, 2004, pp. 1–6. * [9] K. Vanthournout, G. Deconinck, and R. Belmans, “Building dependable peer-to-peer systems,” in _DSN 2004 Workshop on Architecting Dependable Systems_ , 2004, pp. 297–301. * [10] J.-W. Lin, M.-F. Yang, and J. Tsai, “Fault tolerance for super-peers of p2p systems,” in _13th Pacific Rim International Symposium on Dependable Computing (PRDC 2007)_. IEEE, 2007, pp. 107–114. * [11] A.-R. Mawlood-Yunis, M. Weiss, and N. Santoro, “From p2p to reliable semantic p2p systems,” _Peer-to-peer networking and applications_ , vol. 3, no. 4, pp. 363–381, 2010. * [12] Y. Li and Z. Lan, “Exploit failure prediction for adaptive fault-tolerance in cluster computing,” in _Sixth IEEE International Symposium on Cluster Computing and the Grid (CCGRID’06)_ , vol. 1. IEEE, 2006, pp. 1–8. * [13] G. M. Weiss and H. Hirsh, “Learning to predict rare events in event sequences,” in _KDD_ , 1998, pp. 359–363. * [14] R. Vilalta and S. Ma, “Predicting rare events in temporal domains,” in _2002 IEEE International Conference on Data Mining, 2002. Proceedings._ IEEE, 2002, pp. 474–481. * [15] R. K. Sahoo, A. J. Oliner, I. Rish, M. Gupta, J. E. Moreira, S. Ma, R. Vilalta, and A. Sivasubramaniam, “Critical event prediction for proactive management in large-scale computer clusters,” in _Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining_. ACM, 2003, pp. 426–435. * [16] R. Vilalta and S. Ma, “Predicting rare events in temporal domains using associative classification rules,” _Technical Report_ , pp. 426–435, 2002\. * [17] G. A. Hoffmann, F. Salfner, and M. Malek, _Advanced Failure Prediction in Complex Software Systems_. Humboldt-Universität zu Berlin, Mathematisch-Naturwissenschaftliche Fakultät II, Institut für Informatik, 2011. * [18] C. Leangsuksun, T. Liu, T. Rao, S. Scott, and R. Libby, “A failure predictive and policy-based high availability strategy for linux high performance computing cluster,” in _The 5th LCI International Conference on Linux Clusters: The HPC Revolution_. Citeseer, 2004, pp. 18–20. * [19] V. Castelli, R. E. Harper, P. Heidelberger, S. W. Hunter, K. S. Trivedi, K. Vaidyanathan, and W. P. Zeggert, “Proactive management of software aging,” _IBM Journal of Research and Development_ , vol. 45, no. 2, pp. 311–332, 2001. * [20] A. J. Oliner, L. Rudolph, R. K. Sahoo, J. E. Moreira, and M. Gupta, “Probabilistic qos guarantees for supercomputing systems,” in _2005 International Conference on Dependable Systems and Networks (DSN’05)_. IEEE, 2005, pp. 634–643. * [21] S. Chakravorty, C. Mendes, and L. Kalé, “Proactive fault tolerance in large systems,” in _HPCRI Workshop in conjunction with HPCA_ , 2005, pp. 1–7. * [22] S. Hwang and C. Kesselman, “Grid workflow: a flexible failure handling framework for the grid,” in _High Performance Distributed Computing, 2003\. Proceedings. 12th IEEE International Symposium on_. IEEE, 2003, pp. 126–137. * [23] H. Jin, D. Zou, H. Chen, J. Sun, and S. Wu, “Fault-tolerant grid architecture and practice,” _Journal of Computer Science and Technology_ , vol. 18, no. 4, p. 423, 2003. * [24] H. Lee, K. Chung, S. Chin, J. Lee, D. Lee, S. Park, and H. Yu, “A resource management and fault tolerance services in grid computing,” _Journal of Parallel and Distributed Computing_ , vol. 65, no. 11, pp. 1305–1317, 2005. * [25] G. Kandaswamy, A. Mandal, and D. A. Reed, “Fault tolerance and recovery of scientific workflows on computational grids,” in _2008 Eighth IEEE International Symposium on Cluster Computing and the Grid (CCGRID)_. IEEE, 2008, pp. 777–782. * [26] B. B. Khoo and B. Veeravalli, “Pro-active failure handling mechanisms for scheduling in grid computing environments,” _Journal of Parallel and Distributed Computing_ , vol. 70, no. 3, pp. 189–200, 2010. * [27] Y. Sharma, W. Si, D. Sun, and B. Javadi, “Failure-aware energy-efficient vm consolidation in cloud computing systems,” _Future Generation Computer Systems_ , vol. 94, pp. 620–633, 2019. * [28] L. Luo, S. Meng, X. Qiu, and Y. Dai, “Improving failure tolerance in large-scale cloud computing systems,” _IEEE Transactions on Reliability_ , vol. 68, no. 2, pp. 620–632, 2019. * [29] R. Buyya, S. N. Srirama, G. Casale, R. Calheiros, Y. Simmhan, B. Varghese, E. Gelenbe, B. Javadi, L. M. Vaquero, M. A. Netto _et al._ , “A manifesto for future generation cloud computing: research directions for the next decade,” _ACM computing surveys (CSUR)_ , vol. 51, no. 5, p. 105, 2019\. * [30] Y. Liu, J. E. Fieldsend, and G. Min, “A framework of fog computing: Architecture, challenges, and optimization,” _IEEE Access_ , vol. 5, pp. 25 445–25 454, 2017. * [31] A. Alarifi, F. Abdelsamie, and M. Amoon, “A fault-tolerant aware scheduling method for fog-cloud environments,” _PLOS ONE_ , vol. 14, no. 10, pp. 1–24, 2019. * [32] M. M. Tajiki, M. Shojafar, B. Akbari, S. Salsano, and M. Conti, “Software defined service function chaining with failure consideration for fog computing,” _Concurrency and Computation: Practice and Experience_ , vol. 31, no. 8, pp. 1–14, 2019. * [33] S. K. Battula, S. Garg, J. Montgomery, and B. H. Kang, “An efficient resource monitoring service for fog computing environments,” _IEEE Transactions on Services Computing_ , 2019. * [34] K. Kai, W. Cong, and L. Tao, “Fog computing for vehicular ad-hoc networks: paradigms, scenarios, and issues,” _the journal of China Universities of Posts and Telecommunications_ , vol. 23, no. 2, pp. 56–96, 2016. * [35] H. Madsen, B. Burtschy, G. Albeanu, and F. Popentiu-Vladicescu, “Reliability in the utility computing era: Towards reliable fog computing,” in _2013 20th International Conference on Systems, Signals and Image Processing (IWSSIP)_. IEEE, 2013, pp. 43–46. * [36] J. P. de Araujo Neto, D. M. Pianto, and C. G. Ralha, “A fault-tolerant agent-based architecture for transient servers in fog computing,” in _2018 30th International Symposium on Computer Architecture and High Performance Computing (SBAC-PAD)_. IEEE, 2018, pp. 282–289. * [37] C. Puliafito, E. Mingozzi, C. Vallati, F. Longo, and G. Merlino, “Companion fog computing: Supporting things mobility through container migration at the edge,” in _2018 IEEE International Conference on Smart Computing (SMARTCOMP)_. IEEE, 2018, pp. 97–105. * [38] B. Suleiman and S. Venugopal, “Modeling performance of elasticity rules for cloud-based applications,” in _2013 17th IEEE International Enterprise Distributed Object Computing Conference_. IEEE, 2013, pp. 201–206. * [39] C. Liu, K. Li, J. Liang, and K. Li, “Service reliability in an hc: Considering from the perspective of scheduling with load-dependent machine reliability,” _IEEE Transactions on Reliability_ , vol. 68, no. 2, pp. 476–495, 2019. * [40] F. Al-Haidari, M. Sqalli, and K. Salah, “Impact of cpu utilization thresholds and scaling size on autoscaling cloud resources,” in _2013 IEEE 5th International Conference on Cloud Computing Technology and Science_ , vol. 2. IEEE, 2013, pp. 256–261. * [41] P. Cingolani and J. Alcalá-Fdez, “jfuzzylogic: a java library to design fuzzy logic controllers according to the standard for fuzzy control programming,” _International Journal of Computational Intelligence Systems_ , vol. 6, no. sup1, pp. 61–75, 2013. * [42] B. Javadi, D. Kondo, A. Iosup, and D. Epema, “The failure trace archive: Enabling the comparison of failure measurements and models of distributed systems,” _Journal of Parallel and Distributed Computing_ , vol. 73, no. 8, pp. 1208–1223, 2013. * [43] B. Schroeder and G. Gibson, “A large-scale study of failures in high-performance computing systems,” _IEEE transactions on Dependable and Secure Computing_ , vol. 7, no. 4, pp. 337–350, 2009. * [44] R. K. Naha, S. Garg, A. Chan, and S. K. Battula, “Deadline-based dynamic resource allocation and provisioning algorithms in fog-cloud environment,” _Future Generation Computer Systems_ , vol. 104, pp. 131–141, 2020. * [45] R. K. Naha and S. Garg, “Multi-criteria-based dynamic user behaviour aware resource allocation in fog computing,” _arXiv preprint_ , 2019. * [46] R. N. Calheiros, R. Ranjan, A. Beloglazov, C. A. De Rose, and R. Buyya, “Cloudsim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms,” _Software: Practice and experience_ , vol. 41, no. 1, pp. 23–50, 2011. | Ranesh Kumar Naha Ranesh Kumar Naha completed his Ph.D. studies on reliable resource allocation and scheduling in the Fog computing environment with the University of Tasmania, Australia. He has been awarded a Tasmania Graduate Research Scholarship (TGRS) to support his studies. He received his Master of Science (M.Sc.) degree from Universiti Putra Malaysia, in 2015. He has been awarded a prestigious Commonwealth Scholarship provided by the Ministry of Higher Education, Malaysia. His research interests include wired and wireless networks, parallel and distributed computing, security, Blockchain, Cloud computing, Internet of Things (IoT) and Fog/Edge computing. ---|--- | Dr. Saurabh Garg is currently a Lecturer with the University of Tasmania, Australia. He is one of the few Ph.D. students who completed in less than three years from the University of Melbourne. He has authored over 40 papers in highly cited journals and conferences. During his Ph.D., he has been received various special scholarships for his Ph.D. candidature. His research interests include resource management, scheduling, utility and grid computing, Cloud computing, green computing, wireless networks, and ad hoc networks. ---|--- | Dr. Muhammad Bilal Amin received the M.S. degree from DePaul University, Chicago, IL, USA, in 2006, and the Ph.D. degree from Kyung Hee University, South Korea, in 2015. He is currently a Korea Research Fellow and serving as a Lecturer with the Department of ICT, University of Tasmania, Australia. He has an experience of more than ten years in the software industry, working for Fortune 500 companies in the USA. He is an author of more than 50 publications. His research interests include blockchain, distributed systems, software engineering and architecture, and performance-based cloud applications ---|--- | Prof. Dr. Rajiv Ranjan (Senior Member, IEEE) is currently a chair and professor of computing science and Internet of Things with Newcastle University (from August 2018), United Kingdom. He is an internationally established scientist with more than 300 scientific publications and expertise in cloud computing, big data, and the Internet of Things. He has secured more than $24 Million AUD in the form of competitive research grants from both public and private agencies. He is an innovator with strong and sustained academic and industrial impact and a globally recognized R&D leader with the proven track record. His work has been extensively cited (17K+, Google Scholar; 9K+ Scopus; 6K+ Web of Science). He serves on the editorial boards of top-quality international journals including the IEEE Transactions on Computers, IEEE Transactions on Cloud Computing, IEEE Cloud Computing, and Future Generation Computer Systems. ---|---
* (60) P. De Lange, A. Maloney and E. Verlinde, _Monstrous Product CFTs in the Grand Canonical Ensemble_ , 1807.06200. * (61) R. Dijkgraaf, G. W. Moore, E. P. Verlinde and H. L. Verlinde, _Elliptic Genera of Symmetric Products and Second Quantized Strings_ , _Commun. Math. Phys._ 185 (1997) 197 [hep-th/9608096]. * (62) J. M. Maldacena, G. W. Moore and A. Strominger, _Counting BPS Black Holes in Toroidal Type II String Theory_ , hep-th/9903163. * (63) A. Maloney and E. Witten, _Quantum Gravity Partition Functions in Three Dimensions_ , _JHEP_ 02 (2010) 029 [0712.0155]. * (64) R. Dijkgraaf, J. M. Maldacena, G. W. Moore and E. P. Verlinde, _A Black Hole Farey Tail_ , hep-th/0005003. * (65) J. Manschot and G. W. Moore, _A Modern Farey Tail_ , _Commun. Num. Theor. Phys._ 4 (2010) 103 [0712.0573]. * (66) C. A. Keller and A. Maloney, _Poincare Series, 3D Gravity and CFT Spectroscopy_ , _JHEP_ 02 (2015) 080 [1407.6008]. * (67) J. Maldacena and L. Susskind, _Cool Horizons for Entangled Black Holes_ , _Fortsch. Phys._ 61 (2013) 781 [1306.0533]. * (68) N. Engelhardt, S. Fischetti and A. Maloney, _Free Energy from Replica Wormholes_ , 2007.07444. * (69) N. Seiberg and E. Witten, _The ${\rm D}1$ / ${\rm D}5$ system and singular CFT_, _JHEP_ 04 (1999) 017 [hep-th/9903224]. * (70) R. Gopakumar, _From Free Fields to AdS_ , _Phys. Rev. D_ 70 (2004) 025009 [hep-th/0308184]. * (71) R. Gopakumar, _From Free Fields to AdS. 2._ , _Phys. Rev. D_ 70 (2004) 025010 [hep-th/0402063]. * (72) R. Gopakumar, _From Free Fields to AdS: III_ , _Phys. Rev. D_ 72 (2005) 066008 [hep-th/0504229]. * (73) S. S. Razamat, _On a Worldsheet Dual of the Gaussian Matrix Model_ , _JHEP_ 07 (2008) 026 [0803.2681]. * (74) M. R. Gaberdiel, R. Gopakumar, B. Knighton and P. Maity, _From Symmetric Product CFTs to ${\rm AdS}_{3}$_, 2011.10038. * (75) H. Maxfield and G. J. Turiaci, _The Path Integral of 3D Gravity Near Extremality; Or, Jt Gravity with Defects as a Matrix Integral_ , 2006.11317. * (76) L. Susskind, _Some Speculations About Black Hole Entropy in String Theory_ , hep-th/9309145. * (77) G. T. Horowitz and J. Polchinski, _A Correspondence Principle for Black Holes and Strings_ , _Phys. Rev. D_ 55 (1997) 6189 [hep-th/9612146]. * (78) A. Giveon, D. Kutasov, E. Rabinovici and A. Sever, _Phases of Quantum Gravity in $\mathrm{AdS}_{3}$ and Linear Dilaton Backgrounds_, _Nucl. Phys. B_ 719 (2005) 3 [hep-th/0503121]. * (79) V. Balasubramanian, B. Craps, B. Czech and G. Sárosi, _Echoes of chaos from string theory black holes_ , _JHEP_ 03 (2017) 154 [1612.04334]. * (80) A. Belin, _Permutation Orbifolds and Chaos_ , _JHEP_ 11 (2017) 131 [1705.08451]. * (81) S. R. Coleman, _Black Holes as Red Herrings: Topological Fluctuations and the Loss of Quantum Coherence_ , _Nucl. Phys. B_ 307 (1988) 867. * (82) S. B. Giddings and A. Strominger, _Loss of Incoherence and Determination of Coupling Constants in Quantum Gravity_ , _Nucl. Phys. B_ 307 (1988) 854. * (83) S. B. Giddings and A. Strominger, _Baby Universes, Third Quantization and the Cosmological Constant_ , _Nucl. Phys. B_ 321 (1989) 481. * (84) O. Aharony, S. M. Chester and E. Y. Urbach, _A Derivation of AdS/CFT for Vector Models_ , 2011.06328. * (85) N. Berkovits and C. Vafa, _Towards a Worldsheet Derivation of the Maldacena Conjecture_ , _AIP Conf. Proc._ 1031 (2008) 21 [0711.1799]. * (86) N. Berkovits, _Sketching a Proof of the Maldacena Conjecture at Small Radius_ , _JHEP_ 06 (2019) 111 [1903.08264]. * (87) B. Sundborg, _Stringy gravity, interacting tensionless strings and massless higher spins_ , _Nucl. Phys. Proc. Suppl._ 102 (2001) 113 [hep-th/0103247]. * (88) H. Maxfield, S. Ross and B. Way, _Holographic Partition Functions and Phases for Higher Genus Riemann Surfaces_ , _Class. Quant. Grav._ 33 (2016) 125018 [1601.00980]. * (89) N. Hitchin, _Harmonic spinors_ , _Advances in Mathematics_ 14 (1974) 1 . * (90) L. Bers, _Simultaneous uniformization_ , _Bull. Amer. Math. Soc._ 66 (1960) 94. * (91) D. Gallo, M. Kapovich and A. Marden, _The monodromy groups of schwarzian equations on closed riemann surfaces_ , _Annals of Mathematics_ (2000) 625\. * (92) R. Mandelbaum, _Branched structures and affine and projective bundles on riemann surfaces_ , _Transactions of the American Mathematical Society_ 183 (1973) 37. * (93) A. Gerasimov, A. Morozov, M. Olshanetsky, A. Marshakov and S. L. Shatashvili, _Wess-Zumino-Witten Model as a Theory of Free Fields_ , _Int. J. Mod. Phys. A_ 5 (1990) 2495. * (94) R. Iengo and B. Ivanov, _Fermionization of chiral string determinants in factorizable metrics_ , _Physics Letters B_ 203 (1988) 89 .
# Density functional theory plus dynamical mean field theory within the framework of linear combination of numerical atomic orbitals: Formulation and benchmarks Xin Qu Rocket Force University of Engineering, Xi’an, Shaanxi 710025, China CAS Key Laboratory of Quantum Information, University of Science and Technology of China, Hefei, Anhui 230026, China Peng Xu Rocket Force University of Engineering, Xi’an, Shaanxi 710025, China Rusong Li College of Nuclear Science and Technology, Xi’an Jiaotong University, Xi’an, Shaanxi 710049, China Gang Li<EMAIL_ADDRESS>School of Physical Science and Technology, ShanghaiTech University, Shanghai 201210, China ShanghaiTech Laboratory for Topological Physics, ShanghaiTech University, Shanghai 201210, China Lixin He<EMAIL_ADDRESS>CAS Key Laboratory of Quantum Information, University of Science and Technology of China, Hefei, Anhui 230026, China Xinguo Ren<EMAIL_ADDRESS>Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China Songshan Lake Materials Laboratory, Dongguan 523808, Guangdong, China ###### Abstract The combination of density functional theory with dynamical mean-field theory (DFT+DMFT) has become a powerful first-principles approach to tackle strongly correlated materials in condensed matter physics. The wide use of this approach relies on robust and easy-to-use implementations, and its implementation in various numerical frameworks will increase its applicability on the one hand and help crosscheck the validity of the obtained results on the other. In the work, we develop a formalism within the linear combination of numerical atomic orbital (NAO) basis set framework, which allows for merging NAO-based DFT codes with DMFT quantum impurity solvers. The formalism is implemented by interfacing two NAO-based DFT codes with three DMFT impurity solvers, and its validity is testified by benchmark calculations for a wide range of strongly correlated materials, including 3d transition metal compounds, lanthanides, and actinides. Our work not only enables DFT+DMFT calculations using popular and rapidly developing NAO-based DFT code packages, but also facilitates the combination of more advanced beyond-DFT methodologies available in this codes with the DMFT machinery. ## I INTRODUCTION The understanding of strong correlations among electrons in realistic materials is of great importance for both fundamental science and technological applications. The strong electron correlations trigger a variety of exotic phenomena, which provide a fruitful avenue for designing novel materials [1, 2, 3]. Kohn-Sham (KS) density functional theory (DFT) [4, 5] under local-density approximation (LDA) and generalized gradient approximation (GGA) has achieved remarkable success in describing a very wide range of materials. However, these approximations are found to be inadequate for correctly describing strongly correlated materials, e.g., 3d transitional metal compounds, lanthanides, actinides, quantitatively or even qualitatively [6, 7, 8, 9]. To overcome this limitation, the so-called DFT+DMFT method that merges DFT and many-body technique based dynamical mean-field theory (DMFT) has been developed [10, 11, 12, 13, 7, 14, 15, 8, 9]. Various studies have shown the prominent strength of DFT+DMFT in describing electronic structures of strongly correlated materials. DFT+DMFT has thus becoming a promising method of choice in studying realistic strongly correlated materials with certain predictive power. [16, 17, 18, 19, 8, 9] Comparing to DFT+U [6, 20, 21], a method that shares similar spirit as DFT+DMFT and is available in nearly all popular DFT code packages, DFT+DMFT is not only computationally more expensive, but also relatively more difficult to access for normal users. Implementing DFT+DMFT with well-defined local orbital basis and in a way that doesn’t requires much expertise from users is, thus, of practical importance for promoting the methodology in studying realistic correlated materials. One of the key issues in the implementation of DFT+DMFT scheme is the definition of local correlated orbitals, in terms of which the DMFT impurity problem is defined and solved. Many-body corrections to the KS Hamiltonian exist only in this subspace. As a result, the choice of local correlated orbitals has a noticeable influence on the obtained DFT+DMFT results at the quantitative level. In early days, DFT+DMFT was implemented within the linear muffin-tin orbital (LMTO) basis set framework [14, 15, 16, 22, 23], where the local LMTOs were chosen as the basis orbitals hosting strongly correlated d or f electrons. Later, different Wannier-type orbitals, such as projective Wannier functions [24] and maximally localized Wannier functions, were used to define the local orbitals [25, 26]. As for plane-wave based DFT, Amadon et al. [27, 28] used the all-electron atomic partial waves within the projector augmented wave (PAW) framework or pseudoatomic wave functions as the local correlated orbitals. Within the linearized augmented plane-wave (LAPW) method, Aichhorn et al. [29], and Haule et al. and Kim [30] independently developed projector schemes that convert the KS orbital space to local correlated subspace. We note that, in recent years, the underlying embedding idea of DMFT has been used in broader sense whereby the embedded cluster is solved in a ab initio way using beyond-DFT approaches such as $GW$ or coupled-cluster method, whereas the environment is treated using less advanced approaches [31, 32, 33]. In recent years, numerical atomic orbitals (NAOs) have been employed as the basis set choice in several first-principles software packages [34, 35, 36, 37, 38]. Unlike most other basis sets, the linear combination of NAOs represents a versatile framework that can be used in both all-electron and pseudopotential-based DFT calculations. Past experience indicates that NAOs are an efficient basis set choice not only for conventional ground-state calculations, but also for functionalities that go beyond conventional DFT calculations whereby explicit two-electron Coulomb integrals and/or unoccupied KS states are needed [39, 40, 41, 42]. In this regard, it is highly desirable to develop computational schemes within the NAO basis set framework that enables a convenient merging of first-principles and model-Hamiltonian based approaches. In case of DFT+$U$, several NAO-based implementations have been reported previously [43, 44, 45]. However, in case of DFT+DMFT, only one recent work within the pseudoatomic orbital basis sets (i.e., NAOs tailored for pseudopotential-based calculations) was reported by Sim and Han [46]. In their work, these authors proposed to use the so-called natural atomic orbitals – the eigen-orbitals for the local density matrix – to define the local correlated orbitals. In the present work, we develop a projection scheme that allows to conveniently interface the NAO-based DFT codes and the DMFT impurity solvers and thus enables NAO-based DFT+DMFT calculations. We test the efficacy of such a scheme for the all-electron FHI-aims code [37] and the norm-conserving pseudopotential based ABACUS code [38], interfaced with three different DMFT impurity solver packages. Consistent results are obtained for transition metal compounds, including both correlated metals (SrVO3) and Mott insulators (NiO and MnO), rare-earth systems (Ce metal and Ce2O3), and actinides (Pu2O3 and PuO2). Our work thus extends the reach of the NAO-based numerical framework to treat strongly correlated materials. This paper is organized as follows. Sec. II focuses on the DFT+DMFT formalism. After introducing the general DFT+DMFT formulation, we present our definition of local correlated orbitals within the NAO framework. This is followed by a discussion of the self-consistency scheme of our DFT+DMFT implementation. In Sec. III, we benchmark the validity of our DFT+DMFT formalism and implementation over a wide range of strongly correlated materials, i.e., 3d-systems SrVO3, MnO and NiO, 4f-systems Ce metal and Ce2O3, and 5f-systems Pu2O3 and PuO2. Sec. IV concludes this paper. ## II THEORETICAL FORMULATION ### II.1 General DFT+DMFT formalism Combining DFT with DMFT is not straightforward conceptually as the two theories are formulated in different forms and in different Hilbert spaces. In KS-DFT, one needs to self-consistently solve an effective single-particle problem, whereby (for periodic systems) a KS Hamiltonian is constructed and solved separately at each individual ${\bf k}$ point in the Brillouin zone (BZ). On the other hand, DMFT is a real-space approach developed to solve the lattice Hamiltonian models in which the on-site two-electron interactions are explicitly included. Furthermore, as a first-principles approach, DFT accounts for all the chemical details in the systems and include all (at least valence) electrons in the calculations. In contrast, only the strongly correlated electrons sitting energetically in the vicinity of the Fermi level, relevant for the low-energy physics, are included in the model Hamiltonian and treated explicitly in DMFT. As such, key quantities like charge density, density matrix, and Green functions all have different representations in DFT and DMFT, impeding a straightforward combination of the two approaches. One solution to address this challenge, as detailed below, is to define a suitable projector to convert the key quantities back and forth between the two representations. As a side note, in the following we only focus on the general formalism of DFT+DMFT, without going into a detailed account of DMFT itself in this paper. Interested readers are referred to Refs. [10, 12, 7, 8, 9, 47] for a detailed discussion of this theory. Briefly, in the DFT+DMFT formalism, the Green functions defined in the correlated subspace are the central objects gluing the two theories. There are two kinds of interacting Green functions defined in the corrected subspace: One is the impurity Green function $G^{imp}$ determined by solving the impurity problem in Anderson impurity model, describing a single site with on- site Coulomb interactions and coupled to a mean-field electron bath. The another one, denoted as $G^{loc}$, is the on-site term of the lattice Green function, which is essentially the the Green function of the correlated site described by the Hubbard model and is obtained by projecting the Green function in the full Hilbert space into the correlated subspace. The fundamental requirement in DMFT is that the “on-site” Green function of the lattice problem and that of the Anderson impurity model are equal, $G^{loc}=G^{imp}\,,$ (1) which is achieved via the DMFT self-consistency cycle. In the DFT+DMFT scheme, the single-particle effective Hamiltonian in the full Hilbert space can be expressed as $\hat{H}=\hat{H}_{\mathrm{KS}}+\hat{\Sigma}-\hat{H}_{\mathrm{dc}}\,,$ (2) where $\hat{H}_{\mathrm{KS}}$ is the non-interacting KS Hamiltonian, $\hat{\Sigma}$ the self-energy which encodes all the many-body complexities arising from strong electron correlations and are nonzero only for correlated orbitals. Finally, $\hat{H}_{\mathrm{dc}}$ is the double-counting term, which is introduced to discount the interactions among correlated electrons that have been included in a static mean-field level in $\hat{H}_{\mathrm{KS}}$. Starting from the Hamiltonian Eq. (2), the key issue in the combination of DFT and DMFT is the downfolding (projecting) of the physical quantities from the full KS space to the local correlated space, and the upfolding (embedding) of these quantities from the local space to the full space. Mathematically, the local interacting Green function $G^{loc}$ can be obtained through a projection procedure. In the literature, two different projection procedures have been used. The first one can be seen as a “Hamiltonian projection”, whereby a tight-binding Hamiltonian is first obtained from the KS Hamiltonian by projecting the latter into the correlated subspace, $H_{mm^{\prime}}^{TB}({\bf k})=\langle\phi_{m{\bf k}}|\hat{H}_{\mathrm{KS}}|\phi_{m^{\prime}{\bf k}}\rangle\,.$ (3) Here $\phi_{m{\bf k}}$ is the Bloch-summed local orbitals with $m$ denoting an orbital index in the correlated subspace. In the second step, the local interacting Green function can be directly calculated from this tight-binding Hamiltonian as $\displaystyle G_{mm^{\prime}}^{loc}(i\omega_{n})=$ $\displaystyle\sum_{{\bf k}}\frac{1}{N_{\bf k}}\left[(i\omega_{n}+\mu)I-H^{TB}({\bf k})\right.$ (4) $\displaystyle\left.-\Sigma(i\omega_{n})+H_{\mathrm{dc}}\right]_{mm^{\prime}}^{-1}\,,$ where $N_{\bf k}$ is the is the number of ${\bf k}$ points in the first Brillouin zone, equivalent to the number of unit cells in the Born-von Kármán supercell. In Eq. (4), $\mu$ is the chemical potential, $I$ an identity matrix, and $[\cdots]^{-1}$ denotes the matrix inversion which is taken within the correlated subspace. Here and below, we employ the Matsubara Green function formalism where $\omega_{n}=(2n+1)\pi/\beta$ are the discrete frequency points along the imaginary frequency axis, with $\beta=1/k_{B}T$ being the inverse temperature. The spectroscopy results on the real frequency axis are obtained via analytical continuations. The above procedure to compute the local interacting Green function is the so-called tight-binding Hamiltonian method. In the second scheme, one first constructs the interacting Green function in the full Hilbert space as $G_{ij}({\bf k},i\omega_{n})=\left[(i\omega_{n}+\mu)I-H({\bf k})\right]_{ij}^{-1}\,,$ (5) where the Hamiltonian matrix $H_{ij}({\bf k})=\langle\chi_{i,{\bf k}}\mid\hat{H}\mid\chi_{j,{\bf k}}\rangle$ is the representation of the interacting Hamiltonian operator introduced in Eq. (2) within an arbitrary orthonormal basis set $\mid\chi_{j,{\bf k}}\rangle$ expanding the full Hilbert space. While the matrix elements of the KS Hamiltonian $\hat{H}_{\mathrm{KS}}$ within such a basis set are given straightforwardly, those of $\hat{\Sigma}$ and $\hat{H}_{\mathrm{dc}}$ will be discussed later in Sec. II.3. Once the full-space lattice Green function is obtained, the local interacting Green function can be obtained through a projection procedure $G_{mm^{\prime}}^{loc}(i\omega_{n})=\sum_{{\bf k}}\frac{1}{N_{\bf k}}\sum_{ij}P_{ij}({\bf k},mm^{\prime})G_{ji}({\bf k},i\omega_{n})~{},$ (6) where the projector above is given by $P_{ij}({\bf k},mm^{\prime})=\langle\phi_{m}|\chi_{i{\bf k}}\rangle\langle\chi_{j{\bf k}}|\phi_{m^{\prime}}\rangle\,.$ (7) The construction of the local Green function via Eqs. (5)-(7) is known as the projector method. The two methods discussed above to construct the local Green function differ in that the tight-binding Hamiltonian method projects the Hamiltonian matrix while the projector method projects the Green function. In the projector method, the matrix inversion is carried out in the full Hilbert space, as indicated in Eq. (5), and thus the interactions between the correlated electrons and the rest electrons are retained to some extent. On the other hand, within the tight-binding method, if one wants to describe the interaction between correlated electrons and the rest ones, e.g., studying the charge transfer process between the correlated orbitals and the bath, one has to enlarge the Hilbert space of the tight-binding Hamiltonian to encompass the extra itinerant electrons, and it will inevitably increases the computational complexity. In this work, we choose the projector method in our DFT+DMFT implementation. The other key quantity in DFT+DMFT calculations is the interacting impurity Green function $G^{imp}$ in Eq. (1), corresponding to the local propagator of the effective single-impurity Anderson model, which describes a single site coupled to a bath that mimics the lattice environment at a mean-field level. Formally $G^{imp}$ satisfies the following relationship $\displaystyle G^{imp}_{mm^{\prime}}\left(i\omega_{n}\right)=$ $\displaystyle[(i\omega_{n}+\mu)I-{\cal E}^{imp}-\Delta\left(i\omega_{n}\right)$ (8) $\displaystyle-\Sigma^{imp}\left(i\omega_{n}\right)]_{mm^{\prime}}^{-1}\,,$ where ${\cal E}^{imp}$ is the impurity energy level, $\Delta\left(i\omega_{n}\right)$ the so-called hybridization function characterizing the influence of the environment on the embedded impurity, and $\Sigma^{imp}\left(i\omega_{n}\right)$ the impurity self-energy. It should be noted that ${\cal E}^{imp}$ and $\Delta\left(i\omega_{n}\right)$ are matrices when there are multiple orbitals in the correlated subspace, which is typical in DFT+DMFT calculations. In this context, it is also customary to define a so-called Weiss Green function $\displaystyle\mathcal{G}_{mm^{\prime}}^{-1}(i\omega_{n})=$ $\displaystyle G_{mm^{\prime}}^{-1}(i\omega_{n})+\Sigma_{mm^{\prime}}$ (9) $\displaystyle=$ $\displaystyle[(i\omega_{n}+\mu)I-{\cal E}^{imp}-\Delta\left(i\omega_{n}\right)]_{mm^{\prime}}~{},$ which acts as the dynamical (energy-dependent) mean field that the impurity electrons experience, and encodes essentially the same information as the hybridization function $\Delta\left(i\omega_{n}\right)$. When the self- consistency in the DMFT loop is reached, the local Green function $G_{mm^{\prime}}^{loc}(i\omega_{n})$ and the local self-energy $\Sigma_{mm^{\prime}}(i\omega_{n})$ in Eq. (4) and (6) are equal to the impurity Green function $G^{imp}_{mm^{\prime}}(i\omega_{n})$ and the impurity self-energy $\Sigma^{imp}_{mm^{\prime}}(i\omega_{n})$, respectively. The Weiss Green function together with the local Coulomb interactions defines the Anderson impurity model, which can be expressed in terms of an action [7] $\displaystyle S=$ $\displaystyle\int_{0}^{\beta}d\tau\sum_{mm^{\prime}}c_{m}^{\dagger}(\tau)\mathcal{G}_{mm^{\prime}}^{-1}\left(\tau\right)c_{m^{\prime}}\left(\tau\right)$ (10) $\displaystyle-\sum_{lmno}U_{lmno}\int_{0}^{\beta}d\tau c_{l}^{\dagger}(\tau)c_{n}(\tau)c_{m}^{\dagger}(\tau)c_{o}(\tau)~{},$ where $c_{l}^{\dagger}(\tau)$, $c_{n}(\tau)$, etc., should be understood as the Grassmann variables, and $U_{lmno}$ is the on-site Coulomb interaction expressed within a set of local orbitals (labelled by $l,m,n,o$) spanning the correlated subspace. The action $S$ is essentially the integration of the Lagrangian over the imaginary time. For given $S$, the impurity Green function can be calculated as $G^{imp}_{mm^{\prime}}=-\frac{1}{\mathcal{Z}}\int\mathcal{D}\prod_{i}\left[c^{\dagger},c\right]c_{m}c_{m^{\prime}}^{\dagger}e^{-S}~{},$ (11) where $i$ runs over all $m$ indices, and $\mathcal{Z}$ is the partition function $\mathcal{Z}=\int\mathcal{D}\prod_{i}\left[c^{\dagger},c\right]e^{-S}\,.$ (12) The interacting impurity Green’s function defined via Eqs. (10-12) can then be obtained through a variety of numerical approaches – usually termed as impurity solvers. Up to date, several types of impurity solvers have been developed, including the quantum Monte Carlo (QMC) [13, 48, 49, 50], non- crossing approximation (NCA) [51, 52, 53], one-crossing approximation (OCA) [54, 55], exact diagonalization (ED) [56, 57], numerical renormalization group (NRG) [58, 59], etc. Among all these impurity solvers, continuous-time quantum Monte Carlo (CTQMC) [60, 61, 62, 63] provides access to both high and low energy scales and is effective for a wide class of realistic material calculations. Nowadays, the CTQMC, especially the hybridization expansion based CTQMC (CT-HYB), is the most popular impurity solver employed in DFT+DMFT calculations. Once the interacting impurity Green function is determined, the impurity self- energy $\Sigma^{imp}(i\omega_{n})$ can be obtained via the Dyson equation $\Sigma^{imp}(i\omega_{n})=\left[\mathcal{G}(i\omega_{n})\right]^{-1}-\left[G^{imp}(i\omega_{n})\right]^{-1}$, which is usually done within the impurity solvers. This impurity self-energy will be taken as the updated local self-energy and fed into Eq. (3) and (4) or (2) and (5), from which a new local Green function and consequently a new Weiss Green’s function can be obtained. This is where a new iteration starts. This self-consistency loop keeps going until the self-energy reaches convergence or the local and impurity interacting Green function satisfies the self-consistency condition, i.e., Eq. (1). ### II.2 Construction of the projector Within the NAO basis set framework, it’s natural to take the $d$ or $f$-type NAOs that contribute most to the electronic states around the Fermi level as the the local correlated orbitals. These NAOs by construction are localized and atom-centered, and thus satisfy the usual requirement of correlated orbitals. In early DFT+DMFT implementations, analogous atomic-like orbitals like LMTOs were used. LMTOs are minimal basis sets in the sense that for each angular momentum channel there is only one radial basis function. In contrast with LMTOs, the NAO basis sets are of multi-$\zeta$ character meaning that there are more than one radial functions per angular momentum, thus offering a more accurate description of the electronic structure. In the past, the DFT+U method has been successfully implemented within NAO-based DFT codes [43, 44, 45], whereby it turns out to be a good practice to choose the most localized $d$ or $f$ basis functions as the correlated orbitals to apply the Hubbard $U$ correction. Thus the most localized $d$ or $f$ orbitals seems to span the correlated subspace rather well. In practice, since NAOs centering on neighboring atoms are non-orthogonal to each other, certain orthorgonalization procedure is needed to generate an orthonormal local basis set, which is convenient for DMFT calculations. Below we shall discuss our procedure to construct the projector and local correlated orbitals to facilitate DFT+DMFT calculations within the NAO basis set framework. In analogy to the transformation between Bloch orbitals and Wannier orbitals, we define the following Bloch-summed atomic orbitals as $\Phi_{I,m}^{\bf{k}}({\bf r})=\frac{1}{\sqrt{N_{\bf k}}}\sum_{\mathbf{R}}e^{i\mathbf{k}\cdot\mathbf{R}}\phi_{I,m}\left({\bf r}-{\bm{\tau}}_{I}-\bf{R}\right)\,,$ (13) where $\phi_{I,m}\left({\bf r}-{\bm{\tau}}_{I}-\bf{R}\right)$ is a NAO located at atomic site $I$ in cell ${\bf R}$. Here the magnetic quantum number $m$ labels the different orbitals in the correlated angular moment channel, such as the five $d$ orbitals or seven $f$ orbitals. Since NAOs on neighboring atomic sites are non-orthogonal to each other, it is obvious that the $\Phi_{I,m}^{{\bf k}}({\bf r})$’s defined in Eq. (13) with different $m$ are also non-orthogonal. The key next step is to employ the L$\ddot{\text{o}}$wdin orthonormalization procedure to $\Phi_{I,m}^{\mathbf{k}}$, i.e., $\mid\tilde{\Phi}_{I,m}^{\mathbf{k}}\rangle=\sum_{I^{\prime}m^{\prime}}O^{-\frac{1}{2}}_{Im,I^{\prime}m^{\prime}}\left(\mathbf{k}\right)\mid\Phi_{I^{\prime},m^{\prime}}^{\mathbf{k}}\rangle\,,$ (14) where $O_{Im,I^{\prime}m^{\prime}}\left(\mathbf{k}\right)=\langle\Phi_{I,m}^{\mathbf{k}}\mid\Phi_{I^{\prime},m^{\prime}}^{\mathbf{k}}\rangle$ (15) is the overlap matrix. The newly obtained $\tilde{\Phi}_{I,m}^{\mathbf{k}}$ orbitals are then orthonormal by construction. Afterwards, a Fourier transform is applied to get the correlated orbital in real space, i.e., $\mid W_{I,m}^{{\bf R}}\rangle=\frac{1}{\sqrt{N_{\bf k}}}\sum_{\mathbf{k}}e^{-i{\bf k}\cdot\mathbf{R}}\mid\tilde{\Phi}_{I,m}^{\mathbf{k}}\rangle~{}.$ (16) The orthonormality of $W_{I,m}^{{\bf R}}$ is also guaranteed by construction. We choose the KS states $|\Psi_{i{\bf k}}\rangle$ as the basis sets $|\chi_{i{\bf k}}\rangle$ (cf. Eq. (7)) to span the full Hilbert space, and then Eq. (6) becomes $\displaystyle G_{I,mm^{\prime}}^{loc}=$ $\displaystyle\sum_{\mathbf{k}}\frac{1}{N_{\bf k}}\sum_{ij}P^{I}_{ij}(\mathbf{k},mm^{\prime})$ (17) $\displaystyle\left[\frac{1}{i\omega_{n}+\mu-\epsilon({\mathbf{k}})-\bar{\Sigma}(\mathbf{k},i\omega_{n})}\right]_{ji},$ where $\epsilon_{ji}({\mathbf{k}})=\langle\Psi_{j\mathbf{k}}|\hat{H}_{\textrm{KS}}|\Psi_{i\mathbf{k}}\rangle=\epsilon_{i{\bf k}}\delta_{ij}$ and $\bar{\Sigma}_{ji}(\mathbf{k},i\omega_{n})=\langle\Psi_{j\mathbf{k}}|\hat{\Sigma}(i\omega_{n})-\hat{H}_{\mathrm{dc}}|\Psi_{i\mathbf{k}}\rangle$, with $\epsilon_{i{\bf k}}$ being the KS eigenvalues. The projector, Eq. (7), then becomes $P_{ij}^{I}\left({\bf k},mm^{\prime}\right)=\langle\Psi_{i\mathbf{k}}\mid W_{I,m}^{0}\rangle\langle W_{I,m^{\prime}}^{0}\mid\Psi_{j\mathbf{k}}\rangle\,,$ (18) where the superscript $0$ denotes the central unit cell. Since in DMFT calculations, only the “on-site” Green function, where $m$ and $m^{\prime}$ orbitals are located in the same unit cell is needed, the projector is designed to project the full Green function into the central unit cell, without losing generality. Formally, the projector for a given correlated atom $I$ and a wave vector ${\bf k}$ is a fourth-order tensor, but since it is separable and symmetric, only a second-order tensor, i.e., the overlap matrix $\langle\Psi_{i\mathbf{k}}\mid W_{I,m}^{0}\rangle$ needs to be computed and stored in practical implementations. The whole DFT+DMFT scheme requires the orthonormality of local orbitals representing the correlated subspace. In the language of the projector, it requires the projector to satisfy the following orthonormal condition $\displaystyle\sum_{i}P_{ii}^{I}\left({\bf k},mm^{\prime}\right)=$ $\displaystyle\sum_{i}\langle W_{I,m^{\prime}}^{0}\mid\Psi_{i\mathbf{k}}\rangle\langle\Psi_{i\mathbf{k}}\mid W_{I,m}^{0}\rangle$ (19) $\displaystyle=$ $\displaystyle\delta_{mm^{\prime}}.$ In principle, this condition is automatically satisfied if the summation over $i$ goes over all the KS bands. In practical DFT+DMFT implementations, however, one truncates the full KS states into a small subset around the Fermi level, which means $i$ just runs over bands that are located in a chosen energy window around the Fermi level (in the following, these subsets of bands are denoted as $\mathcal{C}$). This truncation destroys the completeness of $|\Psi_{ik}\rangle$ and thus the orthonormality of the projector. To deal with this issue, one can introduce an extra transformation $\displaystyle\tilde{P}_{ij}^{I}\left(\mathbf{k},mm^{\prime}\right)=$ $\displaystyle\sum_{m^{\prime\prime\prime}}\tilde{O}_{m^{\prime}m^{\prime\prime\prime}}^{-\frac{1}{2}}(\mathbf{k})\langle W_{I,m^{\prime\prime\prime}}^{0}\mid\Psi_{j\mathbf{k}}\rangle$ (20) $\displaystyle\sum_{m^{\prime\prime}}\langle\Psi_{i\mathbf{k}}\mid W_{I,m^{\prime\prime}}^{0}\rangle\tilde{O}_{m^{\prime\prime}m}^{-\frac{1}{2}}(\mathbf{k})$ to orthonormalize the projector. The transformation matrix in the above equation is given by $\displaystyle\tilde{O}_{mm^{\prime}}(\mathbf{k})$ $\displaystyle=\sum_{i\in\mathcal{C}}P_{ii}^{I}\left(\mathbf{k},mm^{\prime}\right)$ $\displaystyle=\sum_{i\in\mathcal{C}}\langle W_{I,m}^{0}|\Psi_{i\mathbf{k}}\rangle\langle\Psi_{i\mathbf{k}}|W_{I,m^{\prime}}^{0}\rangle\,,$ (21) which is nothing but the overlap between the projections of the orthonormalized local orbitals $|W_{I,m}^{0}\rangle$’s within the subspace $\mathcal{C}$. Mathematically, the local correlated orbitals we use above to construct the projector can be explicitly expressed as $\displaystyle\mid\tilde{W}_{I,m}^{{\bf R}}\rangle=$ $\displaystyle\frac{1}{\sqrt{N_{\mathbf{k}}}}\sum_{\mathbf{k}}e^{-i\mathbf{k}{{\bf R}}}\sum_{m^{\prime}}\tilde{O}_{mm^{\prime}}^{-\frac{1}{2}}(\mathbf{k})$ (22) $\displaystyle\sum_{i\in\mathcal{C}}\langle\Psi_{i\mathbf{k}}|\tilde{\Phi}_{I,m^{\prime}}^{\mathbf{k}}\rangle|\Psi_{i\mathbf{k}}\rangle.$ In this form, our scheme is similar in spirit to the projective Wannier- orbital scheme proposed by Anisimov et al. [24] in the context of LDA+DMFT calculations. In our case, the most localized $d$ or $f$ NAO plays the role of the LMTO in the work of Ref. [24] There are a few advantages of using NAOs to define the local correlated space. Firstly, this choice is physically intuitive and technologically straightforward within the NAO basis set framework. We do not need to spend extra efforts to generate a set of local orbitals and make sure they are physically reasonable atomic-like orbitals. Secondly, from both the theoretical and technical perspectives, the choice of NAOs to define the local correlated space and the resulting projection scheme are suitable for all NAO- based packages. Especially, the key quantities that are required in this formalism, e.g., the KS wave functions and the overlap matrix of the basis functions, exist naturally in NAO-based DFT code packages, and hence no additional efforts are required to calculate these quantities. Thirdly, due to its high flexibility, our DFT+DMFT infrastructure can be interfaced with a new NAO-based DFT code without much effort. We hope it can serve as a platform to enable NAO-based DFT codes to do DFT+DMFT researches on strongly correlated materials. In this work, we implement the DFT+DMFT interface and test it with two NAO-based DFT codes using different techniques, i.e., the pseudopotential- based ABACUS code [38] and full-potential all-electron FHI-aims code [37]. ### II.3 DFT+DMFT self-consistency scheme In this section, we will explain step by step our DFT+DMFT calculation procedure, according to the workflow outlined in the flow diagram depicted in Fig. 1. Step 1\. The DFT+DMFT calculation starts from well-converged DFT band structures. To get high-quality KS orbitals $|\Psi_{i\bf{k}}\rangle$, a dense ${\bf k}$-point mesh is usually needed. Step 2\. With $|\Psi_{i\bf{k}}\rangle$, the projector defined in Eq. (20) can be straightforwardly constructed. To this end, the most localized $d$ or $f$ orbital of the correlated atoms in the NAO basis set is used to construct $|W_{I,m}^{0}\rangle$. Step 3\. In the DFT+DMFT iteration loop, the frequency-dependent self-energy $\Sigma(i\omega_{n})$ is determined by the impurity solver at each iteration step. To start with, the initial self-energy is set to be equal to the double- counting term, i.e., $\bar{\Sigma}_{ij}(\mathbf{k},i\omega_{n})=0$. Here the following choice of the double-counting term $H_{\mathrm{dc},mm^{\prime}}^{I,\sigma}=\left[U(n_{I}-1/2)-1/2J(n_{I}-1)\right]\delta_{mm^{\prime}}$ (23) is used. Here $n_{I}$ is the total number of correlated electrons associated with the correlated atom $I$ and is fixed during the DMFT cycles. In the spirit of reducing the necessity of introducing additional empirical parameters, $n_{I}$ is given by projecting the KS orbitals in the subset $\mathcal{C}$ to the local subspace as $n_{I}=\sum_{m}\sum_{\mathbf{k}}\frac{1}{N_{\mathbf{k}}}\sum_{i\in\mathbf{C}}f_{i\mathbf{k}}\tilde{P}_{ii}^{I}\left(\mathbf{k},mm\right)\,,$ (24) where $f_{i\mathbf{k}}$ is the occupation number of KS orbital $\Psi_{i\mathbf{k}}$. This double-counting scheme is similar to the so-called fixed double-counting [22, 64] scheme, which is considered to be able to improve the stability of DFT+DMFT self-consistency loop [64] by fixing the value of $n_{I}$. The difference is that the nominal number of strongly correlated electrons is specified manually in the fixed double-counting scheme, whereas in our case this number is determined using Eq. (24). Step 4\. Using the projector constructed in Step 2, we embed (upfold) the self-energy back to the selected KS space subset $\mathcal{C}$, which is expressed as $\displaystyle\bar{\Sigma}_{ij}(\mathbf{k},i\omega_{n})=$ $\displaystyle\sum_{I}\sum_{mm^{\prime}}\tilde{P}_{ij}^{I}\left(\mathbf{k},mm^{\prime}\right)$ (25) $\displaystyle\left(\Sigma_{mm^{\prime}}^{I}(i\omega_{n})-H_{\mathrm{dc},mm^{\prime}}^{I}\right).$ Step 5\. During the DFT+DMFT self-consistency iteration, the electronic chemical potential needs to be adjusted according to the newly obtained self- energy at each iteration, to keep the number of electrons hosted by KS bands in $\mathcal{C}$ $N_{\mathcal{C}}^{\mathrm{KS}}=\sum_{\mathbf{k},i\in\mathcal{C}}\frac{1}{N_{\mathbf{k}}}f_{i\mathbf{k}}$ (26) conserved. Within the DMFT cycle, this condition means that $\displaystyle N_{\mathcal{C}}^{\mathrm{KS}}=$ $\displaystyle\frac{1}{\beta}\sum_{\omega_{n}}\sum_{\mathbf{k},i\in\mathcal{C}}\frac{1}{N_{\mathbf{k}}}$ (27) $\displaystyle\left[\frac{1}{i\omega_{n}+\mu-\epsilon_{\mathbf{k}}-\bar{\Sigma}(\mathbf{k},i\omega_{n})}\right]_{ii}\,,$ where $\beta$ is again the inverse temperature $1/{k_{B}T}$. The summation of imaginary frequency $\omega_{n}$ should run from negative infinity to positive infinity. However, in realistic calculations, to save computational cost, the explicit summation over Matsubara frequency points is only carried within a frequency window $[-\omega_{N},\omega_{N}]$, where the contribution from the frequency points outside the window is treated approximately. This is enabled by making use of the asymptotic behavior of the self-energy, i.e., $\lim_{n\to\infty}\bar{\Sigma}_{ii}(\mathbf{k},i\omega_{n})=\bar{\Sigma}_{ii}(\mathbf{k},\infty)$, where $\bar{\Sigma}_{ii}(\mathbf{k},\infty)$ is a real value. Then the Eq. (27) is approximated by $\displaystyle N_{\mathcal{C}}^{\mathrm{KS}}=$ $\displaystyle\frac{1}{\beta}\sum_{\mathbf{k},i\in\mathcal{C}}\frac{1}{N_{\mathbf{k}}}\left\\{\sum_{\omega_{n}=-\omega_{N}}^{\omega_{N}}\right.$ (28) $\displaystyle\left(\left[\frac{1}{i\omega_{n}+\mu-\epsilon_{\mathbf{k}}-\bar{\Sigma}(\mathbf{k},i\omega_{n})}\right]_{ii}\right.$ $\displaystyle\left.-\left[\frac{1}{i\omega_{n}+\mu-\epsilon_{k}-\bar{\Sigma}(\mathbf{k},\infty)}\right]_{ii}\right)$ $\displaystyle\left.+\frac{1}{1+e^{\beta(\epsilon_{i\mathbf{k}}+\bar{\Sigma}_{ii}(\mathbf{k},\infty)-\mu)}}\right\\}.$ When the chosen cutoff Matsubara frequency $\omega_{N}$ is high enough so that the asymptotic behavior of the self-energy is correct, this approximation is of good accuracy. Step 6\. As all information is secured, the local interacting Green function is constructed from Eq. (17), where, of course, the renormalized projector Eq. (20) is used. Under the DFT+DMFT self-consistency condition, Eq. (1) and (8), the matrices of the impurity level and hybridization function are determined by ${\cal E}_{I,mm^{\prime}}^{imp}=-H_{dc,mm^{\prime}}^{I}+\sum_{\mathbf{k},i\in\mathcal{C}}\frac{1}{N_{\mathbf{k}}}\tilde{P}_{ij}^{I}\left(\mathbf{k},mm^{\prime}\right)\varepsilon_{i\mathbf{k}}$ (29) and $\displaystyle\Delta_{mm\prime}^{I}\left(i\omega_{n}\right)=$ $\displaystyle(i\omega_{n}+\mu)\delta_{mm^{\prime}}-{\cal E}^{imp}_{I,mm^{\prime}}$ (30) $\displaystyle-\Sigma_{mm^{\prime}}^{I}\left(i\omega_{n}\right)-[G^{loc}]^{-1}_{I,mm^{\prime}}.$ For some CT-HYB impurity solver, the imaginary time hybridization function is needed. Figure 1: Flow diagram of the major steps in our DFT+DMFT implementation. Step 7\. Solve the impurity problem with the determined $\mu$, ${\cal E}_{I}^{imp}$, $\Delta^{I}\left(i\omega_{n}\right)$ and the given on-site Coulomb interaction through the impurity solver to obtain the new self-energy and impurity Green function. In this paper, we use the Kanamori [65, 66] form Coulomb interaction in which only the density-density term is included, so there are no sign problem in our CTQMC calculations. Step 8\. Check whether the self-consistency is reached. If the self- consistency is reached, stop the DFT+DMFT calcution, else return to Step 4. ## III Results and discussion ### III.1 d-electron systems We first benchmark our DFT+DMFT implementation on three prototypical strongly correlated d-electron systems – SrVO3, NiO, and MnO. For the DFT part, we carried out GGA calculations using two NAO-based code packages – FHI-aims [37] and ABACUS [38]. In FHI-aims, the default tight tier 1 basis set is used for V, Ni, Mn, O, and Sr atoms, and the corresponding cutoff radii of the basis functions are $6.0\text{\,}\mathrm{\SIUnitSymbolAngstrom}$, $6.0\text{\,}\mathrm{\SIUnitSymbolAngstrom}$, $6.0\text{\,}\mathrm{\SIUnitSymbolAngstrom}$, $6.0\text{\,}\mathrm{\SIUnitSymbolAngstrom}$, and $8.0\text{\,}\mathrm{\SIUnitSymbolAngstrom}$, respectively. In ABACUS we use the SG15 optimized norm-conserving Vanderbilt (ONCV) multi-projector pseudo- potentials [67, 68, 69] and the corresponding optimized double-$\zeta$ plus polarization (DZP) atomic basis sets, which comprise $4s2p2d1f$ basis functions with a cutoff radius of $9.0\text{\,}\mathrm{B}\mathrm{o}\mathrm{h}\mathrm{r}$ for transition metal atoms and $2s2p1d$ basis functions with cutoff radii of $7.0\text{\,}\mathrm{B}\mathrm{o}\mathrm{h}\mathrm{r}$ for O atoms and $10.0\text{\,}\mathrm{B}\mathrm{o}\mathrm{h}\mathrm{r}$ for Sr atoms. In all DFT calculations, the Perdew-Burke-Ernzerhof (PBE) GGA exchange-correlation functional was used [70]. As for the DMFT part, we employed three different CT-HYB impurity solvers to solve the single-site impurity problem. These are PACS 111The segment implementation [60] of the ct-qmc impurity solver is a part of the PACS@sdf package (Package for Analyzing Correlated Systems with Spatial and Dynamical fluctuations). PACS@sdf aims at providing an integrated framework for the study of strongly correlated models and materials beyond the local approximation of the DMFT [7]. It takes DMFT as the zero-order approximation and systematically provides non-local corrections to it [Li2015]. developed by one of the present authors Gang Li, iQIST developed by Huang et al. [72, 73, 74], and the one used in eDMFT [30] developed by Haule at Rutgers University [62, 75], which is referred to as “Rutgers” below in this paper. #### III.1.1 SrVO3 SrVO3 has a simple cubic perovskite structure without magnetism. DFT fails to reproduce the upper and lower Hubbard bands observed in experiments, and this calls for advanced theoretical and computational techniques [8]. In the past, this material has been extensively studied both theoretically and experimentally [25, 76, 77, 78, 79, 80, 81, 82], which provides abundant reference results. Therefore, SrVO3 is an ideal test example for DFT+DMFT implementations [24, 26, 27, 66, 46]. The V4+ cation with only one 3d electron is located in the center of the octahedron formed by its six surrouding axial O2- ligand anions. In the presence of an octahedral crystal field, the five degenerate d orbitals split into two subsets: Three-fold degenerate t2g orbitals (i.e., $d_{xy}$, $d_{yz}$ and $d_{xz}$), and two-fold degenerate eg orbitals (i.e., $d_{z^{2}}$ and $d_{x^{2}-y^{2}}$). The only one 3d electron of V4+ occupies the lower-energy t2g orbitals with the higher-energy eg orbitals being empty. As a common practice, we only consider three degenerate t2g orbitals in our DFT+DMFT calculation. In DFT calculations, we use 11$\times$11$\times$11 ${\bf k}$-point mesh generated by the Monkhorst-Pack method [83]. The Hubbard U and Hund J parameters are set to be $4.0\text{\,}\mathrm{eV}$ and $0.65\text{\,}\mathrm{eV}$, respectively, following the choice in the literature [26, 27, 46]. The DMFT calculation is carried out at a temperature of $300\text{\,}\mathrm{K}$. DFT calculations give a group of bands around the Fermi level with substantial 3d characterizes, which are well separated from other bands [26, 27]. We enclose the six (twelve if spin degree of freedom is taken into account) bands crossing the Fermi level in the subset of KS bands $\mathcal{C}$. Figure 2: The single-particle energy spectral function of SrVO3 3d electrons given from direct analytical continuation of the impurity Green function. The results presented in the six panels are obtained by six computational schemes combining two DFT codes – FHI-aims (upper panels) and ABACUS (lower panels) and three different impurity solvers – PACS (left panels), iQIST (middle panels), and Rutgers (right panels). One of the powerful strengths of Green-function based DFT+DMFT approach is that it can deliver physically meaningful single particle excitation spectral functions, as given by the the imaginary part of Green function. Such spectral functions can be directly measured by photoemission and/or inverse photoemission experiments. In Fig. 2, the spectral functions of 3d electrons of SrVO3 calculated by combining two NAO-based DFT packages – FHI-aims and ABACUS, and three impurity solvers – PACS, iQIST and Rutgers are presented. Despite the considerable differences underlying the implementations of the DFT codes, as well as the DMFT impurity solvers, the calculated spectral functions are remarkably similar. Most importantly, they all successfully reproduce the typical three-peak structure: A quasi-particle band located at the Fermi level with the lower and upper Hubbard bands on the two sides, arising from the strong electron correlations which traditional LDA and GGA fail to capture. The main features of the spectral functions given by our DFT+DMFT calculations are in good agreement with previous theoretical results except for some small details such as the exact peak positions and intensities [78, 25, 79, 24, 76, 26, 27, 66, 46]. Our results also correctly describe the strong particle-hole asymmetry of the lower and upper Hubbard band as the intensity of the occupied lower Hubbard band is much smaller than the empty upper Hubbard band. This is the consequence that the V4+, with only one 3d electron, is far away from half-filling. Similar to previous DFT+DMFT studies [25, 79, 24, 76, 26, 27, 66, 46], our spectral function results are in excellent agreement with the experimental photoemission spectrum [78]. The first main feature lies in the reproduction of the lower Hubbard band at around $-2.0\text{\,}\mathrm{eV}$, which is the manifestation of the strong correlation among 3d electrons. The second feature, in quantitative agreement with experiment, is that the lower Hubbard band vanishes at $\sim$-1.0\text{\,}\mathrm{eV}$$, and then the quasiparticle band starts to rise sharply. Figure 3: The spectral function of MnO 3d electrons obtained from direct analytical continuation of the impurity Green function. Similar to Fig. 2, results presented in the six panels are obtained by two different DFT codes combined with three different impurity solvers. Each panel is labelled by the specific combination. #### III.1.2 MnO and NiO Next we test our DFT+DMFT scheme on the late transition metal oxides MnO and NiO. Both MnO and NiO crystallize in face-centered cubic (FCC) NaCl-type structures where 3d orbitals are split into threefold degenerate $t_{2g}$ and twofold degenerate $e_{g}$ orbitals as in the SrVO3 case. In the MnO system, we adopt the parameters of U$=$$5.0\text{\,}\mathrm{eV}$, J$=$$1.0\text{\,}\mathrm{eV}$, and the temperature $T=$$300\text{\,}\mathrm{K}$. For MnO, the KS bands within the energy window of [$-3.0\text{\,}\mathrm{eV}$, $3.0\text{\,}\mathrm{eV}$] are included in the subset $\mathcal{C}$. For NiO, another set of parameters with the temperature of $1160\text{\,}\mathrm{K}$, U of $8.0\text{\,}\mathrm{eV}$ and J of $1.0\text{\,}\mathrm{eV}$ [6, 84, 85, 46] is used. The energy window enclosing the subset of KS bands $\mathcal{C}$ is chosen to be [$-2.0\text{\,}\mathrm{eV}$, $1.0\text{\,}\mathrm{eV}$]. Figure 4: The single-particle spectral function of NiO 3d electron calculated from direct analytical continuation of the impurity Green function. The meaning of the six panels is same as Fig. 2 and 3. Figure 5: The ${\bf k}$-resolved spectral function $A(\mathbf{k},\omega)$ of MnO, obtained by (a) FHI-aims+Rutgers and (b) ABACUS+Rutgers, respectively. Figure 6: The k-resolved spectral function $A(\mathbf{k},\omega)$ of NiO, obtained by (a) FHI-aims+Rutgers and (b) ABACUS+Rutgers, respectively. MnO and NiO are classical examples that clearly demonstrate the failure of traditional band theory, e.g., LDA and GGA. LDA and GGA predict MnO and NiO to be metallic while they are wide-gap insulators experimentally. Even though a gap will appear if the spin-symmetry-broken antiferromagnetic state is considered, the calculated gap is however still one order of magnitude smaller than the experimental values. By merging the Hubbard model and ab-initio DFT, static DFT+U [6, 20, 21] and the more advanced dynamical DFT+DMFT [84, 85] successfully open the gaps in those prototypical transitional metal oxides. Then obtained band gaps are in quantitative agreement with experiments, if proper parameters are used. We then apply our DFT+DMFT implementation to both MnO and NiO, and the obtained spectral functions are presented in Fig. 3 and Fig. 4, respectively. Again in each figure, we present six panels containing results obtained by combining two NAO-based DFT codes and three DMFT impurity solvers. The most prominent feature in the DFT+DMFT spectra of both MnO and NiO is that a sizeable gap is opened up, arising apparently from the strong Coulomb interactions whose effect is properly captured by DMFT. One can further see from Fig. 3 that the intensity of the lower and upper Hubbard band of Mn 3d electrons are nearly the same. This result is consistent with the nominally half-filling 3d orbitals of Mn given by chemical analysis, which is confirmed by the fact that the number of 3d electrons of Mn given by $n=-\sum_{m}G^{imp}_{mm}(\beta)$ is 5. Similar to the SrVO3 case, the results presented in all six panels are very close, except that the peaks given by the Rutgers impurity solver is slightly more pronounced. For NiO, the number of nominal $3d$ electrons is approximately 8, and DFT+DMFT calculations lead to a splitting of the $3d$ bands, with fully occupied $t_{2g}$ orbitals and half-filled $e_{g}$ orbitals. Our results in general agree well with previous DFT+DMFT results reported in the literature [84, 85]. However, Fig. 4 also reveals that both the DFT codes and the impurity solvers have certain influence on the obtained DFT+DMFT energy spectra. Although there are no qualitative differences, the shape and width of the left peak, and the depth of the dip between the first and middle peaks show noticeable quantitative differences. It is not entirely clear to us yet what factors caused such differences. We note that, however, NiO is a prototypical charge transfer insulator [86], and the competition between the strong Coulomb interactions among the Ni $3d$ electrons and the hybridization between the Ni $3d$ and O $2p$ electrons governs its underlying physics. Thus a complete DFT+DMFT treatment of NiO should also include O $2p$ electrons in the game. With the O $2p$ excluded, the DFT+DMFT calculations are probably more sensitive to the details in the band structures and numerical techniques behind the impurity solvers. Figure 3 and 4 only contain the spectral information of 3d electrons. Our DFT+DMFT implementation also allows for calculating the ${\bf k}$-resolved spectral function in the KS orbital space, as determined by $A(\mathbf{k},\omega)=-\frac{1}{\pi}\sum_{i\in\mathcal{C}}\operatorname{Im}\left[(\omega+\mu-\epsilon({\mathbf{k}}))I-\bar{\Sigma}(\mathbf{k},\omega)\right]^{-1}_{ii},$ (31) where the real-frequency self-energy is evaluated by analytical continuation of the imaginary frequency-self-energy through the maximum entropy formalism in Ref. [87]. The corresponding results for MnO and NiO are presented in Figs. 5 and 6, respectively. Now for simplicity only results obtained using the two DFT codes combined with the Rutgers impurity solver are presented. The agreement between our theoretical gaps and experimental gaps [88, 89, 90] is rather satisfactory. The two ${\bf k}$-resolved spectral functions of MnO is nearly identical whereas again noticeable differences, in particular regarding the spectral weights of certain bands, exist for NiO. Nevertheless the main features, e.g., the energy positions and dispersion of the bands, are reasonably similar in Fig. 6(a) and 6(b). NiO has been extensively studied experimentally, and reliable experimental photoemission spectra are available. In Fig. 7, we directly compare our DFT+DMFT spectrum of NiO with the experimental spectrum of Sawatzky and Allen [89] measured by x-ray-photoemission (XPS) and bremsstrahlung-isochromat- spectroscopy (BIS) techniques. Again, the two sets of theoretical spectra are obtained using FHI-aims and ABACUS codes interfaced with the Rutgers impurity solver, respectively. For the spectrum below the Fermi level, corresponding to the (negative of) energy cost for removing a particle from the occupied levels, we reproduce the sharp peak around $-2.0\text{\,}\mathrm{eV}$ and a local minimum at $-3.0\text{\,}\mathrm{eV}$ given by XPS. For the spectrum above the Fermi level, corresponding to the (negative of) energy cost for adding a particle to the system, our curves match perfectly with the BIS result. Although small deviations of the shoulder peak at approximately $-3.5\text{\,}\mathrm{eV}$ are visible, our DFT+DMFT spectrum, in general, agrees well with previous theoretical work [84] and the experiment data[89]. Figure 7: Comparison of theoretical and experimental spectral functions of NiO. The theoretical results are given by the imaginary part of the impurity Green function on the real frequency axis, analytically continued from the Matsubara impurity Green function. The experimental data are obtained from XPS+BIS measurements, taken from Ref. [89]. In ABACUS, the pseudopotential method is used and only the valence electrons are explicitly included. Furthermore, as given in the beginning of this section, the spatial cutoffs of NAOs used in ABACUS are much smaller than those used in FHI-aims. By contrast, in FHI-aims full potential is used with all electrons being included in DFT calculations. Although these two NAO schemes are rather different, we get nearly identical DFT+DMFT results of SrVO3 and MnO and reasonably similar results for NiO. It proves the efficacy and robustness of our DFT+DMFT formalism and implementation. We expect that the DFT+DMFT formalism presented in this paper should work for other NAO-based DFT codes as well. ### III.2 f-electron systems The f-electron systems, including 4f lanthanides and 5f actinides, are an important class of strongly correlated materials characterized by partially filled f-type orbitals. These systems exhibit a variety of exotic phenomena, such as the heavy-fermion behavior, metal-insulator transition, Kondo physics, volume collapses accompanying phase transitions, etc. It has been generally accepted that traditional DFT calculations based on static mean-field-type approximations such as LDA and GGAs do not provide adequate accuracy for describing these physical scenarios. DFT+DMFT has proved to be a powerful approach to tackle these systems and achieved remarkable successes in the last two decades. In this section we test our DFT+DMFT implementation on several typical f-electron systems, including 4f systems like $\alpha$ and $\gamma$ cerium (Ce) metal and Ce2O3, and 5f systems like PuO2 and Pu2O3. DFT calculations are carried out using FHI-aims with the default tight tier1 basis sets. For these basis sets, the cutoff radii of Ce, Pu, and O elements are $7\text{\,}\mathrm{\SIUnitSymbolAngstrom}$, $6.0\text{\,}\mathrm{\SIUnitSymbolAngstrom}$, and $6.0\text{\,}\mathrm{\SIUnitSymbolAngstrom}$, respectively. Analogous to the case of $3d$-electron systems, the GGA-PBE functional [70] is used in the DFT calculations and 11$\times$11$\times$11 Monkhorst-Pack ${\bf k}$-point mesh [83] is used for Brillouin-zone integration, which is deemed dense enough to obtain accurate DFT band structures. Unlike the practice in previous sections where results obtained using three DMFT impurity solvers are shown for comparison, here for simplicity only the results obtained using the Rutgers solver will be presented. #### III.2.1 Ce metal Despite its simple FCC structure and the fact that only one 4f valence electron is present per Ce atom, Ce exhibits spectacular physical properties, which has attracted considerable research interest over the last forty decades [91, 92, 93, 94, 95, 96, 97, 66, 98, 99]. At low temperature and ambient pressure, Ce crystalizes in the $\alpha$ phase (smaller volume), where the system is paramagnetic and shows Pauli-like magnetic susceptibility without forming local magnetic moments. At high temperature, Ce transforms into the $\gamma$ phase (larger volume), which instead carries local magnetic moments and exhibits Curie-Weiss behavior of magnetic susceptibility. When increasing pressure or decreasing temperature, Ce undergoes the famous isostructural $\gamma$-$\alpha$ phase transition accompanied by a 14% volume collapse and a drastic change of magnetic properties [93, 94, 95, 96, 97, 66, 98, 99]. A theoretical understanding of this behavior poses a great challenge to condensed matter physics and has motivated lots of experimental and theoretical investigations. From the perspective of first-principles computations, standard DFT calculations are unable to give a proper description of this phase transition, and the approaches going beyond conventional DFT are required. In this paper, the crystal structures of $\alpha$ and $\gamma$ Ce phases are discerned by different volumes, i.e., $29\text{\,}\mathrm{\SIUnitSymbolAngstrom}$3 for $\alpha$ and $34\text{\,}\mathrm{\SIUnitSymbolAngstrom}$3 for $\gamma$ phase, respectively, which covers the volume ranges during the $\alpha$-$\gamma$ transition [94]. The chosen energy window for KS subset $\mathcal{C}$ is [$-5.0\text{\,}\mathrm{eV}$, $5.0\text{\,}\mathrm{eV}$]. The Coulomb interaction parameter U given by constrained DFT is $6.0\text{\,}\mathrm{eV}$ [66, 98]. Here we used the value of U of $6.0\text{\,}\mathrm{eV}$ and Hund parameter J of $0.5\text{\,}\mathrm{eV}$ [96, 98]. DFT+DMFT calculations are done at the temperature of $800\text{\,}\mathrm{K}$. Figure 8: Comparison of theoretical and experimental spectral functions of $\alpha$ and $\gamma$ Ce. The DFT+DMFT calculation is carried out by FHI-aims interfaced with the Rutgers impurity solver and the corresponding spectral results are given by direct analytical continuation of the impurity Green function. The resonant inverse photoemission spectroscopies (RIPES) are taken from Ref. [91], and the photoemission (PE) spectra are taken from Ref. [92]. Our DFT+DMFT spectra for Ce are shown in Fig. 8. First of all, the energy spectra of both phases show the characteristic three-peak structures of strongly correlated metals. One can also see that, going from the $\alpha$ phase to the $\gamma$ phase, the central quasiparticle peak gets substantially suppressed, and its spectral weights transfer to the lower and upper Hubbard bands . Such results are in agreement with experimental observations [91, 92] and previous DFT+DMFT calculations for Ce [94, 95, 96, 98]. The relative strengths of the quasi-particle peak and the upper Hubbard peak of both $\alpha$ and $\gamma$ phases are also captured qualitatively. However, a direct comparison of the calculated spectra with the experimental spectra reveals that the energy separation between the quasi-particle peak and upper Hubbard band is underestimated in our calculations, while that between the lower Hubbard band and quasi-particle peak is overestimated. Compared to previous DFT+DMFT calculations, the location of the lower Hubbard peak of our results is close to that reported in the literature [94, 98] where the maximum of lower Hubbard peaks is around -4.0$\sim$$-3.0\text{\,}\mathrm{eV}$. Concerning the energy positions of the upper Hubbard band, we found that the results reported in the literature also vary quite a bit, with those reported in Ref. [98] agreeing best with experiment. Our results are close to those reported in Ref. [96], where the upper Hubbard peaks are located at around $2.5\text{\,}\mathrm{eV}$. In general, our DFT+DMFT spectrums are in quantitative agreement with experimental and previous theoretical results. The discrepancies between our results and experimental and previous theoretical results indicate that details of the DFT calculations, the definition of the projector, as well as the numerical implementation of the impurity solvers will still have appreciable influence on the outcomes of DFT+DMFT calculations. Further investigations along these lines are still needed. #### III.2.2 Ce2O3 Ce2O3 is a Mott insulator with a gap of about $2.4\text{\,}\mathrm{eV}$ [100, 22] arising from the strong correlation among 4f electrons. This system has been calculated using a variety of approaches such as DFT+U, GW [101], and DFT+DMFT [22, 28]. Ce2O3 crystalizes in the hexagonal lattice with space group $P\bar{3}m1$. The experimental lattice parameters $a=b=$3.891\text{\,}\mathrm{\SIUnitSymbolAngstrom}$$, $c=$6.059\text{\,}\mathrm{\SIUnitSymbolAngstrom}$$ [102] are used in our calculations. The energy window for KS subset $\mathcal{C}$ is [$-3.0\text{\,}\mathrm{eV}$, $2.0\text{\,}\mathrm{eV}$]. Constrained DFT calculation predicted the parameter U varying from 5.5 to $8.0\text{\,}\mathrm{eV}$, and we adopt U as $6.5\text{\,}\mathrm{eV}$ and J as $0.5\text{\,}\mathrm{eV}$. The temperature is set to be $300\text{\,}\mathrm{K}$. Figure 9: DOS of Ce2O3. The DFT+DMFT calculation is carried out by combination FHI-aims and impurity solver Rutgers. Here we calculate the total density of states (TDOS) of Ce2O3 through $\rho(\omega)=\frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{k}}A(\mathbf{k},\omega),$ (32) where $A(\mathbf{k},\omega)$ is defined in Eq. (31). The result is shown in Fig. 9, our DFT+DMFT calculation gives a band gap of about $2.5\text{\,}\mathrm{eV}$, which agrees well with the experimental and previous theoretical results [100, 22, 101, 28]. For example, all predict that the lower Hubbard peaks are sharp and narrow whereas the upper Hubbard peaks are strong and wide. The interesting feature about the upper Hubbard part is that it emerges as a low plateau between 1 and 3 eV above the Fermi level, continued by a pronounced peak between 3 and 4 eV. Note that the peaks mainly composed of O p bands sitting below the lower Hubbard peaks does not show up here due to the fact that those O p characteristic bands are not included in the subset $\mathcal{C}$ in the DFT+DMFT calculations whereas they are retained in Refs. [22, 101, 28]. #### III.2.3 Pu2O3 and PuO2 Plutonium (Pu)-based oxides such as Pu2O3 and PuO2 are essential for the fuel components in current nuclear reactors, as well as transmutation of the minor actinides from spent nuclear fuels [103, 104]. A clear understanding of the physio-chemical properties of plutonium-based oxides is of key importance for the safe operation and development of nuclear reactor systems and nuclear waste reprocessing [105, 106], and the correct description of the oxidation and reduction processes [107, 108]. The physical, chemical, and thermodynamical properties of Pu-based oxides such as the chemical bonding and electronic structure are intimately related to the states of strongly correlated 5f electrons. Conversely, the relative tendency of delocalization versus localization of strongly correlated 5f electrons is extremely sensitive to the physical and chemical environment of the Pu atom [109]. The description of complicated behaviors of 5f electrons, e.g., whether they are settled in the delocalized or localized states, or in the intermediate regime, is out of the reach of standard DFT. In contrast, DFT+DMFT is becoming a promising approach that facilitates an in-depth understanding of this type of materials. In this paper, we look at two plutonium oxides, i.e., Pu2O3 in its $\beta$ phase [110] (simply called Pu2O3 below) and PuO2 [111] with our DFT+DMFT implementation. We use the Hubbard parameter U of $4.0\text{\,}\mathrm{eV}$ and the Hund parameter J of $0.5\text{\,}\mathrm{eV}$ for Pu2O3. As for PuO2, U and J are chosen to be $5.0\text{\,}\mathrm{eV}$ and $0.6\text{\,}\mathrm{eV}$. The electronic temperatures for both systems are set at $300\text{\,}\mathrm{K}$. The chosen energy windows for the KS subsets $\mathcal{C}$ are [$-3.0\text{\,}\mathrm{eV}$, $2.0\text{\,}\mathrm{eV}$] for Pu2O3, and [$-3.0\text{\,}\mathrm{eV}$, $3.0\text{\,}\mathrm{eV}$] for PuO2. We calculated DFT+DMFT TDOS of Pu2O3 through the method introduced in the case of Ce2O3. The TDOS here contains contributions from both the correlated orbitals and the rest (here $spd$) orbitals. The projected density of states (PDOS) belonging to correlated ($5f$) orbitals, is evaluated by $\displaystyle\rho_{f}(\omega)=$ $\displaystyle\sum_{m}\frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{k}}-\frac{1}{\pi}\operatorname{Im}\left\\{\tilde{P}^{I}_{ij}(\bm{k},mm)\right.$ (33) $\displaystyle\left.\left[(\omega+\mu-\epsilon({\mathbf{k}}))I-\bar{\Sigma}(\mathbf{k},\omega)\right]^{-1}_{ij}\right\\}\,$ whereas the PDOS for the $spd$ orbitals is obtained by taking the difference between the TDOS and $\rho_{f}(\omega)$. The results are depicted in Fig. 10. Figure 10 indicates that the band gap of Pu2O3 as determined by DFT+DMFT is about $1.7\text{\,}\mathrm{eV}$, which is in good agreement with the experimental bands gap of $1.8\text{\,}\mathrm{eV}$ [112, 28]. Additionally, the occupation analysis yields an average occupation number nf=5.0 for Pu 5f electrons through $n=-\sum_{m}G^{imp}_{mm}(\beta)$, which is consistent with the chemical environment of Pu3+ in Pu2O3. Figure 10: TDOS and PDOS of Pu2O3. The TDOS is given by Eq. (32). The PDOS of Pu 5f electrons is evaluated through Eq. (33). The PDOS for spd electrons is obtained by subtracting the PDOS of Pu 5f electrons from the TDOS. The DFT+DMFT calculation is carried out by combining FHI-aims and the Rutgers impurity solver. In Fig. 11, the DFT+DMFT TDOS of PuO2 is plotted. The bands gap is predicted to be about $2.5\text{\,}\mathrm{eV}$, which is in good agreement with previous theoretical work [113] and the experimental band gap of $2.8\text{\,}\mathrm{eV}$ [114]. In PuO2, the peak of occupied Pu 5f electrons is noticeably sharper than that of Pu2O3, which indicates that the Pu 5f electrons in PuO2 are likely more localized than the case of Pu2O3. This localization picture is in agreement with the bigger gaps in PuO2. The occupation analysis gives an average occupation number of nf=4.0 for Pu 5f electrons, which agrees with the chemical state of Pu4+ in PuO2. Figure 11: DFT+DMFT TDOS of PuO2. The DFT+DMFT calculation is carried out by a combination FHI-aims and the Rutgers impurity solver. ## IV Summary In summary, we developed and implemented a formalism that allows us to carry out DFT+DMFT calculations within the NAO basis set framework. For transition metal compounds and $f$-electron systems, the most localized $d$ or $f$-type NAO is used to define the local correlated subspace. Such a choice is physically intuitive and implementationally convenient for NAO basis sets. Following what is usually done in the literature, only a subset of the KS bands $\mathcal{C}$ out of the full KS space around the Fermi energy, which hosts the majority of electrons with strong correlations, are enclosed in the definition of the projector. Our projector scheme is mathematically equivalent to the projective-Wannier-function approach adopted in Ref. [24]. We implemented such a DFT+DMFT formalism by interfacing two NAO-based DFT codes, i.e., FHI-aims [37] and ABACUS [38] with three DMFT quantum impurity solvers, i.e., PACS, iQIST [72, 73, 74], and the Rutgers impurity solver [30, 62, 75]. In particular, interfacing with the all-electron FHI-aims code allows one to study all types of strongly correlated materials over the periodic table. Our DFT+DMFT formalism and implementation are testified for three typical series of strongly correlated materials, namely the 3d transitional metal compounds SrVO3, MnO and NiO, 4f materials including the Ce metal and Ce2O3, as well as the 5f actinides Pu2O3 and PuO2. For SrVO3 and MnO, our calculated one-electron removal and addition spectra are in good agreement with previously reported DFT+DMFT results and experimental data. Furthermore, the calculated results are rather robust against the use of different DFT codes or different impurity solvers. For NiO, the obtained DFT+DMFT results show noticeable dependence on the chosen DFT codes and/or impurity solvers, although the key spectroscopic features are captured by all calculations. For $f$-electron systems, the characteristic three-peak structures are obtained for the Ce metal, whereas for the correlated insulators, the obtained DFT+DMFT band gaps are in good agreement with experiments. However, there remain issues calling for further investigations, like the energy separation of the quasiparticle peak and the upper Hubbard band for the Ce metal. Admittedly, our scheme is still in its infancy, and at the quantitative level there are still issue to be sorted out. Further improvement would be necessary for more reliable descriptions of charge-transfer-type Mott insulators, and the intricate lanthanides and actinides. However, we consider that our attempt of developing an infrastructure that merge NAO-based DFT codes and DMFT-based techniques is rather rewarding. This will not only enable standard DFT+DMFT calculations within the NAO basis set framework, but also paves the way for developing more advanced schemes by combining beyond-DFT approaches like hybrid functionals [40, 115, 41] or $GW$ [42], recently available in NAO-based DFT codes, with the DMFT machinery. ###### Acknowledgements. This work is supported by National Natural Science Foundation of China (Grant Nos. 12134012, 11874335, 11874263), the Strategic Priority Research Program of Chinese Academy of Sciences (Grant No. XDPB25), and The Max Planck Parter Group for Advanced Electronic Structure Methods. We thank valuable help from Li Huang on the impurity solver iQIST [72, 73, 74], E. Gull and S. Iskakov on the impurity solver ALPS-CTHYB-SEGMENTS [116, 117, 118], H. Shinaoka on the impurity solver ALPS-CTHYB [119, 120], and M. J. Han and J. H. Sim on the DFT+DMFT package DMFT-pack [46, 121]. ## References * [1] Paul R. C. Kent and Gabriel Kotliar. Toward a predictive theory of correlated materials. Science, 361(6400):348–354, jul 2018. * [2] Arpita Paul and Turan Birol. Applications of DFT+DMFT in Materials Science. Annu. Rev. Mater. Res., 49(1):31–52, jul 2019. * [3] Ran Adler, Chang-Jong Kang, Chuck-Hou Yee, and Gabriel Kotliar. Correlated materials design: prospects and challenges. Reports Prog. Phys., 82(1):012504, jan 2019. * [4] P. Hohenberg and W. Kohn. Inhomogeneous electron gas. Phys. Rev., 136:B864–B871, 1964. * [5] W. Kohn and L. J. Sham. Self-Consistent Equations Including Exchange and Correlation Effects. Phys. Rev., 140(4A):A1133–A1138, 1965. * [6] Vi Vi Anisimov, J. Zaanen, and O. K. Andersen. Band theory and Mott insulators: Hubbard U instead of Stoner I. Phys Rev B: Condens Matter, 44(3):943–954, 1991. * [7] Antoine Georges, Gabriel Kotliar, Werner Krauth, and Marcelo J. Rozenberg. Dynamical mean-field theory of strongly correlated fermion systems and the limit of infinite dimensions. Rev. Mod. Phys., 68(1):13–125, 1996. * [8] G. Kotliar, S. Y. Savrasov, K. Haule, V. S. Oudovenko, O. Parcollet, and C. A. Marianetti. Electronic structure calculations with dynamical mean-field theory. Rev. Mod. Phys., 78(3):865–951, 2006. * [9] K. Held. Electronic structure calculations using dynamical mean field theory. Adv. Phys., 56(6):829–926, 2007. * [10] Walter Metzner and Dieter Vollhardt. Correlated Lattice Fermions in $d$=$\infty$ Dimensions. Phys. Rev. Lett., 62(3):324–327, 1989. * [11] Fusayoshi J. Ohkawa. Electron Correlation in the Hubbard Model in $d$=$\infty$ Dimension. J. Phys. Soc. Jpn., 60(10):3218–3221, 1991. * [12] Antoine Georges and Gabriel Kotliar. Hubbard model in infinite dimensions. Phys. Rev. B, 45(12):6479–6483, mar 1992. * [13] M. Jarrell. Hubbard model in infinite dimensions: A quantum Monte Carlo study. Phys. Rev. Lett., 69(1):168–171, 1992. * [14] Vladimir I. Anisimov, F. Aryasetiawan, and A. I. Lichtenstein. First-principles calculations of the electronic structure and spectra of strongly correlated systems: the LDA + U method. J. Phys. Condens. Matter, 9(4):767–808, 1997. * [15] A. I. Lichtenstein and M. I. Katsnelson. Ab initio calculations of quasiparticle band structure in correlated systems: LDA++ approach. Phys. Rev. B, 57(12):6884–6895, 1998. * [16] S. Y. Savrasov, G. Kotliar, and E. Abrahams. Correlated electrons in $\delta$-plutonium within a dynamical mean-field picture. Nature, 410(6830):793–795, 2001. * [17] X. Dai, S. Y. Savrasov, G. Kotliar, A. Migliori, H. Ledbetter, and E. Abrahams. Calculated phonon spectra of plutonium at high temperatures. Science, 300(5621):953–955, 2003. * [18] J. H. Shim, K. Haule, and G. Kotliar. Fluctuating valence in a correlated solid and the anomalous properties of $\delta$-plutonium. Nature, 446(7135):513–516, 2007. * [19] J. H. Shim, K. Haule, and G. Kotliar. Modeling the Localized-to-Itinerant Electronic Transition in the Heavy Fermion System CeIrIn5. Science, 318(5856):1615–1617, dec 2007. * [20] V. I. Anisimov, I. V. Solovyev, M. A. Korotin, M. T. Czyżyk, and G. A. Sawatzky. Density-functional theory and NiO photoemission spectra. Phys. Rev. B, 48(23):16929–16934, 1993. * [21] Vladimir I. Anisimov, F. Aryasetiawan, and A. I. Lichtenstein. First-principles calculations of the electronic structure and spectra of strongly correlated systems: the LDA+U method. J. Phys. Condens. Matter, 9(4):767–808, 1997. * [22] L. V. Pourovskii, B. Amadon, S. Biermann, and A. Georges. Self-consistency over the charge density in dynamical mean-field theory: A linear muffin-tin implementation and some physical implications. Phys. Rev. B, 76(23):235101, dec 2007. * [23] O. Grånäs, I. Di Marco, P. Thunström, L. Nordström, O. Eriksson, T. Björkman, and J.M. Wills. Charge self-consistent dynamical mean-field theory based on the full-potential linear muffin-tin orbital method: Methodology and applications. Comput. Mater. Sci., 55:295–302, apr 2012. * [24] V. I. Anisimov, D. E. Kondakov, A. V. Kozhevnikov, I. A. Nekrasov, Z. V. Pchelkina, J. W. Allen, S.-K. Mo, H.-D. Kim, P. Metcalf, S. Suga, A. Sekiyama, G. Keller, I. Leonov, X. Ren, and D. Vollhardt. Full orbital calculation scheme for materials with strongly correlated electrons. Phys. Rev. B, 71(12):125119, mar 2005. * [25] E. Pavarini, S. Biermann, A. Poteryaev, A. I. Lichtenstein, A. Georges, and O. K. Andersen. Mott transition and suppression of orbital fluctuations in orthorhombic $3d^{1}$ perovskites. Phys. Rev. Lett., 92(17):176403, 2004. * [26] F. Lechermann, A. Georges, A. Poteryaev, S. Biermann, M. Posternak, A. Yamasaki, and O. K. Andersen. Dynamical mean-field theory using Wannier functions: A flexible route to electronic structure calculations of strongly correlated materials. Phys. Rev. B, 74(12):125120, 2006. * [27] B. Amadon, F. Lechermann, A. Georges, F. Jollet, T. O. Wehling, and A. I. Lichtenstein. Plane-wave based electronic structure calculations for correlated materials using dynamical mean-field theory and projected local orbitals. Phys. Rev. B, 77(20), 2008. * [28] B. Amadon. A self-consistent DFT+DMFT scheme in the projector augmented wave method: applications to cerium, Ce2O3 and Pu2O3 with the Hubbard I solver and comparison to DFT+U. J. Phys. Condens. Matter, 24(7):075604, 2012. * [29] Markus Aichhorn, Leonid Pourovskii, Veronica Vildosola, Michel Ferrero, Olivier Parcollet, Takashi Miyake, Antoine Georges, and Silke Biermann. Dynamical mean-field theory within an augmented plane-wave framework: Assessing electronic correlations in the iron pnictide LaFeAsO. Phys. Rev. B, 80(8):085101, aug 2009. * [30] Kristjan Haule, Chuck-Hou Yee, and Kyoo Kim. Dynamical mean-field theory within the full-potential methods: Electronic structure of CeIrIn5, CeCoIn5, and CeRhIn5. Phys. Rev. B, 81(19):195107, may 2010. * [31] Dominika Zgid and Garnet Kin-Lic Chan. Dynamical mean-field theory from a quantum chemical perspective. J. Chem. Phys., 134(9):094115, mar 2011. * [32] Wael Chibani, Xinguo Ren, Matthias Scheffler, and Patrick Rinke. Self-consistent Green’s function embedding for advanced electronic structure methods based on a dynamical mean-field concept. Phys. Rev. B, 93(16):165106, apr 2016. * [33] Tianyu Zhu and Garnet Kin-Lic Chan. Ab Initio Full Cell GW+DMFT for Correlated Materials. Phys. Rev. X, 11(2):021006, apr 2021. * [34] José M. Soler, Emilio Artacho, Julian D. Gale, Alberto García, Javier Junquera, Pablo Ordejón, and Daniel Sánchez-Portal. The SIESTA method for ab initio order-$N$ materials simulation. J. Phys. Condens. Matter, 14(11):2745–2779, mar 2002. * [35] T. Ozaki, H. Kino, J. Yu, M. Han, N. Kobayashi, M. Ohfuti, F. Ishii, and T. Ohwaki. User’s manual of OpenMX, http://www.openmx-square.org, 2008. * [36] B. Delley. From molecules to solids with the DMol3 approach. J. Chem. Phys., 113(18):7756–7764, nov 2000. * [37] Volker Blum, Ralf Gehrke, Felix Hanke, Paula Havu, Ville Havu, Xinguo Ren, Karsten Reuter, and Matthias Scheffler. Ab initio molecular simulations with numeric atom-centered orbitals. Comput. Phys. Commun., 180(11):2175–2196, nov 2009. * [38] Pengfei Li, Xiaohui Liu, Mohan Chen, Peize Lin, Xinguo Ren, Lin Lin, Chao Yang, and Lixin He. Large-scale ab initio simulations based on systematically improvable atomic basis. Comput. Mater. Sci., 112:503–517, 2016. * [39] Xinguo Ren, Patrick Rinke, Volker Blum, Jürgen Wieferink, Alexandre Tkatchenko, Andrea Sanfilippo, Karsten Reuter, and Matthias Scheffler. Resolution-of-identity approach to Hartree–Fock, hybrid density functionals, RPA, MP2 and GW with numeric atom-centered orbital basis functions. New J. Phys., 14(5):053020, may 2012. * [40] Sergey V. Levchenko, Xinguo Ren, Jürgen Wieferink, Rainer Johanni, Patrick Rinke, Volker Blum, and Matthias Scheffler. Hybrid functionals for large periodic systems in an all-electron, numeric atom-centered basis framework. Comput. Phys. Commun., 192:60–69, 2015. * [41] P. Lin, X. Ren, and L. He. Efficient hybrid density functional calculations for large periodic systems using numerical atomic orbitals. J. Chem. Theory Comput., 17(1):222–239, 2021. * [42] Xinguo Ren, Florian Merz, Hong Jiang, Yi Yao, Markus Rampp, Hermann Lederer, Volker Blum, and Matthias Scheffler. All-electron periodic $g_{0}w_{0}$ implementation with numerical atomic orbital basis functions: Algorithm and benchmarks. Phys. Rev. Mater., 5(1):013807, 2021. * [43] Myung Joon Han, Taisuke Ozaki, and Jaejun Yu. O$\left(N\right)$ LDA+U electronic structure calculation method based on the nonorthogonal pseudoatomic orbital basis. Phys. Rev. B, 73(4):045110, 2006. * [44] M. Kick, K. Reuter, and H. Oberhofer. Intricacies of DFT+U, Not Only in a Numeric Atom Centered Orbital Framework. J. Chem. Theory Comput., 15(3):1705–1718, 2019. * [45] Xin Qu, Peng Xu, Hong Jiang, Lixin He, and Xinguo Ren. DFT+U within the framework of linear combination of numerical atomic orbitals. arXiv e-prints, page arXiv:2202.05409, 2022. * [46] J. H. Sim and M. J. Han. Density functional theory plus dynamical mean-field theory with natural atomic orbital projectors. Phys. Rev. B, 100(11):115151, 2019. * [47] Dieter Vollhardt. Dynamical mean-field theory of electronic correlations in models and materials. AIP Conf. Proc., 1297:339–403, 2010. * [48] J. E. Hirsch and R. M. Fye. Monte carlo method for magnetic impurities in metals. Phys. Rev. Lett., 56(23):2521–2524, 1986. * [49] M. J. Rozenberg, X. Y. Zhang, and G. Kotliar. Mott-hubbard transition in infinite dimensions. Phys. Rev. Lett., 69(8):1236–1239, 1992. * [50] A. Georges and W. Krauth. Numerical solution of the $d$=$\infty$ Hubbard model: evidence for a Mott transition. Phys. Rev. Lett., 69(8):1240–1243, 1992. * [51] Th Pruschke, D. L. Cox, and M. Jarrell. Transport Properties of the Infinite-Dimensional Hubbard Model. Europhys. Lett., 21(5):593–598, feb 1993. * [52] Th. Pruschke, D. L. Cox, and M Jarrell. Hubbard model at infinite dimensions: Thermodynamic and transport properties. Phys. Rev. B, 47(7):3553–3565, feb 1993. * [53] M. Jarrell and T. Pruschke. Anomalous properties of the Hubbard model in infinite dimensions. Phys. Rev. B, 49(2):1458–1461, 1994. * [54] K. Haule, S. Kirchner, J. Kroha, and P. Wölfle. Anderson impurity model at finite Coulomb interaction U : Generalized noncrossing approximation. Phys. Rev. B, 64(15):155111, sep 2001. * [55] Kristjan Haule, Viktor Oudovenko, Sergej Y. Savrasov, and Gabriel Kotliar. The $\alpha\rightarrow\gamma$ Transition in Ce: A Theoretical View from Optical Spectroscopy. Phys. Rev. Lett., 94(3):036401, jan 2005. * [56] Q. Si, M. J. Rozenberg, G. Kotliar, and A. E. Ruckenstein. Correlation induced insulator to metal transitions. Phys. Rev. Lett., 72(17):2761–2764, 1994. * [57] M. Caffarel and W. Krauth. Exact diagonalization approach to correlated fermions in infinite dimensions: Mott transition and superconductivity. Phys. Rev. Lett., 72(10):1545–1548, 1994. * [58] O. Sakai and Y. Kuramoto. Application of the numerical renormalization group method to the hubbard model in infinite dimensions. Solid State Commun., 89(4):307–311, jan 1994. * [59] R. Bulla, A. C. Hewson, and Th Pruschke. Numerical renormalization group calculations for the self-energy of the impurity anderson model. J. Phys. Condens. Matter, 10(37):8365–8380, 1998. * [60] P. Werner, A. Comanac, L. De’ Medici, M. Troyer, and A. J. Millis. Continuous-time solver for quantum impurity models. Phys. Rev. Lett., 97(7):076405, 2006. * [61] Emanuel Gull, Philipp Werner, Andrew Millis, and Matthias Troyer. Performance analysis of continuous-time solvers for quantum impurity models. Phys. Rev. B, 76(23), 2007. * [62] K. Haule. Quantum Monte Carlo impurity solver for cluster dynamical mean-field theory and electronic structure calculations with adjustable cluster base. Phys. Rev. B, 75(15):155113, 2007. * [63] Emanuel Gull, Andrew J. Millis, Alexander I. Lichtenstein, Alexey N. Rubtsov, Matthias Troyer, and Philipp Werner. Continuous-time monte carlo methods for quantum impurity models. Rev. Mod. Phys., 83(2):349–404, 2011. * [64] Kristjan Haule, Turan Birol, and Gabriel Kotliar. Covalency in transition-metal oxides within all-electron dynamical mean-field theory. Phys. Rev. B, 90(7):075136, aug 2014. * [65] Junjiro Kanamori. Electron correlation and ferromagnetism of transition metals. Prog. Theor. Phys., 30(3):275–289, 1963. * [66] V. I. Anisimov and A. V. Lukoyanov. Investigation of real materials with strong electronic correlations by the LDA+DMFT method. Acta Crystallogr. Sect. C Struct. Chem., 70(2):137–159, feb 2014. * [67] D. R. Hamann. Optimized norm-conserving Vanderbilt pseudopotentials. Phys. Rev. B, 88(8):085117, 2013. * [68] Martin Schlipf and François Gygi. Optimization algorithm for the generation of ONCV pseudopotentials. Comput. Phys. Commun., 196:36–44, 2015. * [69] Peter Scherpelz, Marco Govoni, Ikutaro Hamada, and Giulia Galli. Implementation and Validation of Fully Relativistic GW Calculations: Spin–Orbit Coupling in Molecules, Nanocrystals, and Solids. J. Chem. Theory Comput., 12(8):3523–3544, 2016. * [70] John P. Perdew, Kieron Burke, and Matthias Ernzerhof. Generalized Gradient Approximation Made Simple. Phys. Rev. Lett., 77(18):3865–3868, 1996. * [71] The segment implementation [60] of the ct-qmc impurity solver is a part of the PACS@sdf package (Package for Analyzing Correlated Systems with Spatial and Dynamical fluctuations). PACS@sdf aims at providing an integrated framework for the study of strongly correlated models and materials beyond the local approximation of the DMFT [7]. It takes DMFT as the zero-order approximation and systematically provides non-local corrections to it [Li2015]. * [72] L. Huang, Y. L. Wang, Z. Y. Meng, L. Du, P. Werner, and X. Dai. iqist: An open source continuous-time quantum monte carlo impurity solver toolkit. Comput. Phys. Commun., 195:140–160, 2015. * [73] L. Huang. iqist v0.7: An open source continuous-time quantum monte carlo impurity solver toolkit. Comput. Phys. Commun., 221:423–424, 2017. * [74] https://github.com/huangli712/iQIST. * [75] http://hauleweb.rutgers.edu/tutorials/Tutorial0.html. * [76] I. A. Nekrasov, K. Held, G. Keller, D. E. Kondakov, Th. Pruschke, M. Kollar, O. K. Andersen, V. I. Anisimov, and D. Vollhardt. Momentum-resolved spectral functions of SrVO3 calculatd by LDA+DMFT. Phys. Rev. B, 73(15):155112, apr 2006. * [77] A. Liebsch. Surface versus Bulk Coulomb Correlations in Photoemission Spectra of SrVO3 and CaVO3. Phys. Rev. Lett., 90(9):096401, mar 2003. * [78] A. Sekiyama, H. Fujiwara, S. Imada, S. Suga, H. Eisaki, S. I. Uchida, K. Takegahara, H. Harima, Y. Saitoh, I. A. Nekrasov, G. Keller, D. E. Kondakov, A. V. Kozhevnikov, Th. Pruschke, K. Held, D. Vollhardt, and V. I. Anisimov. Mutual Experimental and Theoretical Validation of Bulk Photoemission Spectra of Sr1-xCaxVO3. Phys. Rev. Lett., 93(15):156402, oct 2004. * [79] I. A. Nekrasov, G. Keller, D. E. Kondakov, A. V. Kozhevnikov, Th. Pruschke, K. Held, D. Vollhardt, and V. I. Anisimov. Comparative study of correlation effects in CaVO3 and SrVO3. Phys. Rev. B, 72(15):155106, oct 2005. * [80] T. Yoshida, K. Tanaka, H. Yagi, A. Ino, H. Eisaki, A. Fujimori, and Z.-X. Shen. Direct Observation of the Mass Renormalization in SrVO3 by Angle Resolved Photoemission Spectroscopy. Phys. Rev. Lett., 95(14):146404, sep 2005. * [81] R. Eguchi, T. Kiss, S. Tsuda, T. Shimojima, T. Mizokami, T. Yokoya, A. Chainani, S. Shin, I. H. Inoue, T. Togashi, S. Watanabe, C. Q. Zhang, C. T. Chen, M. Arita, K. Shimada, H. Namatame, and M. Taniguchi. Bulk- and Surface-Sensitive High-Resolution Photoemission Study of Two Mott-Hubbard Systems: SrVO3 and CaVO3. Phys. Rev. Lett., 96(7):076402, feb 2006. * [82] K Maiti, D. D Sarma, M. J Rozenberg, I. H Inoue, H Makino, O Goto, M Pedio, and R Cimino. Electronic structure of Ca1-xSrxVO3 : A tale of two energy scales. Europhys. Lett., 55(2):246–252, jul 2001. * [83] Hendrik J. Monkhorst and James D. Pack. Special points for Brillouin-zone integrations. Phys. Rev. B, 13(12):5188–5192, jun 1976. * [84] X. Ren, I. Leonov, G. Keller, M. Kollar, I. Nekrasov, and D. Vollhardt. LDA+DMFT computation of the electronic spectrum of NiO. Phys. Rev. B, 74(19):195114, nov 2006. * [85] J. Kuneš, V. I. Anisimov, A. V. Lukoyanov, and D. Vollhardt. Local correlations and hole doping in NiO: A dynamical mean-field study. Phys. Rev. B, 75(16):165115, apr 2007. * [86] J. Zaanen, G. A. Sawatzky, and J. W. Allen. Band gaps and electronic structure of transition-metal compounds. Phys. Rev. Lett., 55(4):418–421, jul 1985. * [87] Gernot J. Kraberger, Robert Triebl, Manuel Zingl, and Markus Aichhorn. Maximum entropy formalism for the analytic continuation of matrix-valued Green’s functions. Phys. Rev. B, 96(15):155128, oct 2017. * [88] L. Messick, W. C. Walker, and R. Glosser. Direct and Temperature-Modulated Reflectance Spectra of MnO, CoO, and NiO. Phys. Rev. B, 6(10):3941–3949, 1972. * [89] G. A. Sawatzky and J. W. Allen. Magnitude and Origin of the Band Gap in NiO. Phys. Rev. Lett., 53(24):2339–2342, 1984. * [90] S. Hüfner, J. Osterwalder, T. Riesterer, and F. Hulliger. Photoemission and inverse photoemission spectroscopy of NiO. Solid State Commun., 52(9):793–796, 1984. * [91] M. Grioni, P. Weibel, D. Malterre, Y. Baer, and L. Du‘o. Resonant inverse photoemission in cerium-based materials. Phys. Rev. B, 55(4):2056–2067, jan 1997. * [92] E. Weschke, C. Laubschat, T. Simmons, M. Domke, O. Strebel, and G. Kaindl. Surface and bulk electronic structure of Ce metal studied by high-resolution resonant photoemission. Phys. Rev. B, 44(15):8304–8307, oct 1991. * [93] A. K. McMahan, C. Huscroft, R. T. Scalettar, and E. L. Pollock. Volume-collapse transitions in the rare earth metals. J. Comput. Aided Mater. Des., 5(2):131–162, 1998. * [94] K. Held, A. K. McMahan, and R. T. Scalettar. Cerium Volume Collapse: Results from the Merger of Dynamical Mean-Field Theory and Local Density Approximation. Phys. Rev. Lett., 87(27):276404, dec 2001. * [95] A. K. McMahan, K. Held, and R. T. Scalettar. Thermodynamic and spectral properties of compressed Ce calculated using a combined local-density approximation and dynamical mean-field theory. Phys. Rev. B, 67(7):075108, feb 2003. * [96] B. Amadon, S. Biermann, A. Georges, and F. Aryasetiawan. The $\alpha$-$\gamma$ Transition of Cerium Is Entropy Driven. Phys. Rev. Lett., 96(6):066402, feb 2006. * [97] M. J. Lipp, A. P. Sorini, J. Bradley, B. Maddox, K. T. Moore, H. Cynn, T. P. Devereaux, Y. Xiao, P. Chow, and W. J. Evans. X-ray Emission Spectroscopy of Cerium Across the $\gamma$-$\alpha$ Volume Collapse Transition. Phys. Rev. Lett., 109(19):195705, nov 2012. * [98] J. Bieder and B. Amadon. Thermodynamics of the $\alpha$-$\gamma$ transition in cerium from first principles. Phys. Rev. B, 89(19):195132, 2014. * [99] Marco Casadei, Xinguo Ren, Patrick Rinke, Angel Rubio, and Matthias Scheffler. Density functional theory study of the $\alpha$-$\gamma$ phase transition in cerium: Role of electron correlation and f-orbital localization. Phys. Rev. B, 93(7):075153, feb 2016. * [100] A.V. Prokofiev, A.I. Shelykh, and B.T. Melekh. Periodicity in the band gap variation of Ln2X3 (X = O, S, Se) in the lanthanide series. J. Alloys Compd., 242(1-2):41–44, sep 1996. * [101] Hong Jiang, Ricardo I. Gomez-Abal, Patrick Rinke, and Matthias Scheffler. Localized and Itinerant States in Lanthanide Oxides United by GW@LDA+U. Phys. Rev. Lett., 102(12):126403, mar 2009. * [102] H. Bärnighausen and G. Schiller. The crystal structure of A-Ce2O3. J. Less Common Met., 110(1-2):385–390, aug 1985. * [103] Juan J Carbajo, Gradyon L Yoder, Sergey G Popov, and Victor K Ivanov. A review of the thermophysical properties of MOX and UO2 fuels. J. Nucl. Mater., 299(3):181–198, dec 2001. * [104] Laurent Claparede, Mireille Guigue, Gauthier Jouan, Nassima Nadah, Nicolas Dacheux, and Philippe Moisy. Long-term behavior of refractory thorium-plutonium dioxide solid solutions. J. Nucl. Mater., 483:158–166, jan 2017. * [105] Tatsumi Arima, Sho Yamasaki, Yaohiro Inagaki, and Kazuya Idemitsu. Evaluation of thermal properties of UO2 and PuO2 by equilibrium molecular dynamics simulations from 300 to 2000K. J. Alloys Compd., 400(1-2):43–50, sep 2005. * [106] L. Petit, A. Svane, Z. Szotek, W. M. Temmerman, and G. M. Stocks. Electronic structure and ionicity of actinide oxides from first principles. Phys. Rev. B, 81(4):045108, jan 2010. * [107] Emily Moore, Christine Guéneau, and Jean-Paul Crocombette. Oxygen diffusion model of the mixed (U,Pu)O2±x: Assessment and application. J. Nucl. Mater., 485:216–230, mar 2017. * [108] Ru-song Li, Zheng Xie, Ling-Yun Kong, Su-xia Hou, Ji-jun Luo, and Du-qiang Xin. Intermediate occupation numbers for 5f electrons in a Pu and U mixed oxide. Phys. Chem. Chem. Phys., 23(27):14725–14736, 2021. * [109] Kevin T. Moore and Gerrit van der Laan. Nature of the 5f states in actinide metals. Rev. Mod. Phys., 81(1):235–298, feb 2009. * [110] M. Wulff and G. H. Lander. Magnetic structure and Pu ground state in $\beta$‐Pu2O3. J. Chem. Phys., 89(5):3295–3299, sep 1988. * [111] Renaud C. Belin, Philippe J. Valenza, Muriel A. Reynaud, and Philippe E. Raison. New hermetic sample holder for radioactive materials fitting to Siemens D5000 and Bruker D8 X-ray diffractometers: application to the Rietveld analysis of plutonium dioxide. J. Appl. Crystallogr., 37(6):1034–1037, dec 2004. * [112] C.E. McNeilly. The electrical properties of plutonium oxides. J. Nucl. Mater., 11(1):53–58, jan 1964. * [113] Jindřich Kolorenč, Alexander B. Shick, and Alexander I. Lichtenstein. Electronic structure and core-level spectra of light actinide dioxides in the dynamical mean-field theory. Phys. Rev. B, 92(8):085125, aug 2015. * [114] T. Mark McCleskey, Eve Bauer, Quanxi Jia, Anthony K. Burrell, Brian L. Scott, Steven D. Conradson, Alex Mueller, Lindsay Roy, Xiaodong Wen, Gustavo E. Scuseria, and Richard L. Martin. Optical band gap of NpO2 and PuO2 from optical absorbance of epitaxial films. J. Appl. Phys., 113(1):013515, jan 2013. * [115] P. Lin, X. Ren, and L. He. Accuracy of localized resolution of the identity in periodic hybrid functional calculations with numerical atomic orbitals. J. Phys. Chem. Lett., 11(8):3082–3088, 2020. * [116] E. Gull, P. Werner, S. Fuchs, B. Surer, T. Pruschke, and M. Troyer. Continuous-time quantum monte carlo impurity solvers. Comput. Phys. Commun., 182(4):1078–1082, 2011. * [117] Hartmut Hafermann, Philipp Werner, and Emanuel Gull. Efficient implementation of the continuous-time hybridization expansion quantum impurity solver. Comput. Phys. Commun., 184(4):1280–1286, apr 2013. * [118] https://github.com/ALPSCore/CT-HYB-SEGMEN. * [119] H. Shinaoka, E. Gull, and P. Werner. Continuous-time hybridization expansion quantum impurity solver for multi-orbital systems with complex hybridizations. Comput. Phys. Commun., 215:128–136, 2017. * [120] https://github.com/ALPSCore/CT-HYB. * [121] https://kaist-elst.github.io/DMFTpack/.
11institutetext: CAS Key Laboratory of Optical Astronomy, National Astronomical Observatories, Chinese Academy of Sciences, Beijing 100101, People’s Republic of China 11email<EMAIL_ADDRESS>22institutetext: School of Astronomy and Space Science, University of Chinese Academy of Sciences, Beijing 100049, People’s Republic of China 33institutetext: National Astronomical Observatory of Japan, 2-21-1 Osawa, Mitaka, Tokyo 181-8588, Japan 44institutetext: Astronomical Institute, Tohoku University, 6-3, Aramaki, Aoba-ku, Sendai, Miyagi 980-8578, Japan 55institutetext: Kavli Institute for the Physics and Mathematics of the Universe (WPI), The University of Tokyo Institutes for Advanced Study, The University of Tokyo, 5-1-5 Kashiwanoha, Kashiwa, Chiba 277-8583, Japan 66institutetext: Key Laboratory for Research in Galaxies and Cosmology, Shanghai Astronomical Observatory, 80 Nandan Road, Shanghai 200030, People’s Republic of China # Existence of Tidal Tails for the Globular Cluster NGC 5824 Yong Yang 1122 Jing-Kun Zhao 11 Miho N. Ishigaki 334455 Masashi Chiba 44 Cheng-Qun Yang 66 Xiang-Xiang Xue 11 Xian-Hao Ye 1122 Gang Zhao 1122 ###### Abstract Context. Several dynamically cold streams have been associated with certain globular clusters (GCs) based on orbital energies and angular momenta. Some of these streams are surprisingly far apart from their progenitors and one such pair is Triangulum and NGC 5824. Triangulum can be considered as a piece of NGC 5824 leading tail since the cluster’s future orbit matches with the stream’s track well. The existence of the leading tail for NGC 5824 is the motivation behind the search for its trailing tail. Aims. Our goal is to confirm the connection between Triangulum and NGC 5824 and seek the trailing tail of the cluster. Methods. The selection of member stars of Triangulum is made through various cuts in metallicity, proper motions (PMs), radial velocity and color-magnitude diagram (CMD). The selected members are compared in phase space to a mock stream which models the disruption of NGC 5824. We then try to detect the trailing tail of the cluster based on a modified matched-filter technique. Stars are assigned weights using their color differences from the cluster’s locus in CMD. These weights are further scaled based on stars’ departures from expected PMs of the model stream. Results. A total of 26 member stars for Triangulum are obtained and 16 of them are newly identified. These members are consistent with the mock stream in the phase space and their metalicity and position on the CMD are in good agreements with NGC 5824. By applying the matched-filter, a tenuous trailing tail of the cluster is detected, spanning $\sim$ 50$\degr$ long on sky. The signature matches with the mock stream’s trajectory well. Conclusions. Our results support that Triangulum stream acts as a part of the leading tail for NGC 5824. On the trailing side, we have detected a 50$\degr$ tail extended from the cluster. The existence of both leading and trailing tails for the GC NGC 5824 is verified. ###### Key Words.: globular clusters: individual: NGC 5824 – Galaxy: structure – Galaxy: kinematics and dynamics – Galaxy: halo ## 1 Introduction Increasing amount of data from various revolutionary surveys are revealing mysteries of stellar streams in the Milky Way and providing unprecedented details of the Galactic halo (e.g., Bell et al. 2008; Zhao et al. 2009; Law & Majewski 2010; Bowden et al. 2015; Bernard et al. 2016; Liang et al. 2017; Zhao et al. 2018; Malhan et al. 2018; Yang et al. 2019a, b; Zhao et al. 2020; Yang et al. 2021; Ye et al. 2021; Zhao & Chen 2021). Tidal streams extending from extant globular clusters (GCs) are usually thin and dynamically cold (e.g., Odenkirchen et al. 2003; Grillmair & Johnson 2006; Palau & Miralda- Escudé 2019; Grillmair 2019). Some narrow streams without explicit cores are generally also attributed to GC origins (e.g., Grillmair 2009; Koposov et al. 2010; Bonaca et al. 2012; Koposov et al. 2014; Shipp et al. 2018; Malhan et al. 2018). The progenitors of most of those streams are still unknown but several streams have been recently associated with extant GCs (Ibata et al. 2021). The connections between $\omega$ Centauri and Fimbulthul (Ibata et al. 2019), NGC 3201 and Gjöll (Palau & Miralda-Escudé 2021), and NGC 4590 and Fjörm (Palau & Miralda-Escudé 2019) have been reported, which suggest that the associations between a stream and a GC, where the GC does not connect directly to the stream, are present in the Milky Way. By exploring the orbits, Bonaca et al. (2021) further attributed 5 more streams to extant GCs (Table 1 therein), and one pair is Triangulum and NGC 5824. Triangulum stream was first detected by Bonaca et al. (2012) with a matched-filter technique (Rockosi et al. 2002). Thereafter, Martin et al. (2013) kinematically discovered a part of the stream and provided 11 possible member stars. The stream is in the direction of M31 and M33 galaxies, and far apart from NGC 5824. However, the cluster’s future orbit passes through the stream well, implying a connection between them (Fig. 4 in Bonaca et al. 2021). Li et al. (2022) further confirmed this connection by comparing a model stream of NGC 5824 in phase space to the Triangulum member stars from Martin et al. (2013). Therefore, Triangulum stream could be treated as a piece of NGC 5824 leading tail. Based on the picture that tidal tails are developed symmetrically around GCs (Küpper et al. 2010), the existence of leading tail for NGC 5824 motivates us to search for its trailing tail. In this work, we provide a confirmation of the connection between Triangulum and NGC 5824, which is similar to that of Li et al. (2022) but with member stars that span a wider sky extent ($\sim$ 16$\degr$). We further apply a modified match-filter method (Grillmair 2019) to look for the trailing tail of NGC 5824. The paper is organized as follows. In Sect. 2, we introduce the data. In Sect. 3, we show the selection of Triangulum member stars and compare them to a model stream of NGC 5824. The detection of the cluster’s trailing tail is given in Sect. 4. We present a discussion in Sect. 5 and draw our conclusion in Sect. 6. ## 2 Data We base our search on high-quality astrometric and photometric data provided by the $Gaia$ EDR3 (Gaia Collaboration et al. 2021; Lindegren et al. 2021; Riello et al. 2021), along with the spectroscopic data from the Sloan Extension for Galactic Understanding and Exploration (SEGUE; Yanny et al. 2009) and the Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST; Cui et al. 2012; Zhao et al. 2006, 2012; Liu et al. 2015) surveys. To obtain the individual members of Triangulum, we retrieve stars from the $Gaia$ EDR3 gaia_source catalog overlapping with the stream region on the celestial sphere. The stream region is determined by limiting 22$\degr$ $<$ $\delta$ $<$ 41$\degr$ and moving $\delta=-4.4\alpha+128.5$ by $\pm 1\degr$ along the $\alpha$ direction (green area in Fig. 1), where the equation was defined in Bonaca et al. (2012) to describe the stream coordinates. Note that Bonaca et al. (2012) traced Triangulum to $\delta\simeq 23\degr-35\degr$, and Martin et al. (2014) extended the stream to further north $\delta\simeq 40\degr$. Our choice of $\delta$ extent is based on both of them. The zero- point correction in the parallax is implemented using the code provided by Lindegren et al. (2021), which requires astrometric_params_solved $>$ 3\. The corrections of G-band magnitude and BP/RP excess factor are applied as instructed in Riello et al. (2021). In order to ensure good astrometric and photometric solutions, only stars with ruwe $<$ 1.4 and absolute corrected BP/RP excess factor smaller than 3 times the associated uncertainty (see Sect. 9.4 in Riello et al. 2021) are retained. Given that both of estimated distances of the stream in Bonaca et al. (2012) and Martin et al. (2013) are farther than 20 kpc, we remove foreground stars that satisfy the criterion $\varpi-3\sigma_{\varpi}>0.05$ mas. The remaining stars are cross-matched with SDSS/SEGUE DR16 (Ahn et al. 2012) and LAMOST DR8, by which the metallicity and heliocentric radial velocity are obtained. For stars that are common in both datasets, we adopt measurements from SEGUE because signal-to-noise ratios of spectra in SEGUE are mostly higher than those in LAMOST. The data for detecting trailing tail of NGC 5824 are also obtained from $Gaia$ EDR3. Stars within the sky box of 210$\degr$ $<$ $\alpha$ $<$ 250$\degr$ and -40$\degr$ $<$ $\delta$ $<$ 30$\degr$ are retrieved (orange area in Fig. 1) and reduced with the same procedures as above (including the foreground stars removing111Removing foreground stars within 20 kpc will not affect results since if the cluster’s trailing tail exists, it would be farther than 30 kpc from the sun (see Fig. 6).). Since the spectroscopic surveys are unavailable in this sky region, only $Gaia$ data are used. In Fig. 1, we show projections of the data (green and orange areas), along with a mock stream (red dots) which will be described in Sect. 3.2. The black line represents the Galactic plane, and the blue (inverted) triangle represents the direction of Galactic (anti-) center. It should be noted that the NGC 5824 field is exactly designed based on the mock stream. Figure 1: Sky projections of the data (green and orange areas) and a mock stream (red dots). The black line represents the Galactic plane, and the blue (inverted) triangle represents the direction of Galactic (anti-) center. The black circle denotes the GC NGC 5824. ## 3 Connection between Triangulum and NGC 5824 ### 3.1 Triangulum Member Stars Cross-matching between $Gaia$ sources and spectroscopic data yields 1,968 stars. Bonaca et al. (2012) presented an estimate of Triangulum’s [Fe/H] to be $\sim$ -1.0 dex, while Martin et al. (2013) contended a poorer metallicity $\simeq$ -2.2 dex for the stream. In order to obtain as many member stars as possible, we adopt [Fe/H] $<$ -1.0 dex as the selection criterion and are left with 451 candidates. After this cut, an overdensity can be seen clearly in proper motion (PM) space (top panel of Fig. 2, only the local region around the overdensity is shown). We overplot the member candidates provided by Martin et al. (2013) (cross-matched with $Gaia$ EDR3) and verify that this overdensity exactly corresponds to Triangulum stream. To pick out stream stars, we define a dispersion ellipse whose center and semi-axes are determined based on the known candidates from Martin et al. (2013). The center (1.014, 0.012) mas yr-1 is the mean PM of the members in $\alpha$ and $\delta$, and the semi-axes (0.777, 1.116) mas yr-1 are three times the PM dispersions in respective directions. 47 stars enclosed within the ellipse are selected. These stars are then plotted in $\delta$ \- $V_{r}$ plane (middle panel of Fig. 2) and a dominant monotonic sequence is clearly discernable. Generally, the radial velocities of a halo stream are supposed to change monotonically along coordinates as long as there is no turning point contained (like apogalacticon), such as Pal 5 (Ishigaki et al. 2016), GD-1 (Bovy et al. 2016), NGC 5466 (Yang et al. 2022), Hríd and Gjöll stream (Ibata et al. 2021). Hence we consider that this dominant sequence should correspond to Triangulum stream. We fit a straight line to the sequence where weights are given by the uncertainties of $V_{r}$. The relation can be described with the equation $V_{r}=-4.6\delta+86.5$. 31 stars with $V_{r}$ consistent to the fit in 3$\sigma$ range are retained. Finally, we reject 4 more outliers on the basis of color-magnitude diagram (CMD) and 27 member stars follow a typical GC isochrone (bottom panel of Fig. 2). All sources here have been extinction-corrected using the Schlegel et al. (1998) maps as re-calibrated by Schlafly & Finkbeiner (2011) with RV = 3.1, assuming $A_{G}/A_{V}=0.83627$, $A_{BP}/A_{V}=1.08337$, $A_{RP}/A_{V}=0.63439$222These extinction ratios are listed on the Padova model site http://stev.oapd.inaf.it/cgi-bin/cmd.. The detailed information of 27 member stars is summarized in Table 1. Table 1: Triangulum stream member stars. No. | $\alpha_{J2000}$ | $\delta_{J2000}$ | $\mu_{\alpha}^{*}$ | $\sigma_{\mu_{\alpha}^{*}}$ | $\mu_{\delta}$ | $\sigma_{\mu_{\delta}}$ | $V_{r}$ | $\sigma_{V_{r}}$ | [Fe/H] | $\sigma_{\rm[Fe/H]}$ | $G$ | $G_{bp}$ | $G_{rp}$ | Survey ---|---|---|---|---|---|---|---|---|---|---|---|---|---|--- | (°) | (°) | (mas yr-1) | (mas yr-1) | (mas yr-1) | (mas yr-1) | (km s-1) | (km s-1) | (dex) | (dex) | (mag) | (mag) | (mag) | 1⋆ | 23.8285 | 22.8031 | 1.0098 | 0.0986 | -0.1037 | 0.0784 | -16.84 | 3.11 | -1.913 | 0.068 | 16.759 | 17.222 | 16.106 | SEGUE 2${}^{\star}\times$ | 24.1829 | 22.9364 | 1.0127 | 0.1831 | -0.9748 | 0.1339 | 2.78 | 7.84 | -2.545 | 0.048 | 17.617 | 18.049 | 16.997 | SEGUE 3⋆ | 24.2157 | 22.9598 | 0.8790 | 0.3456 | 0.2562 | 0.2288 | -17.60 | 8.73 | -1.867 | 0.104 | 18.530 | 19.005 | 17.998 | SEGUE 4⋆ (BHB) | 23.2433 | 23.1934 | 0.6858 | 0.1844 | -0.2179 | 0.1546 | -32.59 | 6.76 | -2.121 | 0.078 | 17.967 | 18.031 | 17.815 | SEGUE 5⋆ | 24.1515 | 23.3639 | 1.2849 | 0.1128 | 0.2147 | 0.0938 | -23.76 | 4.55 | -2.348 | 0.114 | 17.335 | 17.803 | 16.718 | SEGUE 6⋆ | 23.9200 | 23.3903 | 0.5511 | 0.2549 | 0.4586 | 0.1641 | -26.84 | 7.61 | -2.128 | 0.114 | 18.251 | 18.672 | 17.667 | SEGUE 7⋆ | 23.8055 | 23.4783 | 1.1128 | 0.2314 | 0.0493 | 0.1889 | -16.15 | 8.19 | -2.371 | 0.056 | 18.457 | 18.891 | 17.927 | SEGUE 8 (BHB) | 24.5968 | 23.6575 | 1.1237 | 0.1755 | 0.1393 | 0.1465 | -28.03 | 4.77 | -2.278 | 0.152 | 17.834 | 17.903 | 17.692 | SEGUE 9⋆ | 23.7817 | 23.8800 | 1.0403 | 0.2248 | 0.2268 | 0.1843 | -36.43 | 5.51 | -2.383 | 0.095 | 18.256 | 18.669 | 17.716 | SEGUE 10⋆ | 23.9193 | 24.1185 | 0.8721 | 0.2242 | -0.1813 | 0.1614 | -31.52 | 5.94 | -1.953 | 0.128 | 18.220 | 18.705 | 17.689 | SEGUE 11 | 24.0593 | 24.1634 | 1.0918 | 0.0460 | 0.1548 | 0.0351 | -27.93 | 8.18 | -2.249 | 0.061 | 15.279 | 15.832 | 14.573 | LAMOST 12⋆ | 23.4145 | 24.3451 | 1.5329 | 0.3380 | 0.0838 | 0.1762 | -22.60 | 9.99 | -2.072 | 0.072 | 18.648 | 18.981 | 18.075 | SEGUE 13⋆ | 23.5792 | 24.3911 | 1.1712 | 0.2725 | 0.3187 | 0.1921 | -13.69 | 9.70 | -2.653 | 0.147 | 18.648 | 19.030 | 18.073 | SEGUE 14 | 23.5445 | 24.5692 | 1.0955 | 0.0584 | 0.1400 | 0.0363 | -32.73 | 9.18 | -1.962 | 0.135 | 16.025 | 16.537 | 15.347 | LAMOST 15 | 23.8850 | 24.7038 | 0.9478 | 0.1592 | 0.0680 | 0.1383 | -24.41 | 6.23 | -1.830 | 0.140 | 17.839 | 18.239 | 17.246 | SEGUE 16 | 22.3385 | 28.4684 | 0.9975 | 0.0801 | 0.3481 | 0.0569 | -41.53 | 12.97 | -2.311 | 0.111 | 16.581 | 17.045 | 15.961 | LAMOST 17 | 22.4622 | 30.0863 | 0.9254 | 0.1081 | 0.0873 | 0.0825 | -34.82 | 9.93 | -1.716 | 0.272 | 17.156 | 17.596 | 16.563 | LAMOST 18 | 22.3181 | 31.2075 | 0.9126 | 0.0968 | 0.0614 | 0.0716 | -57.01 | 16.48 | -2.143 | 0.241 | 17.152 | 17.616 | 16.521 | LAMOST 19 | 21.4457 | 34.2315 | 0.8154 | 0.0552 | 0.2357 | 0.0419 | -68.02 | 13.15 | -2.439 | 0.161 | 15.964 | 16.453 | 15.313 | LAMOST 20 | 20.8085 | 34.9178 | 0.7522 | 0.0638 | 0.4043 | 0.0443 | -75.26 | 12.41 | -2.056 | 0.136 | 16.213 | 16.680 | 15.569 | LAMOST 21 | 21.2461 | 35.1423 | 0.9246 | 0.1032 | 0.4356 | 0.0884 | -74.81 | 12.45 | -2.322 | 0.204 | 17.493 | 17.885 | 16.919 | LAMOST 22 | 21.3470 | 35.2527 | 0.8679 | 0.0563 | 0.2653 | 0.0435 | -74.42 | 11.61 | -2.141 | 0.211 | 16.273 | 16.759 | 15.613 | LAMOST 23 | 20.6703 | 37.3033 | 0.5967 | 0.2600 | 0.3732 | 0.2044 | -76.24 | 12.45 | -2.145 | 0.038 | 19.007 | 19.514 | 18.485 | SEGUE 24 (BHB) | 20.4665 | 37.9600 | 0.7622 | 0.1674 | 0.3160 | 0.1455 | -90.22 | 5.21 | -1.740 | 0.055 | 18.101 | 18.157 | 18.010 | SEGUE 25 | 20.6305 | 38.4133 | 0.9838 | 0.1966 | 0.2130 | 0.1741 | -87.51 | 7.20 | -2.042 | 0.089 | 18.524 | 18.966 | 17.931 | SEGUE 26 (BHB) | 20.7368 | 38.4931 | 1.0037 | 0.1538 | 0.2964 | 0.1236 | -105.28 | 7.22 | -1.378 | 0.040 | 17.908 | 18.044 | 17.637 | SEGUE 27 | 20.1517 | 38.6458 | 0.8931 | 0.1369 | 0.2047 | 0.0976 | -101.34 | 5.60 | -2.036 | 0.041 | 17.676 | 18.130 | 17.124 | SEGUE 333Identified member stars of Triangulum stream, sorted by $\delta$. Common stars of Martin et al. (2013) are marked with “$\star$.” An outlier identifies in $\mu_{\delta}$ panel of Fig. 3 is further labeled in “$\times$.” Cols. 12-14 are $Gaia$ magnitudes which have been extinction-corrected (see text). The last column indicates which survey the radial velocity and metallicity come from. Figure 2: The selections of Triangulum member stars. The gray dots represent rejected stars and the red ones represent the selected stars during each step. The member candidates identified by Martin et al. (2013) are marked in the green points. The top panel shows the local region of the overdensity in PM space, where the ellipse is defined to select member candidates in this step. The middle panel shows stars in $\alpha$ \- $V_{r}$ plane, where the error bars represent three times uncertainties of $V_{r}$ and the red line is a linear fit to the stream sequence. The bottom panel shows those candidates in CMD. ### 3.2 NGC 5824 Model Stream Li et al. (2022) have modeled the disruption of NGC 5824 in a static Milky Way potential plus a moving Large Magellanic Cloud (LMC). As the authors pointed, the model stream matched with observations of Triangulum well. Motivated by this, we also generate our own mock stream to make a similar comparison between the model and data, using the identified member stars above which span a wider sky extent. The model body is nearly identical to that of Li et al. (2022), but specific configurations are different, such as the Milky Way potential, the adopted mass and radius of LMC, the velocity dispersion and integration time (see details below). We use the Python package GALA (Price-Whelan 2017), which is designed for performing common tasks needed in Galactic Dynamics, to model the disruption of NGC 5824. The procedure closely follows that of Yang et al. (2022) as applied to NGC 5466. The adopted Milky Way potential consists of a Plummer bulge (Plummer 1911), $\Phi_{\rm bulge}$, two Miyamoto-Nagai disks (Miyamoto & Nagai 1975), $\Phi_{\rm thin}$ and $\Phi_{\rm thick}$, and a spherical NFW halo (Navarro et al. 1996), $\Phi_{\rm halo}$: $\Phi_{\rm bulge}(r)=\frac{-GM_{\rm bulge}}{\sqrt{r^{2}+b_{\rm bulge}^{2}}}$ (1) $\Phi_{\rm thin/thick}(R,z)=\frac{-GM_{\rm thin/thick}}{\sqrt{R^{2}+(a_{\rm thin/thick}+\sqrt{z^{2}+b_{\rm thin/thick}^{2}})^{2}}}$ (2) $\Phi_{\rm halo}(r)=\frac{-4\pi G\rho_{s}r_{s}^{3}}{r}{\rm ln}(1+\frac{r}{r_{s}})$ (3) where $r$ is the Galactocentric radius, $R$ is the cylindrical radius and $z$ is the vertical height. For the bulge and disks, we adopt the parameters from Pouliasis et al. (2017, Model I). The virial mass $M_{\rm virial}$ and concentration $c$ used to initialize the NFW halo are from McMillan (2017). Those chosen parameters are summarized in Table 2. Table 2: Adopted parameters for the Galactic potential. Parameter | Value ---|--- $M_{\rm bulge}$ | $1.0672\times 10^{10}M_{\sun}$ $b_{\rm bulge}$ | 0.3 kpc $M_{\rm thin}$ | $3.944\times 10^{10}M_{\sun}$ $a_{\rm thin}$ | 5.3 kpc $b_{\rm thin}$ | 0.25 kpc $M_{\rm thick}$ | $3.944\times 10^{10}M_{\sun}$ $a_{\rm thick}$ | 2.6 kpc $b_{\rm thick}$ | 0.8 kpc $M_{\rm virial}$ | $1.37\times 10^{12}M_{\sun}$ $c$ | 15.4 Following El-Falou & Webb (2022), we take a Hernquist Potential (Hernquist 1990) as the internal potential of LMC: $\Phi_{\rm LMC}(r^{\prime})=\frac{-GM_{\rm LMC}}{r^{\prime}+a_{\rm LMC}}$ (4) where $r^{\prime}$ is the distance to the LMC center and $M_{\rm LMC}$ and $a_{\rm LMC}$ are set to $10^{11}M_{\sun}$ and 10.2 kpc as well. The position and velocity of LMC are taken from Gaia Collaboration et al. (2018). As for the internal gravity of the GC NGC 5824, we choose a Plummer potential: $\Phi_{\rm GC}(r^{\prime\prime})=\frac{-GM_{\rm GC}}{\sqrt{r^{\prime\prime 2}+b_{\rm GC}^{2}}}$ (5) with a $M_{\rm GC}$ of $7.6\times 10^{5}M_{\sun}$ and a $b_{\rm GC}$ of 6.51 pc (half-mass radius) (Baumgardt & Hilker 2018). Here $r^{\prime\prime}$ denotes the distance to the cluster’s center. The position and velocity of NGC 5824 come from Vasiliev & Baumgardt (2021) and Harris (1996, 2010 edition). The solar distance to the Galactic center, circular velocity at the Sun and solar velocities relative to the Local Standard of Rest are set to 8 kpc, 220 km s-1 (Bovy et al. 2012) and (11.1, 12.24, 7.25) km s-1 (Schönrich et al. 2010), respectively. In the static Milky Way potential accompanied with a moving LMC, the cluster is initialized 2 Gyr ago444This integration time is chosen such that the generated mock tidal tail is long enough to completely cover the data. and integrated forward from then on, releasing two particles (leading and trailing directions respectively) at Lagrange points (Gibbons et al. 2014) per 0.05 Myr with a total of 40000 steps. The velocity dispersion is set to 11.9 km s-1 (Baumgardt & Hilker 2018) and the cluster mass is fixed during this process. By doing so, a mock stream for NGC 5824 is obtained as illustrated with the red dots in Fig. 1. We note that the observed Triangulum (green area) deviates a little from the locus of the mock stream, which also happened in Bonaca et al. (2021, Fig. 4 therein) and Li et al. (2022, Fig. 8 therein). We consider that this deviation between the observation and simulation might be common. ### 3.3 Phase Space We compare the Triangulum member stars to the model stream of NGC 5824 in phase space. In Fig. 3, right ascension $\alpha$, PMs $\mu^{*}_{\alpha}$ and $\mu_{\delta}$, and radial velocity $V_{r}$ as a function of declination $\delta$ are presented from top to bottom. The gray dots represent the stream particles within the same sky area as Triangulum. The member stars are shown in the red and green points. It can be seen that even though the selection process of member stars in Sect. 3.1 is completely independent of the model, the stream particles show good consistency with the observations in phase space. We note an outlier that falls too far from the others in $\mu_{\delta}$ plane. This star was selected by Martin et al. (2013) based on sky position, radial velocity, metallicity and CMD, when PM measurements were unavailable. We mark it with “$\times$” in Table 1 and remove it in subsequent analysis. Furthermore, we do not show distance plane here because there is some confusion, and we present a discussion about it in Sect. 5. Figure 3: Right ascension $\alpha$, PMs $\mu^{*}_{\alpha}$ and $\mu_{\delta}$, and radial velocity $V_{r}$ as a function of declination $\delta$ are presented from top to bottom. The gray dots represent the stream particles within the same sky area as Triangulum. The green and red points represent the stream member stars. ### 3.4 Metallicity and CMD To further examine whether Triangulum is stripped from GC NGC 5824, we compare them on the basis of metallicity and CMD. The metallicity distribution of Triangulum members is presented in Fig. 4. There are 4 blue horizontal branch (BHB) stars and 22 red giant branch (RGB) stars. For the whole sample, the mean value $\langle\rm[Fe/H]\rangle$ = -2.10 and standard deviation $\sigma_{\rm[Fe/H]}$ = 0.26 dex are consistent with those of Martin et al. (2013) ($\langle\rm[Fe/H]\rangle$ = -2.2, $\sigma_{\rm[Fe/H]}$ = 0.3 dex). Picking out RGB stars separately is aimed for a comparison to some chemical researches on NGC 5824. Mucciarelli et al. (2018) analyzed 87 RGB stars of the cluster and obtained a metallicity distribution peaked at [Fe/H] = -2.11 dex, which is very similar to $\langle\rm[Fe/H]\rangle$ = -2.14 dex here. The observed scatter $\sigma_{\rm[Fe/H]}$ = 0.22 dex could probably be caused by observational uncertainties in low-resolution spectra ($R\sim$1800). Figure 4: The metallicity distribution of Triangulum member stars. The red bars represent the whole sample and the green bars correspond to only RGB stars. To compare Triangulum with GC NGC 5824 in CMD, we need to know the stream’s distance. Xue et al. (2011) estimated distances of $\sim$ 5000 BHB stars by matching them in ($u-g$, $g-r$) space to theoretical colors for BHB stars with a series of absolute magnitudes. The individual distances of 4 BHB stars in our sample can be obtained from this catalog: 28.8, 26.9, 30.6 and 26.0 kpc for stars with No. 4, 8, 24 and 26 in Table 1, respectively. This yields a median distance of 27.85 kpc, close to 26 kpc proposed by Bonaca et al. (2012). In addition, we also estimate distances to all 26 stars (see Fig. 10) using the method from Carlin et al. (2015), which is a Bayesian approach with likelihood estimated via comparison of spectroscopically derived atmospheric parameters to a grid of stellar isochrones, and returns a posterior probability density function for star’s absolute magnitude. This yields a median value at 33 kpc similar to 35 kpc estimated by Martin et al. (2013). We adopt the distance to Triangulum stream as $\sim$ 30 kpc, which is a median value between BHB distance and our estimate. In CMD, we move the member stars from 30 to 32.1 kpc, where GC NGC 5824 is located (Harris 1996, 2010 edition), and find that they match well as shown in Fig. 5. The cluster stars here marked in the orange dots are obtained through sky and PM selections as instructed by Kundu et al. (2021). Specifically, we retrieve stars within the tidal radius $r_{t}$ = 5.73′ of NGC 5824 (Harris 1996, 2010 edition) and clean the data following procedures as described in Sect. 2. A 2D Gaussian mixture model consisting of two Gaussians is then fitted in PM space to decompose the cluster and field stars apart. For the cluster component, we get the center ($\mu^{*}_{\alpha}$, $\mu_{\delta}$) = (-1.193, -2.235) with the intrinsic dispersion ($\sigma^{\rm in}_{\mu^{*}_{\alpha}}$, $\sigma^{\rm in}_{\mu_{\delta}}$) = (0.424, 0.360) mas yr-1, where the center is very close to (-1.189, -2.234) mas yr-1 measured by Vasiliev & Baumgardt (2021). The cluster stars are selected as those whose PMs, within uncertainties, match the PMs and dispersion of NGC 5824: $\\{\mu^{*}_{\alpha}\pm\sigma_{\mu^{*}_{\alpha}},\mu_{\delta}\pm\sigma_{\mu_{\delta}}\\}^{\rm star}$ $\leq$ $\\{\mu^{*}_{\alpha}\pm\sigma^{\rm in}_{\mu^{*}_{\alpha}},\mu_{\delta}\pm\sigma^{\rm in}_{\mu_{\delta}}\\}^{\rm cluster}$. The black line denotes the RGB locus obtained by fitting the RGB stars directly with a third-order polynomial, which is used in Sect. 4.1 to assign weights in CMD. Figure 5: The orange dots represent GC NGC 5824 stars. The red and green dots represent Triangulum members. The black line denotes the RGB locus obtained by directly fitting the RGB stars with a third-order polynomial. The connection between the stream and the cluster, based on the three aspects above, confirms that Triangulum was disrupted from GC NGC 5824. In other words, the stream can be treated as a part of the cluster’s leading tail. ## 4 Detecting the Trailing Tail Motivated by the existence of leading tail for NGC 5824, in this section we aim to search for its trailing tail. ### 4.1 A Modified Matched Filter Method Combining PMs and CMD together to search for extra-tidal structures of GCs has proved to be an effective way (e.g., Kundu et al. 2019a, b, 2021). Here we adopt the method from Grillmair (2019) who applied a modified matched filter technique and successfully detected a 50$\degr$ long tidal tail for GC M5. Stars fetched in Sect. 2 are assigned weights based on their locations in CMD and PM space. In CMD, individual stars in the NGC 5824 field are assigned weights according to their color differences from the cluster locus, assuming a Gaussian error distribution: $w_{\rm CMD}=\frac{1}{\sqrt{2\pi}\sigma_{color}}{\rm exp}\left[-\frac{1}{2}\left(\frac{color- color_{0}}{\sigma_{color}}\right)^{2}\right].$ (6) Here $color$ and $\sigma_{color}$ denote $G_{BP}$ \- $G_{RP}$ and corresponding errors. Color errors are simply calculated through $\sqrt{\sigma^{2}_{G_{BP}}+\sigma^{2}_{G_{RP}}}$ where $\sigma_{G_{BP}}$ and $\sigma_{G_{RP}}$ are obtained with a propagation of flux errors (see CDS website 555https://vizier.u-strasbg.fr/viz- bin/VizieR-n?-source=METAnot&catid=1350&notid=63&-out=text.). $color_{0}$ is determined by the cluster RGB locus (the black line in Fig. 5) at a given $G$ magnitude of a star. During assigning weights, we do not include $\sigma_{G}$ since uncertainties in $G$ band are much smaller than those in $G_{BP}$ and $G_{RP}$ (on the order of $\sim$ 0.1) for $Gaia$ photometry. Stars from $G$ = 15 mag (tip of the cluster’s RGB) to the $Gaia$ limit $G\simeq$ 21 mag are investigated. The PMs of the model stream generated in Sect. 3.2 are further employed to weight stars. Fig. 6 shows the stream particles within the NGC 5824 field in phase space, which serves as an estimate to the real stream. In PM space, weights are computed as: $w_{\rm PMs}=\frac{1}{2\pi n^{2}\sigma_{\mu^{*}_{\alpha}}\sigma_{\mu_{\delta}}}{\rm exp}\left\\{-\frac{1}{2}\left[\left(\frac{\mu^{*}_{\alpha}-\mu^{*}_{\alpha,0}}{n\sigma_{\mu^{*}_{\alpha}}}\right)^{2}+\left(\frac{\mu_{\delta}-\mu_{\delta,0}}{n\sigma_{\mu_{\delta}}}\right)^{2}\right]\right\\}.$ (7) $\mu^{*}_{\alpha}$, $\mu_{\delta}$, $\sigma_{\mu^{*}_{\alpha}}$ and $\sigma_{\mu_{\delta}}$ are measured PMs and corresponding errors. $\mu^{*}_{\alpha,0}$ and $\mu_{\delta,0}$ are the components of PMs predicted at each star’s $\delta$ based on the model stream’s locus (blue lines of PM panels in Fig. 6). The locus is obtained by dividing the particles into $\delta$ bins (bin width = 1$\degr$) and calculating medians of PMs in each bin. It is worth noting that PM errors are multiplied by $n$ and we choose a moderate $n$ = 2 here, which is designed to allow some deviations between the model and observations. This can be illustrated using a one-dimensional example (see Fig. 7). Assume that we are going to assign a weight to a stream star (if exist) with $\mu_{\delta}$ = $x$ and $\sigma_{\mu_{\delta}}$ = 0.4 mas yr-1. The $\mu_{\delta,0}$ predicted by the model stream at the star’s $\delta$ is 2 mas yr-1. The star’s weight will be determined by a Gaussian with mean = 2 and sigma = 0.4 ($n$ = 1, red line) or 0.8 ($n$ = 2, green line). If the model predicts the stream very well, that is $x$ is very close to 2 mas yr-1, the red line ($n$ = 1) will give a higher weight to the star apparently. However, the model stream is just an approximation to the real one and it is likely that there are small deviations between them, which might lead to that $x$ falls out of the blue dashed lines. When this happens, the green line ($n$ = 2) gives a higher weight. We have compared results using different $n$ values and verified that $n$ = 2 is the most favorable. Figure 6: The planes of $\alpha$, heliocentric distance, proper motion in $\alpha$ and $\delta$, and radial velocity as a function of $\delta$, are shown from the top to the bottom, respectively. The pink dots represent the model stream particles within the NGC 5824 field. The red circle represents GC NGC 5824. The blue lines denote medians of y-axis values in each $\delta$ bin with a bin width of 1$\degr$. Figure 7: Illustration for using $n$ = 2 in Eq. (7). The red and green lines represent Gaussians centered at 2 with sigma = 0.4 and 0.8 mas yr-1, respectively. Finally, stars weights are obtained by multiplying $w_{\rm CMD}$ and $w_{\rm PMs}$, and then summed in $0.2\degr\times 0.2\degr$ sky pixels to expose structures. ### 4.2 Results A weighted sky map is obtained after applying the above method to data in the cluster field and shown in the left panel of Fig. 8. To make the stream look more prominent, pixels with summed weights $>$ 80 and $<$ 2 are masked such that too strong noises and weak background are not shown. The map is then smoothed with a Gaussian kernel of $\sigma$ = 0.5$\degr$. The stretch is logarithmic, with brighter areas corresponding to higher weight regions. The blue circle on the bottom marks the location of NGC 5824. The white bottom- right corner is due to being close to the Galactic disk, which is further masked in the middle and right panels. Figure 8: Log stretch of a matched filter map in the NGC 5824 field. The sky pixel width is 0.2$\degr$ and the map is smoothed with a Gaussian kernel of $\sigma$ = 0.5$\degr$. Three panels present the same map. The white arrows in left panel point the stream features. The locus of the model stream is overplotted in the small red dots in middle panel. The right panel illustrates the way of creating the stream’s lateral profile (see text). Bottom-right region close to the Galactic disk is masked in the middle and right panels. Due to the photometric depth of $Gaia$, the cluster’s main sequence stars are not observable and only RGB stars can be used to trace the underlying trailing tail, which are much fewer than the former. However, some stream-like signals are still detected. In the left panel, it is clear that there are several structures (marked with arrows) with higher weights between $\delta\simeq$ -21 $-$ -4$\degr$ that could be connected smoothly and likely extended from NGC 5824. In the middle panel, we overplot the trajectory of the model stream (small red dots) and find that it passes through the structures well. An additional segment of $\delta\simeq$ 6 $-$ 16$\degr$ is a farther extension of the stream. There is a gap in the middle at $\delta\simeq$ -4 $-$ 6$\degr$ corresponding to the most distant range of the model stream (see the distance panel in Fig. 6), where many RGB stars might have been darker than 21 mag. The detected signature traces the cluster’s trailing tail to $\sim 50\degr$ whose path can be roughly fitted using $\alpha=4.07\times 10^{-5}\delta^{3}+6.68\times 10^{-3}\delta^{2}+0.37\delta+232.45$ (8) where -33$\degr$ $<$ $\delta$ $<$ 16$\degr$. In the right panel of Fig. 8, stars enclosed by the red lines are selected to calculate the statistical significance of the stream. The $\delta$ range is $-22\degr--3\degr$. The central dashed line represents a more precise description to the stream of this region, which is given by $\alpha=7.15\times 10^{-3}\delta^{2}+0.38\delta+232.58+{\rm offset}$ (9) with offset = 0$\degr$. The left and right boundaries correspond to offset = -4 and 4$\degr$, respectively. A bin width = 0.2$\degr$ is used and at offset = -4, -3.8, -3.6…, weights of stars around Eq. (9) $\pm 0.1\degr$ are integrated to create a lateral profile of the stream as displayed in Fig. 9. The central peak at offset = 0$\degr$ represents the stream feature. The larger random counts at positive side are caused by higher stellar density near the disk. The significance is defined as $S=(w_{stream}-w_{background})/\sigma_{background}$, where $w_{stream}$ is the stream signal and $w_{background}$ and $\sigma_{background}$ are the mean and standard deviation of weights for off-stream regions 0.5$\degr$ $<$ —offset— $<$ 4$\degr$. We get S = 7.5 and 3.6 for negative and positive sides, respectively, and S is 4.3 if both are considered. It can be inferred from Fig. 9 that the stream’s width is expected to be $\lesssim 0.2\degr$ because signals drop back to the level of background when —offset— $>$ 0.1$\degr$ which means that there are few stream signals beyond this range. If we adopt $d$ = 39 kpc for this segment based on the model, the physical width is $\lesssim$ 136 pc. Figure 9: The stream one-dimensional profile. The offset coordinate is defined as deviation from the stream along $\alpha$ direction (see Eq. (9)). ### 4.3 A Part of Cetus? Bonaca et al. (2021) pointed that GC NGC 5824 and Cetus (Newberg et al. 2009), which is a stellar stream with a dwarf galaxy origin, have very close orbital energies and angular momenta. Similar orbital trajectories between them are also demonstrated in Chang et al. (2020). This arises a question: do those features on the trailing side of the cluster belong to Cetus stream? Combining the results here with previous researches on Cetus, we present 4 reasons for that the detected features are indeed related to the trailing tail of NGC 5824. 1. 1. The width of features in Fig. 8 is only $\lesssim 0.2\degr$, which is thin compared to a stream produced by a dwarf galaxy. 2. 2. Cetus stars should have a relatively spread distribution in CMD. However, the stream features indicated with arrows in Fig. 8 disappear if the RGB locus used to weight stars in CMD is shifted either blueward or redward by 0.1 mag, which means that they are exactly corresponding to NGC 5824. 3. 3. Chang et al. (2020) pointed that GC NGC 5824 should not be the core of Cetus, implying that there is no direct connection between the cluster and Cetus stream. Furthermore, Yuan et al. (2021) concentrated on searching for Cetus’s members using data covering the cluster but they did not detect any densely populated structure around NGC 5824. Hence the features should not be a part of Cetus. 4. 4. Triangulum as a piece of the leading tail also provides a weak evidence of existence for the trailing tail. ## 5 Discussion During comparing Triangulum to the model stream in distance, we find some incompatibility and show them in Fig. 10. Bonaca et al. (2012) estimated a Triangulum’s distance of 26 $\pm$ 4 kpc (the lower black error bar) while Martin et al. (2013) proposed 35 $\pm$ 3 kpc (the upper black error bar) for the stream. As mentioned above, we adopt a distance of 30 kpc (the green solid line) and find that the member stars match with GC NGC 5824 well in CMD. However, under a static Milky Way potential plus a moving LMC, the resulting model stream predicts that Triangulum’s distance should be about 20 - 25 kpc (gray dots), which is true in both this work and Li et al. (2022) (the second panel of Fig. 8 therein). This arises a confusion: why is there such a difference? Figure 10: Heliocentric distance as a function of $\delta$. The gray dots represent the stream particles. The blue points represent 4 BHB stars in our samples, whose distances come from Xue et al. (2011). The red dots and error bars present distances and corresponding errors for all member stars estimated using the method of Carlin et al. (2015). The red dashed line corresponds to their median value 33 kpc. The green solid line marks adopted distance 30 kpc. Two black error bars represent 26 $\pm$ 4 kpc (Bonaca et al. 2012) and 35 $\pm$ 3 kpc (Martin et al. 2013). Sheffield et al. (2014) presented an analysis on TriAnd1 ($d\sim$ 20 kpc) and TriAnd2 ($d\sim$ 28 kpc) (Martin et al. 2007), other two stellar substructures in the direction of M31 and M33. They show that even though the two structures are separated by more than 5 kpc in distance, they are indistinguishable in radial velocity and PMs. We note that this kinematic feature is very similar to that of Triangulum when compared to the model stream. The real and mock streams are separated by more than 5 kpc as well but their trends in phase space are still in concordance. Considering that the stream and those structures are exactly in the same region, it is very likely that Triangulum has been affected by the mechanism which leads to TriAnd1 and TriAnd2. Specifically, either related to a dwarf galaxy (Sheffield et al. 2014) or the Galactic disk (Xu et al. 2015), some process that created TriAnd1 and TriAnd2 might push Triangulum farther away (30 kpc) from where it should be (20 - 25 kpc). We anticipate that this prediction could be proved by later simulations on the formation of TriAnd overdensities. It is also worth nothing that there is another stream segment named Turbio (Shipp et al. 2018) between Triangulum and GC NGC 5824 that was considered to be disrupted from the cluster based on their similar dynamics in Bonaca et al. (2021) and Li et al. (2022). We do not inspect this stream due to lack of spectroscopic data. It is expected that upcoming observations will be able to provide more details on connections between Turbio and the cluster, and even more opportunities of searching for other stream segments on the leading side. If these can be confirmed, NGC 5824 tidal tails would be the longest cold stream ever discovered in the Milky Way. ## 6 Conclusions We first validate the connection between Triangulum stream and NGC 5824. A total of 26 stream member stars are selected and 16 of them are newly identified. We model the cluster’s disruption under a static Milky Way potential accompanied with a moving LMC. The real stream is compared to the mock one in phase space and consistent trends can be found. In metallicity and CMD, the member stars and the cluster are also in good agreement. These results support the previous statement that Triangulum originates from GC NGC 5824 (Bonaca et al. 2021; Li et al. 2022). Given that Triangulum can be considered as a segment of the cluster’s leading tail, we examine the existence of its trailing tail. Using a matched-filer method that combines CMD and PMs to weight stars, we find a $\sim$ 50$\degr$ trailing tail for GC NGC 5824. The features match with the model stream well. Although the signals are tenuous and discrete, a peak of $>$ 3$\sigma$ over background noises can be still discerned in the lateral stream profile, from which we estimate that its width is $\lesssim$ 0.2$\degr$. We expect that follow-up observations will be able to provide more details about the NGC 5824 stream. ###### Acknowledgements. We thank the anonymous referee, whose comments greatly improved this publication. This study was supported by the National Natural Science Foundation of China under grant nos 11988101, 11973048, 11927804, 11890694 and 11873052, and the National Key R&D Program of China, grant no. 2019YFA0405500. This work (MNI) is also supported by JSPS KAKENHI Grant Number 20H05855 and the GHfund A (202202018107). We acknowledge the support from the 2m Chinese Space Station Telescope project: CMS-CSST-2021-B05. Guoshoujing Telescope (the Large Sky Area Multi-Object Fiber Spectroscopic Telescope LAMOST) is a National Major Scientific Project built by the Chinese Academy of Sciences. Funding for the project has been provided by the National Development and Reform Commission. LAMOST is operated and managed by the National Astronomical Observatories, Chinese Academy of Sciences. This work presents results from the European Space Agency (ESA) space mission $Gaia$. $Gaia$ data are being processed by the $Gaia$ Data Processing and Analysis Consortium (DPAC). Funding for the DPAC is provided by national institutions, in particular the institutions participating in the $Gaia$ MultiLateral Agreement (MLA). The $Gaia$ mission website is https://www.cosmos.esa.int/gaia. The $Gaia$ archive website is https://archives.esac.esa.int/gaia. Funding for the Sloan Digital Sky Survey IV has been provided by the Alfred P. Sloan Foundation, the U.S. Department of Energy Office of Science, and the Participating Institutions. SDSS-IV acknowledges support and resources from the Center for High Performance Computing at the University of Utah. The SDSS website is www.sdss.org. SDSS-IV is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS Collaboration including the Brazilian Participation Group, the Carnegie Institution for Science, Carnegie Mellon University, Center for Astrophysics — Harvard & Smithsonian, the Chilean Participation Group, the French Participation Group, Instituto de Astrofísica de Canarias, The Johns Hopkins University, Kavli Institute for the Physics and Mathematics of the Universe (IPMU) / University of Tokyo, the Korean Participation Group, Lawrence Berkeley National Laboratory, Leibniz Institut für Astrophysik Potsdam (AIP), Max-Planck-Institut für Astronomie (MPIA Heidelberg), Max-Planck-Institut für Astrophysik (MPA Garching), Max- Planck-Institut für Extraterrestrische Physik (MPE), National Astronomical Observatories of China, New Mexico State University, New York University, University of Notre Dame, Observatário Nacional / MCTI, The Ohio State University, Pennsylvania State University, Shanghai Astronomical Observatory, United Kingdom Participation Group, Universidad Nacional Autónoma de México, University of Arizona, University of Colorado Boulder, University of Oxford, University of Portsmouth, University of Utah, University of Virginia, University of Washington, University of Wisconsin, Vanderbilt University, and Yale University. ## References * Ahn et al. (2012) Ahn, C. P., Alexandroff, R., Allende Prieto, C., et al. 2012, ApJS, 203, 21 * Baumgardt & Hilker (2018) Baumgardt, H. & Hilker, M. 2018, MNRAS, 478, 1520 * Bell et al. (2008) Bell, E. F., Zucker, D. B., Belokurov, V., et al. 2008, ApJ, 680, 295 * Bernard et al. (2016) Bernard, E. J., Ferguson, A. M. N., Schlafly, E. F., et al. 2016, MNRAS, 463, 1759 * Bonaca et al. (2012) Bonaca, A., Geha, M., & Kallivayalil, N. 2012, ApJ, 760, L6 * Bonaca et al. (2021) Bonaca, A., Naidu, R. P., Conroy, C., et al. 2021, ApJ, 909, L26 * Bovy et al. (2012) Bovy, J., Allende Prieto, C., Beers, T. C., et al. 2012, ApJ, 759, 131 * Bovy et al. (2016) Bovy, J., Bahmanyar, A., Fritz, T. K., & Kallivayalil, N. 2016, ApJ, 833, 31 * Bowden et al. (2015) Bowden, A., Belokurov, V., & Evans, N. W. 2015, MNRAS, 449, 1391 * Carlin et al. (2015) Carlin, J. L., Liu, C., Newberg, H. J., et al. 2015, AJ, 150, 4 * Chang et al. (2020) Chang, J., Yuan, Z., Xue, X.-X., et al. 2020, ApJ, 905, 100 * Cui et al. (2012) Cui, X.-Q., Zhao, Y.-H., Chu, Y.-Q., et al. 2012, RAA, 12, 1197 * El-Falou & Webb (2022) El-Falou, N. & Webb, J. J. 2022, MNRAS, 510, 2437 * Gaia Collaboration et al. (2021) Gaia Collaboration, Brown, A. G. A., Vallenari, A., et al. 2021, A&A, 649, A1 * Gaia Collaboration et al. (2018) Gaia Collaboration, Helmi, A., van Leeuwen, F., et al. 2018, A&A, 616, A12 * Gibbons et al. (2014) Gibbons, S. L. J., Belokurov, V., & Evans, N. W. 2014, MNRAS, 445, 3788 * Grillmair (2009) Grillmair, C. J. 2009, ApJ, 693, 1118 * Grillmair (2019) Grillmair, C. J. 2019, ApJ, 884, 174 * Grillmair & Johnson (2006) Grillmair, C. J. & Johnson, R. 2006, ApJ, 639, L17 * Harris (1996) Harris, W. E. 1996, AJ, 112, 1487 * Hernquist (1990) Hernquist, L. 1990, ApJ, 356, 359 * Ibata et al. (2021) Ibata, R., Malhan, K., Martin, N., et al. 2021, ApJ, 914, 123 * Ibata et al. (2019) Ibata, R. A., Bellazzini, M., Malhan, K., Martin, N., & Bianchini, P. 2019, Nature Astronomy, 3, 667 * Ishigaki et al. (2016) Ishigaki, M. N., Hwang, N., Chiba, M., & Aoki, W. 2016, ApJ, 823, 157 * Koposov et al. (2014) Koposov, S. E., Irwin, M., Belokurov, V., et al. 2014, MNRAS, 442, L85 * Koposov et al. (2010) Koposov, S. E., Rix, H.-W., & Hogg, D. W. 2010, ApJ, 712, 260 * Kundu et al. (2019a) Kundu, R., Fernández-Trincado, J. G., Minniti, D., et al. 2019a, MNRAS, 489, 4565 * Kundu et al. (2019b) Kundu, R., Minniti, D., & Singh, H. P. 2019b, MNRAS, 483, 1737 * Kundu et al. (2021) Kundu, R., Navarrete, C., Fernández-Trincado, J. G., et al. 2021, A&A, 645, A116 * Küpper et al. (2010) Küpper, A. H. W., Kroupa, P., Baumgardt, H., & Heggie, D. C. 2010, MNRAS, 401, 105 * Law & Majewski (2010) Law, D. R. & Majewski, S. R. 2010, ApJ, 714, 229 * Li et al. (2022) Li, T. S., Ji, A. P., Pace, A. B., et al. 2022, ApJ, 928, 30 * Liang et al. (2017) Liang, X. L., Zhao, J. K., Oswalt, T. D., et al. 2017, ApJ, 844, 152 * Lindegren et al. (2021) Lindegren, L., Bastian, U., Biermann, M., et al. 2021, A&A, 649, A4 * Liu et al. (2015) Liu, X.-W., Zhao, G., & Hou, J.-L. 2015, Research in Astronomy and Astrophysics, 15, 1089 * Malhan et al. (2018) Malhan, K., Ibata, R. A., & Martin, N. F. 2018, MNRAS, 481, 3442 * Martin et al. (2013) Martin, C., Carlin, J. L., Newberg, H. J., & Grillmair, C. 2013, ApJ, 765, L39 * Martin et al. (2007) Martin, N. F., Ibata, R. A., & Irwin, M. 2007, ApJ, 668, L123 * Martin et al. (2014) Martin, N. F., Ibata, R. A., Rich, R. M., et al. 2014, ApJ, 787, 19 * McMillan (2017) McMillan, P. J. 2017, MNRAS, 465, 76 * Miyamoto & Nagai (1975) Miyamoto, M. & Nagai, R. 1975, PASJ, 27, 533 * Mucciarelli et al. (2018) Mucciarelli, A., Lapenna, E., Ferraro, F. R., & Lanzoni, B. 2018, ApJ, 859, 75 * Navarro et al. (1996) Navarro, J. F., Frenk, C. S., & White, S. D. M. 1996, ApJ, 462, 563 * Newberg et al. (2009) Newberg, H. J., Yanny, B., & Willett, B. A. 2009, ApJ, 700, L61 * Odenkirchen et al. (2003) Odenkirchen, M., Grebel, E. K., Dehnen, W., et al. 2003, AJ, 126, 2385 * Palau & Miralda-Escudé (2019) Palau, C. G. & Miralda-Escudé, J. 2019, MNRAS, 488, 1535 * Palau & Miralda-Escudé (2021) Palau, C. G. & Miralda-Escudé, J. 2021, MNRAS, 504, 2727 * Plummer (1911) Plummer, H. C. 1911, MNRAS, 71, 460 * Pouliasis et al. (2017) Pouliasis, E., Di Matteo, P., & Haywood, M. 2017, A&A, 598, A66 * Price-Whelan (2017) Price-Whelan, A. M. 2017, The Journal of Open Source Software, 2, 388 * Riello et al. (2021) Riello, M., De Angeli, F., Evans, D. W., et al. 2021, A&A, 649, A3 * Rockosi et al. (2002) Rockosi, C. M., Odenkirchen, M., Grebel, E. K., et al. 2002, AJ, 124, 349 * Schlafly & Finkbeiner (2011) Schlafly, E. F. & Finkbeiner, D. P. 2011, ApJ, 737, 103 * Schlegel et al. (1998) Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ, 500, 525 * Schönrich et al. (2010) Schönrich, R., Binney, J., & Dehnen, W. 2010, MNRAS, 403, 1829 * Sheffield et al. (2014) Sheffield, A. A., Johnston, K. V., Majewski, S. R., et al. 2014, ApJ, 793, 62 * Shipp et al. (2018) Shipp, N., Drlica-Wagner, A., Balbinot, E., et al. 2018, ApJ, 862, 114 * Vasiliev & Baumgardt (2021) Vasiliev, E. & Baumgardt, H. 2021, MNRAS, 505, 5978 * Xu et al. (2015) Xu, Y., Newberg, H. J., Carlin, J. L., et al. 2015, ApJ, 801, 105 * Xue et al. (2011) Xue, X.-X., Rix, H.-W., Yanny, B., et al. 2011, ApJ, 738, 79 * Yang et al. (2019a) Yang, C., Xue, X.-X., Li, J., et al. 2019a, ApJ, 886, 154 * Yang et al. (2019b) Yang, C., Xue, X.-X., Li, J., et al. 2019b, ApJ, 880, 65 * Yang et al. (2021) Yang, Y., Zhao, J., Zhang, J., Ye, X., & Zhao, G. 2021, ApJ, 922, 105 * Yang et al. (2022) Yang, Y., Zhao, J.-K., Ishigaki, M. N., et al. 2022, MNRAS, 513, 853 * Yanny et al. (2009) Yanny, B., Rockosi, C., Newberg, H. J., et al. 2009, AJ, 137, 4377 * Ye et al. (2021) Ye, X., Zhao, J., Zhang, J., Yang, Y., & Zhao, G. 2021, AJ, 162, 171 * Yuan et al. (2021) Yuan, Z., Malhan, K., Sestito, F., et al. 2021, arXiv e-prints, arXiv:2112.05775 * Zhao & Chen (2021) Zhao, G. & Chen, Y. 2021, Science China Physics, Mechanics, and Astronomy, 64, 239562 * Zhao et al. (2006) Zhao, G., Chen, Y.-Q., Shi, J.-R., et al. 2006, Chinese J. Astron. Astrophys., 6, 265 * Zhao et al. (2012) Zhao, G., Zhao, Y.-H., Chu, Y.-Q., Jing, Y.-P., & Deng, L.-C. 2012, RAA, 12, 723 * Zhao et al. (2009) Zhao, J., Zhao, G., & Chen, Y. 2009, ApJ, 692, L113 * Zhao et al. (2020) Zhao, J. K., Ye, X. H., Wu, H., et al. 2020, ApJ, 904, 61 * Zhao et al. (2018) Zhao, J. K., Zhao, G., Aoki, W., et al. 2018, ApJ, 868, 105
# TypeScript’s Evolution: An Analysis of Feature Adoption Over Time ††thanks: This research was supported by an Australian Government Research Training Program Scholarship. Joshua D. Scarsbrook, Mark Utting, Ryan K. L. Ko School of Information Technology and Electrical Engineering The University of Queensland Brisbane, Australia <EMAIL_ADDRESS> ###### Abstract TypeScript is a quickly evolving superset of JavaScript with active development of new features. Our paper seeks to understand how quickly these features are adopted by the developer community. Existing work in JavaScript shows the adoption of dynamic language features can be a major hindrance to static analysis. As TypeScript evolves the addition of features makes the underlying standard more and more difficult to keep up with. In our work we present an analysis of 454 open source TypeScript repositories and study the adoption of 13 language features over the past three years. We show that while new versions of the TypeScript compiler are aggressively adopted by the community, the same cannot be said for language features. While some experience strong growth others are rarely adopted by projects. Our work serves as a starting point for future study of the adoption of features in TypeScript. We also release our analysis and data gathering software as open source in the hope it helps the programming languages community. ###### Index Terms: TypeScript, JavaScript, Data Mining ## I Introduction TypeScript[1] is a fast evolving superset of JavaScript implementing static type checking. From 2020 to 2022, there have been ten releases each bringing additional features and most adding new syntax to the language. With the rapid pace of evolution the question becomes, how quickly are these features being picked up by the developer community? Are some features more popular than others? This question has already been asked about other programming languages such as JavaScript, Java, and Python. In JavaScript the work by Richards et al. [2] explores the use of dynamic language features in JavaScript and concludes that production applications often use dynamic features making static analysis challenging. Similar work has also been done in Java where Parnin et al. [3] discovered that most uses of generics were covered by a small number of classes but the usage varies between developers. TypeScript has an evolving standard without a formal specification. Our paper seeks to understand how quickly new features in TypeScript are adopted, to determine how important it is for tools to stay up to date with the latest release. We hypothesize that it is unnecessary for program analysis tools to support the entire language and a smaller subset is sufficient for most applications. In this paper, we focus specifically on syntactic features (features implemented in the Abstract Syntax Tree without modifying language semantics) introduced by TypeScript versions between 4.0 and 4.9 in popular TypeScript libraries and applications. TypeScript also sees regular improvements to type inference and language features that are expressed through the type checker. These features are not a focus for our study. In this paper, we aimed to answer three research questions about the adoption of TypeScript features: * • (RQ1) What are the most popular features recently introduced in TypeScript? * • (RQ2) How quickly are new TypeScript features adopted by projects that use TypeScript? * • (RQ3) How quickly are new TypeScript language versions adopted by projects that use TypeScript? Results are presented in Section III. In this paper, we contribute: * • A dataset of current popular TypeScript repositories collected from GitHub[4]. * • A open source framework for TypeScript feature/version adoption studies. * • The first study of the rate of language feature/version adoption for TypeScript. * • Recommendations for how important it is for tools to adopt new language features in TypeScript. Our paper is organized into four further sections. We start with our methodology for analysis (Section II) before presenting our results (Section III). We then make a brief review of related work (Section IV) before concluding with a discussion including some future research directions (Section V). ## II Methodology We ran our study on top-starred/rated repositories containing TypeScript code on GitHub. We extracted all commits between 2020 and 2022 (inclusive) and extracted a series of boolean flags indicating the usage of each language feature. The analysis code and data sets used for this analysis are available in our repository.111See https://github.com/Vbitz/jsdata_msr. ### II-A Dataset We started by downloading a list of the top $500$ TypeScript repositories from GitHub. The repositories are sorted according to the number of users that have starred the repository. These repositories include languages besides TypeScript code but only TypeScript is considered in our paper. We collected the list of repositories on January 4 2023 and included the list as part of our dataset. Our analysis includes all commits attached to a given repository. These projects often use feature branches which may include a feature well before its released on the main branch. For all calculations we used the date the feature first turned up in the repository rather than the date it was included in a release. Of those $500$ repositories $23$ had no commits extracted and an additional $23$ recorded no versions of TypeScript. Therefore there are $454$ repositories with at least one version of TypeScript recorded. ### II-B Analysis Our pipeline consists of an open source program written in Go[5] that extracts every unique TypeScript file from every commit in each repository. This includes all branches and all tags. We only consider commits made between January 1st 2020 and December 31st 2022 inclusive. We filtered by dates and only selected TypeScript features released between 2020 and 2022 because including commits outside this time span will not yield useful results. In our dataset of $454$ repositories, $87\%$ contain less than $1000$ TypeScript files. We consider $1,325,810$ total commits in our analysis. Git commits contain multiple dates such as when the commit was authored versus when the commit was committed. For our analysis we choose the latest possible date included in the commit. Extracted TypeScript files are parsed by TypeScript and usage of different language features are detected according to their presence in the Abstract Syntax Tree (AST). With extensive caching and duplicate detection, the entire analysis takes approximately one hour. ### II-C Version Detection We parse the package.json file in the root of the repository to detect the TypeScript version from the installed dependencies. ### II-D Feature List We focused on syntactic features which are exposed in the AST exported by TypeScript. We chose to focus on features released in the last three years (between 2020 and 2022). TypeScript versions are released as Beta and a Release Candidate before they are formally released. In our paper, we consider the full release to be Day Zero, as listed below. Projects adopting betas will show up as adopting features or versions before they were formally released (a negative number of days relative to Day Zero). TypeScript 4.8 and 4.6 did not make syntactic changes to the language and only included semantic and inference changes. TABLE I: A list of the 8 TypeScript versions and 13 TypeScript features studied in our paper. Version | Release Date | | Name ---|---|---|--- 4.9 | 2022-11-15 | $f_{0}$ | satisfies operator | | $f_{1}$ | accessor property 4.7 | 2022-05-24 | $f_{2}$ | extends constraint on infer | | $f_{3}$ | Variance Annotations in and out 4.5 | 2021-11-17 | $f_{4}$ | type import modifier | | $f_{5}$ | Import assertions 4.4 | 2021-08-26 | $f_{6}$ | static blocks in classes 4.3 | 2021-05-26 | $f_{7}$ | override modifier on methods 4.2 | 2021-02-23 | $f_{8}$ | Abstract constructs signature 4.1 | 2020-11-19 | $f_{9}$ | Template literal types | | $f_{10}$ | Key Remapping in Mapped Types 4.0 | 2020-08-20 | $f_{11}$ | Labeled Tuple Elements | | $f_{12}$ | Short-Circuiting Assignment Table I lists the $8$ versions and $13$ features in our study. This is not an exhaustive list of features introduced since we excluded features requiring type inference or type checking to identify. Our dataset includes a few special repositories that have different characteristics to other projects: * • TypeScript[1]: The source code of TypeScript is included as part of this analysis. * • Babel[6]: Babel is a compiler for JavaScript. It includes both ECMAScript[7] features and some TypeScript features since it has support for TypeScript syntax. ## III Results To address our research questions, we started with the adoption of TypeScript versions before moving onto the adoption of TypeScript features. ### III-A RQ1: Feature Adoption Rating We can categorize the adoption of features into two major groups based on their adoption slopes. Group one contains $f_{4}$, $f_{9}$, $f_{11}$, $f_{7}$, $f_{12}$, $f_{10}$ and includes features that have more than $20$ repositories adopting them within one year after release. The most popular feature in this dataset is $f_{4}$ (type modifiers on import) and the second most popular is $f_{9}$ (Template Literal Types). Both of these features were necessary to solve some missing gaps in the TypeScript language. type modifiers ensure imports are used only for type definitions, and are erased when the code is compiled. This allows including other libraries and files without including runtime dependencies. It can also help to break import loops in some cases where a module needs a type from a module that imports it. Template Literal Types similarly give more flexibility in how types are described and open up new avenues of meta programming. Group two contains $f_{2}$, $f_{1}$, $f_{8}$, $f_{3}$, $f_{0}$, $f_{5}$, $f_{6}$ and includes features that have less than $20$ repositories implementing them one year after release. $f_{6}$ (static blocks in classes) has the lowest adoption rate among our dataset with only four repositories adopting it in a year. Unlike $f_{4}$ and $f_{9}$ static blocks have equivalents in existing code so they are only used in niche circumstances. ### III-B RQ2: Feature Adoption Figure 1: How quickly is each TypeScript feature adopted relative to one another. Note the release date of each feature as some features have not been released for all $800$ days. Figure 1 shows the adoption curve of each of the TypeScript feature we looked at. Unlike Figure 2 we can immediately see two major differences. Different features have significantly different adoption rates with some reaching high levels of adoption and some barely being adopted at all. Secondly all features have mostly linear adoption rates. Features were detected across any file ending with .ts that can successfully be parsed as TypeScript. This means features that are only used in unit tests are also included here. In addition we include every branch of the repository so some features are adopted first in a feature branch before being included in the main branch. TABLE II: The number of days before/after release where features were introduced into TypeScript and Babel. | $f_{0}$ | $f_{1}$ | $f_{2}$ | $f_{3}$ | $f_{4}$ | $f_{5}$ | $f_{6}$ | $f_{7}$ | $f_{8}$ | $f_{9}$ | $f_{10}$ | $f_{11}$ | $f_{12}$ ---|---|---|---|---|---|---|---|---|---|---|---|---|--- TypeScript | $-73$ | $-80$ | $-80$ | $-62$ | $-50$ | $-57$ | $-61$ | $-60$ | $-46$ | $-70$ | $-70$ | $-92$ | $-96$ Babel | $37$ | $-19$ | $-6$ | $-6$ | $-515$ | N/A | $-114$ | $-27$ | $-1$ | $-322$ | $-35$ | $-21$ | $211$ Both Babel and TypeScript were major outliers in the feature adoption rates. Table II focuses on these two repositories. Babel adopted some features well before TypeScript introduced them and TypeScript adopted all features before they were released. The behavior of TypeScript is easy to explain. A high coverage rate for unit tests means TypeScript starts adopting features as soon as they are implemented into the repository. Some features are part of the ECMAScript standard rather than TypeScript so Babel may include these features before TypeScript adds support for them. That explains why Babel adopts some features well before TypeScript. ### III-C RQ3: TypeScript Versions Figure 2: The adoption curves of different versions. Version 4.9 was released 50 days before the data collection ended (31st December 2022) so the data stops there. Adoption rates asympote to 180 projects, which is around $40\%$ of projects. Other projects jump versions, rather than adopting every version. Figure 2 shows the adoption curve of each of the TypeScript versions we pulled features from. We can see here that all versions follow a similar adoption curve, with an initial slow adoption of pre-release versions, then a rapid adoption in the three months after release, followed by slower late adoption by a small number of repositories. Roughly $1/3$ of projects ($160$ out of $454$) adopt the latest release within the first three months after release (except for TypeScript 4.9, which was released less than 50 days before data collection ended). These fast adoption curves are not surprising since JavaScript/TypeScript projects regularly update dependencies to the latest revision and TypeScript releases do not introduce significant breaking changes. Most adoption happens in the first three months after release (Roughly $35\%$ of projects in our dataset) with a small tail at the end for projects that update after a new version is already released. At the time of writing, TypeScript releases new versions every three months so some projects may not have adopted a version before the new version is released. These are results aggregated over $454$ different repositories and we do not see all repositories accounted for here. This comes down to two major reasons. While some repositories (Visual Studio Code for example) adopt new versions within a few days of release some take a few months to adopt new versions or do not adopt them at all. The adoption averages out to the same curve though. The other reason is we detected the TypeScript version using the package.json file. The maximum adoption for any version is $185$ out of $454$ ($46$ have no TypeScript version recorded). The reason is not all repositories adopt all versions of TypeScript and most skip versions as they do not regularly update. A few repositories adopted versions before they were formally released. TypeScript depends on itself but overrides that with the local version. It therefore adopts new versions before they are released. ## IV Related Work Some existing work has already investigated adoption of language features in JavaScript [2, 8, 9], Java [3, 10, 11], and Python [12, 13]. JavaScript allows for self-modifying code and code generated and evaluated at runtime. These features make tracking the control flow over a programs execution difficult so some previous works exclude them from analysis. The work by Richards et al. [2] questions this approach by looking at the prevalence of these features in production code. Due to the nature of those features most analysis there is based on dynamic analysis rather than the static analysis we use in our work. A large amount of work in this area has been done in Java [3, 10, 11]. Firstly the work by Parnin et al. [3] discovered that most uses of generics were covered by a small number of classes but the usage varies between developers. The work by Dyer et al. [10] broadened this by looking at $31,432$ Java projects on SourceForge [14] and studying the adoption of $18$ language features introduced in three versions of Java. The work by Peng et al. [12] performs a similar study to our work focusing on Python projects instead of TypeScript projects. They perform a smaller study on $35$ different projects across a range of sectors. They make the interesting observation that larger projects tend to use less involved language features like safety checks rather than more advanced features like diamond inheritance. This lines up with our outcome since the most popular features we observed increase safety and the least popular feature (static blocks in classes) can make control flow more difficult to read. Another work by Yang et al. [15] follows a similar direction to the work by Richards et al. [2] looking at the impact of dynamic features on static analysis of Python code. The work by Cristiani and Thiemann[16] includes a brief analysis of feature usage in DefinitelyTyped[17]. The work is limited to types in type declaration files rather than our study looking at TypeScript source code. The static analysis field leverages this study to inform the language features they implement support for. For instance the work by Rastogi et al.[18] seeks to improve the safety of TypeScript programs and uses a smaller subset of TypeScript called ”Safe TypeScript”. This work was done before prior to the release of TypeScript 1.1 (October 6, 2014) and lacks may of the features introduced after. In addition the work by Feldthaus and Møller[19] uses a version of the TypeScript language to detect faults in JavaScript interfaces. Like the work by Cristiani and Thiemann[16] it focuses on declaration files rather than TypeScript source code. Overall the related work covers two different kinds of study. Some work [2] uses dynamic analysis to study the prevalence of dynamic features. The other group of studies [12] look at the usage of features across different types of project. Another further field[16, 19, 18] uses static analysis to perform code analysis on TypeScript language features. Our work extends on the second field of work by looking at a series of different versions. ## V Discussion & Concluding Remarks The answer to RQ1 is that the most popular new language features are type modifiers on imports and template literal types. While type modifiers solve an existing issue of unintended side effects from imported modules, template literal types give additional flexibility in how types are constructed. The answer to RQ2 is more involved. Different features are adopted at different rates which is an expected outcome. Some features are very niche and are only used by a small number of libraries. The unexpected outcome is that adoption rates are static over time and no features sees a large initial peak as developers race to adopt them. Our interpretation of this is that very few projects need a new feature, so they are adopted as developers learn about them and gradually utilize them in new code and in code rewrites. Finally, the answer to RQ3 is straightforward. Most projects adopt new versions of TypeScript quickly with an expected long tail as remaining projects update to new versions. ### V-A Conclusions We observed a simple adoption curve for language versions, with most adoption happening shortly after release with ${1}/{3}$ of repositories updating before the next TypeScript version is released. However, the adoption of new language features into repositories is much more gradual. A project can update to a new version of TypeScript without changing their code at all, so without adopting any new features. So adopting a new language feature may require adopting a new TypeScript version, but not vice versa. We can draw the conclusion that while a project has a feature available it may not adopt it until much later. Returning to our overall goal of specifying a useful subset of TypeScript for program analysis tools we can see that although new language versions are adopted quickly by the ecosystem (${1}/{3}$ over $3$ months) the adoption of new features is a lot more variable with some features never being adopted outside of a few projects. This shows that it is important for tools to keep up to date with language versions but it is less important to support all language features (e.g. Group 2 features are used by only a few projects). ### V-B Future Work Currently our analysis focuses on syntactic changes to TypeScript, which misses improvements made to type inference and to the developer experience. It would be useful in future research to expand the list of features and look at semantic changes. Our paper focuses on the features introduced in the 4.x versions of TypeScript to make timely analysis possible. Future work could look at additional TypeScript versions. Additionally it would be interesting to run our analysis on a wider body of repositories to see how the results change with less popular projects. ## References * [1] “Typescript.” [Online]. Available: https://www.typescriptlang.org/ * Richards et al. [2010] G. Richards, S. Lebresne, B. Burg, and J. Vitek, “An Analysis of the Dynamic Behavior of JavaScript Programs,” in _Proceedings of the 31st ACM SIGPLAN Conference on Programming Language Design and Implementation_ , ser. PLDI ’10. New York, NY, USA: Association for Computing Machinery, 2010, p. 1–12. [Online]. Available: https://doi.org/10.1145/1806596.1806598 * Parnin et al. [2011] C. Parnin, C. Bird, and E. Murphy-Hill, “Java Generics Adoption: How New Features Are Introduced, Championed, or Ignored,” in _Proceedings of the 8th Working Conference on Mining Software Repositories_ , ser. MSR ’11. New York, NY, USA: Association for Computing Machinery, 2011, p. 3–12. [Online]. Available: https://doi.org/10.1145/1985441.1985446 * [4] “Github.” [Online]. Available: https://github.com/ * [5] “Golang.” [Online]. Available: https://go.dev/ * [6] “Babel.” [Online]. Available: https://babeljs.io/ * [7] ECMA TC39, “Ecmascript.” [Online]. Available: https://www.ecma-international.org/publications-and-standards/standards/ecma-262/ * Wei et al. [2016] S. Wei, F. Xhakaj, and B. G. Ryder, “Empirical study of the dynamic behavior of JavaScript objects,” _Software: Practice and Experience_ , vol. 46, no. 7, pp. 867–889, 2016. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/spe.2334 * Richards et al. [2011] G. Richards, C. Hammer, B. Burg, and J. Vitek, “The Eval That Men Do,” in _ECOOP 2011 – Object-Oriented Programming_ , M. Mezini, Ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 2011, pp. 52–78. * Dyer et al. [2014] R. Dyer, H. Rajan, H. A. Nguyen, and T. N. Nguyen, “Mining Billions of AST Nodes to Study Actual and Potential Usage of Java Language Features,” in _Proceedings of the 36th International Conference on Software Engineering_ , ser. ICSE 2014. New York, NY, USA: Association for Computing Machinery, 2014, p. 779–790. [Online]. Available: https://doi.org/10.1145/2568225.2568295 * Tempero et al. [2008] E. Tempero, J. Noble, and H. Melton, “How Do Java Programs Use Inheritance? An Empirical Study of Inheritance in Java Software,” in _ECOOP 2008 – Object-Oriented Programming_ , J. Vitek, Ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 2008, pp. 667–691. * Peng et al. [2021] Y. Peng, Y. Zhang, and M. Hu, “An Empirical Study for Common Language Features Used in Python Projects,” in _2021 IEEE International Conference on Software Analysis, Evolution and Reengineering (SANER)_ , 2021, pp. 24–35. * Åkerblom et al. [2014] B. Åkerblom, J. Stendahl, M. Tumlin, and T. Wrigstad, “Tracing Dynamic Features in Python Programs,” in _Proceedings of the 11th Working Conference on Mining Software Repositories_ , ser. MSR 2014. New York, NY, USA: Association for Computing Machinery, 2014, p. 292–295. [Online]. Available: https://doi.org/10.1145/2597073.2597103 * [14] “Sourceforge.” [Online]. Available: https://sourceforge.net/ * Yang et al. [2022] Y. Yang, A. Milanova, and M. Hirzel, “Complex Python Features in the Wild,” in _Proceedings of the 19th International Conference on Mining Software Repositories_ , ser. MSR ’22. New York, NY, USA: Association for Computing Machinery, 2022, p. 282–293. [Online]. Available: https://doi.org/10.1145/3524842.3528467 * Cristiani and Thiemann [2021] F. Cristiani and P. Thiemann, “Generation of typescript declaration files from javascript code,” in _Proceedings of the 18th ACM SIGPLAN International Conference on Managed Programming Languages and Runtimes_ , ser. MPLR 2021\. New York, NY, USA: Association for Computing Machinery, 2021, p. 97–112. [Online]. Available: https://doi.org/10.1145/3475738.3480941 * [17] “Definitelytyped.” [Online]. Available: http://definitelytyped.org/ * Rastogi et al. [2015] A. Rastogi, N. Swamy, C. Fournet, G. Bierman, and P. Vekris, “Safe & efficient gradual typing for typescript,” in _Proceedings of the 42nd Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages_ , ser. POPL ’15. New York, NY, USA: Association for Computing Machinery, 2015, p. 167–180. [Online]. Available: https://doi.org/10.1145/2676726.2676971 * Feldthaus and Møller [2014] A. Feldthaus and A. Møller, “Checking correctness of typescript interfaces for javascript libraries,” _SIGPLAN Not._ , vol. 49, no. 10, p. 1–16, oct 2014. [Online]. Available: https://doi.org/10.1145/2714064.2660215
# Walking for short distances and turning in lower-limb amputees: a study in low-cost prosthesis users Nidhi Seethapathi1,3, Anil Kumar Jain2 and Manoj Srinivasan1 1 Mechanical and Aerospace Engineering, The Ohio State University, Columbus OH 43210, USA 2Santokba Durlabhji Memorial Hospital, Jaipur, Rajasthan 302015, India 3 Department of Bioengineering, University of Pennsylvania, Philadelphia PA 19104, USA ###### Abstract Preferred walking speed is a widely-used performance measure for people with mobility issues. Often these speeds are measured walking in a straight line and walking over short distances. However, daily walking involves walking for bouts of different distances and walking with turning. Here, we observe walking for short distances and walking in circles in subjects with unilateral lower-limb amputation using a Jaipur Foot prosthetic leg. We find that the preferred walking speeds are lower for unilateral amputees for short distances, but the distance-dependence of preferred walking speed is less steep than for non-amputees. In circle walking, unilateral amputees slow down when walking in circles of smaller radii and above-knee amputees in particular walk faster when the prosthetic leg is outside the circle. Using simple mathematical models, we show that the observed preferred walking speeds are predicted by minimal energy cost with increased costs for walking, changing speeds, and turning for amputees compared to non-amputees. These results predict that the cost of changing speeds and turning may be substantial in amputees but still not as high as the constant-speed cost. These findings will inform prosthesis design and rehabilitation therapy to better assist changing speeds and turning tasks in amputee walking. ## 1 Introduction Preferred overground walking speed is commonly used to quantify a human subject’s mobility improvement after being fit with a new prosthetic leg or after undergoing physical therapy or rehabilitation from stroke, other injury, or movement disorder [1, 2]. Such walking speeds are estimated using a variety of tests in the lab, most commonly using the 6 minute walk test [3] or the 10 m walk test [4], but also by having subjects walk other short distances such as 3 m [5], 4 m [6], 5 m [7], and 15 m [8, 9]. In healthy adults with no movement disorders, the preferred walking speed for walking in a straight line is distance-dependent [10]: the speed is systematically lower for shorter distances and this distance-dependence can be explained by the larger energetic cost of speeding up and slowing down for shorter distances [10]. Here, we propose this distance dependence of walking speed as a more complete measure of preferred walking speed (than measuring the speed at just one distance) and characterize it in unilateral amputees. This distance dependence of walking speed can also help shed light on daily walking behavior in humans, in which a considerable percentage of walking occurs in short bouts [11], especially in amputees [12]. Another important aspect of daily walking behavior is that people often need to walk along curves or with some turning. Indeed, for subjects in one previous study, between 8% and 50% of all walking steps in daily life involved turning [13]. So effective mobility requires ability to walk with turning. As a way of quantifying turning ability and as an additional measure of the effectiveness of a prosthetic leg or a physio-therapeutic intervention, here we propose the measurement of preferred speeds while walking in circles of different radii and characterize such walking speeds in unilateral amputees. Although such curved walking interventions have been used for other populations [14, 15, 16], there is less literature on the subject for amputee- walking. In healthy non-amputee adults, the tangential speed of walking depends on the curvature of the circle walked: slower walking for smaller circles or higher curvature, explained by the increased energetic cost of walking with turning [17, 18]. Here, we test the hypothesis that walking speeds depend on circle radius in unilateral amputees and whether the prosthesis foot being the outer or inner foot affects circle-walking speeds. Metabolic energy optimality has been used to make predictions for a number or aspects of overground and treadmill walking behaviors in healthy non-amputee adults [19, 20, 21]. Indeed, as noted above, the human preference slower walking speeds for shorter distances and slower walking speeds for smaller radii and higher curvatures have been attributed to optimization of the corresponding energy costs [10, 17, 18]. However, there is less work making such energy optimality based predictions for other population with amputations or other movement disorders [22, 23, 24]. Here, we examine this energy minimization hypothesis in the context of amputee walking for short distances and in curves. Amputees walking with passive prosthetic legs usually have a higher steady state metabolic cost [25] compared to non-amputees (but see recent work that find no significant cost difference in some amputee sub-populations [26]). However, it is not known whether amputees also have a higher cost of speeding up and slowing down when walking. The interactions between the walking cost and the changing speed costs may decide how amputees walk over short distances. The higher the cost of speeding up and slowing down, the greater it will affect the optimal walking speed at short distances. Similarly, while it is known that walking along curves costs more energy than walking in a straight line in non-amputee adults [17, 18], the energy cost of turning has not been characterized in amputees. Thus, by viewing preferred walking behavior in amputees through the lens of energy minimization, these short distance and circle walking tasks provide insight into how their cost of changing speeds and the cost of turning may compare to that of non-amputee individuals. Here, we studied a small population of unilateral amputees, both above- and below- knee, using the Jaipur foot prosthesis, a low cost prosthesis used widely in the developing world [27]. This prosthesis, developed by P. K. Sethi and co-workers [27], was designed for facilitating practices common in India, such as bare foot or shod walking over unpaved uneven terrain and squatting or cross-legged sitting on the floor. The Jaipur foot is used in over 22 countries and by hundreds of thousands of amputees, most widely used second only to the SACH foot [28]. Thus, this study adds to the small number of biomechanical studies (e.g., [29, 30]) on the Jaipur foot, which is under- studied despite being so widely used. ## 2 Methods ### Subject population. The experimental protocol was approved by the Ohio State University Institution Review Board and all subjects participated with informed verbal consent. All subjects ($N=12$ with 11 males, 1 female, $65.75\pm 12.6$ kg with prosthesis and shoes, height $1.67\pm 0.09$ meters and age $39.67\pm 15.31$ years, mean $\pm$ s.d.) were unilateral amputees, out of which 7 subjects were above-knee amputees and 5 were below-knee amputees. All subjects had a Jaipur Foot prosthesis [27], manufactured and fit in the SDMH hospital in Jaipur; all walking trials were also conducted at this location. Subjects walked independently without using canes, crutches, hand rails, or other assistive devices. The subjects did not carry any additional instrumentation. The subjects performed two kinds of walking trials: (i) walking for short distances and (ii) walking in circles, as described below (Figure 1). ### Walking for short distances. Subjects were instructed to walk in a straight-line for five different short distances: 4 m, 6 m, 8 m, 10 m and 23 m (Figure 1a). There were four trials for each distance and trial order was randomized. Subjects were asked to “walk the way they usually walk” and they had to start and end each trial standing still, so they had to speed up from and slow down to rest. Average walking speeds were estimated by measuring the time duration for each trial. Figure 1: Overground walking experiment setup. We measured the preferred walking speed of walking for unilateral amputees wearing a passive prosthetic leg in two conditions: a) walking a range of short distances, starting and stopping each bout at rest and b) walking in circles of different radii, both clockwise and anti-clockwise. ### Walking in circles. Subjects were asked to walk in circles of three different radii: 1 m, 2 m and 3 m (Figure 1b), completing 5, 4 and 3 laps, respectively, for these radii. For each radius, subjects performed two trials, once with the prosthetic leg inside the perimeter and once with the prosthetic leg outside the perimeter of the circle. Trial order was randomized over the circle radii and walking directions. The average speed was obtained by measuring the total walking duration and averaging over all laps for each trial. Subjects walked with the circle between their two feet, maintaining a non-zero step width, rather than step on the circle with both feet. ### Mathematical model: walking for short distances. For short distance walking, we compare the experimentally observed preferred walking speed results to the walking speed predicted by minimizing the total metabolic cost of the walking bout. For simplicity, we assume that people walking a distance $D$ start from rest (specified in experiment), then instantaneously speed up to some speed $v$, continue at that speed for the whole distance and then instantaneously come to rest again. Thus, the total cost of walking the distance includes the cost of accelerating from rest to speed $v$ at the start, walking at constant speed $v$ and then decelerating to rest at the end of the walking bout, given by the following equation (analogous to the approach in [10]): $E_{\mathrm{met}}=(a_{0}+a_{1}v+a_{2}v^{2})\frac{D}{v}+\lambda\left(\frac{1}{\eta_{\mathrm{pos}}}+\frac{1}{\eta_{\mathrm{neg}}}\right)\left(\frac{1}{2}mv^{2}\right).$ (1) Here, $\dot{E}=a_{0}+a_{1}v+a_{2}v^{2}$ in Wkg-1, with $v$ in ms-1 models the metabolic rate of walking $\dot{E}$ at a constant speed $v$ for both amputees and non-amputees, with $a_{0}=4.97$, $a_{1}=-5.78$, and $a_{2}=5.62$ for above-knee amputees [31], $a_{0}=3.24$, $a_{1}=-2.19$, and $a_{2}=2.89$ for below-knee amputees [32], and $a_{0}=2.22$, $a_{1}=0$, and $a_{2}=1.155$ for non-amputees [31]. All of these relations for $\dot{E}$ result in a classical U-shaped relationship between energy cost per unit distance and speed of walking. The quantity $\lambda$ provides a scaling factor between the kinetic energy $mv^{2}/2$ and the energy cost required to achieve it, with $\eta_{\mathrm{pos}}=0.25$ and $\eta_{\mathrm{neg}}=1.2$ are traditional muscle efficiencies for performing positive and negative mechanical work, respectively [33]. Because we did not directly estimate the cost of changing speeds in amputees here (as in [10]), we tested different $\lambda$ values to best explain the observed walking behavior. The speed that minimizes the short-distance cost of walking $E_{\mathrm{met}}$ is given by the implicit equation: $\lambda v^{3}\left({\eta_{\mathrm{pos}}}^{-1}+{\eta_{\mathrm{neg}}}^{-1}\right)/(a_{0}-a_{2}v^{2})=D$. This relation implies that shorter distance bouts should have lower speeds. ### Mathematical model: Walking in circles. Following [17, 18], who directly measured the cost for walking in circles for non-amputee subjects, we analogously propose that the metabolic rate of walking in a circle for amputees is $\dot{E}=a_{0}+a_{1}v+a_{2}v^{2}+b_{2}(v/R)^{2}$, where $R$ is the circle radius, with $a_{0,1,2}$ values as above. The cost per distance is given by $\dot{E}/v=a_{0}/v+a_{1}+a_{2}v+b_{2}v/R^{2}$, which is minimized by the speed $v=\sqrt{a_{0}/(a_{2}+b_{2}/R^{2})}$. Thus, the prediction is that the optimal speed is smaller for smaller radius $R$. In the following, we chose $b_{2}$ to best explain the speed reduction exhibited by our amputee subjects. For both short-distance walking and circle walking, we also determine the speeds that are within 1% of the minimum energy cost, because the energy landscapes are usually flat and a small change in speed near the minimum energy usually results in a much smaller energy change. ## 3 Results ### Preferred walking speed for unilateral amputees depends on distance walked. Both above- and below-knee amputees showed, on average, a decrease in preferred walking speed for short distances. Pooled across all above and below knee subjects, the preferred walking speed for each of the short distances (4 m, 6 m, 8 m and 10 m) was significantly lower than the preferred walking speed for the long-distance 23 m trial ($p<0.0025$ for each distance, left-tailed paired $t$-test). The percentage decreases in the amputees’ preferred walking speeds, compared to the long distance 23 m trial, are shown in Figure 2a. The percentage decrease is higher for shorter distances: in above knee amputees, the percentage decreases ranged from 7% for the 10 m walk to 13% for the 4 m walk, on average. In below knee amputees, the percentage decreases ranged from 7% for the 10 m walk to 18% for the 4 m walk. ### The rate of speed decrease with distance is lower for amputees compared to non-amputees. Although we find that the preferred walking speed for amputees decreases with distance walked, the rate of decrease is flatter compared to that for non- amputee individuals (data repeated from [10]). As seen in Figure 2b, the best- fit slope of the distance dependence of speed for non-amputee subjects is about three times steeper than that for the above-knee amputee subjects and about 1.5 times that for the below-knee amputees. Thus, the above-knee amputees slowed down by a much smaller percentage for short-distance walking trials, compared to both below knee amputees and non-amputee subjects. Figure 2: Decrease in preferred walking speed with distance walked for amputees. a) Amputees showed a greater decrease in preferred walking speed for short distances speed compared to able-bodied individuals. b) The rate of change in preferred walking speed with distance is steeper for able-bodied individuals than for the unilateral amputees. ### Increased cost of walking and changing speeds is consistent with flatter speed-distance relationships. Our simple optimization-based model of short distance walking predicts a flatter distance-dependence of optimal walking speed for both above and below knee amputees (Figure 2,3), when we take into account the increased constant- speed metabolic cost of walking previously measured in experiments [31, 32] and an increased cost of changing speeds compared to able-bodied subjects [10]. Compared to non-amputees, the scaling factor $\lambda$ for the cost of changing speeds was increased by a factor of 1.87 for below-knee amputees and 1.90 above-knee amputees. Nevertheless, mechanistically, our model suggests that the amputees do not slow down as much for short distances because the cost of moving at a reduced speed for the whole distance outweighs the cost of changing speeds from rest. As seen in Figure 3, the mean amputee preferred walking speeds plus one standard error is within 1% of the optimal energy costs from this model. Figure 3: Minimization of total metabolic cost captures slower short-distance walking speeds. The total cost of the walking a short distance includes a term due to constant-speed cost and a changing-speed cost. We find that minimizing this total cost predicts the observed trends in changing preferred walking speed with distance for both amputees and non-amputees. The error bars for human data represent standard errors, and the filled bands represent the set of all speeds within 1% of the energy optimal energy cost. ### Preferred walking speeds for unilateral amputees walking in circles depends on the circle radii. We find that, irrespective of type of amputation, all the unilateral amputees that we observed slowed down when walking along circles of smaller radii (Figure 5), compared to walking a longer distance in a straight line. The slope of decrease in preferred walking speed with radius of circle walked, is quite similar for the able-bodied (0.1 s-1) and amputee populations (0.08 s-1). Using our optimization-based model to fit this decreasing speed trend, the best-fit scaling coefficient $b_{2}$ for the cost of turning was about six times more than non-amputees for below-knee amputees and eight times more than non-amputees for above-knee amputees. Figure 4: Preferred walking speeds for circle walking. a) The preferred walking speed for all the unilateral amputees showed a decrease with radius of the circle walked. b) Amputees, when pooled together, did not show a significant difference in preferred walking speed when walking with the prosthesis-leg inside versus outside the circle. c) Above knee amputees show a greater walking speed on average when the prosthesis leg is outside the circle. Figure 5: Optimal walking speeds for circle walking. Minimizing the energy cost of walking in a circle predicts slower walking for smaller circles. Error bars shown for the data correspond to one standard error about the mean, and these are generally within the set of all speeds within 1% of the optimal energy costs (the shaded bands shown, as in Figure 3. The non- amputee data and model used for comparison is from [17, 18]. ### Preferred walking speeds for above-knee amputees walking in circles is dependent on turning direction. We had subjects walk both clockwise and anti-clockwise along circles drawn on the ground. We did this so as to check for any effects due to having the prosthesis-leg as the pivot, as opposed to the intact leg as the pivot. Considering all amputees together, we did not find significant differences between the two conditions ($p=0.096$). Just considering the above-knee amputees, with trials for all radii pooled, we found that they walked slightly faster when the prosthetic leg was outside the circle ($p=0.0008$, mean difference about 0.03 m/s on average, median 0.04 m/s). ### Preferred walking speeds are proportional to degree of amputation. For all the cases above, that is, straight-line walking for different distances and circle-walking at different radii, we found that the preferred walking speeds for the below-knee amputees are higher than above-knee and lower than non-amputee preferred speeds (Figures 2-3). This result for straight-line long-distance walking over long distances is known from other work in the past [34]. However, the result for short-distance and circle- walking is new, even if unsurprising. In addition, we found, as have others in the past , that the preferred walking speed for the below-knee unilateral amputees were higher than those for the above-knee unilateral amputees. ### Preferred gait initiation swing is usually with the affected limb. In addition to measuring the preferred walking speeds, we noted whether the subjects stepped forward with their affected or unaffected limb for their very first step. Stepping forward with the affected limb corresponds to the first swing phase being with the affected limb and the first stance phase being with the unaffected limb. We found that 9 out of 12 subjects had over 80% of their first steps be their prosthetic foot; the other three subjects had 69%, 36%, and 0% of their steps start with swinging the affected limb. These leading limb preferences are similar to those found in [35]. ## 4 Discussion Preferred walking speed is often used as a measure of progress in walking rehabilitation for various populations, for instance persons with neuromuscular disorders and amputees. An implicit assumption made when relying on such measures is that higher speeds means more improvement. However, past theoretical and experimental work on able-bodied subjects shows that the speed at which people choose to move depends on the constraints of the motion itself like the distance walked [10] and the curvature of the motion [17, 18]. So, depending on the situation, people may sometimes move at a lower speed than physically possible to satisfy some other objective, like minimizing energy. Here, we find that these observations extend to amputee populations as well and may have significant implications for the usage of preferred walking speeds as a measure of performance. These implications are detailed in the paragraphs below. We find that above-knee and below-knee amputees slow down when walking short distances, as do non-amputees. This implies that the distance over which the preferred walking speed is measured and interpreted during rehabilitation may systematically overestimate or underestimate the progress that the patient has made. Many studies about measure progress in people with walking disorders after rehabilitation, measure the preferred walking speed over very short distances. In order to circumvent these distance-effects, we suggest measuring the speed over a few distances, not just one or two. Also, when comparing the preferred speed values for amputees to the able-bodied values, we suggest comparing to the values for the same distance walked. This is because the difference in walking speed between able-bodied and above-knee amputees over short distances is lesser than the corresponding difference over longer distances. An alternative to in-lab measurements of preferred speeds is to track subjects’ speeds and movements all day using body-worn sensors such as pedometers, IMUs, GPS, etc [36, 37, 38]. Ultimately, it is these speeds during daily living that is of relevance to quantifying mobility. Such ambulatory measurements provide an opportunity to independently corroborate the results in this study, by characterizing the speeds over bouts of different lengths and walks with turns that naturally occur during daily life. Everyday walking consists of not just straight line walking but also turning. Here, we find that unilateral amputees also slow down when taking sharp turns (circles of smaller radii) similar to non-amputees [17, 18]. We also find that the corresponding walking speeds are lower for the amputees for all radii, compared to those for able-bodied individuals. For all these reasons, we propose that circle-walking lends itself as an additional useful measure of walking performance during rehabilitation. Unlike the straight line walking trials, we were not interested in capturing the effects of speeding up from and slowing down to rest for circle walking. We believe that the multiple laps used makes such distance-dependence effects negligible. In the past, people have predicted aspects of walking behavior such as walking speed, walking step width and walking step frequency using energy minimization as a hypothesis. However, most of these studies look at constant-speed straight line walking on treadmills. Moreover, a majority of these studies attempt to predict non-amputee walking behavior. Here, we provide evidence that energy-minimization can predict aspects of non-steady overground walking behavior in a disabled population. Specifically, we have found that minimizing a combined cost of constant-speed walking and a cost for changing speeds captures the short-distance walking speed trends in amputees; similarly, adding a cost for turning captures the circle-walking speed trends. We have studied the preferred walking speeds of unilateral amputees wearing a passive prosthetic leg (Jaipur Foot) used in a number of developing countries. Using preferred walking speeds as a performance measure may be more relevant where the resources available are limited, where access to other measures of performance, such as using a gait lab with motion capture and force plates, may be limited. Thus, we feel that our conclusions regarding the qualification of preferred walking speeds as a performance measure would be more relevant where walking-speed-based mobility measures would be more exclusively used. Fitting the short-distance walking model and circle-walking model to the amputee walking speed behavior, we found that the scaling factors for the cost of changing speeds and for the cost of turning ($\lambda$ and $b_{2}$) are much higher than for non-amputees. To test whether these increased cost estimates, one could directly measure the metabolic cost of changing speed and the cost of turning, as in [10, 17, 18]. Stability considerations may be an alternative to metabolic cost being the determinant of reduced speeds, especially for turning in a circle. One could examine this alternative hypothesis by having subjects use different speeds and estimating simple measures of stability [39, 40]. For short-distance walking, we measured the average speed over the whole bout of the amputees. This average speed includes the acceleration and deceleration periods. So the reduction in average speed is partly due to a greater portion of the bout being spent in acceleration-deceleration and partly due to reduced walking speed. This was the case in the earlier non-amputee study as well [10]. In that study, walking slower for shorter distances was a real choice, as the subjects could certainly cover the distance in shorter time. While we did not test this hypothesis here by asking the amputees to walk faster than they preferred to do, we suggest, based on prior studies with amputee populations, that they can typically walk much faster than their preferred speeds (e.g., [41, 31]). Specifically, we do not know of any demonstrated instance in which someone’s preferred speed is also their maximum possible speed, although it may be possible. One limitation of this study is the relatively small sample size of subjects. The experiments were conducted in a hospital in India and not all the patients had the time or resources to participate in the study. Small sample sizes are common with such amputee populations. Despite the small sample size, the qualitative results for the amputees are robust. A follow-up study will be conducted to investigate the population with a larger sample size. Another potential limitation is that we did not limit our amputee sample by years since amputation and age. However, we believe that this limitation is offset by the fact that we focus here on the change in preferred speeds of each subject under different conditions relative to his/her own long-distance straight-line walking speed. So, the individual differences in terms of years of being accustomed to wearing the prosthesis will not affect our results. We compare the non-amputee walking speeds of a diverse group of college or graduate student-age subjects [10, 17, 18] in the USA to those of Indian amputees. It would perhaps be more appropriate to compare with age and size- matched Indian non-amputee adults. However, past studies of variations in walking speed due to location or country are much lesser than the differences between non-amputee and amputee populations that we observe here [42]. So, our fundamental findings will not change even with the more appropriate Indian able-bodied population. It would also be useful to repeat these experiments in other subject populations, including amputees wearing other prostheses with different mechanical properties. The Jaipur foot is a unique prosthesis, designed for Indian life and has features that make it mechanically different compared to other prostheses. Specifically, it has more mobility in the ankle and subtalar joint to allow squatting, sitting cross legged, and walking over uneven unpaved surfaces. It has three main pieces: a heel block and fore-foot-toe block made out of micro-cellular rubber bonded to a wooden ankle block. Tread rubber is used on the undersurface to provide traction and the entire assembly is covered in a skin colored compound to provide cosmesis and water proofing. We speculate that these construction features give more compliance, allowing its users to perform more three dimensional tasks; future work should consider more detailed biomechanical analyses for such 3D tasks. Finally, given the simplicity of these tasks, we propose that these distance- dependence of walking speeds and radius-dependence of walking speeds be used as routine measures of mobility not just in amputees, but also other subject populations such as the elderly and those with or recovering from other movement disorders. ### Acknowledgments. The authors thank Dr. Harlal Singh Mali for hosting N.S. in his lab briefly during her visit to Jaipur and for initially facilitating interactions. ### Ethics Statement All subjects participated with informed consent and the experimental protocol was approved by the the Ohio State University Institution Review Board. ### Funding Statement. For this study, N.S. received funding from Schlumberger Foundation Faculty for the Future Fellowship, The Ohio State University Global Gateway Grant and the Alumni Grant for Graduate Research and Scholarship. MS was supported in part by NSF CMMI grant . ### Data Accessibility. All data and codes form will be made available upon acceptance. ### Competing Interests. We have no competing interests. ### Authors’ Contributions. N.S. conceived the study, collected the data, analyzed the results, performed the mathematical analyses, and wrote the paper. A.J. provided guidance for data collection and edited the draft. M.S. provided ideas for designing the study, suggested additional analyses, and edited the draft. ## References * [1] Richard W Bohannon. Comfortable and maximum walking speed of adults aged 20?79 years: reference values and determinants. Age and ageing, 26(1):15–19, 1997. * [2] AM Boonstra, V Fidler, and WH Eisma. Walking speed of normal subjects and amputees: aspects of validity of gait analysis. Prosthetics and Orthotics International, 17(2):78–82, 1993. * [3] Nancy D Harada, Vicki Chiu, and Anita L Stewart. Mobility-related function in older adults: assessment with a 6-minute walk test. Archives of physical medicine and rehabilitation, 80(7):837–841, 1999. * [4] S Amatachaya, S Naewla, K Srisim, P Arrayawichanon, and W Siritaratiwat. Concurrent validity of the 10-meter walk test as compared with the 6-minute walk test in patients with spinal cord injury at various levels of ability. Spinal Cord, 52(4):333, 2014. * [5] James E Graham, Glenn V Ostir, Steven R Fisher, and Kenneth J Ottenbacher. Assessing walking speed in clinical research: a systematic review. Journal of evaluation in clinical practice, 14(4):552–562, 2008\. * [6] Denise M Peters, Stacy L Fritz, and Debra E Krotish. Assessing the reliability and validity of a shorter walk test compared with the 10-meter walk test for measurements of gait speed in healthy, older adults. Journal of geriatric physical therapy, 36(1):24–30, 2013. * [7] Christopher M Wilson, Stephanie R Kostsuca, and Judith A Boura. Utilization of a 5-meter walk test in evaluating self-selected gait speed during preoperative screening of patients scheduled for cardiac surgery. Cardiopulmonary physical therapy journal, 24(3):36, 2013. * [8] Bruce H Dobkin. Short-distance walking speed and timed walking distance: redundant measures for clinical trials? Neurology, 66(4):584–586, 2006. * [9] Ruth Dickstein. Rehabilitation of gait speed after stroke: a critical review of intervention approaches. Neurorehabilitation and neural repair, 22(6):649–660, 2008. * [10] Nidhi Seethapathi and Manoj Srinivasan. The metabolic cost of changing walking speeds is significant, implies lower optimal speeds for shorter distances, and increases daily energy estimates. Biology letters, 11(9):20150486, 2015. * [11] Michael S Orendurff, Jason A Schoen, Greta C Bernatz, Ava D Segal, and Glenn K Klute. How humans walk: bout duration, steps per bout, and rest duration. Journal of Rehabilitation Research & Development, 45(7), 2008. * [12] Glenn K Klute, Jocelyn S Berge, Michael S Orendurff, Rhonda M Williams, and Joseph M Czerniecki. Prosthetic intervention effects on activity of lower-extremity amputees. Archives of physical medicine and rehabilitation, 87(5):717–722, 2006. * [13] Brian C Glaister, Greta C Bernatz, Glenn K Klute, and Michael S Orendurff. Video task analysis of turning during activities of daily living. Gait & posture, 25(2):289–294, 2007. * [14] Kristin A Lowry, Jennifer S Brach, Robert D Nebes, Stephanie A Studenski, and Jessie M VanSwearingen. Contributions of cognitive function to straight-and curved-path walking in older adults. Archives of physical medicine and rehabilitation, 93(5):802–807, 2012. * [15] Marco Godi, Antonio Nardone, and Marco Schieppati. Curved walking in hemiparetic patients. Journal of rehabilitation medicine, 42(9):858–865, 2010. * [16] Rebecca J Hess, Jennifer S Brach, Sara R Piva, and Jessie M VanSwearingen. Walking skill can be assessed in older adults: validity of the figure-of-8 walk test. Physical therapy, 90(1):89–99, 2010. * [17] G. L. Brown. Nonlinear Locomotion: Mechanics, energetics, and optimality of walking in circles and other curved paths. PhD thesis, Ohio State University, 2012. * [18] G. L. Brown and M. Srinivasan. Walking with turning: Metabolic energy, energy optimality, and preferred speeds while walking in circles, walking on complex paths, and path planning. submitted, in revision, 2019. * [19] MY Zarrugh, FN Todd, and HJ Ralston. Optimization of energy expenditure during level walking. European journal of applied physiology and occupational physiology, 33(4):293–306, 1974. * [20] Manoj Srinivasan. Optimal speeds for walking and running, and walking on a moving walkway. Chaos: An Interdisciplinary Journal of Nonlinear Science, 19(2):026112, 2009. * [21] J Maxwell Donelan, Rodger Kram, et al. Mechanical and metabolic determinants of the preferred step width in human walking. Proceedings of the Royal Society of London B: Biological Sciences, 268(1480):1985–1992, 2001. * [22] James M Finley, Amy J Bastian, and Jinger S Gottschall. Learning to be economical: the energy cost of walking tracks motor adaptation. The Journal of physiology, 591(4):1081–1095, 2013. * [23] Nicholas P Fey, Glenn K Klute, and Richard R Neptune. Optimization of prosthetic foot stiffness to reduce metabolic cost and intact knee loading during below-knee amputee walking: a theoretical study. Journal of biomechanical engineering, 134(11):111005, 2012. * [24] Matthew L Handford and Manoj Srinivasan. Robotic lower limb prosthesis design through simultaneous computer optimizations of human and prosthesis costs. Scientific reports, 6:19983, 2016. * [25] RL Waters, Jacquelin Perry, DANIEL Antonelli, and Helen Hislop. Energy cost of walking of amputees: the influence of level of amputation. J Bone Joint Surg Am, 58(1):42–46, 1976. * [26] Elizabeth Russell Esposito, Kelly M Rodriguez, Christopher A Ràbago, and Jason M Wilken. Does unilateral transtibial amputation lead to greater metabolic demand during walking. J Rehabil Res Dev, 51(8):1287–96, 2014. * [27] PK Sethi, MP Udawat, SC Kasliwal, and R Chandra. Vulcanized rubber foot for lower limb amputees. Prosthetics and orthotics international, 2(3):125–136, 1978. * [28] Peter Howitt, Ara Darzi, Guang-Zhong Yang, Hutan Ashrafian, Rifat Atun, James Barlow, Alex Blakemore, Anthony MJ Bull, Josip Car, Lesong Conteh, et al. Technologies for global health. The Lancet, 380(9840):507–535, 2012. * [29] AP Arya, A Lees, HC Nerula, and L Klenerman. A biomechanical comparison of the sach, seattle and jaipur feet using ground reaction forces. Prosthetics and Orthotics International, 19(1):37–45, 1995. * [30] P Lenka and R Kumar. Gait comparisons of trans tibial amputees with six different prosthetic feet in developing countries. IPJMR, pages 8–14, 2010. * [31] S. M. H. J. Jaegers, L. D. W. Vos, P. Rispens, and A. L. Hof. The relationship between comfortable and most metabolically efficient walking speed in persons with unilateral above-knee amputation. Arch Phys Med Rehabil, 74:521–525, 1993. * [32] J. J. Genin, G. J. Bastien, B. Franck, C. Detrembleur, and P. A. Willems. Effect of speed on the energy cost of walking in unilateral traumatic lower limb amputees. Eur J Appl Physiol, 103:655–663, 2008. * [33] Manoj Srinivasan. Fifteen observations on the structure of energy-minimizing gaits in many simple biped models. Journal of The Royal Society Interface, 8(54):74–98, 2010. * [34] Sarah J Mattes, Philip E Martin, and Todd D Royer. Walking symmetry and energy cost in persons with unilateral transtibial amputations: matching prosthetic and intact limb inertial properties. Archives of physical medicine and rehabilitation, 81(5):561–568, 2000. * [35] Aline H Vrieling, HG Van Keeken, T Schoppen, E Otten, JPK Halbertsma, AL Hof, and K Postema. Gait initiation in lower limb amputees. Gait & posture, 27(3):423–430, 2008. * [36] Nancy L Dudek, Omar D Khan, Edward D Lemaire, Meridith B Marks, and Leyana Saville. Ambulation monitoring of transtibial amputation subjects with patient activity monitor versus pedometer. Journal of Rehabilitation Research & Development, 45(4), 2008. * [37] John R Rebula, Lauro V Ojeda, Peter G Adamczyk, and Arthur D Kuo. Measurement of foot placement and its variability with inertial sensors. Gait & posture, 38(4):974–980, 2013. * [38] Brenton Hordacre, Christopher Barr, and Maria Crotty. Use of an activity monitor and gps device to assess community activity and participation in transtibial amputees. Sensors, 14(4):5845–5859, 2014. * [39] Yang Wang and Manoj Srinivasan. Stepping in the direction of the fall: the next foot placement can be predicted from current upper body state in steady-state walking. Biology letters, 10(9):20140405, 2014. * [40] Nidhi Seethapathi and Manoj Srinivasan. Step-to-step variations in human running reveal how humans run without falling. eLife, 8:e38371, 2019. * [41] Henry J Ralston. Energy-speed relation and optimal speed during level walking. Internationale Zeitschrift für Angewandte Physiologie Einschliesslich Arbeitsphysiologie, 17(4):277–283, 1958. * [42] Marc H Bornstein and Helen G Bornstein. The pace of life. Nature, 259(5544):557, 1976.